1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
81 #include <linux/tracehook.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
93 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
95 /* 512 entries per page on 64-bit archs, 64 pages max */
96 #define IORING_MAX_FIXED_FILES (1U << 15)
97 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
98 IORING_REGISTER_LAST + IORING_OP_LAST)
100 #define IO_RSRC_TAG_TABLE_SHIFT 9
101 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
102 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
104 #define IORING_MAX_REG_BUFFERS (1U << 14)
106 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
107 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
109 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
112 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
115 u32 head ____cacheline_aligned_in_smp;
116 u32 tail ____cacheline_aligned_in_smp;
120 * This data is shared with the application through the mmap at offsets
121 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
123 * The offsets to the member fields are published through struct
124 * io_sqring_offsets when calling io_uring_setup.
128 * Head and tail offsets into the ring; the offsets need to be
129 * masked to get valid indices.
131 * The kernel controls head of the sq ring and the tail of the cq ring,
132 * and the application controls tail of the sq ring and the head of the
135 struct io_uring sq, cq;
137 * Bitmasks to apply to head and tail offsets (constant, equals
140 u32 sq_ring_mask, cq_ring_mask;
141 /* Ring sizes (constant, power of 2) */
142 u32 sq_ring_entries, cq_ring_entries;
144 * Number of invalid entries dropped by the kernel due to
145 * invalid index stored in array
147 * Written by the kernel, shouldn't be modified by the
148 * application (i.e. get number of "new events" by comparing to
151 * After a new SQ head value was read by the application this
152 * counter includes all submissions that were dropped reaching
153 * the new SQ head (and possibly more).
159 * Written by the kernel, shouldn't be modified by the
162 * The application needs a full memory barrier before checking
163 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
169 * Written by the application, shouldn't be modified by the
174 * Number of completion events lost because the queue was full;
175 * this should be avoided by the application by making sure
176 * there are not more requests pending than there is space in
177 * the completion queue.
179 * Written by the kernel, shouldn't be modified by the
180 * application (i.e. get number of "new events" by comparing to
183 * As completion events come in out of order this counter is not
184 * ordered with any other data.
188 * Ring buffer of completion events.
190 * The kernel writes completion events fresh every time they are
191 * produced, so the application is allowed to modify pending
194 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
197 enum io_uring_cmd_flags {
198 IO_URING_F_NONBLOCK = 1,
199 IO_URING_F_COMPLETE_DEFER = 2,
202 struct io_mapped_ubuf {
205 unsigned int nr_bvecs;
206 unsigned long acct_pages;
207 struct bio_vec bvec[];
212 struct io_overflow_cqe {
213 struct io_uring_cqe cqe;
214 struct list_head list;
217 struct io_fixed_file {
218 /* file * with additional FFS_* flags */
219 unsigned long file_ptr;
223 struct list_head list;
228 struct io_mapped_ubuf *buf;
232 struct io_file_table {
233 struct io_fixed_file *files;
236 struct io_rsrc_node {
237 struct percpu_ref refs;
238 struct list_head node;
239 struct list_head rsrc_list;
240 struct io_rsrc_data *rsrc_data;
241 struct llist_node llist;
245 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
247 struct io_rsrc_data {
248 struct io_ring_ctx *ctx;
254 struct completion done;
259 struct list_head list;
265 struct io_restriction {
266 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
267 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
268 u8 sqe_flags_allowed;
269 u8 sqe_flags_required;
274 IO_SQ_THREAD_SHOULD_STOP = 0,
275 IO_SQ_THREAD_SHOULD_PARK,
280 atomic_t park_pending;
283 /* ctx's that are using this sqd */
284 struct list_head ctx_list;
286 struct task_struct *thread;
287 struct wait_queue_head wait;
289 unsigned sq_thread_idle;
295 struct completion exited;
298 #define IO_COMPL_BATCH 32
299 #define IO_REQ_CACHE_SIZE 32
300 #define IO_REQ_ALLOC_BATCH 8
302 struct io_comp_state {
303 struct io_kiocb *reqs[IO_COMPL_BATCH];
305 /* inline/task_work completion list, under ->uring_lock */
306 struct list_head free_list;
309 struct io_submit_link {
310 struct io_kiocb *head;
311 struct io_kiocb *last;
314 struct io_submit_state {
315 struct blk_plug plug;
316 struct io_submit_link link;
319 * io_kiocb alloc cache
321 void *reqs[IO_REQ_CACHE_SIZE];
322 unsigned int free_reqs;
327 * Batch completion logic
329 struct io_comp_state comp;
332 * File reference cache
336 unsigned int file_refs;
337 unsigned int ios_left;
341 /* const or read-mostly hot data */
343 struct percpu_ref refs;
345 struct io_rings *rings;
347 unsigned int compat: 1;
348 unsigned int drain_next: 1;
349 unsigned int eventfd_async: 1;
350 unsigned int restricted: 1;
351 unsigned int off_timeout_used: 1;
352 unsigned int drain_active: 1;
353 } ____cacheline_aligned_in_smp;
355 /* submission data */
357 struct mutex uring_lock;
360 * Ring buffer of indices into array of io_uring_sqe, which is
361 * mmapped by the application using the IORING_OFF_SQES offset.
363 * This indirection could e.g. be used to assign fixed
364 * io_uring_sqe entries to operations and only submit them to
365 * the queue when needed.
367 * The kernel modifies neither the indices array nor the entries
371 struct io_uring_sqe *sq_sqes;
372 unsigned cached_sq_head;
374 struct list_head defer_list;
377 * Fixed resources fast path, should be accessed only under
378 * uring_lock, and updated through io_uring_register(2)
380 struct io_rsrc_node *rsrc_node;
381 struct io_file_table file_table;
382 unsigned nr_user_files;
383 unsigned nr_user_bufs;
384 struct io_mapped_ubuf **user_bufs;
386 struct io_submit_state submit_state;
387 struct list_head timeout_list;
388 struct list_head cq_overflow_list;
389 struct xarray io_buffers;
390 struct xarray personalities;
392 unsigned sq_thread_idle;
393 } ____cacheline_aligned_in_smp;
395 /* IRQ completion list, under ->completion_lock */
396 struct list_head locked_free_list;
397 unsigned int locked_free_nr;
399 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
400 struct io_sq_data *sq_data; /* if using sq thread polling */
402 struct wait_queue_head sqo_sq_wait;
403 struct list_head sqd_list;
405 unsigned long check_cq_overflow;
408 unsigned cached_cq_tail;
410 struct eventfd_ctx *cq_ev_fd;
411 struct wait_queue_head poll_wait;
412 struct wait_queue_head cq_wait;
414 atomic_t cq_timeouts;
415 struct fasync_struct *cq_fasync;
416 unsigned cq_last_tm_flush;
417 } ____cacheline_aligned_in_smp;
420 spinlock_t completion_lock;
423 * ->iopoll_list is protected by the ctx->uring_lock for
424 * io_uring instances that don't use IORING_SETUP_SQPOLL.
425 * For SQPOLL, only the single threaded io_sq_thread() will
426 * manipulate the list, hence no extra locking is needed there.
428 struct list_head iopoll_list;
429 struct hlist_head *cancel_hash;
430 unsigned cancel_hash_bits;
431 bool poll_multi_queue;
432 } ____cacheline_aligned_in_smp;
434 struct io_restriction restrictions;
436 /* slow path rsrc auxilary data, used by update/register */
438 struct io_rsrc_node *rsrc_backup_node;
439 struct io_mapped_ubuf *dummy_ubuf;
440 struct io_rsrc_data *file_data;
441 struct io_rsrc_data *buf_data;
443 struct delayed_work rsrc_put_work;
444 struct llist_head rsrc_put_llist;
445 struct list_head rsrc_ref_list;
446 spinlock_t rsrc_ref_lock;
449 /* Keep this last, we don't need it for the fast path */
451 #if defined(CONFIG_UNIX)
452 struct socket *ring_sock;
454 /* hashed buffered write serialization */
455 struct io_wq_hash *hash_map;
457 /* Only used for accounting purposes */
458 struct user_struct *user;
459 struct mm_struct *mm_account;
461 /* ctx exit and cancelation */
462 struct llist_head fallback_llist;
463 struct delayed_work fallback_work;
464 struct work_struct exit_work;
465 struct list_head tctx_list;
466 struct completion ref_comp;
470 struct io_uring_task {
471 /* submission side */
474 struct wait_queue_head wait;
475 const struct io_ring_ctx *last;
477 struct percpu_counter inflight;
478 atomic_t inflight_tracked;
481 spinlock_t task_lock;
482 struct io_wq_work_list task_list;
483 unsigned long task_state;
484 struct callback_head task_work;
488 * First field must be the file pointer in all the
489 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
491 struct io_poll_iocb {
493 struct wait_queue_head *head;
497 struct wait_queue_entry wait;
500 struct io_poll_update {
506 bool update_user_data;
514 struct io_timeout_data {
515 struct io_kiocb *req;
516 struct hrtimer timer;
517 struct timespec64 ts;
518 enum hrtimer_mode mode;
523 struct sockaddr __user *addr;
524 int __user *addr_len;
526 unsigned long nofile;
546 struct list_head list;
547 /* head of the link, used by linked timeouts only */
548 struct io_kiocb *head;
551 struct io_timeout_rem {
556 struct timespec64 ts;
561 /* NOTE: kiocb has the file as the first member, so don't do it here */
569 struct sockaddr __user *addr;
576 struct compat_msghdr __user *umsg_compat;
577 struct user_msghdr __user *umsg;
583 struct io_buffer *kbuf;
589 struct filename *filename;
591 unsigned long nofile;
594 struct io_rsrc_update {
620 struct epoll_event event;
624 struct file *file_out;
625 struct file *file_in;
632 struct io_provide_buf {
646 const char __user *filename;
647 struct statx __user *buffer;
659 struct filename *oldpath;
660 struct filename *newpath;
668 struct filename *filename;
671 struct io_completion {
673 struct list_head list;
677 struct io_async_connect {
678 struct sockaddr_storage address;
681 struct io_async_msghdr {
682 struct iovec fast_iov[UIO_FASTIOV];
683 /* points to an allocated iov, if NULL we use fast_iov instead */
684 struct iovec *free_iov;
685 struct sockaddr __user *uaddr;
687 struct sockaddr_storage addr;
691 struct iovec fast_iov[UIO_FASTIOV];
692 const struct iovec *free_iovec;
693 struct iov_iter iter;
695 struct wait_page_queue wpq;
699 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
700 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
701 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
702 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
703 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
704 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
706 /* first byte is taken by user flags, shift it to not overlap */
711 REQ_F_LINK_TIMEOUT_BIT,
712 REQ_F_NEED_CLEANUP_BIT,
714 REQ_F_BUFFER_SELECTED_BIT,
715 REQ_F_LTIMEOUT_ACTIVE_BIT,
716 REQ_F_COMPLETE_INLINE_BIT,
718 REQ_F_DONT_REISSUE_BIT,
720 /* keep async read/write and isreg together and in order */
721 REQ_F_NOWAIT_READ_BIT,
722 REQ_F_NOWAIT_WRITE_BIT,
725 /* not a real bit, just to check we're not overflowing the space */
731 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
732 /* drain existing IO first */
733 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
735 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
736 /* doesn't sever on completion < 0 */
737 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
739 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
740 /* IOSQE_BUFFER_SELECT */
741 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
743 /* fail rest of links */
744 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
745 /* on inflight list, should be cancelled and waited on exit reliably */
746 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
747 /* read/write uses file position */
748 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
749 /* must not punt to workers */
750 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
751 /* has or had linked timeout */
752 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
754 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
755 /* already went through poll handler */
756 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
757 /* buffer already selected */
758 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
759 /* linked timeout is active, i.e. prepared by link's head */
760 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
761 /* completion is deferred through io_comp_state */
762 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
763 /* caller should reissue async */
764 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
765 /* don't attempt request reissue, see io_rw_reissue() */
766 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
767 /* supports async reads */
768 REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT),
769 /* supports async writes */
770 REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT),
772 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
773 /* has creds assigned */
774 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
778 struct io_poll_iocb poll;
779 struct io_poll_iocb *double_poll;
782 typedef void (*io_req_tw_func_t)(struct io_kiocb *req);
784 struct io_task_work {
786 struct io_wq_work_node node;
787 struct llist_node fallback_node;
789 io_req_tw_func_t func;
793 IORING_RSRC_FILE = 0,
794 IORING_RSRC_BUFFER = 1,
798 * NOTE! Each of the iocb union members has the file pointer
799 * as the first entry in their struct definition. So you can
800 * access the file pointer through any of the sub-structs,
801 * or directly as just 'ki_filp' in this struct.
807 struct io_poll_iocb poll;
808 struct io_poll_update poll_update;
809 struct io_accept accept;
811 struct io_cancel cancel;
812 struct io_timeout timeout;
813 struct io_timeout_rem timeout_rem;
814 struct io_connect connect;
815 struct io_sr_msg sr_msg;
817 struct io_close close;
818 struct io_rsrc_update rsrc_update;
819 struct io_fadvise fadvise;
820 struct io_madvise madvise;
821 struct io_epoll epoll;
822 struct io_splice splice;
823 struct io_provide_buf pbuf;
824 struct io_statx statx;
825 struct io_shutdown shutdown;
826 struct io_rename rename;
827 struct io_unlink unlink;
828 /* use only after cleaning per-op data, see io_clean_op() */
829 struct io_completion compl;
832 /* opcode allocated if it needs to store data for async defer */
835 /* polled IO has completed */
841 struct io_ring_ctx *ctx;
844 struct task_struct *task;
847 struct io_kiocb *link;
848 struct percpu_ref *fixed_rsrc_refs;
850 /* used with ctx->iopoll_list with reads/writes */
851 struct list_head inflight_entry;
852 struct io_task_work io_task_work;
853 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
854 struct hlist_node hash_node;
855 struct async_poll *apoll;
856 struct io_wq_work work;
857 const struct cred *creds;
859 /* store used ubuf, so we can prevent reloading */
860 struct io_mapped_ubuf *imu;
863 struct io_tctx_node {
864 struct list_head ctx_node;
865 struct task_struct *task;
866 struct io_ring_ctx *ctx;
869 struct io_defer_entry {
870 struct list_head list;
871 struct io_kiocb *req;
876 /* needs req->file assigned */
877 unsigned needs_file : 1;
878 /* hash wq insertion if file is a regular file */
879 unsigned hash_reg_file : 1;
880 /* unbound wq insertion if file is a non-regular file */
881 unsigned unbound_nonreg_file : 1;
882 /* opcode is not supported by this kernel */
883 unsigned not_supported : 1;
884 /* set if opcode supports polled "wait" */
886 unsigned pollout : 1;
887 /* op supports buffer selection */
888 unsigned buffer_select : 1;
889 /* do prep async if is going to be punted */
890 unsigned needs_async_setup : 1;
891 /* should block plug */
893 /* size of async data needed, if any */
894 unsigned short async_size;
897 static const struct io_op_def io_op_defs[] = {
898 [IORING_OP_NOP] = {},
899 [IORING_OP_READV] = {
901 .unbound_nonreg_file = 1,
904 .needs_async_setup = 1,
906 .async_size = sizeof(struct io_async_rw),
908 [IORING_OP_WRITEV] = {
911 .unbound_nonreg_file = 1,
913 .needs_async_setup = 1,
915 .async_size = sizeof(struct io_async_rw),
917 [IORING_OP_FSYNC] = {
920 [IORING_OP_READ_FIXED] = {
922 .unbound_nonreg_file = 1,
925 .async_size = sizeof(struct io_async_rw),
927 [IORING_OP_WRITE_FIXED] = {
930 .unbound_nonreg_file = 1,
933 .async_size = sizeof(struct io_async_rw),
935 [IORING_OP_POLL_ADD] = {
937 .unbound_nonreg_file = 1,
939 [IORING_OP_POLL_REMOVE] = {},
940 [IORING_OP_SYNC_FILE_RANGE] = {
943 [IORING_OP_SENDMSG] = {
945 .unbound_nonreg_file = 1,
947 .needs_async_setup = 1,
948 .async_size = sizeof(struct io_async_msghdr),
950 [IORING_OP_RECVMSG] = {
952 .unbound_nonreg_file = 1,
955 .needs_async_setup = 1,
956 .async_size = sizeof(struct io_async_msghdr),
958 [IORING_OP_TIMEOUT] = {
959 .async_size = sizeof(struct io_timeout_data),
961 [IORING_OP_TIMEOUT_REMOVE] = {
962 /* used by timeout updates' prep() */
964 [IORING_OP_ACCEPT] = {
966 .unbound_nonreg_file = 1,
969 [IORING_OP_ASYNC_CANCEL] = {},
970 [IORING_OP_LINK_TIMEOUT] = {
971 .async_size = sizeof(struct io_timeout_data),
973 [IORING_OP_CONNECT] = {
975 .unbound_nonreg_file = 1,
977 .needs_async_setup = 1,
978 .async_size = sizeof(struct io_async_connect),
980 [IORING_OP_FALLOCATE] = {
983 [IORING_OP_OPENAT] = {},
984 [IORING_OP_CLOSE] = {},
985 [IORING_OP_FILES_UPDATE] = {},
986 [IORING_OP_STATX] = {},
989 .unbound_nonreg_file = 1,
993 .async_size = sizeof(struct io_async_rw),
995 [IORING_OP_WRITE] = {
997 .unbound_nonreg_file = 1,
1000 .async_size = sizeof(struct io_async_rw),
1002 [IORING_OP_FADVISE] = {
1005 [IORING_OP_MADVISE] = {},
1006 [IORING_OP_SEND] = {
1008 .unbound_nonreg_file = 1,
1011 [IORING_OP_RECV] = {
1013 .unbound_nonreg_file = 1,
1017 [IORING_OP_OPENAT2] = {
1019 [IORING_OP_EPOLL_CTL] = {
1020 .unbound_nonreg_file = 1,
1022 [IORING_OP_SPLICE] = {
1025 .unbound_nonreg_file = 1,
1027 [IORING_OP_PROVIDE_BUFFERS] = {},
1028 [IORING_OP_REMOVE_BUFFERS] = {},
1032 .unbound_nonreg_file = 1,
1034 [IORING_OP_SHUTDOWN] = {
1037 [IORING_OP_RENAMEAT] = {},
1038 [IORING_OP_UNLINKAT] = {},
1041 static bool io_disarm_next(struct io_kiocb *req);
1042 static void io_uring_del_tctx_node(unsigned long index);
1043 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1044 struct task_struct *task,
1046 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1048 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1049 long res, unsigned int cflags);
1050 static void io_put_req(struct io_kiocb *req);
1051 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1052 static void io_dismantle_req(struct io_kiocb *req);
1053 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1054 static void io_queue_linked_timeout(struct io_kiocb *req);
1055 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1056 struct io_uring_rsrc_update2 *up,
1058 static void io_clean_op(struct io_kiocb *req);
1059 static struct file *io_file_get(struct io_ring_ctx *ctx,
1060 struct io_submit_state *state,
1061 struct io_kiocb *req, int fd, bool fixed);
1062 static void __io_queue_sqe(struct io_kiocb *req);
1063 static void io_rsrc_put_work(struct work_struct *work);
1065 static void io_req_task_queue(struct io_kiocb *req);
1066 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1067 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1068 static int io_req_prep_async(struct io_kiocb *req);
1070 static struct kmem_cache *req_cachep;
1072 static const struct file_operations io_uring_fops;
1074 struct sock *io_uring_get_socket(struct file *file)
1076 #if defined(CONFIG_UNIX)
1077 if (file->f_op == &io_uring_fops) {
1078 struct io_ring_ctx *ctx = file->private_data;
1080 return ctx->ring_sock->sk;
1085 EXPORT_SYMBOL(io_uring_get_socket);
1087 #define io_for_each_link(pos, head) \
1088 for (pos = (head); pos; pos = pos->link)
1090 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1092 struct io_ring_ctx *ctx = req->ctx;
1094 if (!req->fixed_rsrc_refs) {
1095 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1096 percpu_ref_get(req->fixed_rsrc_refs);
1100 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1102 bool got = percpu_ref_tryget(ref);
1104 /* already at zero, wait for ->release() */
1106 wait_for_completion(compl);
1107 percpu_ref_resurrect(ref);
1109 percpu_ref_put(ref);
1112 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1115 struct io_kiocb *req;
1117 if (task && head->task != task)
1122 io_for_each_link(req, head) {
1123 if (req->flags & REQ_F_INFLIGHT)
1129 static inline void req_set_fail(struct io_kiocb *req)
1131 req->flags |= REQ_F_FAIL;
1134 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1136 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1138 complete(&ctx->ref_comp);
1141 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1143 return !req->timeout.off;
1146 static void io_fallback_req_func(struct work_struct *work)
1148 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1149 fallback_work.work);
1150 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1151 struct io_kiocb *req, *tmp;
1153 percpu_ref_get(&ctx->refs);
1154 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1155 req->io_task_work.func(req);
1156 percpu_ref_put(&ctx->refs);
1159 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1161 struct io_ring_ctx *ctx;
1164 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1169 * Use 5 bits less than the max cq entries, that should give us around
1170 * 32 entries per hash list if totally full and uniformly spread.
1172 hash_bits = ilog2(p->cq_entries);
1176 ctx->cancel_hash_bits = hash_bits;
1177 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1179 if (!ctx->cancel_hash)
1181 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1183 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1184 if (!ctx->dummy_ubuf)
1186 /* set invalid range, so io_import_fixed() fails meeting it */
1187 ctx->dummy_ubuf->ubuf = -1UL;
1189 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1190 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1193 ctx->flags = p->flags;
1194 init_waitqueue_head(&ctx->sqo_sq_wait);
1195 INIT_LIST_HEAD(&ctx->sqd_list);
1196 init_waitqueue_head(&ctx->poll_wait);
1197 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1198 init_completion(&ctx->ref_comp);
1199 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1200 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1201 mutex_init(&ctx->uring_lock);
1202 init_waitqueue_head(&ctx->cq_wait);
1203 spin_lock_init(&ctx->completion_lock);
1204 INIT_LIST_HEAD(&ctx->iopoll_list);
1205 INIT_LIST_HEAD(&ctx->defer_list);
1206 INIT_LIST_HEAD(&ctx->timeout_list);
1207 spin_lock_init(&ctx->rsrc_ref_lock);
1208 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1209 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1210 init_llist_head(&ctx->rsrc_put_llist);
1211 INIT_LIST_HEAD(&ctx->tctx_list);
1212 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1213 INIT_LIST_HEAD(&ctx->locked_free_list);
1214 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1217 kfree(ctx->dummy_ubuf);
1218 kfree(ctx->cancel_hash);
1223 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1225 struct io_rings *r = ctx->rings;
1227 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1231 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1233 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1234 struct io_ring_ctx *ctx = req->ctx;
1236 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1242 #define FFS_ASYNC_READ 0x1UL
1243 #define FFS_ASYNC_WRITE 0x2UL
1245 #define FFS_ISREG 0x4UL
1247 #define FFS_ISREG 0x0UL
1249 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1251 static inline bool io_req_ffs_set(struct io_kiocb *req)
1253 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1256 static void io_req_track_inflight(struct io_kiocb *req)
1258 if (!(req->flags & REQ_F_INFLIGHT)) {
1259 req->flags |= REQ_F_INFLIGHT;
1260 atomic_inc(¤t->io_uring->inflight_tracked);
1264 static void io_prep_async_work(struct io_kiocb *req)
1266 const struct io_op_def *def = &io_op_defs[req->opcode];
1267 struct io_ring_ctx *ctx = req->ctx;
1269 if (!(req->flags & REQ_F_CREDS)) {
1270 req->flags |= REQ_F_CREDS;
1271 req->creds = get_current_cred();
1274 req->work.list.next = NULL;
1275 req->work.flags = 0;
1276 if (req->flags & REQ_F_FORCE_ASYNC)
1277 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1279 if (req->flags & REQ_F_ISREG) {
1280 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1281 io_wq_hash_work(&req->work, file_inode(req->file));
1282 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1283 if (def->unbound_nonreg_file)
1284 req->work.flags |= IO_WQ_WORK_UNBOUND;
1287 switch (req->opcode) {
1288 case IORING_OP_SPLICE:
1290 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1291 req->work.flags |= IO_WQ_WORK_UNBOUND;
1296 static void io_prep_async_link(struct io_kiocb *req)
1298 struct io_kiocb *cur;
1300 if (req->flags & REQ_F_LINK_TIMEOUT) {
1301 struct io_ring_ctx *ctx = req->ctx;
1303 spin_lock_irq(&ctx->completion_lock);
1304 io_for_each_link(cur, req)
1305 io_prep_async_work(cur);
1306 spin_unlock_irq(&ctx->completion_lock);
1308 io_for_each_link(cur, req)
1309 io_prep_async_work(cur);
1313 static void io_queue_async_work(struct io_kiocb *req)
1315 struct io_ring_ctx *ctx = req->ctx;
1316 struct io_kiocb *link = io_prep_linked_timeout(req);
1317 struct io_uring_task *tctx = req->task->io_uring;
1320 BUG_ON(!tctx->io_wq);
1322 /* init ->work of the whole link before punting */
1323 io_prep_async_link(req);
1326 * Not expected to happen, but if we do have a bug where this _can_
1327 * happen, catch it here and ensure the request is marked as
1328 * canceled. That will make io-wq go through the usual work cancel
1329 * procedure rather than attempt to run this request (or create a new
1332 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1333 req->work.flags |= IO_WQ_WORK_CANCEL;
1335 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1336 &req->work, req->flags);
1337 io_wq_enqueue(tctx->io_wq, &req->work);
1339 io_queue_linked_timeout(link);
1342 static void io_kill_timeout(struct io_kiocb *req, int status)
1343 __must_hold(&req->ctx->completion_lock)
1345 struct io_timeout_data *io = req->async_data;
1347 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1348 atomic_set(&req->ctx->cq_timeouts,
1349 atomic_read(&req->ctx->cq_timeouts) + 1);
1350 list_del_init(&req->timeout.list);
1351 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1352 io_put_req_deferred(req, 1);
1356 static void io_queue_deferred(struct io_ring_ctx *ctx)
1358 while (!list_empty(&ctx->defer_list)) {
1359 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1360 struct io_defer_entry, list);
1362 if (req_need_defer(de->req, de->seq))
1364 list_del_init(&de->list);
1365 io_req_task_queue(de->req);
1370 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1372 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1374 while (!list_empty(&ctx->timeout_list)) {
1375 u32 events_needed, events_got;
1376 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1377 struct io_kiocb, timeout.list);
1379 if (io_is_timeout_noseq(req))
1383 * Since seq can easily wrap around over time, subtract
1384 * the last seq at which timeouts were flushed before comparing.
1385 * Assuming not more than 2^31-1 events have happened since,
1386 * these subtractions won't have wrapped, so we can check if
1387 * target is in [last_seq, current_seq] by comparing the two.
1389 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1390 events_got = seq - ctx->cq_last_tm_flush;
1391 if (events_got < events_needed)
1394 list_del_init(&req->timeout.list);
1395 io_kill_timeout(req, 0);
1397 ctx->cq_last_tm_flush = seq;
1400 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1402 if (ctx->off_timeout_used)
1403 io_flush_timeouts(ctx);
1404 if (ctx->drain_active)
1405 io_queue_deferred(ctx);
1408 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1410 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1411 __io_commit_cqring_flush(ctx);
1412 /* order cqe stores with ring update */
1413 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1416 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1418 struct io_rings *r = ctx->rings;
1420 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1423 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1425 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1428 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1430 struct io_rings *rings = ctx->rings;
1431 unsigned tail, mask = ctx->cq_entries - 1;
1434 * writes to the cq entry need to come after reading head; the
1435 * control dependency is enough as we're using WRITE_ONCE to
1438 if (__io_cqring_events(ctx) == ctx->cq_entries)
1441 tail = ctx->cached_cq_tail++;
1442 return &rings->cqes[tail & mask];
1445 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1447 if (likely(!ctx->cq_ev_fd))
1449 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1451 return !ctx->eventfd_async || io_wq_current_is_worker();
1454 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1457 * wake_up_all() may seem excessive, but io_wake_function() and
1458 * io_should_wake() handle the termination of the loop and only
1459 * wake as many waiters as we need to.
1461 if (wq_has_sleeper(&ctx->cq_wait))
1462 wake_up_all(&ctx->cq_wait);
1463 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1464 wake_up(&ctx->sq_data->wait);
1465 if (io_should_trigger_evfd(ctx))
1466 eventfd_signal(ctx->cq_ev_fd, 1);
1467 if (waitqueue_active(&ctx->poll_wait)) {
1468 wake_up_interruptible(&ctx->poll_wait);
1469 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1473 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1475 if (ctx->flags & IORING_SETUP_SQPOLL) {
1476 if (wq_has_sleeper(&ctx->cq_wait))
1477 wake_up_all(&ctx->cq_wait);
1479 if (io_should_trigger_evfd(ctx))
1480 eventfd_signal(ctx->cq_ev_fd, 1);
1481 if (waitqueue_active(&ctx->poll_wait)) {
1482 wake_up_interruptible(&ctx->poll_wait);
1483 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1487 /* Returns true if there are no backlogged entries after the flush */
1488 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1490 unsigned long flags;
1491 bool all_flushed, posted;
1493 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1497 spin_lock_irqsave(&ctx->completion_lock, flags);
1498 while (!list_empty(&ctx->cq_overflow_list)) {
1499 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1500 struct io_overflow_cqe *ocqe;
1504 ocqe = list_first_entry(&ctx->cq_overflow_list,
1505 struct io_overflow_cqe, list);
1507 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1509 io_account_cq_overflow(ctx);
1512 list_del(&ocqe->list);
1516 all_flushed = list_empty(&ctx->cq_overflow_list);
1518 clear_bit(0, &ctx->check_cq_overflow);
1519 WRITE_ONCE(ctx->rings->sq_flags,
1520 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1524 io_commit_cqring(ctx);
1525 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1527 io_cqring_ev_posted(ctx);
1531 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1535 if (test_bit(0, &ctx->check_cq_overflow)) {
1536 /* iopoll syncs against uring_lock, not completion_lock */
1537 if (ctx->flags & IORING_SETUP_IOPOLL)
1538 mutex_lock(&ctx->uring_lock);
1539 ret = __io_cqring_overflow_flush(ctx, force);
1540 if (ctx->flags & IORING_SETUP_IOPOLL)
1541 mutex_unlock(&ctx->uring_lock);
1548 * Shamelessly stolen from the mm implementation of page reference checking,
1549 * see commit f958d7b528b1 for details.
1551 #define req_ref_zero_or_close_to_overflow(req) \
1552 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1554 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1556 return atomic_inc_not_zero(&req->refs);
1559 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1561 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1562 return atomic_sub_and_test(refs, &req->refs);
1565 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1567 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1568 return atomic_dec_and_test(&req->refs);
1571 static inline void req_ref_put(struct io_kiocb *req)
1573 WARN_ON_ONCE(req_ref_put_and_test(req));
1576 static inline void req_ref_get(struct io_kiocb *req)
1578 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1579 atomic_inc(&req->refs);
1582 /* must to be called somewhat shortly after putting a request */
1583 static inline void io_put_task(struct task_struct *task, int nr)
1585 struct io_uring_task *tctx = task->io_uring;
1587 percpu_counter_sub(&tctx->inflight, nr);
1588 if (unlikely(atomic_read(&tctx->in_idle)))
1589 wake_up(&tctx->wait);
1590 put_task_struct_many(task, nr);
1593 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1594 long res, unsigned int cflags)
1596 struct io_overflow_cqe *ocqe;
1598 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1601 * If we're in ring overflow flush mode, or in task cancel mode,
1602 * or cannot allocate an overflow entry, then we need to drop it
1605 io_account_cq_overflow(ctx);
1608 if (list_empty(&ctx->cq_overflow_list)) {
1609 set_bit(0, &ctx->check_cq_overflow);
1610 WRITE_ONCE(ctx->rings->sq_flags,
1611 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1614 ocqe->cqe.user_data = user_data;
1615 ocqe->cqe.res = res;
1616 ocqe->cqe.flags = cflags;
1617 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1621 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1622 long res, unsigned int cflags)
1624 struct io_uring_cqe *cqe;
1626 trace_io_uring_complete(ctx, user_data, res, cflags);
1629 * If we can't get a cq entry, userspace overflowed the
1630 * submission (by quite a lot). Increment the overflow count in
1633 cqe = io_get_cqe(ctx);
1635 WRITE_ONCE(cqe->user_data, user_data);
1636 WRITE_ONCE(cqe->res, res);
1637 WRITE_ONCE(cqe->flags, cflags);
1640 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1643 /* not as hot to bloat with inlining */
1644 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1645 long res, unsigned int cflags)
1647 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1650 static void io_req_complete_post(struct io_kiocb *req, long res,
1651 unsigned int cflags)
1653 struct io_ring_ctx *ctx = req->ctx;
1654 unsigned long flags;
1656 spin_lock_irqsave(&ctx->completion_lock, flags);
1657 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1659 * If we're the last reference to this request, add to our locked
1662 if (req_ref_put_and_test(req)) {
1663 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1664 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1665 io_disarm_next(req);
1667 io_req_task_queue(req->link);
1671 io_dismantle_req(req);
1672 io_put_task(req->task, 1);
1673 list_add(&req->compl.list, &ctx->locked_free_list);
1674 ctx->locked_free_nr++;
1676 if (!percpu_ref_tryget(&ctx->refs))
1679 io_commit_cqring(ctx);
1680 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1683 io_cqring_ev_posted(ctx);
1684 percpu_ref_put(&ctx->refs);
1688 static inline bool io_req_needs_clean(struct io_kiocb *req)
1690 return req->flags & IO_REQ_CLEAN_FLAGS;
1693 static void io_req_complete_state(struct io_kiocb *req, long res,
1694 unsigned int cflags)
1696 if (io_req_needs_clean(req))
1699 req->compl.cflags = cflags;
1700 req->flags |= REQ_F_COMPLETE_INLINE;
1703 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1704 long res, unsigned cflags)
1706 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1707 io_req_complete_state(req, res, cflags);
1709 io_req_complete_post(req, res, cflags);
1712 static inline void io_req_complete(struct io_kiocb *req, long res)
1714 __io_req_complete(req, 0, res, 0);
1717 static void io_req_complete_failed(struct io_kiocb *req, long res)
1721 io_req_complete_post(req, res, 0);
1725 * Don't initialise the fields below on every allocation, but do that in
1726 * advance and keep them valid across allocations.
1728 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1732 req->async_data = NULL;
1733 /* not necessary, but safer to zero */
1737 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1738 struct io_comp_state *cs)
1740 spin_lock_irq(&ctx->completion_lock);
1741 list_splice_init(&ctx->locked_free_list, &cs->free_list);
1742 ctx->locked_free_nr = 0;
1743 spin_unlock_irq(&ctx->completion_lock);
1746 /* Returns true IFF there are requests in the cache */
1747 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1749 struct io_submit_state *state = &ctx->submit_state;
1750 struct io_comp_state *cs = &state->comp;
1754 * If we have more than a batch's worth of requests in our IRQ side
1755 * locked cache, grab the lock and move them over to our submission
1758 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1759 io_flush_cached_locked_reqs(ctx, cs);
1761 nr = state->free_reqs;
1762 while (!list_empty(&cs->free_list)) {
1763 struct io_kiocb *req = list_first_entry(&cs->free_list,
1764 struct io_kiocb, compl.list);
1766 list_del(&req->compl.list);
1767 state->reqs[nr++] = req;
1768 if (nr == ARRAY_SIZE(state->reqs))
1772 state->free_reqs = nr;
1776 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1778 struct io_submit_state *state = &ctx->submit_state;
1779 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1782 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1784 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1787 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1791 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1792 * retry single alloc to be on the safe side.
1794 if (unlikely(ret <= 0)) {
1795 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1796 if (!state->reqs[0])
1801 for (i = 0; i < ret; i++)
1802 io_preinit_req(state->reqs[i], ctx);
1803 state->free_reqs = ret;
1806 return state->reqs[state->free_reqs];
1809 static inline void io_put_file(struct file *file)
1815 static void io_dismantle_req(struct io_kiocb *req)
1817 unsigned int flags = req->flags;
1819 if (io_req_needs_clean(req))
1821 if (!(flags & REQ_F_FIXED_FILE))
1822 io_put_file(req->file);
1823 if (req->fixed_rsrc_refs)
1824 percpu_ref_put(req->fixed_rsrc_refs);
1825 if (req->async_data) {
1826 kfree(req->async_data);
1827 req->async_data = NULL;
1831 static void __io_free_req(struct io_kiocb *req)
1833 struct io_ring_ctx *ctx = req->ctx;
1834 unsigned long flags;
1836 io_dismantle_req(req);
1837 io_put_task(req->task, 1);
1839 spin_lock_irqsave(&ctx->completion_lock, flags);
1840 list_add(&req->compl.list, &ctx->locked_free_list);
1841 ctx->locked_free_nr++;
1842 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1844 percpu_ref_put(&ctx->refs);
1847 static inline void io_remove_next_linked(struct io_kiocb *req)
1849 struct io_kiocb *nxt = req->link;
1851 req->link = nxt->link;
1855 static bool io_kill_linked_timeout(struct io_kiocb *req)
1856 __must_hold(&req->ctx->completion_lock)
1858 struct io_kiocb *link = req->link;
1861 * Can happen if a linked timeout fired and link had been like
1862 * req -> link t-out -> link t-out [-> ...]
1864 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1865 struct io_timeout_data *io = link->async_data;
1867 io_remove_next_linked(req);
1868 link->timeout.head = NULL;
1869 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1870 io_cqring_fill_event(link->ctx, link->user_data,
1872 io_put_req_deferred(link, 1);
1879 static void io_fail_links(struct io_kiocb *req)
1880 __must_hold(&req->ctx->completion_lock)
1882 struct io_kiocb *nxt, *link = req->link;
1889 trace_io_uring_fail_link(req, link);
1890 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1891 io_put_req_deferred(link, 2);
1896 static bool io_disarm_next(struct io_kiocb *req)
1897 __must_hold(&req->ctx->completion_lock)
1899 bool posted = false;
1901 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1902 posted = io_kill_linked_timeout(req);
1903 if (unlikely((req->flags & REQ_F_FAIL) &&
1904 !(req->flags & REQ_F_HARDLINK))) {
1905 posted |= (req->link != NULL);
1911 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1913 struct io_kiocb *nxt;
1916 * If LINK is set, we have dependent requests in this chain. If we
1917 * didn't fail this request, queue the first one up, moving any other
1918 * dependencies to the next request. In case of failure, fail the rest
1921 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1922 struct io_ring_ctx *ctx = req->ctx;
1923 unsigned long flags;
1926 spin_lock_irqsave(&ctx->completion_lock, flags);
1927 posted = io_disarm_next(req);
1929 io_commit_cqring(req->ctx);
1930 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1932 io_cqring_ev_posted(ctx);
1939 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1941 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1943 return __io_req_find_next(req);
1946 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1950 if (ctx->submit_state.comp.nr) {
1951 mutex_lock(&ctx->uring_lock);
1952 io_submit_flush_completions(ctx);
1953 mutex_unlock(&ctx->uring_lock);
1955 percpu_ref_put(&ctx->refs);
1958 static void tctx_task_work(struct callback_head *cb)
1960 struct io_ring_ctx *ctx = NULL;
1961 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1965 struct io_wq_work_node *node;
1967 spin_lock_irq(&tctx->task_lock);
1968 node = tctx->task_list.first;
1969 INIT_WQ_LIST(&tctx->task_list);
1970 spin_unlock_irq(&tctx->task_lock);
1973 struct io_wq_work_node *next = node->next;
1974 struct io_kiocb *req = container_of(node, struct io_kiocb,
1977 if (req->ctx != ctx) {
1978 ctx_flush_and_put(ctx);
1980 percpu_ref_get(&ctx->refs);
1982 req->io_task_work.func(req);
1985 if (wq_list_empty(&tctx->task_list)) {
1986 spin_lock_irq(&tctx->task_lock);
1987 clear_bit(0, &tctx->task_state);
1988 if (wq_list_empty(&tctx->task_list)) {
1989 spin_unlock_irq(&tctx->task_lock);
1992 spin_unlock_irq(&tctx->task_lock);
1993 /* another tctx_task_work() is enqueued, yield */
1994 if (test_and_set_bit(0, &tctx->task_state))
2000 ctx_flush_and_put(ctx);
2003 static void io_req_task_work_add(struct io_kiocb *req)
2005 struct task_struct *tsk = req->task;
2006 struct io_uring_task *tctx = tsk->io_uring;
2007 enum task_work_notify_mode notify;
2008 struct io_wq_work_node *node;
2009 unsigned long flags;
2011 WARN_ON_ONCE(!tctx);
2013 spin_lock_irqsave(&tctx->task_lock, flags);
2014 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2015 spin_unlock_irqrestore(&tctx->task_lock, flags);
2017 /* task_work already pending, we're done */
2018 if (test_bit(0, &tctx->task_state) ||
2019 test_and_set_bit(0, &tctx->task_state))
2023 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2024 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2025 * processing task_work. There's no reliable way to tell if TWA_RESUME
2028 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2029 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2030 wake_up_process(tsk);
2034 clear_bit(0, &tctx->task_state);
2035 spin_lock_irqsave(&tctx->task_lock, flags);
2036 node = tctx->task_list.first;
2037 INIT_WQ_LIST(&tctx->task_list);
2038 spin_unlock_irqrestore(&tctx->task_lock, flags);
2041 req = container_of(node, struct io_kiocb, io_task_work.node);
2043 if (llist_add(&req->io_task_work.fallback_node,
2044 &req->ctx->fallback_llist))
2045 schedule_delayed_work(&req->ctx->fallback_work, 1);
2049 static void io_req_task_cancel(struct io_kiocb *req)
2051 struct io_ring_ctx *ctx = req->ctx;
2053 /* ctx is guaranteed to stay alive while we hold uring_lock */
2054 mutex_lock(&ctx->uring_lock);
2055 io_req_complete_failed(req, req->result);
2056 mutex_unlock(&ctx->uring_lock);
2059 static void io_req_task_submit(struct io_kiocb *req)
2061 struct io_ring_ctx *ctx = req->ctx;
2063 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2064 mutex_lock(&ctx->uring_lock);
2065 if (likely(!(req->task->flags & PF_EXITING)))
2066 __io_queue_sqe(req);
2068 io_req_complete_failed(req, -EFAULT);
2069 mutex_unlock(&ctx->uring_lock);
2072 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2075 req->io_task_work.func = io_req_task_cancel;
2076 io_req_task_work_add(req);
2079 static void io_req_task_queue(struct io_kiocb *req)
2081 req->io_task_work.func = io_req_task_submit;
2082 io_req_task_work_add(req);
2085 static void io_req_task_queue_reissue(struct io_kiocb *req)
2087 req->io_task_work.func = io_queue_async_work;
2088 io_req_task_work_add(req);
2091 static inline void io_queue_next(struct io_kiocb *req)
2093 struct io_kiocb *nxt = io_req_find_next(req);
2096 io_req_task_queue(nxt);
2099 static void io_free_req(struct io_kiocb *req)
2106 struct task_struct *task;
2111 static inline void io_init_req_batch(struct req_batch *rb)
2118 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2119 struct req_batch *rb)
2122 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2123 if (rb->task == current)
2124 current->io_uring->cached_refs += rb->task_refs;
2126 io_put_task(rb->task, rb->task_refs);
2129 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2130 struct io_submit_state *state)
2133 io_dismantle_req(req);
2135 if (req->task != rb->task) {
2137 io_put_task(rb->task, rb->task_refs);
2138 rb->task = req->task;
2144 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2145 state->reqs[state->free_reqs++] = req;
2147 list_add(&req->compl.list, &state->comp.free_list);
2150 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2151 __must_hold(&req->ctx->uring_lock)
2153 struct io_comp_state *cs = &ctx->submit_state.comp;
2155 struct req_batch rb;
2157 spin_lock_irq(&ctx->completion_lock);
2158 for (i = 0; i < nr; i++) {
2159 struct io_kiocb *req = cs->reqs[i];
2161 __io_cqring_fill_event(ctx, req->user_data, req->result,
2164 io_commit_cqring(ctx);
2165 spin_unlock_irq(&ctx->completion_lock);
2166 io_cqring_ev_posted(ctx);
2168 io_init_req_batch(&rb);
2169 for (i = 0; i < nr; i++) {
2170 struct io_kiocb *req = cs->reqs[i];
2172 /* submission and completion refs */
2173 if (req_ref_sub_and_test(req, 2))
2174 io_req_free_batch(&rb, req, &ctx->submit_state);
2177 io_req_free_batch_finish(ctx, &rb);
2182 * Drop reference to request, return next in chain (if there is one) if this
2183 * was the last reference to this request.
2185 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2187 struct io_kiocb *nxt = NULL;
2189 if (req_ref_put_and_test(req)) {
2190 nxt = io_req_find_next(req);
2196 static inline void io_put_req(struct io_kiocb *req)
2198 if (req_ref_put_and_test(req))
2202 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2204 if (req_ref_sub_and_test(req, refs)) {
2205 req->io_task_work.func = io_free_req;
2206 io_req_task_work_add(req);
2210 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2212 /* See comment at the top of this file */
2214 return __io_cqring_events(ctx);
2217 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2219 struct io_rings *rings = ctx->rings;
2221 /* make sure SQ entry isn't read before tail */
2222 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2225 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2227 unsigned int cflags;
2229 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2230 cflags |= IORING_CQE_F_BUFFER;
2231 req->flags &= ~REQ_F_BUFFER_SELECTED;
2236 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2238 struct io_buffer *kbuf;
2240 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2241 return io_put_kbuf(req, kbuf);
2244 static inline bool io_run_task_work(void)
2246 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2247 __set_current_state(TASK_RUNNING);
2248 tracehook_notify_signal();
2256 * Find and free completed poll iocbs
2258 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2259 struct list_head *done, bool resubmit)
2261 struct req_batch rb;
2262 struct io_kiocb *req;
2264 /* order with ->result store in io_complete_rw_iopoll() */
2267 io_init_req_batch(&rb);
2268 while (!list_empty(done)) {
2271 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2272 list_del(&req->inflight_entry);
2274 if (READ_ONCE(req->result) == -EAGAIN && resubmit &&
2275 !(req->flags & REQ_F_DONT_REISSUE)) {
2276 req->iopoll_completed = 0;
2278 io_req_task_queue_reissue(req);
2282 if (req->flags & REQ_F_BUFFER_SELECTED)
2283 cflags = io_put_rw_kbuf(req);
2285 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2288 if (req_ref_put_and_test(req))
2289 io_req_free_batch(&rb, req, &ctx->submit_state);
2292 io_commit_cqring(ctx);
2293 io_cqring_ev_posted_iopoll(ctx);
2294 io_req_free_batch_finish(ctx, &rb);
2297 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2298 long min, bool resubmit)
2300 struct io_kiocb *req, *tmp;
2305 * Only spin for completions if we don't have multiple devices hanging
2306 * off our complete list, and we're under the requested amount.
2308 spin = !ctx->poll_multi_queue && *nr_events < min;
2310 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2311 struct kiocb *kiocb = &req->rw.kiocb;
2315 * Move completed and retryable entries to our local lists.
2316 * If we find a request that requires polling, break out
2317 * and complete those lists first, if we have entries there.
2319 if (READ_ONCE(req->iopoll_completed)) {
2320 list_move_tail(&req->inflight_entry, &done);
2323 if (!list_empty(&done))
2326 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2327 if (unlikely(ret < 0))
2332 /* iopoll may have completed current req */
2333 if (READ_ONCE(req->iopoll_completed))
2334 list_move_tail(&req->inflight_entry, &done);
2337 if (!list_empty(&done))
2338 io_iopoll_complete(ctx, nr_events, &done, resubmit);
2344 * We can't just wait for polled events to come to us, we have to actively
2345 * find and complete them.
2347 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2349 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2352 mutex_lock(&ctx->uring_lock);
2353 while (!list_empty(&ctx->iopoll_list)) {
2354 unsigned int nr_events = 0;
2356 io_do_iopoll(ctx, &nr_events, 0, false);
2358 /* let it sleep and repeat later if can't complete a request */
2362 * Ensure we allow local-to-the-cpu processing to take place,
2363 * in this case we need to ensure that we reap all events.
2364 * Also let task_work, etc. to progress by releasing the mutex
2366 if (need_resched()) {
2367 mutex_unlock(&ctx->uring_lock);
2369 mutex_lock(&ctx->uring_lock);
2372 mutex_unlock(&ctx->uring_lock);
2375 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2377 unsigned int nr_events = 0;
2381 * We disallow the app entering submit/complete with polling, but we
2382 * still need to lock the ring to prevent racing with polled issue
2383 * that got punted to a workqueue.
2385 mutex_lock(&ctx->uring_lock);
2387 * Don't enter poll loop if we already have events pending.
2388 * If we do, we can potentially be spinning for commands that
2389 * already triggered a CQE (eg in error).
2391 if (test_bit(0, &ctx->check_cq_overflow))
2392 __io_cqring_overflow_flush(ctx, false);
2393 if (io_cqring_events(ctx))
2397 * If a submit got punted to a workqueue, we can have the
2398 * application entering polling for a command before it gets
2399 * issued. That app will hold the uring_lock for the duration
2400 * of the poll right here, so we need to take a breather every
2401 * now and then to ensure that the issue has a chance to add
2402 * the poll to the issued list. Otherwise we can spin here
2403 * forever, while the workqueue is stuck trying to acquire the
2406 if (list_empty(&ctx->iopoll_list)) {
2407 u32 tail = ctx->cached_cq_tail;
2409 mutex_unlock(&ctx->uring_lock);
2411 mutex_lock(&ctx->uring_lock);
2413 /* some requests don't go through iopoll_list */
2414 if (tail != ctx->cached_cq_tail ||
2415 list_empty(&ctx->iopoll_list))
2418 ret = io_do_iopoll(ctx, &nr_events, min, true);
2419 } while (!ret && nr_events < min && !need_resched());
2421 mutex_unlock(&ctx->uring_lock);
2425 static void kiocb_end_write(struct io_kiocb *req)
2428 * Tell lockdep we inherited freeze protection from submission
2431 if (req->flags & REQ_F_ISREG) {
2432 struct super_block *sb = file_inode(req->file)->i_sb;
2434 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2440 static bool io_resubmit_prep(struct io_kiocb *req)
2442 struct io_async_rw *rw = req->async_data;
2445 return !io_req_prep_async(req);
2446 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2447 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2451 static bool io_rw_should_reissue(struct io_kiocb *req)
2453 umode_t mode = file_inode(req->file)->i_mode;
2454 struct io_ring_ctx *ctx = req->ctx;
2456 if (!S_ISBLK(mode) && !S_ISREG(mode))
2458 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2459 !(ctx->flags & IORING_SETUP_IOPOLL)))
2462 * If ref is dying, we might be running poll reap from the exit work.
2463 * Don't attempt to reissue from that path, just let it fail with
2466 if (percpu_ref_is_dying(&ctx->refs))
2469 * Play it safe and assume not safe to re-import and reissue if we're
2470 * not in the original thread group (or in task context).
2472 if (!same_thread_group(req->task, current) || !in_task())
2477 static bool io_resubmit_prep(struct io_kiocb *req)
2481 static bool io_rw_should_reissue(struct io_kiocb *req)
2487 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2488 unsigned int issue_flags)
2492 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2493 kiocb_end_write(req);
2494 if (res != req->result) {
2495 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2496 io_rw_should_reissue(req)) {
2497 req->flags |= REQ_F_REISSUE;
2502 if (req->flags & REQ_F_BUFFER_SELECTED)
2503 cflags = io_put_rw_kbuf(req);
2504 __io_req_complete(req, issue_flags, res, cflags);
2507 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2509 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2511 __io_complete_rw(req, res, res2, 0);
2514 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2516 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2518 if (kiocb->ki_flags & IOCB_WRITE)
2519 kiocb_end_write(req);
2520 if (unlikely(res != req->result)) {
2521 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2522 io_resubmit_prep(req))) {
2524 req->flags |= REQ_F_DONT_REISSUE;
2528 WRITE_ONCE(req->result, res);
2529 /* order with io_iopoll_complete() checking ->result */
2531 WRITE_ONCE(req->iopoll_completed, 1);
2535 * After the iocb has been issued, it's safe to be found on the poll list.
2536 * Adding the kiocb to the list AFTER submission ensures that we don't
2537 * find it from a io_do_iopoll() thread before the issuer is done
2538 * accessing the kiocb cookie.
2540 static void io_iopoll_req_issued(struct io_kiocb *req)
2542 struct io_ring_ctx *ctx = req->ctx;
2543 const bool in_async = io_wq_current_is_worker();
2545 /* workqueue context doesn't hold uring_lock, grab it now */
2546 if (unlikely(in_async))
2547 mutex_lock(&ctx->uring_lock);
2550 * Track whether we have multiple files in our lists. This will impact
2551 * how we do polling eventually, not spinning if we're on potentially
2552 * different devices.
2554 if (list_empty(&ctx->iopoll_list)) {
2555 ctx->poll_multi_queue = false;
2556 } else if (!ctx->poll_multi_queue) {
2557 struct io_kiocb *list_req;
2558 unsigned int queue_num0, queue_num1;
2560 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2563 if (list_req->file != req->file) {
2564 ctx->poll_multi_queue = true;
2566 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2567 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2568 if (queue_num0 != queue_num1)
2569 ctx->poll_multi_queue = true;
2574 * For fast devices, IO may have already completed. If it has, add
2575 * it to the front so we find it first.
2577 if (READ_ONCE(req->iopoll_completed))
2578 list_add(&req->inflight_entry, &ctx->iopoll_list);
2580 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2582 if (unlikely(in_async)) {
2584 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2585 * in sq thread task context or in io worker task context. If
2586 * current task context is sq thread, we don't need to check
2587 * whether should wake up sq thread.
2589 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2590 wq_has_sleeper(&ctx->sq_data->wait))
2591 wake_up(&ctx->sq_data->wait);
2593 mutex_unlock(&ctx->uring_lock);
2597 static inline void io_state_file_put(struct io_submit_state *state)
2599 if (state->file_refs) {
2600 fput_many(state->file, state->file_refs);
2601 state->file_refs = 0;
2606 * Get as many references to a file as we have IOs left in this submission,
2607 * assuming most submissions are for one file, or at least that each file
2608 * has more than one submission.
2610 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2615 if (state->file_refs) {
2616 if (state->fd == fd) {
2620 io_state_file_put(state);
2622 state->file = fget_many(fd, state->ios_left);
2623 if (unlikely(!state->file))
2627 state->file_refs = state->ios_left - 1;
2631 static bool io_bdev_nowait(struct block_device *bdev)
2633 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2637 * If we tracked the file through the SCM inflight mechanism, we could support
2638 * any file. For now, just ensure that anything potentially problematic is done
2641 static bool __io_file_supports_nowait(struct file *file, int rw)
2643 umode_t mode = file_inode(file)->i_mode;
2645 if (S_ISBLK(mode)) {
2646 if (IS_ENABLED(CONFIG_BLOCK) &&
2647 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2653 if (S_ISREG(mode)) {
2654 if (IS_ENABLED(CONFIG_BLOCK) &&
2655 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2656 file->f_op != &io_uring_fops)
2661 /* any ->read/write should understand O_NONBLOCK */
2662 if (file->f_flags & O_NONBLOCK)
2665 if (!(file->f_mode & FMODE_NOWAIT))
2669 return file->f_op->read_iter != NULL;
2671 return file->f_op->write_iter != NULL;
2674 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2676 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2678 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2681 return __io_file_supports_nowait(req->file, rw);
2684 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2686 struct io_ring_ctx *ctx = req->ctx;
2687 struct kiocb *kiocb = &req->rw.kiocb;
2688 struct file *file = req->file;
2692 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2693 req->flags |= REQ_F_ISREG;
2695 kiocb->ki_pos = READ_ONCE(sqe->off);
2696 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2697 req->flags |= REQ_F_CUR_POS;
2698 kiocb->ki_pos = file->f_pos;
2700 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2701 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2702 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2706 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2707 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2708 req->flags |= REQ_F_NOWAIT;
2710 ioprio = READ_ONCE(sqe->ioprio);
2712 ret = ioprio_check_cap(ioprio);
2716 kiocb->ki_ioprio = ioprio;
2718 kiocb->ki_ioprio = get_current_ioprio();
2720 if (ctx->flags & IORING_SETUP_IOPOLL) {
2721 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2722 !kiocb->ki_filp->f_op->iopoll)
2725 kiocb->ki_flags |= IOCB_HIPRI;
2726 kiocb->ki_complete = io_complete_rw_iopoll;
2727 req->iopoll_completed = 0;
2729 if (kiocb->ki_flags & IOCB_HIPRI)
2731 kiocb->ki_complete = io_complete_rw;
2734 if (req->opcode == IORING_OP_READ_FIXED ||
2735 req->opcode == IORING_OP_WRITE_FIXED) {
2737 io_req_set_rsrc_node(req);
2740 req->rw.addr = READ_ONCE(sqe->addr);
2741 req->rw.len = READ_ONCE(sqe->len);
2742 req->buf_index = READ_ONCE(sqe->buf_index);
2746 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2752 case -ERESTARTNOINTR:
2753 case -ERESTARTNOHAND:
2754 case -ERESTART_RESTARTBLOCK:
2756 * We can't just restart the syscall, since previously
2757 * submitted sqes may already be in progress. Just fail this
2763 kiocb->ki_complete(kiocb, ret, 0);
2767 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2768 unsigned int issue_flags)
2770 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2771 struct io_async_rw *io = req->async_data;
2772 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2774 /* add previously done IO, if any */
2775 if (io && io->bytes_done > 0) {
2777 ret = io->bytes_done;
2779 ret += io->bytes_done;
2782 if (req->flags & REQ_F_CUR_POS)
2783 req->file->f_pos = kiocb->ki_pos;
2784 if (ret >= 0 && check_reissue)
2785 __io_complete_rw(req, ret, 0, issue_flags);
2787 io_rw_done(kiocb, ret);
2789 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2790 req->flags &= ~REQ_F_REISSUE;
2791 if (io_resubmit_prep(req)) {
2793 io_req_task_queue_reissue(req);
2798 if (req->flags & REQ_F_BUFFER_SELECTED)
2799 cflags = io_put_rw_kbuf(req);
2800 __io_req_complete(req, issue_flags, ret, cflags);
2805 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2806 struct io_mapped_ubuf *imu)
2808 size_t len = req->rw.len;
2809 u64 buf_end, buf_addr = req->rw.addr;
2812 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2814 /* not inside the mapped region */
2815 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2819 * May not be a start of buffer, set size appropriately
2820 * and advance us to the beginning.
2822 offset = buf_addr - imu->ubuf;
2823 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2827 * Don't use iov_iter_advance() here, as it's really slow for
2828 * using the latter parts of a big fixed buffer - it iterates
2829 * over each segment manually. We can cheat a bit here, because
2832 * 1) it's a BVEC iter, we set it up
2833 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2834 * first and last bvec
2836 * So just find our index, and adjust the iterator afterwards.
2837 * If the offset is within the first bvec (or the whole first
2838 * bvec, just use iov_iter_advance(). This makes it easier
2839 * since we can just skip the first segment, which may not
2840 * be PAGE_SIZE aligned.
2842 const struct bio_vec *bvec = imu->bvec;
2844 if (offset <= bvec->bv_len) {
2845 iov_iter_advance(iter, offset);
2847 unsigned long seg_skip;
2849 /* skip first vec */
2850 offset -= bvec->bv_len;
2851 seg_skip = 1 + (offset >> PAGE_SHIFT);
2853 iter->bvec = bvec + seg_skip;
2854 iter->nr_segs -= seg_skip;
2855 iter->count -= bvec->bv_len + offset;
2856 iter->iov_offset = offset & ~PAGE_MASK;
2863 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2865 struct io_ring_ctx *ctx = req->ctx;
2866 struct io_mapped_ubuf *imu = req->imu;
2867 u16 index, buf_index = req->buf_index;
2870 if (unlikely(buf_index >= ctx->nr_user_bufs))
2872 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2873 imu = READ_ONCE(ctx->user_bufs[index]);
2876 return __io_import_fixed(req, rw, iter, imu);
2879 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2882 mutex_unlock(&ctx->uring_lock);
2885 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2888 * "Normal" inline submissions always hold the uring_lock, since we
2889 * grab it from the system call. Same is true for the SQPOLL offload.
2890 * The only exception is when we've detached the request and issue it
2891 * from an async worker thread, grab the lock for that case.
2894 mutex_lock(&ctx->uring_lock);
2897 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2898 int bgid, struct io_buffer *kbuf,
2901 struct io_buffer *head;
2903 if (req->flags & REQ_F_BUFFER_SELECTED)
2906 io_ring_submit_lock(req->ctx, needs_lock);
2908 lockdep_assert_held(&req->ctx->uring_lock);
2910 head = xa_load(&req->ctx->io_buffers, bgid);
2912 if (!list_empty(&head->list)) {
2913 kbuf = list_last_entry(&head->list, struct io_buffer,
2915 list_del(&kbuf->list);
2918 xa_erase(&req->ctx->io_buffers, bgid);
2920 if (*len > kbuf->len)
2923 kbuf = ERR_PTR(-ENOBUFS);
2926 io_ring_submit_unlock(req->ctx, needs_lock);
2931 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2934 struct io_buffer *kbuf;
2937 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2938 bgid = req->buf_index;
2939 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2942 req->rw.addr = (u64) (unsigned long) kbuf;
2943 req->flags |= REQ_F_BUFFER_SELECTED;
2944 return u64_to_user_ptr(kbuf->addr);
2947 #ifdef CONFIG_COMPAT
2948 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2951 struct compat_iovec __user *uiov;
2952 compat_ssize_t clen;
2956 uiov = u64_to_user_ptr(req->rw.addr);
2957 if (!access_ok(uiov, sizeof(*uiov)))
2959 if (__get_user(clen, &uiov->iov_len))
2965 buf = io_rw_buffer_select(req, &len, needs_lock);
2967 return PTR_ERR(buf);
2968 iov[0].iov_base = buf;
2969 iov[0].iov_len = (compat_size_t) len;
2974 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2977 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2981 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2984 len = iov[0].iov_len;
2987 buf = io_rw_buffer_select(req, &len, needs_lock);
2989 return PTR_ERR(buf);
2990 iov[0].iov_base = buf;
2991 iov[0].iov_len = len;
2995 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2998 if (req->flags & REQ_F_BUFFER_SELECTED) {
2999 struct io_buffer *kbuf;
3001 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3002 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3003 iov[0].iov_len = kbuf->len;
3006 if (req->rw.len != 1)
3009 #ifdef CONFIG_COMPAT
3010 if (req->ctx->compat)
3011 return io_compat_import(req, iov, needs_lock);
3014 return __io_iov_buffer_select(req, iov, needs_lock);
3017 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3018 struct iov_iter *iter, bool needs_lock)
3020 void __user *buf = u64_to_user_ptr(req->rw.addr);
3021 size_t sqe_len = req->rw.len;
3022 u8 opcode = req->opcode;
3025 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3027 return io_import_fixed(req, rw, iter);
3030 /* buffer index only valid with fixed read/write, or buffer select */
3031 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3034 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3035 if (req->flags & REQ_F_BUFFER_SELECT) {
3036 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3038 return PTR_ERR(buf);
3039 req->rw.len = sqe_len;
3042 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3047 if (req->flags & REQ_F_BUFFER_SELECT) {
3048 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3050 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3055 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3059 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3061 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3065 * For files that don't have ->read_iter() and ->write_iter(), handle them
3066 * by looping over ->read() or ->write() manually.
3068 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3070 struct kiocb *kiocb = &req->rw.kiocb;
3071 struct file *file = req->file;
3075 * Don't support polled IO through this interface, and we can't
3076 * support non-blocking either. For the latter, this just causes
3077 * the kiocb to be handled from an async context.
3079 if (kiocb->ki_flags & IOCB_HIPRI)
3081 if (kiocb->ki_flags & IOCB_NOWAIT)
3084 while (iov_iter_count(iter)) {
3088 if (!iov_iter_is_bvec(iter)) {
3089 iovec = iov_iter_iovec(iter);
3091 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3092 iovec.iov_len = req->rw.len;
3096 nr = file->f_op->read(file, iovec.iov_base,
3097 iovec.iov_len, io_kiocb_ppos(kiocb));
3099 nr = file->f_op->write(file, iovec.iov_base,
3100 iovec.iov_len, io_kiocb_ppos(kiocb));
3109 if (nr != iovec.iov_len)
3113 iov_iter_advance(iter, nr);
3119 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3120 const struct iovec *fast_iov, struct iov_iter *iter)
3122 struct io_async_rw *rw = req->async_data;
3124 memcpy(&rw->iter, iter, sizeof(*iter));
3125 rw->free_iovec = iovec;
3127 /* can only be fixed buffers, no need to do anything */
3128 if (iov_iter_is_bvec(iter))
3131 unsigned iov_off = 0;
3133 rw->iter.iov = rw->fast_iov;
3134 if (iter->iov != fast_iov) {
3135 iov_off = iter->iov - fast_iov;
3136 rw->iter.iov += iov_off;
3138 if (rw->fast_iov != fast_iov)
3139 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3140 sizeof(struct iovec) * iter->nr_segs);
3142 req->flags |= REQ_F_NEED_CLEANUP;
3146 static inline int io_alloc_async_data(struct io_kiocb *req)
3148 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3149 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3150 return req->async_data == NULL;
3153 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3154 const struct iovec *fast_iov,
3155 struct iov_iter *iter, bool force)
3157 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3159 if (!req->async_data) {
3160 if (io_alloc_async_data(req)) {
3165 io_req_map_rw(req, iovec, fast_iov, iter);
3170 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3172 struct io_async_rw *iorw = req->async_data;
3173 struct iovec *iov = iorw->fast_iov;
3176 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3177 if (unlikely(ret < 0))
3180 iorw->bytes_done = 0;
3181 iorw->free_iovec = iov;
3183 req->flags |= REQ_F_NEED_CLEANUP;
3187 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3189 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3191 return io_prep_rw(req, sqe);
3195 * This is our waitqueue callback handler, registered through lock_page_async()
3196 * when we initially tried to do the IO with the iocb armed our waitqueue.
3197 * This gets called when the page is unlocked, and we generally expect that to
3198 * happen when the page IO is completed and the page is now uptodate. This will
3199 * queue a task_work based retry of the operation, attempting to copy the data
3200 * again. If the latter fails because the page was NOT uptodate, then we will
3201 * do a thread based blocking retry of the operation. That's the unexpected
3204 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3205 int sync, void *arg)
3207 struct wait_page_queue *wpq;
3208 struct io_kiocb *req = wait->private;
3209 struct wait_page_key *key = arg;
3211 wpq = container_of(wait, struct wait_page_queue, wait);
3213 if (!wake_page_match(wpq, key))
3216 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3217 list_del_init(&wait->entry);
3219 /* submit ref gets dropped, acquire a new one */
3221 io_req_task_queue(req);
3226 * This controls whether a given IO request should be armed for async page
3227 * based retry. If we return false here, the request is handed to the async
3228 * worker threads for retry. If we're doing buffered reads on a regular file,
3229 * we prepare a private wait_page_queue entry and retry the operation. This
3230 * will either succeed because the page is now uptodate and unlocked, or it
3231 * will register a callback when the page is unlocked at IO completion. Through
3232 * that callback, io_uring uses task_work to setup a retry of the operation.
3233 * That retry will attempt the buffered read again. The retry will generally
3234 * succeed, or in rare cases where it fails, we then fall back to using the
3235 * async worker threads for a blocking retry.
3237 static bool io_rw_should_retry(struct io_kiocb *req)
3239 struct io_async_rw *rw = req->async_data;
3240 struct wait_page_queue *wait = &rw->wpq;
3241 struct kiocb *kiocb = &req->rw.kiocb;
3243 /* never retry for NOWAIT, we just complete with -EAGAIN */
3244 if (req->flags & REQ_F_NOWAIT)
3247 /* Only for buffered IO */
3248 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3252 * just use poll if we can, and don't attempt if the fs doesn't
3253 * support callback based unlocks
3255 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3258 wait->wait.func = io_async_buf_func;
3259 wait->wait.private = req;
3260 wait->wait.flags = 0;
3261 INIT_LIST_HEAD(&wait->wait.entry);
3262 kiocb->ki_flags |= IOCB_WAITQ;
3263 kiocb->ki_flags &= ~IOCB_NOWAIT;
3264 kiocb->ki_waitq = wait;
3268 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3270 if (req->file->f_op->read_iter)
3271 return call_read_iter(req->file, &req->rw.kiocb, iter);
3272 else if (req->file->f_op->read)
3273 return loop_rw_iter(READ, req, iter);
3278 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3280 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3281 struct kiocb *kiocb = &req->rw.kiocb;
3282 struct iov_iter __iter, *iter = &__iter;
3283 struct io_async_rw *rw = req->async_data;
3284 ssize_t io_size, ret, ret2;
3285 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3291 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3295 io_size = iov_iter_count(iter);
3296 req->result = io_size;
3298 /* Ensure we clear previously set non-block flag */
3299 if (!force_nonblock)
3300 kiocb->ki_flags &= ~IOCB_NOWAIT;
3302 kiocb->ki_flags |= IOCB_NOWAIT;
3304 /* If the file doesn't support async, just async punt */
3305 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3306 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3307 return ret ?: -EAGAIN;
3310 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3311 if (unlikely(ret)) {
3316 ret = io_iter_do_read(req, iter);
3318 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3319 req->flags &= ~REQ_F_REISSUE;
3320 /* IOPOLL retry should happen for io-wq threads */
3321 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3323 /* no retry on NONBLOCK nor RWF_NOWAIT */
3324 if (req->flags & REQ_F_NOWAIT)
3326 /* some cases will consume bytes even on error returns */
3327 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3329 } else if (ret == -EIOCBQUEUED) {
3331 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3332 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3333 /* read all, failed, already did sync or don't want to retry */
3337 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3342 rw = req->async_data;
3343 /* now use our persistent iterator, if we aren't already */
3348 rw->bytes_done += ret;
3349 /* if we can retry, do so with the callbacks armed */
3350 if (!io_rw_should_retry(req)) {
3351 kiocb->ki_flags &= ~IOCB_WAITQ;
3356 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3357 * we get -EIOCBQUEUED, then we'll get a notification when the
3358 * desired page gets unlocked. We can also get a partial read
3359 * here, and if we do, then just retry at the new offset.
3361 ret = io_iter_do_read(req, iter);
3362 if (ret == -EIOCBQUEUED)
3364 /* we got some bytes, but not all. retry. */
3365 kiocb->ki_flags &= ~IOCB_WAITQ;
3366 } while (ret > 0 && ret < io_size);
3368 kiocb_done(kiocb, ret, issue_flags);
3370 /* it's faster to check here then delegate to kfree */
3376 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3378 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3380 return io_prep_rw(req, sqe);
3383 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3385 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3386 struct kiocb *kiocb = &req->rw.kiocb;
3387 struct iov_iter __iter, *iter = &__iter;
3388 struct io_async_rw *rw = req->async_data;
3389 ssize_t ret, ret2, io_size;
3390 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3396 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3400 io_size = iov_iter_count(iter);
3401 req->result = io_size;
3403 /* Ensure we clear previously set non-block flag */
3404 if (!force_nonblock)
3405 kiocb->ki_flags &= ~IOCB_NOWAIT;
3407 kiocb->ki_flags |= IOCB_NOWAIT;
3409 /* If the file doesn't support async, just async punt */
3410 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3413 /* file path doesn't support NOWAIT for non-direct_IO */
3414 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3415 (req->flags & REQ_F_ISREG))
3418 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3423 * Open-code file_start_write here to grab freeze protection,
3424 * which will be released by another thread in
3425 * io_complete_rw(). Fool lockdep by telling it the lock got
3426 * released so that it doesn't complain about the held lock when
3427 * we return to userspace.
3429 if (req->flags & REQ_F_ISREG) {
3430 sb_start_write(file_inode(req->file)->i_sb);
3431 __sb_writers_release(file_inode(req->file)->i_sb,
3434 kiocb->ki_flags |= IOCB_WRITE;
3436 if (req->file->f_op->write_iter)
3437 ret2 = call_write_iter(req->file, kiocb, iter);
3438 else if (req->file->f_op->write)
3439 ret2 = loop_rw_iter(WRITE, req, iter);
3443 if (req->flags & REQ_F_REISSUE) {
3444 req->flags &= ~REQ_F_REISSUE;
3449 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3450 * retry them without IOCB_NOWAIT.
3452 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3454 /* no retry on NONBLOCK nor RWF_NOWAIT */
3455 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3457 if (!force_nonblock || ret2 != -EAGAIN) {
3458 /* IOPOLL retry should happen for io-wq threads */
3459 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3462 kiocb_done(kiocb, ret2, issue_flags);
3465 /* some cases will consume bytes even on error returns */
3466 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3467 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3468 return ret ?: -EAGAIN;
3471 /* it's reportedly faster than delegating the null check to kfree() */
3477 static int io_renameat_prep(struct io_kiocb *req,
3478 const struct io_uring_sqe *sqe)
3480 struct io_rename *ren = &req->rename;
3481 const char __user *oldf, *newf;
3483 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3485 if (sqe->ioprio || sqe->buf_index)
3487 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3490 ren->old_dfd = READ_ONCE(sqe->fd);
3491 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3492 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3493 ren->new_dfd = READ_ONCE(sqe->len);
3494 ren->flags = READ_ONCE(sqe->rename_flags);
3496 ren->oldpath = getname(oldf);
3497 if (IS_ERR(ren->oldpath))
3498 return PTR_ERR(ren->oldpath);
3500 ren->newpath = getname(newf);
3501 if (IS_ERR(ren->newpath)) {
3502 putname(ren->oldpath);
3503 return PTR_ERR(ren->newpath);
3506 req->flags |= REQ_F_NEED_CLEANUP;
3510 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3512 struct io_rename *ren = &req->rename;
3515 if (issue_flags & IO_URING_F_NONBLOCK)
3518 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3519 ren->newpath, ren->flags);
3521 req->flags &= ~REQ_F_NEED_CLEANUP;
3524 io_req_complete(req, ret);
3528 static int io_unlinkat_prep(struct io_kiocb *req,
3529 const struct io_uring_sqe *sqe)
3531 struct io_unlink *un = &req->unlink;
3532 const char __user *fname;
3534 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3536 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3538 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3541 un->dfd = READ_ONCE(sqe->fd);
3543 un->flags = READ_ONCE(sqe->unlink_flags);
3544 if (un->flags & ~AT_REMOVEDIR)
3547 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3548 un->filename = getname(fname);
3549 if (IS_ERR(un->filename))
3550 return PTR_ERR(un->filename);
3552 req->flags |= REQ_F_NEED_CLEANUP;
3556 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3558 struct io_unlink *un = &req->unlink;
3561 if (issue_flags & IO_URING_F_NONBLOCK)
3564 if (un->flags & AT_REMOVEDIR)
3565 ret = do_rmdir(un->dfd, un->filename);
3567 ret = do_unlinkat(un->dfd, un->filename);
3569 req->flags &= ~REQ_F_NEED_CLEANUP;
3572 io_req_complete(req, ret);
3576 static int io_shutdown_prep(struct io_kiocb *req,
3577 const struct io_uring_sqe *sqe)
3579 #if defined(CONFIG_NET)
3580 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3582 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3586 req->shutdown.how = READ_ONCE(sqe->len);
3593 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3595 #if defined(CONFIG_NET)
3596 struct socket *sock;
3599 if (issue_flags & IO_URING_F_NONBLOCK)
3602 sock = sock_from_file(req->file);
3603 if (unlikely(!sock))
3606 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3609 io_req_complete(req, ret);
3616 static int __io_splice_prep(struct io_kiocb *req,
3617 const struct io_uring_sqe *sqe)
3619 struct io_splice *sp = &req->splice;
3620 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3622 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3626 sp->len = READ_ONCE(sqe->len);
3627 sp->flags = READ_ONCE(sqe->splice_flags);
3629 if (unlikely(sp->flags & ~valid_flags))
3632 sp->file_in = io_file_get(req->ctx, NULL, req,
3633 READ_ONCE(sqe->splice_fd_in),
3634 (sp->flags & SPLICE_F_FD_IN_FIXED));
3637 req->flags |= REQ_F_NEED_CLEANUP;
3641 static int io_tee_prep(struct io_kiocb *req,
3642 const struct io_uring_sqe *sqe)
3644 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3646 return __io_splice_prep(req, sqe);
3649 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3651 struct io_splice *sp = &req->splice;
3652 struct file *in = sp->file_in;
3653 struct file *out = sp->file_out;
3654 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3657 if (issue_flags & IO_URING_F_NONBLOCK)
3660 ret = do_tee(in, out, sp->len, flags);
3662 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3664 req->flags &= ~REQ_F_NEED_CLEANUP;
3668 io_req_complete(req, ret);
3672 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3674 struct io_splice *sp = &req->splice;
3676 sp->off_in = READ_ONCE(sqe->splice_off_in);
3677 sp->off_out = READ_ONCE(sqe->off);
3678 return __io_splice_prep(req, sqe);
3681 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3683 struct io_splice *sp = &req->splice;
3684 struct file *in = sp->file_in;
3685 struct file *out = sp->file_out;
3686 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3687 loff_t *poff_in, *poff_out;
3690 if (issue_flags & IO_URING_F_NONBLOCK)
3693 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3694 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3697 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3699 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3701 req->flags &= ~REQ_F_NEED_CLEANUP;
3705 io_req_complete(req, ret);
3710 * IORING_OP_NOP just posts a completion event, nothing else.
3712 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3714 struct io_ring_ctx *ctx = req->ctx;
3716 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3719 __io_req_complete(req, issue_flags, 0, 0);
3723 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3725 struct io_ring_ctx *ctx = req->ctx;
3730 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3732 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3735 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3736 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3739 req->sync.off = READ_ONCE(sqe->off);
3740 req->sync.len = READ_ONCE(sqe->len);
3744 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3746 loff_t end = req->sync.off + req->sync.len;
3749 /* fsync always requires a blocking context */
3750 if (issue_flags & IO_URING_F_NONBLOCK)
3753 ret = vfs_fsync_range(req->file, req->sync.off,
3754 end > 0 ? end : LLONG_MAX,
3755 req->sync.flags & IORING_FSYNC_DATASYNC);
3758 io_req_complete(req, ret);
3762 static int io_fallocate_prep(struct io_kiocb *req,
3763 const struct io_uring_sqe *sqe)
3765 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3767 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3770 req->sync.off = READ_ONCE(sqe->off);
3771 req->sync.len = READ_ONCE(sqe->addr);
3772 req->sync.mode = READ_ONCE(sqe->len);
3776 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3780 /* fallocate always requiring blocking context */
3781 if (issue_flags & IO_URING_F_NONBLOCK)
3783 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3787 io_req_complete(req, ret);
3791 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3793 const char __user *fname;
3796 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3798 if (unlikely(sqe->ioprio || sqe->buf_index))
3800 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3803 /* open.how should be already initialised */
3804 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3805 req->open.how.flags |= O_LARGEFILE;
3807 req->open.dfd = READ_ONCE(sqe->fd);
3808 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3809 req->open.filename = getname(fname);
3810 if (IS_ERR(req->open.filename)) {
3811 ret = PTR_ERR(req->open.filename);
3812 req->open.filename = NULL;
3815 req->open.nofile = rlimit(RLIMIT_NOFILE);
3816 req->flags |= REQ_F_NEED_CLEANUP;
3820 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3822 u64 mode = READ_ONCE(sqe->len);
3823 u64 flags = READ_ONCE(sqe->open_flags);
3825 req->open.how = build_open_how(flags, mode);
3826 return __io_openat_prep(req, sqe);
3829 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3831 struct open_how __user *how;
3835 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3836 len = READ_ONCE(sqe->len);
3837 if (len < OPEN_HOW_SIZE_VER0)
3840 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3845 return __io_openat_prep(req, sqe);
3848 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3850 struct open_flags op;
3853 bool resolve_nonblock;
3856 ret = build_open_flags(&req->open.how, &op);
3859 nonblock_set = op.open_flag & O_NONBLOCK;
3860 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3861 if (issue_flags & IO_URING_F_NONBLOCK) {
3863 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3864 * it'll always -EAGAIN
3866 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3868 op.lookup_flags |= LOOKUP_CACHED;
3869 op.open_flag |= O_NONBLOCK;
3872 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3876 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3879 * We could hang on to this 'fd' on retrying, but seems like
3880 * marginal gain for something that is now known to be a slower
3881 * path. So just put it, and we'll get a new one when we retry.
3885 ret = PTR_ERR(file);
3886 /* only retry if RESOLVE_CACHED wasn't already set by application */
3887 if (ret == -EAGAIN &&
3888 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3893 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3894 file->f_flags &= ~O_NONBLOCK;
3895 fsnotify_open(file);
3896 fd_install(ret, file);
3898 putname(req->open.filename);
3899 req->flags &= ~REQ_F_NEED_CLEANUP;
3902 __io_req_complete(req, issue_flags, ret, 0);
3906 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3908 return io_openat2(req, issue_flags);
3911 static int io_remove_buffers_prep(struct io_kiocb *req,
3912 const struct io_uring_sqe *sqe)
3914 struct io_provide_buf *p = &req->pbuf;
3917 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3920 tmp = READ_ONCE(sqe->fd);
3921 if (!tmp || tmp > USHRT_MAX)
3924 memset(p, 0, sizeof(*p));
3926 p->bgid = READ_ONCE(sqe->buf_group);
3930 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3931 int bgid, unsigned nbufs)
3935 /* shouldn't happen */
3939 /* the head kbuf is the list itself */
3940 while (!list_empty(&buf->list)) {
3941 struct io_buffer *nxt;
3943 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3944 list_del(&nxt->list);
3951 xa_erase(&ctx->io_buffers, bgid);
3956 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3958 struct io_provide_buf *p = &req->pbuf;
3959 struct io_ring_ctx *ctx = req->ctx;
3960 struct io_buffer *head;
3962 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3964 io_ring_submit_lock(ctx, !force_nonblock);
3966 lockdep_assert_held(&ctx->uring_lock);
3969 head = xa_load(&ctx->io_buffers, p->bgid);
3971 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3975 /* complete before unlock, IOPOLL may need the lock */
3976 __io_req_complete(req, issue_flags, ret, 0);
3977 io_ring_submit_unlock(ctx, !force_nonblock);
3981 static int io_provide_buffers_prep(struct io_kiocb *req,
3982 const struct io_uring_sqe *sqe)
3984 unsigned long size, tmp_check;
3985 struct io_provide_buf *p = &req->pbuf;
3988 if (sqe->ioprio || sqe->rw_flags)
3991 tmp = READ_ONCE(sqe->fd);
3992 if (!tmp || tmp > USHRT_MAX)
3995 p->addr = READ_ONCE(sqe->addr);
3996 p->len = READ_ONCE(sqe->len);
3998 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4001 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4004 size = (unsigned long)p->len * p->nbufs;
4005 if (!access_ok(u64_to_user_ptr(p->addr), size))
4008 p->bgid = READ_ONCE(sqe->buf_group);
4009 tmp = READ_ONCE(sqe->off);
4010 if (tmp > USHRT_MAX)
4016 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4018 struct io_buffer *buf;
4019 u64 addr = pbuf->addr;
4020 int i, bid = pbuf->bid;
4022 for (i = 0; i < pbuf->nbufs; i++) {
4023 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4028 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4033 INIT_LIST_HEAD(&buf->list);
4036 list_add_tail(&buf->list, &(*head)->list);
4040 return i ? i : -ENOMEM;
4043 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4045 struct io_provide_buf *p = &req->pbuf;
4046 struct io_ring_ctx *ctx = req->ctx;
4047 struct io_buffer *head, *list;
4049 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4051 io_ring_submit_lock(ctx, !force_nonblock);
4053 lockdep_assert_held(&ctx->uring_lock);
4055 list = head = xa_load(&ctx->io_buffers, p->bgid);
4057 ret = io_add_buffers(p, &head);
4058 if (ret >= 0 && !list) {
4059 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4061 __io_remove_buffers(ctx, head, p->bgid, -1U);
4065 /* complete before unlock, IOPOLL may need the lock */
4066 __io_req_complete(req, issue_flags, ret, 0);
4067 io_ring_submit_unlock(ctx, !force_nonblock);
4071 static int io_epoll_ctl_prep(struct io_kiocb *req,
4072 const struct io_uring_sqe *sqe)
4074 #if defined(CONFIG_EPOLL)
4075 if (sqe->ioprio || sqe->buf_index)
4077 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4080 req->epoll.epfd = READ_ONCE(sqe->fd);
4081 req->epoll.op = READ_ONCE(sqe->len);
4082 req->epoll.fd = READ_ONCE(sqe->off);
4084 if (ep_op_has_event(req->epoll.op)) {
4085 struct epoll_event __user *ev;
4087 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4088 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4098 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4100 #if defined(CONFIG_EPOLL)
4101 struct io_epoll *ie = &req->epoll;
4103 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4105 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4106 if (force_nonblock && ret == -EAGAIN)
4111 __io_req_complete(req, issue_flags, ret, 0);
4118 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4120 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4121 if (sqe->ioprio || sqe->buf_index || sqe->off)
4123 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4126 req->madvise.addr = READ_ONCE(sqe->addr);
4127 req->madvise.len = READ_ONCE(sqe->len);
4128 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4135 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4137 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4138 struct io_madvise *ma = &req->madvise;
4141 if (issue_flags & IO_URING_F_NONBLOCK)
4144 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4147 io_req_complete(req, ret);
4154 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4156 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4158 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4161 req->fadvise.offset = READ_ONCE(sqe->off);
4162 req->fadvise.len = READ_ONCE(sqe->len);
4163 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4167 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4169 struct io_fadvise *fa = &req->fadvise;
4172 if (issue_flags & IO_URING_F_NONBLOCK) {
4173 switch (fa->advice) {
4174 case POSIX_FADV_NORMAL:
4175 case POSIX_FADV_RANDOM:
4176 case POSIX_FADV_SEQUENTIAL:
4183 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4186 __io_req_complete(req, issue_flags, ret, 0);
4190 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4192 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4194 if (sqe->ioprio || sqe->buf_index)
4196 if (req->flags & REQ_F_FIXED_FILE)
4199 req->statx.dfd = READ_ONCE(sqe->fd);
4200 req->statx.mask = READ_ONCE(sqe->len);
4201 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4202 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4203 req->statx.flags = READ_ONCE(sqe->statx_flags);
4208 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4210 struct io_statx *ctx = &req->statx;
4213 if (issue_flags & IO_URING_F_NONBLOCK)
4216 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4221 io_req_complete(req, ret);
4225 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4227 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4229 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4230 sqe->rw_flags || sqe->buf_index)
4232 if (req->flags & REQ_F_FIXED_FILE)
4235 req->close.fd = READ_ONCE(sqe->fd);
4239 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4241 struct files_struct *files = current->files;
4242 struct io_close *close = &req->close;
4243 struct fdtable *fdt;
4244 struct file *file = NULL;
4247 spin_lock(&files->file_lock);
4248 fdt = files_fdtable(files);
4249 if (close->fd >= fdt->max_fds) {
4250 spin_unlock(&files->file_lock);
4253 file = fdt->fd[close->fd];
4254 if (!file || file->f_op == &io_uring_fops) {
4255 spin_unlock(&files->file_lock);
4260 /* if the file has a flush method, be safe and punt to async */
4261 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4262 spin_unlock(&files->file_lock);
4266 ret = __close_fd_get_file(close->fd, &file);
4267 spin_unlock(&files->file_lock);
4274 /* No ->flush() or already async, safely close from here */
4275 ret = filp_close(file, current->files);
4281 __io_req_complete(req, issue_flags, ret, 0);
4285 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4287 struct io_ring_ctx *ctx = req->ctx;
4289 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4291 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4294 req->sync.off = READ_ONCE(sqe->off);
4295 req->sync.len = READ_ONCE(sqe->len);
4296 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4300 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4304 /* sync_file_range always requires a blocking context */
4305 if (issue_flags & IO_URING_F_NONBLOCK)
4308 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4312 io_req_complete(req, ret);
4316 #if defined(CONFIG_NET)
4317 static int io_setup_async_msg(struct io_kiocb *req,
4318 struct io_async_msghdr *kmsg)
4320 struct io_async_msghdr *async_msg = req->async_data;
4324 if (io_alloc_async_data(req)) {
4325 kfree(kmsg->free_iov);
4328 async_msg = req->async_data;
4329 req->flags |= REQ_F_NEED_CLEANUP;
4330 memcpy(async_msg, kmsg, sizeof(*kmsg));
4331 async_msg->msg.msg_name = &async_msg->addr;
4332 /* if were using fast_iov, set it to the new one */
4333 if (!async_msg->free_iov)
4334 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4339 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4340 struct io_async_msghdr *iomsg)
4342 iomsg->msg.msg_name = &iomsg->addr;
4343 iomsg->free_iov = iomsg->fast_iov;
4344 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4345 req->sr_msg.msg_flags, &iomsg->free_iov);
4348 static int io_sendmsg_prep_async(struct io_kiocb *req)
4352 ret = io_sendmsg_copy_hdr(req, req->async_data);
4354 req->flags |= REQ_F_NEED_CLEANUP;
4358 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4360 struct io_sr_msg *sr = &req->sr_msg;
4362 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4365 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4366 sr->len = READ_ONCE(sqe->len);
4367 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4368 if (sr->msg_flags & MSG_DONTWAIT)
4369 req->flags |= REQ_F_NOWAIT;
4371 #ifdef CONFIG_COMPAT
4372 if (req->ctx->compat)
4373 sr->msg_flags |= MSG_CMSG_COMPAT;
4378 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4380 struct io_async_msghdr iomsg, *kmsg;
4381 struct socket *sock;
4386 sock = sock_from_file(req->file);
4387 if (unlikely(!sock))
4390 kmsg = req->async_data;
4392 ret = io_sendmsg_copy_hdr(req, &iomsg);
4398 flags = req->sr_msg.msg_flags;
4399 if (issue_flags & IO_URING_F_NONBLOCK)
4400 flags |= MSG_DONTWAIT;
4401 if (flags & MSG_WAITALL)
4402 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4404 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4405 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4406 return io_setup_async_msg(req, kmsg);
4407 if (ret == -ERESTARTSYS)
4410 /* fast path, check for non-NULL to avoid function call */
4412 kfree(kmsg->free_iov);
4413 req->flags &= ~REQ_F_NEED_CLEANUP;
4416 __io_req_complete(req, issue_flags, ret, 0);
4420 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4422 struct io_sr_msg *sr = &req->sr_msg;
4425 struct socket *sock;
4430 sock = sock_from_file(req->file);
4431 if (unlikely(!sock))
4434 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4438 msg.msg_name = NULL;
4439 msg.msg_control = NULL;
4440 msg.msg_controllen = 0;
4441 msg.msg_namelen = 0;
4443 flags = req->sr_msg.msg_flags;
4444 if (issue_flags & IO_URING_F_NONBLOCK)
4445 flags |= MSG_DONTWAIT;
4446 if (flags & MSG_WAITALL)
4447 min_ret = iov_iter_count(&msg.msg_iter);
4449 msg.msg_flags = flags;
4450 ret = sock_sendmsg(sock, &msg);
4451 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4453 if (ret == -ERESTARTSYS)
4458 __io_req_complete(req, issue_flags, ret, 0);
4462 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4463 struct io_async_msghdr *iomsg)
4465 struct io_sr_msg *sr = &req->sr_msg;
4466 struct iovec __user *uiov;
4470 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4471 &iomsg->uaddr, &uiov, &iov_len);
4475 if (req->flags & REQ_F_BUFFER_SELECT) {
4478 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4480 sr->len = iomsg->fast_iov[0].iov_len;
4481 iomsg->free_iov = NULL;
4483 iomsg->free_iov = iomsg->fast_iov;
4484 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4485 &iomsg->free_iov, &iomsg->msg.msg_iter,
4494 #ifdef CONFIG_COMPAT
4495 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4496 struct io_async_msghdr *iomsg)
4498 struct io_sr_msg *sr = &req->sr_msg;
4499 struct compat_iovec __user *uiov;
4504 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4509 uiov = compat_ptr(ptr);
4510 if (req->flags & REQ_F_BUFFER_SELECT) {
4511 compat_ssize_t clen;
4515 if (!access_ok(uiov, sizeof(*uiov)))
4517 if (__get_user(clen, &uiov->iov_len))
4522 iomsg->free_iov = NULL;
4524 iomsg->free_iov = iomsg->fast_iov;
4525 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4526 UIO_FASTIOV, &iomsg->free_iov,
4527 &iomsg->msg.msg_iter, true);
4536 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4537 struct io_async_msghdr *iomsg)
4539 iomsg->msg.msg_name = &iomsg->addr;
4541 #ifdef CONFIG_COMPAT
4542 if (req->ctx->compat)
4543 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4546 return __io_recvmsg_copy_hdr(req, iomsg);
4549 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4552 struct io_sr_msg *sr = &req->sr_msg;
4553 struct io_buffer *kbuf;
4555 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4560 req->flags |= REQ_F_BUFFER_SELECTED;
4564 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4566 return io_put_kbuf(req, req->sr_msg.kbuf);
4569 static int io_recvmsg_prep_async(struct io_kiocb *req)
4573 ret = io_recvmsg_copy_hdr(req, req->async_data);
4575 req->flags |= REQ_F_NEED_CLEANUP;
4579 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4581 struct io_sr_msg *sr = &req->sr_msg;
4583 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4586 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4587 sr->len = READ_ONCE(sqe->len);
4588 sr->bgid = READ_ONCE(sqe->buf_group);
4589 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4590 if (sr->msg_flags & MSG_DONTWAIT)
4591 req->flags |= REQ_F_NOWAIT;
4593 #ifdef CONFIG_COMPAT
4594 if (req->ctx->compat)
4595 sr->msg_flags |= MSG_CMSG_COMPAT;
4600 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4602 struct io_async_msghdr iomsg, *kmsg;
4603 struct socket *sock;
4604 struct io_buffer *kbuf;
4607 int ret, cflags = 0;
4608 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4610 sock = sock_from_file(req->file);
4611 if (unlikely(!sock))
4614 kmsg = req->async_data;
4616 ret = io_recvmsg_copy_hdr(req, &iomsg);
4622 if (req->flags & REQ_F_BUFFER_SELECT) {
4623 kbuf = io_recv_buffer_select(req, !force_nonblock);
4625 return PTR_ERR(kbuf);
4626 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4627 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4628 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4629 1, req->sr_msg.len);
4632 flags = req->sr_msg.msg_flags;
4634 flags |= MSG_DONTWAIT;
4635 if (flags & MSG_WAITALL)
4636 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4638 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4639 kmsg->uaddr, flags);
4640 if (force_nonblock && ret == -EAGAIN)
4641 return io_setup_async_msg(req, kmsg);
4642 if (ret == -ERESTARTSYS)
4645 if (req->flags & REQ_F_BUFFER_SELECTED)
4646 cflags = io_put_recv_kbuf(req);
4647 /* fast path, check for non-NULL to avoid function call */
4649 kfree(kmsg->free_iov);
4650 req->flags &= ~REQ_F_NEED_CLEANUP;
4651 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4653 __io_req_complete(req, issue_flags, ret, cflags);
4657 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4659 struct io_buffer *kbuf;
4660 struct io_sr_msg *sr = &req->sr_msg;
4662 void __user *buf = sr->buf;
4663 struct socket *sock;
4667 int ret, cflags = 0;
4668 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4670 sock = sock_from_file(req->file);
4671 if (unlikely(!sock))
4674 if (req->flags & REQ_F_BUFFER_SELECT) {
4675 kbuf = io_recv_buffer_select(req, !force_nonblock);
4677 return PTR_ERR(kbuf);
4678 buf = u64_to_user_ptr(kbuf->addr);
4681 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4685 msg.msg_name = NULL;
4686 msg.msg_control = NULL;
4687 msg.msg_controllen = 0;
4688 msg.msg_namelen = 0;
4689 msg.msg_iocb = NULL;
4692 flags = req->sr_msg.msg_flags;
4694 flags |= MSG_DONTWAIT;
4695 if (flags & MSG_WAITALL)
4696 min_ret = iov_iter_count(&msg.msg_iter);
4698 ret = sock_recvmsg(sock, &msg, flags);
4699 if (force_nonblock && ret == -EAGAIN)
4701 if (ret == -ERESTARTSYS)
4704 if (req->flags & REQ_F_BUFFER_SELECTED)
4705 cflags = io_put_recv_kbuf(req);
4706 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4708 __io_req_complete(req, issue_flags, ret, cflags);
4712 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4714 struct io_accept *accept = &req->accept;
4716 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4718 if (sqe->ioprio || sqe->len || sqe->buf_index)
4721 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4722 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4723 accept->flags = READ_ONCE(sqe->accept_flags);
4724 accept->nofile = rlimit(RLIMIT_NOFILE);
4728 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4730 struct io_accept *accept = &req->accept;
4731 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4732 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4735 if (req->file->f_flags & O_NONBLOCK)
4736 req->flags |= REQ_F_NOWAIT;
4738 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4739 accept->addr_len, accept->flags,
4741 if (ret == -EAGAIN && force_nonblock)
4744 if (ret == -ERESTARTSYS)
4748 __io_req_complete(req, issue_flags, ret, 0);
4752 static int io_connect_prep_async(struct io_kiocb *req)
4754 struct io_async_connect *io = req->async_data;
4755 struct io_connect *conn = &req->connect;
4757 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4760 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4762 struct io_connect *conn = &req->connect;
4764 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4766 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4769 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4770 conn->addr_len = READ_ONCE(sqe->addr2);
4774 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4776 struct io_async_connect __io, *io;
4777 unsigned file_flags;
4779 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4781 if (req->async_data) {
4782 io = req->async_data;
4784 ret = move_addr_to_kernel(req->connect.addr,
4785 req->connect.addr_len,
4792 file_flags = force_nonblock ? O_NONBLOCK : 0;
4794 ret = __sys_connect_file(req->file, &io->address,
4795 req->connect.addr_len, file_flags);
4796 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4797 if (req->async_data)
4799 if (io_alloc_async_data(req)) {
4803 memcpy(req->async_data, &__io, sizeof(__io));
4806 if (ret == -ERESTARTSYS)
4811 __io_req_complete(req, issue_flags, ret, 0);
4814 #else /* !CONFIG_NET */
4815 #define IO_NETOP_FN(op) \
4816 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4818 return -EOPNOTSUPP; \
4821 #define IO_NETOP_PREP(op) \
4823 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4825 return -EOPNOTSUPP; \
4828 #define IO_NETOP_PREP_ASYNC(op) \
4830 static int io_##op##_prep_async(struct io_kiocb *req) \
4832 return -EOPNOTSUPP; \
4835 IO_NETOP_PREP_ASYNC(sendmsg);
4836 IO_NETOP_PREP_ASYNC(recvmsg);
4837 IO_NETOP_PREP_ASYNC(connect);
4838 IO_NETOP_PREP(accept);
4841 #endif /* CONFIG_NET */
4843 struct io_poll_table {
4844 struct poll_table_struct pt;
4845 struct io_kiocb *req;
4850 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4851 __poll_t mask, io_req_tw_func_t func)
4853 /* for instances that support it check for an event match first: */
4854 if (mask && !(mask & poll->events))
4857 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4859 list_del_init(&poll->wait.entry);
4862 req->io_task_work.func = func;
4865 * If this fails, then the task is exiting. When a task exits, the
4866 * work gets canceled, so just cancel this request as well instead
4867 * of executing it. We can't safely execute it anyway, as we may not
4868 * have the needed state needed for it anyway.
4870 io_req_task_work_add(req);
4874 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4875 __acquires(&req->ctx->completion_lock)
4877 struct io_ring_ctx *ctx = req->ctx;
4879 if (unlikely(req->task->flags & PF_EXITING))
4880 WRITE_ONCE(poll->canceled, true);
4882 if (!req->result && !READ_ONCE(poll->canceled)) {
4883 struct poll_table_struct pt = { ._key = poll->events };
4885 req->result = vfs_poll(req->file, &pt) & poll->events;
4888 spin_lock_irq(&ctx->completion_lock);
4889 if (!req->result && !READ_ONCE(poll->canceled)) {
4890 add_wait_queue(poll->head, &poll->wait);
4897 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4899 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4900 if (req->opcode == IORING_OP_POLL_ADD)
4901 return req->async_data;
4902 return req->apoll->double_poll;
4905 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4907 if (req->opcode == IORING_OP_POLL_ADD)
4909 return &req->apoll->poll;
4912 static void io_poll_remove_double(struct io_kiocb *req)
4913 __must_hold(&req->ctx->completion_lock)
4915 struct io_poll_iocb *poll = io_poll_get_double(req);
4917 lockdep_assert_held(&req->ctx->completion_lock);
4919 if (poll && poll->head) {
4920 struct wait_queue_head *head = poll->head;
4922 spin_lock(&head->lock);
4923 list_del_init(&poll->wait.entry);
4924 if (poll->wait.private)
4927 spin_unlock(&head->lock);
4931 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4932 __must_hold(&req->ctx->completion_lock)
4934 struct io_ring_ctx *ctx = req->ctx;
4935 unsigned flags = IORING_CQE_F_MORE;
4938 if (READ_ONCE(req->poll.canceled)) {
4940 req->poll.events |= EPOLLONESHOT;
4942 error = mangle_poll(mask);
4944 if (req->poll.events & EPOLLONESHOT)
4946 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4947 req->poll.done = true;
4950 if (flags & IORING_CQE_F_MORE)
4953 io_commit_cqring(ctx);
4954 return !(flags & IORING_CQE_F_MORE);
4957 static void io_poll_task_func(struct io_kiocb *req)
4959 struct io_ring_ctx *ctx = req->ctx;
4960 struct io_kiocb *nxt;
4962 if (io_poll_rewait(req, &req->poll)) {
4963 spin_unlock_irq(&ctx->completion_lock);
4967 done = io_poll_complete(req, req->result);
4969 io_poll_remove_double(req);
4970 hash_del(&req->hash_node);
4973 add_wait_queue(req->poll.head, &req->poll.wait);
4975 spin_unlock_irq(&ctx->completion_lock);
4976 io_cqring_ev_posted(ctx);
4979 nxt = io_put_req_find_next(req);
4981 io_req_task_submit(nxt);
4986 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4987 int sync, void *key)
4989 struct io_kiocb *req = wait->private;
4990 struct io_poll_iocb *poll = io_poll_get_single(req);
4991 __poll_t mask = key_to_poll(key);
4993 /* for instances that support it check for an event match first: */
4994 if (mask && !(mask & poll->events))
4996 if (!(poll->events & EPOLLONESHOT))
4997 return poll->wait.func(&poll->wait, mode, sync, key);
4999 list_del_init(&wait->entry);
5004 spin_lock(&poll->head->lock);
5005 done = list_empty(&poll->wait.entry);
5007 list_del_init(&poll->wait.entry);
5008 /* make sure double remove sees this as being gone */
5009 wait->private = NULL;
5010 spin_unlock(&poll->head->lock);
5012 /* use wait func handler, so it matches the rq type */
5013 poll->wait.func(&poll->wait, mode, sync, key);
5020 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5021 wait_queue_func_t wake_func)
5025 poll->canceled = false;
5026 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5027 /* mask in events that we always want/need */
5028 poll->events = events | IO_POLL_UNMASK;
5029 INIT_LIST_HEAD(&poll->wait.entry);
5030 init_waitqueue_func_entry(&poll->wait, wake_func);
5033 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5034 struct wait_queue_head *head,
5035 struct io_poll_iocb **poll_ptr)
5037 struct io_kiocb *req = pt->req;
5040 * The file being polled uses multiple waitqueues for poll handling
5041 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5044 if (unlikely(pt->nr_entries)) {
5045 struct io_poll_iocb *poll_one = poll;
5047 /* already have a 2nd entry, fail a third attempt */
5049 pt->error = -EINVAL;
5053 * Can't handle multishot for double wait for now, turn it
5054 * into one-shot mode.
5056 if (!(poll_one->events & EPOLLONESHOT))
5057 poll_one->events |= EPOLLONESHOT;
5058 /* double add on the same waitqueue head, ignore */
5059 if (poll_one->head == head)
5061 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5063 pt->error = -ENOMEM;
5066 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5068 poll->wait.private = req;
5075 if (poll->events & EPOLLEXCLUSIVE)
5076 add_wait_queue_exclusive(head, &poll->wait);
5078 add_wait_queue(head, &poll->wait);
5081 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5082 struct poll_table_struct *p)
5084 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5085 struct async_poll *apoll = pt->req->apoll;
5087 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5090 static void io_async_task_func(struct io_kiocb *req)
5092 struct async_poll *apoll = req->apoll;
5093 struct io_ring_ctx *ctx = req->ctx;
5095 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5097 if (io_poll_rewait(req, &apoll->poll)) {
5098 spin_unlock_irq(&ctx->completion_lock);
5102 hash_del(&req->hash_node);
5103 io_poll_remove_double(req);
5104 spin_unlock_irq(&ctx->completion_lock);
5106 if (!READ_ONCE(apoll->poll.canceled))
5107 io_req_task_submit(req);
5109 io_req_complete_failed(req, -ECANCELED);
5112 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5115 struct io_kiocb *req = wait->private;
5116 struct io_poll_iocb *poll = &req->apoll->poll;
5118 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5121 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5124 static void io_poll_req_insert(struct io_kiocb *req)
5126 struct io_ring_ctx *ctx = req->ctx;
5127 struct hlist_head *list;
5129 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5130 hlist_add_head(&req->hash_node, list);
5133 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5134 struct io_poll_iocb *poll,
5135 struct io_poll_table *ipt, __poll_t mask,
5136 wait_queue_func_t wake_func)
5137 __acquires(&ctx->completion_lock)
5139 struct io_ring_ctx *ctx = req->ctx;
5140 bool cancel = false;
5142 INIT_HLIST_NODE(&req->hash_node);
5143 io_init_poll_iocb(poll, mask, wake_func);
5144 poll->file = req->file;
5145 poll->wait.private = req;
5147 ipt->pt._key = mask;
5150 ipt->nr_entries = 0;
5152 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5153 if (unlikely(!ipt->nr_entries) && !ipt->error)
5154 ipt->error = -EINVAL;
5156 spin_lock_irq(&ctx->completion_lock);
5157 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5158 io_poll_remove_double(req);
5159 if (likely(poll->head)) {
5160 spin_lock(&poll->head->lock);
5161 if (unlikely(list_empty(&poll->wait.entry))) {
5167 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5168 list_del_init(&poll->wait.entry);
5170 WRITE_ONCE(poll->canceled, true);
5171 else if (!poll->done) /* actually waiting for an event */
5172 io_poll_req_insert(req);
5173 spin_unlock(&poll->head->lock);
5185 static int io_arm_poll_handler(struct io_kiocb *req)
5187 const struct io_op_def *def = &io_op_defs[req->opcode];
5188 struct io_ring_ctx *ctx = req->ctx;
5189 struct async_poll *apoll;
5190 struct io_poll_table ipt;
5191 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5194 if (!req->file || !file_can_poll(req->file))
5195 return IO_APOLL_ABORTED;
5196 if (req->flags & REQ_F_POLLED)
5197 return IO_APOLL_ABORTED;
5198 if (!def->pollin && !def->pollout)
5199 return IO_APOLL_ABORTED;
5203 mask |= POLLIN | POLLRDNORM;
5205 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5206 if ((req->opcode == IORING_OP_RECVMSG) &&
5207 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5211 mask |= POLLOUT | POLLWRNORM;
5214 /* if we can't nonblock try, then no point in arming a poll handler */
5215 if (!io_file_supports_nowait(req, rw))
5216 return IO_APOLL_ABORTED;
5218 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5219 if (unlikely(!apoll))
5220 return IO_APOLL_ABORTED;
5221 apoll->double_poll = NULL;
5223 req->flags |= REQ_F_POLLED;
5224 ipt.pt._qproc = io_async_queue_proc;
5226 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5228 if (ret || ipt.error) {
5229 spin_unlock_irq(&ctx->completion_lock);
5231 return IO_APOLL_READY;
5232 return IO_APOLL_ABORTED;
5234 spin_unlock_irq(&ctx->completion_lock);
5235 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5236 mask, apoll->poll.events);
5240 static bool __io_poll_remove_one(struct io_kiocb *req,
5241 struct io_poll_iocb *poll, bool do_cancel)
5242 __must_hold(&req->ctx->completion_lock)
5244 bool do_complete = false;
5248 spin_lock(&poll->head->lock);
5250 WRITE_ONCE(poll->canceled, true);
5251 if (!list_empty(&poll->wait.entry)) {
5252 list_del_init(&poll->wait.entry);
5255 spin_unlock(&poll->head->lock);
5256 hash_del(&req->hash_node);
5260 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5261 __must_hold(&req->ctx->completion_lock)
5265 io_poll_remove_double(req);
5266 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5268 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5269 /* non-poll requests have submit ref still */
5275 static bool io_poll_remove_one(struct io_kiocb *req)
5276 __must_hold(&req->ctx->completion_lock)
5280 do_complete = io_poll_remove_waitqs(req);
5282 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5283 io_commit_cqring(req->ctx);
5285 io_put_req_deferred(req, 1);
5292 * Returns true if we found and killed one or more poll requests
5294 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5297 struct hlist_node *tmp;
5298 struct io_kiocb *req;
5301 spin_lock_irq(&ctx->completion_lock);
5302 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5303 struct hlist_head *list;
5305 list = &ctx->cancel_hash[i];
5306 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5307 if (io_match_task(req, tsk, cancel_all))
5308 posted += io_poll_remove_one(req);
5311 spin_unlock_irq(&ctx->completion_lock);
5314 io_cqring_ev_posted(ctx);
5319 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5321 __must_hold(&ctx->completion_lock)
5323 struct hlist_head *list;
5324 struct io_kiocb *req;
5326 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5327 hlist_for_each_entry(req, list, hash_node) {
5328 if (sqe_addr != req->user_data)
5330 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5337 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5339 __must_hold(&ctx->completion_lock)
5341 struct io_kiocb *req;
5343 req = io_poll_find(ctx, sqe_addr, poll_only);
5346 if (io_poll_remove_one(req))
5352 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5357 events = READ_ONCE(sqe->poll32_events);
5359 events = swahw32(events);
5361 if (!(flags & IORING_POLL_ADD_MULTI))
5362 events |= EPOLLONESHOT;
5363 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5366 static int io_poll_update_prep(struct io_kiocb *req,
5367 const struct io_uring_sqe *sqe)
5369 struct io_poll_update *upd = &req->poll_update;
5372 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5374 if (sqe->ioprio || sqe->buf_index)
5376 flags = READ_ONCE(sqe->len);
5377 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5378 IORING_POLL_ADD_MULTI))
5380 /* meaningless without update */
5381 if (flags == IORING_POLL_ADD_MULTI)
5384 upd->old_user_data = READ_ONCE(sqe->addr);
5385 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5386 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5388 upd->new_user_data = READ_ONCE(sqe->off);
5389 if (!upd->update_user_data && upd->new_user_data)
5391 if (upd->update_events)
5392 upd->events = io_poll_parse_events(sqe, flags);
5393 else if (sqe->poll32_events)
5399 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5402 struct io_kiocb *req = wait->private;
5403 struct io_poll_iocb *poll = &req->poll;
5405 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5408 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5409 struct poll_table_struct *p)
5411 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5413 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5416 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5418 struct io_poll_iocb *poll = &req->poll;
5421 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5423 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5425 flags = READ_ONCE(sqe->len);
5426 if (flags & ~IORING_POLL_ADD_MULTI)
5429 poll->events = io_poll_parse_events(sqe, flags);
5433 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5435 struct io_poll_iocb *poll = &req->poll;
5436 struct io_ring_ctx *ctx = req->ctx;
5437 struct io_poll_table ipt;
5440 ipt.pt._qproc = io_poll_queue_proc;
5442 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5445 if (mask) { /* no async, we'd stolen it */
5447 io_poll_complete(req, mask);
5449 spin_unlock_irq(&ctx->completion_lock);
5452 io_cqring_ev_posted(ctx);
5453 if (poll->events & EPOLLONESHOT)
5459 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5461 struct io_ring_ctx *ctx = req->ctx;
5462 struct io_kiocb *preq;
5466 spin_lock_irq(&ctx->completion_lock);
5467 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5473 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5475 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5480 * Don't allow racy completion with singleshot, as we cannot safely
5481 * update those. For multishot, if we're racing with completion, just
5482 * let completion re-add it.
5484 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5485 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5489 /* we now have a detached poll request. reissue. */
5493 spin_unlock_irq(&ctx->completion_lock);
5495 io_req_complete(req, ret);
5498 /* only mask one event flags, keep behavior flags */
5499 if (req->poll_update.update_events) {
5500 preq->poll.events &= ~0xffff;
5501 preq->poll.events |= req->poll_update.events & 0xffff;
5502 preq->poll.events |= IO_POLL_UNMASK;
5504 if (req->poll_update.update_user_data)
5505 preq->user_data = req->poll_update.new_user_data;
5506 spin_unlock_irq(&ctx->completion_lock);
5508 /* complete update request, we're done with it */
5509 io_req_complete(req, ret);
5512 ret = io_poll_add(preq, issue_flags);
5515 io_req_complete(preq, ret);
5521 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5523 struct io_timeout_data *data = container_of(timer,
5524 struct io_timeout_data, timer);
5525 struct io_kiocb *req = data->req;
5526 struct io_ring_ctx *ctx = req->ctx;
5527 unsigned long flags;
5529 spin_lock_irqsave(&ctx->completion_lock, flags);
5530 list_del_init(&req->timeout.list);
5531 atomic_set(&req->ctx->cq_timeouts,
5532 atomic_read(&req->ctx->cq_timeouts) + 1);
5534 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5535 io_commit_cqring(ctx);
5536 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5538 io_cqring_ev_posted(ctx);
5541 return HRTIMER_NORESTART;
5544 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5546 __must_hold(&ctx->completion_lock)
5548 struct io_timeout_data *io;
5549 struct io_kiocb *req;
5552 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5553 found = user_data == req->user_data;
5558 return ERR_PTR(-ENOENT);
5560 io = req->async_data;
5561 if (hrtimer_try_to_cancel(&io->timer) == -1)
5562 return ERR_PTR(-EALREADY);
5563 list_del_init(&req->timeout.list);
5567 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5568 __must_hold(&ctx->completion_lock)
5570 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5573 return PTR_ERR(req);
5576 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5577 io_put_req_deferred(req, 1);
5581 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5582 struct timespec64 *ts, enum hrtimer_mode mode)
5583 __must_hold(&ctx->completion_lock)
5585 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5586 struct io_timeout_data *data;
5589 return PTR_ERR(req);
5591 req->timeout.off = 0; /* noseq */
5592 data = req->async_data;
5593 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5594 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5595 data->timer.function = io_timeout_fn;
5596 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5600 static int io_timeout_remove_prep(struct io_kiocb *req,
5601 const struct io_uring_sqe *sqe)
5603 struct io_timeout_rem *tr = &req->timeout_rem;
5605 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5607 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5609 if (sqe->ioprio || sqe->buf_index || sqe->len)
5612 tr->addr = READ_ONCE(sqe->addr);
5613 tr->flags = READ_ONCE(sqe->timeout_flags);
5614 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5615 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5617 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5619 } else if (tr->flags) {
5620 /* timeout removal doesn't support flags */
5627 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5629 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5634 * Remove or update an existing timeout command
5636 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5638 struct io_timeout_rem *tr = &req->timeout_rem;
5639 struct io_ring_ctx *ctx = req->ctx;
5642 spin_lock_irq(&ctx->completion_lock);
5643 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5644 ret = io_timeout_cancel(ctx, tr->addr);
5646 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5647 io_translate_timeout_mode(tr->flags));
5649 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5650 io_commit_cqring(ctx);
5651 spin_unlock_irq(&ctx->completion_lock);
5652 io_cqring_ev_posted(ctx);
5659 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5660 bool is_timeout_link)
5662 struct io_timeout_data *data;
5664 u32 off = READ_ONCE(sqe->off);
5666 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5668 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5670 if (off && is_timeout_link)
5672 flags = READ_ONCE(sqe->timeout_flags);
5673 if (flags & ~IORING_TIMEOUT_ABS)
5676 req->timeout.off = off;
5677 if (unlikely(off && !req->ctx->off_timeout_used))
5678 req->ctx->off_timeout_used = true;
5680 if (!req->async_data && io_alloc_async_data(req))
5683 data = req->async_data;
5686 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5689 data->mode = io_translate_timeout_mode(flags);
5690 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5691 if (is_timeout_link)
5692 io_req_track_inflight(req);
5696 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5698 struct io_ring_ctx *ctx = req->ctx;
5699 struct io_timeout_data *data = req->async_data;
5700 struct list_head *entry;
5701 u32 tail, off = req->timeout.off;
5703 spin_lock_irq(&ctx->completion_lock);
5706 * sqe->off holds how many events that need to occur for this
5707 * timeout event to be satisfied. If it isn't set, then this is
5708 * a pure timeout request, sequence isn't used.
5710 if (io_is_timeout_noseq(req)) {
5711 entry = ctx->timeout_list.prev;
5715 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5716 req->timeout.target_seq = tail + off;
5718 /* Update the last seq here in case io_flush_timeouts() hasn't.
5719 * This is safe because ->completion_lock is held, and submissions
5720 * and completions are never mixed in the same ->completion_lock section.
5722 ctx->cq_last_tm_flush = tail;
5725 * Insertion sort, ensuring the first entry in the list is always
5726 * the one we need first.
5728 list_for_each_prev(entry, &ctx->timeout_list) {
5729 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5732 if (io_is_timeout_noseq(nxt))
5734 /* nxt.seq is behind @tail, otherwise would've been completed */
5735 if (off >= nxt->timeout.target_seq - tail)
5739 list_add(&req->timeout.list, entry);
5740 data->timer.function = io_timeout_fn;
5741 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5742 spin_unlock_irq(&ctx->completion_lock);
5746 struct io_cancel_data {
5747 struct io_ring_ctx *ctx;
5751 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5753 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5754 struct io_cancel_data *cd = data;
5756 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5759 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5760 struct io_ring_ctx *ctx)
5762 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5763 enum io_wq_cancel cancel_ret;
5766 if (!tctx || !tctx->io_wq)
5769 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5770 switch (cancel_ret) {
5771 case IO_WQ_CANCEL_OK:
5774 case IO_WQ_CANCEL_RUNNING:
5777 case IO_WQ_CANCEL_NOTFOUND:
5785 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5786 struct io_kiocb *req, __u64 sqe_addr,
5789 unsigned long flags;
5792 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5793 spin_lock_irqsave(&ctx->completion_lock, flags);
5796 ret = io_timeout_cancel(ctx, sqe_addr);
5799 ret = io_poll_cancel(ctx, sqe_addr, false);
5803 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5804 io_commit_cqring(ctx);
5805 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5806 io_cqring_ev_posted(ctx);
5812 static int io_async_cancel_prep(struct io_kiocb *req,
5813 const struct io_uring_sqe *sqe)
5815 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5817 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5819 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5822 req->cancel.addr = READ_ONCE(sqe->addr);
5826 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5828 struct io_ring_ctx *ctx = req->ctx;
5829 u64 sqe_addr = req->cancel.addr;
5830 struct io_tctx_node *node;
5833 /* tasks should wait for their io-wq threads, so safe w/o sync */
5834 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5835 spin_lock_irq(&ctx->completion_lock);
5838 ret = io_timeout_cancel(ctx, sqe_addr);
5841 ret = io_poll_cancel(ctx, sqe_addr, false);
5844 spin_unlock_irq(&ctx->completion_lock);
5846 /* slow path, try all io-wq's */
5847 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5849 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5850 struct io_uring_task *tctx = node->task->io_uring;
5852 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5856 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5858 spin_lock_irq(&ctx->completion_lock);
5860 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5861 io_commit_cqring(ctx);
5862 spin_unlock_irq(&ctx->completion_lock);
5863 io_cqring_ev_posted(ctx);
5871 static int io_rsrc_update_prep(struct io_kiocb *req,
5872 const struct io_uring_sqe *sqe)
5874 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5876 if (sqe->ioprio || sqe->rw_flags)
5879 req->rsrc_update.offset = READ_ONCE(sqe->off);
5880 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5881 if (!req->rsrc_update.nr_args)
5883 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5887 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5889 struct io_ring_ctx *ctx = req->ctx;
5890 struct io_uring_rsrc_update2 up;
5893 if (issue_flags & IO_URING_F_NONBLOCK)
5896 up.offset = req->rsrc_update.offset;
5897 up.data = req->rsrc_update.arg;
5902 mutex_lock(&ctx->uring_lock);
5903 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5904 &up, req->rsrc_update.nr_args);
5905 mutex_unlock(&ctx->uring_lock);
5909 __io_req_complete(req, issue_flags, ret, 0);
5913 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5915 switch (req->opcode) {
5918 case IORING_OP_READV:
5919 case IORING_OP_READ_FIXED:
5920 case IORING_OP_READ:
5921 return io_read_prep(req, sqe);
5922 case IORING_OP_WRITEV:
5923 case IORING_OP_WRITE_FIXED:
5924 case IORING_OP_WRITE:
5925 return io_write_prep(req, sqe);
5926 case IORING_OP_POLL_ADD:
5927 return io_poll_add_prep(req, sqe);
5928 case IORING_OP_POLL_REMOVE:
5929 return io_poll_update_prep(req, sqe);
5930 case IORING_OP_FSYNC:
5931 return io_fsync_prep(req, sqe);
5932 case IORING_OP_SYNC_FILE_RANGE:
5933 return io_sfr_prep(req, sqe);
5934 case IORING_OP_SENDMSG:
5935 case IORING_OP_SEND:
5936 return io_sendmsg_prep(req, sqe);
5937 case IORING_OP_RECVMSG:
5938 case IORING_OP_RECV:
5939 return io_recvmsg_prep(req, sqe);
5940 case IORING_OP_CONNECT:
5941 return io_connect_prep(req, sqe);
5942 case IORING_OP_TIMEOUT:
5943 return io_timeout_prep(req, sqe, false);
5944 case IORING_OP_TIMEOUT_REMOVE:
5945 return io_timeout_remove_prep(req, sqe);
5946 case IORING_OP_ASYNC_CANCEL:
5947 return io_async_cancel_prep(req, sqe);
5948 case IORING_OP_LINK_TIMEOUT:
5949 return io_timeout_prep(req, sqe, true);
5950 case IORING_OP_ACCEPT:
5951 return io_accept_prep(req, sqe);
5952 case IORING_OP_FALLOCATE:
5953 return io_fallocate_prep(req, sqe);
5954 case IORING_OP_OPENAT:
5955 return io_openat_prep(req, sqe);
5956 case IORING_OP_CLOSE:
5957 return io_close_prep(req, sqe);
5958 case IORING_OP_FILES_UPDATE:
5959 return io_rsrc_update_prep(req, sqe);
5960 case IORING_OP_STATX:
5961 return io_statx_prep(req, sqe);
5962 case IORING_OP_FADVISE:
5963 return io_fadvise_prep(req, sqe);
5964 case IORING_OP_MADVISE:
5965 return io_madvise_prep(req, sqe);
5966 case IORING_OP_OPENAT2:
5967 return io_openat2_prep(req, sqe);
5968 case IORING_OP_EPOLL_CTL:
5969 return io_epoll_ctl_prep(req, sqe);
5970 case IORING_OP_SPLICE:
5971 return io_splice_prep(req, sqe);
5972 case IORING_OP_PROVIDE_BUFFERS:
5973 return io_provide_buffers_prep(req, sqe);
5974 case IORING_OP_REMOVE_BUFFERS:
5975 return io_remove_buffers_prep(req, sqe);
5977 return io_tee_prep(req, sqe);
5978 case IORING_OP_SHUTDOWN:
5979 return io_shutdown_prep(req, sqe);
5980 case IORING_OP_RENAMEAT:
5981 return io_renameat_prep(req, sqe);
5982 case IORING_OP_UNLINKAT:
5983 return io_unlinkat_prep(req, sqe);
5986 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5991 static int io_req_prep_async(struct io_kiocb *req)
5993 if (!io_op_defs[req->opcode].needs_async_setup)
5995 if (WARN_ON_ONCE(req->async_data))
5997 if (io_alloc_async_data(req))
6000 switch (req->opcode) {
6001 case IORING_OP_READV:
6002 return io_rw_prep_async(req, READ);
6003 case IORING_OP_WRITEV:
6004 return io_rw_prep_async(req, WRITE);
6005 case IORING_OP_SENDMSG:
6006 return io_sendmsg_prep_async(req);
6007 case IORING_OP_RECVMSG:
6008 return io_recvmsg_prep_async(req);
6009 case IORING_OP_CONNECT:
6010 return io_connect_prep_async(req);
6012 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6017 static u32 io_get_sequence(struct io_kiocb *req)
6019 u32 seq = req->ctx->cached_sq_head;
6021 /* need original cached_sq_head, but it was increased for each req */
6022 io_for_each_link(req, req)
6027 static bool io_drain_req(struct io_kiocb *req)
6029 struct io_kiocb *pos;
6030 struct io_ring_ctx *ctx = req->ctx;
6031 struct io_defer_entry *de;
6036 * If we need to drain a request in the middle of a link, drain the
6037 * head request and the next request/link after the current link.
6038 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6039 * maintained for every request of our link.
6041 if (ctx->drain_next) {
6042 req->flags |= REQ_F_IO_DRAIN;
6043 ctx->drain_next = false;
6045 /* not interested in head, start from the first linked */
6046 io_for_each_link(pos, req->link) {
6047 if (pos->flags & REQ_F_IO_DRAIN) {
6048 ctx->drain_next = true;
6049 req->flags |= REQ_F_IO_DRAIN;
6054 /* Still need defer if there is pending req in defer list. */
6055 if (likely(list_empty_careful(&ctx->defer_list) &&
6056 !(req->flags & REQ_F_IO_DRAIN))) {
6057 ctx->drain_active = false;
6061 seq = io_get_sequence(req);
6062 /* Still a chance to pass the sequence check */
6063 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6066 ret = io_req_prep_async(req);
6069 io_prep_async_link(req);
6070 de = kmalloc(sizeof(*de), GFP_KERNEL);
6074 io_req_complete_failed(req, ret);
6078 spin_lock_irq(&ctx->completion_lock);
6079 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6080 spin_unlock_irq(&ctx->completion_lock);
6082 io_queue_async_work(req);
6086 trace_io_uring_defer(ctx, req, req->user_data);
6089 list_add_tail(&de->list, &ctx->defer_list);
6090 spin_unlock_irq(&ctx->completion_lock);
6094 static void io_clean_op(struct io_kiocb *req)
6096 if (req->flags & REQ_F_BUFFER_SELECTED) {
6097 switch (req->opcode) {
6098 case IORING_OP_READV:
6099 case IORING_OP_READ_FIXED:
6100 case IORING_OP_READ:
6101 kfree((void *)(unsigned long)req->rw.addr);
6103 case IORING_OP_RECVMSG:
6104 case IORING_OP_RECV:
6105 kfree(req->sr_msg.kbuf);
6110 if (req->flags & REQ_F_NEED_CLEANUP) {
6111 switch (req->opcode) {
6112 case IORING_OP_READV:
6113 case IORING_OP_READ_FIXED:
6114 case IORING_OP_READ:
6115 case IORING_OP_WRITEV:
6116 case IORING_OP_WRITE_FIXED:
6117 case IORING_OP_WRITE: {
6118 struct io_async_rw *io = req->async_data;
6120 kfree(io->free_iovec);
6123 case IORING_OP_RECVMSG:
6124 case IORING_OP_SENDMSG: {
6125 struct io_async_msghdr *io = req->async_data;
6127 kfree(io->free_iov);
6130 case IORING_OP_SPLICE:
6132 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6133 io_put_file(req->splice.file_in);
6135 case IORING_OP_OPENAT:
6136 case IORING_OP_OPENAT2:
6137 if (req->open.filename)
6138 putname(req->open.filename);
6140 case IORING_OP_RENAMEAT:
6141 putname(req->rename.oldpath);
6142 putname(req->rename.newpath);
6144 case IORING_OP_UNLINKAT:
6145 putname(req->unlink.filename);
6149 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6150 kfree(req->apoll->double_poll);
6154 if (req->flags & REQ_F_INFLIGHT) {
6155 struct io_uring_task *tctx = req->task->io_uring;
6157 atomic_dec(&tctx->inflight_tracked);
6159 if (req->flags & REQ_F_CREDS)
6160 put_cred(req->creds);
6162 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6165 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6167 struct io_ring_ctx *ctx = req->ctx;
6168 const struct cred *creds = NULL;
6171 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6172 creds = override_creds(req->creds);
6174 switch (req->opcode) {
6176 ret = io_nop(req, issue_flags);
6178 case IORING_OP_READV:
6179 case IORING_OP_READ_FIXED:
6180 case IORING_OP_READ:
6181 ret = io_read(req, issue_flags);
6183 case IORING_OP_WRITEV:
6184 case IORING_OP_WRITE_FIXED:
6185 case IORING_OP_WRITE:
6186 ret = io_write(req, issue_flags);
6188 case IORING_OP_FSYNC:
6189 ret = io_fsync(req, issue_flags);
6191 case IORING_OP_POLL_ADD:
6192 ret = io_poll_add(req, issue_flags);
6194 case IORING_OP_POLL_REMOVE:
6195 ret = io_poll_update(req, issue_flags);
6197 case IORING_OP_SYNC_FILE_RANGE:
6198 ret = io_sync_file_range(req, issue_flags);
6200 case IORING_OP_SENDMSG:
6201 ret = io_sendmsg(req, issue_flags);
6203 case IORING_OP_SEND:
6204 ret = io_send(req, issue_flags);
6206 case IORING_OP_RECVMSG:
6207 ret = io_recvmsg(req, issue_flags);
6209 case IORING_OP_RECV:
6210 ret = io_recv(req, issue_flags);
6212 case IORING_OP_TIMEOUT:
6213 ret = io_timeout(req, issue_flags);
6215 case IORING_OP_TIMEOUT_REMOVE:
6216 ret = io_timeout_remove(req, issue_flags);
6218 case IORING_OP_ACCEPT:
6219 ret = io_accept(req, issue_flags);
6221 case IORING_OP_CONNECT:
6222 ret = io_connect(req, issue_flags);
6224 case IORING_OP_ASYNC_CANCEL:
6225 ret = io_async_cancel(req, issue_flags);
6227 case IORING_OP_FALLOCATE:
6228 ret = io_fallocate(req, issue_flags);
6230 case IORING_OP_OPENAT:
6231 ret = io_openat(req, issue_flags);
6233 case IORING_OP_CLOSE:
6234 ret = io_close(req, issue_flags);
6236 case IORING_OP_FILES_UPDATE:
6237 ret = io_files_update(req, issue_flags);
6239 case IORING_OP_STATX:
6240 ret = io_statx(req, issue_flags);
6242 case IORING_OP_FADVISE:
6243 ret = io_fadvise(req, issue_flags);
6245 case IORING_OP_MADVISE:
6246 ret = io_madvise(req, issue_flags);
6248 case IORING_OP_OPENAT2:
6249 ret = io_openat2(req, issue_flags);
6251 case IORING_OP_EPOLL_CTL:
6252 ret = io_epoll_ctl(req, issue_flags);
6254 case IORING_OP_SPLICE:
6255 ret = io_splice(req, issue_flags);
6257 case IORING_OP_PROVIDE_BUFFERS:
6258 ret = io_provide_buffers(req, issue_flags);
6260 case IORING_OP_REMOVE_BUFFERS:
6261 ret = io_remove_buffers(req, issue_flags);
6264 ret = io_tee(req, issue_flags);
6266 case IORING_OP_SHUTDOWN:
6267 ret = io_shutdown(req, issue_flags);
6269 case IORING_OP_RENAMEAT:
6270 ret = io_renameat(req, issue_flags);
6272 case IORING_OP_UNLINKAT:
6273 ret = io_unlinkat(req, issue_flags);
6281 revert_creds(creds);
6284 /* If the op doesn't have a file, we're not polling for it */
6285 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6286 io_iopoll_req_issued(req);
6291 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6293 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6295 req = io_put_req_find_next(req);
6296 return req ? &req->work : NULL;
6299 static void io_wq_submit_work(struct io_wq_work *work)
6301 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6302 struct io_kiocb *timeout;
6305 timeout = io_prep_linked_timeout(req);
6307 io_queue_linked_timeout(timeout);
6309 if (work->flags & IO_WQ_WORK_CANCEL)
6314 ret = io_issue_sqe(req, 0);
6316 * We can get EAGAIN for polled IO even though we're
6317 * forcing a sync submission from here, since we can't
6318 * wait for request slots on the block side.
6326 /* avoid locking problems by failing it from a clean context */
6328 /* io-wq is going to take one down */
6330 io_req_task_queue_fail(req, ret);
6334 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6337 return &table->files[i];
6340 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6343 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6345 return (struct file *) (slot->file_ptr & FFS_MASK);
6348 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6350 unsigned long file_ptr = (unsigned long) file;
6352 if (__io_file_supports_nowait(file, READ))
6353 file_ptr |= FFS_ASYNC_READ;
6354 if (__io_file_supports_nowait(file, WRITE))
6355 file_ptr |= FFS_ASYNC_WRITE;
6356 if (S_ISREG(file_inode(file)->i_mode))
6357 file_ptr |= FFS_ISREG;
6358 file_slot->file_ptr = file_ptr;
6361 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6362 struct io_kiocb *req, int fd)
6365 unsigned long file_ptr;
6367 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6369 fd = array_index_nospec(fd, ctx->nr_user_files);
6370 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6371 file = (struct file *) (file_ptr & FFS_MASK);
6372 file_ptr &= ~FFS_MASK;
6373 /* mask in overlapping REQ_F and FFS bits */
6374 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6375 io_req_set_rsrc_node(req);
6379 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6380 struct io_submit_state *state,
6381 struct io_kiocb *req, int fd)
6383 struct file *file = __io_file_get(state, fd);
6385 trace_io_uring_file_get(ctx, fd);
6387 /* we don't allow fixed io_uring files */
6388 if (file && unlikely(file->f_op == &io_uring_fops))
6389 io_req_track_inflight(req);
6393 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6394 struct io_submit_state *state,
6395 struct io_kiocb *req, int fd, bool fixed)
6398 return io_file_get_fixed(ctx, req, fd);
6400 return io_file_get_normal(ctx, state, req, fd);
6403 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6405 struct io_timeout_data *data = container_of(timer,
6406 struct io_timeout_data, timer);
6407 struct io_kiocb *prev, *req = data->req;
6408 struct io_ring_ctx *ctx = req->ctx;
6409 unsigned long flags;
6411 spin_lock_irqsave(&ctx->completion_lock, flags);
6412 prev = req->timeout.head;
6413 req->timeout.head = NULL;
6416 * We don't expect the list to be empty, that will only happen if we
6417 * race with the completion of the linked work.
6420 io_remove_next_linked(prev);
6421 if (!req_ref_inc_not_zero(prev))
6424 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6427 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6428 io_put_req_deferred(prev, 1);
6429 io_put_req_deferred(req, 1);
6431 io_req_complete_post(req, -ETIME, 0);
6433 return HRTIMER_NORESTART;
6436 static void io_queue_linked_timeout(struct io_kiocb *req)
6438 struct io_ring_ctx *ctx = req->ctx;
6440 spin_lock_irq(&ctx->completion_lock);
6442 * If the back reference is NULL, then our linked request finished
6443 * before we got a chance to setup the timer
6445 if (req->timeout.head) {
6446 struct io_timeout_data *data = req->async_data;
6448 data->timer.function = io_link_timeout_fn;
6449 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6452 spin_unlock_irq(&ctx->completion_lock);
6453 /* drop submission reference */
6457 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6459 struct io_kiocb *nxt = req->link;
6461 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6462 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6465 nxt->timeout.head = req;
6466 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6467 req->flags |= REQ_F_LINK_TIMEOUT;
6471 static void __io_queue_sqe(struct io_kiocb *req)
6472 __must_hold(&req->ctx->uring_lock)
6474 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6478 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6481 * We async punt it if the file wasn't marked NOWAIT, or if the file
6482 * doesn't support non-blocking read/write attempts
6485 /* drop submission reference */
6486 if (req->flags & REQ_F_COMPLETE_INLINE) {
6487 struct io_ring_ctx *ctx = req->ctx;
6488 struct io_comp_state *cs = &ctx->submit_state.comp;
6490 cs->reqs[cs->nr++] = req;
6491 if (cs->nr == ARRAY_SIZE(cs->reqs))
6492 io_submit_flush_completions(ctx);
6496 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6497 switch (io_arm_poll_handler(req)) {
6498 case IO_APOLL_READY:
6500 case IO_APOLL_ABORTED:
6502 * Queued up for async execution, worker will release
6503 * submit reference when the iocb is actually submitted.
6505 io_queue_async_work(req);
6509 io_req_complete_failed(req, ret);
6512 io_queue_linked_timeout(linked_timeout);
6515 static inline void io_queue_sqe(struct io_kiocb *req)
6516 __must_hold(&req->ctx->uring_lock)
6518 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6521 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6522 __io_queue_sqe(req);
6524 int ret = io_req_prep_async(req);
6527 io_req_complete_failed(req, ret);
6529 io_queue_async_work(req);
6534 * Check SQE restrictions (opcode and flags).
6536 * Returns 'true' if SQE is allowed, 'false' otherwise.
6538 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6539 struct io_kiocb *req,
6540 unsigned int sqe_flags)
6542 if (likely(!ctx->restricted))
6545 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6548 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6549 ctx->restrictions.sqe_flags_required)
6552 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6553 ctx->restrictions.sqe_flags_required))
6559 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6560 const struct io_uring_sqe *sqe)
6561 __must_hold(&ctx->uring_lock)
6563 struct io_submit_state *state;
6564 unsigned int sqe_flags;
6565 int personality, ret = 0;
6567 /* req is partially pre-initialised, see io_preinit_req() */
6568 req->opcode = READ_ONCE(sqe->opcode);
6569 /* same numerical values with corresponding REQ_F_*, safe to copy */
6570 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6571 req->user_data = READ_ONCE(sqe->user_data);
6573 req->fixed_rsrc_refs = NULL;
6574 /* one is dropped after submission, the other at completion */
6575 atomic_set(&req->refs, 2);
6576 req->task = current;
6578 /* enforce forwards compatibility on users */
6579 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6581 if (unlikely(req->opcode >= IORING_OP_LAST))
6583 if (!io_check_restriction(ctx, req, sqe_flags))
6586 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6587 !io_op_defs[req->opcode].buffer_select)
6589 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6590 ctx->drain_active = true;
6592 personality = READ_ONCE(sqe->personality);
6594 req->creds = xa_load(&ctx->personalities, personality);
6597 get_cred(req->creds);
6598 req->flags |= REQ_F_CREDS;
6600 state = &ctx->submit_state;
6603 * Plug now if we have more than 1 IO left after this, and the target
6604 * is potentially a read/write to block based storage.
6606 if (!state->plug_started && state->ios_left > 1 &&
6607 io_op_defs[req->opcode].plug) {
6608 blk_start_plug(&state->plug);
6609 state->plug_started = true;
6612 if (io_op_defs[req->opcode].needs_file) {
6613 req->file = io_file_get(ctx, state, req, READ_ONCE(sqe->fd),
6614 (sqe_flags & IOSQE_FIXED_FILE));
6615 if (unlikely(!req->file))
6623 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6624 const struct io_uring_sqe *sqe)
6625 __must_hold(&ctx->uring_lock)
6627 struct io_submit_link *link = &ctx->submit_state.link;
6630 ret = io_init_req(ctx, req, sqe);
6631 if (unlikely(ret)) {
6634 /* fail even hard links since we don't submit */
6635 req_set_fail(link->head);
6636 io_req_complete_failed(link->head, -ECANCELED);
6639 io_req_complete_failed(req, ret);
6643 ret = io_req_prep(req, sqe);
6647 /* don't need @sqe from now on */
6648 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6650 ctx->flags & IORING_SETUP_SQPOLL);
6653 * If we already have a head request, queue this one for async
6654 * submittal once the head completes. If we don't have a head but
6655 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6656 * submitted sync once the chain is complete. If none of those
6657 * conditions are true (normal request), then just queue it.
6660 struct io_kiocb *head = link->head;
6662 ret = io_req_prep_async(req);
6665 trace_io_uring_link(ctx, req, head);
6666 link->last->link = req;
6669 /* last request of a link, enqueue the link */
6670 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6675 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6687 * Batched submission is done, ensure local IO is flushed out.
6689 static void io_submit_state_end(struct io_submit_state *state,
6690 struct io_ring_ctx *ctx)
6692 if (state->link.head)
6693 io_queue_sqe(state->link.head);
6695 io_submit_flush_completions(ctx);
6696 if (state->plug_started)
6697 blk_finish_plug(&state->plug);
6698 io_state_file_put(state);
6702 * Start submission side cache.
6704 static void io_submit_state_start(struct io_submit_state *state,
6705 unsigned int max_ios)
6707 state->plug_started = false;
6708 state->ios_left = max_ios;
6709 /* set only head, no need to init link_last in advance */
6710 state->link.head = NULL;
6713 static void io_commit_sqring(struct io_ring_ctx *ctx)
6715 struct io_rings *rings = ctx->rings;
6718 * Ensure any loads from the SQEs are done at this point,
6719 * since once we write the new head, the application could
6720 * write new data to them.
6722 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6726 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6727 * that is mapped by userspace. This means that care needs to be taken to
6728 * ensure that reads are stable, as we cannot rely on userspace always
6729 * being a good citizen. If members of the sqe are validated and then later
6730 * used, it's important that those reads are done through READ_ONCE() to
6731 * prevent a re-load down the line.
6733 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6735 unsigned head, mask = ctx->sq_entries - 1;
6736 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6739 * The cached sq head (or cq tail) serves two purposes:
6741 * 1) allows us to batch the cost of updating the user visible
6743 * 2) allows the kernel side to track the head on its own, even
6744 * though the application is the one updating it.
6746 head = READ_ONCE(ctx->sq_array[sq_idx]);
6747 if (likely(head < ctx->sq_entries))
6748 return &ctx->sq_sqes[head];
6750 /* drop invalid entries */
6752 WRITE_ONCE(ctx->rings->sq_dropped,
6753 READ_ONCE(ctx->rings->sq_dropped) + 1);
6757 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6758 __must_hold(&ctx->uring_lock)
6760 struct io_uring_task *tctx;
6763 /* make sure SQ entry isn't read before tail */
6764 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6765 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6768 tctx = current->io_uring;
6769 tctx->cached_refs -= nr;
6770 if (unlikely(tctx->cached_refs < 0)) {
6771 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6773 percpu_counter_add(&tctx->inflight, refill);
6774 refcount_add(refill, ¤t->usage);
6775 tctx->cached_refs += refill;
6777 io_submit_state_start(&ctx->submit_state, nr);
6779 while (submitted < nr) {
6780 const struct io_uring_sqe *sqe;
6781 struct io_kiocb *req;
6783 req = io_alloc_req(ctx);
6784 if (unlikely(!req)) {
6786 submitted = -EAGAIN;
6789 sqe = io_get_sqe(ctx);
6790 if (unlikely(!sqe)) {
6791 kmem_cache_free(req_cachep, req);
6794 /* will complete beyond this point, count as submitted */
6796 if (io_submit_sqe(ctx, req, sqe))
6800 if (unlikely(submitted != nr)) {
6801 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6802 int unused = nr - ref_used;
6804 current->io_uring->cached_refs += unused;
6805 percpu_ref_put_many(&ctx->refs, unused);
6808 io_submit_state_end(&ctx->submit_state, ctx);
6809 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6810 io_commit_sqring(ctx);
6815 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6817 return READ_ONCE(sqd->state);
6820 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6822 /* Tell userspace we may need a wakeup call */
6823 spin_lock_irq(&ctx->completion_lock);
6824 WRITE_ONCE(ctx->rings->sq_flags,
6825 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
6826 spin_unlock_irq(&ctx->completion_lock);
6829 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6831 spin_lock_irq(&ctx->completion_lock);
6832 WRITE_ONCE(ctx->rings->sq_flags,
6833 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
6834 spin_unlock_irq(&ctx->completion_lock);
6837 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6839 unsigned int to_submit;
6842 to_submit = io_sqring_entries(ctx);
6843 /* if we're handling multiple rings, cap submit size for fairness */
6844 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6845 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6847 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6848 unsigned nr_events = 0;
6849 const struct cred *creds = NULL;
6851 if (ctx->sq_creds != current_cred())
6852 creds = override_creds(ctx->sq_creds);
6854 mutex_lock(&ctx->uring_lock);
6855 if (!list_empty(&ctx->iopoll_list))
6856 io_do_iopoll(ctx, &nr_events, 0, true);
6859 * Don't submit if refs are dying, good for io_uring_register(),
6860 * but also it is relied upon by io_ring_exit_work()
6862 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6863 !(ctx->flags & IORING_SETUP_R_DISABLED))
6864 ret = io_submit_sqes(ctx, to_submit);
6865 mutex_unlock(&ctx->uring_lock);
6867 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6868 wake_up(&ctx->sqo_sq_wait);
6870 revert_creds(creds);
6876 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6878 struct io_ring_ctx *ctx;
6879 unsigned sq_thread_idle = 0;
6881 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6882 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6883 sqd->sq_thread_idle = sq_thread_idle;
6886 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6888 bool did_sig = false;
6889 struct ksignal ksig;
6891 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6892 signal_pending(current)) {
6893 mutex_unlock(&sqd->lock);
6894 if (signal_pending(current))
6895 did_sig = get_signal(&ksig);
6897 mutex_lock(&sqd->lock);
6899 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6902 static int io_sq_thread(void *data)
6904 struct io_sq_data *sqd = data;
6905 struct io_ring_ctx *ctx;
6906 unsigned long timeout = 0;
6907 char buf[TASK_COMM_LEN];
6910 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6911 set_task_comm(current, buf);
6913 if (sqd->sq_cpu != -1)
6914 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6916 set_cpus_allowed_ptr(current, cpu_online_mask);
6917 current->flags |= PF_NO_SETAFFINITY;
6919 mutex_lock(&sqd->lock);
6921 bool cap_entries, sqt_spin = false;
6923 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6924 if (io_sqd_handle_event(sqd))
6926 timeout = jiffies + sqd->sq_thread_idle;
6929 cap_entries = !list_is_singular(&sqd->ctx_list);
6930 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6931 int ret = __io_sq_thread(ctx, cap_entries);
6933 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6936 if (io_run_task_work())
6939 if (sqt_spin || !time_after(jiffies, timeout)) {
6942 timeout = jiffies + sqd->sq_thread_idle;
6946 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6947 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6948 bool needs_sched = true;
6950 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6951 io_ring_set_wakeup_flag(ctx);
6953 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6954 !list_empty_careful(&ctx->iopoll_list)) {
6955 needs_sched = false;
6958 if (io_sqring_entries(ctx)) {
6959 needs_sched = false;
6965 mutex_unlock(&sqd->lock);
6967 mutex_lock(&sqd->lock);
6969 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6970 io_ring_clear_wakeup_flag(ctx);
6973 finish_wait(&sqd->wait, &wait);
6974 timeout = jiffies + sqd->sq_thread_idle;
6977 io_uring_cancel_generic(true, sqd);
6979 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6980 io_ring_set_wakeup_flag(ctx);
6982 mutex_unlock(&sqd->lock);
6984 complete(&sqd->exited);
6988 struct io_wait_queue {
6989 struct wait_queue_entry wq;
6990 struct io_ring_ctx *ctx;
6992 unsigned nr_timeouts;
6995 static inline bool io_should_wake(struct io_wait_queue *iowq)
6997 struct io_ring_ctx *ctx = iowq->ctx;
6998 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7001 * Wake up if we have enough events, or if a timeout occurred since we
7002 * started waiting. For timeouts, we always want to return to userspace,
7003 * regardless of event count.
7005 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7008 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7009 int wake_flags, void *key)
7011 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7015 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7016 * the task, and the next invocation will do it.
7018 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7019 return autoremove_wake_function(curr, mode, wake_flags, key);
7023 static int io_run_task_work_sig(void)
7025 if (io_run_task_work())
7027 if (!signal_pending(current))
7029 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7030 return -ERESTARTSYS;
7034 /* when returns >0, the caller should retry */
7035 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7036 struct io_wait_queue *iowq,
7037 signed long *timeout)
7041 /* make sure we run task_work before checking for signals */
7042 ret = io_run_task_work_sig();
7043 if (ret || io_should_wake(iowq))
7045 /* let the caller flush overflows, retry */
7046 if (test_bit(0, &ctx->check_cq_overflow))
7049 *timeout = schedule_timeout(*timeout);
7050 return !*timeout ? -ETIME : 1;
7054 * Wait until events become available, if we don't already have some. The
7055 * application must reap them itself, as they reside on the shared cq ring.
7057 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7058 const sigset_t __user *sig, size_t sigsz,
7059 struct __kernel_timespec __user *uts)
7061 struct io_wait_queue iowq;
7062 struct io_rings *rings = ctx->rings;
7063 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7067 io_cqring_overflow_flush(ctx, false);
7068 if (io_cqring_events(ctx) >= min_events)
7070 if (!io_run_task_work())
7075 #ifdef CONFIG_COMPAT
7076 if (in_compat_syscall())
7077 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7081 ret = set_user_sigmask(sig, sigsz);
7088 struct timespec64 ts;
7090 if (get_timespec64(&ts, uts))
7092 timeout = timespec64_to_jiffies(&ts);
7095 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7096 iowq.wq.private = current;
7097 INIT_LIST_HEAD(&iowq.wq.entry);
7099 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7100 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7102 trace_io_uring_cqring_wait(ctx, min_events);
7104 /* if we can't even flush overflow, don't wait for more */
7105 if (!io_cqring_overflow_flush(ctx, false)) {
7109 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7110 TASK_INTERRUPTIBLE);
7111 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7112 finish_wait(&ctx->cq_wait, &iowq.wq);
7116 restore_saved_sigmask_unless(ret == -EINTR);
7118 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7121 static void io_free_page_table(void **table, size_t size)
7123 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7125 for (i = 0; i < nr_tables; i++)
7130 static void **io_alloc_page_table(size_t size)
7132 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7133 size_t init_size = size;
7136 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7140 for (i = 0; i < nr_tables; i++) {
7141 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7143 table[i] = kzalloc(this_size, GFP_KERNEL);
7145 io_free_page_table(table, init_size);
7153 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7155 percpu_ref_exit(&ref_node->refs);
7159 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7161 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7162 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7163 unsigned long flags;
7164 bool first_add = false;
7166 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7169 while (!list_empty(&ctx->rsrc_ref_list)) {
7170 node = list_first_entry(&ctx->rsrc_ref_list,
7171 struct io_rsrc_node, node);
7172 /* recycle ref nodes in order */
7175 list_del(&node->node);
7176 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7178 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7181 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7184 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7186 struct io_rsrc_node *ref_node;
7188 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7192 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7197 INIT_LIST_HEAD(&ref_node->node);
7198 INIT_LIST_HEAD(&ref_node->rsrc_list);
7199 ref_node->done = false;
7203 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7204 struct io_rsrc_data *data_to_kill)
7206 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7207 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7210 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7212 rsrc_node->rsrc_data = data_to_kill;
7213 spin_lock_irq(&ctx->rsrc_ref_lock);
7214 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7215 spin_unlock_irq(&ctx->rsrc_ref_lock);
7217 atomic_inc(&data_to_kill->refs);
7218 percpu_ref_kill(&rsrc_node->refs);
7219 ctx->rsrc_node = NULL;
7222 if (!ctx->rsrc_node) {
7223 ctx->rsrc_node = ctx->rsrc_backup_node;
7224 ctx->rsrc_backup_node = NULL;
7228 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7230 if (ctx->rsrc_backup_node)
7232 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7233 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7236 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7240 /* As we may drop ->uring_lock, other task may have started quiesce */
7244 data->quiesce = true;
7246 ret = io_rsrc_node_switch_start(ctx);
7249 io_rsrc_node_switch(ctx, data);
7251 /* kill initial ref, already quiesced if zero */
7252 if (atomic_dec_and_test(&data->refs))
7254 mutex_unlock(&ctx->uring_lock);
7255 flush_delayed_work(&ctx->rsrc_put_work);
7256 ret = wait_for_completion_interruptible(&data->done);
7258 mutex_lock(&ctx->uring_lock);
7262 atomic_inc(&data->refs);
7263 /* wait for all works potentially completing data->done */
7264 flush_delayed_work(&ctx->rsrc_put_work);
7265 reinit_completion(&data->done);
7267 ret = io_run_task_work_sig();
7268 mutex_lock(&ctx->uring_lock);
7270 data->quiesce = false;
7275 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7277 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7278 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7280 return &data->tags[table_idx][off];
7283 static void io_rsrc_data_free(struct io_rsrc_data *data)
7285 size_t size = data->nr * sizeof(data->tags[0][0]);
7288 io_free_page_table((void **)data->tags, size);
7292 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7293 u64 __user *utags, unsigned nr,
7294 struct io_rsrc_data **pdata)
7296 struct io_rsrc_data *data;
7300 data = kzalloc(sizeof(*data), GFP_KERNEL);
7303 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7311 data->do_put = do_put;
7314 for (i = 0; i < nr; i++) {
7315 u64 *tag_slot = io_get_tag_slot(data, i);
7317 if (copy_from_user(tag_slot, &utags[i],
7323 atomic_set(&data->refs, 1);
7324 init_completion(&data->done);
7328 io_rsrc_data_free(data);
7332 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7334 table->files = kvcalloc(nr_files, sizeof(table->files[0]), GFP_KERNEL);
7335 return !!table->files;
7338 static void io_free_file_tables(struct io_file_table *table)
7340 kvfree(table->files);
7341 table->files = NULL;
7344 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7346 #if defined(CONFIG_UNIX)
7347 if (ctx->ring_sock) {
7348 struct sock *sock = ctx->ring_sock->sk;
7349 struct sk_buff *skb;
7351 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7357 for (i = 0; i < ctx->nr_user_files; i++) {
7360 file = io_file_from_index(ctx, i);
7365 io_free_file_tables(&ctx->file_table);
7366 io_rsrc_data_free(ctx->file_data);
7367 ctx->file_data = NULL;
7368 ctx->nr_user_files = 0;
7371 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7375 if (!ctx->file_data)
7377 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7379 __io_sqe_files_unregister(ctx);
7383 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7384 __releases(&sqd->lock)
7386 WARN_ON_ONCE(sqd->thread == current);
7389 * Do the dance but not conditional clear_bit() because it'd race with
7390 * other threads incrementing park_pending and setting the bit.
7392 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7393 if (atomic_dec_return(&sqd->park_pending))
7394 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7395 mutex_unlock(&sqd->lock);
7398 static void io_sq_thread_park(struct io_sq_data *sqd)
7399 __acquires(&sqd->lock)
7401 WARN_ON_ONCE(sqd->thread == current);
7403 atomic_inc(&sqd->park_pending);
7404 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7405 mutex_lock(&sqd->lock);
7407 wake_up_process(sqd->thread);
7410 static void io_sq_thread_stop(struct io_sq_data *sqd)
7412 WARN_ON_ONCE(sqd->thread == current);
7413 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7415 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7416 mutex_lock(&sqd->lock);
7418 wake_up_process(sqd->thread);
7419 mutex_unlock(&sqd->lock);
7420 wait_for_completion(&sqd->exited);
7423 static void io_put_sq_data(struct io_sq_data *sqd)
7425 if (refcount_dec_and_test(&sqd->refs)) {
7426 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7428 io_sq_thread_stop(sqd);
7433 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7435 struct io_sq_data *sqd = ctx->sq_data;
7438 io_sq_thread_park(sqd);
7439 list_del_init(&ctx->sqd_list);
7440 io_sqd_update_thread_idle(sqd);
7441 io_sq_thread_unpark(sqd);
7443 io_put_sq_data(sqd);
7444 ctx->sq_data = NULL;
7448 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7450 struct io_ring_ctx *ctx_attach;
7451 struct io_sq_data *sqd;
7454 f = fdget(p->wq_fd);
7456 return ERR_PTR(-ENXIO);
7457 if (f.file->f_op != &io_uring_fops) {
7459 return ERR_PTR(-EINVAL);
7462 ctx_attach = f.file->private_data;
7463 sqd = ctx_attach->sq_data;
7466 return ERR_PTR(-EINVAL);
7468 if (sqd->task_tgid != current->tgid) {
7470 return ERR_PTR(-EPERM);
7473 refcount_inc(&sqd->refs);
7478 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7481 struct io_sq_data *sqd;
7484 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7485 sqd = io_attach_sq_data(p);
7490 /* fall through for EPERM case, setup new sqd/task */
7491 if (PTR_ERR(sqd) != -EPERM)
7495 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7497 return ERR_PTR(-ENOMEM);
7499 atomic_set(&sqd->park_pending, 0);
7500 refcount_set(&sqd->refs, 1);
7501 INIT_LIST_HEAD(&sqd->ctx_list);
7502 mutex_init(&sqd->lock);
7503 init_waitqueue_head(&sqd->wait);
7504 init_completion(&sqd->exited);
7508 #if defined(CONFIG_UNIX)
7510 * Ensure the UNIX gc is aware of our file set, so we are certain that
7511 * the io_uring can be safely unregistered on process exit, even if we have
7512 * loops in the file referencing.
7514 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7516 struct sock *sk = ctx->ring_sock->sk;
7517 struct scm_fp_list *fpl;
7518 struct sk_buff *skb;
7521 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7525 skb = alloc_skb(0, GFP_KERNEL);
7534 fpl->user = get_uid(current_user());
7535 for (i = 0; i < nr; i++) {
7536 struct file *file = io_file_from_index(ctx, i + offset);
7540 fpl->fp[nr_files] = get_file(file);
7541 unix_inflight(fpl->user, fpl->fp[nr_files]);
7546 fpl->max = SCM_MAX_FD;
7547 fpl->count = nr_files;
7548 UNIXCB(skb).fp = fpl;
7549 skb->destructor = unix_destruct_scm;
7550 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7551 skb_queue_head(&sk->sk_receive_queue, skb);
7553 for (i = 0; i < nr_files; i++)
7564 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7565 * causes regular reference counting to break down. We rely on the UNIX
7566 * garbage collection to take care of this problem for us.
7568 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7570 unsigned left, total;
7574 left = ctx->nr_user_files;
7576 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7578 ret = __io_sqe_files_scm(ctx, this_files, total);
7582 total += this_files;
7588 while (total < ctx->nr_user_files) {
7589 struct file *file = io_file_from_index(ctx, total);
7599 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7605 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7607 struct file *file = prsrc->file;
7608 #if defined(CONFIG_UNIX)
7609 struct sock *sock = ctx->ring_sock->sk;
7610 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7611 struct sk_buff *skb;
7614 __skb_queue_head_init(&list);
7617 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7618 * remove this entry and rearrange the file array.
7620 skb = skb_dequeue(head);
7622 struct scm_fp_list *fp;
7624 fp = UNIXCB(skb).fp;
7625 for (i = 0; i < fp->count; i++) {
7628 if (fp->fp[i] != file)
7631 unix_notinflight(fp->user, fp->fp[i]);
7632 left = fp->count - 1 - i;
7634 memmove(&fp->fp[i], &fp->fp[i + 1],
7635 left * sizeof(struct file *));
7642 __skb_queue_tail(&list, skb);
7652 __skb_queue_tail(&list, skb);
7654 skb = skb_dequeue(head);
7657 if (skb_peek(&list)) {
7658 spin_lock_irq(&head->lock);
7659 while ((skb = __skb_dequeue(&list)) != NULL)
7660 __skb_queue_tail(head, skb);
7661 spin_unlock_irq(&head->lock);
7668 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7670 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7671 struct io_ring_ctx *ctx = rsrc_data->ctx;
7672 struct io_rsrc_put *prsrc, *tmp;
7674 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7675 list_del(&prsrc->list);
7678 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7680 io_ring_submit_lock(ctx, lock_ring);
7681 spin_lock_irq(&ctx->completion_lock);
7682 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7684 io_commit_cqring(ctx);
7685 spin_unlock_irq(&ctx->completion_lock);
7686 io_cqring_ev_posted(ctx);
7687 io_ring_submit_unlock(ctx, lock_ring);
7690 rsrc_data->do_put(ctx, prsrc);
7694 io_rsrc_node_destroy(ref_node);
7695 if (atomic_dec_and_test(&rsrc_data->refs))
7696 complete(&rsrc_data->done);
7699 static void io_rsrc_put_work(struct work_struct *work)
7701 struct io_ring_ctx *ctx;
7702 struct llist_node *node;
7704 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7705 node = llist_del_all(&ctx->rsrc_put_llist);
7708 struct io_rsrc_node *ref_node;
7709 struct llist_node *next = node->next;
7711 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7712 __io_rsrc_put_work(ref_node);
7717 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7718 unsigned nr_args, u64 __user *tags)
7720 __s32 __user *fds = (__s32 __user *) arg;
7729 if (nr_args > IORING_MAX_FIXED_FILES)
7731 ret = io_rsrc_node_switch_start(ctx);
7734 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7740 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7743 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7744 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7748 /* allow sparse sets */
7751 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7758 if (unlikely(!file))
7762 * Don't allow io_uring instances to be registered. If UNIX
7763 * isn't enabled, then this causes a reference cycle and this
7764 * instance can never get freed. If UNIX is enabled we'll
7765 * handle it just fine, but there's still no point in allowing
7766 * a ring fd as it doesn't support regular read/write anyway.
7768 if (file->f_op == &io_uring_fops) {
7772 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7775 ret = io_sqe_files_scm(ctx);
7777 __io_sqe_files_unregister(ctx);
7781 io_rsrc_node_switch(ctx, NULL);
7784 for (i = 0; i < ctx->nr_user_files; i++) {
7785 file = io_file_from_index(ctx, i);
7789 io_free_file_tables(&ctx->file_table);
7790 ctx->nr_user_files = 0;
7792 io_rsrc_data_free(ctx->file_data);
7793 ctx->file_data = NULL;
7797 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7800 #if defined(CONFIG_UNIX)
7801 struct sock *sock = ctx->ring_sock->sk;
7802 struct sk_buff_head *head = &sock->sk_receive_queue;
7803 struct sk_buff *skb;
7806 * See if we can merge this file into an existing skb SCM_RIGHTS
7807 * file set. If there's no room, fall back to allocating a new skb
7808 * and filling it in.
7810 spin_lock_irq(&head->lock);
7811 skb = skb_peek(head);
7813 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7815 if (fpl->count < SCM_MAX_FD) {
7816 __skb_unlink(skb, head);
7817 spin_unlock_irq(&head->lock);
7818 fpl->fp[fpl->count] = get_file(file);
7819 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7821 spin_lock_irq(&head->lock);
7822 __skb_queue_head(head, skb);
7827 spin_unlock_irq(&head->lock);
7834 return __io_sqe_files_scm(ctx, 1, index);
7840 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7841 struct io_rsrc_node *node, void *rsrc)
7843 struct io_rsrc_put *prsrc;
7845 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7849 prsrc->tag = *io_get_tag_slot(data, idx);
7851 list_add(&prsrc->list, &node->rsrc_list);
7855 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7856 struct io_uring_rsrc_update2 *up,
7859 u64 __user *tags = u64_to_user_ptr(up->tags);
7860 __s32 __user *fds = u64_to_user_ptr(up->data);
7861 struct io_rsrc_data *data = ctx->file_data;
7862 struct io_fixed_file *file_slot;
7866 bool needs_switch = false;
7868 if (!ctx->file_data)
7870 if (up->offset + nr_args > ctx->nr_user_files)
7873 for (done = 0; done < nr_args; done++) {
7876 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7877 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7881 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7885 if (fd == IORING_REGISTER_FILES_SKIP)
7888 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7889 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7891 if (file_slot->file_ptr) {
7892 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7893 err = io_queue_rsrc_removal(data, up->offset + done,
7894 ctx->rsrc_node, file);
7897 file_slot->file_ptr = 0;
7898 needs_switch = true;
7907 * Don't allow io_uring instances to be registered. If
7908 * UNIX isn't enabled, then this causes a reference
7909 * cycle and this instance can never get freed. If UNIX
7910 * is enabled we'll handle it just fine, but there's
7911 * still no point in allowing a ring fd as it doesn't
7912 * support regular read/write anyway.
7914 if (file->f_op == &io_uring_fops) {
7919 *io_get_tag_slot(data, up->offset + done) = tag;
7920 io_fixed_file_set(file_slot, file);
7921 err = io_sqe_file_register(ctx, file, i);
7923 file_slot->file_ptr = 0;
7931 io_rsrc_node_switch(ctx, data);
7932 return done ? done : err;
7935 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7936 struct task_struct *task)
7938 struct io_wq_hash *hash;
7939 struct io_wq_data data;
7940 unsigned int concurrency;
7942 mutex_lock(&ctx->uring_lock);
7943 hash = ctx->hash_map;
7945 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7947 mutex_unlock(&ctx->uring_lock);
7948 return ERR_PTR(-ENOMEM);
7950 refcount_set(&hash->refs, 1);
7951 init_waitqueue_head(&hash->wait);
7952 ctx->hash_map = hash;
7954 mutex_unlock(&ctx->uring_lock);
7958 data.free_work = io_wq_free_work;
7959 data.do_work = io_wq_submit_work;
7961 /* Do QD, or 4 * CPUS, whatever is smallest */
7962 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7964 return io_wq_create(concurrency, &data);
7967 static int io_uring_alloc_task_context(struct task_struct *task,
7968 struct io_ring_ctx *ctx)
7970 struct io_uring_task *tctx;
7973 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7974 if (unlikely(!tctx))
7977 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7978 if (unlikely(ret)) {
7983 tctx->io_wq = io_init_wq_offload(ctx, task);
7984 if (IS_ERR(tctx->io_wq)) {
7985 ret = PTR_ERR(tctx->io_wq);
7986 percpu_counter_destroy(&tctx->inflight);
7992 init_waitqueue_head(&tctx->wait);
7993 atomic_set(&tctx->in_idle, 0);
7994 atomic_set(&tctx->inflight_tracked, 0);
7995 task->io_uring = tctx;
7996 spin_lock_init(&tctx->task_lock);
7997 INIT_WQ_LIST(&tctx->task_list);
7998 init_task_work(&tctx->task_work, tctx_task_work);
8002 void __io_uring_free(struct task_struct *tsk)
8004 struct io_uring_task *tctx = tsk->io_uring;
8006 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8007 WARN_ON_ONCE(tctx->io_wq);
8008 WARN_ON_ONCE(tctx->cached_refs);
8010 percpu_counter_destroy(&tctx->inflight);
8012 tsk->io_uring = NULL;
8015 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8016 struct io_uring_params *p)
8020 /* Retain compatibility with failing for an invalid attach attempt */
8021 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8022 IORING_SETUP_ATTACH_WQ) {
8025 f = fdget(p->wq_fd);
8028 if (f.file->f_op != &io_uring_fops) {
8034 if (ctx->flags & IORING_SETUP_SQPOLL) {
8035 struct task_struct *tsk;
8036 struct io_sq_data *sqd;
8039 sqd = io_get_sq_data(p, &attached);
8045 ctx->sq_creds = get_current_cred();
8047 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8048 if (!ctx->sq_thread_idle)
8049 ctx->sq_thread_idle = HZ;
8051 io_sq_thread_park(sqd);
8052 list_add(&ctx->sqd_list, &sqd->ctx_list);
8053 io_sqd_update_thread_idle(sqd);
8054 /* don't attach to a dying SQPOLL thread, would be racy */
8055 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8056 io_sq_thread_unpark(sqd);
8063 if (p->flags & IORING_SETUP_SQ_AFF) {
8064 int cpu = p->sq_thread_cpu;
8067 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8074 sqd->task_pid = current->pid;
8075 sqd->task_tgid = current->tgid;
8076 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8083 ret = io_uring_alloc_task_context(tsk, ctx);
8084 wake_up_new_task(tsk);
8087 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8088 /* Can't have SQ_AFF without SQPOLL */
8095 complete(&ctx->sq_data->exited);
8097 io_sq_thread_finish(ctx);
8101 static inline void __io_unaccount_mem(struct user_struct *user,
8102 unsigned long nr_pages)
8104 atomic_long_sub(nr_pages, &user->locked_vm);
8107 static inline int __io_account_mem(struct user_struct *user,
8108 unsigned long nr_pages)
8110 unsigned long page_limit, cur_pages, new_pages;
8112 /* Don't allow more pages than we can safely lock */
8113 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8116 cur_pages = atomic_long_read(&user->locked_vm);
8117 new_pages = cur_pages + nr_pages;
8118 if (new_pages > page_limit)
8120 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8121 new_pages) != cur_pages);
8126 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8129 __io_unaccount_mem(ctx->user, nr_pages);
8131 if (ctx->mm_account)
8132 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8135 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8140 ret = __io_account_mem(ctx->user, nr_pages);
8145 if (ctx->mm_account)
8146 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8151 static void io_mem_free(void *ptr)
8158 page = virt_to_head_page(ptr);
8159 if (put_page_testzero(page))
8160 free_compound_page(page);
8163 static void *io_mem_alloc(size_t size)
8165 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8166 __GFP_NORETRY | __GFP_ACCOUNT;
8168 return (void *) __get_free_pages(gfp_flags, get_order(size));
8171 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8174 struct io_rings *rings;
8175 size_t off, sq_array_size;
8177 off = struct_size(rings, cqes, cq_entries);
8178 if (off == SIZE_MAX)
8182 off = ALIGN(off, SMP_CACHE_BYTES);
8190 sq_array_size = array_size(sizeof(u32), sq_entries);
8191 if (sq_array_size == SIZE_MAX)
8194 if (check_add_overflow(off, sq_array_size, &off))
8200 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8202 struct io_mapped_ubuf *imu = *slot;
8205 if (imu != ctx->dummy_ubuf) {
8206 for (i = 0; i < imu->nr_bvecs; i++)
8207 unpin_user_page(imu->bvec[i].bv_page);
8208 if (imu->acct_pages)
8209 io_unaccount_mem(ctx, imu->acct_pages);
8215 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8217 io_buffer_unmap(ctx, &prsrc->buf);
8221 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8225 for (i = 0; i < ctx->nr_user_bufs; i++)
8226 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8227 kfree(ctx->user_bufs);
8228 io_rsrc_data_free(ctx->buf_data);
8229 ctx->user_bufs = NULL;
8230 ctx->buf_data = NULL;
8231 ctx->nr_user_bufs = 0;
8234 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8241 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8243 __io_sqe_buffers_unregister(ctx);
8247 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8248 void __user *arg, unsigned index)
8250 struct iovec __user *src;
8252 #ifdef CONFIG_COMPAT
8254 struct compat_iovec __user *ciovs;
8255 struct compat_iovec ciov;
8257 ciovs = (struct compat_iovec __user *) arg;
8258 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8261 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8262 dst->iov_len = ciov.iov_len;
8266 src = (struct iovec __user *) arg;
8267 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8273 * Not super efficient, but this is just a registration time. And we do cache
8274 * the last compound head, so generally we'll only do a full search if we don't
8277 * We check if the given compound head page has already been accounted, to
8278 * avoid double accounting it. This allows us to account the full size of the
8279 * page, not just the constituent pages of a huge page.
8281 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8282 int nr_pages, struct page *hpage)
8286 /* check current page array */
8287 for (i = 0; i < nr_pages; i++) {
8288 if (!PageCompound(pages[i]))
8290 if (compound_head(pages[i]) == hpage)
8294 /* check previously registered pages */
8295 for (i = 0; i < ctx->nr_user_bufs; i++) {
8296 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8298 for (j = 0; j < imu->nr_bvecs; j++) {
8299 if (!PageCompound(imu->bvec[j].bv_page))
8301 if (compound_head(imu->bvec[j].bv_page) == hpage)
8309 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8310 int nr_pages, struct io_mapped_ubuf *imu,
8311 struct page **last_hpage)
8315 imu->acct_pages = 0;
8316 for (i = 0; i < nr_pages; i++) {
8317 if (!PageCompound(pages[i])) {
8322 hpage = compound_head(pages[i]);
8323 if (hpage == *last_hpage)
8325 *last_hpage = hpage;
8326 if (headpage_already_acct(ctx, pages, i, hpage))
8328 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8332 if (!imu->acct_pages)
8335 ret = io_account_mem(ctx, imu->acct_pages);
8337 imu->acct_pages = 0;
8341 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8342 struct io_mapped_ubuf **pimu,
8343 struct page **last_hpage)
8345 struct io_mapped_ubuf *imu = NULL;
8346 struct vm_area_struct **vmas = NULL;
8347 struct page **pages = NULL;
8348 unsigned long off, start, end, ubuf;
8350 int ret, pret, nr_pages, i;
8352 if (!iov->iov_base) {
8353 *pimu = ctx->dummy_ubuf;
8357 ubuf = (unsigned long) iov->iov_base;
8358 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8359 start = ubuf >> PAGE_SHIFT;
8360 nr_pages = end - start;
8365 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8369 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8374 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8379 mmap_read_lock(current->mm);
8380 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8382 if (pret == nr_pages) {
8383 /* don't support file backed memory */
8384 for (i = 0; i < nr_pages; i++) {
8385 struct vm_area_struct *vma = vmas[i];
8387 if (vma_is_shmem(vma))
8390 !is_file_hugepages(vma->vm_file)) {
8396 ret = pret < 0 ? pret : -EFAULT;
8398 mmap_read_unlock(current->mm);
8401 * if we did partial map, or found file backed vmas,
8402 * release any pages we did get
8405 unpin_user_pages(pages, pret);
8409 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8411 unpin_user_pages(pages, pret);
8415 off = ubuf & ~PAGE_MASK;
8416 size = iov->iov_len;
8417 for (i = 0; i < nr_pages; i++) {
8420 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8421 imu->bvec[i].bv_page = pages[i];
8422 imu->bvec[i].bv_len = vec_len;
8423 imu->bvec[i].bv_offset = off;
8427 /* store original address for later verification */
8429 imu->ubuf_end = ubuf + iov->iov_len;
8430 imu->nr_bvecs = nr_pages;
8441 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8443 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8444 return ctx->user_bufs ? 0 : -ENOMEM;
8447 static int io_buffer_validate(struct iovec *iov)
8449 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8452 * Don't impose further limits on the size and buffer
8453 * constraints here, we'll -EINVAL later when IO is
8454 * submitted if they are wrong.
8457 return iov->iov_len ? -EFAULT : 0;
8461 /* arbitrary limit, but we need something */
8462 if (iov->iov_len > SZ_1G)
8465 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8471 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8472 unsigned int nr_args, u64 __user *tags)
8474 struct page *last_hpage = NULL;
8475 struct io_rsrc_data *data;
8481 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8483 ret = io_rsrc_node_switch_start(ctx);
8486 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8489 ret = io_buffers_map_alloc(ctx, nr_args);
8491 io_rsrc_data_free(data);
8495 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8496 ret = io_copy_iov(ctx, &iov, arg, i);
8499 ret = io_buffer_validate(&iov);
8502 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8507 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8513 WARN_ON_ONCE(ctx->buf_data);
8515 ctx->buf_data = data;
8517 __io_sqe_buffers_unregister(ctx);
8519 io_rsrc_node_switch(ctx, NULL);
8523 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8524 struct io_uring_rsrc_update2 *up,
8525 unsigned int nr_args)
8527 u64 __user *tags = u64_to_user_ptr(up->tags);
8528 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8529 struct page *last_hpage = NULL;
8530 bool needs_switch = false;
8536 if (up->offset + nr_args > ctx->nr_user_bufs)
8539 for (done = 0; done < nr_args; done++) {
8540 struct io_mapped_ubuf *imu;
8541 int offset = up->offset + done;
8544 err = io_copy_iov(ctx, &iov, iovs, done);
8547 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8551 err = io_buffer_validate(&iov);
8554 if (!iov.iov_base && tag) {
8558 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8562 i = array_index_nospec(offset, ctx->nr_user_bufs);
8563 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8564 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8565 ctx->rsrc_node, ctx->user_bufs[i]);
8566 if (unlikely(err)) {
8567 io_buffer_unmap(ctx, &imu);
8570 ctx->user_bufs[i] = NULL;
8571 needs_switch = true;
8574 ctx->user_bufs[i] = imu;
8575 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8579 io_rsrc_node_switch(ctx, ctx->buf_data);
8580 return done ? done : err;
8583 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8585 __s32 __user *fds = arg;
8591 if (copy_from_user(&fd, fds, sizeof(*fds)))
8594 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8595 if (IS_ERR(ctx->cq_ev_fd)) {
8596 int ret = PTR_ERR(ctx->cq_ev_fd);
8598 ctx->cq_ev_fd = NULL;
8605 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8607 if (ctx->cq_ev_fd) {
8608 eventfd_ctx_put(ctx->cq_ev_fd);
8609 ctx->cq_ev_fd = NULL;
8616 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8618 struct io_buffer *buf;
8619 unsigned long index;
8621 xa_for_each(&ctx->io_buffers, index, buf)
8622 __io_remove_buffers(ctx, buf, index, -1U);
8625 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8627 struct io_kiocb *req, *nxt;
8629 list_for_each_entry_safe(req, nxt, list, compl.list) {
8630 if (tsk && req->task != tsk)
8632 list_del(&req->compl.list);
8633 kmem_cache_free(req_cachep, req);
8637 static void io_req_caches_free(struct io_ring_ctx *ctx)
8639 struct io_submit_state *submit_state = &ctx->submit_state;
8640 struct io_comp_state *cs = &ctx->submit_state.comp;
8642 mutex_lock(&ctx->uring_lock);
8644 if (submit_state->free_reqs) {
8645 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8646 submit_state->reqs);
8647 submit_state->free_reqs = 0;
8650 io_flush_cached_locked_reqs(ctx, cs);
8651 io_req_cache_free(&cs->free_list, NULL);
8652 mutex_unlock(&ctx->uring_lock);
8655 static void io_wait_rsrc_data(struct io_rsrc_data *data)
8657 if (data && !atomic_dec_and_test(&data->refs))
8658 wait_for_completion(&data->done);
8661 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8663 io_sq_thread_finish(ctx);
8665 if (ctx->mm_account) {
8666 mmdrop(ctx->mm_account);
8667 ctx->mm_account = NULL;
8670 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
8671 io_wait_rsrc_data(ctx->buf_data);
8672 io_wait_rsrc_data(ctx->file_data);
8674 mutex_lock(&ctx->uring_lock);
8676 __io_sqe_buffers_unregister(ctx);
8678 __io_sqe_files_unregister(ctx);
8680 __io_cqring_overflow_flush(ctx, true);
8681 mutex_unlock(&ctx->uring_lock);
8682 io_eventfd_unregister(ctx);
8683 io_destroy_buffers(ctx);
8685 put_cred(ctx->sq_creds);
8687 /* there are no registered resources left, nobody uses it */
8689 io_rsrc_node_destroy(ctx->rsrc_node);
8690 if (ctx->rsrc_backup_node)
8691 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8692 flush_delayed_work(&ctx->rsrc_put_work);
8694 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8695 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8697 #if defined(CONFIG_UNIX)
8698 if (ctx->ring_sock) {
8699 ctx->ring_sock->file = NULL; /* so that iput() is called */
8700 sock_release(ctx->ring_sock);
8704 io_mem_free(ctx->rings);
8705 io_mem_free(ctx->sq_sqes);
8707 percpu_ref_exit(&ctx->refs);
8708 free_uid(ctx->user);
8709 io_req_caches_free(ctx);
8711 io_wq_put_hash(ctx->hash_map);
8712 kfree(ctx->cancel_hash);
8713 kfree(ctx->dummy_ubuf);
8717 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8719 struct io_ring_ctx *ctx = file->private_data;
8722 poll_wait(file, &ctx->poll_wait, wait);
8724 * synchronizes with barrier from wq_has_sleeper call in
8728 if (!io_sqring_full(ctx))
8729 mask |= EPOLLOUT | EPOLLWRNORM;
8732 * Don't flush cqring overflow list here, just do a simple check.
8733 * Otherwise there could possible be ABBA deadlock:
8736 * lock(&ctx->uring_lock);
8738 * lock(&ctx->uring_lock);
8741 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8742 * pushs them to do the flush.
8744 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8745 mask |= EPOLLIN | EPOLLRDNORM;
8750 static int io_uring_fasync(int fd, struct file *file, int on)
8752 struct io_ring_ctx *ctx = file->private_data;
8754 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8757 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8759 const struct cred *creds;
8761 creds = xa_erase(&ctx->personalities, id);
8770 struct io_tctx_exit {
8771 struct callback_head task_work;
8772 struct completion completion;
8773 struct io_ring_ctx *ctx;
8776 static void io_tctx_exit_cb(struct callback_head *cb)
8778 struct io_uring_task *tctx = current->io_uring;
8779 struct io_tctx_exit *work;
8781 work = container_of(cb, struct io_tctx_exit, task_work);
8783 * When @in_idle, we're in cancellation and it's racy to remove the
8784 * node. It'll be removed by the end of cancellation, just ignore it.
8786 if (!atomic_read(&tctx->in_idle))
8787 io_uring_del_tctx_node((unsigned long)work->ctx);
8788 complete(&work->completion);
8791 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8793 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8795 return req->ctx == data;
8798 static void io_ring_exit_work(struct work_struct *work)
8800 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8801 unsigned long timeout = jiffies + HZ * 60 * 5;
8802 unsigned long interval = HZ / 20;
8803 struct io_tctx_exit exit;
8804 struct io_tctx_node *node;
8808 * If we're doing polled IO and end up having requests being
8809 * submitted async (out-of-line), then completions can come in while
8810 * we're waiting for refs to drop. We need to reap these manually,
8811 * as nobody else will be looking for them.
8814 io_uring_try_cancel_requests(ctx, NULL, true);
8816 struct io_sq_data *sqd = ctx->sq_data;
8817 struct task_struct *tsk;
8819 io_sq_thread_park(sqd);
8821 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8822 io_wq_cancel_cb(tsk->io_uring->io_wq,
8823 io_cancel_ctx_cb, ctx, true);
8824 io_sq_thread_unpark(sqd);
8827 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
8828 /* there is little hope left, don't run it too often */
8831 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
8833 init_completion(&exit.completion);
8834 init_task_work(&exit.task_work, io_tctx_exit_cb);
8837 * Some may use context even when all refs and requests have been put,
8838 * and they are free to do so while still holding uring_lock or
8839 * completion_lock, see io_req_task_submit(). Apart from other work,
8840 * this lock/unlock section also waits them to finish.
8842 mutex_lock(&ctx->uring_lock);
8843 while (!list_empty(&ctx->tctx_list)) {
8844 WARN_ON_ONCE(time_after(jiffies, timeout));
8846 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8848 /* don't spin on a single task if cancellation failed */
8849 list_rotate_left(&ctx->tctx_list);
8850 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8851 if (WARN_ON_ONCE(ret))
8853 wake_up_process(node->task);
8855 mutex_unlock(&ctx->uring_lock);
8856 wait_for_completion(&exit.completion);
8857 mutex_lock(&ctx->uring_lock);
8859 mutex_unlock(&ctx->uring_lock);
8860 spin_lock_irq(&ctx->completion_lock);
8861 spin_unlock_irq(&ctx->completion_lock);
8863 io_ring_ctx_free(ctx);
8866 /* Returns true if we found and killed one or more timeouts */
8867 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8870 struct io_kiocb *req, *tmp;
8873 spin_lock_irq(&ctx->completion_lock);
8874 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8875 if (io_match_task(req, tsk, cancel_all)) {
8876 io_kill_timeout(req, -ECANCELED);
8881 io_commit_cqring(ctx);
8882 spin_unlock_irq(&ctx->completion_lock);
8884 io_cqring_ev_posted(ctx);
8885 return canceled != 0;
8888 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8890 unsigned long index;
8891 struct creds *creds;
8893 mutex_lock(&ctx->uring_lock);
8894 percpu_ref_kill(&ctx->refs);
8896 __io_cqring_overflow_flush(ctx, true);
8897 xa_for_each(&ctx->personalities, index, creds)
8898 io_unregister_personality(ctx, index);
8899 mutex_unlock(&ctx->uring_lock);
8901 io_kill_timeouts(ctx, NULL, true);
8902 io_poll_remove_all(ctx, NULL, true);
8904 /* if we failed setting up the ctx, we might not have any rings */
8905 io_iopoll_try_reap_events(ctx);
8907 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8909 * Use system_unbound_wq to avoid spawning tons of event kworkers
8910 * if we're exiting a ton of rings at the same time. It just adds
8911 * noise and overhead, there's no discernable change in runtime
8912 * over using system_wq.
8914 queue_work(system_unbound_wq, &ctx->exit_work);
8917 static int io_uring_release(struct inode *inode, struct file *file)
8919 struct io_ring_ctx *ctx = file->private_data;
8921 file->private_data = NULL;
8922 io_ring_ctx_wait_and_kill(ctx);
8926 struct io_task_cancel {
8927 struct task_struct *task;
8931 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8933 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8934 struct io_task_cancel *cancel = data;
8937 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8938 unsigned long flags;
8939 struct io_ring_ctx *ctx = req->ctx;
8941 /* protect against races with linked timeouts */
8942 spin_lock_irqsave(&ctx->completion_lock, flags);
8943 ret = io_match_task(req, cancel->task, cancel->all);
8944 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8946 ret = io_match_task(req, cancel->task, cancel->all);
8951 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8952 struct task_struct *task, bool cancel_all)
8954 struct io_defer_entry *de;
8957 spin_lock_irq(&ctx->completion_lock);
8958 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8959 if (io_match_task(de->req, task, cancel_all)) {
8960 list_cut_position(&list, &ctx->defer_list, &de->list);
8964 spin_unlock_irq(&ctx->completion_lock);
8965 if (list_empty(&list))
8968 while (!list_empty(&list)) {
8969 de = list_first_entry(&list, struct io_defer_entry, list);
8970 list_del_init(&de->list);
8971 io_req_complete_failed(de->req, -ECANCELED);
8977 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8979 struct io_tctx_node *node;
8980 enum io_wq_cancel cret;
8983 mutex_lock(&ctx->uring_lock);
8984 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8985 struct io_uring_task *tctx = node->task->io_uring;
8988 * io_wq will stay alive while we hold uring_lock, because it's
8989 * killed after ctx nodes, which requires to take the lock.
8991 if (!tctx || !tctx->io_wq)
8993 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8994 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8996 mutex_unlock(&ctx->uring_lock);
9001 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9002 struct task_struct *task,
9005 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9006 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9009 enum io_wq_cancel cret;
9013 ret |= io_uring_try_cancel_iowq(ctx);
9014 } else if (tctx && tctx->io_wq) {
9016 * Cancels requests of all rings, not only @ctx, but
9017 * it's fine as the task is in exit/exec.
9019 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9021 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9024 /* SQPOLL thread does its own polling */
9025 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9026 (ctx->sq_data && ctx->sq_data->thread == current)) {
9027 while (!list_empty_careful(&ctx->iopoll_list)) {
9028 io_iopoll_try_reap_events(ctx);
9033 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9034 ret |= io_poll_remove_all(ctx, task, cancel_all);
9035 ret |= io_kill_timeouts(ctx, task, cancel_all);
9037 ret |= io_run_task_work();
9044 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9046 struct io_uring_task *tctx = current->io_uring;
9047 struct io_tctx_node *node;
9050 if (unlikely(!tctx)) {
9051 ret = io_uring_alloc_task_context(current, ctx);
9054 tctx = current->io_uring;
9056 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9057 node = kmalloc(sizeof(*node), GFP_KERNEL);
9061 node->task = current;
9063 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9070 mutex_lock(&ctx->uring_lock);
9071 list_add(&node->ctx_node, &ctx->tctx_list);
9072 mutex_unlock(&ctx->uring_lock);
9079 * Note that this task has used io_uring. We use it for cancelation purposes.
9081 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9083 struct io_uring_task *tctx = current->io_uring;
9085 if (likely(tctx && tctx->last == ctx))
9087 return __io_uring_add_tctx_node(ctx);
9091 * Remove this io_uring_file -> task mapping.
9093 static void io_uring_del_tctx_node(unsigned long index)
9095 struct io_uring_task *tctx = current->io_uring;
9096 struct io_tctx_node *node;
9100 node = xa_erase(&tctx->xa, index);
9104 WARN_ON_ONCE(current != node->task);
9105 WARN_ON_ONCE(list_empty(&node->ctx_node));
9107 mutex_lock(&node->ctx->uring_lock);
9108 list_del(&node->ctx_node);
9109 mutex_unlock(&node->ctx->uring_lock);
9111 if (tctx->last == node->ctx)
9116 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9118 struct io_wq *wq = tctx->io_wq;
9119 struct io_tctx_node *node;
9120 unsigned long index;
9122 xa_for_each(&tctx->xa, index, node)
9123 io_uring_del_tctx_node(index);
9126 * Must be after io_uring_del_task_file() (removes nodes under
9127 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9130 io_wq_put_and_exit(wq);
9134 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9137 return atomic_read(&tctx->inflight_tracked);
9138 return percpu_counter_sum(&tctx->inflight);
9141 static void io_uring_drop_tctx_refs(struct task_struct *task)
9143 struct io_uring_task *tctx = task->io_uring;
9144 unsigned int refs = tctx->cached_refs;
9147 tctx->cached_refs = 0;
9148 percpu_counter_sub(&tctx->inflight, refs);
9149 put_task_struct_many(task, refs);
9154 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9155 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9157 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9159 struct io_uring_task *tctx = current->io_uring;
9160 struct io_ring_ctx *ctx;
9164 WARN_ON_ONCE(sqd && sqd->thread != current);
9166 if (!current->io_uring)
9169 io_wq_exit_start(tctx->io_wq);
9171 atomic_inc(&tctx->in_idle);
9173 io_uring_drop_tctx_refs(current);
9174 /* read completions before cancelations */
9175 inflight = tctx_inflight(tctx, !cancel_all);
9180 struct io_tctx_node *node;
9181 unsigned long index;
9183 xa_for_each(&tctx->xa, index, node) {
9184 /* sqpoll task will cancel all its requests */
9185 if (node->ctx->sq_data)
9187 io_uring_try_cancel_requests(node->ctx, current,
9191 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9192 io_uring_try_cancel_requests(ctx, current,
9196 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9197 io_uring_drop_tctx_refs(current);
9199 * If we've seen completions, retry without waiting. This
9200 * avoids a race where a completion comes in before we did
9201 * prepare_to_wait().
9203 if (inflight == tctx_inflight(tctx, !cancel_all))
9205 finish_wait(&tctx->wait, &wait);
9207 atomic_dec(&tctx->in_idle);
9209 io_uring_clean_tctx(tctx);
9211 /* for exec all current's requests should be gone, kill tctx */
9212 __io_uring_free(current);
9216 void __io_uring_cancel(struct files_struct *files)
9218 io_uring_cancel_generic(!files, NULL);
9221 static void *io_uring_validate_mmap_request(struct file *file,
9222 loff_t pgoff, size_t sz)
9224 struct io_ring_ctx *ctx = file->private_data;
9225 loff_t offset = pgoff << PAGE_SHIFT;
9230 case IORING_OFF_SQ_RING:
9231 case IORING_OFF_CQ_RING:
9234 case IORING_OFF_SQES:
9238 return ERR_PTR(-EINVAL);
9241 page = virt_to_head_page(ptr);
9242 if (sz > page_size(page))
9243 return ERR_PTR(-EINVAL);
9250 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9252 size_t sz = vma->vm_end - vma->vm_start;
9256 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9258 return PTR_ERR(ptr);
9260 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9261 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9264 #else /* !CONFIG_MMU */
9266 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9268 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9271 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9273 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9276 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9277 unsigned long addr, unsigned long len,
9278 unsigned long pgoff, unsigned long flags)
9282 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9284 return PTR_ERR(ptr);
9286 return (unsigned long) ptr;
9289 #endif /* !CONFIG_MMU */
9291 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9296 if (!io_sqring_full(ctx))
9298 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9300 if (!io_sqring_full(ctx))
9303 } while (!signal_pending(current));
9305 finish_wait(&ctx->sqo_sq_wait, &wait);
9309 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9310 struct __kernel_timespec __user **ts,
9311 const sigset_t __user **sig)
9313 struct io_uring_getevents_arg arg;
9316 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9317 * is just a pointer to the sigset_t.
9319 if (!(flags & IORING_ENTER_EXT_ARG)) {
9320 *sig = (const sigset_t __user *) argp;
9326 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9327 * timespec and sigset_t pointers if good.
9329 if (*argsz != sizeof(arg))
9331 if (copy_from_user(&arg, argp, sizeof(arg)))
9333 *sig = u64_to_user_ptr(arg.sigmask);
9334 *argsz = arg.sigmask_sz;
9335 *ts = u64_to_user_ptr(arg.ts);
9339 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9340 u32, min_complete, u32, flags, const void __user *, argp,
9343 struct io_ring_ctx *ctx;
9350 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9351 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9355 if (unlikely(!f.file))
9359 if (unlikely(f.file->f_op != &io_uring_fops))
9363 ctx = f.file->private_data;
9364 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9368 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9372 * For SQ polling, the thread will do all submissions and completions.
9373 * Just return the requested submit count, and wake the thread if
9377 if (ctx->flags & IORING_SETUP_SQPOLL) {
9378 io_cqring_overflow_flush(ctx, false);
9380 if (unlikely(ctx->sq_data->thread == NULL)) {
9384 if (flags & IORING_ENTER_SQ_WAKEUP)
9385 wake_up(&ctx->sq_data->wait);
9386 if (flags & IORING_ENTER_SQ_WAIT) {
9387 ret = io_sqpoll_wait_sq(ctx);
9391 submitted = to_submit;
9392 } else if (to_submit) {
9393 ret = io_uring_add_tctx_node(ctx);
9396 mutex_lock(&ctx->uring_lock);
9397 submitted = io_submit_sqes(ctx, to_submit);
9398 mutex_unlock(&ctx->uring_lock);
9400 if (submitted != to_submit)
9403 if (flags & IORING_ENTER_GETEVENTS) {
9404 const sigset_t __user *sig;
9405 struct __kernel_timespec __user *ts;
9407 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9411 min_complete = min(min_complete, ctx->cq_entries);
9414 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9415 * space applications don't need to do io completion events
9416 * polling again, they can rely on io_sq_thread to do polling
9417 * work, which can reduce cpu usage and uring_lock contention.
9419 if (ctx->flags & IORING_SETUP_IOPOLL &&
9420 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9421 ret = io_iopoll_check(ctx, min_complete);
9423 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9428 percpu_ref_put(&ctx->refs);
9431 return submitted ? submitted : ret;
9434 #ifdef CONFIG_PROC_FS
9435 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9436 const struct cred *cred)
9438 struct user_namespace *uns = seq_user_ns(m);
9439 struct group_info *gi;
9444 seq_printf(m, "%5d\n", id);
9445 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9446 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9447 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9448 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9449 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9450 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9451 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9452 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9453 seq_puts(m, "\n\tGroups:\t");
9454 gi = cred->group_info;
9455 for (g = 0; g < gi->ngroups; g++) {
9456 seq_put_decimal_ull(m, g ? " " : "",
9457 from_kgid_munged(uns, gi->gid[g]));
9459 seq_puts(m, "\n\tCapEff:\t");
9460 cap = cred->cap_effective;
9461 CAP_FOR_EACH_U32(__capi)
9462 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9467 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9469 struct io_sq_data *sq = NULL;
9474 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9475 * since fdinfo case grabs it in the opposite direction of normal use
9476 * cases. If we fail to get the lock, we just don't iterate any
9477 * structures that could be going away outside the io_uring mutex.
9479 has_lock = mutex_trylock(&ctx->uring_lock);
9481 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9487 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9488 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9489 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9490 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9491 struct file *f = io_file_from_index(ctx, i);
9494 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9496 seq_printf(m, "%5u: <none>\n", i);
9498 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9499 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9500 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9501 unsigned int len = buf->ubuf_end - buf->ubuf;
9503 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9505 if (has_lock && !xa_empty(&ctx->personalities)) {
9506 unsigned long index;
9507 const struct cred *cred;
9509 seq_printf(m, "Personalities:\n");
9510 xa_for_each(&ctx->personalities, index, cred)
9511 io_uring_show_cred(m, index, cred);
9513 seq_printf(m, "PollList:\n");
9514 spin_lock_irq(&ctx->completion_lock);
9515 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9516 struct hlist_head *list = &ctx->cancel_hash[i];
9517 struct io_kiocb *req;
9519 hlist_for_each_entry(req, list, hash_node)
9520 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9521 req->task->task_works != NULL);
9523 spin_unlock_irq(&ctx->completion_lock);
9525 mutex_unlock(&ctx->uring_lock);
9528 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9530 struct io_ring_ctx *ctx = f->private_data;
9532 if (percpu_ref_tryget(&ctx->refs)) {
9533 __io_uring_show_fdinfo(ctx, m);
9534 percpu_ref_put(&ctx->refs);
9539 static const struct file_operations io_uring_fops = {
9540 .release = io_uring_release,
9541 .mmap = io_uring_mmap,
9543 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9544 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9546 .poll = io_uring_poll,
9547 .fasync = io_uring_fasync,
9548 #ifdef CONFIG_PROC_FS
9549 .show_fdinfo = io_uring_show_fdinfo,
9553 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9554 struct io_uring_params *p)
9556 struct io_rings *rings;
9557 size_t size, sq_array_offset;
9559 /* make sure these are sane, as we already accounted them */
9560 ctx->sq_entries = p->sq_entries;
9561 ctx->cq_entries = p->cq_entries;
9563 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9564 if (size == SIZE_MAX)
9567 rings = io_mem_alloc(size);
9572 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9573 rings->sq_ring_mask = p->sq_entries - 1;
9574 rings->cq_ring_mask = p->cq_entries - 1;
9575 rings->sq_ring_entries = p->sq_entries;
9576 rings->cq_ring_entries = p->cq_entries;
9578 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9579 if (size == SIZE_MAX) {
9580 io_mem_free(ctx->rings);
9585 ctx->sq_sqes = io_mem_alloc(size);
9586 if (!ctx->sq_sqes) {
9587 io_mem_free(ctx->rings);
9595 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9599 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9603 ret = io_uring_add_tctx_node(ctx);
9608 fd_install(fd, file);
9613 * Allocate an anonymous fd, this is what constitutes the application
9614 * visible backing of an io_uring instance. The application mmaps this
9615 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9616 * we have to tie this fd to a socket for file garbage collection purposes.
9618 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9621 #if defined(CONFIG_UNIX)
9624 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9627 return ERR_PTR(ret);
9630 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9631 O_RDWR | O_CLOEXEC);
9632 #if defined(CONFIG_UNIX)
9634 sock_release(ctx->ring_sock);
9635 ctx->ring_sock = NULL;
9637 ctx->ring_sock->file = file;
9643 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9644 struct io_uring_params __user *params)
9646 struct io_ring_ctx *ctx;
9652 if (entries > IORING_MAX_ENTRIES) {
9653 if (!(p->flags & IORING_SETUP_CLAMP))
9655 entries = IORING_MAX_ENTRIES;
9659 * Use twice as many entries for the CQ ring. It's possible for the
9660 * application to drive a higher depth than the size of the SQ ring,
9661 * since the sqes are only used at submission time. This allows for
9662 * some flexibility in overcommitting a bit. If the application has
9663 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9664 * of CQ ring entries manually.
9666 p->sq_entries = roundup_pow_of_two(entries);
9667 if (p->flags & IORING_SETUP_CQSIZE) {
9669 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9670 * to a power-of-two, if it isn't already. We do NOT impose
9671 * any cq vs sq ring sizing.
9675 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9676 if (!(p->flags & IORING_SETUP_CLAMP))
9678 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9680 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9681 if (p->cq_entries < p->sq_entries)
9684 p->cq_entries = 2 * p->sq_entries;
9687 ctx = io_ring_ctx_alloc(p);
9690 ctx->compat = in_compat_syscall();
9691 if (!capable(CAP_IPC_LOCK))
9692 ctx->user = get_uid(current_user());
9695 * This is just grabbed for accounting purposes. When a process exits,
9696 * the mm is exited and dropped before the files, hence we need to hang
9697 * on to this mm purely for the purposes of being able to unaccount
9698 * memory (locked/pinned vm). It's not used for anything else.
9700 mmgrab(current->mm);
9701 ctx->mm_account = current->mm;
9703 ret = io_allocate_scq_urings(ctx, p);
9707 ret = io_sq_offload_create(ctx, p);
9710 /* always set a rsrc node */
9711 ret = io_rsrc_node_switch_start(ctx);
9714 io_rsrc_node_switch(ctx, NULL);
9716 memset(&p->sq_off, 0, sizeof(p->sq_off));
9717 p->sq_off.head = offsetof(struct io_rings, sq.head);
9718 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9719 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9720 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9721 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9722 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9723 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9725 memset(&p->cq_off, 0, sizeof(p->cq_off));
9726 p->cq_off.head = offsetof(struct io_rings, cq.head);
9727 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9728 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9729 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9730 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9731 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9732 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9734 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9735 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9736 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9737 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9738 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9739 IORING_FEAT_RSRC_TAGS;
9741 if (copy_to_user(params, p, sizeof(*p))) {
9746 file = io_uring_get_file(ctx);
9748 ret = PTR_ERR(file);
9753 * Install ring fd as the very last thing, so we don't risk someone
9754 * having closed it before we finish setup
9756 ret = io_uring_install_fd(ctx, file);
9758 /* fput will clean it up */
9763 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9766 io_ring_ctx_wait_and_kill(ctx);
9771 * Sets up an aio uring context, and returns the fd. Applications asks for a
9772 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9773 * params structure passed in.
9775 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9777 struct io_uring_params p;
9780 if (copy_from_user(&p, params, sizeof(p)))
9782 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9787 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9788 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9789 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9790 IORING_SETUP_R_DISABLED))
9793 return io_uring_create(entries, &p, params);
9796 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9797 struct io_uring_params __user *, params)
9799 return io_uring_setup(entries, params);
9802 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9804 struct io_uring_probe *p;
9808 size = struct_size(p, ops, nr_args);
9809 if (size == SIZE_MAX)
9811 p = kzalloc(size, GFP_KERNEL);
9816 if (copy_from_user(p, arg, size))
9819 if (memchr_inv(p, 0, size))
9822 p->last_op = IORING_OP_LAST - 1;
9823 if (nr_args > IORING_OP_LAST)
9824 nr_args = IORING_OP_LAST;
9826 for (i = 0; i < nr_args; i++) {
9828 if (!io_op_defs[i].not_supported)
9829 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9834 if (copy_to_user(arg, p, size))
9841 static int io_register_personality(struct io_ring_ctx *ctx)
9843 const struct cred *creds;
9847 creds = get_current_cred();
9849 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9850 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9858 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9859 unsigned int nr_args)
9861 struct io_uring_restriction *res;
9865 /* Restrictions allowed only if rings started disabled */
9866 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9869 /* We allow only a single restrictions registration */
9870 if (ctx->restrictions.registered)
9873 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9876 size = array_size(nr_args, sizeof(*res));
9877 if (size == SIZE_MAX)
9880 res = memdup_user(arg, size);
9882 return PTR_ERR(res);
9886 for (i = 0; i < nr_args; i++) {
9887 switch (res[i].opcode) {
9888 case IORING_RESTRICTION_REGISTER_OP:
9889 if (res[i].register_op >= IORING_REGISTER_LAST) {
9894 __set_bit(res[i].register_op,
9895 ctx->restrictions.register_op);
9897 case IORING_RESTRICTION_SQE_OP:
9898 if (res[i].sqe_op >= IORING_OP_LAST) {
9903 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9905 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9906 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9908 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9909 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9918 /* Reset all restrictions if an error happened */
9920 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9922 ctx->restrictions.registered = true;
9928 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9930 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9933 if (ctx->restrictions.registered)
9934 ctx->restricted = 1;
9936 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9937 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9938 wake_up(&ctx->sq_data->wait);
9942 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9943 struct io_uring_rsrc_update2 *up,
9951 if (check_add_overflow(up->offset, nr_args, &tmp))
9953 err = io_rsrc_node_switch_start(ctx);
9958 case IORING_RSRC_FILE:
9959 return __io_sqe_files_update(ctx, up, nr_args);
9960 case IORING_RSRC_BUFFER:
9961 return __io_sqe_buffers_update(ctx, up, nr_args);
9966 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9969 struct io_uring_rsrc_update2 up;
9973 memset(&up, 0, sizeof(up));
9974 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9976 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9979 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9980 unsigned size, unsigned type)
9982 struct io_uring_rsrc_update2 up;
9984 if (size != sizeof(up))
9986 if (copy_from_user(&up, arg, sizeof(up)))
9988 if (!up.nr || up.resv)
9990 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9993 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9994 unsigned int size, unsigned int type)
9996 struct io_uring_rsrc_register rr;
9998 /* keep it extendible */
9999 if (size != sizeof(rr))
10002 memset(&rr, 0, sizeof(rr));
10003 if (copy_from_user(&rr, arg, size))
10005 if (!rr.nr || rr.resv || rr.resv2)
10009 case IORING_RSRC_FILE:
10010 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10011 rr.nr, u64_to_user_ptr(rr.tags));
10012 case IORING_RSRC_BUFFER:
10013 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10014 rr.nr, u64_to_user_ptr(rr.tags));
10019 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10022 struct io_uring_task *tctx = current->io_uring;
10023 cpumask_var_t new_mask;
10026 if (!tctx || !tctx->io_wq)
10029 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10032 cpumask_clear(new_mask);
10033 if (len > cpumask_size())
10034 len = cpumask_size();
10036 if (copy_from_user(new_mask, arg, len)) {
10037 free_cpumask_var(new_mask);
10041 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10042 free_cpumask_var(new_mask);
10046 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10048 struct io_uring_task *tctx = current->io_uring;
10050 if (!tctx || !tctx->io_wq)
10053 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10056 static bool io_register_op_must_quiesce(int op)
10059 case IORING_REGISTER_BUFFERS:
10060 case IORING_UNREGISTER_BUFFERS:
10061 case IORING_REGISTER_FILES:
10062 case IORING_UNREGISTER_FILES:
10063 case IORING_REGISTER_FILES_UPDATE:
10064 case IORING_REGISTER_PROBE:
10065 case IORING_REGISTER_PERSONALITY:
10066 case IORING_UNREGISTER_PERSONALITY:
10067 case IORING_REGISTER_FILES2:
10068 case IORING_REGISTER_FILES_UPDATE2:
10069 case IORING_REGISTER_BUFFERS2:
10070 case IORING_REGISTER_BUFFERS_UPDATE:
10071 case IORING_REGISTER_IOWQ_AFF:
10072 case IORING_UNREGISTER_IOWQ_AFF:
10079 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10083 percpu_ref_kill(&ctx->refs);
10086 * Drop uring mutex before waiting for references to exit. If another
10087 * thread is currently inside io_uring_enter() it might need to grab the
10088 * uring_lock to make progress. If we hold it here across the drain
10089 * wait, then we can deadlock. It's safe to drop the mutex here, since
10090 * no new references will come in after we've killed the percpu ref.
10092 mutex_unlock(&ctx->uring_lock);
10094 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10097 ret = io_run_task_work_sig();
10098 } while (ret >= 0);
10099 mutex_lock(&ctx->uring_lock);
10102 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10106 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10107 void __user *arg, unsigned nr_args)
10108 __releases(ctx->uring_lock)
10109 __acquires(ctx->uring_lock)
10114 * We're inside the ring mutex, if the ref is already dying, then
10115 * someone else killed the ctx or is already going through
10116 * io_uring_register().
10118 if (percpu_ref_is_dying(&ctx->refs))
10121 if (ctx->restricted) {
10122 if (opcode >= IORING_REGISTER_LAST)
10124 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10125 if (!test_bit(opcode, ctx->restrictions.register_op))
10129 if (io_register_op_must_quiesce(opcode)) {
10130 ret = io_ctx_quiesce(ctx);
10136 case IORING_REGISTER_BUFFERS:
10137 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10139 case IORING_UNREGISTER_BUFFERS:
10141 if (arg || nr_args)
10143 ret = io_sqe_buffers_unregister(ctx);
10145 case IORING_REGISTER_FILES:
10146 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10148 case IORING_UNREGISTER_FILES:
10150 if (arg || nr_args)
10152 ret = io_sqe_files_unregister(ctx);
10154 case IORING_REGISTER_FILES_UPDATE:
10155 ret = io_register_files_update(ctx, arg, nr_args);
10157 case IORING_REGISTER_EVENTFD:
10158 case IORING_REGISTER_EVENTFD_ASYNC:
10162 ret = io_eventfd_register(ctx, arg);
10165 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10166 ctx->eventfd_async = 1;
10168 ctx->eventfd_async = 0;
10170 case IORING_UNREGISTER_EVENTFD:
10172 if (arg || nr_args)
10174 ret = io_eventfd_unregister(ctx);
10176 case IORING_REGISTER_PROBE:
10178 if (!arg || nr_args > 256)
10180 ret = io_probe(ctx, arg, nr_args);
10182 case IORING_REGISTER_PERSONALITY:
10184 if (arg || nr_args)
10186 ret = io_register_personality(ctx);
10188 case IORING_UNREGISTER_PERSONALITY:
10192 ret = io_unregister_personality(ctx, nr_args);
10194 case IORING_REGISTER_ENABLE_RINGS:
10196 if (arg || nr_args)
10198 ret = io_register_enable_rings(ctx);
10200 case IORING_REGISTER_RESTRICTIONS:
10201 ret = io_register_restrictions(ctx, arg, nr_args);
10203 case IORING_REGISTER_FILES2:
10204 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10206 case IORING_REGISTER_FILES_UPDATE2:
10207 ret = io_register_rsrc_update(ctx, arg, nr_args,
10210 case IORING_REGISTER_BUFFERS2:
10211 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10213 case IORING_REGISTER_BUFFERS_UPDATE:
10214 ret = io_register_rsrc_update(ctx, arg, nr_args,
10215 IORING_RSRC_BUFFER);
10217 case IORING_REGISTER_IOWQ_AFF:
10219 if (!arg || !nr_args)
10221 ret = io_register_iowq_aff(ctx, arg, nr_args);
10223 case IORING_UNREGISTER_IOWQ_AFF:
10225 if (arg || nr_args)
10227 ret = io_unregister_iowq_aff(ctx);
10234 if (io_register_op_must_quiesce(opcode)) {
10235 /* bring the ctx back to life */
10236 percpu_ref_reinit(&ctx->refs);
10237 reinit_completion(&ctx->ref_comp);
10242 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10243 void __user *, arg, unsigned int, nr_args)
10245 struct io_ring_ctx *ctx;
10254 if (f.file->f_op != &io_uring_fops)
10257 ctx = f.file->private_data;
10259 io_run_task_work();
10261 mutex_lock(&ctx->uring_lock);
10262 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10263 mutex_unlock(&ctx->uring_lock);
10264 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10265 ctx->cq_ev_fd != NULL, ret);
10271 static int __init io_uring_init(void)
10273 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10274 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10275 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10278 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10279 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10280 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10281 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10282 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10283 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10284 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10285 BUILD_BUG_SQE_ELEM(8, __u64, off);
10286 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10287 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10288 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10289 BUILD_BUG_SQE_ELEM(24, __u32, len);
10290 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10291 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10292 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10293 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10294 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10295 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10296 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10297 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10298 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10299 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10300 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10301 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10302 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10303 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10304 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10305 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10306 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10307 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10308 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10309 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10311 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10312 sizeof(struct io_uring_rsrc_update));
10313 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10314 sizeof(struct io_uring_rsrc_update2));
10315 /* should fit into one byte */
10316 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10318 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10319 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10321 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10325 __initcall(io_uring_init);