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/blk-mq.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/audit.h>
82 #include <linux/security.h>
83 #include <linux/xattr.h>
85 #define CREATE_TRACE_POINTS
86 #include <trace/events/io_uring.h>
88 #include <uapi/linux/io_uring.h>
93 #define IORING_MAX_ENTRIES 32768
94 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
98 #define IORING_MAX_FIXED_FILES (1U << 20)
99 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
100 IORING_REGISTER_LAST + IORING_OP_LAST)
102 #define IO_RSRC_TAG_TABLE_SHIFT (PAGE_SHIFT - 3)
103 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
104 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
106 #define IORING_MAX_REG_BUFFERS (1U << 14)
108 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
109 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
111 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
112 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
114 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
115 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
118 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
121 #define IO_APOLL_MULTI_POLLED (REQ_F_APOLL_MULTISHOT | REQ_F_POLLED)
123 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
126 u32 head ____cacheline_aligned_in_smp
;
127 u32 tail ____cacheline_aligned_in_smp
;
131 * This data is shared with the application through the mmap at offsets
132 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
134 * The offsets to the member fields are published through struct
135 * io_sqring_offsets when calling io_uring_setup.
139 * Head and tail offsets into the ring; the offsets need to be
140 * masked to get valid indices.
142 * The kernel controls head of the sq ring and the tail of the cq ring,
143 * and the application controls tail of the sq ring and the head of the
146 struct io_uring sq
, cq
;
148 * Bitmasks to apply to head and tail offsets (constant, equals
151 u32 sq_ring_mask
, cq_ring_mask
;
152 /* Ring sizes (constant, power of 2) */
153 u32 sq_ring_entries
, cq_ring_entries
;
155 * Number of invalid entries dropped by the kernel due to
156 * invalid index stored in array
158 * Written by the kernel, shouldn't be modified by the
159 * application (i.e. get number of "new events" by comparing to
162 * After a new SQ head value was read by the application this
163 * counter includes all submissions that were dropped reaching
164 * the new SQ head (and possibly more).
170 * Written by the kernel, shouldn't be modified by the
173 * The application needs a full memory barrier before checking
174 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
180 * Written by the application, shouldn't be modified by the
185 * Number of completion events lost because the queue was full;
186 * this should be avoided by the application by making sure
187 * there are not more requests pending than there is space in
188 * the completion queue.
190 * Written by the kernel, shouldn't be modified by the
191 * application (i.e. get number of "new events" by comparing to
194 * As completion events come in out of order this counter is not
195 * ordered with any other data.
199 * Ring buffer of completion events.
201 * The kernel writes completion events fresh every time they are
202 * produced, so the application is allowed to modify pending
205 struct io_uring_cqe cqes
[] ____cacheline_aligned_in_smp
;
208 struct io_mapped_ubuf
{
211 unsigned int nr_bvecs
;
212 unsigned long acct_pages
;
213 struct bio_vec bvec
[];
218 struct io_overflow_cqe
{
219 struct list_head list
;
220 struct io_uring_cqe cqe
;
224 * FFS_SCM is only available on 64-bit archs, for 32-bit we just define it as 0
225 * and define IO_URING_SCM_ALL. For this case, we use SCM for all files as we
226 * can't safely always dereference the file when the task has exited and ring
227 * cleanup is done. If a file is tracked and part of SCM, then unix gc on
228 * process exit may reap it before __io_sqe_files_unregister() is run.
230 #define FFS_NOWAIT 0x1UL
231 #define FFS_ISREG 0x2UL
232 #if defined(CONFIG_64BIT)
233 #define FFS_SCM 0x4UL
235 #define IO_URING_SCM_ALL
236 #define FFS_SCM 0x0UL
238 #define FFS_MASK ~(FFS_NOWAIT|FFS_ISREG|FFS_SCM)
240 struct io_fixed_file
{
241 /* file * with additional FFS_* flags */
242 unsigned long file_ptr
;
246 struct list_head list
;
251 struct io_mapped_ubuf
*buf
;
255 struct io_file_table
{
256 struct io_fixed_file
*files
;
257 unsigned long *bitmap
;
258 unsigned int alloc_hint
;
261 struct io_rsrc_node
{
262 struct percpu_ref refs
;
263 struct list_head node
;
264 struct list_head rsrc_list
;
265 struct io_rsrc_data
*rsrc_data
;
266 struct llist_node llist
;
270 typedef void (rsrc_put_fn
)(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
);
272 struct io_rsrc_data
{
273 struct io_ring_ctx
*ctx
;
279 struct completion done
;
283 #define IO_BUFFER_LIST_BUF_PER_PAGE (PAGE_SIZE / sizeof(struct io_uring_buf))
284 struct io_buffer_list
{
286 * If ->buf_nr_pages is set, then buf_pages/buf_ring are used. If not,
287 * then these are classic provided buffers and ->buf_list is used.
290 struct list_head buf_list
;
292 struct page
**buf_pages
;
293 struct io_uring_buf_ring
*buf_ring
;
298 /* below is for ring provided buffers */
306 struct list_head list
;
313 struct io_restriction
{
314 DECLARE_BITMAP(register_op
, IORING_REGISTER_LAST
);
315 DECLARE_BITMAP(sqe_op
, IORING_OP_LAST
);
316 u8 sqe_flags_allowed
;
317 u8 sqe_flags_required
;
322 IO_SQ_THREAD_SHOULD_STOP
= 0,
323 IO_SQ_THREAD_SHOULD_PARK
,
328 atomic_t park_pending
;
331 /* ctx's that are using this sqd */
332 struct list_head ctx_list
;
334 struct task_struct
*thread
;
335 struct wait_queue_head wait
;
337 unsigned sq_thread_idle
;
343 struct completion exited
;
346 #define IO_COMPL_BATCH 32
347 #define IO_REQ_CACHE_SIZE 32
348 #define IO_REQ_ALLOC_BATCH 8
350 struct io_submit_link
{
351 struct io_kiocb
*head
;
352 struct io_kiocb
*last
;
355 struct io_submit_state
{
356 /* inline/task_work completion list, under ->uring_lock */
357 struct io_wq_work_node free_list
;
358 /* batch completion logic */
359 struct io_wq_work_list compl_reqs
;
360 struct io_submit_link link
;
365 unsigned short submit_nr
;
366 struct blk_plug plug
;
370 struct eventfd_ctx
*cq_ev_fd
;
371 unsigned int eventfd_async
: 1;
375 #define BGID_ARRAY 64
378 /* const or read-mostly hot data */
380 struct percpu_ref refs
;
382 struct io_rings
*rings
;
384 enum task_work_notify_mode notify_method
;
385 unsigned int compat
: 1;
386 unsigned int drain_next
: 1;
387 unsigned int restricted
: 1;
388 unsigned int off_timeout_used
: 1;
389 unsigned int drain_active
: 1;
390 unsigned int drain_disabled
: 1;
391 unsigned int has_evfd
: 1;
392 unsigned int syscall_iopoll
: 1;
393 } ____cacheline_aligned_in_smp
;
395 /* submission data */
397 struct mutex uring_lock
;
400 * Ring buffer of indices into array of io_uring_sqe, which is
401 * mmapped by the application using the IORING_OFF_SQES offset.
403 * This indirection could e.g. be used to assign fixed
404 * io_uring_sqe entries to operations and only submit them to
405 * the queue when needed.
407 * The kernel modifies neither the indices array nor the entries
411 struct io_uring_sqe
*sq_sqes
;
412 unsigned cached_sq_head
;
414 struct list_head defer_list
;
417 * Fixed resources fast path, should be accessed only under
418 * uring_lock, and updated through io_uring_register(2)
420 struct io_rsrc_node
*rsrc_node
;
421 int rsrc_cached_refs
;
423 struct io_file_table file_table
;
424 unsigned nr_user_files
;
425 unsigned nr_user_bufs
;
426 struct io_mapped_ubuf
**user_bufs
;
428 struct io_submit_state submit_state
;
430 struct io_buffer_list
*io_bl
;
431 struct xarray io_bl_xa
;
432 struct list_head io_buffers_cache
;
434 struct list_head timeout_list
;
435 struct list_head ltimeout_list
;
436 struct list_head cq_overflow_list
;
437 struct list_head apoll_cache
;
438 struct xarray personalities
;
440 unsigned sq_thread_idle
;
441 } ____cacheline_aligned_in_smp
;
443 /* IRQ completion list, under ->completion_lock */
444 struct io_wq_work_list locked_free_list
;
445 unsigned int locked_free_nr
;
447 const struct cred
*sq_creds
; /* cred used for __io_sq_thread() */
448 struct io_sq_data
*sq_data
; /* if using sq thread polling */
450 struct wait_queue_head sqo_sq_wait
;
451 struct list_head sqd_list
;
453 unsigned long check_cq
;
457 * We cache a range of free CQEs we can use, once exhausted it
458 * should go through a slower range setup, see __io_get_cqe()
460 struct io_uring_cqe
*cqe_cached
;
461 struct io_uring_cqe
*cqe_sentinel
;
463 unsigned cached_cq_tail
;
465 struct io_ev_fd __rcu
*io_ev_fd
;
466 struct wait_queue_head cq_wait
;
468 atomic_t cq_timeouts
;
469 unsigned cq_last_tm_flush
;
470 } ____cacheline_aligned_in_smp
;
473 spinlock_t completion_lock
;
475 spinlock_t timeout_lock
;
478 * ->iopoll_list is protected by the ctx->uring_lock for
479 * io_uring instances that don't use IORING_SETUP_SQPOLL.
480 * For SQPOLL, only the single threaded io_sq_thread() will
481 * manipulate the list, hence no extra locking is needed there.
483 struct io_wq_work_list iopoll_list
;
484 struct hlist_head
*cancel_hash
;
485 unsigned cancel_hash_bits
;
486 bool poll_multi_queue
;
488 struct list_head io_buffers_comp
;
489 } ____cacheline_aligned_in_smp
;
491 struct io_restriction restrictions
;
493 /* slow path rsrc auxilary data, used by update/register */
495 struct io_rsrc_node
*rsrc_backup_node
;
496 struct io_mapped_ubuf
*dummy_ubuf
;
497 struct io_rsrc_data
*file_data
;
498 struct io_rsrc_data
*buf_data
;
500 struct delayed_work rsrc_put_work
;
501 struct llist_head rsrc_put_llist
;
502 struct list_head rsrc_ref_list
;
503 spinlock_t rsrc_ref_lock
;
505 struct list_head io_buffers_pages
;
508 /* Keep this last, we don't need it for the fast path */
510 #if defined(CONFIG_UNIX)
511 struct socket
*ring_sock
;
513 /* hashed buffered write serialization */
514 struct io_wq_hash
*hash_map
;
516 /* Only used for accounting purposes */
517 struct user_struct
*user
;
518 struct mm_struct
*mm_account
;
520 /* ctx exit and cancelation */
521 struct llist_head fallback_llist
;
522 struct delayed_work fallback_work
;
523 struct work_struct exit_work
;
524 struct list_head tctx_list
;
525 struct completion ref_comp
;
527 bool iowq_limits_set
;
532 * Arbitrary limit, can be raised if need be
534 #define IO_RINGFD_REG_MAX 16
536 struct io_uring_task
{
537 /* submission side */
540 struct wait_queue_head wait
;
541 const struct io_ring_ctx
*last
;
543 struct percpu_counter inflight
;
544 atomic_t inflight_tracked
;
547 spinlock_t task_lock
;
548 struct io_wq_work_list task_list
;
549 struct io_wq_work_list prio_task_list
;
550 struct callback_head task_work
;
551 struct file
**registered_rings
;
556 * First field must be the file pointer in all the
557 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
559 struct io_poll_iocb
{
561 struct wait_queue_head
*head
;
563 struct wait_queue_entry wait
;
566 struct io_poll_update
{
572 bool update_user_data
;
581 struct io_timeout_data
{
582 struct io_kiocb
*req
;
583 struct hrtimer timer
;
584 struct timespec64 ts
;
585 enum hrtimer_mode mode
;
591 struct sockaddr __user
*addr
;
592 int __user
*addr_len
;
595 unsigned long nofile
;
605 unsigned long nofile
;
627 struct list_head list
;
628 /* head of the link, used by linked timeouts only */
629 struct io_kiocb
*head
;
630 /* for linked completions */
631 struct io_kiocb
*prev
;
634 struct io_timeout_rem
{
639 struct timespec64 ts
;
645 /* NOTE: kiocb has the file as the first member, so don't do it here */
654 struct sockaddr __user
*addr
;
661 struct compat_msghdr __user
*umsg_compat
;
662 struct user_msghdr __user
*umsg
;
675 struct filename
*filename
;
677 unsigned long nofile
;
680 struct io_rsrc_update
{
706 struct epoll_event event
;
710 struct file
*file_out
;
718 struct io_provide_buf
{
732 struct filename
*filename
;
733 struct statx __user
*buffer
;
745 struct filename
*oldpath
;
746 struct filename
*newpath
;
754 struct filename
*filename
;
761 struct filename
*filename
;
767 struct filename
*oldpath
;
768 struct filename
*newpath
;
775 struct filename
*oldpath
;
776 struct filename
*newpath
;
786 struct io_async_connect
{
787 struct sockaddr_storage address
;
790 struct io_async_msghdr
{
791 struct iovec fast_iov
[UIO_FASTIOV
];
792 /* points to an allocated iov, if NULL we use fast_iov instead */
793 struct iovec
*free_iov
;
794 struct sockaddr __user
*uaddr
;
796 struct sockaddr_storage addr
;
800 struct iov_iter iter
;
801 struct iov_iter_state iter_state
;
802 struct iovec fast_iov
[UIO_FASTIOV
];
806 struct io_rw_state s
;
807 const struct iovec
*free_iovec
;
809 struct wait_page_queue wpq
;
814 struct xattr_ctx ctx
;
815 struct filename
*filename
;
819 REQ_F_FIXED_FILE_BIT
= IOSQE_FIXED_FILE_BIT
,
820 REQ_F_IO_DRAIN_BIT
= IOSQE_IO_DRAIN_BIT
,
821 REQ_F_LINK_BIT
= IOSQE_IO_LINK_BIT
,
822 REQ_F_HARDLINK_BIT
= IOSQE_IO_HARDLINK_BIT
,
823 REQ_F_FORCE_ASYNC_BIT
= IOSQE_ASYNC_BIT
,
824 REQ_F_BUFFER_SELECT_BIT
= IOSQE_BUFFER_SELECT_BIT
,
825 REQ_F_CQE_SKIP_BIT
= IOSQE_CQE_SKIP_SUCCESS_BIT
,
827 /* first byte is taken by user flags, shift it to not overlap */
832 REQ_F_LINK_TIMEOUT_BIT
,
833 REQ_F_NEED_CLEANUP_BIT
,
835 REQ_F_BUFFER_SELECTED_BIT
,
836 REQ_F_BUFFER_RING_BIT
,
837 REQ_F_COMPLETE_INLINE_BIT
,
841 REQ_F_ARM_LTIMEOUT_BIT
,
842 REQ_F_ASYNC_DATA_BIT
,
843 REQ_F_SKIP_LINK_CQES_BIT
,
844 REQ_F_SINGLE_POLL_BIT
,
845 REQ_F_DOUBLE_POLL_BIT
,
846 REQ_F_PARTIAL_IO_BIT
,
847 REQ_F_CQE32_INIT_BIT
,
848 REQ_F_APOLL_MULTISHOT_BIT
,
849 /* keep async read/write and isreg together and in order */
850 REQ_F_SUPPORT_NOWAIT_BIT
,
853 /* not a real bit, just to check we're not overflowing the space */
859 REQ_F_FIXED_FILE
= BIT(REQ_F_FIXED_FILE_BIT
),
860 /* drain existing IO first */
861 REQ_F_IO_DRAIN
= BIT(REQ_F_IO_DRAIN_BIT
),
863 REQ_F_LINK
= BIT(REQ_F_LINK_BIT
),
864 /* doesn't sever on completion < 0 */
865 REQ_F_HARDLINK
= BIT(REQ_F_HARDLINK_BIT
),
867 REQ_F_FORCE_ASYNC
= BIT(REQ_F_FORCE_ASYNC_BIT
),
868 /* IOSQE_BUFFER_SELECT */
869 REQ_F_BUFFER_SELECT
= BIT(REQ_F_BUFFER_SELECT_BIT
),
870 /* IOSQE_CQE_SKIP_SUCCESS */
871 REQ_F_CQE_SKIP
= BIT(REQ_F_CQE_SKIP_BIT
),
873 /* fail rest of links */
874 REQ_F_FAIL
= BIT(REQ_F_FAIL_BIT
),
875 /* on inflight list, should be cancelled and waited on exit reliably */
876 REQ_F_INFLIGHT
= BIT(REQ_F_INFLIGHT_BIT
),
877 /* read/write uses file position */
878 REQ_F_CUR_POS
= BIT(REQ_F_CUR_POS_BIT
),
879 /* must not punt to workers */
880 REQ_F_NOWAIT
= BIT(REQ_F_NOWAIT_BIT
),
881 /* has or had linked timeout */
882 REQ_F_LINK_TIMEOUT
= BIT(REQ_F_LINK_TIMEOUT_BIT
),
884 REQ_F_NEED_CLEANUP
= BIT(REQ_F_NEED_CLEANUP_BIT
),
885 /* already went through poll handler */
886 REQ_F_POLLED
= BIT(REQ_F_POLLED_BIT
),
887 /* buffer already selected */
888 REQ_F_BUFFER_SELECTED
= BIT(REQ_F_BUFFER_SELECTED_BIT
),
889 /* buffer selected from ring, needs commit */
890 REQ_F_BUFFER_RING
= BIT(REQ_F_BUFFER_RING_BIT
),
891 /* completion is deferred through io_comp_state */
892 REQ_F_COMPLETE_INLINE
= BIT(REQ_F_COMPLETE_INLINE_BIT
),
893 /* caller should reissue async */
894 REQ_F_REISSUE
= BIT(REQ_F_REISSUE_BIT
),
895 /* supports async reads/writes */
896 REQ_F_SUPPORT_NOWAIT
= BIT(REQ_F_SUPPORT_NOWAIT_BIT
),
898 REQ_F_ISREG
= BIT(REQ_F_ISREG_BIT
),
899 /* has creds assigned */
900 REQ_F_CREDS
= BIT(REQ_F_CREDS_BIT
),
901 /* skip refcounting if not set */
902 REQ_F_REFCOUNT
= BIT(REQ_F_REFCOUNT_BIT
),
903 /* there is a linked timeout that has to be armed */
904 REQ_F_ARM_LTIMEOUT
= BIT(REQ_F_ARM_LTIMEOUT_BIT
),
905 /* ->async_data allocated */
906 REQ_F_ASYNC_DATA
= BIT(REQ_F_ASYNC_DATA_BIT
),
907 /* don't post CQEs while failing linked requests */
908 REQ_F_SKIP_LINK_CQES
= BIT(REQ_F_SKIP_LINK_CQES_BIT
),
909 /* single poll may be active */
910 REQ_F_SINGLE_POLL
= BIT(REQ_F_SINGLE_POLL_BIT
),
911 /* double poll may active */
912 REQ_F_DOUBLE_POLL
= BIT(REQ_F_DOUBLE_POLL_BIT
),
913 /* request has already done partial IO */
914 REQ_F_PARTIAL_IO
= BIT(REQ_F_PARTIAL_IO_BIT
),
915 /* fast poll multishot mode */
916 REQ_F_APOLL_MULTISHOT
= BIT(REQ_F_APOLL_MULTISHOT_BIT
),
917 /* ->extra1 and ->extra2 are initialised */
918 REQ_F_CQE32_INIT
= BIT(REQ_F_CQE32_INIT_BIT
),
922 struct io_poll_iocb poll
;
923 struct io_poll_iocb
*double_poll
;
926 typedef void (*io_req_tw_func_t
)(struct io_kiocb
*req
, bool *locked
);
928 struct io_task_work
{
930 struct io_wq_work_node node
;
931 struct llist_node fallback_node
;
933 io_req_tw_func_t func
;
937 IORING_RSRC_FILE
= 0,
938 IORING_RSRC_BUFFER
= 1,
944 /* fd initially, then cflags for completion */
952 IO_CHECK_CQ_OVERFLOW_BIT
,
953 IO_CHECK_CQ_DROPPED_BIT
,
957 * NOTE! Each of the iocb union members has the file pointer
958 * as the first entry in their struct definition. So you can
959 * access the file pointer through any of the sub-structs,
960 * or directly as just 'file' in this struct.
966 struct io_poll_iocb poll
;
967 struct io_poll_update poll_update
;
968 struct io_accept accept
;
970 struct io_cancel cancel
;
971 struct io_timeout timeout
;
972 struct io_timeout_rem timeout_rem
;
973 struct io_connect connect
;
974 struct io_sr_msg sr_msg
;
976 struct io_close close
;
977 struct io_rsrc_update rsrc_update
;
978 struct io_fadvise fadvise
;
979 struct io_madvise madvise
;
980 struct io_epoll epoll
;
981 struct io_splice splice
;
982 struct io_provide_buf pbuf
;
983 struct io_statx statx
;
984 struct io_shutdown shutdown
;
985 struct io_rename rename
;
986 struct io_unlink unlink
;
987 struct io_mkdir mkdir
;
988 struct io_symlink symlink
;
989 struct io_hardlink hardlink
;
991 struct io_xattr xattr
;
992 struct io_socket sock
;
993 struct io_uring_cmd uring_cmd
;
997 /* polled IO has completed */
1000 * Can be either a fixed buffer index, or used with provided buffers.
1001 * For the latter, before issue it points to the buffer group ID,
1002 * and after selection it points to the buffer ID itself.
1009 struct io_ring_ctx
*ctx
;
1010 struct task_struct
*task
;
1012 struct io_rsrc_node
*rsrc_node
;
1015 /* store used ubuf, so we can prevent reloading */
1016 struct io_mapped_ubuf
*imu
;
1018 /* stores selected buf, valid IFF REQ_F_BUFFER_SELECTED is set */
1019 struct io_buffer
*kbuf
;
1022 * stores buffer ID for ring provided buffers, valid IFF
1023 * REQ_F_BUFFER_RING is set.
1025 struct io_buffer_list
*buf_list
;
1029 /* used by request caches, completion batching and iopoll */
1030 struct io_wq_work_node comp_list
;
1031 /* cache ->apoll->events */
1032 __poll_t apoll_events
;
1036 struct io_task_work io_task_work
;
1037 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
1039 struct hlist_node hash_node
;
1045 /* internal polling, see IORING_FEAT_FAST_POLL */
1046 struct async_poll
*apoll
;
1047 /* opcode allocated if it needs to store data for async defer */
1049 /* linked requests, IFF REQ_F_HARDLINK or REQ_F_LINK are set */
1050 struct io_kiocb
*link
;
1051 /* custom credentials, valid IFF REQ_F_CREDS is set */
1052 const struct cred
*creds
;
1053 struct io_wq_work work
;
1056 struct io_tctx_node
{
1057 struct list_head ctx_node
;
1058 struct task_struct
*task
;
1059 struct io_ring_ctx
*ctx
;
1062 struct io_defer_entry
{
1063 struct list_head list
;
1064 struct io_kiocb
*req
;
1068 struct io_cancel_data
{
1069 struct io_ring_ctx
*ctx
;
1079 * The URING_CMD payload starts at 'cmd' in the first sqe, and continues into
1080 * the following sqe if SQE128 is used.
1082 #define uring_cmd_pdu_size(is_sqe128) \
1083 ((1 + !!(is_sqe128)) * sizeof(struct io_uring_sqe) - \
1084 offsetof(struct io_uring_sqe, cmd))
1087 /* needs req->file assigned */
1088 unsigned needs_file
: 1;
1089 /* should block plug */
1091 /* hash wq insertion if file is a regular file */
1092 unsigned hash_reg_file
: 1;
1093 /* unbound wq insertion if file is a non-regular file */
1094 unsigned unbound_nonreg_file
: 1;
1095 /* set if opcode supports polled "wait" */
1096 unsigned pollin
: 1;
1097 unsigned pollout
: 1;
1098 unsigned poll_exclusive
: 1;
1099 /* op supports buffer selection */
1100 unsigned buffer_select
: 1;
1101 /* do prep async if is going to be punted */
1102 unsigned needs_async_setup
: 1;
1103 /* opcode is not supported by this kernel */
1104 unsigned not_supported
: 1;
1106 unsigned audit_skip
: 1;
1107 /* supports ioprio */
1108 unsigned ioprio
: 1;
1109 /* supports iopoll */
1110 unsigned iopoll
: 1;
1111 /* size of async data needed, if any */
1112 unsigned short async_size
;
1115 static const struct io_op_def io_op_defs
[] = {
1120 [IORING_OP_READV
] = {
1122 .unbound_nonreg_file
= 1,
1125 .needs_async_setup
= 1,
1130 .async_size
= sizeof(struct io_async_rw
),
1132 [IORING_OP_WRITEV
] = {
1135 .unbound_nonreg_file
= 1,
1137 .needs_async_setup
= 1,
1142 .async_size
= sizeof(struct io_async_rw
),
1144 [IORING_OP_FSYNC
] = {
1148 [IORING_OP_READ_FIXED
] = {
1150 .unbound_nonreg_file
= 1,
1156 .async_size
= sizeof(struct io_async_rw
),
1158 [IORING_OP_WRITE_FIXED
] = {
1161 .unbound_nonreg_file
= 1,
1167 .async_size
= sizeof(struct io_async_rw
),
1169 [IORING_OP_POLL_ADD
] = {
1171 .unbound_nonreg_file
= 1,
1174 [IORING_OP_POLL_REMOVE
] = {
1177 [IORING_OP_SYNC_FILE_RANGE
] = {
1181 [IORING_OP_SENDMSG
] = {
1183 .unbound_nonreg_file
= 1,
1185 .needs_async_setup
= 1,
1186 .async_size
= sizeof(struct io_async_msghdr
),
1188 [IORING_OP_RECVMSG
] = {
1190 .unbound_nonreg_file
= 1,
1193 .needs_async_setup
= 1,
1194 .async_size
= sizeof(struct io_async_msghdr
),
1196 [IORING_OP_TIMEOUT
] = {
1198 .async_size
= sizeof(struct io_timeout_data
),
1200 [IORING_OP_TIMEOUT_REMOVE
] = {
1201 /* used by timeout updates' prep() */
1204 [IORING_OP_ACCEPT
] = {
1206 .unbound_nonreg_file
= 1,
1208 .poll_exclusive
= 1,
1209 .ioprio
= 1, /* used for flags */
1211 [IORING_OP_ASYNC_CANCEL
] = {
1214 [IORING_OP_LINK_TIMEOUT
] = {
1216 .async_size
= sizeof(struct io_timeout_data
),
1218 [IORING_OP_CONNECT
] = {
1220 .unbound_nonreg_file
= 1,
1222 .needs_async_setup
= 1,
1223 .async_size
= sizeof(struct io_async_connect
),
1225 [IORING_OP_FALLOCATE
] = {
1228 [IORING_OP_OPENAT
] = {},
1229 [IORING_OP_CLOSE
] = {},
1230 [IORING_OP_FILES_UPDATE
] = {
1234 [IORING_OP_STATX
] = {
1237 [IORING_OP_READ
] = {
1239 .unbound_nonreg_file
= 1,
1246 .async_size
= sizeof(struct io_async_rw
),
1248 [IORING_OP_WRITE
] = {
1251 .unbound_nonreg_file
= 1,
1257 .async_size
= sizeof(struct io_async_rw
),
1259 [IORING_OP_FADVISE
] = {
1263 [IORING_OP_MADVISE
] = {},
1264 [IORING_OP_SEND
] = {
1266 .unbound_nonreg_file
= 1,
1270 [IORING_OP_RECV
] = {
1272 .unbound_nonreg_file
= 1,
1277 [IORING_OP_OPENAT2
] = {
1279 [IORING_OP_EPOLL_CTL
] = {
1280 .unbound_nonreg_file
= 1,
1283 [IORING_OP_SPLICE
] = {
1286 .unbound_nonreg_file
= 1,
1289 [IORING_OP_PROVIDE_BUFFERS
] = {
1293 [IORING_OP_REMOVE_BUFFERS
] = {
1300 .unbound_nonreg_file
= 1,
1303 [IORING_OP_SHUTDOWN
] = {
1306 [IORING_OP_RENAMEAT
] = {},
1307 [IORING_OP_UNLINKAT
] = {},
1308 [IORING_OP_MKDIRAT
] = {},
1309 [IORING_OP_SYMLINKAT
] = {},
1310 [IORING_OP_LINKAT
] = {},
1311 [IORING_OP_MSG_RING
] = {
1315 [IORING_OP_FSETXATTR
] = {
1318 [IORING_OP_SETXATTR
] = {},
1319 [IORING_OP_FGETXATTR
] = {
1322 [IORING_OP_GETXATTR
] = {},
1323 [IORING_OP_SOCKET
] = {
1326 [IORING_OP_URING_CMD
] = {
1329 .needs_async_setup
= 1,
1330 .async_size
= uring_cmd_pdu_size(1),
1334 /* requests with any of those set should undergo io_disarm_next() */
1335 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1336 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
1338 static bool io_disarm_next(struct io_kiocb
*req
);
1339 static void io_uring_del_tctx_node(unsigned long index
);
1340 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
1341 struct task_struct
*task
,
1343 static void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
);
1345 static void __io_req_complete_post(struct io_kiocb
*req
, s32 res
, u32 cflags
);
1346 static void io_dismantle_req(struct io_kiocb
*req
);
1347 static void io_queue_linked_timeout(struct io_kiocb
*req
);
1348 static int __io_register_rsrc_update(struct io_ring_ctx
*ctx
, unsigned type
,
1349 struct io_uring_rsrc_update2
*up
,
1351 static void io_clean_op(struct io_kiocb
*req
);
1352 static inline struct file
*io_file_get_fixed(struct io_kiocb
*req
, int fd
,
1353 unsigned issue_flags
);
1354 static struct file
*io_file_get_normal(struct io_kiocb
*req
, int fd
);
1355 static void io_queue_sqe(struct io_kiocb
*req
);
1356 static void io_rsrc_put_work(struct work_struct
*work
);
1358 static void io_req_task_queue(struct io_kiocb
*req
);
1359 static void __io_submit_flush_completions(struct io_ring_ctx
*ctx
);
1360 static int io_req_prep_async(struct io_kiocb
*req
);
1362 static int io_install_fixed_file(struct io_kiocb
*req
, struct file
*file
,
1363 unsigned int issue_flags
, u32 slot_index
);
1364 static int __io_close_fixed(struct io_kiocb
*req
, unsigned int issue_flags
,
1365 unsigned int offset
);
1366 static inline int io_close_fixed(struct io_kiocb
*req
, unsigned int issue_flags
);
1368 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
);
1369 static void io_eventfd_signal(struct io_ring_ctx
*ctx
);
1370 static void io_req_tw_post_queue(struct io_kiocb
*req
, s32 res
, u32 cflags
);
1372 static struct kmem_cache
*req_cachep
;
1374 static const struct file_operations io_uring_fops
;
1376 const char *io_uring_get_opcode(u8 opcode
)
1378 switch ((enum io_uring_op
)opcode
) {
1381 case IORING_OP_READV
:
1383 case IORING_OP_WRITEV
:
1385 case IORING_OP_FSYNC
:
1387 case IORING_OP_READ_FIXED
:
1388 return "READ_FIXED";
1389 case IORING_OP_WRITE_FIXED
:
1390 return "WRITE_FIXED";
1391 case IORING_OP_POLL_ADD
:
1393 case IORING_OP_POLL_REMOVE
:
1394 return "POLL_REMOVE";
1395 case IORING_OP_SYNC_FILE_RANGE
:
1396 return "SYNC_FILE_RANGE";
1397 case IORING_OP_SENDMSG
:
1399 case IORING_OP_RECVMSG
:
1401 case IORING_OP_TIMEOUT
:
1403 case IORING_OP_TIMEOUT_REMOVE
:
1404 return "TIMEOUT_REMOVE";
1405 case IORING_OP_ACCEPT
:
1407 case IORING_OP_ASYNC_CANCEL
:
1408 return "ASYNC_CANCEL";
1409 case IORING_OP_LINK_TIMEOUT
:
1410 return "LINK_TIMEOUT";
1411 case IORING_OP_CONNECT
:
1413 case IORING_OP_FALLOCATE
:
1415 case IORING_OP_OPENAT
:
1417 case IORING_OP_CLOSE
:
1419 case IORING_OP_FILES_UPDATE
:
1420 return "FILES_UPDATE";
1421 case IORING_OP_STATX
:
1423 case IORING_OP_READ
:
1425 case IORING_OP_WRITE
:
1427 case IORING_OP_FADVISE
:
1429 case IORING_OP_MADVISE
:
1431 case IORING_OP_SEND
:
1433 case IORING_OP_RECV
:
1435 case IORING_OP_OPENAT2
:
1437 case IORING_OP_EPOLL_CTL
:
1439 case IORING_OP_SPLICE
:
1441 case IORING_OP_PROVIDE_BUFFERS
:
1442 return "PROVIDE_BUFFERS";
1443 case IORING_OP_REMOVE_BUFFERS
:
1444 return "REMOVE_BUFFERS";
1447 case IORING_OP_SHUTDOWN
:
1449 case IORING_OP_RENAMEAT
:
1451 case IORING_OP_UNLINKAT
:
1453 case IORING_OP_MKDIRAT
:
1455 case IORING_OP_SYMLINKAT
:
1457 case IORING_OP_LINKAT
:
1459 case IORING_OP_MSG_RING
:
1461 case IORING_OP_FSETXATTR
:
1463 case IORING_OP_SETXATTR
:
1465 case IORING_OP_FGETXATTR
:
1467 case IORING_OP_GETXATTR
:
1469 case IORING_OP_SOCKET
:
1471 case IORING_OP_URING_CMD
:
1473 case IORING_OP_LAST
:
1479 struct sock
*io_uring_get_socket(struct file
*file
)
1481 #if defined(CONFIG_UNIX)
1482 if (file
->f_op
== &io_uring_fops
) {
1483 struct io_ring_ctx
*ctx
= file
->private_data
;
1485 return ctx
->ring_sock
->sk
;
1490 EXPORT_SYMBOL(io_uring_get_socket
);
1492 #if defined(CONFIG_UNIX)
1493 static inline bool io_file_need_scm(struct file
*filp
)
1495 #if defined(IO_URING_SCM_ALL)
1498 return !!unix_get_socket(filp
);
1502 static inline bool io_file_need_scm(struct file
*filp
)
1508 static void io_ring_submit_unlock(struct io_ring_ctx
*ctx
, unsigned issue_flags
)
1510 lockdep_assert_held(&ctx
->uring_lock
);
1511 if (issue_flags
& IO_URING_F_UNLOCKED
)
1512 mutex_unlock(&ctx
->uring_lock
);
1515 static void io_ring_submit_lock(struct io_ring_ctx
*ctx
, unsigned issue_flags
)
1518 * "Normal" inline submissions always hold the uring_lock, since we
1519 * grab it from the system call. Same is true for the SQPOLL offload.
1520 * The only exception is when we've detached the request and issue it
1521 * from an async worker thread, grab the lock for that case.
1523 if (issue_flags
& IO_URING_F_UNLOCKED
)
1524 mutex_lock(&ctx
->uring_lock
);
1525 lockdep_assert_held(&ctx
->uring_lock
);
1528 static inline void io_tw_lock(struct io_ring_ctx
*ctx
, bool *locked
)
1531 mutex_lock(&ctx
->uring_lock
);
1536 #define io_for_each_link(pos, head) \
1537 for (pos = (head); pos; pos = pos->link)
1540 * Shamelessly stolen from the mm implementation of page reference checking,
1541 * see commit f958d7b528b1 for details.
1543 #define req_ref_zero_or_close_to_overflow(req) \
1544 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1546 static inline bool req_ref_inc_not_zero(struct io_kiocb
*req
)
1548 WARN_ON_ONCE(!(req
->flags
& REQ_F_REFCOUNT
));
1549 return atomic_inc_not_zero(&req
->refs
);
1552 static inline bool req_ref_put_and_test(struct io_kiocb
*req
)
1554 if (likely(!(req
->flags
& REQ_F_REFCOUNT
)))
1557 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req
));
1558 return atomic_dec_and_test(&req
->refs
);
1561 static inline void req_ref_get(struct io_kiocb
*req
)
1563 WARN_ON_ONCE(!(req
->flags
& REQ_F_REFCOUNT
));
1564 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req
));
1565 atomic_inc(&req
->refs
);
1568 static inline void io_submit_flush_completions(struct io_ring_ctx
*ctx
)
1570 if (!wq_list_empty(&ctx
->submit_state
.compl_reqs
))
1571 __io_submit_flush_completions(ctx
);
1574 static inline void __io_req_set_refcount(struct io_kiocb
*req
, int nr
)
1576 if (!(req
->flags
& REQ_F_REFCOUNT
)) {
1577 req
->flags
|= REQ_F_REFCOUNT
;
1578 atomic_set(&req
->refs
, nr
);
1582 static inline void io_req_set_refcount(struct io_kiocb
*req
)
1584 __io_req_set_refcount(req
, 1);
1587 #define IO_RSRC_REF_BATCH 100
1589 static void io_rsrc_put_node(struct io_rsrc_node
*node
, int nr
)
1591 percpu_ref_put_many(&node
->refs
, nr
);
1594 static inline void io_req_put_rsrc_locked(struct io_kiocb
*req
,
1595 struct io_ring_ctx
*ctx
)
1596 __must_hold(&ctx
->uring_lock
)
1598 struct io_rsrc_node
*node
= req
->rsrc_node
;
1601 if (node
== ctx
->rsrc_node
)
1602 ctx
->rsrc_cached_refs
++;
1604 io_rsrc_put_node(node
, 1);
1608 static inline void io_req_put_rsrc(struct io_kiocb
*req
)
1611 io_rsrc_put_node(req
->rsrc_node
, 1);
1614 static __cold
void io_rsrc_refs_drop(struct io_ring_ctx
*ctx
)
1615 __must_hold(&ctx
->uring_lock
)
1617 if (ctx
->rsrc_cached_refs
) {
1618 io_rsrc_put_node(ctx
->rsrc_node
, ctx
->rsrc_cached_refs
);
1619 ctx
->rsrc_cached_refs
= 0;
1623 static void io_rsrc_refs_refill(struct io_ring_ctx
*ctx
)
1624 __must_hold(&ctx
->uring_lock
)
1626 ctx
->rsrc_cached_refs
+= IO_RSRC_REF_BATCH
;
1627 percpu_ref_get_many(&ctx
->rsrc_node
->refs
, IO_RSRC_REF_BATCH
);
1630 static inline void io_req_set_rsrc_node(struct io_kiocb
*req
,
1631 struct io_ring_ctx
*ctx
,
1632 unsigned int issue_flags
)
1634 if (!req
->rsrc_node
) {
1635 req
->rsrc_node
= ctx
->rsrc_node
;
1637 if (!(issue_flags
& IO_URING_F_UNLOCKED
)) {
1638 lockdep_assert_held(&ctx
->uring_lock
);
1639 ctx
->rsrc_cached_refs
--;
1640 if (unlikely(ctx
->rsrc_cached_refs
< 0))
1641 io_rsrc_refs_refill(ctx
);
1643 percpu_ref_get(&req
->rsrc_node
->refs
);
1648 static unsigned int __io_put_kbuf(struct io_kiocb
*req
, struct list_head
*list
)
1650 if (req
->flags
& REQ_F_BUFFER_RING
) {
1652 req
->buf_list
->head
++;
1653 req
->flags
&= ~REQ_F_BUFFER_RING
;
1655 list_add(&req
->kbuf
->list
, list
);
1656 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
1659 return IORING_CQE_F_BUFFER
| (req
->buf_index
<< IORING_CQE_BUFFER_SHIFT
);
1662 static inline unsigned int io_put_kbuf_comp(struct io_kiocb
*req
)
1664 lockdep_assert_held(&req
->ctx
->completion_lock
);
1666 if (!(req
->flags
& (REQ_F_BUFFER_SELECTED
|REQ_F_BUFFER_RING
)))
1668 return __io_put_kbuf(req
, &req
->ctx
->io_buffers_comp
);
1671 static inline unsigned int io_put_kbuf(struct io_kiocb
*req
,
1672 unsigned issue_flags
)
1674 unsigned int cflags
;
1676 if (!(req
->flags
& (REQ_F_BUFFER_SELECTED
|REQ_F_BUFFER_RING
)))
1680 * We can add this buffer back to two lists:
1682 * 1) The io_buffers_cache list. This one is protected by the
1683 * ctx->uring_lock. If we already hold this lock, add back to this
1684 * list as we can grab it from issue as well.
1685 * 2) The io_buffers_comp list. This one is protected by the
1686 * ctx->completion_lock.
1688 * We migrate buffers from the comp_list to the issue cache list
1691 if (req
->flags
& REQ_F_BUFFER_RING
) {
1692 /* no buffers to recycle for this case */
1693 cflags
= __io_put_kbuf(req
, NULL
);
1694 } else if (issue_flags
& IO_URING_F_UNLOCKED
) {
1695 struct io_ring_ctx
*ctx
= req
->ctx
;
1697 spin_lock(&ctx
->completion_lock
);
1698 cflags
= __io_put_kbuf(req
, &ctx
->io_buffers_comp
);
1699 spin_unlock(&ctx
->completion_lock
);
1701 lockdep_assert_held(&req
->ctx
->uring_lock
);
1703 cflags
= __io_put_kbuf(req
, &req
->ctx
->io_buffers_cache
);
1709 static struct io_buffer_list
*io_buffer_get_list(struct io_ring_ctx
*ctx
,
1712 if (ctx
->io_bl
&& bgid
< BGID_ARRAY
)
1713 return &ctx
->io_bl
[bgid
];
1715 return xa_load(&ctx
->io_bl_xa
, bgid
);
1718 static void io_kbuf_recycle(struct io_kiocb
*req
, unsigned issue_flags
)
1720 struct io_ring_ctx
*ctx
= req
->ctx
;
1721 struct io_buffer_list
*bl
;
1722 struct io_buffer
*buf
;
1724 if (!(req
->flags
& (REQ_F_BUFFER_SELECTED
|REQ_F_BUFFER_RING
)))
1727 * For legacy provided buffer mode, don't recycle if we already did
1728 * IO to this buffer. For ring-mapped provided buffer mode, we should
1729 * increment ring->head to explicitly monopolize the buffer to avoid
1732 if ((req
->flags
& REQ_F_BUFFER_SELECTED
) &&
1733 (req
->flags
& REQ_F_PARTIAL_IO
))
1737 * We don't need to recycle for REQ_F_BUFFER_RING, we can just clear
1738 * the flag and hence ensure that bl->head doesn't get incremented.
1739 * If the tail has already been incremented, hang on to it.
1741 if (req
->flags
& REQ_F_BUFFER_RING
) {
1742 if (req
->buf_list
) {
1743 if (req
->flags
& REQ_F_PARTIAL_IO
) {
1744 req
->buf_list
->head
++;
1745 req
->buf_list
= NULL
;
1747 req
->buf_index
= req
->buf_list
->bgid
;
1748 req
->flags
&= ~REQ_F_BUFFER_RING
;
1754 io_ring_submit_lock(ctx
, issue_flags
);
1757 bl
= io_buffer_get_list(ctx
, buf
->bgid
);
1758 list_add(&buf
->list
, &bl
->buf_list
);
1759 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
1760 req
->buf_index
= buf
->bgid
;
1762 io_ring_submit_unlock(ctx
, issue_flags
);
1765 static bool io_match_task(struct io_kiocb
*head
, struct task_struct
*task
,
1767 __must_hold(&req
->ctx
->timeout_lock
)
1769 struct io_kiocb
*req
;
1771 if (task
&& head
->task
!= task
)
1776 io_for_each_link(req
, head
) {
1777 if (req
->flags
& REQ_F_INFLIGHT
)
1783 static bool io_match_linked(struct io_kiocb
*head
)
1785 struct io_kiocb
*req
;
1787 io_for_each_link(req
, head
) {
1788 if (req
->flags
& REQ_F_INFLIGHT
)
1795 * As io_match_task() but protected against racing with linked timeouts.
1796 * User must not hold timeout_lock.
1798 static bool io_match_task_safe(struct io_kiocb
*head
, struct task_struct
*task
,
1803 if (task
&& head
->task
!= task
)
1808 if (head
->flags
& REQ_F_LINK_TIMEOUT
) {
1809 struct io_ring_ctx
*ctx
= head
->ctx
;
1811 /* protect against races with linked timeouts */
1812 spin_lock_irq(&ctx
->timeout_lock
);
1813 matched
= io_match_linked(head
);
1814 spin_unlock_irq(&ctx
->timeout_lock
);
1816 matched
= io_match_linked(head
);
1821 static inline bool req_has_async_data(struct io_kiocb
*req
)
1823 return req
->flags
& REQ_F_ASYNC_DATA
;
1826 static inline void req_set_fail(struct io_kiocb
*req
)
1828 req
->flags
|= REQ_F_FAIL
;
1829 if (req
->flags
& REQ_F_CQE_SKIP
) {
1830 req
->flags
&= ~REQ_F_CQE_SKIP
;
1831 req
->flags
|= REQ_F_SKIP_LINK_CQES
;
1835 static inline void req_fail_link_node(struct io_kiocb
*req
, int res
)
1841 static inline void io_req_add_to_cache(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
1843 wq_stack_add_head(&req
->comp_list
, &ctx
->submit_state
.free_list
);
1846 static __cold
void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
1848 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
1850 complete(&ctx
->ref_comp
);
1853 static inline bool io_is_timeout_noseq(struct io_kiocb
*req
)
1855 return !req
->timeout
.off
;
1858 static __cold
void io_fallback_req_func(struct work_struct
*work
)
1860 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
,
1861 fallback_work
.work
);
1862 struct llist_node
*node
= llist_del_all(&ctx
->fallback_llist
);
1863 struct io_kiocb
*req
, *tmp
;
1864 bool locked
= false;
1866 percpu_ref_get(&ctx
->refs
);
1867 llist_for_each_entry_safe(req
, tmp
, node
, io_task_work
.fallback_node
)
1868 req
->io_task_work
.func(req
, &locked
);
1871 io_submit_flush_completions(ctx
);
1872 mutex_unlock(&ctx
->uring_lock
);
1874 percpu_ref_put(&ctx
->refs
);
1877 static __cold
struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
1879 struct io_ring_ctx
*ctx
;
1882 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
1886 xa_init(&ctx
->io_bl_xa
);
1889 * Use 5 bits less than the max cq entries, that should give us around
1890 * 32 entries per hash list if totally full and uniformly spread.
1892 hash_bits
= ilog2(p
->cq_entries
);
1896 ctx
->cancel_hash_bits
= hash_bits
;
1897 ctx
->cancel_hash
= kmalloc((1U << hash_bits
) * sizeof(struct hlist_head
),
1899 if (!ctx
->cancel_hash
)
1901 __hash_init(ctx
->cancel_hash
, 1U << hash_bits
);
1903 ctx
->dummy_ubuf
= kzalloc(sizeof(*ctx
->dummy_ubuf
), GFP_KERNEL
);
1904 if (!ctx
->dummy_ubuf
)
1906 /* set invalid range, so io_import_fixed() fails meeting it */
1907 ctx
->dummy_ubuf
->ubuf
= -1UL;
1909 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
1910 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
1913 ctx
->flags
= p
->flags
;
1914 init_waitqueue_head(&ctx
->sqo_sq_wait
);
1915 INIT_LIST_HEAD(&ctx
->sqd_list
);
1916 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
1917 INIT_LIST_HEAD(&ctx
->io_buffers_cache
);
1918 INIT_LIST_HEAD(&ctx
->apoll_cache
);
1919 init_completion(&ctx
->ref_comp
);
1920 xa_init_flags(&ctx
->personalities
, XA_FLAGS_ALLOC1
);
1921 mutex_init(&ctx
->uring_lock
);
1922 init_waitqueue_head(&ctx
->cq_wait
);
1923 spin_lock_init(&ctx
->completion_lock
);
1924 spin_lock_init(&ctx
->timeout_lock
);
1925 INIT_WQ_LIST(&ctx
->iopoll_list
);
1926 INIT_LIST_HEAD(&ctx
->io_buffers_pages
);
1927 INIT_LIST_HEAD(&ctx
->io_buffers_comp
);
1928 INIT_LIST_HEAD(&ctx
->defer_list
);
1929 INIT_LIST_HEAD(&ctx
->timeout_list
);
1930 INIT_LIST_HEAD(&ctx
->ltimeout_list
);
1931 spin_lock_init(&ctx
->rsrc_ref_lock
);
1932 INIT_LIST_HEAD(&ctx
->rsrc_ref_list
);
1933 INIT_DELAYED_WORK(&ctx
->rsrc_put_work
, io_rsrc_put_work
);
1934 init_llist_head(&ctx
->rsrc_put_llist
);
1935 INIT_LIST_HEAD(&ctx
->tctx_list
);
1936 ctx
->submit_state
.free_list
.next
= NULL
;
1937 INIT_WQ_LIST(&ctx
->locked_free_list
);
1938 INIT_DELAYED_WORK(&ctx
->fallback_work
, io_fallback_req_func
);
1939 INIT_WQ_LIST(&ctx
->submit_state
.compl_reqs
);
1942 kfree(ctx
->dummy_ubuf
);
1943 kfree(ctx
->cancel_hash
);
1945 xa_destroy(&ctx
->io_bl_xa
);
1950 static void io_account_cq_overflow(struct io_ring_ctx
*ctx
)
1952 struct io_rings
*r
= ctx
->rings
;
1954 WRITE_ONCE(r
->cq_overflow
, READ_ONCE(r
->cq_overflow
) + 1);
1958 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
1960 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
1961 struct io_ring_ctx
*ctx
= req
->ctx
;
1963 return seq
+ READ_ONCE(ctx
->cq_extra
) != ctx
->cached_cq_tail
;
1969 static inline bool io_req_ffs_set(struct io_kiocb
*req
)
1971 return req
->flags
& REQ_F_FIXED_FILE
;
1974 static inline void io_req_track_inflight(struct io_kiocb
*req
)
1976 if (!(req
->flags
& REQ_F_INFLIGHT
)) {
1977 req
->flags
|= REQ_F_INFLIGHT
;
1978 atomic_inc(&req
->task
->io_uring
->inflight_tracked
);
1982 static struct io_kiocb
*__io_prep_linked_timeout(struct io_kiocb
*req
)
1984 if (WARN_ON_ONCE(!req
->link
))
1987 req
->flags
&= ~REQ_F_ARM_LTIMEOUT
;
1988 req
->flags
|= REQ_F_LINK_TIMEOUT
;
1990 /* linked timeouts should have two refs once prep'ed */
1991 io_req_set_refcount(req
);
1992 __io_req_set_refcount(req
->link
, 2);
1996 static inline struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
1998 if (likely(!(req
->flags
& REQ_F_ARM_LTIMEOUT
)))
2000 return __io_prep_linked_timeout(req
);
2003 static noinline
void __io_arm_ltimeout(struct io_kiocb
*req
)
2005 io_queue_linked_timeout(__io_prep_linked_timeout(req
));
2008 static inline void io_arm_ltimeout(struct io_kiocb
*req
)
2010 if (unlikely(req
->flags
& REQ_F_ARM_LTIMEOUT
))
2011 __io_arm_ltimeout(req
);
2014 static void io_prep_async_work(struct io_kiocb
*req
)
2016 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
2017 struct io_ring_ctx
*ctx
= req
->ctx
;
2019 if (!(req
->flags
& REQ_F_CREDS
)) {
2020 req
->flags
|= REQ_F_CREDS
;
2021 req
->creds
= get_current_cred();
2024 req
->work
.list
.next
= NULL
;
2025 req
->work
.flags
= 0;
2026 req
->work
.cancel_seq
= atomic_read(&ctx
->cancel_seq
);
2027 if (req
->flags
& REQ_F_FORCE_ASYNC
)
2028 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
2030 if (req
->flags
& REQ_F_ISREG
) {
2031 if (def
->hash_reg_file
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
2032 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
2033 } else if (!req
->file
|| !S_ISBLK(file_inode(req
->file
)->i_mode
)) {
2034 if (def
->unbound_nonreg_file
)
2035 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
2039 static void io_prep_async_link(struct io_kiocb
*req
)
2041 struct io_kiocb
*cur
;
2043 if (req
->flags
& REQ_F_LINK_TIMEOUT
) {
2044 struct io_ring_ctx
*ctx
= req
->ctx
;
2046 spin_lock_irq(&ctx
->timeout_lock
);
2047 io_for_each_link(cur
, req
)
2048 io_prep_async_work(cur
);
2049 spin_unlock_irq(&ctx
->timeout_lock
);
2051 io_for_each_link(cur
, req
)
2052 io_prep_async_work(cur
);
2056 static inline void io_req_add_compl_list(struct io_kiocb
*req
)
2058 struct io_submit_state
*state
= &req
->ctx
->submit_state
;
2060 if (!(req
->flags
& REQ_F_CQE_SKIP
))
2061 state
->flush_cqes
= true;
2062 wq_list_add_tail(&req
->comp_list
, &state
->compl_reqs
);
2065 static void io_queue_iowq(struct io_kiocb
*req
, bool *dont_use
)
2067 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
2068 struct io_uring_task
*tctx
= req
->task
->io_uring
;
2071 BUG_ON(!tctx
->io_wq
);
2073 /* init ->work of the whole link before punting */
2074 io_prep_async_link(req
);
2077 * Not expected to happen, but if we do have a bug where this _can_
2078 * happen, catch it here and ensure the request is marked as
2079 * canceled. That will make io-wq go through the usual work cancel
2080 * procedure rather than attempt to run this request (or create a new
2083 if (WARN_ON_ONCE(!same_thread_group(req
->task
, current
)))
2084 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
2086 trace_io_uring_queue_async_work(req
->ctx
, req
, req
->cqe
.user_data
,
2087 req
->opcode
, req
->flags
, &req
->work
,
2088 io_wq_is_hashed(&req
->work
));
2089 io_wq_enqueue(tctx
->io_wq
, &req
->work
);
2091 io_queue_linked_timeout(link
);
2094 static void io_kill_timeout(struct io_kiocb
*req
, int status
)
2095 __must_hold(&req
->ctx
->completion_lock
)
2096 __must_hold(&req
->ctx
->timeout_lock
)
2098 struct io_timeout_data
*io
= req
->async_data
;
2100 if (hrtimer_try_to_cancel(&io
->timer
) != -1) {
2103 atomic_set(&req
->ctx
->cq_timeouts
,
2104 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
2105 list_del_init(&req
->timeout
.list
);
2106 io_req_tw_post_queue(req
, status
, 0);
2110 static __cold
void io_queue_deferred(struct io_ring_ctx
*ctx
)
2112 while (!list_empty(&ctx
->defer_list
)) {
2113 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
2114 struct io_defer_entry
, list
);
2116 if (req_need_defer(de
->req
, de
->seq
))
2118 list_del_init(&de
->list
);
2119 io_req_task_queue(de
->req
);
2124 static __cold
void io_flush_timeouts(struct io_ring_ctx
*ctx
)
2125 __must_hold(&ctx
->completion_lock
)
2127 u32 seq
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
2128 struct io_kiocb
*req
, *tmp
;
2130 spin_lock_irq(&ctx
->timeout_lock
);
2131 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
2132 u32 events_needed
, events_got
;
2134 if (io_is_timeout_noseq(req
))
2138 * Since seq can easily wrap around over time, subtract
2139 * the last seq at which timeouts were flushed before comparing.
2140 * Assuming not more than 2^31-1 events have happened since,
2141 * these subtractions won't have wrapped, so we can check if
2142 * target is in [last_seq, current_seq] by comparing the two.
2144 events_needed
= req
->timeout
.target_seq
- ctx
->cq_last_tm_flush
;
2145 events_got
= seq
- ctx
->cq_last_tm_flush
;
2146 if (events_got
< events_needed
)
2149 io_kill_timeout(req
, 0);
2151 ctx
->cq_last_tm_flush
= seq
;
2152 spin_unlock_irq(&ctx
->timeout_lock
);
2155 static inline void io_commit_cqring(struct io_ring_ctx
*ctx
)
2157 /* order cqe stores with ring update */
2158 smp_store_release(&ctx
->rings
->cq
.tail
, ctx
->cached_cq_tail
);
2161 static void __io_commit_cqring_flush(struct io_ring_ctx
*ctx
)
2163 if (ctx
->off_timeout_used
|| ctx
->drain_active
) {
2164 spin_lock(&ctx
->completion_lock
);
2165 if (ctx
->off_timeout_used
)
2166 io_flush_timeouts(ctx
);
2167 if (ctx
->drain_active
)
2168 io_queue_deferred(ctx
);
2169 io_commit_cqring(ctx
);
2170 spin_unlock(&ctx
->completion_lock
);
2173 io_eventfd_signal(ctx
);
2176 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
2178 struct io_rings
*r
= ctx
->rings
;
2180 return READ_ONCE(r
->sq
.tail
) - ctx
->cached_sq_head
== ctx
->sq_entries
;
2183 static inline unsigned int __io_cqring_events(struct io_ring_ctx
*ctx
)
2185 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
2189 * writes to the cq entry need to come after reading head; the
2190 * control dependency is enough as we're using WRITE_ONCE to
2193 static noinline
struct io_uring_cqe
*__io_get_cqe(struct io_ring_ctx
*ctx
)
2195 struct io_rings
*rings
= ctx
->rings
;
2196 unsigned int off
= ctx
->cached_cq_tail
& (ctx
->cq_entries
- 1);
2197 unsigned int shift
= 0;
2198 unsigned int free
, queued
, len
;
2200 if (ctx
->flags
& IORING_SETUP_CQE32
)
2203 /* userspace may cheat modifying the tail, be safe and do min */
2204 queued
= min(__io_cqring_events(ctx
), ctx
->cq_entries
);
2205 free
= ctx
->cq_entries
- queued
;
2206 /* we need a contiguous range, limit based on the current array offset */
2207 len
= min(free
, ctx
->cq_entries
- off
);
2211 ctx
->cached_cq_tail
++;
2212 ctx
->cqe_cached
= &rings
->cqes
[off
];
2213 ctx
->cqe_sentinel
= ctx
->cqe_cached
+ len
;
2215 return &rings
->cqes
[off
<< shift
];
2218 static inline struct io_uring_cqe
*io_get_cqe(struct io_ring_ctx
*ctx
)
2220 if (likely(ctx
->cqe_cached
< ctx
->cqe_sentinel
)) {
2221 struct io_uring_cqe
*cqe
= ctx
->cqe_cached
;
2223 if (ctx
->flags
& IORING_SETUP_CQE32
) {
2224 unsigned int off
= ctx
->cqe_cached
- ctx
->rings
->cqes
;
2229 ctx
->cached_cq_tail
++;
2234 return __io_get_cqe(ctx
);
2237 static void io_eventfd_signal(struct io_ring_ctx
*ctx
)
2239 struct io_ev_fd
*ev_fd
;
2243 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
2244 * and eventfd_signal
2246 ev_fd
= rcu_dereference(ctx
->io_ev_fd
);
2249 * Check again if ev_fd exists incase an io_eventfd_unregister call
2250 * completed between the NULL check of ctx->io_ev_fd at the start of
2251 * the function and rcu_read_lock.
2253 if (unlikely(!ev_fd
))
2255 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
2258 if (!ev_fd
->eventfd_async
|| io_wq_current_is_worker())
2259 eventfd_signal(ev_fd
->cq_ev_fd
, 1);
2264 static inline void io_cqring_wake(struct io_ring_ctx
*ctx
)
2267 * wake_up_all() may seem excessive, but io_wake_function() and
2268 * io_should_wake() handle the termination of the loop and only
2269 * wake as many waiters as we need to.
2271 if (wq_has_sleeper(&ctx
->cq_wait
))
2272 wake_up_all(&ctx
->cq_wait
);
2276 * This should only get called when at least one event has been posted.
2277 * Some applications rely on the eventfd notification count only changing
2278 * IFF a new CQE has been added to the CQ ring. There's no depedency on
2279 * 1:1 relationship between how many times this function is called (and
2280 * hence the eventfd count) and number of CQEs posted to the CQ ring.
2282 static inline void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
2284 if (unlikely(ctx
->off_timeout_used
|| ctx
->drain_active
||
2286 __io_commit_cqring_flush(ctx
);
2288 io_cqring_wake(ctx
);
2291 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx
*ctx
)
2293 if (unlikely(ctx
->off_timeout_used
|| ctx
->drain_active
||
2295 __io_commit_cqring_flush(ctx
);
2297 if (ctx
->flags
& IORING_SETUP_SQPOLL
)
2298 io_cqring_wake(ctx
);
2301 /* Returns true if there are no backlogged entries after the flush */
2302 static bool __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
)
2304 bool all_flushed
, posted
;
2305 size_t cqe_size
= sizeof(struct io_uring_cqe
);
2307 if (!force
&& __io_cqring_events(ctx
) == ctx
->cq_entries
)
2310 if (ctx
->flags
& IORING_SETUP_CQE32
)
2314 spin_lock(&ctx
->completion_lock
);
2315 while (!list_empty(&ctx
->cq_overflow_list
)) {
2316 struct io_uring_cqe
*cqe
= io_get_cqe(ctx
);
2317 struct io_overflow_cqe
*ocqe
;
2321 ocqe
= list_first_entry(&ctx
->cq_overflow_list
,
2322 struct io_overflow_cqe
, list
);
2324 memcpy(cqe
, &ocqe
->cqe
, cqe_size
);
2326 io_account_cq_overflow(ctx
);
2329 list_del(&ocqe
->list
);
2333 all_flushed
= list_empty(&ctx
->cq_overflow_list
);
2335 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
2336 atomic_andnot(IORING_SQ_CQ_OVERFLOW
, &ctx
->rings
->sq_flags
);
2339 io_commit_cqring(ctx
);
2340 spin_unlock(&ctx
->completion_lock
);
2342 io_cqring_ev_posted(ctx
);
2346 static bool io_cqring_overflow_flush(struct io_ring_ctx
*ctx
)
2350 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
)) {
2351 /* iopoll syncs against uring_lock, not completion_lock */
2352 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
2353 mutex_lock(&ctx
->uring_lock
);
2354 ret
= __io_cqring_overflow_flush(ctx
, false);
2355 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
2356 mutex_unlock(&ctx
->uring_lock
);
2362 static void __io_put_task(struct task_struct
*task
, int nr
)
2364 struct io_uring_task
*tctx
= task
->io_uring
;
2366 percpu_counter_sub(&tctx
->inflight
, nr
);
2367 if (unlikely(atomic_read(&tctx
->in_idle
)))
2368 wake_up(&tctx
->wait
);
2369 put_task_struct_many(task
, nr
);
2372 /* must to be called somewhat shortly after putting a request */
2373 static inline void io_put_task(struct task_struct
*task
, int nr
)
2375 if (likely(task
== current
))
2376 task
->io_uring
->cached_refs
+= nr
;
2378 __io_put_task(task
, nr
);
2381 static void io_task_refs_refill(struct io_uring_task
*tctx
)
2383 unsigned int refill
= -tctx
->cached_refs
+ IO_TCTX_REFS_CACHE_NR
;
2385 percpu_counter_add(&tctx
->inflight
, refill
);
2386 refcount_add(refill
, ¤t
->usage
);
2387 tctx
->cached_refs
+= refill
;
2390 static inline void io_get_task_refs(int nr
)
2392 struct io_uring_task
*tctx
= current
->io_uring
;
2394 tctx
->cached_refs
-= nr
;
2395 if (unlikely(tctx
->cached_refs
< 0))
2396 io_task_refs_refill(tctx
);
2399 static __cold
void io_uring_drop_tctx_refs(struct task_struct
*task
)
2401 struct io_uring_task
*tctx
= task
->io_uring
;
2402 unsigned int refs
= tctx
->cached_refs
;
2405 tctx
->cached_refs
= 0;
2406 percpu_counter_sub(&tctx
->inflight
, refs
);
2407 put_task_struct_many(task
, refs
);
2411 static bool io_cqring_event_overflow(struct io_ring_ctx
*ctx
, u64 user_data
,
2412 s32 res
, u32 cflags
, u64 extra1
,
2415 struct io_overflow_cqe
*ocqe
;
2416 size_t ocq_size
= sizeof(struct io_overflow_cqe
);
2417 bool is_cqe32
= (ctx
->flags
& IORING_SETUP_CQE32
);
2420 ocq_size
+= sizeof(struct io_uring_cqe
);
2422 ocqe
= kmalloc(ocq_size
, GFP_ATOMIC
| __GFP_ACCOUNT
);
2423 trace_io_uring_cqe_overflow(ctx
, user_data
, res
, cflags
, ocqe
);
2426 * If we're in ring overflow flush mode, or in task cancel mode,
2427 * or cannot allocate an overflow entry, then we need to drop it
2430 io_account_cq_overflow(ctx
);
2431 set_bit(IO_CHECK_CQ_DROPPED_BIT
, &ctx
->check_cq
);
2434 if (list_empty(&ctx
->cq_overflow_list
)) {
2435 set_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
2436 atomic_or(IORING_SQ_CQ_OVERFLOW
, &ctx
->rings
->sq_flags
);
2439 ocqe
->cqe
.user_data
= user_data
;
2440 ocqe
->cqe
.res
= res
;
2441 ocqe
->cqe
.flags
= cflags
;
2443 ocqe
->cqe
.big_cqe
[0] = extra1
;
2444 ocqe
->cqe
.big_cqe
[1] = extra2
;
2446 list_add_tail(&ocqe
->list
, &ctx
->cq_overflow_list
);
2450 static inline bool __io_fill_cqe_req(struct io_ring_ctx
*ctx
,
2451 struct io_kiocb
*req
)
2453 struct io_uring_cqe
*cqe
;
2455 if (!(ctx
->flags
& IORING_SETUP_CQE32
)) {
2456 trace_io_uring_complete(req
->ctx
, req
, req
->cqe
.user_data
,
2457 req
->cqe
.res
, req
->cqe
.flags
, 0, 0);
2460 * If we can't get a cq entry, userspace overflowed the
2461 * submission (by quite a lot). Increment the overflow count in
2464 cqe
= io_get_cqe(ctx
);
2466 memcpy(cqe
, &req
->cqe
, sizeof(*cqe
));
2470 return io_cqring_event_overflow(ctx
, req
->cqe
.user_data
,
2471 req
->cqe
.res
, req
->cqe
.flags
,
2474 u64 extra1
= 0, extra2
= 0;
2476 if (req
->flags
& REQ_F_CQE32_INIT
) {
2477 extra1
= req
->extra1
;
2478 extra2
= req
->extra2
;
2481 trace_io_uring_complete(req
->ctx
, req
, req
->cqe
.user_data
,
2482 req
->cqe
.res
, req
->cqe
.flags
, extra1
, extra2
);
2485 * If we can't get a cq entry, userspace overflowed the
2486 * submission (by quite a lot). Increment the overflow count in
2489 cqe
= io_get_cqe(ctx
);
2491 memcpy(cqe
, &req
->cqe
, sizeof(struct io_uring_cqe
));
2492 WRITE_ONCE(cqe
->big_cqe
[0], extra1
);
2493 WRITE_ONCE(cqe
->big_cqe
[1], extra2
);
2497 return io_cqring_event_overflow(ctx
, req
->cqe
.user_data
,
2498 req
->cqe
.res
, req
->cqe
.flags
,
2503 static noinline
bool io_fill_cqe_aux(struct io_ring_ctx
*ctx
, u64 user_data
,
2504 s32 res
, u32 cflags
)
2506 struct io_uring_cqe
*cqe
;
2509 trace_io_uring_complete(ctx
, NULL
, user_data
, res
, cflags
, 0, 0);
2512 * If we can't get a cq entry, userspace overflowed the
2513 * submission (by quite a lot). Increment the overflow count in
2516 cqe
= io_get_cqe(ctx
);
2518 WRITE_ONCE(cqe
->user_data
, user_data
);
2519 WRITE_ONCE(cqe
->res
, res
);
2520 WRITE_ONCE(cqe
->flags
, cflags
);
2522 if (ctx
->flags
& IORING_SETUP_CQE32
) {
2523 WRITE_ONCE(cqe
->big_cqe
[0], 0);
2524 WRITE_ONCE(cqe
->big_cqe
[1], 0);
2528 return io_cqring_event_overflow(ctx
, user_data
, res
, cflags
, 0, 0);
2531 static void __io_req_complete_put(struct io_kiocb
*req
)
2534 * If we're the last reference to this request, add to our locked
2537 if (req_ref_put_and_test(req
)) {
2538 struct io_ring_ctx
*ctx
= req
->ctx
;
2540 if (req
->flags
& IO_REQ_LINK_FLAGS
) {
2541 if (req
->flags
& IO_DISARM_MASK
)
2542 io_disarm_next(req
);
2544 io_req_task_queue(req
->link
);
2548 io_req_put_rsrc(req
);
2550 * Selected buffer deallocation in io_clean_op() assumes that
2551 * we don't hold ->completion_lock. Clean them here to avoid
2554 io_put_kbuf_comp(req
);
2555 io_dismantle_req(req
);
2556 io_put_task(req
->task
, 1);
2557 wq_list_add_head(&req
->comp_list
, &ctx
->locked_free_list
);
2558 ctx
->locked_free_nr
++;
2562 static void __io_req_complete_post(struct io_kiocb
*req
, s32 res
,
2565 if (!(req
->flags
& REQ_F_CQE_SKIP
)) {
2567 req
->cqe
.flags
= cflags
;
2568 __io_fill_cqe_req(req
->ctx
, req
);
2570 __io_req_complete_put(req
);
2573 static void io_req_complete_post(struct io_kiocb
*req
, s32 res
, u32 cflags
)
2575 struct io_ring_ctx
*ctx
= req
->ctx
;
2577 spin_lock(&ctx
->completion_lock
);
2578 __io_req_complete_post(req
, res
, cflags
);
2579 io_commit_cqring(ctx
);
2580 spin_unlock(&ctx
->completion_lock
);
2581 io_cqring_ev_posted(ctx
);
2584 static inline void io_req_complete_state(struct io_kiocb
*req
, s32 res
,
2588 req
->cqe
.flags
= cflags
;
2589 req
->flags
|= REQ_F_COMPLETE_INLINE
;
2592 static inline void __io_req_complete(struct io_kiocb
*req
, unsigned issue_flags
,
2593 s32 res
, u32 cflags
)
2595 if (issue_flags
& IO_URING_F_COMPLETE_DEFER
)
2596 io_req_complete_state(req
, res
, cflags
);
2598 io_req_complete_post(req
, res
, cflags
);
2601 static inline void io_req_complete(struct io_kiocb
*req
, s32 res
)
2605 __io_req_complete(req
, 0, res
, 0);
2608 static void io_req_complete_failed(struct io_kiocb
*req
, s32 res
)
2611 io_req_complete_post(req
, res
, io_put_kbuf(req
, IO_URING_F_UNLOCKED
));
2615 * Don't initialise the fields below on every allocation, but do that in
2616 * advance and keep them valid across allocations.
2618 static void io_preinit_req(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
2622 req
->async_data
= NULL
;
2623 /* not necessary, but safer to zero */
2627 static void io_flush_cached_locked_reqs(struct io_ring_ctx
*ctx
,
2628 struct io_submit_state
*state
)
2630 spin_lock(&ctx
->completion_lock
);
2631 wq_list_splice(&ctx
->locked_free_list
, &state
->free_list
);
2632 ctx
->locked_free_nr
= 0;
2633 spin_unlock(&ctx
->completion_lock
);
2636 static inline bool io_req_cache_empty(struct io_ring_ctx
*ctx
)
2638 return !ctx
->submit_state
.free_list
.next
;
2642 * A request might get retired back into the request caches even before opcode
2643 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
2644 * Because of that, io_alloc_req() should be called only under ->uring_lock
2645 * and with extra caution to not get a request that is still worked on.
2647 static __cold
bool __io_alloc_req_refill(struct io_ring_ctx
*ctx
)
2648 __must_hold(&ctx
->uring_lock
)
2650 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
2651 void *reqs
[IO_REQ_ALLOC_BATCH
];
2655 * If we have more than a batch's worth of requests in our IRQ side
2656 * locked cache, grab the lock and move them over to our submission
2659 if (data_race(ctx
->locked_free_nr
) > IO_COMPL_BATCH
) {
2660 io_flush_cached_locked_reqs(ctx
, &ctx
->submit_state
);
2661 if (!io_req_cache_empty(ctx
))
2665 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, ARRAY_SIZE(reqs
), reqs
);
2668 * Bulk alloc is all-or-nothing. If we fail to get a batch,
2669 * retry single alloc to be on the safe side.
2671 if (unlikely(ret
<= 0)) {
2672 reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
2678 percpu_ref_get_many(&ctx
->refs
, ret
);
2679 for (i
= 0; i
< ret
; i
++) {
2680 struct io_kiocb
*req
= reqs
[i
];
2682 io_preinit_req(req
, ctx
);
2683 io_req_add_to_cache(req
, ctx
);
2688 static inline bool io_alloc_req_refill(struct io_ring_ctx
*ctx
)
2690 if (unlikely(io_req_cache_empty(ctx
)))
2691 return __io_alloc_req_refill(ctx
);
2695 static inline struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
)
2697 struct io_wq_work_node
*node
;
2699 node
= wq_stack_extract(&ctx
->submit_state
.free_list
);
2700 return container_of(node
, struct io_kiocb
, comp_list
);
2703 static inline void io_put_file(struct file
*file
)
2709 static inline void io_dismantle_req(struct io_kiocb
*req
)
2711 unsigned int flags
= req
->flags
;
2713 if (unlikely(flags
& IO_REQ_CLEAN_FLAGS
))
2715 if (!(flags
& REQ_F_FIXED_FILE
))
2716 io_put_file(req
->file
);
2719 static __cold
void io_free_req(struct io_kiocb
*req
)
2721 struct io_ring_ctx
*ctx
= req
->ctx
;
2723 io_req_put_rsrc(req
);
2724 io_dismantle_req(req
);
2725 io_put_task(req
->task
, 1);
2727 spin_lock(&ctx
->completion_lock
);
2728 wq_list_add_head(&req
->comp_list
, &ctx
->locked_free_list
);
2729 ctx
->locked_free_nr
++;
2730 spin_unlock(&ctx
->completion_lock
);
2733 static inline void io_remove_next_linked(struct io_kiocb
*req
)
2735 struct io_kiocb
*nxt
= req
->link
;
2737 req
->link
= nxt
->link
;
2741 static struct io_kiocb
*io_disarm_linked_timeout(struct io_kiocb
*req
)
2742 __must_hold(&req
->ctx
->completion_lock
)
2743 __must_hold(&req
->ctx
->timeout_lock
)
2745 struct io_kiocb
*link
= req
->link
;
2747 if (link
&& link
->opcode
== IORING_OP_LINK_TIMEOUT
) {
2748 struct io_timeout_data
*io
= link
->async_data
;
2750 io_remove_next_linked(req
);
2751 link
->timeout
.head
= NULL
;
2752 if (hrtimer_try_to_cancel(&io
->timer
) != -1) {
2753 list_del(&link
->timeout
.list
);
2760 static void io_fail_links(struct io_kiocb
*req
)
2761 __must_hold(&req
->ctx
->completion_lock
)
2763 struct io_kiocb
*nxt
, *link
= req
->link
;
2764 bool ignore_cqes
= req
->flags
& REQ_F_SKIP_LINK_CQES
;
2768 long res
= -ECANCELED
;
2770 if (link
->flags
& REQ_F_FAIL
)
2771 res
= link
->cqe
.res
;
2776 trace_io_uring_fail_link(req
->ctx
, req
, req
->cqe
.user_data
,
2780 link
->flags
|= REQ_F_CQE_SKIP
;
2782 link
->flags
&= ~REQ_F_CQE_SKIP
;
2783 __io_req_complete_post(link
, res
, 0);
2788 static bool io_disarm_next(struct io_kiocb
*req
)
2789 __must_hold(&req
->ctx
->completion_lock
)
2791 struct io_kiocb
*link
= NULL
;
2792 bool posted
= false;
2794 if (req
->flags
& REQ_F_ARM_LTIMEOUT
) {
2796 req
->flags
&= ~REQ_F_ARM_LTIMEOUT
;
2797 if (link
&& link
->opcode
== IORING_OP_LINK_TIMEOUT
) {
2798 io_remove_next_linked(req
);
2799 io_req_tw_post_queue(link
, -ECANCELED
, 0);
2802 } else if (req
->flags
& REQ_F_LINK_TIMEOUT
) {
2803 struct io_ring_ctx
*ctx
= req
->ctx
;
2805 spin_lock_irq(&ctx
->timeout_lock
);
2806 link
= io_disarm_linked_timeout(req
);
2807 spin_unlock_irq(&ctx
->timeout_lock
);
2810 io_req_tw_post_queue(link
, -ECANCELED
, 0);
2813 if (unlikely((req
->flags
& REQ_F_FAIL
) &&
2814 !(req
->flags
& REQ_F_HARDLINK
))) {
2815 posted
|= (req
->link
!= NULL
);
2821 static void __io_req_find_next_prep(struct io_kiocb
*req
)
2823 struct io_ring_ctx
*ctx
= req
->ctx
;
2826 spin_lock(&ctx
->completion_lock
);
2827 posted
= io_disarm_next(req
);
2828 io_commit_cqring(ctx
);
2829 spin_unlock(&ctx
->completion_lock
);
2831 io_cqring_ev_posted(ctx
);
2834 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
2836 struct io_kiocb
*nxt
;
2839 * If LINK is set, we have dependent requests in this chain. If we
2840 * didn't fail this request, queue the first one up, moving any other
2841 * dependencies to the next request. In case of failure, fail the rest
2844 if (unlikely(req
->flags
& IO_DISARM_MASK
))
2845 __io_req_find_next_prep(req
);
2851 static void ctx_flush_and_put(struct io_ring_ctx
*ctx
, bool *locked
)
2855 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
2856 atomic_andnot(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
2858 io_submit_flush_completions(ctx
);
2859 mutex_unlock(&ctx
->uring_lock
);
2862 percpu_ref_put(&ctx
->refs
);
2865 static inline void ctx_commit_and_unlock(struct io_ring_ctx
*ctx
)
2867 io_commit_cqring(ctx
);
2868 spin_unlock(&ctx
->completion_lock
);
2869 io_cqring_ev_posted(ctx
);
2872 static void handle_prev_tw_list(struct io_wq_work_node
*node
,
2873 struct io_ring_ctx
**ctx
, bool *uring_locked
)
2875 if (*ctx
&& !*uring_locked
)
2876 spin_lock(&(*ctx
)->completion_lock
);
2879 struct io_wq_work_node
*next
= node
->next
;
2880 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
2883 prefetch(container_of(next
, struct io_kiocb
, io_task_work
.node
));
2885 if (req
->ctx
!= *ctx
) {
2886 if (unlikely(!*uring_locked
&& *ctx
))
2887 ctx_commit_and_unlock(*ctx
);
2889 ctx_flush_and_put(*ctx
, uring_locked
);
2891 /* if not contended, grab and improve batching */
2892 *uring_locked
= mutex_trylock(&(*ctx
)->uring_lock
);
2893 percpu_ref_get(&(*ctx
)->refs
);
2894 if (unlikely(!*uring_locked
))
2895 spin_lock(&(*ctx
)->completion_lock
);
2897 if (likely(*uring_locked
))
2898 req
->io_task_work
.func(req
, uring_locked
);
2900 __io_req_complete_post(req
, req
->cqe
.res
,
2901 io_put_kbuf_comp(req
));
2905 if (unlikely(!*uring_locked
))
2906 ctx_commit_and_unlock(*ctx
);
2909 static void handle_tw_list(struct io_wq_work_node
*node
,
2910 struct io_ring_ctx
**ctx
, bool *locked
)
2913 struct io_wq_work_node
*next
= node
->next
;
2914 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
2917 prefetch(container_of(next
, struct io_kiocb
, io_task_work
.node
));
2919 if (req
->ctx
!= *ctx
) {
2920 ctx_flush_and_put(*ctx
, locked
);
2922 /* if not contended, grab and improve batching */
2923 *locked
= mutex_trylock(&(*ctx
)->uring_lock
);
2924 percpu_ref_get(&(*ctx
)->refs
);
2926 req
->io_task_work
.func(req
, locked
);
2931 static void tctx_task_work(struct callback_head
*cb
)
2933 bool uring_locked
= false;
2934 struct io_ring_ctx
*ctx
= NULL
;
2935 struct io_uring_task
*tctx
= container_of(cb
, struct io_uring_task
,
2939 struct io_wq_work_node
*node1
, *node2
;
2941 spin_lock_irq(&tctx
->task_lock
);
2942 node1
= tctx
->prio_task_list
.first
;
2943 node2
= tctx
->task_list
.first
;
2944 INIT_WQ_LIST(&tctx
->task_list
);
2945 INIT_WQ_LIST(&tctx
->prio_task_list
);
2946 if (!node2
&& !node1
)
2947 tctx
->task_running
= false;
2948 spin_unlock_irq(&tctx
->task_lock
);
2949 if (!node2
&& !node1
)
2953 handle_prev_tw_list(node1
, &ctx
, &uring_locked
);
2955 handle_tw_list(node2
, &ctx
, &uring_locked
);
2958 if (data_race(!tctx
->task_list
.first
) &&
2959 data_race(!tctx
->prio_task_list
.first
) && uring_locked
)
2960 io_submit_flush_completions(ctx
);
2963 ctx_flush_and_put(ctx
, &uring_locked
);
2965 /* relaxed read is enough as only the task itself sets ->in_idle */
2966 if (unlikely(atomic_read(&tctx
->in_idle
)))
2967 io_uring_drop_tctx_refs(current
);
2970 static void __io_req_task_work_add(struct io_kiocb
*req
,
2971 struct io_uring_task
*tctx
,
2972 struct io_wq_work_list
*list
)
2974 struct io_ring_ctx
*ctx
= req
->ctx
;
2975 struct io_wq_work_node
*node
;
2976 unsigned long flags
;
2979 spin_lock_irqsave(&tctx
->task_lock
, flags
);
2980 wq_list_add_tail(&req
->io_task_work
.node
, list
);
2981 running
= tctx
->task_running
;
2983 tctx
->task_running
= true;
2984 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
2986 /* task_work already pending, we're done */
2990 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
2991 atomic_or(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
2993 if (likely(!task_work_add(req
->task
, &tctx
->task_work
, ctx
->notify_method
)))
2996 spin_lock_irqsave(&tctx
->task_lock
, flags
);
2997 tctx
->task_running
= false;
2998 node
= wq_list_merge(&tctx
->prio_task_list
, &tctx
->task_list
);
2999 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
3002 req
= container_of(node
, struct io_kiocb
, io_task_work
.node
);
3004 if (llist_add(&req
->io_task_work
.fallback_node
,
3005 &req
->ctx
->fallback_llist
))
3006 schedule_delayed_work(&req
->ctx
->fallback_work
, 1);
3010 static void io_req_task_work_add(struct io_kiocb
*req
)
3012 struct io_uring_task
*tctx
= req
->task
->io_uring
;
3014 __io_req_task_work_add(req
, tctx
, &tctx
->task_list
);
3017 static void io_req_task_prio_work_add(struct io_kiocb
*req
)
3019 struct io_uring_task
*tctx
= req
->task
->io_uring
;
3021 if (req
->ctx
->flags
& IORING_SETUP_SQPOLL
)
3022 __io_req_task_work_add(req
, tctx
, &tctx
->prio_task_list
);
3024 __io_req_task_work_add(req
, tctx
, &tctx
->task_list
);
3027 static void io_req_tw_post(struct io_kiocb
*req
, bool *locked
)
3029 io_req_complete_post(req
, req
->cqe
.res
, req
->cqe
.flags
);
3032 static void io_req_tw_post_queue(struct io_kiocb
*req
, s32 res
, u32 cflags
)
3035 req
->cqe
.flags
= cflags
;
3036 req
->io_task_work
.func
= io_req_tw_post
;
3037 io_req_task_work_add(req
);
3040 static void io_req_task_cancel(struct io_kiocb
*req
, bool *locked
)
3042 /* not needed for normal modes, but SQPOLL depends on it */
3043 io_tw_lock(req
->ctx
, locked
);
3044 io_req_complete_failed(req
, req
->cqe
.res
);
3047 static void io_req_task_submit(struct io_kiocb
*req
, bool *locked
)
3049 io_tw_lock(req
->ctx
, locked
);
3050 /* req->task == current here, checking PF_EXITING is safe */
3051 if (likely(!(req
->task
->flags
& PF_EXITING
)))
3054 io_req_complete_failed(req
, -EFAULT
);
3057 static void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
)
3060 req
->io_task_work
.func
= io_req_task_cancel
;
3061 io_req_task_work_add(req
);
3064 static void io_req_task_queue(struct io_kiocb
*req
)
3066 req
->io_task_work
.func
= io_req_task_submit
;
3067 io_req_task_work_add(req
);
3070 static void io_req_task_queue_reissue(struct io_kiocb
*req
)
3072 req
->io_task_work
.func
= io_queue_iowq
;
3073 io_req_task_work_add(req
);
3076 static void io_queue_next(struct io_kiocb
*req
)
3078 struct io_kiocb
*nxt
= io_req_find_next(req
);
3081 io_req_task_queue(nxt
);
3084 static void io_free_batch_list(struct io_ring_ctx
*ctx
,
3085 struct io_wq_work_node
*node
)
3086 __must_hold(&ctx
->uring_lock
)
3088 struct task_struct
*task
= NULL
;
3092 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
3095 if (unlikely(req
->flags
& IO_REQ_CLEAN_SLOW_FLAGS
)) {
3096 if (req
->flags
& REQ_F_REFCOUNT
) {
3097 node
= req
->comp_list
.next
;
3098 if (!req_ref_put_and_test(req
))
3101 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
3102 struct async_poll
*apoll
= req
->apoll
;
3104 if (apoll
->double_poll
)
3105 kfree(apoll
->double_poll
);
3106 list_add(&apoll
->poll
.wait
.entry
,
3108 req
->flags
&= ~REQ_F_POLLED
;
3110 if (req
->flags
& IO_REQ_LINK_FLAGS
)
3112 if (unlikely(req
->flags
& IO_REQ_CLEAN_FLAGS
))
3115 if (!(req
->flags
& REQ_F_FIXED_FILE
))
3116 io_put_file(req
->file
);
3118 io_req_put_rsrc_locked(req
, ctx
);
3120 if (req
->task
!= task
) {
3122 io_put_task(task
, task_refs
);
3127 node
= req
->comp_list
.next
;
3128 io_req_add_to_cache(req
, ctx
);
3132 io_put_task(task
, task_refs
);
3135 static void __io_submit_flush_completions(struct io_ring_ctx
*ctx
)
3136 __must_hold(&ctx
->uring_lock
)
3138 struct io_wq_work_node
*node
, *prev
;
3139 struct io_submit_state
*state
= &ctx
->submit_state
;
3141 if (state
->flush_cqes
) {
3142 spin_lock(&ctx
->completion_lock
);
3143 wq_list_for_each(node
, prev
, &state
->compl_reqs
) {
3144 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
3147 if (!(req
->flags
& REQ_F_CQE_SKIP
))
3148 __io_fill_cqe_req(ctx
, req
);
3151 io_commit_cqring(ctx
);
3152 spin_unlock(&ctx
->completion_lock
);
3153 io_cqring_ev_posted(ctx
);
3154 state
->flush_cqes
= false;
3157 io_free_batch_list(ctx
, state
->compl_reqs
.first
);
3158 INIT_WQ_LIST(&state
->compl_reqs
);
3162 * Drop reference to request, return next in chain (if there is one) if this
3163 * was the last reference to this request.
3165 static inline struct io_kiocb
*io_put_req_find_next(struct io_kiocb
*req
)
3167 struct io_kiocb
*nxt
= NULL
;
3169 if (req_ref_put_and_test(req
)) {
3170 if (unlikely(req
->flags
& IO_REQ_LINK_FLAGS
))
3171 nxt
= io_req_find_next(req
);
3177 static inline void io_put_req(struct io_kiocb
*req
)
3179 if (req_ref_put_and_test(req
)) {
3185 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
3187 /* See comment at the top of this file */
3189 return __io_cqring_events(ctx
);
3192 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
3194 struct io_rings
*rings
= ctx
->rings
;
3196 /* make sure SQ entry isn't read before tail */
3197 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
3200 static inline bool io_run_task_work(void)
3202 if (test_thread_flag(TIF_NOTIFY_SIGNAL
) || task_work_pending(current
)) {
3203 __set_current_state(TASK_RUNNING
);
3204 clear_notify_signal();
3205 if (task_work_pending(current
))
3213 static int io_do_iopoll(struct io_ring_ctx
*ctx
, bool force_nonspin
)
3215 struct io_wq_work_node
*pos
, *start
, *prev
;
3216 unsigned int poll_flags
= BLK_POLL_NOSLEEP
;
3217 DEFINE_IO_COMP_BATCH(iob
);
3221 * Only spin for completions if we don't have multiple devices hanging
3222 * off our complete list.
3224 if (ctx
->poll_multi_queue
|| force_nonspin
)
3225 poll_flags
|= BLK_POLL_ONESHOT
;
3227 wq_list_for_each(pos
, start
, &ctx
->iopoll_list
) {
3228 struct io_kiocb
*req
= container_of(pos
, struct io_kiocb
, comp_list
);
3229 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3233 * Move completed and retryable entries to our local lists.
3234 * If we find a request that requires polling, break out
3235 * and complete those lists first, if we have entries there.
3237 if (READ_ONCE(req
->iopoll_completed
))
3240 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, &iob
, poll_flags
);
3241 if (unlikely(ret
< 0))
3244 poll_flags
|= BLK_POLL_ONESHOT
;
3246 /* iopoll may have completed current req */
3247 if (!rq_list_empty(iob
.req_list
) ||
3248 READ_ONCE(req
->iopoll_completed
))
3252 if (!rq_list_empty(iob
.req_list
))
3258 wq_list_for_each_resume(pos
, prev
) {
3259 struct io_kiocb
*req
= container_of(pos
, struct io_kiocb
, comp_list
);
3261 /* order with io_complete_rw_iopoll(), e.g. ->result updates */
3262 if (!smp_load_acquire(&req
->iopoll_completed
))
3265 if (unlikely(req
->flags
& REQ_F_CQE_SKIP
))
3268 req
->cqe
.flags
= io_put_kbuf(req
, 0);
3269 __io_fill_cqe_req(req
->ctx
, req
);
3272 if (unlikely(!nr_events
))
3275 io_commit_cqring(ctx
);
3276 io_cqring_ev_posted_iopoll(ctx
);
3277 pos
= start
? start
->next
: ctx
->iopoll_list
.first
;
3278 wq_list_cut(&ctx
->iopoll_list
, prev
, start
);
3279 io_free_batch_list(ctx
, pos
);
3284 * We can't just wait for polled events to come to us, we have to actively
3285 * find and complete them.
3287 static __cold
void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
3289 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
3292 mutex_lock(&ctx
->uring_lock
);
3293 while (!wq_list_empty(&ctx
->iopoll_list
)) {
3294 /* let it sleep and repeat later if can't complete a request */
3295 if (io_do_iopoll(ctx
, true) == 0)
3298 * Ensure we allow local-to-the-cpu processing to take place,
3299 * in this case we need to ensure that we reap all events.
3300 * Also let task_work, etc. to progress by releasing the mutex
3302 if (need_resched()) {
3303 mutex_unlock(&ctx
->uring_lock
);
3305 mutex_lock(&ctx
->uring_lock
);
3308 mutex_unlock(&ctx
->uring_lock
);
3311 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
3313 unsigned int nr_events
= 0;
3315 unsigned long check_cq
;
3318 * Don't enter poll loop if we already have events pending.
3319 * If we do, we can potentially be spinning for commands that
3320 * already triggered a CQE (eg in error).
3322 check_cq
= READ_ONCE(ctx
->check_cq
);
3323 if (check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
))
3324 __io_cqring_overflow_flush(ctx
, false);
3325 if (io_cqring_events(ctx
))
3329 * Similarly do not spin if we have not informed the user of any
3332 if (unlikely(check_cq
& BIT(IO_CHECK_CQ_DROPPED_BIT
)))
3337 * If a submit got punted to a workqueue, we can have the
3338 * application entering polling for a command before it gets
3339 * issued. That app will hold the uring_lock for the duration
3340 * of the poll right here, so we need to take a breather every
3341 * now and then to ensure that the issue has a chance to add
3342 * the poll to the issued list. Otherwise we can spin here
3343 * forever, while the workqueue is stuck trying to acquire the
3346 if (wq_list_empty(&ctx
->iopoll_list
)) {
3347 u32 tail
= ctx
->cached_cq_tail
;
3349 mutex_unlock(&ctx
->uring_lock
);
3351 mutex_lock(&ctx
->uring_lock
);
3353 /* some requests don't go through iopoll_list */
3354 if (tail
!= ctx
->cached_cq_tail
||
3355 wq_list_empty(&ctx
->iopoll_list
))
3358 ret
= io_do_iopoll(ctx
, !min
);
3363 } while (nr_events
< min
&& !need_resched());
3368 static void kiocb_end_write(struct io_kiocb
*req
)
3371 * Tell lockdep we inherited freeze protection from submission
3374 if (req
->flags
& REQ_F_ISREG
) {
3375 struct super_block
*sb
= file_inode(req
->file
)->i_sb
;
3377 __sb_writers_acquired(sb
, SB_FREEZE_WRITE
);
3383 static bool io_resubmit_prep(struct io_kiocb
*req
)
3385 struct io_async_rw
*rw
= req
->async_data
;
3387 if (!req_has_async_data(req
))
3388 return !io_req_prep_async(req
);
3389 iov_iter_restore(&rw
->s
.iter
, &rw
->s
.iter_state
);
3393 static bool io_rw_should_reissue(struct io_kiocb
*req
)
3395 umode_t mode
= file_inode(req
->file
)->i_mode
;
3396 struct io_ring_ctx
*ctx
= req
->ctx
;
3398 if (!S_ISBLK(mode
) && !S_ISREG(mode
))
3400 if ((req
->flags
& REQ_F_NOWAIT
) || (io_wq_current_is_worker() &&
3401 !(ctx
->flags
& IORING_SETUP_IOPOLL
)))
3404 * If ref is dying, we might be running poll reap from the exit work.
3405 * Don't attempt to reissue from that path, just let it fail with
3408 if (percpu_ref_is_dying(&ctx
->refs
))
3411 * Play it safe and assume not safe to re-import and reissue if we're
3412 * not in the original thread group (or in task context).
3414 if (!same_thread_group(req
->task
, current
) || !in_task())
3419 static bool io_resubmit_prep(struct io_kiocb
*req
)
3423 static bool io_rw_should_reissue(struct io_kiocb
*req
)
3429 static bool __io_complete_rw_common(struct io_kiocb
*req
, long res
)
3431 if (req
->rw
.kiocb
.ki_flags
& IOCB_WRITE
) {
3432 kiocb_end_write(req
);
3433 fsnotify_modify(req
->file
);
3435 fsnotify_access(req
->file
);
3437 if (unlikely(res
!= req
->cqe
.res
)) {
3438 if ((res
== -EAGAIN
|| res
== -EOPNOTSUPP
) &&
3439 io_rw_should_reissue(req
)) {
3440 req
->flags
|= REQ_F_REISSUE
| REQ_F_PARTIAL_IO
;
3449 static inline void io_req_task_complete(struct io_kiocb
*req
, bool *locked
)
3451 int res
= req
->cqe
.res
;
3454 io_req_complete_state(req
, res
, io_put_kbuf(req
, 0));
3455 io_req_add_compl_list(req
);
3457 io_req_complete_post(req
, res
,
3458 io_put_kbuf(req
, IO_URING_F_UNLOCKED
));
3462 static void __io_complete_rw(struct io_kiocb
*req
, long res
,
3463 unsigned int issue_flags
)
3465 if (__io_complete_rw_common(req
, res
))
3467 __io_req_complete(req
, issue_flags
, req
->cqe
.res
,
3468 io_put_kbuf(req
, issue_flags
));
3471 static void io_complete_rw(struct kiocb
*kiocb
, long res
)
3473 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
3475 if (__io_complete_rw_common(req
, res
))
3478 req
->io_task_work
.func
= io_req_task_complete
;
3479 io_req_task_prio_work_add(req
);
3482 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
)
3484 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
3486 if (kiocb
->ki_flags
& IOCB_WRITE
)
3487 kiocb_end_write(req
);
3488 if (unlikely(res
!= req
->cqe
.res
)) {
3489 if (res
== -EAGAIN
&& io_rw_should_reissue(req
)) {
3490 req
->flags
|= REQ_F_REISSUE
| REQ_F_PARTIAL_IO
;
3496 /* order with io_iopoll_complete() checking ->iopoll_completed */
3497 smp_store_release(&req
->iopoll_completed
, 1);
3501 * After the iocb has been issued, it's safe to be found on the poll list.
3502 * Adding the kiocb to the list AFTER submission ensures that we don't
3503 * find it from a io_do_iopoll() thread before the issuer is done
3504 * accessing the kiocb cookie.
3506 static void io_iopoll_req_issued(struct io_kiocb
*req
, unsigned int issue_flags
)
3508 struct io_ring_ctx
*ctx
= req
->ctx
;
3509 const bool needs_lock
= issue_flags
& IO_URING_F_UNLOCKED
;
3511 /* workqueue context doesn't hold uring_lock, grab it now */
3512 if (unlikely(needs_lock
))
3513 mutex_lock(&ctx
->uring_lock
);
3516 * Track whether we have multiple files in our lists. This will impact
3517 * how we do polling eventually, not spinning if we're on potentially
3518 * different devices.
3520 if (wq_list_empty(&ctx
->iopoll_list
)) {
3521 ctx
->poll_multi_queue
= false;
3522 } else if (!ctx
->poll_multi_queue
) {
3523 struct io_kiocb
*list_req
;
3525 list_req
= container_of(ctx
->iopoll_list
.first
, struct io_kiocb
,
3527 if (list_req
->file
!= req
->file
)
3528 ctx
->poll_multi_queue
= true;
3532 * For fast devices, IO may have already completed. If it has, add
3533 * it to the front so we find it first.
3535 if (READ_ONCE(req
->iopoll_completed
))
3536 wq_list_add_head(&req
->comp_list
, &ctx
->iopoll_list
);
3538 wq_list_add_tail(&req
->comp_list
, &ctx
->iopoll_list
);
3540 if (unlikely(needs_lock
)) {
3542 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
3543 * in sq thread task context or in io worker task context. If
3544 * current task context is sq thread, we don't need to check
3545 * whether should wake up sq thread.
3547 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
3548 wq_has_sleeper(&ctx
->sq_data
->wait
))
3549 wake_up(&ctx
->sq_data
->wait
);
3551 mutex_unlock(&ctx
->uring_lock
);
3555 static bool io_bdev_nowait(struct block_device
*bdev
)
3557 return !bdev
|| blk_queue_nowait(bdev_get_queue(bdev
));
3561 * If we tracked the file through the SCM inflight mechanism, we could support
3562 * any file. For now, just ensure that anything potentially problematic is done
3565 static bool __io_file_supports_nowait(struct file
*file
, umode_t mode
)
3567 if (S_ISBLK(mode
)) {
3568 if (IS_ENABLED(CONFIG_BLOCK
) &&
3569 io_bdev_nowait(I_BDEV(file
->f_mapping
->host
)))
3575 if (S_ISREG(mode
)) {
3576 if (IS_ENABLED(CONFIG_BLOCK
) &&
3577 io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
3578 file
->f_op
!= &io_uring_fops
)
3583 /* any ->read/write should understand O_NONBLOCK */
3584 if (file
->f_flags
& O_NONBLOCK
)
3586 return file
->f_mode
& FMODE_NOWAIT
;
3590 * If we tracked the file through the SCM inflight mechanism, we could support
3591 * any file. For now, just ensure that anything potentially problematic is done
3594 static unsigned int io_file_get_flags(struct file
*file
)
3596 umode_t mode
= file_inode(file
)->i_mode
;
3597 unsigned int res
= 0;
3601 if (__io_file_supports_nowait(file
, mode
))
3603 if (io_file_need_scm(file
))
3608 static inline bool io_file_supports_nowait(struct io_kiocb
*req
)
3610 return req
->flags
& REQ_F_SUPPORT_NOWAIT
;
3613 static int io_prep_rw(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3615 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3619 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
3620 /* used for fixed read/write too - just read unconditionally */
3621 req
->buf_index
= READ_ONCE(sqe
->buf_index
);
3623 if (req
->opcode
== IORING_OP_READ_FIXED
||
3624 req
->opcode
== IORING_OP_WRITE_FIXED
) {
3625 struct io_ring_ctx
*ctx
= req
->ctx
;
3628 if (unlikely(req
->buf_index
>= ctx
->nr_user_bufs
))
3630 index
= array_index_nospec(req
->buf_index
, ctx
->nr_user_bufs
);
3631 req
->imu
= ctx
->user_bufs
[index
];
3632 io_req_set_rsrc_node(req
, ctx
, 0);
3635 ioprio
= READ_ONCE(sqe
->ioprio
);
3637 ret
= ioprio_check_cap(ioprio
);
3641 kiocb
->ki_ioprio
= ioprio
;
3643 kiocb
->ki_ioprio
= get_current_ioprio();
3646 req
->rw
.addr
= READ_ONCE(sqe
->addr
);
3647 req
->rw
.len
= READ_ONCE(sqe
->len
);
3648 req
->rw
.flags
= READ_ONCE(sqe
->rw_flags
);
3652 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
3658 case -ERESTARTNOINTR
:
3659 case -ERESTARTNOHAND
:
3660 case -ERESTART_RESTARTBLOCK
:
3662 * We can't just restart the syscall, since previously
3663 * submitted sqes may already be in progress. Just fail this
3669 kiocb
->ki_complete(kiocb
, ret
);
3673 static inline loff_t
*io_kiocb_update_pos(struct io_kiocb
*req
)
3675 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3677 if (kiocb
->ki_pos
!= -1)
3678 return &kiocb
->ki_pos
;
3680 if (!(req
->file
->f_mode
& FMODE_STREAM
)) {
3681 req
->flags
|= REQ_F_CUR_POS
;
3682 kiocb
->ki_pos
= req
->file
->f_pos
;
3683 return &kiocb
->ki_pos
;
3690 static void kiocb_done(struct io_kiocb
*req
, ssize_t ret
,
3691 unsigned int issue_flags
)
3693 struct io_async_rw
*io
= req
->async_data
;
3695 /* add previously done IO, if any */
3696 if (req_has_async_data(req
) && io
->bytes_done
> 0) {
3698 ret
= io
->bytes_done
;
3700 ret
+= io
->bytes_done
;
3703 if (req
->flags
& REQ_F_CUR_POS
)
3704 req
->file
->f_pos
= req
->rw
.kiocb
.ki_pos
;
3705 if (ret
>= 0 && (req
->rw
.kiocb
.ki_complete
== io_complete_rw
))
3706 __io_complete_rw(req
, ret
, issue_flags
);
3708 io_rw_done(&req
->rw
.kiocb
, ret
);
3710 if (req
->flags
& REQ_F_REISSUE
) {
3711 req
->flags
&= ~REQ_F_REISSUE
;
3712 if (io_resubmit_prep(req
))
3713 io_req_task_queue_reissue(req
);
3715 io_req_task_queue_fail(req
, ret
);
3719 static int __io_import_fixed(struct io_kiocb
*req
, int rw
, struct iov_iter
*iter
,
3720 struct io_mapped_ubuf
*imu
)
3722 size_t len
= req
->rw
.len
;
3723 u64 buf_end
, buf_addr
= req
->rw
.addr
;
3726 if (unlikely(check_add_overflow(buf_addr
, (u64
)len
, &buf_end
)))
3728 /* not inside the mapped region */
3729 if (unlikely(buf_addr
< imu
->ubuf
|| buf_end
> imu
->ubuf_end
))
3733 * May not be a start of buffer, set size appropriately
3734 * and advance us to the beginning.
3736 offset
= buf_addr
- imu
->ubuf
;
3737 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
3741 * Don't use iov_iter_advance() here, as it's really slow for
3742 * using the latter parts of a big fixed buffer - it iterates
3743 * over each segment manually. We can cheat a bit here, because
3746 * 1) it's a BVEC iter, we set it up
3747 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3748 * first and last bvec
3750 * So just find our index, and adjust the iterator afterwards.
3751 * If the offset is within the first bvec (or the whole first
3752 * bvec, just use iov_iter_advance(). This makes it easier
3753 * since we can just skip the first segment, which may not
3754 * be PAGE_SIZE aligned.
3756 const struct bio_vec
*bvec
= imu
->bvec
;
3758 if (offset
<= bvec
->bv_len
) {
3759 iov_iter_advance(iter
, offset
);
3761 unsigned long seg_skip
;
3763 /* skip first vec */
3764 offset
-= bvec
->bv_len
;
3765 seg_skip
= 1 + (offset
>> PAGE_SHIFT
);
3767 iter
->bvec
= bvec
+ seg_skip
;
3768 iter
->nr_segs
-= seg_skip
;
3769 iter
->count
-= bvec
->bv_len
+ offset
;
3770 iter
->iov_offset
= offset
& ~PAGE_MASK
;
3777 static int io_import_fixed(struct io_kiocb
*req
, int rw
, struct iov_iter
*iter
,
3778 unsigned int issue_flags
)
3780 if (WARN_ON_ONCE(!req
->imu
))
3782 return __io_import_fixed(req
, rw
, iter
, req
->imu
);
3785 static int io_buffer_add_list(struct io_ring_ctx
*ctx
,
3786 struct io_buffer_list
*bl
, unsigned int bgid
)
3789 if (bgid
< BGID_ARRAY
)
3792 return xa_err(xa_store(&ctx
->io_bl_xa
, bgid
, bl
, GFP_KERNEL
));
3795 static void __user
*io_provided_buffer_select(struct io_kiocb
*req
, size_t *len
,
3796 struct io_buffer_list
*bl
)
3798 if (!list_empty(&bl
->buf_list
)) {
3799 struct io_buffer
*kbuf
;
3801 kbuf
= list_first_entry(&bl
->buf_list
, struct io_buffer
, list
);
3802 list_del(&kbuf
->list
);
3803 if (*len
> kbuf
->len
)
3805 req
->flags
|= REQ_F_BUFFER_SELECTED
;
3807 req
->buf_index
= kbuf
->bid
;
3808 return u64_to_user_ptr(kbuf
->addr
);
3813 static void __user
*io_ring_buffer_select(struct io_kiocb
*req
, size_t *len
,
3814 struct io_buffer_list
*bl
,
3815 unsigned int issue_flags
)
3817 struct io_uring_buf_ring
*br
= bl
->buf_ring
;
3818 struct io_uring_buf
*buf
;
3819 __u16 head
= bl
->head
;
3821 if (unlikely(smp_load_acquire(&br
->tail
) == head
))
3825 if (head
< IO_BUFFER_LIST_BUF_PER_PAGE
) {
3826 buf
= &br
->bufs
[head
];
3828 int off
= head
& (IO_BUFFER_LIST_BUF_PER_PAGE
- 1);
3829 int index
= head
/ IO_BUFFER_LIST_BUF_PER_PAGE
;
3830 buf
= page_address(bl
->buf_pages
[index
]);
3833 if (*len
> buf
->len
)
3835 req
->flags
|= REQ_F_BUFFER_RING
;
3837 req
->buf_index
= buf
->bid
;
3839 if (issue_flags
& IO_URING_F_UNLOCKED
|| !file_can_poll(req
->file
)) {
3841 * If we came in unlocked, we have no choice but to consume the
3842 * buffer here. This does mean it'll be pinned until the IO
3843 * completes. But coming in unlocked means we're in io-wq
3844 * context, hence there should be no further retry. For the
3845 * locked case, the caller must ensure to call the commit when
3846 * the transfer completes (or if we get -EAGAIN and must poll
3849 req
->buf_list
= NULL
;
3852 return u64_to_user_ptr(buf
->addr
);
3855 static void __user
*io_buffer_select(struct io_kiocb
*req
, size_t *len
,
3856 unsigned int issue_flags
)
3858 struct io_ring_ctx
*ctx
= req
->ctx
;
3859 struct io_buffer_list
*bl
;
3860 void __user
*ret
= NULL
;
3862 io_ring_submit_lock(req
->ctx
, issue_flags
);
3864 bl
= io_buffer_get_list(ctx
, req
->buf_index
);
3866 if (bl
->buf_nr_pages
)
3867 ret
= io_ring_buffer_select(req
, len
, bl
, issue_flags
);
3869 ret
= io_provided_buffer_select(req
, len
, bl
);
3871 io_ring_submit_unlock(req
->ctx
, issue_flags
);
3875 #ifdef CONFIG_COMPAT
3876 static ssize_t
io_compat_import(struct io_kiocb
*req
, struct iovec
*iov
,
3877 unsigned int issue_flags
)
3879 struct compat_iovec __user
*uiov
;
3880 compat_ssize_t clen
;
3884 uiov
= u64_to_user_ptr(req
->rw
.addr
);
3885 if (!access_ok(uiov
, sizeof(*uiov
)))
3887 if (__get_user(clen
, &uiov
->iov_len
))
3893 buf
= io_buffer_select(req
, &len
, issue_flags
);
3896 req
->rw
.addr
= (unsigned long) buf
;
3897 iov
[0].iov_base
= buf
;
3898 req
->rw
.len
= iov
[0].iov_len
= (compat_size_t
) len
;
3903 static ssize_t
__io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3904 unsigned int issue_flags
)
3906 struct iovec __user
*uiov
= u64_to_user_ptr(req
->rw
.addr
);
3910 if (copy_from_user(iov
, uiov
, sizeof(*uiov
)))
3913 len
= iov
[0].iov_len
;
3916 buf
= io_buffer_select(req
, &len
, issue_flags
);
3919 req
->rw
.addr
= (unsigned long) buf
;
3920 iov
[0].iov_base
= buf
;
3921 req
->rw
.len
= iov
[0].iov_len
= len
;
3925 static ssize_t
io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3926 unsigned int issue_flags
)
3928 if (req
->flags
& (REQ_F_BUFFER_SELECTED
|REQ_F_BUFFER_RING
)) {
3929 iov
[0].iov_base
= u64_to_user_ptr(req
->rw
.addr
);
3930 iov
[0].iov_len
= req
->rw
.len
;
3933 if (req
->rw
.len
!= 1)
3936 #ifdef CONFIG_COMPAT
3937 if (req
->ctx
->compat
)
3938 return io_compat_import(req
, iov
, issue_flags
);
3941 return __io_iov_buffer_select(req
, iov
, issue_flags
);
3944 static inline bool io_do_buffer_select(struct io_kiocb
*req
)
3946 if (!(req
->flags
& REQ_F_BUFFER_SELECT
))
3948 return !(req
->flags
& (REQ_F_BUFFER_SELECTED
|REQ_F_BUFFER_RING
));
3951 static struct iovec
*__io_import_iovec(int rw
, struct io_kiocb
*req
,
3952 struct io_rw_state
*s
,
3953 unsigned int issue_flags
)
3955 struct iov_iter
*iter
= &s
->iter
;
3956 u8 opcode
= req
->opcode
;
3957 struct iovec
*iovec
;
3962 if (opcode
== IORING_OP_READ_FIXED
|| opcode
== IORING_OP_WRITE_FIXED
) {
3963 ret
= io_import_fixed(req
, rw
, iter
, issue_flags
);
3965 return ERR_PTR(ret
);
3969 buf
= u64_to_user_ptr(req
->rw
.addr
);
3970 sqe_len
= req
->rw
.len
;
3972 if (opcode
== IORING_OP_READ
|| opcode
== IORING_OP_WRITE
) {
3973 if (io_do_buffer_select(req
)) {
3974 buf
= io_buffer_select(req
, &sqe_len
, issue_flags
);
3976 return ERR_PTR(-ENOBUFS
);
3977 req
->rw
.addr
= (unsigned long) buf
;
3978 req
->rw
.len
= sqe_len
;
3981 ret
= import_single_range(rw
, buf
, sqe_len
, s
->fast_iov
, iter
);
3983 return ERR_PTR(ret
);
3987 iovec
= s
->fast_iov
;
3988 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3989 ret
= io_iov_buffer_select(req
, iovec
, issue_flags
);
3991 return ERR_PTR(ret
);
3992 iov_iter_init(iter
, rw
, iovec
, 1, iovec
->iov_len
);
3996 ret
= __import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, &iovec
, iter
,
3998 if (unlikely(ret
< 0))
3999 return ERR_PTR(ret
);
4003 static inline int io_import_iovec(int rw
, struct io_kiocb
*req
,
4004 struct iovec
**iovec
, struct io_rw_state
*s
,
4005 unsigned int issue_flags
)
4007 *iovec
= __io_import_iovec(rw
, req
, s
, issue_flags
);
4008 if (unlikely(IS_ERR(*iovec
)))
4009 return PTR_ERR(*iovec
);
4011 iov_iter_save_state(&s
->iter
, &s
->iter_state
);
4015 static inline loff_t
*io_kiocb_ppos(struct kiocb
*kiocb
)
4017 return (kiocb
->ki_filp
->f_mode
& FMODE_STREAM
) ? NULL
: &kiocb
->ki_pos
;
4021 * For files that don't have ->read_iter() and ->write_iter(), handle them
4022 * by looping over ->read() or ->write() manually.
4024 static ssize_t
loop_rw_iter(int rw
, struct io_kiocb
*req
, struct iov_iter
*iter
)
4026 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
4027 struct file
*file
= req
->file
;
4032 * Don't support polled IO through this interface, and we can't
4033 * support non-blocking either. For the latter, this just causes
4034 * the kiocb to be handled from an async context.
4036 if (kiocb
->ki_flags
& IOCB_HIPRI
)
4038 if ((kiocb
->ki_flags
& IOCB_NOWAIT
) &&
4039 !(kiocb
->ki_filp
->f_flags
& O_NONBLOCK
))
4042 ppos
= io_kiocb_ppos(kiocb
);
4044 while (iov_iter_count(iter
)) {
4048 if (!iov_iter_is_bvec(iter
)) {
4049 iovec
= iov_iter_iovec(iter
);
4051 iovec
.iov_base
= u64_to_user_ptr(req
->rw
.addr
);
4052 iovec
.iov_len
= req
->rw
.len
;
4056 nr
= file
->f_op
->read(file
, iovec
.iov_base
,
4057 iovec
.iov_len
, ppos
);
4059 nr
= file
->f_op
->write(file
, iovec
.iov_base
,
4060 iovec
.iov_len
, ppos
);
4069 if (!iov_iter_is_bvec(iter
)) {
4070 iov_iter_advance(iter
, nr
);
4077 if (nr
!= iovec
.iov_len
)
4084 static void io_req_map_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
4085 const struct iovec
*fast_iov
, struct iov_iter
*iter
)
4087 struct io_async_rw
*rw
= req
->async_data
;
4089 memcpy(&rw
->s
.iter
, iter
, sizeof(*iter
));
4090 rw
->free_iovec
= iovec
;
4092 /* can only be fixed buffers, no need to do anything */
4093 if (iov_iter_is_bvec(iter
))
4096 unsigned iov_off
= 0;
4098 rw
->s
.iter
.iov
= rw
->s
.fast_iov
;
4099 if (iter
->iov
!= fast_iov
) {
4100 iov_off
= iter
->iov
- fast_iov
;
4101 rw
->s
.iter
.iov
+= iov_off
;
4103 if (rw
->s
.fast_iov
!= fast_iov
)
4104 memcpy(rw
->s
.fast_iov
+ iov_off
, fast_iov
+ iov_off
,
4105 sizeof(struct iovec
) * iter
->nr_segs
);
4107 req
->flags
|= REQ_F_NEED_CLEANUP
;
4111 static inline bool io_alloc_async_data(struct io_kiocb
*req
)
4113 WARN_ON_ONCE(!io_op_defs
[req
->opcode
].async_size
);
4114 req
->async_data
= kmalloc(io_op_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
4115 if (req
->async_data
) {
4116 req
->flags
|= REQ_F_ASYNC_DATA
;
4122 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
4123 struct io_rw_state
*s
, bool force
)
4125 if (!force
&& !io_op_defs
[req
->opcode
].needs_async_setup
)
4127 if (!req_has_async_data(req
)) {
4128 struct io_async_rw
*iorw
;
4130 if (io_alloc_async_data(req
)) {
4135 io_req_map_rw(req
, iovec
, s
->fast_iov
, &s
->iter
);
4136 iorw
= req
->async_data
;
4137 /* we've copied and mapped the iter, ensure state is saved */
4138 iov_iter_save_state(&iorw
->s
.iter
, &iorw
->s
.iter_state
);
4143 static inline int io_rw_prep_async(struct io_kiocb
*req
, int rw
)
4145 struct io_async_rw
*iorw
= req
->async_data
;
4149 /* submission path, ->uring_lock should already be taken */
4150 ret
= io_import_iovec(rw
, req
, &iov
, &iorw
->s
, 0);
4151 if (unlikely(ret
< 0))
4154 iorw
->bytes_done
= 0;
4155 iorw
->free_iovec
= iov
;
4157 req
->flags
|= REQ_F_NEED_CLEANUP
;
4161 static int io_readv_prep_async(struct io_kiocb
*req
)
4163 return io_rw_prep_async(req
, READ
);
4166 static int io_writev_prep_async(struct io_kiocb
*req
)
4168 return io_rw_prep_async(req
, WRITE
);
4172 * This is our waitqueue callback handler, registered through __folio_lock_async()
4173 * when we initially tried to do the IO with the iocb armed our waitqueue.
4174 * This gets called when the page is unlocked, and we generally expect that to
4175 * happen when the page IO is completed and the page is now uptodate. This will
4176 * queue a task_work based retry of the operation, attempting to copy the data
4177 * again. If the latter fails because the page was NOT uptodate, then we will
4178 * do a thread based blocking retry of the operation. That's the unexpected
4181 static int io_async_buf_func(struct wait_queue_entry
*wait
, unsigned mode
,
4182 int sync
, void *arg
)
4184 struct wait_page_queue
*wpq
;
4185 struct io_kiocb
*req
= wait
->private;
4186 struct wait_page_key
*key
= arg
;
4188 wpq
= container_of(wait
, struct wait_page_queue
, wait
);
4190 if (!wake_page_match(wpq
, key
))
4193 req
->rw
.kiocb
.ki_flags
&= ~IOCB_WAITQ
;
4194 list_del_init(&wait
->entry
);
4195 io_req_task_queue(req
);
4200 * This controls whether a given IO request should be armed for async page
4201 * based retry. If we return false here, the request is handed to the async
4202 * worker threads for retry. If we're doing buffered reads on a regular file,
4203 * we prepare a private wait_page_queue entry and retry the operation. This
4204 * will either succeed because the page is now uptodate and unlocked, or it
4205 * will register a callback when the page is unlocked at IO completion. Through
4206 * that callback, io_uring uses task_work to setup a retry of the operation.
4207 * That retry will attempt the buffered read again. The retry will generally
4208 * succeed, or in rare cases where it fails, we then fall back to using the
4209 * async worker threads for a blocking retry.
4211 static bool io_rw_should_retry(struct io_kiocb
*req
)
4213 struct io_async_rw
*rw
= req
->async_data
;
4214 struct wait_page_queue
*wait
= &rw
->wpq
;
4215 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
4217 /* never retry for NOWAIT, we just complete with -EAGAIN */
4218 if (req
->flags
& REQ_F_NOWAIT
)
4221 /* Only for buffered IO */
4222 if (kiocb
->ki_flags
& (IOCB_DIRECT
| IOCB_HIPRI
))
4226 * just use poll if we can, and don't attempt if the fs doesn't
4227 * support callback based unlocks
4229 if (file_can_poll(req
->file
) || !(req
->file
->f_mode
& FMODE_BUF_RASYNC
))
4232 wait
->wait
.func
= io_async_buf_func
;
4233 wait
->wait
.private = req
;
4234 wait
->wait
.flags
= 0;
4235 INIT_LIST_HEAD(&wait
->wait
.entry
);
4236 kiocb
->ki_flags
|= IOCB_WAITQ
;
4237 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
4238 kiocb
->ki_waitq
= wait
;
4242 static inline int io_iter_do_read(struct io_kiocb
*req
, struct iov_iter
*iter
)
4244 if (likely(req
->file
->f_op
->read_iter
))
4245 return call_read_iter(req
->file
, &req
->rw
.kiocb
, iter
);
4246 else if (req
->file
->f_op
->read
)
4247 return loop_rw_iter(READ
, req
, iter
);
4252 static bool need_read_all(struct io_kiocb
*req
)
4254 return req
->flags
& REQ_F_ISREG
||
4255 S_ISBLK(file_inode(req
->file
)->i_mode
);
4258 static int io_rw_init_file(struct io_kiocb
*req
, fmode_t mode
)
4260 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
4261 struct io_ring_ctx
*ctx
= req
->ctx
;
4262 struct file
*file
= req
->file
;
4265 if (unlikely(!file
|| !(file
->f_mode
& mode
)))
4268 if (!io_req_ffs_set(req
))
4269 req
->flags
|= io_file_get_flags(file
) << REQ_F_SUPPORT_NOWAIT_BIT
;
4271 kiocb
->ki_flags
= iocb_flags(file
);
4272 ret
= kiocb_set_rw_flags(kiocb
, req
->rw
.flags
);
4277 * If the file is marked O_NONBLOCK, still allow retry for it if it
4278 * supports async. Otherwise it's impossible to use O_NONBLOCK files
4279 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
4281 if ((kiocb
->ki_flags
& IOCB_NOWAIT
) ||
4282 ((file
->f_flags
& O_NONBLOCK
) && !io_file_supports_nowait(req
)))
4283 req
->flags
|= REQ_F_NOWAIT
;
4285 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
4286 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) || !file
->f_op
->iopoll
)
4289 kiocb
->private = NULL
;
4290 kiocb
->ki_flags
|= IOCB_HIPRI
| IOCB_ALLOC_CACHE
;
4291 kiocb
->ki_complete
= io_complete_rw_iopoll
;
4292 req
->iopoll_completed
= 0;
4294 if (kiocb
->ki_flags
& IOCB_HIPRI
)
4296 kiocb
->ki_complete
= io_complete_rw
;
4302 static int io_read(struct io_kiocb
*req
, unsigned int issue_flags
)
4304 struct io_rw_state __s
, *s
= &__s
;
4305 struct iovec
*iovec
;
4306 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
4307 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4308 struct io_async_rw
*rw
;
4312 if (!req_has_async_data(req
)) {
4313 ret
= io_import_iovec(READ
, req
, &iovec
, s
, issue_flags
);
4314 if (unlikely(ret
< 0))
4318 * Safe and required to re-import if we're using provided
4319 * buffers, as we dropped the selected one before retry.
4321 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4322 ret
= io_import_iovec(READ
, req
, &iovec
, s
, issue_flags
);
4323 if (unlikely(ret
< 0))
4327 rw
= req
->async_data
;
4330 * We come here from an earlier attempt, restore our state to
4331 * match in case it doesn't. It's cheap enough that we don't
4332 * need to make this conditional.
4334 iov_iter_restore(&s
->iter
, &s
->iter_state
);
4337 ret
= io_rw_init_file(req
, FMODE_READ
);
4338 if (unlikely(ret
)) {
4342 req
->cqe
.res
= iov_iter_count(&s
->iter
);
4344 if (force_nonblock
) {
4345 /* If the file doesn't support async, just async punt */
4346 if (unlikely(!io_file_supports_nowait(req
))) {
4347 ret
= io_setup_async_rw(req
, iovec
, s
, true);
4348 return ret
?: -EAGAIN
;
4350 kiocb
->ki_flags
|= IOCB_NOWAIT
;
4352 /* Ensure we clear previously set non-block flag */
4353 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
4356 ppos
= io_kiocb_update_pos(req
);
4358 ret
= rw_verify_area(READ
, req
->file
, ppos
, req
->cqe
.res
);
4359 if (unlikely(ret
)) {
4364 ret
= io_iter_do_read(req
, &s
->iter
);
4366 if (ret
== -EAGAIN
|| (req
->flags
& REQ_F_REISSUE
)) {
4367 req
->flags
&= ~REQ_F_REISSUE
;
4368 /* if we can poll, just do that */
4369 if (req
->opcode
== IORING_OP_READ
&& file_can_poll(req
->file
))
4371 /* IOPOLL retry should happen for io-wq threads */
4372 if (!force_nonblock
&& !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4374 /* no retry on NONBLOCK nor RWF_NOWAIT */
4375 if (req
->flags
& REQ_F_NOWAIT
)
4378 } else if (ret
== -EIOCBQUEUED
) {
4380 } else if (ret
== req
->cqe
.res
|| ret
<= 0 || !force_nonblock
||
4381 (req
->flags
& REQ_F_NOWAIT
) || !need_read_all(req
)) {
4382 /* read all, failed, already did sync or don't want to retry */
4387 * Don't depend on the iter state matching what was consumed, or being
4388 * untouched in case of error. Restore it and we'll advance it
4389 * manually if we need to.
4391 iov_iter_restore(&s
->iter
, &s
->iter_state
);
4393 ret2
= io_setup_async_rw(req
, iovec
, s
, true);
4398 rw
= req
->async_data
;
4401 * Now use our persistent iterator and state, if we aren't already.
4402 * We've restored and mapped the iter to match.
4407 * We end up here because of a partial read, either from
4408 * above or inside this loop. Advance the iter by the bytes
4409 * that were consumed.
4411 iov_iter_advance(&s
->iter
, ret
);
4412 if (!iov_iter_count(&s
->iter
))
4414 rw
->bytes_done
+= ret
;
4415 iov_iter_save_state(&s
->iter
, &s
->iter_state
);
4417 /* if we can retry, do so with the callbacks armed */
4418 if (!io_rw_should_retry(req
)) {
4419 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
4424 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
4425 * we get -EIOCBQUEUED, then we'll get a notification when the
4426 * desired page gets unlocked. We can also get a partial read
4427 * here, and if we do, then just retry at the new offset.
4429 ret
= io_iter_do_read(req
, &s
->iter
);
4430 if (ret
== -EIOCBQUEUED
)
4432 /* we got some bytes, but not all. retry. */
4433 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
4434 iov_iter_restore(&s
->iter
, &s
->iter_state
);
4437 kiocb_done(req
, ret
, issue_flags
);
4439 /* it's faster to check here then delegate to kfree */
4445 static int io_write(struct io_kiocb
*req
, unsigned int issue_flags
)
4447 struct io_rw_state __s
, *s
= &__s
;
4448 struct iovec
*iovec
;
4449 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
4450 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4454 if (!req_has_async_data(req
)) {
4455 ret
= io_import_iovec(WRITE
, req
, &iovec
, s
, issue_flags
);
4456 if (unlikely(ret
< 0))
4459 struct io_async_rw
*rw
= req
->async_data
;
4462 iov_iter_restore(&s
->iter
, &s
->iter_state
);
4465 ret
= io_rw_init_file(req
, FMODE_WRITE
);
4466 if (unlikely(ret
)) {
4470 req
->cqe
.res
= iov_iter_count(&s
->iter
);
4472 if (force_nonblock
) {
4473 /* If the file doesn't support async, just async punt */
4474 if (unlikely(!io_file_supports_nowait(req
)))
4477 /* file path doesn't support NOWAIT for non-direct_IO */
4478 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
4479 (req
->flags
& REQ_F_ISREG
))
4482 kiocb
->ki_flags
|= IOCB_NOWAIT
;
4484 /* Ensure we clear previously set non-block flag */
4485 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
4488 ppos
= io_kiocb_update_pos(req
);
4490 ret
= rw_verify_area(WRITE
, req
->file
, ppos
, req
->cqe
.res
);
4495 * Open-code file_start_write here to grab freeze protection,
4496 * which will be released by another thread in
4497 * io_complete_rw(). Fool lockdep by telling it the lock got
4498 * released so that it doesn't complain about the held lock when
4499 * we return to userspace.
4501 if (req
->flags
& REQ_F_ISREG
) {
4502 sb_start_write(file_inode(req
->file
)->i_sb
);
4503 __sb_writers_release(file_inode(req
->file
)->i_sb
,
4506 kiocb
->ki_flags
|= IOCB_WRITE
;
4508 if (likely(req
->file
->f_op
->write_iter
))
4509 ret2
= call_write_iter(req
->file
, kiocb
, &s
->iter
);
4510 else if (req
->file
->f_op
->write
)
4511 ret2
= loop_rw_iter(WRITE
, req
, &s
->iter
);
4515 if (req
->flags
& REQ_F_REISSUE
) {
4516 req
->flags
&= ~REQ_F_REISSUE
;
4521 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
4522 * retry them without IOCB_NOWAIT.
4524 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
4526 /* no retry on NONBLOCK nor RWF_NOWAIT */
4527 if (ret2
== -EAGAIN
&& (req
->flags
& REQ_F_NOWAIT
))
4529 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
4530 /* IOPOLL retry should happen for io-wq threads */
4531 if (ret2
== -EAGAIN
&& (req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4534 kiocb_done(req
, ret2
, issue_flags
);
4537 iov_iter_restore(&s
->iter
, &s
->iter_state
);
4538 ret
= io_setup_async_rw(req
, iovec
, s
, false);
4539 return ret
?: -EAGAIN
;
4542 /* it's reportedly faster than delegating the null check to kfree() */
4548 static int io_renameat_prep(struct io_kiocb
*req
,
4549 const struct io_uring_sqe
*sqe
)
4551 struct io_rename
*ren
= &req
->rename
;
4552 const char __user
*oldf
, *newf
;
4554 if (sqe
->buf_index
|| sqe
->splice_fd_in
)
4556 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
4559 ren
->old_dfd
= READ_ONCE(sqe
->fd
);
4560 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4561 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4562 ren
->new_dfd
= READ_ONCE(sqe
->len
);
4563 ren
->flags
= READ_ONCE(sqe
->rename_flags
);
4565 ren
->oldpath
= getname(oldf
);
4566 if (IS_ERR(ren
->oldpath
))
4567 return PTR_ERR(ren
->oldpath
);
4569 ren
->newpath
= getname(newf
);
4570 if (IS_ERR(ren
->newpath
)) {
4571 putname(ren
->oldpath
);
4572 return PTR_ERR(ren
->newpath
);
4575 req
->flags
|= REQ_F_NEED_CLEANUP
;
4579 static int io_renameat(struct io_kiocb
*req
, unsigned int issue_flags
)
4581 struct io_rename
*ren
= &req
->rename
;
4584 if (issue_flags
& IO_URING_F_NONBLOCK
)
4587 ret
= do_renameat2(ren
->old_dfd
, ren
->oldpath
, ren
->new_dfd
,
4588 ren
->newpath
, ren
->flags
);
4590 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4591 io_req_complete(req
, ret
);
4595 static inline void __io_xattr_finish(struct io_kiocb
*req
)
4597 struct io_xattr
*ix
= &req
->xattr
;
4600 putname(ix
->filename
);
4602 kfree(ix
->ctx
.kname
);
4603 kvfree(ix
->ctx
.kvalue
);
4606 static void io_xattr_finish(struct io_kiocb
*req
, int ret
)
4608 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4610 __io_xattr_finish(req
);
4611 io_req_complete(req
, ret
);
4614 static int __io_getxattr_prep(struct io_kiocb
*req
,
4615 const struct io_uring_sqe
*sqe
)
4617 struct io_xattr
*ix
= &req
->xattr
;
4618 const char __user
*name
;
4621 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
4624 ix
->filename
= NULL
;
4625 ix
->ctx
.kvalue
= NULL
;
4626 name
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4627 ix
->ctx
.cvalue
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4628 ix
->ctx
.size
= READ_ONCE(sqe
->len
);
4629 ix
->ctx
.flags
= READ_ONCE(sqe
->xattr_flags
);
4634 ix
->ctx
.kname
= kmalloc(sizeof(*ix
->ctx
.kname
), GFP_KERNEL
);
4638 ret
= strncpy_from_user(ix
->ctx
.kname
->name
, name
,
4639 sizeof(ix
->ctx
.kname
->name
));
4640 if (!ret
|| ret
== sizeof(ix
->ctx
.kname
->name
))
4643 kfree(ix
->ctx
.kname
);
4647 req
->flags
|= REQ_F_NEED_CLEANUP
;
4651 static int io_fgetxattr_prep(struct io_kiocb
*req
,
4652 const struct io_uring_sqe
*sqe
)
4654 return __io_getxattr_prep(req
, sqe
);
4657 static int io_getxattr_prep(struct io_kiocb
*req
,
4658 const struct io_uring_sqe
*sqe
)
4660 struct io_xattr
*ix
= &req
->xattr
;
4661 const char __user
*path
;
4664 ret
= __io_getxattr_prep(req
, sqe
);
4668 path
= u64_to_user_ptr(READ_ONCE(sqe
->addr3
));
4670 ix
->filename
= getname_flags(path
, LOOKUP_FOLLOW
, NULL
);
4671 if (IS_ERR(ix
->filename
)) {
4672 ret
= PTR_ERR(ix
->filename
);
4673 ix
->filename
= NULL
;
4679 static int io_fgetxattr(struct io_kiocb
*req
, unsigned int issue_flags
)
4681 struct io_xattr
*ix
= &req
->xattr
;
4684 if (issue_flags
& IO_URING_F_NONBLOCK
)
4687 ret
= do_getxattr(mnt_user_ns(req
->file
->f_path
.mnt
),
4688 req
->file
->f_path
.dentry
,
4691 io_xattr_finish(req
, ret
);
4695 static int io_getxattr(struct io_kiocb
*req
, unsigned int issue_flags
)
4697 struct io_xattr
*ix
= &req
->xattr
;
4698 unsigned int lookup_flags
= LOOKUP_FOLLOW
;
4702 if (issue_flags
& IO_URING_F_NONBLOCK
)
4706 ret
= filename_lookup(AT_FDCWD
, ix
->filename
, lookup_flags
, &path
, NULL
);
4708 ret
= do_getxattr(mnt_user_ns(path
.mnt
),
4713 if (retry_estale(ret
, lookup_flags
)) {
4714 lookup_flags
|= LOOKUP_REVAL
;
4719 io_xattr_finish(req
, ret
);
4723 static int __io_setxattr_prep(struct io_kiocb
*req
,
4724 const struct io_uring_sqe
*sqe
)
4726 struct io_xattr
*ix
= &req
->xattr
;
4727 const char __user
*name
;
4730 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
4733 ix
->filename
= NULL
;
4734 name
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4735 ix
->ctx
.cvalue
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4736 ix
->ctx
.kvalue
= NULL
;
4737 ix
->ctx
.size
= READ_ONCE(sqe
->len
);
4738 ix
->ctx
.flags
= READ_ONCE(sqe
->xattr_flags
);
4740 ix
->ctx
.kname
= kmalloc(sizeof(*ix
->ctx
.kname
), GFP_KERNEL
);
4744 ret
= setxattr_copy(name
, &ix
->ctx
);
4746 kfree(ix
->ctx
.kname
);
4750 req
->flags
|= REQ_F_NEED_CLEANUP
;
4754 static int io_setxattr_prep(struct io_kiocb
*req
,
4755 const struct io_uring_sqe
*sqe
)
4757 struct io_xattr
*ix
= &req
->xattr
;
4758 const char __user
*path
;
4761 ret
= __io_setxattr_prep(req
, sqe
);
4765 path
= u64_to_user_ptr(READ_ONCE(sqe
->addr3
));
4767 ix
->filename
= getname_flags(path
, LOOKUP_FOLLOW
, NULL
);
4768 if (IS_ERR(ix
->filename
)) {
4769 ret
= PTR_ERR(ix
->filename
);
4770 ix
->filename
= NULL
;
4776 static int io_fsetxattr_prep(struct io_kiocb
*req
,
4777 const struct io_uring_sqe
*sqe
)
4779 return __io_setxattr_prep(req
, sqe
);
4782 static int __io_setxattr(struct io_kiocb
*req
, unsigned int issue_flags
,
4785 struct io_xattr
*ix
= &req
->xattr
;
4788 ret
= mnt_want_write(path
->mnt
);
4790 ret
= do_setxattr(mnt_user_ns(path
->mnt
), path
->dentry
, &ix
->ctx
);
4791 mnt_drop_write(path
->mnt
);
4797 static int io_fsetxattr(struct io_kiocb
*req
, unsigned int issue_flags
)
4801 if (issue_flags
& IO_URING_F_NONBLOCK
)
4804 ret
= __io_setxattr(req
, issue_flags
, &req
->file
->f_path
);
4805 io_xattr_finish(req
, ret
);
4810 static int io_setxattr(struct io_kiocb
*req
, unsigned int issue_flags
)
4812 struct io_xattr
*ix
= &req
->xattr
;
4813 unsigned int lookup_flags
= LOOKUP_FOLLOW
;
4817 if (issue_flags
& IO_URING_F_NONBLOCK
)
4821 ret
= filename_lookup(AT_FDCWD
, ix
->filename
, lookup_flags
, &path
, NULL
);
4823 ret
= __io_setxattr(req
, issue_flags
, &path
);
4825 if (retry_estale(ret
, lookup_flags
)) {
4826 lookup_flags
|= LOOKUP_REVAL
;
4831 io_xattr_finish(req
, ret
);
4835 static int io_unlinkat_prep(struct io_kiocb
*req
,
4836 const struct io_uring_sqe
*sqe
)
4838 struct io_unlink
*un
= &req
->unlink
;
4839 const char __user
*fname
;
4841 if (sqe
->off
|| sqe
->len
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
4843 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
4846 un
->dfd
= READ_ONCE(sqe
->fd
);
4848 un
->flags
= READ_ONCE(sqe
->unlink_flags
);
4849 if (un
->flags
& ~AT_REMOVEDIR
)
4852 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4853 un
->filename
= getname(fname
);
4854 if (IS_ERR(un
->filename
))
4855 return PTR_ERR(un
->filename
);
4857 req
->flags
|= REQ_F_NEED_CLEANUP
;
4861 static int io_unlinkat(struct io_kiocb
*req
, unsigned int issue_flags
)
4863 struct io_unlink
*un
= &req
->unlink
;
4866 if (issue_flags
& IO_URING_F_NONBLOCK
)
4869 if (un
->flags
& AT_REMOVEDIR
)
4870 ret
= do_rmdir(un
->dfd
, un
->filename
);
4872 ret
= do_unlinkat(un
->dfd
, un
->filename
);
4874 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4875 io_req_complete(req
, ret
);
4879 static int io_mkdirat_prep(struct io_kiocb
*req
,
4880 const struct io_uring_sqe
*sqe
)
4882 struct io_mkdir
*mkd
= &req
->mkdir
;
4883 const char __user
*fname
;
4885 if (sqe
->off
|| sqe
->rw_flags
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
4887 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
4890 mkd
->dfd
= READ_ONCE(sqe
->fd
);
4891 mkd
->mode
= READ_ONCE(sqe
->len
);
4893 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4894 mkd
->filename
= getname(fname
);
4895 if (IS_ERR(mkd
->filename
))
4896 return PTR_ERR(mkd
->filename
);
4898 req
->flags
|= REQ_F_NEED_CLEANUP
;
4902 static int io_mkdirat(struct io_kiocb
*req
, unsigned int issue_flags
)
4904 struct io_mkdir
*mkd
= &req
->mkdir
;
4907 if (issue_flags
& IO_URING_F_NONBLOCK
)
4910 ret
= do_mkdirat(mkd
->dfd
, mkd
->filename
, mkd
->mode
);
4912 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4913 io_req_complete(req
, ret
);
4917 static int io_symlinkat_prep(struct io_kiocb
*req
,
4918 const struct io_uring_sqe
*sqe
)
4920 struct io_symlink
*sl
= &req
->symlink
;
4921 const char __user
*oldpath
, *newpath
;
4923 if (sqe
->len
|| sqe
->rw_flags
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
4925 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
4928 sl
->new_dfd
= READ_ONCE(sqe
->fd
);
4929 oldpath
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4930 newpath
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4932 sl
->oldpath
= getname(oldpath
);
4933 if (IS_ERR(sl
->oldpath
))
4934 return PTR_ERR(sl
->oldpath
);
4936 sl
->newpath
= getname(newpath
);
4937 if (IS_ERR(sl
->newpath
)) {
4938 putname(sl
->oldpath
);
4939 return PTR_ERR(sl
->newpath
);
4942 req
->flags
|= REQ_F_NEED_CLEANUP
;
4946 static int io_symlinkat(struct io_kiocb
*req
, unsigned int issue_flags
)
4948 struct io_symlink
*sl
= &req
->symlink
;
4951 if (issue_flags
& IO_URING_F_NONBLOCK
)
4954 ret
= do_symlinkat(sl
->oldpath
, sl
->new_dfd
, sl
->newpath
);
4956 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4957 io_req_complete(req
, ret
);
4961 static int io_linkat_prep(struct io_kiocb
*req
,
4962 const struct io_uring_sqe
*sqe
)
4964 struct io_hardlink
*lnk
= &req
->hardlink
;
4965 const char __user
*oldf
, *newf
;
4967 if (sqe
->rw_flags
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
4969 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
4972 lnk
->old_dfd
= READ_ONCE(sqe
->fd
);
4973 lnk
->new_dfd
= READ_ONCE(sqe
->len
);
4974 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4975 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4976 lnk
->flags
= READ_ONCE(sqe
->hardlink_flags
);
4978 lnk
->oldpath
= getname(oldf
);
4979 if (IS_ERR(lnk
->oldpath
))
4980 return PTR_ERR(lnk
->oldpath
);
4982 lnk
->newpath
= getname(newf
);
4983 if (IS_ERR(lnk
->newpath
)) {
4984 putname(lnk
->oldpath
);
4985 return PTR_ERR(lnk
->newpath
);
4988 req
->flags
|= REQ_F_NEED_CLEANUP
;
4992 static int io_linkat(struct io_kiocb
*req
, unsigned int issue_flags
)
4994 struct io_hardlink
*lnk
= &req
->hardlink
;
4997 if (issue_flags
& IO_URING_F_NONBLOCK
)
5000 ret
= do_linkat(lnk
->old_dfd
, lnk
->oldpath
, lnk
->new_dfd
,
5001 lnk
->newpath
, lnk
->flags
);
5003 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
5004 io_req_complete(req
, ret
);
5008 static void io_uring_cmd_work(struct io_kiocb
*req
, bool *locked
)
5010 req
->uring_cmd
.task_work_cb(&req
->uring_cmd
);
5013 void io_uring_cmd_complete_in_task(struct io_uring_cmd
*ioucmd
,
5014 void (*task_work_cb
)(struct io_uring_cmd
*))
5016 struct io_kiocb
*req
= container_of(ioucmd
, struct io_kiocb
, uring_cmd
);
5018 req
->uring_cmd
.task_work_cb
= task_work_cb
;
5019 req
->io_task_work
.func
= io_uring_cmd_work
;
5020 io_req_task_work_add(req
);
5022 EXPORT_SYMBOL_GPL(io_uring_cmd_complete_in_task
);
5024 static inline void io_req_set_cqe32_extra(struct io_kiocb
*req
,
5025 u64 extra1
, u64 extra2
)
5027 req
->extra1
= extra1
;
5028 req
->extra2
= extra2
;
5029 req
->flags
|= REQ_F_CQE32_INIT
;
5033 * Called by consumers of io_uring_cmd, if they originally returned
5034 * -EIOCBQUEUED upon receiving the command.
5036 void io_uring_cmd_done(struct io_uring_cmd
*ioucmd
, ssize_t ret
, ssize_t res2
)
5038 struct io_kiocb
*req
= container_of(ioucmd
, struct io_kiocb
, uring_cmd
);
5043 if (req
->ctx
->flags
& IORING_SETUP_CQE32
)
5044 io_req_set_cqe32_extra(req
, res2
, 0);
5045 io_req_complete(req
, ret
);
5047 EXPORT_SYMBOL_GPL(io_uring_cmd_done
);
5049 static int io_uring_cmd_prep_async(struct io_kiocb
*req
)
5053 cmd_size
= uring_cmd_pdu_size(req
->ctx
->flags
& IORING_SETUP_SQE128
);
5055 memcpy(req
->async_data
, req
->uring_cmd
.cmd
, cmd_size
);
5059 static int io_uring_cmd_prep(struct io_kiocb
*req
,
5060 const struct io_uring_sqe
*sqe
)
5062 struct io_uring_cmd
*ioucmd
= &req
->uring_cmd
;
5066 ioucmd
->cmd
= sqe
->cmd
;
5067 ioucmd
->cmd_op
= READ_ONCE(sqe
->cmd_op
);
5071 static int io_uring_cmd(struct io_kiocb
*req
, unsigned int issue_flags
)
5073 struct io_uring_cmd
*ioucmd
= &req
->uring_cmd
;
5074 struct io_ring_ctx
*ctx
= req
->ctx
;
5075 struct file
*file
= req
->file
;
5078 if (!req
->file
->f_op
->uring_cmd
)
5081 if (ctx
->flags
& IORING_SETUP_SQE128
)
5082 issue_flags
|= IO_URING_F_SQE128
;
5083 if (ctx
->flags
& IORING_SETUP_CQE32
)
5084 issue_flags
|= IO_URING_F_CQE32
;
5085 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
5086 issue_flags
|= IO_URING_F_IOPOLL
;
5088 if (req_has_async_data(req
))
5089 ioucmd
->cmd
= req
->async_data
;
5091 ret
= file
->f_op
->uring_cmd(ioucmd
, issue_flags
);
5092 if (ret
== -EAGAIN
) {
5093 if (!req_has_async_data(req
)) {
5094 if (io_alloc_async_data(req
))
5096 io_uring_cmd_prep_async(req
);
5101 if (ret
!= -EIOCBQUEUED
)
5102 io_uring_cmd_done(ioucmd
, ret
, 0);
5106 static int __io_splice_prep(struct io_kiocb
*req
,
5107 const struct io_uring_sqe
*sqe
)
5109 struct io_splice
*sp
= &req
->splice
;
5110 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
5112 sp
->len
= READ_ONCE(sqe
->len
);
5113 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
5114 if (unlikely(sp
->flags
& ~valid_flags
))
5116 sp
->splice_fd_in
= READ_ONCE(sqe
->splice_fd_in
);
5120 static int io_tee_prep(struct io_kiocb
*req
,
5121 const struct io_uring_sqe
*sqe
)
5123 if (READ_ONCE(sqe
->splice_off_in
) || READ_ONCE(sqe
->off
))
5125 return __io_splice_prep(req
, sqe
);
5128 static int io_tee(struct io_kiocb
*req
, unsigned int issue_flags
)
5130 struct io_splice
*sp
= &req
->splice
;
5131 struct file
*out
= sp
->file_out
;
5132 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
5136 if (issue_flags
& IO_URING_F_NONBLOCK
)
5139 if (sp
->flags
& SPLICE_F_FD_IN_FIXED
)
5140 in
= io_file_get_fixed(req
, sp
->splice_fd_in
, issue_flags
);
5142 in
= io_file_get_normal(req
, sp
->splice_fd_in
);
5149 ret
= do_tee(in
, out
, sp
->len
, flags
);
5151 if (!(sp
->flags
& SPLICE_F_FD_IN_FIXED
))
5156 __io_req_complete(req
, 0, ret
, 0);
5160 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5162 struct io_splice
*sp
= &req
->splice
;
5164 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
5165 sp
->off_out
= READ_ONCE(sqe
->off
);
5166 return __io_splice_prep(req
, sqe
);
5169 static int io_splice(struct io_kiocb
*req
, unsigned int issue_flags
)
5171 struct io_splice
*sp
= &req
->splice
;
5172 struct file
*out
= sp
->file_out
;
5173 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
5174 loff_t
*poff_in
, *poff_out
;
5178 if (issue_flags
& IO_URING_F_NONBLOCK
)
5181 if (sp
->flags
& SPLICE_F_FD_IN_FIXED
)
5182 in
= io_file_get_fixed(req
, sp
->splice_fd_in
, issue_flags
);
5184 in
= io_file_get_normal(req
, sp
->splice_fd_in
);
5190 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
5191 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
5194 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
5196 if (!(sp
->flags
& SPLICE_F_FD_IN_FIXED
))
5201 __io_req_complete(req
, 0, ret
, 0);
5205 static int io_nop_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5211 * IORING_OP_NOP just posts a completion event, nothing else.
5213 static int io_nop(struct io_kiocb
*req
, unsigned int issue_flags
)
5215 __io_req_complete(req
, issue_flags
, 0, 0);
5219 static int io_msg_ring_prep(struct io_kiocb
*req
,
5220 const struct io_uring_sqe
*sqe
)
5222 if (unlikely(sqe
->addr
|| sqe
->rw_flags
|| sqe
->splice_fd_in
||
5223 sqe
->buf_index
|| sqe
->personality
))
5226 req
->msg
.user_data
= READ_ONCE(sqe
->off
);
5227 req
->msg
.len
= READ_ONCE(sqe
->len
);
5231 static int io_msg_ring(struct io_kiocb
*req
, unsigned int issue_flags
)
5233 struct io_ring_ctx
*target_ctx
;
5234 struct io_msg
*msg
= &req
->msg
;
5239 if (req
->file
->f_op
!= &io_uring_fops
)
5243 target_ctx
= req
->file
->private_data
;
5245 spin_lock(&target_ctx
->completion_lock
);
5246 filled
= io_fill_cqe_aux(target_ctx
, msg
->user_data
, msg
->len
, 0);
5247 io_commit_cqring(target_ctx
);
5248 spin_unlock(&target_ctx
->completion_lock
);
5251 io_cqring_ev_posted(target_ctx
);
5258 __io_req_complete(req
, issue_flags
, ret
, 0);
5259 /* put file to avoid an attempt to IOPOLL the req */
5260 io_put_file(req
->file
);
5265 static int io_fsync_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5267 if (unlikely(sqe
->addr
|| sqe
->buf_index
|| sqe
->splice_fd_in
))
5270 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
5271 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
5274 req
->sync
.off
= READ_ONCE(sqe
->off
);
5275 req
->sync
.len
= READ_ONCE(sqe
->len
);
5279 static int io_fsync(struct io_kiocb
*req
, unsigned int issue_flags
)
5281 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
5284 /* fsync always requires a blocking context */
5285 if (issue_flags
& IO_URING_F_NONBLOCK
)
5288 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
5289 end
> 0 ? end
: LLONG_MAX
,
5290 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
5291 io_req_complete(req
, ret
);
5295 static int io_fallocate_prep(struct io_kiocb
*req
,
5296 const struct io_uring_sqe
*sqe
)
5298 if (sqe
->buf_index
|| sqe
->rw_flags
|| sqe
->splice_fd_in
)
5301 req
->sync
.off
= READ_ONCE(sqe
->off
);
5302 req
->sync
.len
= READ_ONCE(sqe
->addr
);
5303 req
->sync
.mode
= READ_ONCE(sqe
->len
);
5307 static int io_fallocate(struct io_kiocb
*req
, unsigned int issue_flags
)
5311 /* fallocate always requiring blocking context */
5312 if (issue_flags
& IO_URING_F_NONBLOCK
)
5314 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
5317 fsnotify_modify(req
->file
);
5318 io_req_complete(req
, ret
);
5322 static int __io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5324 const char __user
*fname
;
5327 if (unlikely(sqe
->buf_index
))
5329 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
5332 /* open.how should be already initialised */
5333 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
5334 req
->open
.how
.flags
|= O_LARGEFILE
;
5336 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
5337 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5338 req
->open
.filename
= getname(fname
);
5339 if (IS_ERR(req
->open
.filename
)) {
5340 ret
= PTR_ERR(req
->open
.filename
);
5341 req
->open
.filename
= NULL
;
5345 req
->open
.file_slot
= READ_ONCE(sqe
->file_index
);
5346 if (req
->open
.file_slot
&& (req
->open
.how
.flags
& O_CLOEXEC
))
5349 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
5350 req
->flags
|= REQ_F_NEED_CLEANUP
;
5354 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5356 u64 mode
= READ_ONCE(sqe
->len
);
5357 u64 flags
= READ_ONCE(sqe
->open_flags
);
5359 req
->open
.how
= build_open_how(flags
, mode
);
5360 return __io_openat_prep(req
, sqe
);
5363 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5365 struct open_how __user
*how
;
5369 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
5370 len
= READ_ONCE(sqe
->len
);
5371 if (len
< OPEN_HOW_SIZE_VER0
)
5374 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
5379 return __io_openat_prep(req
, sqe
);
5382 static int io_file_bitmap_get(struct io_ring_ctx
*ctx
)
5384 struct io_file_table
*table
= &ctx
->file_table
;
5385 unsigned long nr
= ctx
->nr_user_files
;
5389 ret
= find_next_zero_bit(table
->bitmap
, nr
, table
->alloc_hint
);
5393 if (!table
->alloc_hint
)
5396 nr
= table
->alloc_hint
;
5397 table
->alloc_hint
= 0;
5404 * Note when io_fixed_fd_install() returns error value, it will ensure
5405 * fput() is called correspondingly.
5407 static int io_fixed_fd_install(struct io_kiocb
*req
, unsigned int issue_flags
,
5408 struct file
*file
, unsigned int file_slot
)
5410 bool alloc_slot
= file_slot
== IORING_FILE_INDEX_ALLOC
;
5411 struct io_ring_ctx
*ctx
= req
->ctx
;
5414 io_ring_submit_lock(ctx
, issue_flags
);
5417 ret
= io_file_bitmap_get(ctx
);
5418 if (unlikely(ret
< 0))
5425 ret
= io_install_fixed_file(req
, file
, issue_flags
, file_slot
);
5426 if (!ret
&& alloc_slot
)
5429 io_ring_submit_unlock(ctx
, issue_flags
);
5430 if (unlikely(ret
< 0))
5435 static int io_openat2(struct io_kiocb
*req
, unsigned int issue_flags
)
5437 struct open_flags op
;
5439 bool resolve_nonblock
, nonblock_set
;
5440 bool fixed
= !!req
->open
.file_slot
;
5443 ret
= build_open_flags(&req
->open
.how
, &op
);
5446 nonblock_set
= op
.open_flag
& O_NONBLOCK
;
5447 resolve_nonblock
= req
->open
.how
.resolve
& RESOLVE_CACHED
;
5448 if (issue_flags
& IO_URING_F_NONBLOCK
) {
5450 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
5451 * it'll always -EAGAIN
5453 if (req
->open
.how
.flags
& (O_TRUNC
| O_CREAT
| O_TMPFILE
))
5455 op
.lookup_flags
|= LOOKUP_CACHED
;
5456 op
.open_flag
|= O_NONBLOCK
;
5460 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
5465 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
5468 * We could hang on to this 'fd' on retrying, but seems like
5469 * marginal gain for something that is now known to be a slower
5470 * path. So just put it, and we'll get a new one when we retry.
5475 ret
= PTR_ERR(file
);
5476 /* only retry if RESOLVE_CACHED wasn't already set by application */
5477 if (ret
== -EAGAIN
&&
5478 (!resolve_nonblock
&& (issue_flags
& IO_URING_F_NONBLOCK
)))
5483 if ((issue_flags
& IO_URING_F_NONBLOCK
) && !nonblock_set
)
5484 file
->f_flags
&= ~O_NONBLOCK
;
5485 fsnotify_open(file
);
5488 fd_install(ret
, file
);
5490 ret
= io_fixed_fd_install(req
, issue_flags
, file
,
5491 req
->open
.file_slot
);
5493 putname(req
->open
.filename
);
5494 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
5497 __io_req_complete(req
, issue_flags
, ret
, 0);
5501 static int io_openat(struct io_kiocb
*req
, unsigned int issue_flags
)
5503 return io_openat2(req
, issue_flags
);
5506 static int io_remove_buffers_prep(struct io_kiocb
*req
,
5507 const struct io_uring_sqe
*sqe
)
5509 struct io_provide_buf
*p
= &req
->pbuf
;
5512 if (sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
||
5516 tmp
= READ_ONCE(sqe
->fd
);
5517 if (!tmp
|| tmp
> USHRT_MAX
)
5520 memset(p
, 0, sizeof(*p
));
5522 p
->bgid
= READ_ONCE(sqe
->buf_group
);
5526 static int __io_remove_buffers(struct io_ring_ctx
*ctx
,
5527 struct io_buffer_list
*bl
, unsigned nbufs
)
5531 /* shouldn't happen */
5535 if (bl
->buf_nr_pages
) {
5538 i
= bl
->buf_ring
->tail
- bl
->head
;
5539 for (j
= 0; j
< bl
->buf_nr_pages
; j
++)
5540 unpin_user_page(bl
->buf_pages
[j
]);
5541 kvfree(bl
->buf_pages
);
5542 bl
->buf_pages
= NULL
;
5543 bl
->buf_nr_pages
= 0;
5544 /* make sure it's seen as empty */
5545 INIT_LIST_HEAD(&bl
->buf_list
);
5549 /* the head kbuf is the list itself */
5550 while (!list_empty(&bl
->buf_list
)) {
5551 struct io_buffer
*nxt
;
5553 nxt
= list_first_entry(&bl
->buf_list
, struct io_buffer
, list
);
5554 list_del(&nxt
->list
);
5564 static int io_remove_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
5566 struct io_provide_buf
*p
= &req
->pbuf
;
5567 struct io_ring_ctx
*ctx
= req
->ctx
;
5568 struct io_buffer_list
*bl
;
5571 io_ring_submit_lock(ctx
, issue_flags
);
5574 bl
= io_buffer_get_list(ctx
, p
->bgid
);
5577 /* can't use provide/remove buffers command on mapped buffers */
5578 if (!bl
->buf_nr_pages
)
5579 ret
= __io_remove_buffers(ctx
, bl
, p
->nbufs
);
5584 /* complete before unlock, IOPOLL may need the lock */
5585 __io_req_complete(req
, issue_flags
, ret
, 0);
5586 io_ring_submit_unlock(ctx
, issue_flags
);
5590 static int io_provide_buffers_prep(struct io_kiocb
*req
,
5591 const struct io_uring_sqe
*sqe
)
5593 unsigned long size
, tmp_check
;
5594 struct io_provide_buf
*p
= &req
->pbuf
;
5597 if (sqe
->rw_flags
|| sqe
->splice_fd_in
)
5600 tmp
= READ_ONCE(sqe
->fd
);
5601 if (!tmp
|| tmp
> USHRT_MAX
)
5604 p
->addr
= READ_ONCE(sqe
->addr
);
5605 p
->len
= READ_ONCE(sqe
->len
);
5607 if (check_mul_overflow((unsigned long)p
->len
, (unsigned long)p
->nbufs
,
5610 if (check_add_overflow((unsigned long)p
->addr
, size
, &tmp_check
))
5613 size
= (unsigned long)p
->len
* p
->nbufs
;
5614 if (!access_ok(u64_to_user_ptr(p
->addr
), size
))
5617 p
->bgid
= READ_ONCE(sqe
->buf_group
);
5618 tmp
= READ_ONCE(sqe
->off
);
5619 if (tmp
> USHRT_MAX
)
5625 static int io_refill_buffer_cache(struct io_ring_ctx
*ctx
)
5627 struct io_buffer
*buf
;
5632 * Completions that don't happen inline (eg not under uring_lock) will
5633 * add to ->io_buffers_comp. If we don't have any free buffers, check
5634 * the completion list and splice those entries first.
5636 if (!list_empty_careful(&ctx
->io_buffers_comp
)) {
5637 spin_lock(&ctx
->completion_lock
);
5638 if (!list_empty(&ctx
->io_buffers_comp
)) {
5639 list_splice_init(&ctx
->io_buffers_comp
,
5640 &ctx
->io_buffers_cache
);
5641 spin_unlock(&ctx
->completion_lock
);
5644 spin_unlock(&ctx
->completion_lock
);
5648 * No free buffers and no completion entries either. Allocate a new
5649 * page worth of buffer entries and add those to our freelist.
5651 page
= alloc_page(GFP_KERNEL_ACCOUNT
);
5655 list_add(&page
->lru
, &ctx
->io_buffers_pages
);
5657 buf
= page_address(page
);
5658 bufs_in_page
= PAGE_SIZE
/ sizeof(*buf
);
5659 while (bufs_in_page
) {
5660 list_add_tail(&buf
->list
, &ctx
->io_buffers_cache
);
5668 static int io_add_buffers(struct io_ring_ctx
*ctx
, struct io_provide_buf
*pbuf
,
5669 struct io_buffer_list
*bl
)
5671 struct io_buffer
*buf
;
5672 u64 addr
= pbuf
->addr
;
5673 int i
, bid
= pbuf
->bid
;
5675 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
5676 if (list_empty(&ctx
->io_buffers_cache
) &&
5677 io_refill_buffer_cache(ctx
))
5679 buf
= list_first_entry(&ctx
->io_buffers_cache
, struct io_buffer
,
5681 list_move_tail(&buf
->list
, &bl
->buf_list
);
5683 buf
->len
= min_t(__u32
, pbuf
->len
, MAX_RW_COUNT
);
5685 buf
->bgid
= pbuf
->bgid
;
5691 return i
? 0 : -ENOMEM
;
5694 static __cold
int io_init_bl_list(struct io_ring_ctx
*ctx
)
5698 ctx
->io_bl
= kcalloc(BGID_ARRAY
, sizeof(struct io_buffer_list
),
5703 for (i
= 0; i
< BGID_ARRAY
; i
++) {
5704 INIT_LIST_HEAD(&ctx
->io_bl
[i
].buf_list
);
5705 ctx
->io_bl
[i
].bgid
= i
;
5711 static int io_provide_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
5713 struct io_provide_buf
*p
= &req
->pbuf
;
5714 struct io_ring_ctx
*ctx
= req
->ctx
;
5715 struct io_buffer_list
*bl
;
5718 io_ring_submit_lock(ctx
, issue_flags
);
5720 if (unlikely(p
->bgid
< BGID_ARRAY
&& !ctx
->io_bl
)) {
5721 ret
= io_init_bl_list(ctx
);
5726 bl
= io_buffer_get_list(ctx
, p
->bgid
);
5727 if (unlikely(!bl
)) {
5728 bl
= kzalloc(sizeof(*bl
), GFP_KERNEL
);
5733 INIT_LIST_HEAD(&bl
->buf_list
);
5734 ret
= io_buffer_add_list(ctx
, bl
, p
->bgid
);
5740 /* can't add buffers via this command for a mapped buffer ring */
5741 if (bl
->buf_nr_pages
) {
5746 ret
= io_add_buffers(ctx
, p
, bl
);
5750 /* complete before unlock, IOPOLL may need the lock */
5751 __io_req_complete(req
, issue_flags
, ret
, 0);
5752 io_ring_submit_unlock(ctx
, issue_flags
);
5756 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
5757 const struct io_uring_sqe
*sqe
)
5759 #if defined(CONFIG_EPOLL)
5760 if (sqe
->buf_index
|| sqe
->splice_fd_in
)
5763 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
5764 req
->epoll
.op
= READ_ONCE(sqe
->len
);
5765 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
5767 if (ep_op_has_event(req
->epoll
.op
)) {
5768 struct epoll_event __user
*ev
;
5770 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5771 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
5781 static int io_epoll_ctl(struct io_kiocb
*req
, unsigned int issue_flags
)
5783 #if defined(CONFIG_EPOLL)
5784 struct io_epoll
*ie
= &req
->epoll
;
5786 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5788 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
5789 if (force_nonblock
&& ret
== -EAGAIN
)
5794 __io_req_complete(req
, issue_flags
, ret
, 0);
5801 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5803 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
5804 if (sqe
->buf_index
|| sqe
->off
|| sqe
->splice_fd_in
)
5807 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
5808 req
->madvise
.len
= READ_ONCE(sqe
->len
);
5809 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
5816 static int io_madvise(struct io_kiocb
*req
, unsigned int issue_flags
)
5818 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
5819 struct io_madvise
*ma
= &req
->madvise
;
5822 if (issue_flags
& IO_URING_F_NONBLOCK
)
5825 ret
= do_madvise(current
->mm
, ma
->addr
, ma
->len
, ma
->advice
);
5826 io_req_complete(req
, ret
);
5833 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5835 if (sqe
->buf_index
|| sqe
->addr
|| sqe
->splice_fd_in
)
5838 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
5839 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
5840 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
5844 static int io_fadvise(struct io_kiocb
*req
, unsigned int issue_flags
)
5846 struct io_fadvise
*fa
= &req
->fadvise
;
5849 if (issue_flags
& IO_URING_F_NONBLOCK
) {
5850 switch (fa
->advice
) {
5851 case POSIX_FADV_NORMAL
:
5852 case POSIX_FADV_RANDOM
:
5853 case POSIX_FADV_SEQUENTIAL
:
5860 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
5863 __io_req_complete(req
, issue_flags
, ret
, 0);
5867 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5869 const char __user
*path
;
5871 if (sqe
->buf_index
|| sqe
->splice_fd_in
)
5873 if (req
->flags
& REQ_F_FIXED_FILE
)
5876 req
->statx
.dfd
= READ_ONCE(sqe
->fd
);
5877 req
->statx
.mask
= READ_ONCE(sqe
->len
);
5878 path
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5879 req
->statx
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
5880 req
->statx
.flags
= READ_ONCE(sqe
->statx_flags
);
5882 req
->statx
.filename
= getname_flags(path
,
5883 getname_statx_lookup_flags(req
->statx
.flags
),
5886 if (IS_ERR(req
->statx
.filename
)) {
5887 int ret
= PTR_ERR(req
->statx
.filename
);
5889 req
->statx
.filename
= NULL
;
5893 req
->flags
|= REQ_F_NEED_CLEANUP
;
5897 static int io_statx(struct io_kiocb
*req
, unsigned int issue_flags
)
5899 struct io_statx
*ctx
= &req
->statx
;
5902 if (issue_flags
& IO_URING_F_NONBLOCK
)
5905 ret
= do_statx(ctx
->dfd
, ctx
->filename
, ctx
->flags
, ctx
->mask
,
5907 io_req_complete(req
, ret
);
5911 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5913 if (sqe
->off
|| sqe
->addr
|| sqe
->len
|| sqe
->rw_flags
|| sqe
->buf_index
)
5915 if (req
->flags
& REQ_F_FIXED_FILE
)
5918 req
->close
.fd
= READ_ONCE(sqe
->fd
);
5919 req
->close
.file_slot
= READ_ONCE(sqe
->file_index
);
5920 if (req
->close
.file_slot
&& req
->close
.fd
)
5926 static int io_close(struct io_kiocb
*req
, unsigned int issue_flags
)
5928 struct files_struct
*files
= current
->files
;
5929 struct io_close
*close
= &req
->close
;
5930 struct fdtable
*fdt
;
5934 if (req
->close
.file_slot
) {
5935 ret
= io_close_fixed(req
, issue_flags
);
5939 spin_lock(&files
->file_lock
);
5940 fdt
= files_fdtable(files
);
5941 if (close
->fd
>= fdt
->max_fds
) {
5942 spin_unlock(&files
->file_lock
);
5945 file
= rcu_dereference_protected(fdt
->fd
[close
->fd
],
5946 lockdep_is_held(&files
->file_lock
));
5947 if (!file
|| file
->f_op
== &io_uring_fops
) {
5948 spin_unlock(&files
->file_lock
);
5952 /* if the file has a flush method, be safe and punt to async */
5953 if (file
->f_op
->flush
&& (issue_flags
& IO_URING_F_NONBLOCK
)) {
5954 spin_unlock(&files
->file_lock
);
5958 file
= __close_fd_get_file(close
->fd
);
5959 spin_unlock(&files
->file_lock
);
5963 /* No ->flush() or already async, safely close from here */
5964 ret
= filp_close(file
, current
->files
);
5968 __io_req_complete(req
, issue_flags
, ret
, 0);
5972 static int io_sfr_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5974 if (unlikely(sqe
->addr
|| sqe
->buf_index
|| sqe
->splice_fd_in
))
5977 req
->sync
.off
= READ_ONCE(sqe
->off
);
5978 req
->sync
.len
= READ_ONCE(sqe
->len
);
5979 req
->sync
.flags
= READ_ONCE(sqe
->sync_range_flags
);
5983 static int io_sync_file_range(struct io_kiocb
*req
, unsigned int issue_flags
)
5987 /* sync_file_range always requires a blocking context */
5988 if (issue_flags
& IO_URING_F_NONBLOCK
)
5991 ret
= sync_file_range(req
->file
, req
->sync
.off
, req
->sync
.len
,
5993 io_req_complete(req
, ret
);
5997 #if defined(CONFIG_NET)
5998 static int io_shutdown_prep(struct io_kiocb
*req
,
5999 const struct io_uring_sqe
*sqe
)
6001 if (unlikely(sqe
->off
|| sqe
->addr
|| sqe
->rw_flags
||
6002 sqe
->buf_index
|| sqe
->splice_fd_in
))
6005 req
->shutdown
.how
= READ_ONCE(sqe
->len
);
6009 static int io_shutdown(struct io_kiocb
*req
, unsigned int issue_flags
)
6011 struct socket
*sock
;
6014 if (issue_flags
& IO_URING_F_NONBLOCK
)
6017 sock
= sock_from_file(req
->file
);
6018 if (unlikely(!sock
))
6021 ret
= __sys_shutdown_sock(sock
, req
->shutdown
.how
);
6022 io_req_complete(req
, ret
);
6026 static bool io_net_retry(struct socket
*sock
, int flags
)
6028 if (!(flags
& MSG_WAITALL
))
6030 return sock
->type
== SOCK_STREAM
|| sock
->type
== SOCK_SEQPACKET
;
6033 static int io_setup_async_msg(struct io_kiocb
*req
,
6034 struct io_async_msghdr
*kmsg
)
6036 struct io_async_msghdr
*async_msg
= req
->async_data
;
6040 if (io_alloc_async_data(req
)) {
6041 kfree(kmsg
->free_iov
);
6044 async_msg
= req
->async_data
;
6045 req
->flags
|= REQ_F_NEED_CLEANUP
;
6046 memcpy(async_msg
, kmsg
, sizeof(*kmsg
));
6047 async_msg
->msg
.msg_name
= &async_msg
->addr
;
6048 /* if were using fast_iov, set it to the new one */
6049 if (!async_msg
->free_iov
)
6050 async_msg
->msg
.msg_iter
.iov
= async_msg
->fast_iov
;
6055 static int io_sendmsg_copy_hdr(struct io_kiocb
*req
,
6056 struct io_async_msghdr
*iomsg
)
6058 iomsg
->msg
.msg_name
= &iomsg
->addr
;
6059 iomsg
->free_iov
= iomsg
->fast_iov
;
6060 return sendmsg_copy_msghdr(&iomsg
->msg
, req
->sr_msg
.umsg
,
6061 req
->sr_msg
.msg_flags
, &iomsg
->free_iov
);
6064 static int io_sendmsg_prep_async(struct io_kiocb
*req
)
6068 ret
= io_sendmsg_copy_hdr(req
, req
->async_data
);
6070 req
->flags
|= REQ_F_NEED_CLEANUP
;
6074 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6076 struct io_sr_msg
*sr
= &req
->sr_msg
;
6078 if (unlikely(sqe
->file_index
))
6081 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
6082 sr
->len
= READ_ONCE(sqe
->len
);
6083 sr
->flags
= READ_ONCE(sqe
->addr2
);
6084 if (sr
->flags
& ~IORING_RECVSEND_POLL_FIRST
)
6086 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
) | MSG_NOSIGNAL
;
6087 if (sr
->msg_flags
& MSG_DONTWAIT
)
6088 req
->flags
|= REQ_F_NOWAIT
;
6090 #ifdef CONFIG_COMPAT
6091 if (req
->ctx
->compat
)
6092 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
6098 static int io_sendmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
6100 struct io_async_msghdr iomsg
, *kmsg
;
6101 struct io_sr_msg
*sr
= &req
->sr_msg
;
6102 struct socket
*sock
;
6107 sock
= sock_from_file(req
->file
);
6108 if (unlikely(!sock
))
6111 if (req_has_async_data(req
)) {
6112 kmsg
= req
->async_data
;
6114 ret
= io_sendmsg_copy_hdr(req
, &iomsg
);
6120 if (!(req
->flags
& REQ_F_POLLED
) &&
6121 (sr
->flags
& IORING_RECVSEND_POLL_FIRST
))
6122 return io_setup_async_msg(req
, kmsg
);
6124 flags
= sr
->msg_flags
;
6125 if (issue_flags
& IO_URING_F_NONBLOCK
)
6126 flags
|= MSG_DONTWAIT
;
6127 if (flags
& MSG_WAITALL
)
6128 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
6130 ret
= __sys_sendmsg_sock(sock
, &kmsg
->msg
, flags
);
6132 if (ret
< min_ret
) {
6133 if (ret
== -EAGAIN
&& (issue_flags
& IO_URING_F_NONBLOCK
))
6134 return io_setup_async_msg(req
, kmsg
);
6135 if (ret
== -ERESTARTSYS
)
6137 if (ret
> 0 && io_net_retry(sock
, flags
)) {
6139 req
->flags
|= REQ_F_PARTIAL_IO
;
6140 return io_setup_async_msg(req
, kmsg
);
6144 /* fast path, check for non-NULL to avoid function call */
6146 kfree(kmsg
->free_iov
);
6147 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
6150 else if (sr
->done_io
)
6152 __io_req_complete(req
, issue_flags
, ret
, 0);
6156 static int io_send(struct io_kiocb
*req
, unsigned int issue_flags
)
6158 struct io_sr_msg
*sr
= &req
->sr_msg
;
6161 struct socket
*sock
;
6166 if (!(req
->flags
& REQ_F_POLLED
) &&
6167 (sr
->flags
& IORING_RECVSEND_POLL_FIRST
))
6170 sock
= sock_from_file(req
->file
);
6171 if (unlikely(!sock
))
6174 ret
= import_single_range(WRITE
, sr
->buf
, sr
->len
, &iov
, &msg
.msg_iter
);
6178 msg
.msg_name
= NULL
;
6179 msg
.msg_control
= NULL
;
6180 msg
.msg_controllen
= 0;
6181 msg
.msg_namelen
= 0;
6183 flags
= sr
->msg_flags
;
6184 if (issue_flags
& IO_URING_F_NONBLOCK
)
6185 flags
|= MSG_DONTWAIT
;
6186 if (flags
& MSG_WAITALL
)
6187 min_ret
= iov_iter_count(&msg
.msg_iter
);
6189 msg
.msg_flags
= flags
;
6190 ret
= sock_sendmsg(sock
, &msg
);
6191 if (ret
< min_ret
) {
6192 if (ret
== -EAGAIN
&& (issue_flags
& IO_URING_F_NONBLOCK
))
6194 if (ret
== -ERESTARTSYS
)
6196 if (ret
> 0 && io_net_retry(sock
, flags
)) {
6200 req
->flags
|= REQ_F_PARTIAL_IO
;
6207 else if (sr
->done_io
)
6209 __io_req_complete(req
, issue_flags
, ret
, 0);
6213 static int __io_recvmsg_copy_hdr(struct io_kiocb
*req
,
6214 struct io_async_msghdr
*iomsg
)
6216 struct io_sr_msg
*sr
= &req
->sr_msg
;
6217 struct iovec __user
*uiov
;
6221 ret
= __copy_msghdr_from_user(&iomsg
->msg
, sr
->umsg
,
6222 &iomsg
->uaddr
, &uiov
, &iov_len
);
6226 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
6229 if (copy_from_user(iomsg
->fast_iov
, uiov
, sizeof(*uiov
)))
6231 sr
->len
= iomsg
->fast_iov
[0].iov_len
;
6232 iomsg
->free_iov
= NULL
;
6234 iomsg
->free_iov
= iomsg
->fast_iov
;
6235 ret
= __import_iovec(READ
, uiov
, iov_len
, UIO_FASTIOV
,
6236 &iomsg
->free_iov
, &iomsg
->msg
.msg_iter
,
6245 #ifdef CONFIG_COMPAT
6246 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb
*req
,
6247 struct io_async_msghdr
*iomsg
)
6249 struct io_sr_msg
*sr
= &req
->sr_msg
;
6250 struct compat_iovec __user
*uiov
;
6255 ret
= __get_compat_msghdr(&iomsg
->msg
, sr
->umsg_compat
, &iomsg
->uaddr
,
6260 uiov
= compat_ptr(ptr
);
6261 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
6262 compat_ssize_t clen
;
6266 if (!access_ok(uiov
, sizeof(*uiov
)))
6268 if (__get_user(clen
, &uiov
->iov_len
))
6273 iomsg
->free_iov
= NULL
;
6275 iomsg
->free_iov
= iomsg
->fast_iov
;
6276 ret
= __import_iovec(READ
, (struct iovec __user
*)uiov
, len
,
6277 UIO_FASTIOV
, &iomsg
->free_iov
,
6278 &iomsg
->msg
.msg_iter
, true);
6287 static int io_recvmsg_copy_hdr(struct io_kiocb
*req
,
6288 struct io_async_msghdr
*iomsg
)
6290 iomsg
->msg
.msg_name
= &iomsg
->addr
;
6292 #ifdef CONFIG_COMPAT
6293 if (req
->ctx
->compat
)
6294 return __io_compat_recvmsg_copy_hdr(req
, iomsg
);
6297 return __io_recvmsg_copy_hdr(req
, iomsg
);
6300 static int io_recvmsg_prep_async(struct io_kiocb
*req
)
6304 ret
= io_recvmsg_copy_hdr(req
, req
->async_data
);
6306 req
->flags
|= REQ_F_NEED_CLEANUP
;
6310 static int io_recvmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6312 struct io_sr_msg
*sr
= &req
->sr_msg
;
6314 if (unlikely(sqe
->file_index
))
6317 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
6318 sr
->len
= READ_ONCE(sqe
->len
);
6319 sr
->flags
= READ_ONCE(sqe
->addr2
);
6320 if (sr
->flags
& ~IORING_RECVSEND_POLL_FIRST
)
6322 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
) | MSG_NOSIGNAL
;
6323 if (sr
->msg_flags
& MSG_DONTWAIT
)
6324 req
->flags
|= REQ_F_NOWAIT
;
6326 #ifdef CONFIG_COMPAT
6327 if (req
->ctx
->compat
)
6328 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
6334 static int io_recvmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
6336 struct io_async_msghdr iomsg
, *kmsg
;
6337 struct io_sr_msg
*sr
= &req
->sr_msg
;
6338 struct socket
*sock
;
6339 unsigned int cflags
;
6341 int ret
, min_ret
= 0;
6342 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
6344 sock
= sock_from_file(req
->file
);
6345 if (unlikely(!sock
))
6348 if (req_has_async_data(req
)) {
6349 kmsg
= req
->async_data
;
6351 ret
= io_recvmsg_copy_hdr(req
, &iomsg
);
6357 if (!(req
->flags
& REQ_F_POLLED
) &&
6358 (sr
->flags
& IORING_RECVSEND_POLL_FIRST
))
6359 return io_setup_async_msg(req
, kmsg
);
6361 if (io_do_buffer_select(req
)) {
6364 buf
= io_buffer_select(req
, &sr
->len
, issue_flags
);
6367 kmsg
->fast_iov
[0].iov_base
= buf
;
6368 kmsg
->fast_iov
[0].iov_len
= sr
->len
;
6369 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->fast_iov
, 1,
6373 flags
= sr
->msg_flags
;
6375 flags
|= MSG_DONTWAIT
;
6376 if (flags
& MSG_WAITALL
)
6377 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
6379 kmsg
->msg
.msg_get_inq
= 1;
6380 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, sr
->umsg
, kmsg
->uaddr
, flags
);
6381 if (ret
< min_ret
) {
6382 if (ret
== -EAGAIN
&& force_nonblock
)
6383 return io_setup_async_msg(req
, kmsg
);
6384 if (ret
== -ERESTARTSYS
)
6386 if (ret
> 0 && io_net_retry(sock
, flags
)) {
6388 req
->flags
|= REQ_F_PARTIAL_IO
;
6389 return io_setup_async_msg(req
, kmsg
);
6392 } else if ((flags
& MSG_WAITALL
) && (kmsg
->msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))) {
6396 /* fast path, check for non-NULL to avoid function call */
6398 kfree(kmsg
->free_iov
);
6399 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
6402 else if (sr
->done_io
)
6404 cflags
= io_put_kbuf(req
, issue_flags
);
6405 if (kmsg
->msg
.msg_inq
)
6406 cflags
|= IORING_CQE_F_SOCK_NONEMPTY
;
6407 __io_req_complete(req
, issue_flags
, ret
, cflags
);
6411 static int io_recv(struct io_kiocb
*req
, unsigned int issue_flags
)
6413 struct io_sr_msg
*sr
= &req
->sr_msg
;
6415 struct socket
*sock
;
6417 unsigned int cflags
;
6419 int ret
, min_ret
= 0;
6420 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
6422 if (!(req
->flags
& REQ_F_POLLED
) &&
6423 (sr
->flags
& IORING_RECVSEND_POLL_FIRST
))
6426 sock
= sock_from_file(req
->file
);
6427 if (unlikely(!sock
))
6430 if (io_do_buffer_select(req
)) {
6433 buf
= io_buffer_select(req
, &sr
->len
, issue_flags
);
6439 ret
= import_single_range(READ
, sr
->buf
, sr
->len
, &iov
, &msg
.msg_iter
);
6443 msg
.msg_name
= NULL
;
6444 msg
.msg_namelen
= 0;
6445 msg
.msg_control
= NULL
;
6446 msg
.msg_get_inq
= 1;
6448 msg
.msg_controllen
= 0;
6449 msg
.msg_iocb
= NULL
;
6451 flags
= sr
->msg_flags
;
6453 flags
|= MSG_DONTWAIT
;
6454 if (flags
& MSG_WAITALL
)
6455 min_ret
= iov_iter_count(&msg
.msg_iter
);
6457 ret
= sock_recvmsg(sock
, &msg
, flags
);
6458 if (ret
< min_ret
) {
6459 if (ret
== -EAGAIN
&& force_nonblock
)
6461 if (ret
== -ERESTARTSYS
)
6463 if (ret
> 0 && io_net_retry(sock
, flags
)) {
6467 req
->flags
|= REQ_F_PARTIAL_IO
;
6471 } else if ((flags
& MSG_WAITALL
) && (msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))) {
6478 else if (sr
->done_io
)
6480 cflags
= io_put_kbuf(req
, issue_flags
);
6482 cflags
|= IORING_CQE_F_SOCK_NONEMPTY
;
6483 __io_req_complete(req
, issue_flags
, ret
, cflags
);
6487 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6489 struct io_accept
*accept
= &req
->accept
;
6492 if (sqe
->len
|| sqe
->buf_index
)
6495 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
6496 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
6497 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
6498 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
6499 flags
= READ_ONCE(sqe
->ioprio
);
6500 if (flags
& ~IORING_ACCEPT_MULTISHOT
)
6503 accept
->file_slot
= READ_ONCE(sqe
->file_index
);
6504 if (accept
->file_slot
) {
6505 if (accept
->flags
& SOCK_CLOEXEC
)
6507 if (flags
& IORING_ACCEPT_MULTISHOT
&&
6508 accept
->file_slot
!= IORING_FILE_INDEX_ALLOC
)
6511 if (accept
->flags
& ~(SOCK_CLOEXEC
| SOCK_NONBLOCK
))
6513 if (SOCK_NONBLOCK
!= O_NONBLOCK
&& (accept
->flags
& SOCK_NONBLOCK
))
6514 accept
->flags
= (accept
->flags
& ~SOCK_NONBLOCK
) | O_NONBLOCK
;
6515 if (flags
& IORING_ACCEPT_MULTISHOT
)
6516 req
->flags
|= REQ_F_APOLL_MULTISHOT
;
6520 static int io_accept(struct io_kiocb
*req
, unsigned int issue_flags
)
6522 struct io_ring_ctx
*ctx
= req
->ctx
;
6523 struct io_accept
*accept
= &req
->accept
;
6524 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
6525 unsigned int file_flags
= force_nonblock
? O_NONBLOCK
: 0;
6526 bool fixed
= !!accept
->file_slot
;
6532 fd
= __get_unused_fd_flags(accept
->flags
, accept
->nofile
);
6533 if (unlikely(fd
< 0))
6536 file
= do_accept(req
->file
, file_flags
, accept
->addr
, accept
->addr_len
,
6541 ret
= PTR_ERR(file
);
6542 if (ret
== -EAGAIN
&& force_nonblock
) {
6544 * if it's multishot and polled, we don't need to
6545 * return EAGAIN to arm the poll infra since it
6546 * has already been done
6548 if ((req
->flags
& IO_APOLL_MULTI_POLLED
) ==
6549 IO_APOLL_MULTI_POLLED
)
6553 if (ret
== -ERESTARTSYS
)
6556 } else if (!fixed
) {
6557 fd_install(fd
, file
);
6560 ret
= io_fixed_fd_install(req
, issue_flags
, file
,
6564 if (!(req
->flags
& REQ_F_APOLL_MULTISHOT
)) {
6565 __io_req_complete(req
, issue_flags
, ret
, 0);
6571 spin_lock(&ctx
->completion_lock
);
6572 filled
= io_fill_cqe_aux(ctx
, req
->cqe
.user_data
, ret
,
6574 io_commit_cqring(ctx
);
6575 spin_unlock(&ctx
->completion_lock
);
6577 io_cqring_ev_posted(ctx
);
6586 static int io_socket_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6588 struct io_socket
*sock
= &req
->sock
;
6590 if (sqe
->addr
|| sqe
->rw_flags
|| sqe
->buf_index
)
6593 sock
->domain
= READ_ONCE(sqe
->fd
);
6594 sock
->type
= READ_ONCE(sqe
->off
);
6595 sock
->protocol
= READ_ONCE(sqe
->len
);
6596 sock
->file_slot
= READ_ONCE(sqe
->file_index
);
6597 sock
->nofile
= rlimit(RLIMIT_NOFILE
);
6599 sock
->flags
= sock
->type
& ~SOCK_TYPE_MASK
;
6600 if (sock
->file_slot
&& (sock
->flags
& SOCK_CLOEXEC
))
6602 if (sock
->flags
& ~(SOCK_CLOEXEC
| SOCK_NONBLOCK
))
6607 static int io_socket(struct io_kiocb
*req
, unsigned int issue_flags
)
6609 struct io_socket
*sock
= &req
->sock
;
6610 bool fixed
= !!sock
->file_slot
;
6615 fd
= __get_unused_fd_flags(sock
->flags
, sock
->nofile
);
6616 if (unlikely(fd
< 0))
6619 file
= __sys_socket_file(sock
->domain
, sock
->type
, sock
->protocol
);
6623 ret
= PTR_ERR(file
);
6624 if (ret
== -EAGAIN
&& (issue_flags
& IO_URING_F_NONBLOCK
))
6626 if (ret
== -ERESTARTSYS
)
6629 } else if (!fixed
) {
6630 fd_install(fd
, file
);
6633 ret
= io_fixed_fd_install(req
, issue_flags
, file
,
6636 __io_req_complete(req
, issue_flags
, ret
, 0);
6640 static int io_connect_prep_async(struct io_kiocb
*req
)
6642 struct io_async_connect
*io
= req
->async_data
;
6643 struct io_connect
*conn
= &req
->connect
;
6645 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
, &io
->address
);
6648 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6650 struct io_connect
*conn
= &req
->connect
;
6652 if (sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
|| sqe
->splice_fd_in
)
6655 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
6656 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
6660 static int io_connect(struct io_kiocb
*req
, unsigned int issue_flags
)
6662 struct io_async_connect __io
, *io
;
6663 unsigned file_flags
;
6665 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
6667 if (req_has_async_data(req
)) {
6668 io
= req
->async_data
;
6670 ret
= move_addr_to_kernel(req
->connect
.addr
,
6671 req
->connect
.addr_len
,
6678 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
6680 ret
= __sys_connect_file(req
->file
, &io
->address
,
6681 req
->connect
.addr_len
, file_flags
);
6682 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
6683 if (req_has_async_data(req
))
6685 if (io_alloc_async_data(req
)) {
6689 memcpy(req
->async_data
, &__io
, sizeof(__io
));
6692 if (ret
== -ERESTARTSYS
)
6697 __io_req_complete(req
, issue_flags
, ret
, 0);
6700 #else /* !CONFIG_NET */
6701 #define IO_NETOP_FN(op) \
6702 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
6704 return -EOPNOTSUPP; \
6707 #define IO_NETOP_PREP(op) \
6709 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
6711 return -EOPNOTSUPP; \
6714 #define IO_NETOP_PREP_ASYNC(op) \
6716 static int io_##op##_prep_async(struct io_kiocb *req) \
6718 return -EOPNOTSUPP; \
6721 IO_NETOP_PREP_ASYNC(sendmsg
);
6722 IO_NETOP_PREP_ASYNC(recvmsg
);
6723 IO_NETOP_PREP_ASYNC(connect
);
6724 IO_NETOP_PREP(accept
);
6725 IO_NETOP_PREP(socket
);
6726 IO_NETOP_PREP(shutdown
);
6729 #endif /* CONFIG_NET */
6731 struct io_poll_table
{
6732 struct poll_table_struct pt
;
6733 struct io_kiocb
*req
;
6738 #define IO_POLL_CANCEL_FLAG BIT(31)
6739 #define IO_POLL_REF_MASK GENMASK(30, 0)
6742 * If refs part of ->poll_refs (see IO_POLL_REF_MASK) is 0, it's free. We can
6743 * bump it and acquire ownership. It's disallowed to modify requests while not
6744 * owning it, that prevents from races for enqueueing task_work's and b/w
6745 * arming poll and wakeups.
6747 static inline bool io_poll_get_ownership(struct io_kiocb
*req
)
6749 return !(atomic_fetch_inc(&req
->poll_refs
) & IO_POLL_REF_MASK
);
6752 static void io_poll_mark_cancelled(struct io_kiocb
*req
)
6754 atomic_or(IO_POLL_CANCEL_FLAG
, &req
->poll_refs
);
6757 static struct io_poll_iocb
*io_poll_get_double(struct io_kiocb
*req
)
6759 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
6760 if (req
->opcode
== IORING_OP_POLL_ADD
)
6761 return req
->async_data
;
6762 return req
->apoll
->double_poll
;
6765 static struct io_poll_iocb
*io_poll_get_single(struct io_kiocb
*req
)
6767 if (req
->opcode
== IORING_OP_POLL_ADD
)
6769 return &req
->apoll
->poll
;
6772 static void io_poll_req_insert(struct io_kiocb
*req
)
6774 struct io_ring_ctx
*ctx
= req
->ctx
;
6775 struct hlist_head
*list
;
6777 list
= &ctx
->cancel_hash
[hash_long(req
->cqe
.user_data
, ctx
->cancel_hash_bits
)];
6778 hlist_add_head(&req
->hash_node
, list
);
6781 static void io_init_poll_iocb(struct io_poll_iocb
*poll
, __poll_t events
,
6782 wait_queue_func_t wake_func
)
6785 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
6786 /* mask in events that we always want/need */
6787 poll
->events
= events
| IO_POLL_UNMASK
;
6788 INIT_LIST_HEAD(&poll
->wait
.entry
);
6789 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
6792 static inline void io_poll_remove_entry(struct io_poll_iocb
*poll
)
6794 struct wait_queue_head
*head
= smp_load_acquire(&poll
->head
);
6797 spin_lock_irq(&head
->lock
);
6798 list_del_init(&poll
->wait
.entry
);
6800 spin_unlock_irq(&head
->lock
);
6804 static void io_poll_remove_entries(struct io_kiocb
*req
)
6807 * Nothing to do if neither of those flags are set. Avoid dipping
6808 * into the poll/apoll/double cachelines if we can.
6810 if (!(req
->flags
& (REQ_F_SINGLE_POLL
| REQ_F_DOUBLE_POLL
)))
6814 * While we hold the waitqueue lock and the waitqueue is nonempty,
6815 * wake_up_pollfree() will wait for us. However, taking the waitqueue
6816 * lock in the first place can race with the waitqueue being freed.
6818 * We solve this as eventpoll does: by taking advantage of the fact that
6819 * all users of wake_up_pollfree() will RCU-delay the actual free. If
6820 * we enter rcu_read_lock() and see that the pointer to the queue is
6821 * non-NULL, we can then lock it without the memory being freed out from
6824 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
6825 * case the caller deletes the entry from the queue, leaving it empty.
6826 * In that case, only RCU prevents the queue memory from being freed.
6829 if (req
->flags
& REQ_F_SINGLE_POLL
)
6830 io_poll_remove_entry(io_poll_get_single(req
));
6831 if (req
->flags
& REQ_F_DOUBLE_POLL
)
6832 io_poll_remove_entry(io_poll_get_double(req
));
6836 static int io_issue_sqe(struct io_kiocb
*req
, unsigned int issue_flags
);
6838 * All poll tw should go through this. Checks for poll events, manages
6839 * references, does rewait, etc.
6841 * Returns a negative error on failure. >0 when no action require, which is
6842 * either spurious wakeup or multishot CQE is served. 0 when it's done with
6843 * the request, then the mask is stored in req->cqe.res.
6845 static int io_poll_check_events(struct io_kiocb
*req
, bool *locked
)
6847 struct io_ring_ctx
*ctx
= req
->ctx
;
6850 /* req->task == current here, checking PF_EXITING is safe */
6851 if (unlikely(req
->task
->flags
& PF_EXITING
))
6855 v
= atomic_read(&req
->poll_refs
);
6857 /* tw handler should be the owner, and so have some references */
6858 if (WARN_ON_ONCE(!(v
& IO_POLL_REF_MASK
)))
6860 if (v
& IO_POLL_CANCEL_FLAG
)
6863 if (!req
->cqe
.res
) {
6864 struct poll_table_struct pt
= { ._key
= req
->apoll_events
};
6865 req
->cqe
.res
= vfs_poll(req
->file
, &pt
) & req
->apoll_events
;
6868 if ((unlikely(!req
->cqe
.res
)))
6870 if (req
->apoll_events
& EPOLLONESHOT
)
6873 /* multishot, just fill a CQE and proceed */
6874 if (!(req
->flags
& REQ_F_APOLL_MULTISHOT
)) {
6875 __poll_t mask
= mangle_poll(req
->cqe
.res
&
6879 spin_lock(&ctx
->completion_lock
);
6880 filled
= io_fill_cqe_aux(ctx
, req
->cqe
.user_data
,
6881 mask
, IORING_CQE_F_MORE
);
6882 io_commit_cqring(ctx
);
6883 spin_unlock(&ctx
->completion_lock
);
6885 io_cqring_ev_posted(ctx
);
6891 io_tw_lock(req
->ctx
, locked
);
6892 if (unlikely(req
->task
->flags
& PF_EXITING
))
6894 ret
= io_issue_sqe(req
,
6895 IO_URING_F_NONBLOCK
|IO_URING_F_COMPLETE_DEFER
);
6900 * Release all references, retry if someone tried to restart
6901 * task_work while we were executing it.
6903 } while (atomic_sub_return(v
& IO_POLL_REF_MASK
, &req
->poll_refs
));
6908 static void io_poll_task_func(struct io_kiocb
*req
, bool *locked
)
6910 struct io_ring_ctx
*ctx
= req
->ctx
;
6913 ret
= io_poll_check_events(req
, locked
);
6918 req
->cqe
.res
= mangle_poll(req
->cqe
.res
& req
->poll
.events
);
6924 io_poll_remove_entries(req
);
6925 spin_lock(&ctx
->completion_lock
);
6926 hash_del(&req
->hash_node
);
6927 __io_req_complete_post(req
, req
->cqe
.res
, 0);
6928 io_commit_cqring(ctx
);
6929 spin_unlock(&ctx
->completion_lock
);
6930 io_cqring_ev_posted(ctx
);
6933 static void io_apoll_task_func(struct io_kiocb
*req
, bool *locked
)
6935 struct io_ring_ctx
*ctx
= req
->ctx
;
6938 ret
= io_poll_check_events(req
, locked
);
6942 io_poll_remove_entries(req
);
6943 spin_lock(&ctx
->completion_lock
);
6944 hash_del(&req
->hash_node
);
6945 spin_unlock(&ctx
->completion_lock
);
6948 io_req_task_submit(req
, locked
);
6950 io_req_complete_failed(req
, ret
);
6953 static void __io_poll_execute(struct io_kiocb
*req
, int mask
,
6954 __poll_t __maybe_unused events
)
6956 req
->cqe
.res
= mask
;
6958 * This is useful for poll that is armed on behalf of another
6959 * request, and where the wakeup path could be on a different
6960 * CPU. We want to avoid pulling in req->apoll->events for that
6963 if (req
->opcode
== IORING_OP_POLL_ADD
)
6964 req
->io_task_work
.func
= io_poll_task_func
;
6966 req
->io_task_work
.func
= io_apoll_task_func
;
6968 trace_io_uring_task_add(req
->ctx
, req
, req
->cqe
.user_data
, req
->opcode
, mask
);
6969 io_req_task_work_add(req
);
6972 static inline void io_poll_execute(struct io_kiocb
*req
, int res
,
6975 if (io_poll_get_ownership(req
))
6976 __io_poll_execute(req
, res
, events
);
6979 static void io_poll_cancel_req(struct io_kiocb
*req
)
6981 io_poll_mark_cancelled(req
);
6982 /* kick tw, which should complete the request */
6983 io_poll_execute(req
, 0, 0);
6986 #define wqe_to_req(wait) ((void *)((unsigned long) (wait)->private & ~1))
6987 #define wqe_is_double(wait) ((unsigned long) (wait)->private & 1)
6988 #define IO_ASYNC_POLL_COMMON (EPOLLONESHOT | EPOLLPRI)
6990 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
6993 struct io_kiocb
*req
= wqe_to_req(wait
);
6994 struct io_poll_iocb
*poll
= container_of(wait
, struct io_poll_iocb
,
6996 __poll_t mask
= key_to_poll(key
);
6998 if (unlikely(mask
& POLLFREE
)) {
6999 io_poll_mark_cancelled(req
);
7000 /* we have to kick tw in case it's not already */
7001 io_poll_execute(req
, 0, poll
->events
);
7004 * If the waitqueue is being freed early but someone is already
7005 * holds ownership over it, we have to tear down the request as
7006 * best we can. That means immediately removing the request from
7007 * its waitqueue and preventing all further accesses to the
7008 * waitqueue via the request.
7010 list_del_init(&poll
->wait
.entry
);
7013 * Careful: this *must* be the last step, since as soon
7014 * as req->head is NULL'ed out, the request can be
7015 * completed and freed, since aio_poll_complete_work()
7016 * will no longer need to take the waitqueue lock.
7018 smp_store_release(&poll
->head
, NULL
);
7022 /* for instances that support it check for an event match first */
7023 if (mask
&& !(mask
& (poll
->events
& ~IO_ASYNC_POLL_COMMON
)))
7026 if (io_poll_get_ownership(req
)) {
7027 /* optional, saves extra locking for removal in tw handler */
7028 if (mask
&& poll
->events
& EPOLLONESHOT
) {
7029 list_del_init(&poll
->wait
.entry
);
7031 if (wqe_is_double(wait
))
7032 req
->flags
&= ~REQ_F_DOUBLE_POLL
;
7034 req
->flags
&= ~REQ_F_SINGLE_POLL
;
7036 __io_poll_execute(req
, mask
, poll
->events
);
7041 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
7042 struct wait_queue_head
*head
,
7043 struct io_poll_iocb
**poll_ptr
)
7045 struct io_kiocb
*req
= pt
->req
;
7046 unsigned long wqe_private
= (unsigned long) req
;
7049 * The file being polled uses multiple waitqueues for poll handling
7050 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
7053 if (unlikely(pt
->nr_entries
)) {
7054 struct io_poll_iocb
*first
= poll
;
7056 /* double add on the same waitqueue head, ignore */
7057 if (first
->head
== head
)
7059 /* already have a 2nd entry, fail a third attempt */
7061 if ((*poll_ptr
)->head
== head
)
7063 pt
->error
= -EINVAL
;
7067 poll
= kmalloc(sizeof(*poll
), GFP_ATOMIC
);
7069 pt
->error
= -ENOMEM
;
7072 /* mark as double wq entry */
7074 req
->flags
|= REQ_F_DOUBLE_POLL
;
7075 io_init_poll_iocb(poll
, first
->events
, first
->wait
.func
);
7077 if (req
->opcode
== IORING_OP_POLL_ADD
)
7078 req
->flags
|= REQ_F_ASYNC_DATA
;
7081 req
->flags
|= REQ_F_SINGLE_POLL
;
7084 poll
->wait
.private = (void *) wqe_private
;
7086 if (poll
->events
& EPOLLEXCLUSIVE
)
7087 add_wait_queue_exclusive(head
, &poll
->wait
);
7089 add_wait_queue(head
, &poll
->wait
);
7092 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
7093 struct poll_table_struct
*p
)
7095 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
7097 __io_queue_proc(&pt
->req
->poll
, pt
, head
,
7098 (struct io_poll_iocb
**) &pt
->req
->async_data
);
7101 static int __io_arm_poll_handler(struct io_kiocb
*req
,
7102 struct io_poll_iocb
*poll
,
7103 struct io_poll_table
*ipt
, __poll_t mask
)
7105 struct io_ring_ctx
*ctx
= req
->ctx
;
7108 INIT_HLIST_NODE(&req
->hash_node
);
7109 req
->work
.cancel_seq
= atomic_read(&ctx
->cancel_seq
);
7110 io_init_poll_iocb(poll
, mask
, io_poll_wake
);
7111 poll
->file
= req
->file
;
7113 req
->apoll_events
= poll
->events
;
7115 ipt
->pt
._key
= mask
;
7118 ipt
->nr_entries
= 0;
7121 * Take the ownership to delay any tw execution up until we're done
7122 * with poll arming. see io_poll_get_ownership().
7124 atomic_set(&req
->poll_refs
, 1);
7125 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
7127 if (mask
&& (poll
->events
& EPOLLONESHOT
)) {
7128 io_poll_remove_entries(req
);
7129 /* no one else has access to the req, forget about the ref */
7132 if (!mask
&& unlikely(ipt
->error
|| !ipt
->nr_entries
)) {
7133 io_poll_remove_entries(req
);
7135 ipt
->error
= -EINVAL
;
7139 spin_lock(&ctx
->completion_lock
);
7140 io_poll_req_insert(req
);
7141 spin_unlock(&ctx
->completion_lock
);
7144 /* can't multishot if failed, just queue the event we've got */
7145 if (unlikely(ipt
->error
|| !ipt
->nr_entries
)) {
7146 poll
->events
|= EPOLLONESHOT
;
7147 req
->apoll_events
|= EPOLLONESHOT
;
7150 __io_poll_execute(req
, mask
, poll
->events
);
7155 * Release ownership. If someone tried to queue a tw while it was
7156 * locked, kick it off for them.
7158 v
= atomic_dec_return(&req
->poll_refs
);
7159 if (unlikely(v
& IO_POLL_REF_MASK
))
7160 __io_poll_execute(req
, 0, poll
->events
);
7164 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
7165 struct poll_table_struct
*p
)
7167 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
7168 struct async_poll
*apoll
= pt
->req
->apoll
;
7170 __io_queue_proc(&apoll
->poll
, pt
, head
, &apoll
->double_poll
);
7179 static int io_arm_poll_handler(struct io_kiocb
*req
, unsigned issue_flags
)
7181 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
7182 struct io_ring_ctx
*ctx
= req
->ctx
;
7183 struct async_poll
*apoll
;
7184 struct io_poll_table ipt
;
7185 __poll_t mask
= POLLPRI
| POLLERR
;
7188 if (!def
->pollin
&& !def
->pollout
)
7189 return IO_APOLL_ABORTED
;
7190 if (!file_can_poll(req
->file
))
7191 return IO_APOLL_ABORTED
;
7192 if ((req
->flags
& (REQ_F_POLLED
|REQ_F_PARTIAL_IO
)) == REQ_F_POLLED
)
7193 return IO_APOLL_ABORTED
;
7194 if (!(req
->flags
& REQ_F_APOLL_MULTISHOT
))
7195 mask
|= EPOLLONESHOT
;
7198 mask
|= EPOLLIN
| EPOLLRDNORM
;
7200 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
7201 if ((req
->opcode
== IORING_OP_RECVMSG
) &&
7202 (req
->sr_msg
.msg_flags
& MSG_ERRQUEUE
))
7205 mask
|= EPOLLOUT
| EPOLLWRNORM
;
7207 if (def
->poll_exclusive
)
7208 mask
|= EPOLLEXCLUSIVE
;
7209 if (req
->flags
& REQ_F_POLLED
) {
7211 kfree(apoll
->double_poll
);
7212 } else if (!(issue_flags
& IO_URING_F_UNLOCKED
) &&
7213 !list_empty(&ctx
->apoll_cache
)) {
7214 apoll
= list_first_entry(&ctx
->apoll_cache
, struct async_poll
,
7216 list_del_init(&apoll
->poll
.wait
.entry
);
7218 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
7219 if (unlikely(!apoll
))
7220 return IO_APOLL_ABORTED
;
7222 apoll
->double_poll
= NULL
;
7224 req
->flags
|= REQ_F_POLLED
;
7225 ipt
.pt
._qproc
= io_async_queue_proc
;
7227 io_kbuf_recycle(req
, issue_flags
);
7229 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
);
7230 if (ret
|| ipt
.error
)
7231 return ret
? IO_APOLL_READY
: IO_APOLL_ABORTED
;
7233 trace_io_uring_poll_arm(ctx
, req
, req
->cqe
.user_data
, req
->opcode
,
7234 mask
, apoll
->poll
.events
);
7239 * Returns true if we found and killed one or more poll requests
7241 static __cold
bool io_poll_remove_all(struct io_ring_ctx
*ctx
,
7242 struct task_struct
*tsk
, bool cancel_all
)
7244 struct hlist_node
*tmp
;
7245 struct io_kiocb
*req
;
7249 spin_lock(&ctx
->completion_lock
);
7250 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
7251 struct hlist_head
*list
;
7253 list
= &ctx
->cancel_hash
[i
];
7254 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
) {
7255 if (io_match_task_safe(req
, tsk
, cancel_all
)) {
7256 hlist_del_init(&req
->hash_node
);
7257 io_poll_cancel_req(req
);
7262 spin_unlock(&ctx
->completion_lock
);
7266 static struct io_kiocb
*io_poll_find(struct io_ring_ctx
*ctx
, bool poll_only
,
7267 struct io_cancel_data
*cd
)
7268 __must_hold(&ctx
->completion_lock
)
7270 struct hlist_head
*list
;
7271 struct io_kiocb
*req
;
7273 list
= &ctx
->cancel_hash
[hash_long(cd
->data
, ctx
->cancel_hash_bits
)];
7274 hlist_for_each_entry(req
, list
, hash_node
) {
7275 if (cd
->data
!= req
->cqe
.user_data
)
7277 if (poll_only
&& req
->opcode
!= IORING_OP_POLL_ADD
)
7279 if (cd
->flags
& IORING_ASYNC_CANCEL_ALL
) {
7280 if (cd
->seq
== req
->work
.cancel_seq
)
7282 req
->work
.cancel_seq
= cd
->seq
;
7289 static struct io_kiocb
*io_poll_file_find(struct io_ring_ctx
*ctx
,
7290 struct io_cancel_data
*cd
)
7291 __must_hold(&ctx
->completion_lock
)
7293 struct io_kiocb
*req
;
7296 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
7297 struct hlist_head
*list
;
7299 list
= &ctx
->cancel_hash
[i
];
7300 hlist_for_each_entry(req
, list
, hash_node
) {
7301 if (!(cd
->flags
& IORING_ASYNC_CANCEL_ANY
) &&
7302 req
->file
!= cd
->file
)
7304 if (cd
->seq
== req
->work
.cancel_seq
)
7306 req
->work
.cancel_seq
= cd
->seq
;
7313 static bool io_poll_disarm(struct io_kiocb
*req
)
7314 __must_hold(&ctx
->completion_lock
)
7316 if (!io_poll_get_ownership(req
))
7318 io_poll_remove_entries(req
);
7319 hash_del(&req
->hash_node
);
7323 static int io_poll_cancel(struct io_ring_ctx
*ctx
, struct io_cancel_data
*cd
)
7324 __must_hold(&ctx
->completion_lock
)
7326 struct io_kiocb
*req
;
7328 if (cd
->flags
& (IORING_ASYNC_CANCEL_FD
|IORING_ASYNC_CANCEL_ANY
))
7329 req
= io_poll_file_find(ctx
, cd
);
7331 req
= io_poll_find(ctx
, false, cd
);
7334 io_poll_cancel_req(req
);
7338 static __poll_t
io_poll_parse_events(const struct io_uring_sqe
*sqe
,
7343 events
= READ_ONCE(sqe
->poll32_events
);
7345 events
= swahw32(events
);
7347 if (!(flags
& IORING_POLL_ADD_MULTI
))
7348 events
|= EPOLLONESHOT
;
7349 return demangle_poll(events
) | (events
& (EPOLLEXCLUSIVE
|EPOLLONESHOT
));
7352 static int io_poll_remove_prep(struct io_kiocb
*req
,
7353 const struct io_uring_sqe
*sqe
)
7355 struct io_poll_update
*upd
= &req
->poll_update
;
7358 if (sqe
->buf_index
|| sqe
->splice_fd_in
)
7360 flags
= READ_ONCE(sqe
->len
);
7361 if (flags
& ~(IORING_POLL_UPDATE_EVENTS
| IORING_POLL_UPDATE_USER_DATA
|
7362 IORING_POLL_ADD_MULTI
))
7364 /* meaningless without update */
7365 if (flags
== IORING_POLL_ADD_MULTI
)
7368 upd
->old_user_data
= READ_ONCE(sqe
->addr
);
7369 upd
->update_events
= flags
& IORING_POLL_UPDATE_EVENTS
;
7370 upd
->update_user_data
= flags
& IORING_POLL_UPDATE_USER_DATA
;
7372 upd
->new_user_data
= READ_ONCE(sqe
->off
);
7373 if (!upd
->update_user_data
&& upd
->new_user_data
)
7375 if (upd
->update_events
)
7376 upd
->events
= io_poll_parse_events(sqe
, flags
);
7377 else if (sqe
->poll32_events
)
7383 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
7385 struct io_poll_iocb
*poll
= &req
->poll
;
7388 if (sqe
->buf_index
|| sqe
->off
|| sqe
->addr
)
7390 flags
= READ_ONCE(sqe
->len
);
7391 if (flags
& ~IORING_POLL_ADD_MULTI
)
7393 if ((flags
& IORING_POLL_ADD_MULTI
) && (req
->flags
& REQ_F_CQE_SKIP
))
7396 io_req_set_refcount(req
);
7397 poll
->events
= io_poll_parse_events(sqe
, flags
);
7401 static int io_poll_add(struct io_kiocb
*req
, unsigned int issue_flags
)
7403 struct io_poll_iocb
*poll
= &req
->poll
;
7404 struct io_poll_table ipt
;
7407 ipt
.pt
._qproc
= io_poll_queue_proc
;
7409 ret
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
);
7410 if (!ret
&& ipt
.error
)
7412 ret
= ret
?: ipt
.error
;
7414 __io_req_complete(req
, issue_flags
, ret
, 0);
7418 static int io_poll_remove(struct io_kiocb
*req
, unsigned int issue_flags
)
7420 struct io_cancel_data cd
= { .data
= req
->poll_update
.old_user_data
, };
7421 struct io_ring_ctx
*ctx
= req
->ctx
;
7422 struct io_kiocb
*preq
;
7426 spin_lock(&ctx
->completion_lock
);
7427 preq
= io_poll_find(ctx
, true, &cd
);
7428 if (!preq
|| !io_poll_disarm(preq
)) {
7429 spin_unlock(&ctx
->completion_lock
);
7430 ret
= preq
? -EALREADY
: -ENOENT
;
7433 spin_unlock(&ctx
->completion_lock
);
7435 if (req
->poll_update
.update_events
|| req
->poll_update
.update_user_data
) {
7436 /* only mask one event flags, keep behavior flags */
7437 if (req
->poll_update
.update_events
) {
7438 preq
->poll
.events
&= ~0xffff;
7439 preq
->poll
.events
|= req
->poll_update
.events
& 0xffff;
7440 preq
->poll
.events
|= IO_POLL_UNMASK
;
7442 if (req
->poll_update
.update_user_data
)
7443 preq
->cqe
.user_data
= req
->poll_update
.new_user_data
;
7445 ret2
= io_poll_add(preq
, issue_flags
);
7446 /* successfully updated, don't complete poll request */
7452 preq
->cqe
.res
= -ECANCELED
;
7453 locked
= !(issue_flags
& IO_URING_F_UNLOCKED
);
7454 io_req_task_complete(preq
, &locked
);
7458 /* complete update request, we're done with it */
7459 __io_req_complete(req
, issue_flags
, ret
, 0);
7463 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
7465 struct io_timeout_data
*data
= container_of(timer
,
7466 struct io_timeout_data
, timer
);
7467 struct io_kiocb
*req
= data
->req
;
7468 struct io_ring_ctx
*ctx
= req
->ctx
;
7469 unsigned long flags
;
7471 spin_lock_irqsave(&ctx
->timeout_lock
, flags
);
7472 list_del_init(&req
->timeout
.list
);
7473 atomic_set(&req
->ctx
->cq_timeouts
,
7474 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
7475 spin_unlock_irqrestore(&ctx
->timeout_lock
, flags
);
7477 if (!(data
->flags
& IORING_TIMEOUT_ETIME_SUCCESS
))
7480 req
->cqe
.res
= -ETIME
;
7481 req
->io_task_work
.func
= io_req_task_complete
;
7482 io_req_task_work_add(req
);
7483 return HRTIMER_NORESTART
;
7486 static struct io_kiocb
*io_timeout_extract(struct io_ring_ctx
*ctx
,
7487 struct io_cancel_data
*cd
)
7488 __must_hold(&ctx
->timeout_lock
)
7490 struct io_timeout_data
*io
;
7491 struct io_kiocb
*req
;
7494 list_for_each_entry(req
, &ctx
->timeout_list
, timeout
.list
) {
7495 if (!(cd
->flags
& IORING_ASYNC_CANCEL_ANY
) &&
7496 cd
->data
!= req
->cqe
.user_data
)
7498 if (cd
->flags
& (IORING_ASYNC_CANCEL_ALL
|IORING_ASYNC_CANCEL_ANY
)) {
7499 if (cd
->seq
== req
->work
.cancel_seq
)
7501 req
->work
.cancel_seq
= cd
->seq
;
7507 return ERR_PTR(-ENOENT
);
7509 io
= req
->async_data
;
7510 if (hrtimer_try_to_cancel(&io
->timer
) == -1)
7511 return ERR_PTR(-EALREADY
);
7512 list_del_init(&req
->timeout
.list
);
7516 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, struct io_cancel_data
*cd
)
7517 __must_hold(&ctx
->completion_lock
)
7519 struct io_kiocb
*req
;
7521 spin_lock_irq(&ctx
->timeout_lock
);
7522 req
= io_timeout_extract(ctx
, cd
);
7523 spin_unlock_irq(&ctx
->timeout_lock
);
7526 return PTR_ERR(req
);
7527 io_req_task_queue_fail(req
, -ECANCELED
);
7531 static clockid_t
io_timeout_get_clock(struct io_timeout_data
*data
)
7533 switch (data
->flags
& IORING_TIMEOUT_CLOCK_MASK
) {
7534 case IORING_TIMEOUT_BOOTTIME
:
7535 return CLOCK_BOOTTIME
;
7536 case IORING_TIMEOUT_REALTIME
:
7537 return CLOCK_REALTIME
;
7539 /* can't happen, vetted at prep time */
7543 return CLOCK_MONOTONIC
;
7547 static int io_linked_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
7548 struct timespec64
*ts
, enum hrtimer_mode mode
)
7549 __must_hold(&ctx
->timeout_lock
)
7551 struct io_timeout_data
*io
;
7552 struct io_kiocb
*req
;
7555 list_for_each_entry(req
, &ctx
->ltimeout_list
, timeout
.list
) {
7556 found
= user_data
== req
->cqe
.user_data
;
7563 io
= req
->async_data
;
7564 if (hrtimer_try_to_cancel(&io
->timer
) == -1)
7566 hrtimer_init(&io
->timer
, io_timeout_get_clock(io
), mode
);
7567 io
->timer
.function
= io_link_timeout_fn
;
7568 hrtimer_start(&io
->timer
, timespec64_to_ktime(*ts
), mode
);
7572 static int io_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
7573 struct timespec64
*ts
, enum hrtimer_mode mode
)
7574 __must_hold(&ctx
->timeout_lock
)
7576 struct io_cancel_data cd
= { .data
= user_data
, };
7577 struct io_kiocb
*req
= io_timeout_extract(ctx
, &cd
);
7578 struct io_timeout_data
*data
;
7581 return PTR_ERR(req
);
7583 req
->timeout
.off
= 0; /* noseq */
7584 data
= req
->async_data
;
7585 list_add_tail(&req
->timeout
.list
, &ctx
->timeout_list
);
7586 hrtimer_init(&data
->timer
, io_timeout_get_clock(data
), mode
);
7587 data
->timer
.function
= io_timeout_fn
;
7588 hrtimer_start(&data
->timer
, timespec64_to_ktime(*ts
), mode
);
7592 static int io_timeout_remove_prep(struct io_kiocb
*req
,
7593 const struct io_uring_sqe
*sqe
)
7595 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
7597 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
7599 if (sqe
->buf_index
|| sqe
->len
|| sqe
->splice_fd_in
)
7602 tr
->ltimeout
= false;
7603 tr
->addr
= READ_ONCE(sqe
->addr
);
7604 tr
->flags
= READ_ONCE(sqe
->timeout_flags
);
7605 if (tr
->flags
& IORING_TIMEOUT_UPDATE_MASK
) {
7606 if (hweight32(tr
->flags
& IORING_TIMEOUT_CLOCK_MASK
) > 1)
7608 if (tr
->flags
& IORING_LINK_TIMEOUT_UPDATE
)
7609 tr
->ltimeout
= true;
7610 if (tr
->flags
& ~(IORING_TIMEOUT_UPDATE_MASK
|IORING_TIMEOUT_ABS
))
7612 if (get_timespec64(&tr
->ts
, u64_to_user_ptr(sqe
->addr2
)))
7614 if (tr
->ts
.tv_sec
< 0 || tr
->ts
.tv_nsec
< 0)
7616 } else if (tr
->flags
) {
7617 /* timeout removal doesn't support flags */
7624 static inline enum hrtimer_mode
io_translate_timeout_mode(unsigned int flags
)
7626 return (flags
& IORING_TIMEOUT_ABS
) ? HRTIMER_MODE_ABS
7631 * Remove or update an existing timeout command
7633 static int io_timeout_remove(struct io_kiocb
*req
, unsigned int issue_flags
)
7635 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
7636 struct io_ring_ctx
*ctx
= req
->ctx
;
7639 if (!(req
->timeout_rem
.flags
& IORING_TIMEOUT_UPDATE
)) {
7640 struct io_cancel_data cd
= { .data
= tr
->addr
, };
7642 spin_lock(&ctx
->completion_lock
);
7643 ret
= io_timeout_cancel(ctx
, &cd
);
7644 spin_unlock(&ctx
->completion_lock
);
7646 enum hrtimer_mode mode
= io_translate_timeout_mode(tr
->flags
);
7648 spin_lock_irq(&ctx
->timeout_lock
);
7650 ret
= io_linked_timeout_update(ctx
, tr
->addr
, &tr
->ts
, mode
);
7652 ret
= io_timeout_update(ctx
, tr
->addr
, &tr
->ts
, mode
);
7653 spin_unlock_irq(&ctx
->timeout_lock
);
7658 io_req_complete_post(req
, ret
, 0);
7662 static int __io_timeout_prep(struct io_kiocb
*req
,
7663 const struct io_uring_sqe
*sqe
,
7664 bool is_timeout_link
)
7666 struct io_timeout_data
*data
;
7668 u32 off
= READ_ONCE(sqe
->off
);
7670 if (sqe
->buf_index
|| sqe
->len
!= 1 || sqe
->splice_fd_in
)
7672 if (off
&& is_timeout_link
)
7674 flags
= READ_ONCE(sqe
->timeout_flags
);
7675 if (flags
& ~(IORING_TIMEOUT_ABS
| IORING_TIMEOUT_CLOCK_MASK
|
7676 IORING_TIMEOUT_ETIME_SUCCESS
))
7678 /* more than one clock specified is invalid, obviously */
7679 if (hweight32(flags
& IORING_TIMEOUT_CLOCK_MASK
) > 1)
7682 INIT_LIST_HEAD(&req
->timeout
.list
);
7683 req
->timeout
.off
= off
;
7684 if (unlikely(off
&& !req
->ctx
->off_timeout_used
))
7685 req
->ctx
->off_timeout_used
= true;
7687 if (WARN_ON_ONCE(req_has_async_data(req
)))
7689 if (io_alloc_async_data(req
))
7692 data
= req
->async_data
;
7694 data
->flags
= flags
;
7696 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
7699 if (data
->ts
.tv_sec
< 0 || data
->ts
.tv_nsec
< 0)
7702 INIT_LIST_HEAD(&req
->timeout
.list
);
7703 data
->mode
= io_translate_timeout_mode(flags
);
7704 hrtimer_init(&data
->timer
, io_timeout_get_clock(data
), data
->mode
);
7706 if (is_timeout_link
) {
7707 struct io_submit_link
*link
= &req
->ctx
->submit_state
.link
;
7711 if (link
->last
->opcode
== IORING_OP_LINK_TIMEOUT
)
7713 req
->timeout
.head
= link
->last
;
7714 link
->last
->flags
|= REQ_F_ARM_LTIMEOUT
;
7719 static int io_timeout_prep(struct io_kiocb
*req
,
7720 const struct io_uring_sqe
*sqe
)
7722 return __io_timeout_prep(req
, sqe
, false);
7725 static int io_link_timeout_prep(struct io_kiocb
*req
,
7726 const struct io_uring_sqe
*sqe
)
7728 return __io_timeout_prep(req
, sqe
, true);
7731 static int io_timeout(struct io_kiocb
*req
, unsigned int issue_flags
)
7733 struct io_ring_ctx
*ctx
= req
->ctx
;
7734 struct io_timeout_data
*data
= req
->async_data
;
7735 struct list_head
*entry
;
7736 u32 tail
, off
= req
->timeout
.off
;
7738 spin_lock_irq(&ctx
->timeout_lock
);
7741 * sqe->off holds how many events that need to occur for this
7742 * timeout event to be satisfied. If it isn't set, then this is
7743 * a pure timeout request, sequence isn't used.
7745 if (io_is_timeout_noseq(req
)) {
7746 entry
= ctx
->timeout_list
.prev
;
7750 tail
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
7751 req
->timeout
.target_seq
= tail
+ off
;
7753 /* Update the last seq here in case io_flush_timeouts() hasn't.
7754 * This is safe because ->completion_lock is held, and submissions
7755 * and completions are never mixed in the same ->completion_lock section.
7757 ctx
->cq_last_tm_flush
= tail
;
7760 * Insertion sort, ensuring the first entry in the list is always
7761 * the one we need first.
7763 list_for_each_prev(entry
, &ctx
->timeout_list
) {
7764 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
,
7767 if (io_is_timeout_noseq(nxt
))
7769 /* nxt.seq is behind @tail, otherwise would've been completed */
7770 if (off
>= nxt
->timeout
.target_seq
- tail
)
7774 list_add(&req
->timeout
.list
, entry
);
7775 data
->timer
.function
= io_timeout_fn
;
7776 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
7777 spin_unlock_irq(&ctx
->timeout_lock
);
7781 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
7783 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
7784 struct io_cancel_data
*cd
= data
;
7786 if (req
->ctx
!= cd
->ctx
)
7788 if (cd
->flags
& IORING_ASYNC_CANCEL_ANY
) {
7790 } else if (cd
->flags
& IORING_ASYNC_CANCEL_FD
) {
7791 if (req
->file
!= cd
->file
)
7794 if (req
->cqe
.user_data
!= cd
->data
)
7797 if (cd
->flags
& (IORING_ASYNC_CANCEL_ALL
|IORING_ASYNC_CANCEL_ANY
)) {
7798 if (cd
->seq
== req
->work
.cancel_seq
)
7800 req
->work
.cancel_seq
= cd
->seq
;
7805 static int io_async_cancel_one(struct io_uring_task
*tctx
,
7806 struct io_cancel_data
*cd
)
7808 enum io_wq_cancel cancel_ret
;
7812 if (!tctx
|| !tctx
->io_wq
)
7815 all
= cd
->flags
& (IORING_ASYNC_CANCEL_ALL
|IORING_ASYNC_CANCEL_ANY
);
7816 cancel_ret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_cb
, cd
, all
);
7817 switch (cancel_ret
) {
7818 case IO_WQ_CANCEL_OK
:
7821 case IO_WQ_CANCEL_RUNNING
:
7824 case IO_WQ_CANCEL_NOTFOUND
:
7832 static int io_try_cancel(struct io_kiocb
*req
, struct io_cancel_data
*cd
)
7834 struct io_ring_ctx
*ctx
= req
->ctx
;
7837 WARN_ON_ONCE(!io_wq_current_is_worker() && req
->task
!= current
);
7839 ret
= io_async_cancel_one(req
->task
->io_uring
, cd
);
7841 * Fall-through even for -EALREADY, as we may have poll armed
7842 * that need unarming.
7847 spin_lock(&ctx
->completion_lock
);
7848 ret
= io_poll_cancel(ctx
, cd
);
7851 if (!(cd
->flags
& IORING_ASYNC_CANCEL_FD
))
7852 ret
= io_timeout_cancel(ctx
, cd
);
7854 spin_unlock(&ctx
->completion_lock
);
7858 #define CANCEL_FLAGS (IORING_ASYNC_CANCEL_ALL | IORING_ASYNC_CANCEL_FD | \
7859 IORING_ASYNC_CANCEL_ANY)
7861 static int io_async_cancel_prep(struct io_kiocb
*req
,
7862 const struct io_uring_sqe
*sqe
)
7864 if (unlikely(req
->flags
& REQ_F_BUFFER_SELECT
))
7866 if (sqe
->off
|| sqe
->len
|| sqe
->splice_fd_in
)
7869 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
7870 req
->cancel
.flags
= READ_ONCE(sqe
->cancel_flags
);
7871 if (req
->cancel
.flags
& ~CANCEL_FLAGS
)
7873 if (req
->cancel
.flags
& IORING_ASYNC_CANCEL_FD
) {
7874 if (req
->cancel
.flags
& IORING_ASYNC_CANCEL_ANY
)
7876 req
->cancel
.fd
= READ_ONCE(sqe
->fd
);
7882 static int __io_async_cancel(struct io_cancel_data
*cd
, struct io_kiocb
*req
,
7883 unsigned int issue_flags
)
7885 bool all
= cd
->flags
& (IORING_ASYNC_CANCEL_ALL
|IORING_ASYNC_CANCEL_ANY
);
7886 struct io_ring_ctx
*ctx
= cd
->ctx
;
7887 struct io_tctx_node
*node
;
7891 ret
= io_try_cancel(req
, cd
);
7899 /* slow path, try all io-wq's */
7900 io_ring_submit_lock(ctx
, issue_flags
);
7902 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
7903 struct io_uring_task
*tctx
= node
->task
->io_uring
;
7905 ret
= io_async_cancel_one(tctx
, cd
);
7906 if (ret
!= -ENOENT
) {
7912 io_ring_submit_unlock(ctx
, issue_flags
);
7913 return all
? nr
: ret
;
7916 static int io_async_cancel(struct io_kiocb
*req
, unsigned int issue_flags
)
7918 struct io_cancel_data cd
= {
7920 .data
= req
->cancel
.addr
,
7921 .flags
= req
->cancel
.flags
,
7922 .seq
= atomic_inc_return(&req
->ctx
->cancel_seq
),
7926 if (cd
.flags
& IORING_ASYNC_CANCEL_FD
) {
7927 if (req
->flags
& REQ_F_FIXED_FILE
)
7928 req
->file
= io_file_get_fixed(req
, req
->cancel
.fd
,
7931 req
->file
= io_file_get_normal(req
, req
->cancel
.fd
);
7936 cd
.file
= req
->file
;
7939 ret
= __io_async_cancel(&cd
, req
, issue_flags
);
7943 io_req_complete_post(req
, ret
, 0);
7947 static int io_files_update_prep(struct io_kiocb
*req
,
7948 const struct io_uring_sqe
*sqe
)
7950 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
7952 if (sqe
->rw_flags
|| sqe
->splice_fd_in
)
7955 req
->rsrc_update
.offset
= READ_ONCE(sqe
->off
);
7956 req
->rsrc_update
.nr_args
= READ_ONCE(sqe
->len
);
7957 if (!req
->rsrc_update
.nr_args
)
7959 req
->rsrc_update
.arg
= READ_ONCE(sqe
->addr
);
7963 static int io_files_update_with_index_alloc(struct io_kiocb
*req
,
7964 unsigned int issue_flags
)
7966 __s32 __user
*fds
= u64_to_user_ptr(req
->rsrc_update
.arg
);
7971 for (done
= 0; done
< req
->rsrc_update
.nr_args
; done
++) {
7972 if (copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
7982 ret
= io_fixed_fd_install(req
, issue_flags
, file
,
7983 IORING_FILE_INDEX_ALLOC
);
7986 if (copy_to_user(&fds
[done
], &ret
, sizeof(ret
))) {
7987 __io_close_fixed(req
, issue_flags
, ret
);
7998 static int io_files_update(struct io_kiocb
*req
, unsigned int issue_flags
)
8000 struct io_ring_ctx
*ctx
= req
->ctx
;
8001 struct io_uring_rsrc_update2 up
;
8004 up
.offset
= req
->rsrc_update
.offset
;
8005 up
.data
= req
->rsrc_update
.arg
;
8011 if (req
->rsrc_update
.offset
== IORING_FILE_INDEX_ALLOC
) {
8012 ret
= io_files_update_with_index_alloc(req
, issue_flags
);
8014 io_ring_submit_lock(ctx
, issue_flags
);
8015 ret
= __io_register_rsrc_update(ctx
, IORING_RSRC_FILE
,
8016 &up
, req
->rsrc_update
.nr_args
);
8017 io_ring_submit_unlock(ctx
, issue_flags
);
8022 __io_req_complete(req
, issue_flags
, ret
, 0);
8026 static int io_req_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
8028 switch (req
->opcode
) {
8030 return io_nop_prep(req
, sqe
);
8031 case IORING_OP_READV
:
8032 case IORING_OP_READ_FIXED
:
8033 case IORING_OP_READ
:
8034 case IORING_OP_WRITEV
:
8035 case IORING_OP_WRITE_FIXED
:
8036 case IORING_OP_WRITE
:
8037 return io_prep_rw(req
, sqe
);
8038 case IORING_OP_POLL_ADD
:
8039 return io_poll_add_prep(req
, sqe
);
8040 case IORING_OP_POLL_REMOVE
:
8041 return io_poll_remove_prep(req
, sqe
);
8042 case IORING_OP_FSYNC
:
8043 return io_fsync_prep(req
, sqe
);
8044 case IORING_OP_SYNC_FILE_RANGE
:
8045 return io_sfr_prep(req
, sqe
);
8046 case IORING_OP_SENDMSG
:
8047 case IORING_OP_SEND
:
8048 return io_sendmsg_prep(req
, sqe
);
8049 case IORING_OP_RECVMSG
:
8050 case IORING_OP_RECV
:
8051 return io_recvmsg_prep(req
, sqe
);
8052 case IORING_OP_CONNECT
:
8053 return io_connect_prep(req
, sqe
);
8054 case IORING_OP_TIMEOUT
:
8055 return io_timeout_prep(req
, sqe
);
8056 case IORING_OP_TIMEOUT_REMOVE
:
8057 return io_timeout_remove_prep(req
, sqe
);
8058 case IORING_OP_ASYNC_CANCEL
:
8059 return io_async_cancel_prep(req
, sqe
);
8060 case IORING_OP_LINK_TIMEOUT
:
8061 return io_link_timeout_prep(req
, sqe
);
8062 case IORING_OP_ACCEPT
:
8063 return io_accept_prep(req
, sqe
);
8064 case IORING_OP_FALLOCATE
:
8065 return io_fallocate_prep(req
, sqe
);
8066 case IORING_OP_OPENAT
:
8067 return io_openat_prep(req
, sqe
);
8068 case IORING_OP_CLOSE
:
8069 return io_close_prep(req
, sqe
);
8070 case IORING_OP_FILES_UPDATE
:
8071 return io_files_update_prep(req
, sqe
);
8072 case IORING_OP_STATX
:
8073 return io_statx_prep(req
, sqe
);
8074 case IORING_OP_FADVISE
:
8075 return io_fadvise_prep(req
, sqe
);
8076 case IORING_OP_MADVISE
:
8077 return io_madvise_prep(req
, sqe
);
8078 case IORING_OP_OPENAT2
:
8079 return io_openat2_prep(req
, sqe
);
8080 case IORING_OP_EPOLL_CTL
:
8081 return io_epoll_ctl_prep(req
, sqe
);
8082 case IORING_OP_SPLICE
:
8083 return io_splice_prep(req
, sqe
);
8084 case IORING_OP_PROVIDE_BUFFERS
:
8085 return io_provide_buffers_prep(req
, sqe
);
8086 case IORING_OP_REMOVE_BUFFERS
:
8087 return io_remove_buffers_prep(req
, sqe
);
8089 return io_tee_prep(req
, sqe
);
8090 case IORING_OP_SHUTDOWN
:
8091 return io_shutdown_prep(req
, sqe
);
8092 case IORING_OP_RENAMEAT
:
8093 return io_renameat_prep(req
, sqe
);
8094 case IORING_OP_UNLINKAT
:
8095 return io_unlinkat_prep(req
, sqe
);
8096 case IORING_OP_MKDIRAT
:
8097 return io_mkdirat_prep(req
, sqe
);
8098 case IORING_OP_SYMLINKAT
:
8099 return io_symlinkat_prep(req
, sqe
);
8100 case IORING_OP_LINKAT
:
8101 return io_linkat_prep(req
, sqe
);
8102 case IORING_OP_MSG_RING
:
8103 return io_msg_ring_prep(req
, sqe
);
8104 case IORING_OP_FSETXATTR
:
8105 return io_fsetxattr_prep(req
, sqe
);
8106 case IORING_OP_SETXATTR
:
8107 return io_setxattr_prep(req
, sqe
);
8108 case IORING_OP_FGETXATTR
:
8109 return io_fgetxattr_prep(req
, sqe
);
8110 case IORING_OP_GETXATTR
:
8111 return io_getxattr_prep(req
, sqe
);
8112 case IORING_OP_SOCKET
:
8113 return io_socket_prep(req
, sqe
);
8114 case IORING_OP_URING_CMD
:
8115 return io_uring_cmd_prep(req
, sqe
);
8118 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
8123 static int io_req_prep_async(struct io_kiocb
*req
)
8125 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
8127 /* assign early for deferred execution for non-fixed file */
8128 if (def
->needs_file
&& !(req
->flags
& REQ_F_FIXED_FILE
))
8129 req
->file
= io_file_get_normal(req
, req
->cqe
.fd
);
8130 if (!def
->needs_async_setup
)
8132 if (WARN_ON_ONCE(req_has_async_data(req
)))
8134 if (io_alloc_async_data(req
))
8137 switch (req
->opcode
) {
8138 case IORING_OP_READV
:
8139 return io_readv_prep_async(req
);
8140 case IORING_OP_WRITEV
:
8141 return io_writev_prep_async(req
);
8142 case IORING_OP_SENDMSG
:
8143 return io_sendmsg_prep_async(req
);
8144 case IORING_OP_RECVMSG
:
8145 return io_recvmsg_prep_async(req
);
8146 case IORING_OP_CONNECT
:
8147 return io_connect_prep_async(req
);
8148 case IORING_OP_URING_CMD
:
8149 return io_uring_cmd_prep_async(req
);
8151 printk_once(KERN_WARNING
"io_uring: prep_async() bad opcode %d\n",
8156 static u32
io_get_sequence(struct io_kiocb
*req
)
8158 u32 seq
= req
->ctx
->cached_sq_head
;
8159 struct io_kiocb
*cur
;
8161 /* need original cached_sq_head, but it was increased for each req */
8162 io_for_each_link(cur
, req
)
8167 static __cold
void io_drain_req(struct io_kiocb
*req
)
8169 struct io_ring_ctx
*ctx
= req
->ctx
;
8170 struct io_defer_entry
*de
;
8172 u32 seq
= io_get_sequence(req
);
8174 /* Still need defer if there is pending req in defer list. */
8175 spin_lock(&ctx
->completion_lock
);
8176 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
)) {
8177 spin_unlock(&ctx
->completion_lock
);
8179 ctx
->drain_active
= false;
8180 io_req_task_queue(req
);
8183 spin_unlock(&ctx
->completion_lock
);
8185 ret
= io_req_prep_async(req
);
8188 io_req_complete_failed(req
, ret
);
8191 io_prep_async_link(req
);
8192 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
8198 spin_lock(&ctx
->completion_lock
);
8199 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
8200 spin_unlock(&ctx
->completion_lock
);
8205 trace_io_uring_defer(ctx
, req
, req
->cqe
.user_data
, req
->opcode
);
8208 list_add_tail(&de
->list
, &ctx
->defer_list
);
8209 spin_unlock(&ctx
->completion_lock
);
8212 static void io_clean_op(struct io_kiocb
*req
)
8214 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
8215 spin_lock(&req
->ctx
->completion_lock
);
8216 io_put_kbuf_comp(req
);
8217 spin_unlock(&req
->ctx
->completion_lock
);
8220 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
8221 switch (req
->opcode
) {
8222 case IORING_OP_READV
:
8223 case IORING_OP_READ_FIXED
:
8224 case IORING_OP_READ
:
8225 case IORING_OP_WRITEV
:
8226 case IORING_OP_WRITE_FIXED
:
8227 case IORING_OP_WRITE
: {
8228 struct io_async_rw
*io
= req
->async_data
;
8230 kfree(io
->free_iovec
);
8233 case IORING_OP_RECVMSG
:
8234 case IORING_OP_SENDMSG
: {
8235 struct io_async_msghdr
*io
= req
->async_data
;
8237 kfree(io
->free_iov
);
8240 case IORING_OP_OPENAT
:
8241 case IORING_OP_OPENAT2
:
8242 if (req
->open
.filename
)
8243 putname(req
->open
.filename
);
8245 case IORING_OP_RENAMEAT
:
8246 putname(req
->rename
.oldpath
);
8247 putname(req
->rename
.newpath
);
8249 case IORING_OP_UNLINKAT
:
8250 putname(req
->unlink
.filename
);
8252 case IORING_OP_MKDIRAT
:
8253 putname(req
->mkdir
.filename
);
8255 case IORING_OP_SYMLINKAT
:
8256 putname(req
->symlink
.oldpath
);
8257 putname(req
->symlink
.newpath
);
8259 case IORING_OP_LINKAT
:
8260 putname(req
->hardlink
.oldpath
);
8261 putname(req
->hardlink
.newpath
);
8263 case IORING_OP_STATX
:
8264 if (req
->statx
.filename
)
8265 putname(req
->statx
.filename
);
8267 case IORING_OP_SETXATTR
:
8268 case IORING_OP_FSETXATTR
:
8269 case IORING_OP_GETXATTR
:
8270 case IORING_OP_FGETXATTR
:
8271 __io_xattr_finish(req
);
8275 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
8276 kfree(req
->apoll
->double_poll
);
8280 if (req
->flags
& REQ_F_INFLIGHT
) {
8281 struct io_uring_task
*tctx
= req
->task
->io_uring
;
8283 atomic_dec(&tctx
->inflight_tracked
);
8285 if (req
->flags
& REQ_F_CREDS
)
8286 put_cred(req
->creds
);
8287 if (req
->flags
& REQ_F_ASYNC_DATA
) {
8288 kfree(req
->async_data
);
8289 req
->async_data
= NULL
;
8291 req
->flags
&= ~IO_REQ_CLEAN_FLAGS
;
8294 static bool io_assign_file(struct io_kiocb
*req
, unsigned int issue_flags
)
8296 if (req
->file
|| !io_op_defs
[req
->opcode
].needs_file
)
8299 if (req
->flags
& REQ_F_FIXED_FILE
)
8300 req
->file
= io_file_get_fixed(req
, req
->cqe
.fd
, issue_flags
);
8302 req
->file
= io_file_get_normal(req
, req
->cqe
.fd
);
8307 static int io_issue_sqe(struct io_kiocb
*req
, unsigned int issue_flags
)
8309 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
8310 const struct cred
*creds
= NULL
;
8313 if (unlikely(!io_assign_file(req
, issue_flags
)))
8316 if (unlikely((req
->flags
& REQ_F_CREDS
) && req
->creds
!= current_cred()))
8317 creds
= override_creds(req
->creds
);
8319 if (!def
->audit_skip
)
8320 audit_uring_entry(req
->opcode
);
8322 switch (req
->opcode
) {
8324 ret
= io_nop(req
, issue_flags
);
8326 case IORING_OP_READV
:
8327 case IORING_OP_READ_FIXED
:
8328 case IORING_OP_READ
:
8329 ret
= io_read(req
, issue_flags
);
8331 case IORING_OP_WRITEV
:
8332 case IORING_OP_WRITE_FIXED
:
8333 case IORING_OP_WRITE
:
8334 ret
= io_write(req
, issue_flags
);
8336 case IORING_OP_FSYNC
:
8337 ret
= io_fsync(req
, issue_flags
);
8339 case IORING_OP_POLL_ADD
:
8340 ret
= io_poll_add(req
, issue_flags
);
8342 case IORING_OP_POLL_REMOVE
:
8343 ret
= io_poll_remove(req
, issue_flags
);
8345 case IORING_OP_SYNC_FILE_RANGE
:
8346 ret
= io_sync_file_range(req
, issue_flags
);
8348 case IORING_OP_SENDMSG
:
8349 ret
= io_sendmsg(req
, issue_flags
);
8351 case IORING_OP_SEND
:
8352 ret
= io_send(req
, issue_flags
);
8354 case IORING_OP_RECVMSG
:
8355 ret
= io_recvmsg(req
, issue_flags
);
8357 case IORING_OP_RECV
:
8358 ret
= io_recv(req
, issue_flags
);
8360 case IORING_OP_TIMEOUT
:
8361 ret
= io_timeout(req
, issue_flags
);
8363 case IORING_OP_TIMEOUT_REMOVE
:
8364 ret
= io_timeout_remove(req
, issue_flags
);
8366 case IORING_OP_ACCEPT
:
8367 ret
= io_accept(req
, issue_flags
);
8369 case IORING_OP_CONNECT
:
8370 ret
= io_connect(req
, issue_flags
);
8372 case IORING_OP_ASYNC_CANCEL
:
8373 ret
= io_async_cancel(req
, issue_flags
);
8375 case IORING_OP_FALLOCATE
:
8376 ret
= io_fallocate(req
, issue_flags
);
8378 case IORING_OP_OPENAT
:
8379 ret
= io_openat(req
, issue_flags
);
8381 case IORING_OP_CLOSE
:
8382 ret
= io_close(req
, issue_flags
);
8384 case IORING_OP_FILES_UPDATE
:
8385 ret
= io_files_update(req
, issue_flags
);
8387 case IORING_OP_STATX
:
8388 ret
= io_statx(req
, issue_flags
);
8390 case IORING_OP_FADVISE
:
8391 ret
= io_fadvise(req
, issue_flags
);
8393 case IORING_OP_MADVISE
:
8394 ret
= io_madvise(req
, issue_flags
);
8396 case IORING_OP_OPENAT2
:
8397 ret
= io_openat2(req
, issue_flags
);
8399 case IORING_OP_EPOLL_CTL
:
8400 ret
= io_epoll_ctl(req
, issue_flags
);
8402 case IORING_OP_SPLICE
:
8403 ret
= io_splice(req
, issue_flags
);
8405 case IORING_OP_PROVIDE_BUFFERS
:
8406 ret
= io_provide_buffers(req
, issue_flags
);
8408 case IORING_OP_REMOVE_BUFFERS
:
8409 ret
= io_remove_buffers(req
, issue_flags
);
8412 ret
= io_tee(req
, issue_flags
);
8414 case IORING_OP_SHUTDOWN
:
8415 ret
= io_shutdown(req
, issue_flags
);
8417 case IORING_OP_RENAMEAT
:
8418 ret
= io_renameat(req
, issue_flags
);
8420 case IORING_OP_UNLINKAT
:
8421 ret
= io_unlinkat(req
, issue_flags
);
8423 case IORING_OP_MKDIRAT
:
8424 ret
= io_mkdirat(req
, issue_flags
);
8426 case IORING_OP_SYMLINKAT
:
8427 ret
= io_symlinkat(req
, issue_flags
);
8429 case IORING_OP_LINKAT
:
8430 ret
= io_linkat(req
, issue_flags
);
8432 case IORING_OP_MSG_RING
:
8433 ret
= io_msg_ring(req
, issue_flags
);
8435 case IORING_OP_FSETXATTR
:
8436 ret
= io_fsetxattr(req
, issue_flags
);
8438 case IORING_OP_SETXATTR
:
8439 ret
= io_setxattr(req
, issue_flags
);
8441 case IORING_OP_FGETXATTR
:
8442 ret
= io_fgetxattr(req
, issue_flags
);
8444 case IORING_OP_GETXATTR
:
8445 ret
= io_getxattr(req
, issue_flags
);
8447 case IORING_OP_SOCKET
:
8448 ret
= io_socket(req
, issue_flags
);
8450 case IORING_OP_URING_CMD
:
8451 ret
= io_uring_cmd(req
, issue_flags
);
8458 if (!def
->audit_skip
)
8459 audit_uring_exit(!ret
, ret
);
8462 revert_creds(creds
);
8465 /* If the op doesn't have a file, we're not polling for it */
8466 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
)
8467 io_iopoll_req_issued(req
, issue_flags
);
8472 static struct io_wq_work
*io_wq_free_work(struct io_wq_work
*work
)
8474 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8476 req
= io_put_req_find_next(req
);
8477 return req
? &req
->work
: NULL
;
8480 static void io_wq_submit_work(struct io_wq_work
*work
)
8482 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8483 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
8484 unsigned int issue_flags
= IO_URING_F_UNLOCKED
;
8485 bool needs_poll
= false;
8486 int ret
= 0, err
= -ECANCELED
;
8488 /* one will be dropped by ->io_free_work() after returning to io-wq */
8489 if (!(req
->flags
& REQ_F_REFCOUNT
))
8490 __io_req_set_refcount(req
, 2);
8494 io_arm_ltimeout(req
);
8496 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
8497 if (work
->flags
& IO_WQ_WORK_CANCEL
) {
8499 io_req_task_queue_fail(req
, err
);
8502 if (!io_assign_file(req
, issue_flags
)) {
8504 work
->flags
|= IO_WQ_WORK_CANCEL
;
8508 if (req
->flags
& REQ_F_FORCE_ASYNC
) {
8509 bool opcode_poll
= def
->pollin
|| def
->pollout
;
8511 if (opcode_poll
&& file_can_poll(req
->file
)) {
8513 issue_flags
|= IO_URING_F_NONBLOCK
;
8518 ret
= io_issue_sqe(req
, issue_flags
);
8522 * We can get EAGAIN for iopolled IO even though we're
8523 * forcing a sync submission from here, since we can't
8524 * wait for request slots on the block side.
8527 if (!(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
8533 if (io_arm_poll_handler(req
, issue_flags
) == IO_APOLL_OK
)
8535 /* aborted or ready, in either case retry blocking */
8537 issue_flags
&= ~IO_URING_F_NONBLOCK
;
8540 /* avoid locking problems by failing it from a clean context */
8542 io_req_task_queue_fail(req
, ret
);
8545 static inline struct io_fixed_file
*io_fixed_file_slot(struct io_file_table
*table
,
8548 return &table
->files
[i
];
8551 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
8554 struct io_fixed_file
*slot
= io_fixed_file_slot(&ctx
->file_table
, index
);
8556 return (struct file
*) (slot
->file_ptr
& FFS_MASK
);
8559 static void io_fixed_file_set(struct io_fixed_file
*file_slot
, struct file
*file
)
8561 unsigned long file_ptr
= (unsigned long) file
;
8563 file_ptr
|= io_file_get_flags(file
);
8564 file_slot
->file_ptr
= file_ptr
;
8567 static inline struct file
*io_file_get_fixed(struct io_kiocb
*req
, int fd
,
8568 unsigned int issue_flags
)
8570 struct io_ring_ctx
*ctx
= req
->ctx
;
8571 struct file
*file
= NULL
;
8572 unsigned long file_ptr
;
8574 io_ring_submit_lock(ctx
, issue_flags
);
8576 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
8578 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
8579 file_ptr
= io_fixed_file_slot(&ctx
->file_table
, fd
)->file_ptr
;
8580 file
= (struct file
*) (file_ptr
& FFS_MASK
);
8581 file_ptr
&= ~FFS_MASK
;
8582 /* mask in overlapping REQ_F and FFS bits */
8583 req
->flags
|= (file_ptr
<< REQ_F_SUPPORT_NOWAIT_BIT
);
8584 io_req_set_rsrc_node(req
, ctx
, 0);
8585 WARN_ON_ONCE(file
&& !test_bit(fd
, ctx
->file_table
.bitmap
));
8587 io_ring_submit_unlock(ctx
, issue_flags
);
8591 static struct file
*io_file_get_normal(struct io_kiocb
*req
, int fd
)
8593 struct file
*file
= fget(fd
);
8595 trace_io_uring_file_get(req
->ctx
, req
, req
->cqe
.user_data
, fd
);
8597 /* we don't allow fixed io_uring files */
8598 if (file
&& file
->f_op
== &io_uring_fops
)
8599 io_req_track_inflight(req
);
8603 static void io_req_task_link_timeout(struct io_kiocb
*req
, bool *locked
)
8605 struct io_kiocb
*prev
= req
->timeout
.prev
;
8609 if (!(req
->task
->flags
& PF_EXITING
)) {
8610 struct io_cancel_data cd
= {
8612 .data
= prev
->cqe
.user_data
,
8615 ret
= io_try_cancel(req
, &cd
);
8617 io_req_complete_post(req
, ret
?: -ETIME
, 0);
8620 io_req_complete_post(req
, -ETIME
, 0);
8624 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
8626 struct io_timeout_data
*data
= container_of(timer
,
8627 struct io_timeout_data
, timer
);
8628 struct io_kiocb
*prev
, *req
= data
->req
;
8629 struct io_ring_ctx
*ctx
= req
->ctx
;
8630 unsigned long flags
;
8632 spin_lock_irqsave(&ctx
->timeout_lock
, flags
);
8633 prev
= req
->timeout
.head
;
8634 req
->timeout
.head
= NULL
;
8637 * We don't expect the list to be empty, that will only happen if we
8638 * race with the completion of the linked work.
8641 io_remove_next_linked(prev
);
8642 if (!req_ref_inc_not_zero(prev
))
8645 list_del(&req
->timeout
.list
);
8646 req
->timeout
.prev
= prev
;
8647 spin_unlock_irqrestore(&ctx
->timeout_lock
, flags
);
8649 req
->io_task_work
.func
= io_req_task_link_timeout
;
8650 io_req_task_work_add(req
);
8651 return HRTIMER_NORESTART
;
8654 static void io_queue_linked_timeout(struct io_kiocb
*req
)
8656 struct io_ring_ctx
*ctx
= req
->ctx
;
8658 spin_lock_irq(&ctx
->timeout_lock
);
8660 * If the back reference is NULL, then our linked request finished
8661 * before we got a chance to setup the timer
8663 if (req
->timeout
.head
) {
8664 struct io_timeout_data
*data
= req
->async_data
;
8666 data
->timer
.function
= io_link_timeout_fn
;
8667 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
8669 list_add_tail(&req
->timeout
.list
, &ctx
->ltimeout_list
);
8671 spin_unlock_irq(&ctx
->timeout_lock
);
8672 /* drop submission reference */
8676 static void io_queue_async(struct io_kiocb
*req
, int ret
)
8677 __must_hold(&req
->ctx
->uring_lock
)
8679 struct io_kiocb
*linked_timeout
;
8681 if (ret
!= -EAGAIN
|| (req
->flags
& REQ_F_NOWAIT
)) {
8682 io_req_complete_failed(req
, ret
);
8686 linked_timeout
= io_prep_linked_timeout(req
);
8688 switch (io_arm_poll_handler(req
, 0)) {
8689 case IO_APOLL_READY
:
8690 io_req_task_queue(req
);
8692 case IO_APOLL_ABORTED
:
8694 * Queued up for async execution, worker will release
8695 * submit reference when the iocb is actually submitted.
8697 io_kbuf_recycle(req
, 0);
8698 io_queue_iowq(req
, NULL
);
8705 io_queue_linked_timeout(linked_timeout
);
8708 static inline void io_queue_sqe(struct io_kiocb
*req
)
8709 __must_hold(&req
->ctx
->uring_lock
)
8713 ret
= io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_COMPLETE_DEFER
);
8715 if (req
->flags
& REQ_F_COMPLETE_INLINE
) {
8716 io_req_add_compl_list(req
);
8720 * We async punt it if the file wasn't marked NOWAIT, or if the file
8721 * doesn't support non-blocking read/write attempts
8724 io_arm_ltimeout(req
);
8726 io_queue_async(req
, ret
);
8729 static void io_queue_sqe_fallback(struct io_kiocb
*req
)
8730 __must_hold(&req
->ctx
->uring_lock
)
8732 if (unlikely(req
->flags
& REQ_F_FAIL
)) {
8734 * We don't submit, fail them all, for that replace hardlinks
8735 * with normal links. Extra REQ_F_LINK is tolerated.
8737 req
->flags
&= ~REQ_F_HARDLINK
;
8738 req
->flags
|= REQ_F_LINK
;
8739 io_req_complete_failed(req
, req
->cqe
.res
);
8740 } else if (unlikely(req
->ctx
->drain_active
)) {
8743 int ret
= io_req_prep_async(req
);
8746 io_req_complete_failed(req
, ret
);
8748 io_queue_iowq(req
, NULL
);
8753 * Check SQE restrictions (opcode and flags).
8755 * Returns 'true' if SQE is allowed, 'false' otherwise.
8757 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
8758 struct io_kiocb
*req
,
8759 unsigned int sqe_flags
)
8761 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
8764 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
8765 ctx
->restrictions
.sqe_flags_required
)
8768 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
8769 ctx
->restrictions
.sqe_flags_required
))
8775 static void io_init_req_drain(struct io_kiocb
*req
)
8777 struct io_ring_ctx
*ctx
= req
->ctx
;
8778 struct io_kiocb
*head
= ctx
->submit_state
.link
.head
;
8780 ctx
->drain_active
= true;
8783 * If we need to drain a request in the middle of a link, drain
8784 * the head request and the next request/link after the current
8785 * link. Considering sequential execution of links,
8786 * REQ_F_IO_DRAIN will be maintained for every request of our
8789 head
->flags
|= REQ_F_IO_DRAIN
| REQ_F_FORCE_ASYNC
;
8790 ctx
->drain_next
= true;
8794 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
8795 const struct io_uring_sqe
*sqe
)
8796 __must_hold(&ctx
->uring_lock
)
8798 const struct io_op_def
*def
;
8799 unsigned int sqe_flags
;
8803 /* req is partially pre-initialised, see io_preinit_req() */
8804 req
->opcode
= opcode
= READ_ONCE(sqe
->opcode
);
8805 /* same numerical values with corresponding REQ_F_*, safe to copy */
8806 req
->flags
= sqe_flags
= READ_ONCE(sqe
->flags
);
8807 req
->cqe
.user_data
= READ_ONCE(sqe
->user_data
);
8809 req
->rsrc_node
= NULL
;
8810 req
->task
= current
;
8812 if (unlikely(opcode
>= IORING_OP_LAST
)) {
8816 def
= &io_op_defs
[opcode
];
8817 if (unlikely(sqe_flags
& ~SQE_COMMON_FLAGS
)) {
8818 /* enforce forwards compatibility on users */
8819 if (sqe_flags
& ~SQE_VALID_FLAGS
)
8821 if (sqe_flags
& IOSQE_BUFFER_SELECT
) {
8822 if (!def
->buffer_select
)
8824 req
->buf_index
= READ_ONCE(sqe
->buf_group
);
8826 if (sqe_flags
& IOSQE_CQE_SKIP_SUCCESS
)
8827 ctx
->drain_disabled
= true;
8828 if (sqe_flags
& IOSQE_IO_DRAIN
) {
8829 if (ctx
->drain_disabled
)
8831 io_init_req_drain(req
);
8834 if (unlikely(ctx
->restricted
|| ctx
->drain_active
|| ctx
->drain_next
)) {
8835 if (ctx
->restricted
&& !io_check_restriction(ctx
, req
, sqe_flags
))
8837 /* knock it to the slow queue path, will be drained there */
8838 if (ctx
->drain_active
)
8839 req
->flags
|= REQ_F_FORCE_ASYNC
;
8840 /* if there is no link, we're at "next" request and need to drain */
8841 if (unlikely(ctx
->drain_next
) && !ctx
->submit_state
.link
.head
) {
8842 ctx
->drain_next
= false;
8843 ctx
->drain_active
= true;
8844 req
->flags
|= REQ_F_IO_DRAIN
| REQ_F_FORCE_ASYNC
;
8848 if (!def
->ioprio
&& sqe
->ioprio
)
8850 if (!def
->iopoll
&& (ctx
->flags
& IORING_SETUP_IOPOLL
))
8853 if (def
->needs_file
) {
8854 struct io_submit_state
*state
= &ctx
->submit_state
;
8856 req
->cqe
.fd
= READ_ONCE(sqe
->fd
);
8859 * Plug now if we have more than 2 IO left after this, and the
8860 * target is potentially a read/write to block based storage.
8862 if (state
->need_plug
&& def
->plug
) {
8863 state
->plug_started
= true;
8864 state
->need_plug
= false;
8865 blk_start_plug_nr_ios(&state
->plug
, state
->submit_nr
);
8869 personality
= READ_ONCE(sqe
->personality
);
8873 req
->creds
= xa_load(&ctx
->personalities
, personality
);
8876 get_cred(req
->creds
);
8877 ret
= security_uring_override_creds(req
->creds
);
8879 put_cred(req
->creds
);
8882 req
->flags
|= REQ_F_CREDS
;
8885 return io_req_prep(req
, sqe
);
8888 static __cold
int io_submit_fail_init(const struct io_uring_sqe
*sqe
,
8889 struct io_kiocb
*req
, int ret
)
8891 struct io_ring_ctx
*ctx
= req
->ctx
;
8892 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
8893 struct io_kiocb
*head
= link
->head
;
8895 trace_io_uring_req_failed(sqe
, ctx
, req
, ret
);
8898 * Avoid breaking links in the middle as it renders links with SQPOLL
8899 * unusable. Instead of failing eagerly, continue assembling the link if
8900 * applicable and mark the head with REQ_F_FAIL. The link flushing code
8901 * should find the flag and handle the rest.
8903 req_fail_link_node(req
, ret
);
8904 if (head
&& !(head
->flags
& REQ_F_FAIL
))
8905 req_fail_link_node(head
, -ECANCELED
);
8907 if (!(req
->flags
& IO_REQ_LINK_FLAGS
)) {
8909 link
->last
->link
= req
;
8913 io_queue_sqe_fallback(req
);
8918 link
->last
->link
= req
;
8925 static inline int io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
8926 const struct io_uring_sqe
*sqe
)
8927 __must_hold(&ctx
->uring_lock
)
8929 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
8932 ret
= io_init_req(ctx
, req
, sqe
);
8934 return io_submit_fail_init(sqe
, req
, ret
);
8936 /* don't need @sqe from now on */
8937 trace_io_uring_submit_sqe(ctx
, req
, req
->cqe
.user_data
, req
->opcode
,
8939 ctx
->flags
& IORING_SETUP_SQPOLL
);
8942 * If we already have a head request, queue this one for async
8943 * submittal once the head completes. If we don't have a head but
8944 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
8945 * submitted sync once the chain is complete. If none of those
8946 * conditions are true (normal request), then just queue it.
8948 if (unlikely(link
->head
)) {
8949 ret
= io_req_prep_async(req
);
8951 return io_submit_fail_init(sqe
, req
, ret
);
8953 trace_io_uring_link(ctx
, req
, link
->head
);
8954 link
->last
->link
= req
;
8957 if (req
->flags
& IO_REQ_LINK_FLAGS
)
8959 /* last request of the link, flush it */
8962 if (req
->flags
& (REQ_F_FORCE_ASYNC
| REQ_F_FAIL
))
8965 } else if (unlikely(req
->flags
& (IO_REQ_LINK_FLAGS
|
8966 REQ_F_FORCE_ASYNC
| REQ_F_FAIL
))) {
8967 if (req
->flags
& IO_REQ_LINK_FLAGS
) {
8972 io_queue_sqe_fallback(req
);
8982 * Batched submission is done, ensure local IO is flushed out.
8984 static void io_submit_state_end(struct io_ring_ctx
*ctx
)
8986 struct io_submit_state
*state
= &ctx
->submit_state
;
8988 if (unlikely(state
->link
.head
))
8989 io_queue_sqe_fallback(state
->link
.head
);
8990 /* flush only after queuing links as they can generate completions */
8991 io_submit_flush_completions(ctx
);
8992 if (state
->plug_started
)
8993 blk_finish_plug(&state
->plug
);
8997 * Start submission side cache.
8999 static void io_submit_state_start(struct io_submit_state
*state
,
9000 unsigned int max_ios
)
9002 state
->plug_started
= false;
9003 state
->need_plug
= max_ios
> 2;
9004 state
->submit_nr
= max_ios
;
9005 /* set only head, no need to init link_last in advance */
9006 state
->link
.head
= NULL
;
9009 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
9011 struct io_rings
*rings
= ctx
->rings
;
9014 * Ensure any loads from the SQEs are done at this point,
9015 * since once we write the new head, the application could
9016 * write new data to them.
9018 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
9022 * Fetch an sqe, if one is available. Note this returns a pointer to memory
9023 * that is mapped by userspace. This means that care needs to be taken to
9024 * ensure that reads are stable, as we cannot rely on userspace always
9025 * being a good citizen. If members of the sqe are validated and then later
9026 * used, it's important that those reads are done through READ_ONCE() to
9027 * prevent a re-load down the line.
9029 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
9031 unsigned head
, mask
= ctx
->sq_entries
- 1;
9032 unsigned sq_idx
= ctx
->cached_sq_head
++ & mask
;
9035 * The cached sq head (or cq tail) serves two purposes:
9037 * 1) allows us to batch the cost of updating the user visible
9039 * 2) allows the kernel side to track the head on its own, even
9040 * though the application is the one updating it.
9042 head
= READ_ONCE(ctx
->sq_array
[sq_idx
]);
9043 if (likely(head
< ctx
->sq_entries
)) {
9044 /* double index for 128-byte SQEs, twice as long */
9045 if (ctx
->flags
& IORING_SETUP_SQE128
)
9047 return &ctx
->sq_sqes
[head
];
9050 /* drop invalid entries */
9052 WRITE_ONCE(ctx
->rings
->sq_dropped
,
9053 READ_ONCE(ctx
->rings
->sq_dropped
) + 1);
9057 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
9058 __must_hold(&ctx
->uring_lock
)
9060 unsigned int entries
= io_sqring_entries(ctx
);
9064 if (unlikely(!entries
))
9066 /* make sure SQ entry isn't read before tail */
9067 ret
= left
= min3(nr
, ctx
->sq_entries
, entries
);
9068 io_get_task_refs(left
);
9069 io_submit_state_start(&ctx
->submit_state
, left
);
9072 const struct io_uring_sqe
*sqe
;
9073 struct io_kiocb
*req
;
9075 if (unlikely(!io_alloc_req_refill(ctx
)))
9077 req
= io_alloc_req(ctx
);
9078 sqe
= io_get_sqe(ctx
);
9079 if (unlikely(!sqe
)) {
9080 io_req_add_to_cache(req
, ctx
);
9085 * Continue submitting even for sqe failure if the
9086 * ring was setup with IORING_SETUP_SUBMIT_ALL
9088 if (unlikely(io_submit_sqe(ctx
, req
, sqe
)) &&
9089 !(ctx
->flags
& IORING_SETUP_SUBMIT_ALL
)) {
9095 if (unlikely(left
)) {
9097 /* try again if it submitted nothing and can't allocate a req */
9098 if (!ret
&& io_req_cache_empty(ctx
))
9100 current
->io_uring
->cached_refs
+= left
;
9103 io_submit_state_end(ctx
);
9104 /* Commit SQ ring head once we've consumed and submitted all SQEs */
9105 io_commit_sqring(ctx
);
9109 static inline bool io_sqd_events_pending(struct io_sq_data
*sqd
)
9111 return READ_ONCE(sqd
->state
);
9114 static int __io_sq_thread(struct io_ring_ctx
*ctx
, bool cap_entries
)
9116 unsigned int to_submit
;
9119 to_submit
= io_sqring_entries(ctx
);
9120 /* if we're handling multiple rings, cap submit size for fairness */
9121 if (cap_entries
&& to_submit
> IORING_SQPOLL_CAP_ENTRIES_VALUE
)
9122 to_submit
= IORING_SQPOLL_CAP_ENTRIES_VALUE
;
9124 if (!wq_list_empty(&ctx
->iopoll_list
) || to_submit
) {
9125 const struct cred
*creds
= NULL
;
9127 if (ctx
->sq_creds
!= current_cred())
9128 creds
= override_creds(ctx
->sq_creds
);
9130 mutex_lock(&ctx
->uring_lock
);
9131 if (!wq_list_empty(&ctx
->iopoll_list
))
9132 io_do_iopoll(ctx
, true);
9135 * Don't submit if refs are dying, good for io_uring_register(),
9136 * but also it is relied upon by io_ring_exit_work()
9138 if (to_submit
&& likely(!percpu_ref_is_dying(&ctx
->refs
)) &&
9139 !(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9140 ret
= io_submit_sqes(ctx
, to_submit
);
9141 mutex_unlock(&ctx
->uring_lock
);
9143 if (to_submit
&& wq_has_sleeper(&ctx
->sqo_sq_wait
))
9144 wake_up(&ctx
->sqo_sq_wait
);
9146 revert_creds(creds
);
9152 static __cold
void io_sqd_update_thread_idle(struct io_sq_data
*sqd
)
9154 struct io_ring_ctx
*ctx
;
9155 unsigned sq_thread_idle
= 0;
9157 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
9158 sq_thread_idle
= max(sq_thread_idle
, ctx
->sq_thread_idle
);
9159 sqd
->sq_thread_idle
= sq_thread_idle
;
9162 static bool io_sqd_handle_event(struct io_sq_data
*sqd
)
9164 bool did_sig
= false;
9165 struct ksignal ksig
;
9167 if (test_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
) ||
9168 signal_pending(current
)) {
9169 mutex_unlock(&sqd
->lock
);
9170 if (signal_pending(current
))
9171 did_sig
= get_signal(&ksig
);
9173 mutex_lock(&sqd
->lock
);
9175 return did_sig
|| test_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
);
9178 static int io_sq_thread(void *data
)
9180 struct io_sq_data
*sqd
= data
;
9181 struct io_ring_ctx
*ctx
;
9182 unsigned long timeout
= 0;
9183 char buf
[TASK_COMM_LEN
];
9186 snprintf(buf
, sizeof(buf
), "iou-sqp-%d", sqd
->task_pid
);
9187 set_task_comm(current
, buf
);
9189 if (sqd
->sq_cpu
!= -1)
9190 set_cpus_allowed_ptr(current
, cpumask_of(sqd
->sq_cpu
));
9192 set_cpus_allowed_ptr(current
, cpu_online_mask
);
9193 current
->flags
|= PF_NO_SETAFFINITY
;
9195 audit_alloc_kernel(current
);
9197 mutex_lock(&sqd
->lock
);
9199 bool cap_entries
, sqt_spin
= false;
9201 if (io_sqd_events_pending(sqd
) || signal_pending(current
)) {
9202 if (io_sqd_handle_event(sqd
))
9204 timeout
= jiffies
+ sqd
->sq_thread_idle
;
9207 cap_entries
= !list_is_singular(&sqd
->ctx_list
);
9208 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
9209 int ret
= __io_sq_thread(ctx
, cap_entries
);
9211 if (!sqt_spin
&& (ret
> 0 || !wq_list_empty(&ctx
->iopoll_list
)))
9214 if (io_run_task_work())
9217 if (sqt_spin
|| !time_after(jiffies
, timeout
)) {
9220 timeout
= jiffies
+ sqd
->sq_thread_idle
;
9224 prepare_to_wait(&sqd
->wait
, &wait
, TASK_INTERRUPTIBLE
);
9225 if (!io_sqd_events_pending(sqd
) && !task_work_pending(current
)) {
9226 bool needs_sched
= true;
9228 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
9229 atomic_or(IORING_SQ_NEED_WAKEUP
,
9230 &ctx
->rings
->sq_flags
);
9231 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
9232 !wq_list_empty(&ctx
->iopoll_list
)) {
9233 needs_sched
= false;
9238 * Ensure the store of the wakeup flag is not
9239 * reordered with the load of the SQ tail
9241 smp_mb__after_atomic();
9243 if (io_sqring_entries(ctx
)) {
9244 needs_sched
= false;
9250 mutex_unlock(&sqd
->lock
);
9252 mutex_lock(&sqd
->lock
);
9254 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
9255 atomic_andnot(IORING_SQ_NEED_WAKEUP
,
9256 &ctx
->rings
->sq_flags
);
9259 finish_wait(&sqd
->wait
, &wait
);
9260 timeout
= jiffies
+ sqd
->sq_thread_idle
;
9263 io_uring_cancel_generic(true, sqd
);
9265 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
9266 atomic_or(IORING_SQ_NEED_WAKEUP
, &ctx
->rings
->sq_flags
);
9268 mutex_unlock(&sqd
->lock
);
9270 audit_free(current
);
9272 complete(&sqd
->exited
);
9276 struct io_wait_queue
{
9277 struct wait_queue_entry wq
;
9278 struct io_ring_ctx
*ctx
;
9280 unsigned nr_timeouts
;
9283 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
9285 struct io_ring_ctx
*ctx
= iowq
->ctx
;
9286 int dist
= ctx
->cached_cq_tail
- (int) iowq
->cq_tail
;
9289 * Wake up if we have enough events, or if a timeout occurred since we
9290 * started waiting. For timeouts, we always want to return to userspace,
9291 * regardless of event count.
9293 return dist
>= 0 || atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
9296 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
9297 int wake_flags
, void *key
)
9299 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
9303 * Cannot safely flush overflowed CQEs from here, ensure we wake up
9304 * the task, and the next invocation will do it.
9306 if (io_should_wake(iowq
) ||
9307 test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &iowq
->ctx
->check_cq
))
9308 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
9312 static int io_run_task_work_sig(void)
9314 if (io_run_task_work())
9316 if (test_thread_flag(TIF_NOTIFY_SIGNAL
))
9317 return -ERESTARTSYS
;
9318 if (task_sigpending(current
))
9323 /* when returns >0, the caller should retry */
9324 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
9325 struct io_wait_queue
*iowq
,
9329 unsigned long check_cq
;
9331 /* make sure we run task_work before checking for signals */
9332 ret
= io_run_task_work_sig();
9333 if (ret
|| io_should_wake(iowq
))
9335 check_cq
= READ_ONCE(ctx
->check_cq
);
9336 /* let the caller flush overflows, retry */
9337 if (check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
))
9339 if (unlikely(check_cq
& BIT(IO_CHECK_CQ_DROPPED_BIT
)))
9341 if (!schedule_hrtimeout(&timeout
, HRTIMER_MODE_ABS
))
9347 * Wait until events become available, if we don't already have some. The
9348 * application must reap them itself, as they reside on the shared cq ring.
9350 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
9351 const sigset_t __user
*sig
, size_t sigsz
,
9352 struct __kernel_timespec __user
*uts
)
9354 struct io_wait_queue iowq
;
9355 struct io_rings
*rings
= ctx
->rings
;
9356 ktime_t timeout
= KTIME_MAX
;
9360 io_cqring_overflow_flush(ctx
);
9361 if (io_cqring_events(ctx
) >= min_events
)
9363 if (!io_run_task_work())
9368 #ifdef CONFIG_COMPAT
9369 if (in_compat_syscall())
9370 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
9374 ret
= set_user_sigmask(sig
, sigsz
);
9381 struct timespec64 ts
;
9383 if (get_timespec64(&ts
, uts
))
9385 timeout
= ktime_add_ns(timespec64_to_ktime(ts
), ktime_get_ns());
9388 init_waitqueue_func_entry(&iowq
.wq
, io_wake_function
);
9389 iowq
.wq
.private = current
;
9390 INIT_LIST_HEAD(&iowq
.wq
.entry
);
9392 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
9393 iowq
.cq_tail
= READ_ONCE(ctx
->rings
->cq
.head
) + min_events
;
9395 trace_io_uring_cqring_wait(ctx
, min_events
);
9397 /* if we can't even flush overflow, don't wait for more */
9398 if (!io_cqring_overflow_flush(ctx
)) {
9402 prepare_to_wait_exclusive(&ctx
->cq_wait
, &iowq
.wq
,
9403 TASK_INTERRUPTIBLE
);
9404 ret
= io_cqring_wait_schedule(ctx
, &iowq
, timeout
);
9408 finish_wait(&ctx
->cq_wait
, &iowq
.wq
);
9409 restore_saved_sigmask_unless(ret
== -EINTR
);
9411 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
9414 static void io_free_page_table(void **table
, size_t size
)
9416 unsigned i
, nr_tables
= DIV_ROUND_UP(size
, PAGE_SIZE
);
9418 for (i
= 0; i
< nr_tables
; i
++)
9423 static __cold
void **io_alloc_page_table(size_t size
)
9425 unsigned i
, nr_tables
= DIV_ROUND_UP(size
, PAGE_SIZE
);
9426 size_t init_size
= size
;
9429 table
= kcalloc(nr_tables
, sizeof(*table
), GFP_KERNEL_ACCOUNT
);
9433 for (i
= 0; i
< nr_tables
; i
++) {
9434 unsigned int this_size
= min_t(size_t, size
, PAGE_SIZE
);
9436 table
[i
] = kzalloc(this_size
, GFP_KERNEL_ACCOUNT
);
9438 io_free_page_table(table
, init_size
);
9446 static void io_rsrc_node_destroy(struct io_rsrc_node
*ref_node
)
9448 percpu_ref_exit(&ref_node
->refs
);
9452 static __cold
void io_rsrc_node_ref_zero(struct percpu_ref
*ref
)
9454 struct io_rsrc_node
*node
= container_of(ref
, struct io_rsrc_node
, refs
);
9455 struct io_ring_ctx
*ctx
= node
->rsrc_data
->ctx
;
9456 unsigned long flags
;
9457 bool first_add
= false;
9458 unsigned long delay
= HZ
;
9460 spin_lock_irqsave(&ctx
->rsrc_ref_lock
, flags
);
9463 /* if we are mid-quiesce then do not delay */
9464 if (node
->rsrc_data
->quiesce
)
9467 while (!list_empty(&ctx
->rsrc_ref_list
)) {
9468 node
= list_first_entry(&ctx
->rsrc_ref_list
,
9469 struct io_rsrc_node
, node
);
9470 /* recycle ref nodes in order */
9473 list_del(&node
->node
);
9474 first_add
|= llist_add(&node
->llist
, &ctx
->rsrc_put_llist
);
9476 spin_unlock_irqrestore(&ctx
->rsrc_ref_lock
, flags
);
9479 mod_delayed_work(system_wq
, &ctx
->rsrc_put_work
, delay
);
9482 static struct io_rsrc_node
*io_rsrc_node_alloc(void)
9484 struct io_rsrc_node
*ref_node
;
9486 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
9490 if (percpu_ref_init(&ref_node
->refs
, io_rsrc_node_ref_zero
,
9495 INIT_LIST_HEAD(&ref_node
->node
);
9496 INIT_LIST_HEAD(&ref_node
->rsrc_list
);
9497 ref_node
->done
= false;
9501 static void io_rsrc_node_switch(struct io_ring_ctx
*ctx
,
9502 struct io_rsrc_data
*data_to_kill
)
9503 __must_hold(&ctx
->uring_lock
)
9505 WARN_ON_ONCE(!ctx
->rsrc_backup_node
);
9506 WARN_ON_ONCE(data_to_kill
&& !ctx
->rsrc_node
);
9508 io_rsrc_refs_drop(ctx
);
9511 struct io_rsrc_node
*rsrc_node
= ctx
->rsrc_node
;
9513 rsrc_node
->rsrc_data
= data_to_kill
;
9514 spin_lock_irq(&ctx
->rsrc_ref_lock
);
9515 list_add_tail(&rsrc_node
->node
, &ctx
->rsrc_ref_list
);
9516 spin_unlock_irq(&ctx
->rsrc_ref_lock
);
9518 atomic_inc(&data_to_kill
->refs
);
9519 percpu_ref_kill(&rsrc_node
->refs
);
9520 ctx
->rsrc_node
= NULL
;
9523 if (!ctx
->rsrc_node
) {
9524 ctx
->rsrc_node
= ctx
->rsrc_backup_node
;
9525 ctx
->rsrc_backup_node
= NULL
;
9529 static int io_rsrc_node_switch_start(struct io_ring_ctx
*ctx
)
9531 if (ctx
->rsrc_backup_node
)
9533 ctx
->rsrc_backup_node
= io_rsrc_node_alloc();
9534 return ctx
->rsrc_backup_node
? 0 : -ENOMEM
;
9537 static __cold
int io_rsrc_ref_quiesce(struct io_rsrc_data
*data
,
9538 struct io_ring_ctx
*ctx
)
9542 /* As we may drop ->uring_lock, other task may have started quiesce */
9546 data
->quiesce
= true;
9548 ret
= io_rsrc_node_switch_start(ctx
);
9551 io_rsrc_node_switch(ctx
, data
);
9553 /* kill initial ref, already quiesced if zero */
9554 if (atomic_dec_and_test(&data
->refs
))
9556 mutex_unlock(&ctx
->uring_lock
);
9557 flush_delayed_work(&ctx
->rsrc_put_work
);
9558 ret
= wait_for_completion_interruptible(&data
->done
);
9560 mutex_lock(&ctx
->uring_lock
);
9561 if (atomic_read(&data
->refs
) > 0) {
9563 * it has been revived by another thread while
9566 mutex_unlock(&ctx
->uring_lock
);
9572 atomic_inc(&data
->refs
);
9573 /* wait for all works potentially completing data->done */
9574 flush_delayed_work(&ctx
->rsrc_put_work
);
9575 reinit_completion(&data
->done
);
9577 ret
= io_run_task_work_sig();
9578 mutex_lock(&ctx
->uring_lock
);
9580 data
->quiesce
= false;
9585 static u64
*io_get_tag_slot(struct io_rsrc_data
*data
, unsigned int idx
)
9587 unsigned int off
= idx
& IO_RSRC_TAG_TABLE_MASK
;
9588 unsigned int table_idx
= idx
>> IO_RSRC_TAG_TABLE_SHIFT
;
9590 return &data
->tags
[table_idx
][off
];
9593 static void io_rsrc_data_free(struct io_rsrc_data
*data
)
9595 size_t size
= data
->nr
* sizeof(data
->tags
[0][0]);
9598 io_free_page_table((void **)data
->tags
, size
);
9602 static __cold
int io_rsrc_data_alloc(struct io_ring_ctx
*ctx
, rsrc_put_fn
*do_put
,
9603 u64 __user
*utags
, unsigned nr
,
9604 struct io_rsrc_data
**pdata
)
9606 struct io_rsrc_data
*data
;
9610 data
= kzalloc(sizeof(*data
), GFP_KERNEL
);
9613 data
->tags
= (u64
**)io_alloc_page_table(nr
* sizeof(data
->tags
[0][0]));
9621 data
->do_put
= do_put
;
9624 for (i
= 0; i
< nr
; i
++) {
9625 u64
*tag_slot
= io_get_tag_slot(data
, i
);
9627 if (copy_from_user(tag_slot
, &utags
[i
],
9633 atomic_set(&data
->refs
, 1);
9634 init_completion(&data
->done
);
9638 io_rsrc_data_free(data
);
9642 static bool io_alloc_file_tables(struct io_file_table
*table
, unsigned nr_files
)
9644 table
->files
= kvcalloc(nr_files
, sizeof(table
->files
[0]),
9645 GFP_KERNEL_ACCOUNT
);
9646 if (unlikely(!table
->files
))
9649 table
->bitmap
= bitmap_zalloc(nr_files
, GFP_KERNEL_ACCOUNT
);
9650 if (unlikely(!table
->bitmap
)) {
9651 kvfree(table
->files
);
9658 static void io_free_file_tables(struct io_file_table
*table
)
9660 kvfree(table
->files
);
9661 bitmap_free(table
->bitmap
);
9662 table
->files
= NULL
;
9663 table
->bitmap
= NULL
;
9666 static inline void io_file_bitmap_set(struct io_file_table
*table
, int bit
)
9668 WARN_ON_ONCE(test_bit(bit
, table
->bitmap
));
9669 __set_bit(bit
, table
->bitmap
);
9670 table
->alloc_hint
= bit
+ 1;
9673 static inline void io_file_bitmap_clear(struct io_file_table
*table
, int bit
)
9675 __clear_bit(bit
, table
->bitmap
);
9676 table
->alloc_hint
= bit
;
9679 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
9681 #if !defined(IO_URING_SCM_ALL)
9684 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
9685 struct file
*file
= io_file_from_index(ctx
, i
);
9689 if (io_fixed_file_slot(&ctx
->file_table
, i
)->file_ptr
& FFS_SCM
)
9691 io_file_bitmap_clear(&ctx
->file_table
, i
);
9696 #if defined(CONFIG_UNIX)
9697 if (ctx
->ring_sock
) {
9698 struct sock
*sock
= ctx
->ring_sock
->sk
;
9699 struct sk_buff
*skb
;
9701 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
9705 io_free_file_tables(&ctx
->file_table
);
9706 io_rsrc_data_free(ctx
->file_data
);
9707 ctx
->file_data
= NULL
;
9708 ctx
->nr_user_files
= 0;
9711 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
9713 unsigned nr
= ctx
->nr_user_files
;
9716 if (!ctx
->file_data
)
9720 * Quiesce may unlock ->uring_lock, and while it's not held
9721 * prevent new requests using the table.
9723 ctx
->nr_user_files
= 0;
9724 ret
= io_rsrc_ref_quiesce(ctx
->file_data
, ctx
);
9725 ctx
->nr_user_files
= nr
;
9727 __io_sqe_files_unregister(ctx
);
9731 static void io_sq_thread_unpark(struct io_sq_data
*sqd
)
9732 __releases(&sqd
->lock
)
9734 WARN_ON_ONCE(sqd
->thread
== current
);
9737 * Do the dance but not conditional clear_bit() because it'd race with
9738 * other threads incrementing park_pending and setting the bit.
9740 clear_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
9741 if (atomic_dec_return(&sqd
->park_pending
))
9742 set_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
9743 mutex_unlock(&sqd
->lock
);
9746 static void io_sq_thread_park(struct io_sq_data
*sqd
)
9747 __acquires(&sqd
->lock
)
9749 WARN_ON_ONCE(sqd
->thread
== current
);
9751 atomic_inc(&sqd
->park_pending
);
9752 set_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
9753 mutex_lock(&sqd
->lock
);
9755 wake_up_process(sqd
->thread
);
9758 static void io_sq_thread_stop(struct io_sq_data
*sqd
)
9760 WARN_ON_ONCE(sqd
->thread
== current
);
9761 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
));
9763 set_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
);
9764 mutex_lock(&sqd
->lock
);
9766 wake_up_process(sqd
->thread
);
9767 mutex_unlock(&sqd
->lock
);
9768 wait_for_completion(&sqd
->exited
);
9771 static void io_put_sq_data(struct io_sq_data
*sqd
)
9773 if (refcount_dec_and_test(&sqd
->refs
)) {
9774 WARN_ON_ONCE(atomic_read(&sqd
->park_pending
));
9776 io_sq_thread_stop(sqd
);
9781 static void io_sq_thread_finish(struct io_ring_ctx
*ctx
)
9783 struct io_sq_data
*sqd
= ctx
->sq_data
;
9786 io_sq_thread_park(sqd
);
9787 list_del_init(&ctx
->sqd_list
);
9788 io_sqd_update_thread_idle(sqd
);
9789 io_sq_thread_unpark(sqd
);
9791 io_put_sq_data(sqd
);
9792 ctx
->sq_data
= NULL
;
9796 static struct io_sq_data
*io_attach_sq_data(struct io_uring_params
*p
)
9798 struct io_ring_ctx
*ctx_attach
;
9799 struct io_sq_data
*sqd
;
9802 f
= fdget(p
->wq_fd
);
9804 return ERR_PTR(-ENXIO
);
9805 if (f
.file
->f_op
!= &io_uring_fops
) {
9807 return ERR_PTR(-EINVAL
);
9810 ctx_attach
= f
.file
->private_data
;
9811 sqd
= ctx_attach
->sq_data
;
9814 return ERR_PTR(-EINVAL
);
9816 if (sqd
->task_tgid
!= current
->tgid
) {
9818 return ERR_PTR(-EPERM
);
9821 refcount_inc(&sqd
->refs
);
9826 static struct io_sq_data
*io_get_sq_data(struct io_uring_params
*p
,
9829 struct io_sq_data
*sqd
;
9832 if (p
->flags
& IORING_SETUP_ATTACH_WQ
) {
9833 sqd
= io_attach_sq_data(p
);
9838 /* fall through for EPERM case, setup new sqd/task */
9839 if (PTR_ERR(sqd
) != -EPERM
)
9843 sqd
= kzalloc(sizeof(*sqd
), GFP_KERNEL
);
9845 return ERR_PTR(-ENOMEM
);
9847 atomic_set(&sqd
->park_pending
, 0);
9848 refcount_set(&sqd
->refs
, 1);
9849 INIT_LIST_HEAD(&sqd
->ctx_list
);
9850 mutex_init(&sqd
->lock
);
9851 init_waitqueue_head(&sqd
->wait
);
9852 init_completion(&sqd
->exited
);
9857 * Ensure the UNIX gc is aware of our file set, so we are certain that
9858 * the io_uring can be safely unregistered on process exit, even if we have
9859 * loops in the file referencing. We account only files that can hold other
9860 * files because otherwise they can't form a loop and so are not interesting
9863 static int io_scm_file_account(struct io_ring_ctx
*ctx
, struct file
*file
)
9865 #if defined(CONFIG_UNIX)
9866 struct sock
*sk
= ctx
->ring_sock
->sk
;
9867 struct sk_buff_head
*head
= &sk
->sk_receive_queue
;
9868 struct scm_fp_list
*fpl
;
9869 struct sk_buff
*skb
;
9871 if (likely(!io_file_need_scm(file
)))
9875 * See if we can merge this file into an existing skb SCM_RIGHTS
9876 * file set. If there's no room, fall back to allocating a new skb
9877 * and filling it in.
9879 spin_lock_irq(&head
->lock
);
9880 skb
= skb_peek(head
);
9881 if (skb
&& UNIXCB(skb
).fp
->count
< SCM_MAX_FD
)
9882 __skb_unlink(skb
, head
);
9885 spin_unlock_irq(&head
->lock
);
9888 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
9892 skb
= alloc_skb(0, GFP_KERNEL
);
9898 fpl
->user
= get_uid(current_user());
9899 fpl
->max
= SCM_MAX_FD
;
9902 UNIXCB(skb
).fp
= fpl
;
9904 skb
->destructor
= unix_destruct_scm
;
9905 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
9908 fpl
= UNIXCB(skb
).fp
;
9909 fpl
->fp
[fpl
->count
++] = get_file(file
);
9910 unix_inflight(fpl
->user
, file
);
9911 skb_queue_head(head
, skb
);
9917 static void io_rsrc_file_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
9919 struct file
*file
= prsrc
->file
;
9920 #if defined(CONFIG_UNIX)
9921 struct sock
*sock
= ctx
->ring_sock
->sk
;
9922 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
9923 struct sk_buff
*skb
;
9926 if (!io_file_need_scm(file
)) {
9931 __skb_queue_head_init(&list
);
9934 * Find the skb that holds this file in its SCM_RIGHTS. When found,
9935 * remove this entry and rearrange the file array.
9937 skb
= skb_dequeue(head
);
9939 struct scm_fp_list
*fp
;
9941 fp
= UNIXCB(skb
).fp
;
9942 for (i
= 0; i
< fp
->count
; i
++) {
9945 if (fp
->fp
[i
] != file
)
9948 unix_notinflight(fp
->user
, fp
->fp
[i
]);
9949 left
= fp
->count
- 1 - i
;
9951 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
9952 left
* sizeof(struct file
*));
9959 __skb_queue_tail(&list
, skb
);
9969 __skb_queue_tail(&list
, skb
);
9971 skb
= skb_dequeue(head
);
9974 if (skb_peek(&list
)) {
9975 spin_lock_irq(&head
->lock
);
9976 while ((skb
= __skb_dequeue(&list
)) != NULL
)
9977 __skb_queue_tail(head
, skb
);
9978 spin_unlock_irq(&head
->lock
);
9985 static void __io_rsrc_put_work(struct io_rsrc_node
*ref_node
)
9987 struct io_rsrc_data
*rsrc_data
= ref_node
->rsrc_data
;
9988 struct io_ring_ctx
*ctx
= rsrc_data
->ctx
;
9989 struct io_rsrc_put
*prsrc
, *tmp
;
9991 list_for_each_entry_safe(prsrc
, tmp
, &ref_node
->rsrc_list
, list
) {
9992 list_del(&prsrc
->list
);
9995 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
9996 mutex_lock(&ctx
->uring_lock
);
9998 spin_lock(&ctx
->completion_lock
);
9999 io_fill_cqe_aux(ctx
, prsrc
->tag
, 0, 0);
10000 io_commit_cqring(ctx
);
10001 spin_unlock(&ctx
->completion_lock
);
10002 io_cqring_ev_posted(ctx
);
10004 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
10005 mutex_unlock(&ctx
->uring_lock
);
10008 rsrc_data
->do_put(ctx
, prsrc
);
10012 io_rsrc_node_destroy(ref_node
);
10013 if (atomic_dec_and_test(&rsrc_data
->refs
))
10014 complete(&rsrc_data
->done
);
10017 static void io_rsrc_put_work(struct work_struct
*work
)
10019 struct io_ring_ctx
*ctx
;
10020 struct llist_node
*node
;
10022 ctx
= container_of(work
, struct io_ring_ctx
, rsrc_put_work
.work
);
10023 node
= llist_del_all(&ctx
->rsrc_put_llist
);
10026 struct io_rsrc_node
*ref_node
;
10027 struct llist_node
*next
= node
->next
;
10029 ref_node
= llist_entry(node
, struct io_rsrc_node
, llist
);
10030 __io_rsrc_put_work(ref_node
);
10035 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
10036 unsigned nr_args
, u64 __user
*tags
)
10038 __s32 __user
*fds
= (__s32 __user
*) arg
;
10043 if (ctx
->file_data
)
10047 if (nr_args
> IORING_MAX_FIXED_FILES
)
10049 if (nr_args
> rlimit(RLIMIT_NOFILE
))
10051 ret
= io_rsrc_node_switch_start(ctx
);
10054 ret
= io_rsrc_data_alloc(ctx
, io_rsrc_file_put
, tags
, nr_args
,
10059 if (!io_alloc_file_tables(&ctx
->file_table
, nr_args
)) {
10060 io_rsrc_data_free(ctx
->file_data
);
10061 ctx
->file_data
= NULL
;
10065 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
10066 struct io_fixed_file
*file_slot
;
10068 if (fds
&& copy_from_user(&fd
, &fds
[i
], sizeof(fd
))) {
10072 /* allow sparse sets */
10073 if (!fds
|| fd
== -1) {
10075 if (unlikely(*io_get_tag_slot(ctx
->file_data
, i
)))
10082 if (unlikely(!file
))
10086 * Don't allow io_uring instances to be registered. If UNIX
10087 * isn't enabled, then this causes a reference cycle and this
10088 * instance can never get freed. If UNIX is enabled we'll
10089 * handle it just fine, but there's still no point in allowing
10090 * a ring fd as it doesn't support regular read/write anyway.
10092 if (file
->f_op
== &io_uring_fops
) {
10096 ret
= io_scm_file_account(ctx
, file
);
10101 file_slot
= io_fixed_file_slot(&ctx
->file_table
, i
);
10102 io_fixed_file_set(file_slot
, file
);
10103 io_file_bitmap_set(&ctx
->file_table
, i
);
10106 io_rsrc_node_switch(ctx
, NULL
);
10109 __io_sqe_files_unregister(ctx
);
10113 static int io_queue_rsrc_removal(struct io_rsrc_data
*data
, unsigned idx
,
10114 struct io_rsrc_node
*node
, void *rsrc
)
10116 u64
*tag_slot
= io_get_tag_slot(data
, idx
);
10117 struct io_rsrc_put
*prsrc
;
10119 prsrc
= kzalloc(sizeof(*prsrc
), GFP_KERNEL
);
10123 prsrc
->tag
= *tag_slot
;
10125 prsrc
->rsrc
= rsrc
;
10126 list_add(&prsrc
->list
, &node
->rsrc_list
);
10130 static int io_install_fixed_file(struct io_kiocb
*req
, struct file
*file
,
10131 unsigned int issue_flags
, u32 slot_index
)
10132 __must_hold(&req
->ctx
->uring_lock
)
10134 struct io_ring_ctx
*ctx
= req
->ctx
;
10135 bool needs_switch
= false;
10136 struct io_fixed_file
*file_slot
;
10139 if (file
->f_op
== &io_uring_fops
)
10141 if (!ctx
->file_data
)
10143 if (slot_index
>= ctx
->nr_user_files
)
10146 slot_index
= array_index_nospec(slot_index
, ctx
->nr_user_files
);
10147 file_slot
= io_fixed_file_slot(&ctx
->file_table
, slot_index
);
10149 if (file_slot
->file_ptr
) {
10150 struct file
*old_file
;
10152 ret
= io_rsrc_node_switch_start(ctx
);
10156 old_file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
10157 ret
= io_queue_rsrc_removal(ctx
->file_data
, slot_index
,
10158 ctx
->rsrc_node
, old_file
);
10161 file_slot
->file_ptr
= 0;
10162 io_file_bitmap_clear(&ctx
->file_table
, slot_index
);
10163 needs_switch
= true;
10166 ret
= io_scm_file_account(ctx
, file
);
10168 *io_get_tag_slot(ctx
->file_data
, slot_index
) = 0;
10169 io_fixed_file_set(file_slot
, file
);
10170 io_file_bitmap_set(&ctx
->file_table
, slot_index
);
10174 io_rsrc_node_switch(ctx
, ctx
->file_data
);
10180 static int __io_close_fixed(struct io_kiocb
*req
, unsigned int issue_flags
,
10181 unsigned int offset
)
10183 struct io_ring_ctx
*ctx
= req
->ctx
;
10184 struct io_fixed_file
*file_slot
;
10188 io_ring_submit_lock(ctx
, issue_flags
);
10190 if (unlikely(!ctx
->file_data
))
10193 if (offset
>= ctx
->nr_user_files
)
10195 ret
= io_rsrc_node_switch_start(ctx
);
10199 offset
= array_index_nospec(offset
, ctx
->nr_user_files
);
10200 file_slot
= io_fixed_file_slot(&ctx
->file_table
, offset
);
10202 if (!file_slot
->file_ptr
)
10205 file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
10206 ret
= io_queue_rsrc_removal(ctx
->file_data
, offset
, ctx
->rsrc_node
, file
);
10210 file_slot
->file_ptr
= 0;
10211 io_file_bitmap_clear(&ctx
->file_table
, offset
);
10212 io_rsrc_node_switch(ctx
, ctx
->file_data
);
10215 io_ring_submit_unlock(ctx
, issue_flags
);
10219 static inline int io_close_fixed(struct io_kiocb
*req
, unsigned int issue_flags
)
10221 return __io_close_fixed(req
, issue_flags
, req
->close
.file_slot
- 1);
10224 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
10225 struct io_uring_rsrc_update2
*up
,
10228 u64 __user
*tags
= u64_to_user_ptr(up
->tags
);
10229 __s32 __user
*fds
= u64_to_user_ptr(up
->data
);
10230 struct io_rsrc_data
*data
= ctx
->file_data
;
10231 struct io_fixed_file
*file_slot
;
10233 int fd
, i
, err
= 0;
10235 bool needs_switch
= false;
10237 if (!ctx
->file_data
)
10239 if (up
->offset
+ nr_args
> ctx
->nr_user_files
)
10242 for (done
= 0; done
< nr_args
; done
++) {
10245 if ((tags
&& copy_from_user(&tag
, &tags
[done
], sizeof(tag
))) ||
10246 copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
10250 if ((fd
== IORING_REGISTER_FILES_SKIP
|| fd
== -1) && tag
) {
10254 if (fd
== IORING_REGISTER_FILES_SKIP
)
10257 i
= array_index_nospec(up
->offset
+ done
, ctx
->nr_user_files
);
10258 file_slot
= io_fixed_file_slot(&ctx
->file_table
, i
);
10260 if (file_slot
->file_ptr
) {
10261 file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
10262 err
= io_queue_rsrc_removal(data
, i
, ctx
->rsrc_node
, file
);
10265 file_slot
->file_ptr
= 0;
10266 io_file_bitmap_clear(&ctx
->file_table
, i
);
10267 needs_switch
= true;
10276 * Don't allow io_uring instances to be registered. If
10277 * UNIX isn't enabled, then this causes a reference
10278 * cycle and this instance can never get freed. If UNIX
10279 * is enabled we'll handle it just fine, but there's
10280 * still no point in allowing a ring fd as it doesn't
10281 * support regular read/write anyway.
10283 if (file
->f_op
== &io_uring_fops
) {
10288 err
= io_scm_file_account(ctx
, file
);
10293 *io_get_tag_slot(data
, i
) = tag
;
10294 io_fixed_file_set(file_slot
, file
);
10295 io_file_bitmap_set(&ctx
->file_table
, i
);
10300 io_rsrc_node_switch(ctx
, data
);
10301 return done
? done
: err
;
10304 static struct io_wq
*io_init_wq_offload(struct io_ring_ctx
*ctx
,
10305 struct task_struct
*task
)
10307 struct io_wq_hash
*hash
;
10308 struct io_wq_data data
;
10309 unsigned int concurrency
;
10311 mutex_lock(&ctx
->uring_lock
);
10312 hash
= ctx
->hash_map
;
10314 hash
= kzalloc(sizeof(*hash
), GFP_KERNEL
);
10316 mutex_unlock(&ctx
->uring_lock
);
10317 return ERR_PTR(-ENOMEM
);
10319 refcount_set(&hash
->refs
, 1);
10320 init_waitqueue_head(&hash
->wait
);
10321 ctx
->hash_map
= hash
;
10323 mutex_unlock(&ctx
->uring_lock
);
10327 data
.free_work
= io_wq_free_work
;
10328 data
.do_work
= io_wq_submit_work
;
10330 /* Do QD, or 4 * CPUS, whatever is smallest */
10331 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
10333 return io_wq_create(concurrency
, &data
);
10336 static __cold
int io_uring_alloc_task_context(struct task_struct
*task
,
10337 struct io_ring_ctx
*ctx
)
10339 struct io_uring_task
*tctx
;
10342 tctx
= kzalloc(sizeof(*tctx
), GFP_KERNEL
);
10343 if (unlikely(!tctx
))
10346 tctx
->registered_rings
= kcalloc(IO_RINGFD_REG_MAX
,
10347 sizeof(struct file
*), GFP_KERNEL
);
10348 if (unlikely(!tctx
->registered_rings
)) {
10353 ret
= percpu_counter_init(&tctx
->inflight
, 0, GFP_KERNEL
);
10354 if (unlikely(ret
)) {
10355 kfree(tctx
->registered_rings
);
10360 tctx
->io_wq
= io_init_wq_offload(ctx
, task
);
10361 if (IS_ERR(tctx
->io_wq
)) {
10362 ret
= PTR_ERR(tctx
->io_wq
);
10363 percpu_counter_destroy(&tctx
->inflight
);
10364 kfree(tctx
->registered_rings
);
10369 xa_init(&tctx
->xa
);
10370 init_waitqueue_head(&tctx
->wait
);
10371 atomic_set(&tctx
->in_idle
, 0);
10372 atomic_set(&tctx
->inflight_tracked
, 0);
10373 task
->io_uring
= tctx
;
10374 spin_lock_init(&tctx
->task_lock
);
10375 INIT_WQ_LIST(&tctx
->task_list
);
10376 INIT_WQ_LIST(&tctx
->prio_task_list
);
10377 init_task_work(&tctx
->task_work
, tctx_task_work
);
10381 void __io_uring_free(struct task_struct
*tsk
)
10383 struct io_uring_task
*tctx
= tsk
->io_uring
;
10385 WARN_ON_ONCE(!xa_empty(&tctx
->xa
));
10386 WARN_ON_ONCE(tctx
->io_wq
);
10387 WARN_ON_ONCE(tctx
->cached_refs
);
10389 kfree(tctx
->registered_rings
);
10390 percpu_counter_destroy(&tctx
->inflight
);
10392 tsk
->io_uring
= NULL
;
10395 static __cold
int io_sq_offload_create(struct io_ring_ctx
*ctx
,
10396 struct io_uring_params
*p
)
10400 /* Retain compatibility with failing for an invalid attach attempt */
10401 if ((ctx
->flags
& (IORING_SETUP_ATTACH_WQ
| IORING_SETUP_SQPOLL
)) ==
10402 IORING_SETUP_ATTACH_WQ
) {
10405 f
= fdget(p
->wq_fd
);
10408 if (f
.file
->f_op
!= &io_uring_fops
) {
10414 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
10415 struct task_struct
*tsk
;
10416 struct io_sq_data
*sqd
;
10419 ret
= security_uring_sqpoll();
10423 sqd
= io_get_sq_data(p
, &attached
);
10425 ret
= PTR_ERR(sqd
);
10429 ctx
->sq_creds
= get_current_cred();
10430 ctx
->sq_data
= sqd
;
10431 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
10432 if (!ctx
->sq_thread_idle
)
10433 ctx
->sq_thread_idle
= HZ
;
10435 io_sq_thread_park(sqd
);
10436 list_add(&ctx
->sqd_list
, &sqd
->ctx_list
);
10437 io_sqd_update_thread_idle(sqd
);
10438 /* don't attach to a dying SQPOLL thread, would be racy */
10439 ret
= (attached
&& !sqd
->thread
) ? -ENXIO
: 0;
10440 io_sq_thread_unpark(sqd
);
10447 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
10448 int cpu
= p
->sq_thread_cpu
;
10451 if (cpu
>= nr_cpu_ids
|| !cpu_online(cpu
))
10458 sqd
->task_pid
= current
->pid
;
10459 sqd
->task_tgid
= current
->tgid
;
10460 tsk
= create_io_thread(io_sq_thread
, sqd
, NUMA_NO_NODE
);
10462 ret
= PTR_ERR(tsk
);
10467 ret
= io_uring_alloc_task_context(tsk
, ctx
);
10468 wake_up_new_task(tsk
);
10471 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
10472 /* Can't have SQ_AFF without SQPOLL */
10479 complete(&ctx
->sq_data
->exited
);
10481 io_sq_thread_finish(ctx
);
10485 static inline void __io_unaccount_mem(struct user_struct
*user
,
10486 unsigned long nr_pages
)
10488 atomic_long_sub(nr_pages
, &user
->locked_vm
);
10491 static inline int __io_account_mem(struct user_struct
*user
,
10492 unsigned long nr_pages
)
10494 unsigned long page_limit
, cur_pages
, new_pages
;
10496 /* Don't allow more pages than we can safely lock */
10497 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
10500 cur_pages
= atomic_long_read(&user
->locked_vm
);
10501 new_pages
= cur_pages
+ nr_pages
;
10502 if (new_pages
> page_limit
)
10504 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
10505 new_pages
) != cur_pages
);
10510 static void io_unaccount_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
10513 __io_unaccount_mem(ctx
->user
, nr_pages
);
10515 if (ctx
->mm_account
)
10516 atomic64_sub(nr_pages
, &ctx
->mm_account
->pinned_vm
);
10519 static int io_account_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
10524 ret
= __io_account_mem(ctx
->user
, nr_pages
);
10529 if (ctx
->mm_account
)
10530 atomic64_add(nr_pages
, &ctx
->mm_account
->pinned_vm
);
10535 static void io_mem_free(void *ptr
)
10542 page
= virt_to_head_page(ptr
);
10543 if (put_page_testzero(page
))
10544 free_compound_page(page
);
10547 static void *io_mem_alloc(size_t size
)
10549 gfp_t gfp
= GFP_KERNEL_ACCOUNT
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
;
10551 return (void *) __get_free_pages(gfp
, get_order(size
));
10554 static unsigned long rings_size(struct io_ring_ctx
*ctx
, unsigned int sq_entries
,
10555 unsigned int cq_entries
, size_t *sq_offset
)
10557 struct io_rings
*rings
;
10558 size_t off
, sq_array_size
;
10560 off
= struct_size(rings
, cqes
, cq_entries
);
10561 if (off
== SIZE_MAX
)
10563 if (ctx
->flags
& IORING_SETUP_CQE32
) {
10564 if (check_shl_overflow(off
, 1, &off
))
10569 off
= ALIGN(off
, SMP_CACHE_BYTES
);
10577 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
10578 if (sq_array_size
== SIZE_MAX
)
10581 if (check_add_overflow(off
, sq_array_size
, &off
))
10587 static void io_buffer_unmap(struct io_ring_ctx
*ctx
, struct io_mapped_ubuf
**slot
)
10589 struct io_mapped_ubuf
*imu
= *slot
;
10592 if (imu
!= ctx
->dummy_ubuf
) {
10593 for (i
= 0; i
< imu
->nr_bvecs
; i
++)
10594 unpin_user_page(imu
->bvec
[i
].bv_page
);
10595 if (imu
->acct_pages
)
10596 io_unaccount_mem(ctx
, imu
->acct_pages
);
10602 static void io_rsrc_buf_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
10604 io_buffer_unmap(ctx
, &prsrc
->buf
);
10608 static void __io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
10612 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++)
10613 io_buffer_unmap(ctx
, &ctx
->user_bufs
[i
]);
10614 kfree(ctx
->user_bufs
);
10615 io_rsrc_data_free(ctx
->buf_data
);
10616 ctx
->user_bufs
= NULL
;
10617 ctx
->buf_data
= NULL
;
10618 ctx
->nr_user_bufs
= 0;
10621 static int io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
10623 unsigned nr
= ctx
->nr_user_bufs
;
10626 if (!ctx
->buf_data
)
10630 * Quiesce may unlock ->uring_lock, and while it's not held
10631 * prevent new requests using the table.
10633 ctx
->nr_user_bufs
= 0;
10634 ret
= io_rsrc_ref_quiesce(ctx
->buf_data
, ctx
);
10635 ctx
->nr_user_bufs
= nr
;
10637 __io_sqe_buffers_unregister(ctx
);
10641 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
10642 void __user
*arg
, unsigned index
)
10644 struct iovec __user
*src
;
10646 #ifdef CONFIG_COMPAT
10648 struct compat_iovec __user
*ciovs
;
10649 struct compat_iovec ciov
;
10651 ciovs
= (struct compat_iovec __user
*) arg
;
10652 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
10655 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
10656 dst
->iov_len
= ciov
.iov_len
;
10660 src
= (struct iovec __user
*) arg
;
10661 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
10667 * Not super efficient, but this is just a registration time. And we do cache
10668 * the last compound head, so generally we'll only do a full search if we don't
10671 * We check if the given compound head page has already been accounted, to
10672 * avoid double accounting it. This allows us to account the full size of the
10673 * page, not just the constituent pages of a huge page.
10675 static bool headpage_already_acct(struct io_ring_ctx
*ctx
, struct page
**pages
,
10676 int nr_pages
, struct page
*hpage
)
10680 /* check current page array */
10681 for (i
= 0; i
< nr_pages
; i
++) {
10682 if (!PageCompound(pages
[i
]))
10684 if (compound_head(pages
[i
]) == hpage
)
10688 /* check previously registered pages */
10689 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
10690 struct io_mapped_ubuf
*imu
= ctx
->user_bufs
[i
];
10692 for (j
= 0; j
< imu
->nr_bvecs
; j
++) {
10693 if (!PageCompound(imu
->bvec
[j
].bv_page
))
10695 if (compound_head(imu
->bvec
[j
].bv_page
) == hpage
)
10703 static int io_buffer_account_pin(struct io_ring_ctx
*ctx
, struct page
**pages
,
10704 int nr_pages
, struct io_mapped_ubuf
*imu
,
10705 struct page
**last_hpage
)
10709 imu
->acct_pages
= 0;
10710 for (i
= 0; i
< nr_pages
; i
++) {
10711 if (!PageCompound(pages
[i
])) {
10714 struct page
*hpage
;
10716 hpage
= compound_head(pages
[i
]);
10717 if (hpage
== *last_hpage
)
10719 *last_hpage
= hpage
;
10720 if (headpage_already_acct(ctx
, pages
, i
, hpage
))
10722 imu
->acct_pages
+= page_size(hpage
) >> PAGE_SHIFT
;
10726 if (!imu
->acct_pages
)
10729 ret
= io_account_mem(ctx
, imu
->acct_pages
);
10731 imu
->acct_pages
= 0;
10735 static struct page
**io_pin_pages(unsigned long ubuf
, unsigned long len
,
10738 unsigned long start
, end
, nr_pages
;
10739 struct vm_area_struct
**vmas
= NULL
;
10740 struct page
**pages
= NULL
;
10741 int i
, pret
, ret
= -ENOMEM
;
10743 end
= (ubuf
+ len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
10744 start
= ubuf
>> PAGE_SHIFT
;
10745 nr_pages
= end
- start
;
10747 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*), GFP_KERNEL
);
10751 vmas
= kvmalloc_array(nr_pages
, sizeof(struct vm_area_struct
*),
10757 mmap_read_lock(current
->mm
);
10758 pret
= pin_user_pages(ubuf
, nr_pages
, FOLL_WRITE
| FOLL_LONGTERM
,
10760 if (pret
== nr_pages
) {
10761 /* don't support file backed memory */
10762 for (i
= 0; i
< nr_pages
; i
++) {
10763 struct vm_area_struct
*vma
= vmas
[i
];
10765 if (vma_is_shmem(vma
))
10767 if (vma
->vm_file
&&
10768 !is_file_hugepages(vma
->vm_file
)) {
10773 *npages
= nr_pages
;
10775 ret
= pret
< 0 ? pret
: -EFAULT
;
10777 mmap_read_unlock(current
->mm
);
10780 * if we did partial map, or found file backed vmas,
10781 * release any pages we did get
10784 unpin_user_pages(pages
, pret
);
10792 pages
= ERR_PTR(ret
);
10797 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, struct iovec
*iov
,
10798 struct io_mapped_ubuf
**pimu
,
10799 struct page
**last_hpage
)
10801 struct io_mapped_ubuf
*imu
= NULL
;
10802 struct page
**pages
= NULL
;
10805 int ret
, nr_pages
, i
;
10807 if (!iov
->iov_base
) {
10808 *pimu
= ctx
->dummy_ubuf
;
10815 pages
= io_pin_pages((unsigned long) iov
->iov_base
, iov
->iov_len
,
10817 if (IS_ERR(pages
)) {
10818 ret
= PTR_ERR(pages
);
10823 imu
= kvmalloc(struct_size(imu
, bvec
, nr_pages
), GFP_KERNEL
);
10827 ret
= io_buffer_account_pin(ctx
, pages
, nr_pages
, imu
, last_hpage
);
10829 unpin_user_pages(pages
, nr_pages
);
10833 off
= (unsigned long) iov
->iov_base
& ~PAGE_MASK
;
10834 size
= iov
->iov_len
;
10835 for (i
= 0; i
< nr_pages
; i
++) {
10838 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
10839 imu
->bvec
[i
].bv_page
= pages
[i
];
10840 imu
->bvec
[i
].bv_len
= vec_len
;
10841 imu
->bvec
[i
].bv_offset
= off
;
10845 /* store original address for later verification */
10846 imu
->ubuf
= (unsigned long) iov
->iov_base
;
10847 imu
->ubuf_end
= imu
->ubuf
+ iov
->iov_len
;
10848 imu
->nr_bvecs
= nr_pages
;
10858 static int io_buffers_map_alloc(struct io_ring_ctx
*ctx
, unsigned int nr_args
)
10860 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(*ctx
->user_bufs
), GFP_KERNEL
);
10861 return ctx
->user_bufs
? 0 : -ENOMEM
;
10864 static int io_buffer_validate(struct iovec
*iov
)
10866 unsigned long tmp
, acct_len
= iov
->iov_len
+ (PAGE_SIZE
- 1);
10869 * Don't impose further limits on the size and buffer
10870 * constraints here, we'll -EINVAL later when IO is
10871 * submitted if they are wrong.
10873 if (!iov
->iov_base
)
10874 return iov
->iov_len
? -EFAULT
: 0;
10878 /* arbitrary limit, but we need something */
10879 if (iov
->iov_len
> SZ_1G
)
10882 if (check_add_overflow((unsigned long)iov
->iov_base
, acct_len
, &tmp
))
10888 static int io_sqe_buffers_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
10889 unsigned int nr_args
, u64 __user
*tags
)
10891 struct page
*last_hpage
= NULL
;
10892 struct io_rsrc_data
*data
;
10896 if (ctx
->user_bufs
)
10898 if (!nr_args
|| nr_args
> IORING_MAX_REG_BUFFERS
)
10900 ret
= io_rsrc_node_switch_start(ctx
);
10903 ret
= io_rsrc_data_alloc(ctx
, io_rsrc_buf_put
, tags
, nr_args
, &data
);
10906 ret
= io_buffers_map_alloc(ctx
, nr_args
);
10908 io_rsrc_data_free(data
);
10912 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_bufs
++) {
10914 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
10917 ret
= io_buffer_validate(&iov
);
10921 memset(&iov
, 0, sizeof(iov
));
10924 if (!iov
.iov_base
&& *io_get_tag_slot(data
, i
)) {
10929 ret
= io_sqe_buffer_register(ctx
, &iov
, &ctx
->user_bufs
[i
],
10935 WARN_ON_ONCE(ctx
->buf_data
);
10937 ctx
->buf_data
= data
;
10939 __io_sqe_buffers_unregister(ctx
);
10941 io_rsrc_node_switch(ctx
, NULL
);
10945 static int __io_sqe_buffers_update(struct io_ring_ctx
*ctx
,
10946 struct io_uring_rsrc_update2
*up
,
10947 unsigned int nr_args
)
10949 u64 __user
*tags
= u64_to_user_ptr(up
->tags
);
10950 struct iovec iov
, __user
*iovs
= u64_to_user_ptr(up
->data
);
10951 struct page
*last_hpage
= NULL
;
10952 bool needs_switch
= false;
10956 if (!ctx
->buf_data
)
10958 if (up
->offset
+ nr_args
> ctx
->nr_user_bufs
)
10961 for (done
= 0; done
< nr_args
; done
++) {
10962 struct io_mapped_ubuf
*imu
;
10963 int offset
= up
->offset
+ done
;
10966 err
= io_copy_iov(ctx
, &iov
, iovs
, done
);
10969 if (tags
&& copy_from_user(&tag
, &tags
[done
], sizeof(tag
))) {
10973 err
= io_buffer_validate(&iov
);
10976 if (!iov
.iov_base
&& tag
) {
10980 err
= io_sqe_buffer_register(ctx
, &iov
, &imu
, &last_hpage
);
10984 i
= array_index_nospec(offset
, ctx
->nr_user_bufs
);
10985 if (ctx
->user_bufs
[i
] != ctx
->dummy_ubuf
) {
10986 err
= io_queue_rsrc_removal(ctx
->buf_data
, i
,
10987 ctx
->rsrc_node
, ctx
->user_bufs
[i
]);
10988 if (unlikely(err
)) {
10989 io_buffer_unmap(ctx
, &imu
);
10992 ctx
->user_bufs
[i
] = NULL
;
10993 needs_switch
= true;
10996 ctx
->user_bufs
[i
] = imu
;
10997 *io_get_tag_slot(ctx
->buf_data
, offset
) = tag
;
11001 io_rsrc_node_switch(ctx
, ctx
->buf_data
);
11002 return done
? done
: err
;
11005 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
11006 unsigned int eventfd_async
)
11008 struct io_ev_fd
*ev_fd
;
11009 __s32 __user
*fds
= arg
;
11012 ev_fd
= rcu_dereference_protected(ctx
->io_ev_fd
,
11013 lockdep_is_held(&ctx
->uring_lock
));
11017 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
11020 ev_fd
= kmalloc(sizeof(*ev_fd
), GFP_KERNEL
);
11024 ev_fd
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
11025 if (IS_ERR(ev_fd
->cq_ev_fd
)) {
11026 int ret
= PTR_ERR(ev_fd
->cq_ev_fd
);
11030 ev_fd
->eventfd_async
= eventfd_async
;
11031 ctx
->has_evfd
= true;
11032 rcu_assign_pointer(ctx
->io_ev_fd
, ev_fd
);
11036 static void io_eventfd_put(struct rcu_head
*rcu
)
11038 struct io_ev_fd
*ev_fd
= container_of(rcu
, struct io_ev_fd
, rcu
);
11040 eventfd_ctx_put(ev_fd
->cq_ev_fd
);
11044 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
11046 struct io_ev_fd
*ev_fd
;
11048 ev_fd
= rcu_dereference_protected(ctx
->io_ev_fd
,
11049 lockdep_is_held(&ctx
->uring_lock
));
11051 ctx
->has_evfd
= false;
11052 rcu_assign_pointer(ctx
->io_ev_fd
, NULL
);
11053 call_rcu(&ev_fd
->rcu
, io_eventfd_put
);
11060 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
11062 struct io_buffer_list
*bl
;
11063 unsigned long index
;
11066 for (i
= 0; i
< BGID_ARRAY
; i
++) {
11069 __io_remove_buffers(ctx
, &ctx
->io_bl
[i
], -1U);
11072 xa_for_each(&ctx
->io_bl_xa
, index
, bl
) {
11073 xa_erase(&ctx
->io_bl_xa
, bl
->bgid
);
11074 __io_remove_buffers(ctx
, bl
, -1U);
11078 while (!list_empty(&ctx
->io_buffers_pages
)) {
11081 page
= list_first_entry(&ctx
->io_buffers_pages
, struct page
, lru
);
11082 list_del_init(&page
->lru
);
11087 static void io_req_caches_free(struct io_ring_ctx
*ctx
)
11089 struct io_submit_state
*state
= &ctx
->submit_state
;
11092 mutex_lock(&ctx
->uring_lock
);
11093 io_flush_cached_locked_reqs(ctx
, state
);
11095 while (!io_req_cache_empty(ctx
)) {
11096 struct io_wq_work_node
*node
;
11097 struct io_kiocb
*req
;
11099 node
= wq_stack_extract(&state
->free_list
);
11100 req
= container_of(node
, struct io_kiocb
, comp_list
);
11101 kmem_cache_free(req_cachep
, req
);
11105 percpu_ref_put_many(&ctx
->refs
, nr
);
11106 mutex_unlock(&ctx
->uring_lock
);
11109 static void io_wait_rsrc_data(struct io_rsrc_data
*data
)
11111 if (data
&& !atomic_dec_and_test(&data
->refs
))
11112 wait_for_completion(&data
->done
);
11115 static void io_flush_apoll_cache(struct io_ring_ctx
*ctx
)
11117 struct async_poll
*apoll
;
11119 while (!list_empty(&ctx
->apoll_cache
)) {
11120 apoll
= list_first_entry(&ctx
->apoll_cache
, struct async_poll
,
11122 list_del(&apoll
->poll
.wait
.entry
);
11127 static __cold
void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
11129 io_sq_thread_finish(ctx
);
11131 if (ctx
->mm_account
) {
11132 mmdrop(ctx
->mm_account
);
11133 ctx
->mm_account
= NULL
;
11136 io_rsrc_refs_drop(ctx
);
11137 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
11138 io_wait_rsrc_data(ctx
->buf_data
);
11139 io_wait_rsrc_data(ctx
->file_data
);
11141 mutex_lock(&ctx
->uring_lock
);
11143 __io_sqe_buffers_unregister(ctx
);
11144 if (ctx
->file_data
)
11145 __io_sqe_files_unregister(ctx
);
11147 __io_cqring_overflow_flush(ctx
, true);
11148 io_eventfd_unregister(ctx
);
11149 io_flush_apoll_cache(ctx
);
11150 mutex_unlock(&ctx
->uring_lock
);
11151 io_destroy_buffers(ctx
);
11153 put_cred(ctx
->sq_creds
);
11155 /* there are no registered resources left, nobody uses it */
11156 if (ctx
->rsrc_node
)
11157 io_rsrc_node_destroy(ctx
->rsrc_node
);
11158 if (ctx
->rsrc_backup_node
)
11159 io_rsrc_node_destroy(ctx
->rsrc_backup_node
);
11160 flush_delayed_work(&ctx
->rsrc_put_work
);
11161 flush_delayed_work(&ctx
->fallback_work
);
11163 WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
));
11164 WARN_ON_ONCE(!llist_empty(&ctx
->rsrc_put_llist
));
11166 #if defined(CONFIG_UNIX)
11167 if (ctx
->ring_sock
) {
11168 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
11169 sock_release(ctx
->ring_sock
);
11172 WARN_ON_ONCE(!list_empty(&ctx
->ltimeout_list
));
11174 io_mem_free(ctx
->rings
);
11175 io_mem_free(ctx
->sq_sqes
);
11177 percpu_ref_exit(&ctx
->refs
);
11178 free_uid(ctx
->user
);
11179 io_req_caches_free(ctx
);
11181 io_wq_put_hash(ctx
->hash_map
);
11182 kfree(ctx
->cancel_hash
);
11183 kfree(ctx
->dummy_ubuf
);
11185 xa_destroy(&ctx
->io_bl_xa
);
11189 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
11191 struct io_ring_ctx
*ctx
= file
->private_data
;
11194 poll_wait(file
, &ctx
->cq_wait
, wait
);
11196 * synchronizes with barrier from wq_has_sleeper call in
11200 if (!io_sqring_full(ctx
))
11201 mask
|= EPOLLOUT
| EPOLLWRNORM
;
11204 * Don't flush cqring overflow list here, just do a simple check.
11205 * Otherwise there could possible be ABBA deadlock:
11208 * lock(&ctx->uring_lock);
11210 * lock(&ctx->uring_lock);
11213 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
11214 * pushs them to do the flush.
11216 if (io_cqring_events(ctx
) ||
11217 test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
))
11218 mask
|= EPOLLIN
| EPOLLRDNORM
;
11223 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
11225 const struct cred
*creds
;
11227 creds
= xa_erase(&ctx
->personalities
, id
);
11236 struct io_tctx_exit
{
11237 struct callback_head task_work
;
11238 struct completion completion
;
11239 struct io_ring_ctx
*ctx
;
11242 static __cold
void io_tctx_exit_cb(struct callback_head
*cb
)
11244 struct io_uring_task
*tctx
= current
->io_uring
;
11245 struct io_tctx_exit
*work
;
11247 work
= container_of(cb
, struct io_tctx_exit
, task_work
);
11249 * When @in_idle, we're in cancellation and it's racy to remove the
11250 * node. It'll be removed by the end of cancellation, just ignore it.
11252 if (!atomic_read(&tctx
->in_idle
))
11253 io_uring_del_tctx_node((unsigned long)work
->ctx
);
11254 complete(&work
->completion
);
11257 static __cold
bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
11259 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
11261 return req
->ctx
== data
;
11264 static __cold
void io_ring_exit_work(struct work_struct
*work
)
11266 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
, exit_work
);
11267 unsigned long timeout
= jiffies
+ HZ
* 60 * 5;
11268 unsigned long interval
= HZ
/ 20;
11269 struct io_tctx_exit exit
;
11270 struct io_tctx_node
*node
;
11274 * If we're doing polled IO and end up having requests being
11275 * submitted async (out-of-line), then completions can come in while
11276 * we're waiting for refs to drop. We need to reap these manually,
11277 * as nobody else will be looking for them.
11280 io_uring_try_cancel_requests(ctx
, NULL
, true);
11281 if (ctx
->sq_data
) {
11282 struct io_sq_data
*sqd
= ctx
->sq_data
;
11283 struct task_struct
*tsk
;
11285 io_sq_thread_park(sqd
);
11287 if (tsk
&& tsk
->io_uring
&& tsk
->io_uring
->io_wq
)
11288 io_wq_cancel_cb(tsk
->io_uring
->io_wq
,
11289 io_cancel_ctx_cb
, ctx
, true);
11290 io_sq_thread_unpark(sqd
);
11293 io_req_caches_free(ctx
);
11295 if (WARN_ON_ONCE(time_after(jiffies
, timeout
))) {
11296 /* there is little hope left, don't run it too often */
11297 interval
= HZ
* 60;
11299 } while (!wait_for_completion_timeout(&ctx
->ref_comp
, interval
));
11301 init_completion(&exit
.completion
);
11302 init_task_work(&exit
.task_work
, io_tctx_exit_cb
);
11305 * Some may use context even when all refs and requests have been put,
11306 * and they are free to do so while still holding uring_lock or
11307 * completion_lock, see io_req_task_submit(). Apart from other work,
11308 * this lock/unlock section also waits them to finish.
11310 mutex_lock(&ctx
->uring_lock
);
11311 while (!list_empty(&ctx
->tctx_list
)) {
11312 WARN_ON_ONCE(time_after(jiffies
, timeout
));
11314 node
= list_first_entry(&ctx
->tctx_list
, struct io_tctx_node
,
11316 /* don't spin on a single task if cancellation failed */
11317 list_rotate_left(&ctx
->tctx_list
);
11318 ret
= task_work_add(node
->task
, &exit
.task_work
, TWA_SIGNAL
);
11319 if (WARN_ON_ONCE(ret
))
11322 mutex_unlock(&ctx
->uring_lock
);
11323 wait_for_completion(&exit
.completion
);
11324 mutex_lock(&ctx
->uring_lock
);
11326 mutex_unlock(&ctx
->uring_lock
);
11327 spin_lock(&ctx
->completion_lock
);
11328 spin_unlock(&ctx
->completion_lock
);
11330 io_ring_ctx_free(ctx
);
11333 /* Returns true if we found and killed one or more timeouts */
11334 static __cold
bool io_kill_timeouts(struct io_ring_ctx
*ctx
,
11335 struct task_struct
*tsk
, bool cancel_all
)
11337 struct io_kiocb
*req
, *tmp
;
11340 spin_lock(&ctx
->completion_lock
);
11341 spin_lock_irq(&ctx
->timeout_lock
);
11342 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
11343 if (io_match_task(req
, tsk
, cancel_all
)) {
11344 io_kill_timeout(req
, -ECANCELED
);
11348 spin_unlock_irq(&ctx
->timeout_lock
);
11349 io_commit_cqring(ctx
);
11350 spin_unlock(&ctx
->completion_lock
);
11352 io_cqring_ev_posted(ctx
);
11353 return canceled
!= 0;
11356 static __cold
void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
11358 unsigned long index
;
11359 struct creds
*creds
;
11361 mutex_lock(&ctx
->uring_lock
);
11362 percpu_ref_kill(&ctx
->refs
);
11364 __io_cqring_overflow_flush(ctx
, true);
11365 xa_for_each(&ctx
->personalities
, index
, creds
)
11366 io_unregister_personality(ctx
, index
);
11367 mutex_unlock(&ctx
->uring_lock
);
11369 /* failed during ring init, it couldn't have issued any requests */
11371 io_kill_timeouts(ctx
, NULL
, true);
11372 io_poll_remove_all(ctx
, NULL
, true);
11373 /* if we failed setting up the ctx, we might not have any rings */
11374 io_iopoll_try_reap_events(ctx
);
11377 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
11379 * Use system_unbound_wq to avoid spawning tons of event kworkers
11380 * if we're exiting a ton of rings at the same time. It just adds
11381 * noise and overhead, there's no discernable change in runtime
11382 * over using system_wq.
11384 queue_work(system_unbound_wq
, &ctx
->exit_work
);
11387 static int io_uring_release(struct inode
*inode
, struct file
*file
)
11389 struct io_ring_ctx
*ctx
= file
->private_data
;
11391 file
->private_data
= NULL
;
11392 io_ring_ctx_wait_and_kill(ctx
);
11396 struct io_task_cancel
{
11397 struct task_struct
*task
;
11401 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
11403 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
11404 struct io_task_cancel
*cancel
= data
;
11406 return io_match_task_safe(req
, cancel
->task
, cancel
->all
);
11409 static __cold
bool io_cancel_defer_files(struct io_ring_ctx
*ctx
,
11410 struct task_struct
*task
,
11413 struct io_defer_entry
*de
;
11416 spin_lock(&ctx
->completion_lock
);
11417 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
11418 if (io_match_task_safe(de
->req
, task
, cancel_all
)) {
11419 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
11423 spin_unlock(&ctx
->completion_lock
);
11424 if (list_empty(&list
))
11427 while (!list_empty(&list
)) {
11428 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
11429 list_del_init(&de
->list
);
11430 io_req_complete_failed(de
->req
, -ECANCELED
);
11436 static __cold
bool io_uring_try_cancel_iowq(struct io_ring_ctx
*ctx
)
11438 struct io_tctx_node
*node
;
11439 enum io_wq_cancel cret
;
11442 mutex_lock(&ctx
->uring_lock
);
11443 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
11444 struct io_uring_task
*tctx
= node
->task
->io_uring
;
11447 * io_wq will stay alive while we hold uring_lock, because it's
11448 * killed after ctx nodes, which requires to take the lock.
11450 if (!tctx
|| !tctx
->io_wq
)
11452 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
11453 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
11455 mutex_unlock(&ctx
->uring_lock
);
11460 static __cold
void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
11461 struct task_struct
*task
,
11464 struct io_task_cancel cancel
= { .task
= task
, .all
= cancel_all
, };
11465 struct io_uring_task
*tctx
= task
? task
->io_uring
: NULL
;
11467 /* failed during ring init, it couldn't have issued any requests */
11472 enum io_wq_cancel cret
;
11476 ret
|= io_uring_try_cancel_iowq(ctx
);
11477 } else if (tctx
&& tctx
->io_wq
) {
11479 * Cancels requests of all rings, not only @ctx, but
11480 * it's fine as the task is in exit/exec.
11482 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_task_cb
,
11484 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
11487 /* SQPOLL thread does its own polling */
11488 if ((!(ctx
->flags
& IORING_SETUP_SQPOLL
) && cancel_all
) ||
11489 (ctx
->sq_data
&& ctx
->sq_data
->thread
== current
)) {
11490 while (!wq_list_empty(&ctx
->iopoll_list
)) {
11491 io_iopoll_try_reap_events(ctx
);
11496 ret
|= io_cancel_defer_files(ctx
, task
, cancel_all
);
11497 ret
|= io_poll_remove_all(ctx
, task
, cancel_all
);
11498 ret
|= io_kill_timeouts(ctx
, task
, cancel_all
);
11500 ret
|= io_run_task_work();
11507 static int __io_uring_add_tctx_node(struct io_ring_ctx
*ctx
)
11509 struct io_uring_task
*tctx
= current
->io_uring
;
11510 struct io_tctx_node
*node
;
11513 if (unlikely(!tctx
)) {
11514 ret
= io_uring_alloc_task_context(current
, ctx
);
11518 tctx
= current
->io_uring
;
11519 if (ctx
->iowq_limits_set
) {
11520 unsigned int limits
[2] = { ctx
->iowq_limits
[0],
11521 ctx
->iowq_limits
[1], };
11523 ret
= io_wq_max_workers(tctx
->io_wq
, limits
);
11528 if (!xa_load(&tctx
->xa
, (unsigned long)ctx
)) {
11529 node
= kmalloc(sizeof(*node
), GFP_KERNEL
);
11533 node
->task
= current
;
11535 ret
= xa_err(xa_store(&tctx
->xa
, (unsigned long)ctx
,
11536 node
, GFP_KERNEL
));
11542 mutex_lock(&ctx
->uring_lock
);
11543 list_add(&node
->ctx_node
, &ctx
->tctx_list
);
11544 mutex_unlock(&ctx
->uring_lock
);
11551 * Note that this task has used io_uring. We use it for cancelation purposes.
11553 static inline int io_uring_add_tctx_node(struct io_ring_ctx
*ctx
)
11555 struct io_uring_task
*tctx
= current
->io_uring
;
11557 if (likely(tctx
&& tctx
->last
== ctx
))
11559 return __io_uring_add_tctx_node(ctx
);
11563 * Remove this io_uring_file -> task mapping.
11565 static __cold
void io_uring_del_tctx_node(unsigned long index
)
11567 struct io_uring_task
*tctx
= current
->io_uring
;
11568 struct io_tctx_node
*node
;
11572 node
= xa_erase(&tctx
->xa
, index
);
11576 WARN_ON_ONCE(current
!= node
->task
);
11577 WARN_ON_ONCE(list_empty(&node
->ctx_node
));
11579 mutex_lock(&node
->ctx
->uring_lock
);
11580 list_del(&node
->ctx_node
);
11581 mutex_unlock(&node
->ctx
->uring_lock
);
11583 if (tctx
->last
== node
->ctx
)
11588 static __cold
void io_uring_clean_tctx(struct io_uring_task
*tctx
)
11590 struct io_wq
*wq
= tctx
->io_wq
;
11591 struct io_tctx_node
*node
;
11592 unsigned long index
;
11594 xa_for_each(&tctx
->xa
, index
, node
) {
11595 io_uring_del_tctx_node(index
);
11600 * Must be after io_uring_del_tctx_node() (removes nodes under
11601 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
11603 io_wq_put_and_exit(wq
);
11604 tctx
->io_wq
= NULL
;
11608 static s64
tctx_inflight(struct io_uring_task
*tctx
, bool tracked
)
11611 return atomic_read(&tctx
->inflight_tracked
);
11612 return percpu_counter_sum(&tctx
->inflight
);
11616 * Find any io_uring ctx that this task has registered or done IO on, and cancel
11617 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
11619 static __cold
void io_uring_cancel_generic(bool cancel_all
,
11620 struct io_sq_data
*sqd
)
11622 struct io_uring_task
*tctx
= current
->io_uring
;
11623 struct io_ring_ctx
*ctx
;
11627 WARN_ON_ONCE(sqd
&& sqd
->thread
!= current
);
11629 if (!current
->io_uring
)
11632 io_wq_exit_start(tctx
->io_wq
);
11634 atomic_inc(&tctx
->in_idle
);
11636 io_uring_drop_tctx_refs(current
);
11637 /* read completions before cancelations */
11638 inflight
= tctx_inflight(tctx
, !cancel_all
);
11643 struct io_tctx_node
*node
;
11644 unsigned long index
;
11646 xa_for_each(&tctx
->xa
, index
, node
) {
11647 /* sqpoll task will cancel all its requests */
11648 if (node
->ctx
->sq_data
)
11650 io_uring_try_cancel_requests(node
->ctx
, current
,
11654 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
11655 io_uring_try_cancel_requests(ctx
, current
,
11659 prepare_to_wait(&tctx
->wait
, &wait
, TASK_INTERRUPTIBLE
);
11660 io_run_task_work();
11661 io_uring_drop_tctx_refs(current
);
11664 * If we've seen completions, retry without waiting. This
11665 * avoids a race where a completion comes in before we did
11666 * prepare_to_wait().
11668 if (inflight
== tctx_inflight(tctx
, !cancel_all
))
11670 finish_wait(&tctx
->wait
, &wait
);
11673 io_uring_clean_tctx(tctx
);
11676 * We shouldn't run task_works after cancel, so just leave
11677 * ->in_idle set for normal exit.
11679 atomic_dec(&tctx
->in_idle
);
11680 /* for exec all current's requests should be gone, kill tctx */
11681 __io_uring_free(current
);
11685 void __io_uring_cancel(bool cancel_all
)
11687 io_uring_cancel_generic(cancel_all
, NULL
);
11690 void io_uring_unreg_ringfd(void)
11692 struct io_uring_task
*tctx
= current
->io_uring
;
11695 for (i
= 0; i
< IO_RINGFD_REG_MAX
; i
++) {
11696 if (tctx
->registered_rings
[i
]) {
11697 fput(tctx
->registered_rings
[i
]);
11698 tctx
->registered_rings
[i
] = NULL
;
11703 static int io_ring_add_registered_fd(struct io_uring_task
*tctx
, int fd
,
11704 int start
, int end
)
11709 for (offset
= start
; offset
< end
; offset
++) {
11710 offset
= array_index_nospec(offset
, IO_RINGFD_REG_MAX
);
11711 if (tctx
->registered_rings
[offset
])
11717 } else if (file
->f_op
!= &io_uring_fops
) {
11719 return -EOPNOTSUPP
;
11721 tctx
->registered_rings
[offset
] = file
;
11729 * Register a ring fd to avoid fdget/fdput for each io_uring_enter()
11730 * invocation. User passes in an array of struct io_uring_rsrc_update
11731 * with ->data set to the ring_fd, and ->offset given for the desired
11732 * index. If no index is desired, application may set ->offset == -1U
11733 * and we'll find an available index. Returns number of entries
11734 * successfully processed, or < 0 on error if none were processed.
11736 static int io_ringfd_register(struct io_ring_ctx
*ctx
, void __user
*__arg
,
11739 struct io_uring_rsrc_update __user
*arg
= __arg
;
11740 struct io_uring_rsrc_update reg
;
11741 struct io_uring_task
*tctx
;
11744 if (!nr_args
|| nr_args
> IO_RINGFD_REG_MAX
)
11747 mutex_unlock(&ctx
->uring_lock
);
11748 ret
= io_uring_add_tctx_node(ctx
);
11749 mutex_lock(&ctx
->uring_lock
);
11753 tctx
= current
->io_uring
;
11754 for (i
= 0; i
< nr_args
; i
++) {
11757 if (copy_from_user(®
, &arg
[i
], sizeof(reg
))) {
11767 if (reg
.offset
== -1U) {
11769 end
= IO_RINGFD_REG_MAX
;
11771 if (reg
.offset
>= IO_RINGFD_REG_MAX
) {
11775 start
= reg
.offset
;
11779 ret
= io_ring_add_registered_fd(tctx
, reg
.data
, start
, end
);
11784 if (copy_to_user(&arg
[i
], ®
, sizeof(reg
))) {
11785 fput(tctx
->registered_rings
[reg
.offset
]);
11786 tctx
->registered_rings
[reg
.offset
] = NULL
;
11792 return i
? i
: ret
;
11795 static int io_ringfd_unregister(struct io_ring_ctx
*ctx
, void __user
*__arg
,
11798 struct io_uring_rsrc_update __user
*arg
= __arg
;
11799 struct io_uring_task
*tctx
= current
->io_uring
;
11800 struct io_uring_rsrc_update reg
;
11803 if (!nr_args
|| nr_args
> IO_RINGFD_REG_MAX
)
11808 for (i
= 0; i
< nr_args
; i
++) {
11809 if (copy_from_user(®
, &arg
[i
], sizeof(reg
))) {
11813 if (reg
.resv
|| reg
.data
|| reg
.offset
>= IO_RINGFD_REG_MAX
) {
11818 reg
.offset
= array_index_nospec(reg
.offset
, IO_RINGFD_REG_MAX
);
11819 if (tctx
->registered_rings
[reg
.offset
]) {
11820 fput(tctx
->registered_rings
[reg
.offset
]);
11821 tctx
->registered_rings
[reg
.offset
] = NULL
;
11825 return i
? i
: ret
;
11828 static void *io_uring_validate_mmap_request(struct file
*file
,
11829 loff_t pgoff
, size_t sz
)
11831 struct io_ring_ctx
*ctx
= file
->private_data
;
11832 loff_t offset
= pgoff
<< PAGE_SHIFT
;
11837 case IORING_OFF_SQ_RING
:
11838 case IORING_OFF_CQ_RING
:
11841 case IORING_OFF_SQES
:
11842 ptr
= ctx
->sq_sqes
;
11845 return ERR_PTR(-EINVAL
);
11848 page
= virt_to_head_page(ptr
);
11849 if (sz
> page_size(page
))
11850 return ERR_PTR(-EINVAL
);
11857 static __cold
int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
11859 size_t sz
= vma
->vm_end
- vma
->vm_start
;
11863 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
11865 return PTR_ERR(ptr
);
11867 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
11868 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
11871 #else /* !CONFIG_MMU */
11873 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
11875 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
11878 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
11880 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
11883 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
11884 unsigned long addr
, unsigned long len
,
11885 unsigned long pgoff
, unsigned long flags
)
11889 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
11891 return PTR_ERR(ptr
);
11893 return (unsigned long) ptr
;
11896 #endif /* !CONFIG_MMU */
11898 static int io_sqpoll_wait_sq(struct io_ring_ctx
*ctx
)
11903 if (!io_sqring_full(ctx
))
11905 prepare_to_wait(&ctx
->sqo_sq_wait
, &wait
, TASK_INTERRUPTIBLE
);
11907 if (!io_sqring_full(ctx
))
11910 } while (!signal_pending(current
));
11912 finish_wait(&ctx
->sqo_sq_wait
, &wait
);
11916 static int io_validate_ext_arg(unsigned flags
, const void __user
*argp
, size_t argsz
)
11918 if (flags
& IORING_ENTER_EXT_ARG
) {
11919 struct io_uring_getevents_arg arg
;
11921 if (argsz
!= sizeof(arg
))
11923 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
11929 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
11930 struct __kernel_timespec __user
**ts
,
11931 const sigset_t __user
**sig
)
11933 struct io_uring_getevents_arg arg
;
11936 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
11937 * is just a pointer to the sigset_t.
11939 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
11940 *sig
= (const sigset_t __user
*) argp
;
11946 * EXT_ARG is set - ensure we agree on the size of it and copy in our
11947 * timespec and sigset_t pointers if good.
11949 if (*argsz
!= sizeof(arg
))
11951 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
11955 *sig
= u64_to_user_ptr(arg
.sigmask
);
11956 *argsz
= arg
.sigmask_sz
;
11957 *ts
= u64_to_user_ptr(arg
.ts
);
11961 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
11962 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
11965 struct io_ring_ctx
*ctx
;
11969 io_run_task_work();
11971 if (unlikely(flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
11972 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
|
11973 IORING_ENTER_REGISTERED_RING
)))
11977 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
11978 * need only dereference our task private array to find it.
11980 if (flags
& IORING_ENTER_REGISTERED_RING
) {
11981 struct io_uring_task
*tctx
= current
->io_uring
;
11983 if (!tctx
|| fd
>= IO_RINGFD_REG_MAX
)
11985 fd
= array_index_nospec(fd
, IO_RINGFD_REG_MAX
);
11986 f
.file
= tctx
->registered_rings
[fd
];
11992 if (unlikely(!f
.file
))
11996 if (unlikely(f
.file
->f_op
!= &io_uring_fops
))
12000 ctx
= f
.file
->private_data
;
12001 if (unlikely(!percpu_ref_tryget(&ctx
->refs
)))
12005 if (unlikely(ctx
->flags
& IORING_SETUP_R_DISABLED
))
12009 * For SQ polling, the thread will do all submissions and completions.
12010 * Just return the requested submit count, and wake the thread if
12011 * we were asked to.
12014 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
12015 io_cqring_overflow_flush(ctx
);
12017 if (unlikely(ctx
->sq_data
->thread
== NULL
)) {
12021 if (flags
& IORING_ENTER_SQ_WAKEUP
)
12022 wake_up(&ctx
->sq_data
->wait
);
12023 if (flags
& IORING_ENTER_SQ_WAIT
) {
12024 ret
= io_sqpoll_wait_sq(ctx
);
12029 } else if (to_submit
) {
12030 ret
= io_uring_add_tctx_node(ctx
);
12034 mutex_lock(&ctx
->uring_lock
);
12035 ret
= io_submit_sqes(ctx
, to_submit
);
12036 if (ret
!= to_submit
) {
12037 mutex_unlock(&ctx
->uring_lock
);
12040 if ((flags
& IORING_ENTER_GETEVENTS
) && ctx
->syscall_iopoll
)
12041 goto iopoll_locked
;
12042 mutex_unlock(&ctx
->uring_lock
);
12044 if (flags
& IORING_ENTER_GETEVENTS
) {
12046 if (ctx
->syscall_iopoll
) {
12048 * We disallow the app entering submit/complete with
12049 * polling, but we still need to lock the ring to
12050 * prevent racing with polled issue that got punted to
12053 mutex_lock(&ctx
->uring_lock
);
12055 ret2
= io_validate_ext_arg(flags
, argp
, argsz
);
12056 if (likely(!ret2
)) {
12057 min_complete
= min(min_complete
,
12059 ret2
= io_iopoll_check(ctx
, min_complete
);
12061 mutex_unlock(&ctx
->uring_lock
);
12063 const sigset_t __user
*sig
;
12064 struct __kernel_timespec __user
*ts
;
12066 ret2
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
12067 if (likely(!ret2
)) {
12068 min_complete
= min(min_complete
,
12070 ret2
= io_cqring_wait(ctx
, min_complete
, sig
,
12079 * EBADR indicates that one or more CQE were dropped.
12080 * Once the user has been informed we can clear the bit
12081 * as they are obviously ok with those drops.
12083 if (unlikely(ret2
== -EBADR
))
12084 clear_bit(IO_CHECK_CQ_DROPPED_BIT
,
12090 percpu_ref_put(&ctx
->refs
);
12096 #ifdef CONFIG_PROC_FS
12097 static __cold
int io_uring_show_cred(struct seq_file
*m
, unsigned int id
,
12098 const struct cred
*cred
)
12100 struct user_namespace
*uns
= seq_user_ns(m
);
12101 struct group_info
*gi
;
12106 seq_printf(m
, "%5d\n", id
);
12107 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
12108 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
12109 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
12110 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
12111 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
12112 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
12113 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
12114 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
12115 seq_puts(m
, "\n\tGroups:\t");
12116 gi
= cred
->group_info
;
12117 for (g
= 0; g
< gi
->ngroups
; g
++) {
12118 seq_put_decimal_ull(m
, g
? " " : "",
12119 from_kgid_munged(uns
, gi
->gid
[g
]));
12121 seq_puts(m
, "\n\tCapEff:\t");
12122 cap
= cred
->cap_effective
;
12123 CAP_FOR_EACH_U32(__capi
)
12124 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
12129 static __cold
void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
,
12130 struct seq_file
*m
)
12132 struct io_sq_data
*sq
= NULL
;
12133 struct io_overflow_cqe
*ocqe
;
12134 struct io_rings
*r
= ctx
->rings
;
12135 unsigned int sq_mask
= ctx
->sq_entries
- 1, cq_mask
= ctx
->cq_entries
- 1;
12136 unsigned int sq_head
= READ_ONCE(r
->sq
.head
);
12137 unsigned int sq_tail
= READ_ONCE(r
->sq
.tail
);
12138 unsigned int cq_head
= READ_ONCE(r
->cq
.head
);
12139 unsigned int cq_tail
= READ_ONCE(r
->cq
.tail
);
12140 unsigned int cq_shift
= 0;
12141 unsigned int sq_entries
, cq_entries
;
12143 bool is_cqe32
= (ctx
->flags
& IORING_SETUP_CQE32
);
12150 * we may get imprecise sqe and cqe info if uring is actively running
12151 * since we get cached_sq_head and cached_cq_tail without uring_lock
12152 * and sq_tail and cq_head are changed by userspace. But it's ok since
12153 * we usually use these info when it is stuck.
12155 seq_printf(m
, "SqMask:\t0x%x\n", sq_mask
);
12156 seq_printf(m
, "SqHead:\t%u\n", sq_head
);
12157 seq_printf(m
, "SqTail:\t%u\n", sq_tail
);
12158 seq_printf(m
, "CachedSqHead:\t%u\n", ctx
->cached_sq_head
);
12159 seq_printf(m
, "CqMask:\t0x%x\n", cq_mask
);
12160 seq_printf(m
, "CqHead:\t%u\n", cq_head
);
12161 seq_printf(m
, "CqTail:\t%u\n", cq_tail
);
12162 seq_printf(m
, "CachedCqTail:\t%u\n", ctx
->cached_cq_tail
);
12163 seq_printf(m
, "SQEs:\t%u\n", sq_tail
- ctx
->cached_sq_head
);
12164 sq_entries
= min(sq_tail
- sq_head
, ctx
->sq_entries
);
12165 for (i
= 0; i
< sq_entries
; i
++) {
12166 unsigned int entry
= i
+ sq_head
;
12167 unsigned int sq_idx
= READ_ONCE(ctx
->sq_array
[entry
& sq_mask
]);
12168 struct io_uring_sqe
*sqe
;
12170 if (sq_idx
> sq_mask
)
12172 sqe
= &ctx
->sq_sqes
[sq_idx
];
12173 seq_printf(m
, "%5u: opcode:%d, fd:%d, flags:%x, user_data:%llu\n",
12174 sq_idx
, sqe
->opcode
, sqe
->fd
, sqe
->flags
,
12177 seq_printf(m
, "CQEs:\t%u\n", cq_tail
- cq_head
);
12178 cq_entries
= min(cq_tail
- cq_head
, ctx
->cq_entries
);
12179 for (i
= 0; i
< cq_entries
; i
++) {
12180 unsigned int entry
= i
+ cq_head
;
12181 struct io_uring_cqe
*cqe
= &r
->cqes
[(entry
& cq_mask
) << cq_shift
];
12184 seq_printf(m
, "%5u: user_data:%llu, res:%d, flag:%x\n",
12185 entry
& cq_mask
, cqe
->user_data
, cqe
->res
,
12188 seq_printf(m
, "%5u: user_data:%llu, res:%d, flag:%x, "
12189 "extra1:%llu, extra2:%llu\n",
12190 entry
& cq_mask
, cqe
->user_data
, cqe
->res
,
12191 cqe
->flags
, cqe
->big_cqe
[0], cqe
->big_cqe
[1]);
12196 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
12197 * since fdinfo case grabs it in the opposite direction of normal use
12198 * cases. If we fail to get the lock, we just don't iterate any
12199 * structures that could be going away outside the io_uring mutex.
12201 has_lock
= mutex_trylock(&ctx
->uring_lock
);
12203 if (has_lock
&& (ctx
->flags
& IORING_SETUP_SQPOLL
)) {
12209 seq_printf(m
, "SqThread:\t%d\n", sq
? task_pid_nr(sq
->thread
) : -1);
12210 seq_printf(m
, "SqThreadCpu:\t%d\n", sq
? task_cpu(sq
->thread
) : -1);
12211 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
12212 for (i
= 0; has_lock
&& i
< ctx
->nr_user_files
; i
++) {
12213 struct file
*f
= io_file_from_index(ctx
, i
);
12216 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
12218 seq_printf(m
, "%5u: <none>\n", i
);
12220 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
12221 for (i
= 0; has_lock
&& i
< ctx
->nr_user_bufs
; i
++) {
12222 struct io_mapped_ubuf
*buf
= ctx
->user_bufs
[i
];
12223 unsigned int len
= buf
->ubuf_end
- buf
->ubuf
;
12225 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
, len
);
12227 if (has_lock
&& !xa_empty(&ctx
->personalities
)) {
12228 unsigned long index
;
12229 const struct cred
*cred
;
12231 seq_printf(m
, "Personalities:\n");
12232 xa_for_each(&ctx
->personalities
, index
, cred
)
12233 io_uring_show_cred(m
, index
, cred
);
12236 mutex_unlock(&ctx
->uring_lock
);
12238 seq_puts(m
, "PollList:\n");
12239 spin_lock(&ctx
->completion_lock
);
12240 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
12241 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
12242 struct io_kiocb
*req
;
12244 hlist_for_each_entry(req
, list
, hash_node
)
12245 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
12246 task_work_pending(req
->task
));
12249 seq_puts(m
, "CqOverflowList:\n");
12250 list_for_each_entry(ocqe
, &ctx
->cq_overflow_list
, list
) {
12251 struct io_uring_cqe
*cqe
= &ocqe
->cqe
;
12253 seq_printf(m
, " user_data=%llu, res=%d, flags=%x\n",
12254 cqe
->user_data
, cqe
->res
, cqe
->flags
);
12258 spin_unlock(&ctx
->completion_lock
);
12261 static __cold
void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
12263 struct io_ring_ctx
*ctx
= f
->private_data
;
12265 if (percpu_ref_tryget(&ctx
->refs
)) {
12266 __io_uring_show_fdinfo(ctx
, m
);
12267 percpu_ref_put(&ctx
->refs
);
12272 static const struct file_operations io_uring_fops
= {
12273 .release
= io_uring_release
,
12274 .mmap
= io_uring_mmap
,
12276 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
12277 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
12279 .poll
= io_uring_poll
,
12280 #ifdef CONFIG_PROC_FS
12281 .show_fdinfo
= io_uring_show_fdinfo
,
12285 static __cold
int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
12286 struct io_uring_params
*p
)
12288 struct io_rings
*rings
;
12289 size_t size
, sq_array_offset
;
12291 /* make sure these are sane, as we already accounted them */
12292 ctx
->sq_entries
= p
->sq_entries
;
12293 ctx
->cq_entries
= p
->cq_entries
;
12295 size
= rings_size(ctx
, p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
12296 if (size
== SIZE_MAX
)
12299 rings
= io_mem_alloc(size
);
12303 ctx
->rings
= rings
;
12304 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
12305 rings
->sq_ring_mask
= p
->sq_entries
- 1;
12306 rings
->cq_ring_mask
= p
->cq_entries
- 1;
12307 rings
->sq_ring_entries
= p
->sq_entries
;
12308 rings
->cq_ring_entries
= p
->cq_entries
;
12310 if (p
->flags
& IORING_SETUP_SQE128
)
12311 size
= array_size(2 * sizeof(struct io_uring_sqe
), p
->sq_entries
);
12313 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
12314 if (size
== SIZE_MAX
) {
12315 io_mem_free(ctx
->rings
);
12320 ctx
->sq_sqes
= io_mem_alloc(size
);
12321 if (!ctx
->sq_sqes
) {
12322 io_mem_free(ctx
->rings
);
12330 static int io_uring_install_fd(struct io_ring_ctx
*ctx
, struct file
*file
)
12334 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
12338 ret
= io_uring_add_tctx_node(ctx
);
12343 fd_install(fd
, file
);
12348 * Allocate an anonymous fd, this is what constitutes the application
12349 * visible backing of an io_uring instance. The application mmaps this
12350 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
12351 * we have to tie this fd to a socket for file garbage collection purposes.
12353 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
12356 #if defined(CONFIG_UNIX)
12359 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
12362 return ERR_PTR(ret
);
12365 file
= anon_inode_getfile_secure("[io_uring]", &io_uring_fops
, ctx
,
12366 O_RDWR
| O_CLOEXEC
, NULL
);
12367 #if defined(CONFIG_UNIX)
12368 if (IS_ERR(file
)) {
12369 sock_release(ctx
->ring_sock
);
12370 ctx
->ring_sock
= NULL
;
12372 ctx
->ring_sock
->file
= file
;
12378 static __cold
int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
12379 struct io_uring_params __user
*params
)
12381 struct io_ring_ctx
*ctx
;
12387 if (entries
> IORING_MAX_ENTRIES
) {
12388 if (!(p
->flags
& IORING_SETUP_CLAMP
))
12390 entries
= IORING_MAX_ENTRIES
;
12394 * Use twice as many entries for the CQ ring. It's possible for the
12395 * application to drive a higher depth than the size of the SQ ring,
12396 * since the sqes are only used at submission time. This allows for
12397 * some flexibility in overcommitting a bit. If the application has
12398 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
12399 * of CQ ring entries manually.
12401 p
->sq_entries
= roundup_pow_of_two(entries
);
12402 if (p
->flags
& IORING_SETUP_CQSIZE
) {
12404 * If IORING_SETUP_CQSIZE is set, we do the same roundup
12405 * to a power-of-two, if it isn't already. We do NOT impose
12406 * any cq vs sq ring sizing.
12408 if (!p
->cq_entries
)
12410 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
12411 if (!(p
->flags
& IORING_SETUP_CLAMP
))
12413 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
12415 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
12416 if (p
->cq_entries
< p
->sq_entries
)
12419 p
->cq_entries
= 2 * p
->sq_entries
;
12422 ctx
= io_ring_ctx_alloc(p
);
12427 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
12428 * space applications don't need to do io completion events
12429 * polling again, they can rely on io_sq_thread to do polling
12430 * work, which can reduce cpu usage and uring_lock contention.
12432 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
12433 !(ctx
->flags
& IORING_SETUP_SQPOLL
))
12434 ctx
->syscall_iopoll
= 1;
12436 ctx
->compat
= in_compat_syscall();
12437 if (!capable(CAP_IPC_LOCK
))
12438 ctx
->user
= get_uid(current_user());
12441 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
12442 * COOP_TASKRUN is set, then IPIs are never needed by the app.
12445 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
12446 /* IPI related flags don't make sense with SQPOLL */
12447 if (ctx
->flags
& (IORING_SETUP_COOP_TASKRUN
|
12448 IORING_SETUP_TASKRUN_FLAG
))
12450 ctx
->notify_method
= TWA_SIGNAL_NO_IPI
;
12451 } else if (ctx
->flags
& IORING_SETUP_COOP_TASKRUN
) {
12452 ctx
->notify_method
= TWA_SIGNAL_NO_IPI
;
12454 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
12456 ctx
->notify_method
= TWA_SIGNAL
;
12460 * This is just grabbed for accounting purposes. When a process exits,
12461 * the mm is exited and dropped before the files, hence we need to hang
12462 * on to this mm purely for the purposes of being able to unaccount
12463 * memory (locked/pinned vm). It's not used for anything else.
12465 mmgrab(current
->mm
);
12466 ctx
->mm_account
= current
->mm
;
12468 ret
= io_allocate_scq_urings(ctx
, p
);
12472 ret
= io_sq_offload_create(ctx
, p
);
12475 /* always set a rsrc node */
12476 ret
= io_rsrc_node_switch_start(ctx
);
12479 io_rsrc_node_switch(ctx
, NULL
);
12481 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
12482 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
12483 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
12484 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
12485 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
12486 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
12487 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
12488 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
12490 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
12491 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
12492 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
12493 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
12494 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
12495 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
12496 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
12497 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
12499 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
12500 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
12501 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
12502 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
12503 IORING_FEAT_EXT_ARG
| IORING_FEAT_NATIVE_WORKERS
|
12504 IORING_FEAT_RSRC_TAGS
| IORING_FEAT_CQE_SKIP
|
12505 IORING_FEAT_LINKED_FILE
;
12507 if (copy_to_user(params
, p
, sizeof(*p
))) {
12512 file
= io_uring_get_file(ctx
);
12513 if (IS_ERR(file
)) {
12514 ret
= PTR_ERR(file
);
12519 * Install ring fd as the very last thing, so we don't risk someone
12520 * having closed it before we finish setup
12522 ret
= io_uring_install_fd(ctx
, file
);
12524 /* fput will clean it up */
12529 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
12532 io_ring_ctx_wait_and_kill(ctx
);
12537 * Sets up an aio uring context, and returns the fd. Applications asks for a
12538 * ring size, we return the actual sq/cq ring sizes (among other things) in the
12539 * params structure passed in.
12541 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
12543 struct io_uring_params p
;
12546 if (copy_from_user(&p
, params
, sizeof(p
)))
12548 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
12553 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
12554 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
12555 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
12556 IORING_SETUP_R_DISABLED
| IORING_SETUP_SUBMIT_ALL
|
12557 IORING_SETUP_COOP_TASKRUN
| IORING_SETUP_TASKRUN_FLAG
|
12558 IORING_SETUP_SQE128
| IORING_SETUP_CQE32
))
12561 return io_uring_create(entries
, &p
, params
);
12564 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
12565 struct io_uring_params __user
*, params
)
12567 return io_uring_setup(entries
, params
);
12570 static __cold
int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
,
12573 struct io_uring_probe
*p
;
12577 size
= struct_size(p
, ops
, nr_args
);
12578 if (size
== SIZE_MAX
)
12580 p
= kzalloc(size
, GFP_KERNEL
);
12585 if (copy_from_user(p
, arg
, size
))
12588 if (memchr_inv(p
, 0, size
))
12591 p
->last_op
= IORING_OP_LAST
- 1;
12592 if (nr_args
> IORING_OP_LAST
)
12593 nr_args
= IORING_OP_LAST
;
12595 for (i
= 0; i
< nr_args
; i
++) {
12597 if (!io_op_defs
[i
].not_supported
)
12598 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
12603 if (copy_to_user(arg
, p
, size
))
12610 static int io_register_personality(struct io_ring_ctx
*ctx
)
12612 const struct cred
*creds
;
12616 creds
= get_current_cred();
12618 ret
= xa_alloc_cyclic(&ctx
->personalities
, &id
, (void *)creds
,
12619 XA_LIMIT(0, USHRT_MAX
), &ctx
->pers_next
, GFP_KERNEL
);
12627 static __cold
int io_register_restrictions(struct io_ring_ctx
*ctx
,
12628 void __user
*arg
, unsigned int nr_args
)
12630 struct io_uring_restriction
*res
;
12634 /* Restrictions allowed only if rings started disabled */
12635 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
12638 /* We allow only a single restrictions registration */
12639 if (ctx
->restrictions
.registered
)
12642 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
12645 size
= array_size(nr_args
, sizeof(*res
));
12646 if (size
== SIZE_MAX
)
12649 res
= memdup_user(arg
, size
);
12651 return PTR_ERR(res
);
12655 for (i
= 0; i
< nr_args
; i
++) {
12656 switch (res
[i
].opcode
) {
12657 case IORING_RESTRICTION_REGISTER_OP
:
12658 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
12663 __set_bit(res
[i
].register_op
,
12664 ctx
->restrictions
.register_op
);
12666 case IORING_RESTRICTION_SQE_OP
:
12667 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
12672 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
12674 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
12675 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
12677 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
12678 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
12687 /* Reset all restrictions if an error happened */
12689 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
12691 ctx
->restrictions
.registered
= true;
12697 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
12699 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
12702 if (ctx
->restrictions
.registered
)
12703 ctx
->restricted
= 1;
12705 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
12706 if (ctx
->sq_data
&& wq_has_sleeper(&ctx
->sq_data
->wait
))
12707 wake_up(&ctx
->sq_data
->wait
);
12711 static int __io_register_rsrc_update(struct io_ring_ctx
*ctx
, unsigned type
,
12712 struct io_uring_rsrc_update2
*up
,
12718 if (check_add_overflow(up
->offset
, nr_args
, &tmp
))
12720 err
= io_rsrc_node_switch_start(ctx
);
12725 case IORING_RSRC_FILE
:
12726 return __io_sqe_files_update(ctx
, up
, nr_args
);
12727 case IORING_RSRC_BUFFER
:
12728 return __io_sqe_buffers_update(ctx
, up
, nr_args
);
12733 static int io_register_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
12736 struct io_uring_rsrc_update2 up
;
12740 memset(&up
, 0, sizeof(up
));
12741 if (copy_from_user(&up
, arg
, sizeof(struct io_uring_rsrc_update
)))
12743 if (up
.resv
|| up
.resv2
)
12745 return __io_register_rsrc_update(ctx
, IORING_RSRC_FILE
, &up
, nr_args
);
12748 static int io_register_rsrc_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
12749 unsigned size
, unsigned type
)
12751 struct io_uring_rsrc_update2 up
;
12753 if (size
!= sizeof(up
))
12755 if (copy_from_user(&up
, arg
, sizeof(up
)))
12757 if (!up
.nr
|| up
.resv
|| up
.resv2
)
12759 return __io_register_rsrc_update(ctx
, type
, &up
, up
.nr
);
12762 static __cold
int io_register_rsrc(struct io_ring_ctx
*ctx
, void __user
*arg
,
12763 unsigned int size
, unsigned int type
)
12765 struct io_uring_rsrc_register rr
;
12767 /* keep it extendible */
12768 if (size
!= sizeof(rr
))
12771 memset(&rr
, 0, sizeof(rr
));
12772 if (copy_from_user(&rr
, arg
, size
))
12774 if (!rr
.nr
|| rr
.resv2
)
12776 if (rr
.flags
& ~IORING_RSRC_REGISTER_SPARSE
)
12780 case IORING_RSRC_FILE
:
12781 if (rr
.flags
& IORING_RSRC_REGISTER_SPARSE
&& rr
.data
)
12783 return io_sqe_files_register(ctx
, u64_to_user_ptr(rr
.data
),
12784 rr
.nr
, u64_to_user_ptr(rr
.tags
));
12785 case IORING_RSRC_BUFFER
:
12786 if (rr
.flags
& IORING_RSRC_REGISTER_SPARSE
&& rr
.data
)
12788 return io_sqe_buffers_register(ctx
, u64_to_user_ptr(rr
.data
),
12789 rr
.nr
, u64_to_user_ptr(rr
.tags
));
12794 static __cold
int io_register_iowq_aff(struct io_ring_ctx
*ctx
,
12795 void __user
*arg
, unsigned len
)
12797 struct io_uring_task
*tctx
= current
->io_uring
;
12798 cpumask_var_t new_mask
;
12801 if (!tctx
|| !tctx
->io_wq
)
12804 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
12807 cpumask_clear(new_mask
);
12808 if (len
> cpumask_size())
12809 len
= cpumask_size();
12811 if (in_compat_syscall()) {
12812 ret
= compat_get_bitmap(cpumask_bits(new_mask
),
12813 (const compat_ulong_t __user
*)arg
,
12814 len
* 8 /* CHAR_BIT */);
12816 ret
= copy_from_user(new_mask
, arg
, len
);
12820 free_cpumask_var(new_mask
);
12824 ret
= io_wq_cpu_affinity(tctx
->io_wq
, new_mask
);
12825 free_cpumask_var(new_mask
);
12829 static __cold
int io_unregister_iowq_aff(struct io_ring_ctx
*ctx
)
12831 struct io_uring_task
*tctx
= current
->io_uring
;
12833 if (!tctx
|| !tctx
->io_wq
)
12836 return io_wq_cpu_affinity(tctx
->io_wq
, NULL
);
12839 static __cold
int io_register_iowq_max_workers(struct io_ring_ctx
*ctx
,
12841 __must_hold(&ctx
->uring_lock
)
12843 struct io_tctx_node
*node
;
12844 struct io_uring_task
*tctx
= NULL
;
12845 struct io_sq_data
*sqd
= NULL
;
12846 __u32 new_count
[2];
12849 if (copy_from_user(new_count
, arg
, sizeof(new_count
)))
12851 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
12852 if (new_count
[i
] > INT_MAX
)
12855 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
12856 sqd
= ctx
->sq_data
;
12859 * Observe the correct sqd->lock -> ctx->uring_lock
12860 * ordering. Fine to drop uring_lock here, we hold
12861 * a ref to the ctx.
12863 refcount_inc(&sqd
->refs
);
12864 mutex_unlock(&ctx
->uring_lock
);
12865 mutex_lock(&sqd
->lock
);
12866 mutex_lock(&ctx
->uring_lock
);
12868 tctx
= sqd
->thread
->io_uring
;
12871 tctx
= current
->io_uring
;
12874 BUILD_BUG_ON(sizeof(new_count
) != sizeof(ctx
->iowq_limits
));
12876 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
12878 ctx
->iowq_limits
[i
] = new_count
[i
];
12879 ctx
->iowq_limits_set
= true;
12881 if (tctx
&& tctx
->io_wq
) {
12882 ret
= io_wq_max_workers(tctx
->io_wq
, new_count
);
12886 memset(new_count
, 0, sizeof(new_count
));
12890 mutex_unlock(&sqd
->lock
);
12891 io_put_sq_data(sqd
);
12894 if (copy_to_user(arg
, new_count
, sizeof(new_count
)))
12897 /* that's it for SQPOLL, only the SQPOLL task creates requests */
12901 /* now propagate the restriction to all registered users */
12902 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
12903 struct io_uring_task
*tctx
= node
->task
->io_uring
;
12905 if (WARN_ON_ONCE(!tctx
->io_wq
))
12908 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
12909 new_count
[i
] = ctx
->iowq_limits
[i
];
12910 /* ignore errors, it always returns zero anyway */
12911 (void)io_wq_max_workers(tctx
->io_wq
, new_count
);
12916 mutex_unlock(&sqd
->lock
);
12917 io_put_sq_data(sqd
);
12922 static int io_register_pbuf_ring(struct io_ring_ctx
*ctx
, void __user
*arg
)
12924 struct io_uring_buf_ring
*br
;
12925 struct io_uring_buf_reg reg
;
12926 struct io_buffer_list
*bl
;
12927 struct page
**pages
;
12930 if (copy_from_user(®
, arg
, sizeof(reg
)))
12933 if (reg
.pad
|| reg
.resv
[0] || reg
.resv
[1] || reg
.resv
[2])
12935 if (!reg
.ring_addr
)
12937 if (reg
.ring_addr
& ~PAGE_MASK
)
12939 if (!is_power_of_2(reg
.ring_entries
))
12942 /* cannot disambiguate full vs empty due to head/tail size */
12943 if (reg
.ring_entries
>= 65536)
12946 if (unlikely(reg
.bgid
< BGID_ARRAY
&& !ctx
->io_bl
)) {
12947 int ret
= io_init_bl_list(ctx
);
12952 bl
= io_buffer_get_list(ctx
, reg
.bgid
);
12954 /* if mapped buffer ring OR classic exists, don't allow */
12955 if (bl
->buf_nr_pages
|| !list_empty(&bl
->buf_list
))
12958 bl
= kzalloc(sizeof(*bl
), GFP_KERNEL
);
12963 pages
= io_pin_pages(reg
.ring_addr
,
12964 struct_size(br
, bufs
, reg
.ring_entries
),
12966 if (IS_ERR(pages
)) {
12968 return PTR_ERR(pages
);
12971 br
= page_address(pages
[0]);
12972 bl
->buf_pages
= pages
;
12973 bl
->buf_nr_pages
= nr_pages
;
12974 bl
->nr_entries
= reg
.ring_entries
;
12976 bl
->mask
= reg
.ring_entries
- 1;
12977 io_buffer_add_list(ctx
, bl
, reg
.bgid
);
12981 static int io_unregister_pbuf_ring(struct io_ring_ctx
*ctx
, void __user
*arg
)
12983 struct io_uring_buf_reg reg
;
12984 struct io_buffer_list
*bl
;
12986 if (copy_from_user(®
, arg
, sizeof(reg
)))
12988 if (reg
.pad
|| reg
.resv
[0] || reg
.resv
[1] || reg
.resv
[2])
12991 bl
= io_buffer_get_list(ctx
, reg
.bgid
);
12994 if (!bl
->buf_nr_pages
)
12997 __io_remove_buffers(ctx
, bl
, -1U);
12998 if (bl
->bgid
>= BGID_ARRAY
) {
12999 xa_erase(&ctx
->io_bl_xa
, bl
->bgid
);
13005 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
13006 void __user
*arg
, unsigned nr_args
)
13007 __releases(ctx
->uring_lock
)
13008 __acquires(ctx
->uring_lock
)
13013 * We're inside the ring mutex, if the ref is already dying, then
13014 * someone else killed the ctx or is already going through
13015 * io_uring_register().
13017 if (percpu_ref_is_dying(&ctx
->refs
))
13020 if (ctx
->restricted
) {
13021 if (opcode
>= IORING_REGISTER_LAST
)
13023 opcode
= array_index_nospec(opcode
, IORING_REGISTER_LAST
);
13024 if (!test_bit(opcode
, ctx
->restrictions
.register_op
))
13029 case IORING_REGISTER_BUFFERS
:
13033 ret
= io_sqe_buffers_register(ctx
, arg
, nr_args
, NULL
);
13035 case IORING_UNREGISTER_BUFFERS
:
13037 if (arg
|| nr_args
)
13039 ret
= io_sqe_buffers_unregister(ctx
);
13041 case IORING_REGISTER_FILES
:
13045 ret
= io_sqe_files_register(ctx
, arg
, nr_args
, NULL
);
13047 case IORING_UNREGISTER_FILES
:
13049 if (arg
|| nr_args
)
13051 ret
= io_sqe_files_unregister(ctx
);
13053 case IORING_REGISTER_FILES_UPDATE
:
13054 ret
= io_register_files_update(ctx
, arg
, nr_args
);
13056 case IORING_REGISTER_EVENTFD
:
13060 ret
= io_eventfd_register(ctx
, arg
, 0);
13062 case IORING_REGISTER_EVENTFD_ASYNC
:
13066 ret
= io_eventfd_register(ctx
, arg
, 1);
13068 case IORING_UNREGISTER_EVENTFD
:
13070 if (arg
|| nr_args
)
13072 ret
= io_eventfd_unregister(ctx
);
13074 case IORING_REGISTER_PROBE
:
13076 if (!arg
|| nr_args
> 256)
13078 ret
= io_probe(ctx
, arg
, nr_args
);
13080 case IORING_REGISTER_PERSONALITY
:
13082 if (arg
|| nr_args
)
13084 ret
= io_register_personality(ctx
);
13086 case IORING_UNREGISTER_PERSONALITY
:
13090 ret
= io_unregister_personality(ctx
, nr_args
);
13092 case IORING_REGISTER_ENABLE_RINGS
:
13094 if (arg
|| nr_args
)
13096 ret
= io_register_enable_rings(ctx
);
13098 case IORING_REGISTER_RESTRICTIONS
:
13099 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
13101 case IORING_REGISTER_FILES2
:
13102 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_FILE
);
13104 case IORING_REGISTER_FILES_UPDATE2
:
13105 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
13108 case IORING_REGISTER_BUFFERS2
:
13109 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_BUFFER
);
13111 case IORING_REGISTER_BUFFERS_UPDATE
:
13112 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
13113 IORING_RSRC_BUFFER
);
13115 case IORING_REGISTER_IOWQ_AFF
:
13117 if (!arg
|| !nr_args
)
13119 ret
= io_register_iowq_aff(ctx
, arg
, nr_args
);
13121 case IORING_UNREGISTER_IOWQ_AFF
:
13123 if (arg
|| nr_args
)
13125 ret
= io_unregister_iowq_aff(ctx
);
13127 case IORING_REGISTER_IOWQ_MAX_WORKERS
:
13129 if (!arg
|| nr_args
!= 2)
13131 ret
= io_register_iowq_max_workers(ctx
, arg
);
13133 case IORING_REGISTER_RING_FDS
:
13134 ret
= io_ringfd_register(ctx
, arg
, nr_args
);
13136 case IORING_UNREGISTER_RING_FDS
:
13137 ret
= io_ringfd_unregister(ctx
, arg
, nr_args
);
13139 case IORING_REGISTER_PBUF_RING
:
13141 if (!arg
|| nr_args
!= 1)
13143 ret
= io_register_pbuf_ring(ctx
, arg
);
13145 case IORING_UNREGISTER_PBUF_RING
:
13147 if (!arg
|| nr_args
!= 1)
13149 ret
= io_unregister_pbuf_ring(ctx
, arg
);
13159 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
13160 void __user
*, arg
, unsigned int, nr_args
)
13162 struct io_ring_ctx
*ctx
;
13171 if (f
.file
->f_op
!= &io_uring_fops
)
13174 ctx
= f
.file
->private_data
;
13176 io_run_task_work();
13178 mutex_lock(&ctx
->uring_lock
);
13179 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
13180 mutex_unlock(&ctx
->uring_lock
);
13181 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
, ret
);
13187 static int __init
io_uring_init(void)
13189 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
13190 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
13191 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
13194 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
13195 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
13196 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
13197 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
13198 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
13199 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
13200 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
13201 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
13202 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
13203 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
13204 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
13205 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
13206 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
13207 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
13208 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
13209 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
13210 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
13211 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
13212 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
13213 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
13214 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
13215 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
13216 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
13217 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
13218 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
13219 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
13220 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
13221 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
13222 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
13223 BUILD_BUG_SQE_ELEM(40, __u16
, buf_group
);
13224 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
13225 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
13226 BUILD_BUG_SQE_ELEM(44, __u32
, file_index
);
13227 BUILD_BUG_SQE_ELEM(48, __u64
, addr3
);
13229 BUILD_BUG_ON(sizeof(struct io_uring_files_update
) !=
13230 sizeof(struct io_uring_rsrc_update
));
13231 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update
) >
13232 sizeof(struct io_uring_rsrc_update2
));
13234 /* ->buf_index is u16 */
13235 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS
>= (1u << 16));
13236 BUILD_BUG_ON(BGID_ARRAY
* sizeof(struct io_buffer_list
) > PAGE_SIZE
);
13237 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring
, bufs
) != 0);
13238 BUILD_BUG_ON(offsetof(struct io_uring_buf
, resv
) !=
13239 offsetof(struct io_uring_buf_ring
, tail
));
13241 /* should fit into one byte */
13242 BUILD_BUG_ON(SQE_VALID_FLAGS
>= (1 << 8));
13243 BUILD_BUG_ON(SQE_COMMON_FLAGS
>= (1 << 8));
13244 BUILD_BUG_ON((SQE_VALID_FLAGS
| SQE_COMMON_FLAGS
) != SQE_VALID_FLAGS
);
13246 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
13247 BUILD_BUG_ON(__REQ_F_LAST_BIT
> 8 * sizeof(int));
13249 BUILD_BUG_ON(sizeof(atomic_t
) != sizeof(u32
));
13251 BUILD_BUG_ON(sizeof(struct io_uring_cmd
) > 64);
13253 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
|
13257 __initcall(io_uring_init
);