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;
1835 io_dismantle_req(req);
1836 io_put_task(req->task, 1);
1838 kmem_cache_free(req_cachep, req);
1839 percpu_ref_put(&ctx->refs);
1842 static inline void io_remove_next_linked(struct io_kiocb *req)
1844 struct io_kiocb *nxt = req->link;
1846 req->link = nxt->link;
1850 static bool io_kill_linked_timeout(struct io_kiocb *req)
1851 __must_hold(&req->ctx->completion_lock)
1853 struct io_kiocb *link = req->link;
1856 * Can happen if a linked timeout fired and link had been like
1857 * req -> link t-out -> link t-out [-> ...]
1859 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1860 struct io_timeout_data *io = link->async_data;
1862 io_remove_next_linked(req);
1863 link->timeout.head = NULL;
1864 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1865 io_cqring_fill_event(link->ctx, link->user_data,
1867 io_put_req_deferred(link, 1);
1874 static void io_fail_links(struct io_kiocb *req)
1875 __must_hold(&req->ctx->completion_lock)
1877 struct io_kiocb *nxt, *link = req->link;
1884 trace_io_uring_fail_link(req, link);
1885 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1886 io_put_req_deferred(link, 2);
1891 static bool io_disarm_next(struct io_kiocb *req)
1892 __must_hold(&req->ctx->completion_lock)
1894 bool posted = false;
1896 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1897 posted = io_kill_linked_timeout(req);
1898 if (unlikely((req->flags & REQ_F_FAIL) &&
1899 !(req->flags & REQ_F_HARDLINK))) {
1900 posted |= (req->link != NULL);
1906 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1908 struct io_kiocb *nxt;
1911 * If LINK is set, we have dependent requests in this chain. If we
1912 * didn't fail this request, queue the first one up, moving any other
1913 * dependencies to the next request. In case of failure, fail the rest
1916 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1917 struct io_ring_ctx *ctx = req->ctx;
1918 unsigned long flags;
1921 spin_lock_irqsave(&ctx->completion_lock, flags);
1922 posted = io_disarm_next(req);
1924 io_commit_cqring(req->ctx);
1925 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1927 io_cqring_ev_posted(ctx);
1934 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1936 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1938 return __io_req_find_next(req);
1941 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1945 if (ctx->submit_state.comp.nr) {
1946 mutex_lock(&ctx->uring_lock);
1947 io_submit_flush_completions(ctx);
1948 mutex_unlock(&ctx->uring_lock);
1950 percpu_ref_put(&ctx->refs);
1953 static void tctx_task_work(struct callback_head *cb)
1955 struct io_ring_ctx *ctx = NULL;
1956 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1960 struct io_wq_work_node *node;
1962 spin_lock_irq(&tctx->task_lock);
1963 node = tctx->task_list.first;
1964 INIT_WQ_LIST(&tctx->task_list);
1965 spin_unlock_irq(&tctx->task_lock);
1968 struct io_wq_work_node *next = node->next;
1969 struct io_kiocb *req = container_of(node, struct io_kiocb,
1972 if (req->ctx != ctx) {
1973 ctx_flush_and_put(ctx);
1975 percpu_ref_get(&ctx->refs);
1977 req->io_task_work.func(req);
1980 if (wq_list_empty(&tctx->task_list)) {
1981 spin_lock_irq(&tctx->task_lock);
1982 clear_bit(0, &tctx->task_state);
1983 if (wq_list_empty(&tctx->task_list)) {
1984 spin_unlock_irq(&tctx->task_lock);
1987 spin_unlock_irq(&tctx->task_lock);
1988 /* another tctx_task_work() is enqueued, yield */
1989 if (test_and_set_bit(0, &tctx->task_state))
1995 ctx_flush_and_put(ctx);
1998 static void io_req_task_work_add(struct io_kiocb *req)
2000 struct task_struct *tsk = req->task;
2001 struct io_uring_task *tctx = tsk->io_uring;
2002 enum task_work_notify_mode notify;
2003 struct io_wq_work_node *node;
2004 unsigned long flags;
2006 WARN_ON_ONCE(!tctx);
2008 spin_lock_irqsave(&tctx->task_lock, flags);
2009 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2010 spin_unlock_irqrestore(&tctx->task_lock, flags);
2012 /* task_work already pending, we're done */
2013 if (test_bit(0, &tctx->task_state) ||
2014 test_and_set_bit(0, &tctx->task_state))
2018 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2019 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2020 * processing task_work. There's no reliable way to tell if TWA_RESUME
2023 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2024 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2025 wake_up_process(tsk);
2029 clear_bit(0, &tctx->task_state);
2030 spin_lock_irqsave(&tctx->task_lock, flags);
2031 node = tctx->task_list.first;
2032 INIT_WQ_LIST(&tctx->task_list);
2033 spin_unlock_irqrestore(&tctx->task_lock, flags);
2036 req = container_of(node, struct io_kiocb, io_task_work.node);
2038 if (llist_add(&req->io_task_work.fallback_node,
2039 &req->ctx->fallback_llist))
2040 schedule_delayed_work(&req->ctx->fallback_work, 1);
2044 static void io_req_task_cancel(struct io_kiocb *req)
2046 struct io_ring_ctx *ctx = req->ctx;
2048 /* ctx is guaranteed to stay alive while we hold uring_lock */
2049 mutex_lock(&ctx->uring_lock);
2050 io_req_complete_failed(req, req->result);
2051 mutex_unlock(&ctx->uring_lock);
2054 static void io_req_task_submit(struct io_kiocb *req)
2056 struct io_ring_ctx *ctx = req->ctx;
2058 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2059 mutex_lock(&ctx->uring_lock);
2060 if (likely(!(req->task->flags & PF_EXITING)))
2061 __io_queue_sqe(req);
2063 io_req_complete_failed(req, -EFAULT);
2064 mutex_unlock(&ctx->uring_lock);
2067 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2070 req->io_task_work.func = io_req_task_cancel;
2071 io_req_task_work_add(req);
2074 static void io_req_task_queue(struct io_kiocb *req)
2076 req->io_task_work.func = io_req_task_submit;
2077 io_req_task_work_add(req);
2080 static void io_req_task_queue_reissue(struct io_kiocb *req)
2082 req->io_task_work.func = io_queue_async_work;
2083 io_req_task_work_add(req);
2086 static inline void io_queue_next(struct io_kiocb *req)
2088 struct io_kiocb *nxt = io_req_find_next(req);
2091 io_req_task_queue(nxt);
2094 static void io_free_req(struct io_kiocb *req)
2101 struct task_struct *task;
2106 static inline void io_init_req_batch(struct req_batch *rb)
2113 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2114 struct req_batch *rb)
2117 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2118 if (rb->task == current)
2119 current->io_uring->cached_refs += rb->task_refs;
2121 io_put_task(rb->task, rb->task_refs);
2124 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2125 struct io_submit_state *state)
2128 io_dismantle_req(req);
2130 if (req->task != rb->task) {
2132 io_put_task(rb->task, rb->task_refs);
2133 rb->task = req->task;
2139 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2140 state->reqs[state->free_reqs++] = req;
2142 list_add(&req->compl.list, &state->comp.free_list);
2145 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2146 __must_hold(&req->ctx->uring_lock)
2148 struct io_comp_state *cs = &ctx->submit_state.comp;
2150 struct req_batch rb;
2152 spin_lock_irq(&ctx->completion_lock);
2153 for (i = 0; i < nr; i++) {
2154 struct io_kiocb *req = cs->reqs[i];
2156 __io_cqring_fill_event(ctx, req->user_data, req->result,
2159 io_commit_cqring(ctx);
2160 spin_unlock_irq(&ctx->completion_lock);
2161 io_cqring_ev_posted(ctx);
2163 io_init_req_batch(&rb);
2164 for (i = 0; i < nr; i++) {
2165 struct io_kiocb *req = cs->reqs[i];
2167 /* submission and completion refs */
2168 if (req_ref_sub_and_test(req, 2))
2169 io_req_free_batch(&rb, req, &ctx->submit_state);
2172 io_req_free_batch_finish(ctx, &rb);
2177 * Drop reference to request, return next in chain (if there is one) if this
2178 * was the last reference to this request.
2180 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2182 struct io_kiocb *nxt = NULL;
2184 if (req_ref_put_and_test(req)) {
2185 nxt = io_req_find_next(req);
2191 static inline void io_put_req(struct io_kiocb *req)
2193 if (req_ref_put_and_test(req))
2197 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2199 if (req_ref_sub_and_test(req, refs)) {
2200 req->io_task_work.func = io_free_req;
2201 io_req_task_work_add(req);
2205 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2207 /* See comment at the top of this file */
2209 return __io_cqring_events(ctx);
2212 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2214 struct io_rings *rings = ctx->rings;
2216 /* make sure SQ entry isn't read before tail */
2217 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2220 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2222 unsigned int cflags;
2224 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2225 cflags |= IORING_CQE_F_BUFFER;
2226 req->flags &= ~REQ_F_BUFFER_SELECTED;
2231 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2233 struct io_buffer *kbuf;
2235 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2236 return io_put_kbuf(req, kbuf);
2239 static inline bool io_run_task_work(void)
2241 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2242 __set_current_state(TASK_RUNNING);
2243 tracehook_notify_signal();
2251 * Find and free completed poll iocbs
2253 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2254 struct list_head *done, bool resubmit)
2256 struct req_batch rb;
2257 struct io_kiocb *req;
2259 /* order with ->result store in io_complete_rw_iopoll() */
2262 io_init_req_batch(&rb);
2263 while (!list_empty(done)) {
2266 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2267 list_del(&req->inflight_entry);
2269 if (READ_ONCE(req->result) == -EAGAIN && resubmit &&
2270 !(req->flags & REQ_F_DONT_REISSUE)) {
2271 req->iopoll_completed = 0;
2273 io_req_task_queue_reissue(req);
2277 if (req->flags & REQ_F_BUFFER_SELECTED)
2278 cflags = io_put_rw_kbuf(req);
2280 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2283 if (req_ref_put_and_test(req))
2284 io_req_free_batch(&rb, req, &ctx->submit_state);
2287 io_commit_cqring(ctx);
2288 io_cqring_ev_posted_iopoll(ctx);
2289 io_req_free_batch_finish(ctx, &rb);
2292 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2293 long min, bool resubmit)
2295 struct io_kiocb *req, *tmp;
2300 * Only spin for completions if we don't have multiple devices hanging
2301 * off our complete list, and we're under the requested amount.
2303 spin = !ctx->poll_multi_queue && *nr_events < min;
2305 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2306 struct kiocb *kiocb = &req->rw.kiocb;
2310 * Move completed and retryable entries to our local lists.
2311 * If we find a request that requires polling, break out
2312 * and complete those lists first, if we have entries there.
2314 if (READ_ONCE(req->iopoll_completed)) {
2315 list_move_tail(&req->inflight_entry, &done);
2318 if (!list_empty(&done))
2321 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2322 if (unlikely(ret < 0))
2327 /* iopoll may have completed current req */
2328 if (READ_ONCE(req->iopoll_completed))
2329 list_move_tail(&req->inflight_entry, &done);
2332 if (!list_empty(&done))
2333 io_iopoll_complete(ctx, nr_events, &done, resubmit);
2339 * We can't just wait for polled events to come to us, we have to actively
2340 * find and complete them.
2342 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2344 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2347 mutex_lock(&ctx->uring_lock);
2348 while (!list_empty(&ctx->iopoll_list)) {
2349 unsigned int nr_events = 0;
2351 io_do_iopoll(ctx, &nr_events, 0, false);
2353 /* let it sleep and repeat later if can't complete a request */
2357 * Ensure we allow local-to-the-cpu processing to take place,
2358 * in this case we need to ensure that we reap all events.
2359 * Also let task_work, etc. to progress by releasing the mutex
2361 if (need_resched()) {
2362 mutex_unlock(&ctx->uring_lock);
2364 mutex_lock(&ctx->uring_lock);
2367 mutex_unlock(&ctx->uring_lock);
2370 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2372 unsigned int nr_events = 0;
2376 * We disallow the app entering submit/complete with polling, but we
2377 * still need to lock the ring to prevent racing with polled issue
2378 * that got punted to a workqueue.
2380 mutex_lock(&ctx->uring_lock);
2382 * Don't enter poll loop if we already have events pending.
2383 * If we do, we can potentially be spinning for commands that
2384 * already triggered a CQE (eg in error).
2386 if (test_bit(0, &ctx->check_cq_overflow))
2387 __io_cqring_overflow_flush(ctx, false);
2388 if (io_cqring_events(ctx))
2392 * If a submit got punted to a workqueue, we can have the
2393 * application entering polling for a command before it gets
2394 * issued. That app will hold the uring_lock for the duration
2395 * of the poll right here, so we need to take a breather every
2396 * now and then to ensure that the issue has a chance to add
2397 * the poll to the issued list. Otherwise we can spin here
2398 * forever, while the workqueue is stuck trying to acquire the
2401 if (list_empty(&ctx->iopoll_list)) {
2402 u32 tail = ctx->cached_cq_tail;
2404 mutex_unlock(&ctx->uring_lock);
2406 mutex_lock(&ctx->uring_lock);
2408 /* some requests don't go through iopoll_list */
2409 if (tail != ctx->cached_cq_tail ||
2410 list_empty(&ctx->iopoll_list))
2413 ret = io_do_iopoll(ctx, &nr_events, min, true);
2414 } while (!ret && nr_events < min && !need_resched());
2416 mutex_unlock(&ctx->uring_lock);
2420 static void kiocb_end_write(struct io_kiocb *req)
2423 * Tell lockdep we inherited freeze protection from submission
2426 if (req->flags & REQ_F_ISREG) {
2427 struct super_block *sb = file_inode(req->file)->i_sb;
2429 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2435 static bool io_resubmit_prep(struct io_kiocb *req)
2437 struct io_async_rw *rw = req->async_data;
2440 return !io_req_prep_async(req);
2441 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2442 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2446 static bool io_rw_should_reissue(struct io_kiocb *req)
2448 umode_t mode = file_inode(req->file)->i_mode;
2449 struct io_ring_ctx *ctx = req->ctx;
2451 if (!S_ISBLK(mode) && !S_ISREG(mode))
2453 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2454 !(ctx->flags & IORING_SETUP_IOPOLL)))
2457 * If ref is dying, we might be running poll reap from the exit work.
2458 * Don't attempt to reissue from that path, just let it fail with
2461 if (percpu_ref_is_dying(&ctx->refs))
2464 * Play it safe and assume not safe to re-import and reissue if we're
2465 * not in the original thread group (or in task context).
2467 if (!same_thread_group(req->task, current) || !in_task())
2472 static bool io_resubmit_prep(struct io_kiocb *req)
2476 static bool io_rw_should_reissue(struct io_kiocb *req)
2482 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2483 unsigned int issue_flags)
2487 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2488 kiocb_end_write(req);
2489 if (res != req->result) {
2490 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2491 io_rw_should_reissue(req)) {
2492 req->flags |= REQ_F_REISSUE;
2497 if (req->flags & REQ_F_BUFFER_SELECTED)
2498 cflags = io_put_rw_kbuf(req);
2499 __io_req_complete(req, issue_flags, res, cflags);
2502 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2504 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2506 __io_complete_rw(req, res, res2, 0);
2509 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2511 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2513 if (kiocb->ki_flags & IOCB_WRITE)
2514 kiocb_end_write(req);
2515 if (unlikely(res != req->result)) {
2516 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2517 io_resubmit_prep(req))) {
2519 req->flags |= REQ_F_DONT_REISSUE;
2523 WRITE_ONCE(req->result, res);
2524 /* order with io_iopoll_complete() checking ->result */
2526 WRITE_ONCE(req->iopoll_completed, 1);
2530 * After the iocb has been issued, it's safe to be found on the poll list.
2531 * Adding the kiocb to the list AFTER submission ensures that we don't
2532 * find it from a io_do_iopoll() thread before the issuer is done
2533 * accessing the kiocb cookie.
2535 static void io_iopoll_req_issued(struct io_kiocb *req)
2537 struct io_ring_ctx *ctx = req->ctx;
2538 const bool in_async = io_wq_current_is_worker();
2540 /* workqueue context doesn't hold uring_lock, grab it now */
2541 if (unlikely(in_async))
2542 mutex_lock(&ctx->uring_lock);
2545 * Track whether we have multiple files in our lists. This will impact
2546 * how we do polling eventually, not spinning if we're on potentially
2547 * different devices.
2549 if (list_empty(&ctx->iopoll_list)) {
2550 ctx->poll_multi_queue = false;
2551 } else if (!ctx->poll_multi_queue) {
2552 struct io_kiocb *list_req;
2553 unsigned int queue_num0, queue_num1;
2555 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2558 if (list_req->file != req->file) {
2559 ctx->poll_multi_queue = true;
2561 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2562 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2563 if (queue_num0 != queue_num1)
2564 ctx->poll_multi_queue = true;
2569 * For fast devices, IO may have already completed. If it has, add
2570 * it to the front so we find it first.
2572 if (READ_ONCE(req->iopoll_completed))
2573 list_add(&req->inflight_entry, &ctx->iopoll_list);
2575 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2577 if (unlikely(in_async)) {
2579 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2580 * in sq thread task context or in io worker task context. If
2581 * current task context is sq thread, we don't need to check
2582 * whether should wake up sq thread.
2584 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2585 wq_has_sleeper(&ctx->sq_data->wait))
2586 wake_up(&ctx->sq_data->wait);
2588 mutex_unlock(&ctx->uring_lock);
2592 static inline void io_state_file_put(struct io_submit_state *state)
2594 if (state->file_refs) {
2595 fput_many(state->file, state->file_refs);
2596 state->file_refs = 0;
2601 * Get as many references to a file as we have IOs left in this submission,
2602 * assuming most submissions are for one file, or at least that each file
2603 * has more than one submission.
2605 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2610 if (state->file_refs) {
2611 if (state->fd == fd) {
2615 io_state_file_put(state);
2617 state->file = fget_many(fd, state->ios_left);
2618 if (unlikely(!state->file))
2622 state->file_refs = state->ios_left - 1;
2626 static bool io_bdev_nowait(struct block_device *bdev)
2628 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2632 * If we tracked the file through the SCM inflight mechanism, we could support
2633 * any file. For now, just ensure that anything potentially problematic is done
2636 static bool __io_file_supports_nowait(struct file *file, int rw)
2638 umode_t mode = file_inode(file)->i_mode;
2640 if (S_ISBLK(mode)) {
2641 if (IS_ENABLED(CONFIG_BLOCK) &&
2642 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2648 if (S_ISREG(mode)) {
2649 if (IS_ENABLED(CONFIG_BLOCK) &&
2650 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2651 file->f_op != &io_uring_fops)
2656 /* any ->read/write should understand O_NONBLOCK */
2657 if (file->f_flags & O_NONBLOCK)
2660 if (!(file->f_mode & FMODE_NOWAIT))
2664 return file->f_op->read_iter != NULL;
2666 return file->f_op->write_iter != NULL;
2669 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2671 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2673 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2676 return __io_file_supports_nowait(req->file, rw);
2679 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2681 struct io_ring_ctx *ctx = req->ctx;
2682 struct kiocb *kiocb = &req->rw.kiocb;
2683 struct file *file = req->file;
2687 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2688 req->flags |= REQ_F_ISREG;
2690 kiocb->ki_pos = READ_ONCE(sqe->off);
2691 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2692 req->flags |= REQ_F_CUR_POS;
2693 kiocb->ki_pos = file->f_pos;
2695 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2696 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2697 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2701 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2702 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2703 req->flags |= REQ_F_NOWAIT;
2705 ioprio = READ_ONCE(sqe->ioprio);
2707 ret = ioprio_check_cap(ioprio);
2711 kiocb->ki_ioprio = ioprio;
2713 kiocb->ki_ioprio = get_current_ioprio();
2715 if (ctx->flags & IORING_SETUP_IOPOLL) {
2716 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2717 !kiocb->ki_filp->f_op->iopoll)
2720 kiocb->ki_flags |= IOCB_HIPRI;
2721 kiocb->ki_complete = io_complete_rw_iopoll;
2722 req->iopoll_completed = 0;
2724 if (kiocb->ki_flags & IOCB_HIPRI)
2726 kiocb->ki_complete = io_complete_rw;
2729 if (req->opcode == IORING_OP_READ_FIXED ||
2730 req->opcode == IORING_OP_WRITE_FIXED) {
2732 io_req_set_rsrc_node(req);
2735 req->rw.addr = READ_ONCE(sqe->addr);
2736 req->rw.len = READ_ONCE(sqe->len);
2737 req->buf_index = READ_ONCE(sqe->buf_index);
2741 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2747 case -ERESTARTNOINTR:
2748 case -ERESTARTNOHAND:
2749 case -ERESTART_RESTARTBLOCK:
2751 * We can't just restart the syscall, since previously
2752 * submitted sqes may already be in progress. Just fail this
2758 kiocb->ki_complete(kiocb, ret, 0);
2762 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2763 unsigned int issue_flags)
2765 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2766 struct io_async_rw *io = req->async_data;
2767 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2769 /* add previously done IO, if any */
2770 if (io && io->bytes_done > 0) {
2772 ret = io->bytes_done;
2774 ret += io->bytes_done;
2777 if (req->flags & REQ_F_CUR_POS)
2778 req->file->f_pos = kiocb->ki_pos;
2779 if (ret >= 0 && check_reissue)
2780 __io_complete_rw(req, ret, 0, issue_flags);
2782 io_rw_done(kiocb, ret);
2784 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2785 req->flags &= ~REQ_F_REISSUE;
2786 if (io_resubmit_prep(req)) {
2788 io_req_task_queue_reissue(req);
2793 if (req->flags & REQ_F_BUFFER_SELECTED)
2794 cflags = io_put_rw_kbuf(req);
2795 __io_req_complete(req, issue_flags, ret, cflags);
2800 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2801 struct io_mapped_ubuf *imu)
2803 size_t len = req->rw.len;
2804 u64 buf_end, buf_addr = req->rw.addr;
2807 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2809 /* not inside the mapped region */
2810 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2814 * May not be a start of buffer, set size appropriately
2815 * and advance us to the beginning.
2817 offset = buf_addr - imu->ubuf;
2818 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2822 * Don't use iov_iter_advance() here, as it's really slow for
2823 * using the latter parts of a big fixed buffer - it iterates
2824 * over each segment manually. We can cheat a bit here, because
2827 * 1) it's a BVEC iter, we set it up
2828 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2829 * first and last bvec
2831 * So just find our index, and adjust the iterator afterwards.
2832 * If the offset is within the first bvec (or the whole first
2833 * bvec, just use iov_iter_advance(). This makes it easier
2834 * since we can just skip the first segment, which may not
2835 * be PAGE_SIZE aligned.
2837 const struct bio_vec *bvec = imu->bvec;
2839 if (offset <= bvec->bv_len) {
2840 iov_iter_advance(iter, offset);
2842 unsigned long seg_skip;
2844 /* skip first vec */
2845 offset -= bvec->bv_len;
2846 seg_skip = 1 + (offset >> PAGE_SHIFT);
2848 iter->bvec = bvec + seg_skip;
2849 iter->nr_segs -= seg_skip;
2850 iter->count -= bvec->bv_len + offset;
2851 iter->iov_offset = offset & ~PAGE_MASK;
2858 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2860 struct io_ring_ctx *ctx = req->ctx;
2861 struct io_mapped_ubuf *imu = req->imu;
2862 u16 index, buf_index = req->buf_index;
2865 if (unlikely(buf_index >= ctx->nr_user_bufs))
2867 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2868 imu = READ_ONCE(ctx->user_bufs[index]);
2871 return __io_import_fixed(req, rw, iter, imu);
2874 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2877 mutex_unlock(&ctx->uring_lock);
2880 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2883 * "Normal" inline submissions always hold the uring_lock, since we
2884 * grab it from the system call. Same is true for the SQPOLL offload.
2885 * The only exception is when we've detached the request and issue it
2886 * from an async worker thread, grab the lock for that case.
2889 mutex_lock(&ctx->uring_lock);
2892 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2893 int bgid, struct io_buffer *kbuf,
2896 struct io_buffer *head;
2898 if (req->flags & REQ_F_BUFFER_SELECTED)
2901 io_ring_submit_lock(req->ctx, needs_lock);
2903 lockdep_assert_held(&req->ctx->uring_lock);
2905 head = xa_load(&req->ctx->io_buffers, bgid);
2907 if (!list_empty(&head->list)) {
2908 kbuf = list_last_entry(&head->list, struct io_buffer,
2910 list_del(&kbuf->list);
2913 xa_erase(&req->ctx->io_buffers, bgid);
2915 if (*len > kbuf->len)
2918 kbuf = ERR_PTR(-ENOBUFS);
2921 io_ring_submit_unlock(req->ctx, needs_lock);
2926 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2929 struct io_buffer *kbuf;
2932 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2933 bgid = req->buf_index;
2934 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2937 req->rw.addr = (u64) (unsigned long) kbuf;
2938 req->flags |= REQ_F_BUFFER_SELECTED;
2939 return u64_to_user_ptr(kbuf->addr);
2942 #ifdef CONFIG_COMPAT
2943 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2946 struct compat_iovec __user *uiov;
2947 compat_ssize_t clen;
2951 uiov = u64_to_user_ptr(req->rw.addr);
2952 if (!access_ok(uiov, sizeof(*uiov)))
2954 if (__get_user(clen, &uiov->iov_len))
2960 buf = io_rw_buffer_select(req, &len, needs_lock);
2962 return PTR_ERR(buf);
2963 iov[0].iov_base = buf;
2964 iov[0].iov_len = (compat_size_t) len;
2969 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2972 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2976 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2979 len = iov[0].iov_len;
2982 buf = io_rw_buffer_select(req, &len, needs_lock);
2984 return PTR_ERR(buf);
2985 iov[0].iov_base = buf;
2986 iov[0].iov_len = len;
2990 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2993 if (req->flags & REQ_F_BUFFER_SELECTED) {
2994 struct io_buffer *kbuf;
2996 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2997 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2998 iov[0].iov_len = kbuf->len;
3001 if (req->rw.len != 1)
3004 #ifdef CONFIG_COMPAT
3005 if (req->ctx->compat)
3006 return io_compat_import(req, iov, needs_lock);
3009 return __io_iov_buffer_select(req, iov, needs_lock);
3012 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3013 struct iov_iter *iter, bool needs_lock)
3015 void __user *buf = u64_to_user_ptr(req->rw.addr);
3016 size_t sqe_len = req->rw.len;
3017 u8 opcode = req->opcode;
3020 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3022 return io_import_fixed(req, rw, iter);
3025 /* buffer index only valid with fixed read/write, or buffer select */
3026 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3029 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3030 if (req->flags & REQ_F_BUFFER_SELECT) {
3031 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3033 return PTR_ERR(buf);
3034 req->rw.len = sqe_len;
3037 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3042 if (req->flags & REQ_F_BUFFER_SELECT) {
3043 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3045 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3050 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3054 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3056 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3060 * For files that don't have ->read_iter() and ->write_iter(), handle them
3061 * by looping over ->read() or ->write() manually.
3063 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3065 struct kiocb *kiocb = &req->rw.kiocb;
3066 struct file *file = req->file;
3070 * Don't support polled IO through this interface, and we can't
3071 * support non-blocking either. For the latter, this just causes
3072 * the kiocb to be handled from an async context.
3074 if (kiocb->ki_flags & IOCB_HIPRI)
3076 if (kiocb->ki_flags & IOCB_NOWAIT)
3079 while (iov_iter_count(iter)) {
3083 if (!iov_iter_is_bvec(iter)) {
3084 iovec = iov_iter_iovec(iter);
3086 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3087 iovec.iov_len = req->rw.len;
3091 nr = file->f_op->read(file, iovec.iov_base,
3092 iovec.iov_len, io_kiocb_ppos(kiocb));
3094 nr = file->f_op->write(file, iovec.iov_base,
3095 iovec.iov_len, io_kiocb_ppos(kiocb));
3104 if (nr != iovec.iov_len)
3108 iov_iter_advance(iter, nr);
3114 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3115 const struct iovec *fast_iov, struct iov_iter *iter)
3117 struct io_async_rw *rw = req->async_data;
3119 memcpy(&rw->iter, iter, sizeof(*iter));
3120 rw->free_iovec = iovec;
3122 /* can only be fixed buffers, no need to do anything */
3123 if (iov_iter_is_bvec(iter))
3126 unsigned iov_off = 0;
3128 rw->iter.iov = rw->fast_iov;
3129 if (iter->iov != fast_iov) {
3130 iov_off = iter->iov - fast_iov;
3131 rw->iter.iov += iov_off;
3133 if (rw->fast_iov != fast_iov)
3134 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3135 sizeof(struct iovec) * iter->nr_segs);
3137 req->flags |= REQ_F_NEED_CLEANUP;
3141 static inline int io_alloc_async_data(struct io_kiocb *req)
3143 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3144 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3145 return req->async_data == NULL;
3148 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3149 const struct iovec *fast_iov,
3150 struct iov_iter *iter, bool force)
3152 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3154 if (!req->async_data) {
3155 if (io_alloc_async_data(req)) {
3160 io_req_map_rw(req, iovec, fast_iov, iter);
3165 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3167 struct io_async_rw *iorw = req->async_data;
3168 struct iovec *iov = iorw->fast_iov;
3171 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3172 if (unlikely(ret < 0))
3175 iorw->bytes_done = 0;
3176 iorw->free_iovec = iov;
3178 req->flags |= REQ_F_NEED_CLEANUP;
3182 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3184 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3186 return io_prep_rw(req, sqe);
3190 * This is our waitqueue callback handler, registered through lock_page_async()
3191 * when we initially tried to do the IO with the iocb armed our waitqueue.
3192 * This gets called when the page is unlocked, and we generally expect that to
3193 * happen when the page IO is completed and the page is now uptodate. This will
3194 * queue a task_work based retry of the operation, attempting to copy the data
3195 * again. If the latter fails because the page was NOT uptodate, then we will
3196 * do a thread based blocking retry of the operation. That's the unexpected
3199 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3200 int sync, void *arg)
3202 struct wait_page_queue *wpq;
3203 struct io_kiocb *req = wait->private;
3204 struct wait_page_key *key = arg;
3206 wpq = container_of(wait, struct wait_page_queue, wait);
3208 if (!wake_page_match(wpq, key))
3211 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3212 list_del_init(&wait->entry);
3214 /* submit ref gets dropped, acquire a new one */
3216 io_req_task_queue(req);
3221 * This controls whether a given IO request should be armed for async page
3222 * based retry. If we return false here, the request is handed to the async
3223 * worker threads for retry. If we're doing buffered reads on a regular file,
3224 * we prepare a private wait_page_queue entry and retry the operation. This
3225 * will either succeed because the page is now uptodate and unlocked, or it
3226 * will register a callback when the page is unlocked at IO completion. Through
3227 * that callback, io_uring uses task_work to setup a retry of the operation.
3228 * That retry will attempt the buffered read again. The retry will generally
3229 * succeed, or in rare cases where it fails, we then fall back to using the
3230 * async worker threads for a blocking retry.
3232 static bool io_rw_should_retry(struct io_kiocb *req)
3234 struct io_async_rw *rw = req->async_data;
3235 struct wait_page_queue *wait = &rw->wpq;
3236 struct kiocb *kiocb = &req->rw.kiocb;
3238 /* never retry for NOWAIT, we just complete with -EAGAIN */
3239 if (req->flags & REQ_F_NOWAIT)
3242 /* Only for buffered IO */
3243 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3247 * just use poll if we can, and don't attempt if the fs doesn't
3248 * support callback based unlocks
3250 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3253 wait->wait.func = io_async_buf_func;
3254 wait->wait.private = req;
3255 wait->wait.flags = 0;
3256 INIT_LIST_HEAD(&wait->wait.entry);
3257 kiocb->ki_flags |= IOCB_WAITQ;
3258 kiocb->ki_flags &= ~IOCB_NOWAIT;
3259 kiocb->ki_waitq = wait;
3263 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3265 if (req->file->f_op->read_iter)
3266 return call_read_iter(req->file, &req->rw.kiocb, iter);
3267 else if (req->file->f_op->read)
3268 return loop_rw_iter(READ, req, iter);
3273 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3275 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3276 struct kiocb *kiocb = &req->rw.kiocb;
3277 struct iov_iter __iter, *iter = &__iter;
3278 struct io_async_rw *rw = req->async_data;
3279 ssize_t io_size, ret, ret2;
3280 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3286 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3290 io_size = iov_iter_count(iter);
3291 req->result = io_size;
3293 /* Ensure we clear previously set non-block flag */
3294 if (!force_nonblock)
3295 kiocb->ki_flags &= ~IOCB_NOWAIT;
3297 kiocb->ki_flags |= IOCB_NOWAIT;
3299 /* If the file doesn't support async, just async punt */
3300 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3301 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3302 return ret ?: -EAGAIN;
3305 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3306 if (unlikely(ret)) {
3311 ret = io_iter_do_read(req, iter);
3313 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3314 req->flags &= ~REQ_F_REISSUE;
3315 /* IOPOLL retry should happen for io-wq threads */
3316 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3318 /* no retry on NONBLOCK nor RWF_NOWAIT */
3319 if (req->flags & REQ_F_NOWAIT)
3321 /* some cases will consume bytes even on error returns */
3322 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3324 } else if (ret == -EIOCBQUEUED) {
3326 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3327 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3328 /* read all, failed, already did sync or don't want to retry */
3332 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3337 rw = req->async_data;
3338 /* now use our persistent iterator, if we aren't already */
3343 rw->bytes_done += ret;
3344 /* if we can retry, do so with the callbacks armed */
3345 if (!io_rw_should_retry(req)) {
3346 kiocb->ki_flags &= ~IOCB_WAITQ;
3351 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3352 * we get -EIOCBQUEUED, then we'll get a notification when the
3353 * desired page gets unlocked. We can also get a partial read
3354 * here, and if we do, then just retry at the new offset.
3356 ret = io_iter_do_read(req, iter);
3357 if (ret == -EIOCBQUEUED)
3359 /* we got some bytes, but not all. retry. */
3360 kiocb->ki_flags &= ~IOCB_WAITQ;
3361 } while (ret > 0 && ret < io_size);
3363 kiocb_done(kiocb, ret, issue_flags);
3365 /* it's faster to check here then delegate to kfree */
3371 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3373 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3375 return io_prep_rw(req, sqe);
3378 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3380 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3381 struct kiocb *kiocb = &req->rw.kiocb;
3382 struct iov_iter __iter, *iter = &__iter;
3383 struct io_async_rw *rw = req->async_data;
3384 ssize_t ret, ret2, io_size;
3385 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3391 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3395 io_size = iov_iter_count(iter);
3396 req->result = io_size;
3398 /* Ensure we clear previously set non-block flag */
3399 if (!force_nonblock)
3400 kiocb->ki_flags &= ~IOCB_NOWAIT;
3402 kiocb->ki_flags |= IOCB_NOWAIT;
3404 /* If the file doesn't support async, just async punt */
3405 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3408 /* file path doesn't support NOWAIT for non-direct_IO */
3409 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3410 (req->flags & REQ_F_ISREG))
3413 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3418 * Open-code file_start_write here to grab freeze protection,
3419 * which will be released by another thread in
3420 * io_complete_rw(). Fool lockdep by telling it the lock got
3421 * released so that it doesn't complain about the held lock when
3422 * we return to userspace.
3424 if (req->flags & REQ_F_ISREG) {
3425 sb_start_write(file_inode(req->file)->i_sb);
3426 __sb_writers_release(file_inode(req->file)->i_sb,
3429 kiocb->ki_flags |= IOCB_WRITE;
3431 if (req->file->f_op->write_iter)
3432 ret2 = call_write_iter(req->file, kiocb, iter);
3433 else if (req->file->f_op->write)
3434 ret2 = loop_rw_iter(WRITE, req, iter);
3438 if (req->flags & REQ_F_REISSUE) {
3439 req->flags &= ~REQ_F_REISSUE;
3444 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3445 * retry them without IOCB_NOWAIT.
3447 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3449 /* no retry on NONBLOCK nor RWF_NOWAIT */
3450 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3452 if (!force_nonblock || ret2 != -EAGAIN) {
3453 /* IOPOLL retry should happen for io-wq threads */
3454 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3457 kiocb_done(kiocb, ret2, issue_flags);
3460 /* some cases will consume bytes even on error returns */
3461 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3462 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3463 return ret ?: -EAGAIN;
3466 /* it's reportedly faster than delegating the null check to kfree() */
3472 static int io_renameat_prep(struct io_kiocb *req,
3473 const struct io_uring_sqe *sqe)
3475 struct io_rename *ren = &req->rename;
3476 const char __user *oldf, *newf;
3478 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3480 if (sqe->ioprio || sqe->buf_index)
3482 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3485 ren->old_dfd = READ_ONCE(sqe->fd);
3486 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3487 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3488 ren->new_dfd = READ_ONCE(sqe->len);
3489 ren->flags = READ_ONCE(sqe->rename_flags);
3491 ren->oldpath = getname(oldf);
3492 if (IS_ERR(ren->oldpath))
3493 return PTR_ERR(ren->oldpath);
3495 ren->newpath = getname(newf);
3496 if (IS_ERR(ren->newpath)) {
3497 putname(ren->oldpath);
3498 return PTR_ERR(ren->newpath);
3501 req->flags |= REQ_F_NEED_CLEANUP;
3505 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3507 struct io_rename *ren = &req->rename;
3510 if (issue_flags & IO_URING_F_NONBLOCK)
3513 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3514 ren->newpath, ren->flags);
3516 req->flags &= ~REQ_F_NEED_CLEANUP;
3519 io_req_complete(req, ret);
3523 static int io_unlinkat_prep(struct io_kiocb *req,
3524 const struct io_uring_sqe *sqe)
3526 struct io_unlink *un = &req->unlink;
3527 const char __user *fname;
3529 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3531 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3533 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3536 un->dfd = READ_ONCE(sqe->fd);
3538 un->flags = READ_ONCE(sqe->unlink_flags);
3539 if (un->flags & ~AT_REMOVEDIR)
3542 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3543 un->filename = getname(fname);
3544 if (IS_ERR(un->filename))
3545 return PTR_ERR(un->filename);
3547 req->flags |= REQ_F_NEED_CLEANUP;
3551 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3553 struct io_unlink *un = &req->unlink;
3556 if (issue_flags & IO_URING_F_NONBLOCK)
3559 if (un->flags & AT_REMOVEDIR)
3560 ret = do_rmdir(un->dfd, un->filename);
3562 ret = do_unlinkat(un->dfd, un->filename);
3564 req->flags &= ~REQ_F_NEED_CLEANUP;
3567 io_req_complete(req, ret);
3571 static int io_shutdown_prep(struct io_kiocb *req,
3572 const struct io_uring_sqe *sqe)
3574 #if defined(CONFIG_NET)
3575 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3577 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3581 req->shutdown.how = READ_ONCE(sqe->len);
3588 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3590 #if defined(CONFIG_NET)
3591 struct socket *sock;
3594 if (issue_flags & IO_URING_F_NONBLOCK)
3597 sock = sock_from_file(req->file);
3598 if (unlikely(!sock))
3601 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3604 io_req_complete(req, ret);
3611 static int __io_splice_prep(struct io_kiocb *req,
3612 const struct io_uring_sqe *sqe)
3614 struct io_splice *sp = &req->splice;
3615 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3617 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3621 sp->len = READ_ONCE(sqe->len);
3622 sp->flags = READ_ONCE(sqe->splice_flags);
3624 if (unlikely(sp->flags & ~valid_flags))
3627 sp->file_in = io_file_get(req->ctx, NULL, req,
3628 READ_ONCE(sqe->splice_fd_in),
3629 (sp->flags & SPLICE_F_FD_IN_FIXED));
3632 req->flags |= REQ_F_NEED_CLEANUP;
3636 static int io_tee_prep(struct io_kiocb *req,
3637 const struct io_uring_sqe *sqe)
3639 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3641 return __io_splice_prep(req, sqe);
3644 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3646 struct io_splice *sp = &req->splice;
3647 struct file *in = sp->file_in;
3648 struct file *out = sp->file_out;
3649 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3652 if (issue_flags & IO_URING_F_NONBLOCK)
3655 ret = do_tee(in, out, sp->len, flags);
3657 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3659 req->flags &= ~REQ_F_NEED_CLEANUP;
3663 io_req_complete(req, ret);
3667 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3669 struct io_splice *sp = &req->splice;
3671 sp->off_in = READ_ONCE(sqe->splice_off_in);
3672 sp->off_out = READ_ONCE(sqe->off);
3673 return __io_splice_prep(req, sqe);
3676 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3678 struct io_splice *sp = &req->splice;
3679 struct file *in = sp->file_in;
3680 struct file *out = sp->file_out;
3681 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3682 loff_t *poff_in, *poff_out;
3685 if (issue_flags & IO_URING_F_NONBLOCK)
3688 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3689 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3692 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3694 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3696 req->flags &= ~REQ_F_NEED_CLEANUP;
3700 io_req_complete(req, ret);
3705 * IORING_OP_NOP just posts a completion event, nothing else.
3707 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3709 struct io_ring_ctx *ctx = req->ctx;
3711 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3714 __io_req_complete(req, issue_flags, 0, 0);
3718 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3720 struct io_ring_ctx *ctx = req->ctx;
3725 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3727 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3730 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3731 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3734 req->sync.off = READ_ONCE(sqe->off);
3735 req->sync.len = READ_ONCE(sqe->len);
3739 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3741 loff_t end = req->sync.off + req->sync.len;
3744 /* fsync always requires a blocking context */
3745 if (issue_flags & IO_URING_F_NONBLOCK)
3748 ret = vfs_fsync_range(req->file, req->sync.off,
3749 end > 0 ? end : LLONG_MAX,
3750 req->sync.flags & IORING_FSYNC_DATASYNC);
3753 io_req_complete(req, ret);
3757 static int io_fallocate_prep(struct io_kiocb *req,
3758 const struct io_uring_sqe *sqe)
3760 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3762 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3765 req->sync.off = READ_ONCE(sqe->off);
3766 req->sync.len = READ_ONCE(sqe->addr);
3767 req->sync.mode = READ_ONCE(sqe->len);
3771 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3775 /* fallocate always requiring blocking context */
3776 if (issue_flags & IO_URING_F_NONBLOCK)
3778 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3782 io_req_complete(req, ret);
3786 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3788 const char __user *fname;
3791 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3793 if (unlikely(sqe->ioprio || sqe->buf_index))
3795 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3798 /* open.how should be already initialised */
3799 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3800 req->open.how.flags |= O_LARGEFILE;
3802 req->open.dfd = READ_ONCE(sqe->fd);
3803 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3804 req->open.filename = getname(fname);
3805 if (IS_ERR(req->open.filename)) {
3806 ret = PTR_ERR(req->open.filename);
3807 req->open.filename = NULL;
3810 req->open.nofile = rlimit(RLIMIT_NOFILE);
3811 req->flags |= REQ_F_NEED_CLEANUP;
3815 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3817 u64 mode = READ_ONCE(sqe->len);
3818 u64 flags = READ_ONCE(sqe->open_flags);
3820 req->open.how = build_open_how(flags, mode);
3821 return __io_openat_prep(req, sqe);
3824 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3826 struct open_how __user *how;
3830 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3831 len = READ_ONCE(sqe->len);
3832 if (len < OPEN_HOW_SIZE_VER0)
3835 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3840 return __io_openat_prep(req, sqe);
3843 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3845 struct open_flags op;
3848 bool resolve_nonblock;
3851 ret = build_open_flags(&req->open.how, &op);
3854 nonblock_set = op.open_flag & O_NONBLOCK;
3855 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3856 if (issue_flags & IO_URING_F_NONBLOCK) {
3858 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3859 * it'll always -EAGAIN
3861 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3863 op.lookup_flags |= LOOKUP_CACHED;
3864 op.open_flag |= O_NONBLOCK;
3867 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3871 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3874 * We could hang on to this 'fd' on retrying, but seems like
3875 * marginal gain for something that is now known to be a slower
3876 * path. So just put it, and we'll get a new one when we retry.
3880 ret = PTR_ERR(file);
3881 /* only retry if RESOLVE_CACHED wasn't already set by application */
3882 if (ret == -EAGAIN &&
3883 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3888 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3889 file->f_flags &= ~O_NONBLOCK;
3890 fsnotify_open(file);
3891 fd_install(ret, file);
3893 putname(req->open.filename);
3894 req->flags &= ~REQ_F_NEED_CLEANUP;
3897 __io_req_complete(req, issue_flags, ret, 0);
3901 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3903 return io_openat2(req, issue_flags);
3906 static int io_remove_buffers_prep(struct io_kiocb *req,
3907 const struct io_uring_sqe *sqe)
3909 struct io_provide_buf *p = &req->pbuf;
3912 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3915 tmp = READ_ONCE(sqe->fd);
3916 if (!tmp || tmp > USHRT_MAX)
3919 memset(p, 0, sizeof(*p));
3921 p->bgid = READ_ONCE(sqe->buf_group);
3925 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3926 int bgid, unsigned nbufs)
3930 /* shouldn't happen */
3934 /* the head kbuf is the list itself */
3935 while (!list_empty(&buf->list)) {
3936 struct io_buffer *nxt;
3938 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3939 list_del(&nxt->list);
3946 xa_erase(&ctx->io_buffers, bgid);
3951 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3953 struct io_provide_buf *p = &req->pbuf;
3954 struct io_ring_ctx *ctx = req->ctx;
3955 struct io_buffer *head;
3957 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3959 io_ring_submit_lock(ctx, !force_nonblock);
3961 lockdep_assert_held(&ctx->uring_lock);
3964 head = xa_load(&ctx->io_buffers, p->bgid);
3966 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3970 /* complete before unlock, IOPOLL may need the lock */
3971 __io_req_complete(req, issue_flags, ret, 0);
3972 io_ring_submit_unlock(ctx, !force_nonblock);
3976 static int io_provide_buffers_prep(struct io_kiocb *req,
3977 const struct io_uring_sqe *sqe)
3979 unsigned long size, tmp_check;
3980 struct io_provide_buf *p = &req->pbuf;
3983 if (sqe->ioprio || sqe->rw_flags)
3986 tmp = READ_ONCE(sqe->fd);
3987 if (!tmp || tmp > USHRT_MAX)
3990 p->addr = READ_ONCE(sqe->addr);
3991 p->len = READ_ONCE(sqe->len);
3993 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3996 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3999 size = (unsigned long)p->len * p->nbufs;
4000 if (!access_ok(u64_to_user_ptr(p->addr), size))
4003 p->bgid = READ_ONCE(sqe->buf_group);
4004 tmp = READ_ONCE(sqe->off);
4005 if (tmp > USHRT_MAX)
4011 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4013 struct io_buffer *buf;
4014 u64 addr = pbuf->addr;
4015 int i, bid = pbuf->bid;
4017 for (i = 0; i < pbuf->nbufs; i++) {
4018 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4023 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4028 INIT_LIST_HEAD(&buf->list);
4031 list_add_tail(&buf->list, &(*head)->list);
4035 return i ? i : -ENOMEM;
4038 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4040 struct io_provide_buf *p = &req->pbuf;
4041 struct io_ring_ctx *ctx = req->ctx;
4042 struct io_buffer *head, *list;
4044 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4046 io_ring_submit_lock(ctx, !force_nonblock);
4048 lockdep_assert_held(&ctx->uring_lock);
4050 list = head = xa_load(&ctx->io_buffers, p->bgid);
4052 ret = io_add_buffers(p, &head);
4053 if (ret >= 0 && !list) {
4054 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4056 __io_remove_buffers(ctx, head, p->bgid, -1U);
4060 /* complete before unlock, IOPOLL may need the lock */
4061 __io_req_complete(req, issue_flags, ret, 0);
4062 io_ring_submit_unlock(ctx, !force_nonblock);
4066 static int io_epoll_ctl_prep(struct io_kiocb *req,
4067 const struct io_uring_sqe *sqe)
4069 #if defined(CONFIG_EPOLL)
4070 if (sqe->ioprio || sqe->buf_index)
4072 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4075 req->epoll.epfd = READ_ONCE(sqe->fd);
4076 req->epoll.op = READ_ONCE(sqe->len);
4077 req->epoll.fd = READ_ONCE(sqe->off);
4079 if (ep_op_has_event(req->epoll.op)) {
4080 struct epoll_event __user *ev;
4082 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4083 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4093 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4095 #if defined(CONFIG_EPOLL)
4096 struct io_epoll *ie = &req->epoll;
4098 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4100 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4101 if (force_nonblock && ret == -EAGAIN)
4106 __io_req_complete(req, issue_flags, ret, 0);
4113 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4115 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4116 if (sqe->ioprio || sqe->buf_index || sqe->off)
4118 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4121 req->madvise.addr = READ_ONCE(sqe->addr);
4122 req->madvise.len = READ_ONCE(sqe->len);
4123 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4130 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4132 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4133 struct io_madvise *ma = &req->madvise;
4136 if (issue_flags & IO_URING_F_NONBLOCK)
4139 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4142 io_req_complete(req, ret);
4149 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4151 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4153 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4156 req->fadvise.offset = READ_ONCE(sqe->off);
4157 req->fadvise.len = READ_ONCE(sqe->len);
4158 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4162 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4164 struct io_fadvise *fa = &req->fadvise;
4167 if (issue_flags & IO_URING_F_NONBLOCK) {
4168 switch (fa->advice) {
4169 case POSIX_FADV_NORMAL:
4170 case POSIX_FADV_RANDOM:
4171 case POSIX_FADV_SEQUENTIAL:
4178 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4181 __io_req_complete(req, issue_flags, ret, 0);
4185 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4187 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4189 if (sqe->ioprio || sqe->buf_index)
4191 if (req->flags & REQ_F_FIXED_FILE)
4194 req->statx.dfd = READ_ONCE(sqe->fd);
4195 req->statx.mask = READ_ONCE(sqe->len);
4196 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4197 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4198 req->statx.flags = READ_ONCE(sqe->statx_flags);
4203 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4205 struct io_statx *ctx = &req->statx;
4208 if (issue_flags & IO_URING_F_NONBLOCK)
4211 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4216 io_req_complete(req, ret);
4220 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4222 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4224 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4225 sqe->rw_flags || sqe->buf_index)
4227 if (req->flags & REQ_F_FIXED_FILE)
4230 req->close.fd = READ_ONCE(sqe->fd);
4234 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4236 struct files_struct *files = current->files;
4237 struct io_close *close = &req->close;
4238 struct fdtable *fdt;
4239 struct file *file = NULL;
4242 spin_lock(&files->file_lock);
4243 fdt = files_fdtable(files);
4244 if (close->fd >= fdt->max_fds) {
4245 spin_unlock(&files->file_lock);
4248 file = fdt->fd[close->fd];
4249 if (!file || file->f_op == &io_uring_fops) {
4250 spin_unlock(&files->file_lock);
4255 /* if the file has a flush method, be safe and punt to async */
4256 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4257 spin_unlock(&files->file_lock);
4261 ret = __close_fd_get_file(close->fd, &file);
4262 spin_unlock(&files->file_lock);
4269 /* No ->flush() or already async, safely close from here */
4270 ret = filp_close(file, current->files);
4276 __io_req_complete(req, issue_flags, ret, 0);
4280 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4282 struct io_ring_ctx *ctx = req->ctx;
4284 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4286 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4289 req->sync.off = READ_ONCE(sqe->off);
4290 req->sync.len = READ_ONCE(sqe->len);
4291 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4295 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4299 /* sync_file_range always requires a blocking context */
4300 if (issue_flags & IO_URING_F_NONBLOCK)
4303 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4307 io_req_complete(req, ret);
4311 #if defined(CONFIG_NET)
4312 static int io_setup_async_msg(struct io_kiocb *req,
4313 struct io_async_msghdr *kmsg)
4315 struct io_async_msghdr *async_msg = req->async_data;
4319 if (io_alloc_async_data(req)) {
4320 kfree(kmsg->free_iov);
4323 async_msg = req->async_data;
4324 req->flags |= REQ_F_NEED_CLEANUP;
4325 memcpy(async_msg, kmsg, sizeof(*kmsg));
4326 async_msg->msg.msg_name = &async_msg->addr;
4327 /* if were using fast_iov, set it to the new one */
4328 if (!async_msg->free_iov)
4329 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4334 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4335 struct io_async_msghdr *iomsg)
4337 iomsg->msg.msg_name = &iomsg->addr;
4338 iomsg->free_iov = iomsg->fast_iov;
4339 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4340 req->sr_msg.msg_flags, &iomsg->free_iov);
4343 static int io_sendmsg_prep_async(struct io_kiocb *req)
4347 ret = io_sendmsg_copy_hdr(req, req->async_data);
4349 req->flags |= REQ_F_NEED_CLEANUP;
4353 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4355 struct io_sr_msg *sr = &req->sr_msg;
4357 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4360 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4361 sr->len = READ_ONCE(sqe->len);
4362 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4363 if (sr->msg_flags & MSG_DONTWAIT)
4364 req->flags |= REQ_F_NOWAIT;
4366 #ifdef CONFIG_COMPAT
4367 if (req->ctx->compat)
4368 sr->msg_flags |= MSG_CMSG_COMPAT;
4373 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4375 struct io_async_msghdr iomsg, *kmsg;
4376 struct socket *sock;
4381 sock = sock_from_file(req->file);
4382 if (unlikely(!sock))
4385 kmsg = req->async_data;
4387 ret = io_sendmsg_copy_hdr(req, &iomsg);
4393 flags = req->sr_msg.msg_flags;
4394 if (issue_flags & IO_URING_F_NONBLOCK)
4395 flags |= MSG_DONTWAIT;
4396 if (flags & MSG_WAITALL)
4397 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4399 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4400 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4401 return io_setup_async_msg(req, kmsg);
4402 if (ret == -ERESTARTSYS)
4405 /* fast path, check for non-NULL to avoid function call */
4407 kfree(kmsg->free_iov);
4408 req->flags &= ~REQ_F_NEED_CLEANUP;
4411 __io_req_complete(req, issue_flags, ret, 0);
4415 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4417 struct io_sr_msg *sr = &req->sr_msg;
4420 struct socket *sock;
4425 sock = sock_from_file(req->file);
4426 if (unlikely(!sock))
4429 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4433 msg.msg_name = NULL;
4434 msg.msg_control = NULL;
4435 msg.msg_controllen = 0;
4436 msg.msg_namelen = 0;
4438 flags = req->sr_msg.msg_flags;
4439 if (issue_flags & IO_URING_F_NONBLOCK)
4440 flags |= MSG_DONTWAIT;
4441 if (flags & MSG_WAITALL)
4442 min_ret = iov_iter_count(&msg.msg_iter);
4444 msg.msg_flags = flags;
4445 ret = sock_sendmsg(sock, &msg);
4446 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4448 if (ret == -ERESTARTSYS)
4453 __io_req_complete(req, issue_flags, ret, 0);
4457 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4458 struct io_async_msghdr *iomsg)
4460 struct io_sr_msg *sr = &req->sr_msg;
4461 struct iovec __user *uiov;
4465 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4466 &iomsg->uaddr, &uiov, &iov_len);
4470 if (req->flags & REQ_F_BUFFER_SELECT) {
4473 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4475 sr->len = iomsg->fast_iov[0].iov_len;
4476 iomsg->free_iov = NULL;
4478 iomsg->free_iov = iomsg->fast_iov;
4479 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4480 &iomsg->free_iov, &iomsg->msg.msg_iter,
4489 #ifdef CONFIG_COMPAT
4490 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4491 struct io_async_msghdr *iomsg)
4493 struct io_sr_msg *sr = &req->sr_msg;
4494 struct compat_iovec __user *uiov;
4499 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4504 uiov = compat_ptr(ptr);
4505 if (req->flags & REQ_F_BUFFER_SELECT) {
4506 compat_ssize_t clen;
4510 if (!access_ok(uiov, sizeof(*uiov)))
4512 if (__get_user(clen, &uiov->iov_len))
4517 iomsg->free_iov = NULL;
4519 iomsg->free_iov = iomsg->fast_iov;
4520 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4521 UIO_FASTIOV, &iomsg->free_iov,
4522 &iomsg->msg.msg_iter, true);
4531 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4532 struct io_async_msghdr *iomsg)
4534 iomsg->msg.msg_name = &iomsg->addr;
4536 #ifdef CONFIG_COMPAT
4537 if (req->ctx->compat)
4538 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4541 return __io_recvmsg_copy_hdr(req, iomsg);
4544 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4547 struct io_sr_msg *sr = &req->sr_msg;
4548 struct io_buffer *kbuf;
4550 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4555 req->flags |= REQ_F_BUFFER_SELECTED;
4559 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4561 return io_put_kbuf(req, req->sr_msg.kbuf);
4564 static int io_recvmsg_prep_async(struct io_kiocb *req)
4568 ret = io_recvmsg_copy_hdr(req, req->async_data);
4570 req->flags |= REQ_F_NEED_CLEANUP;
4574 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4576 struct io_sr_msg *sr = &req->sr_msg;
4578 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4581 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4582 sr->len = READ_ONCE(sqe->len);
4583 sr->bgid = READ_ONCE(sqe->buf_group);
4584 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4585 if (sr->msg_flags & MSG_DONTWAIT)
4586 req->flags |= REQ_F_NOWAIT;
4588 #ifdef CONFIG_COMPAT
4589 if (req->ctx->compat)
4590 sr->msg_flags |= MSG_CMSG_COMPAT;
4595 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4597 struct io_async_msghdr iomsg, *kmsg;
4598 struct socket *sock;
4599 struct io_buffer *kbuf;
4602 int ret, cflags = 0;
4603 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4605 sock = sock_from_file(req->file);
4606 if (unlikely(!sock))
4609 kmsg = req->async_data;
4611 ret = io_recvmsg_copy_hdr(req, &iomsg);
4617 if (req->flags & REQ_F_BUFFER_SELECT) {
4618 kbuf = io_recv_buffer_select(req, !force_nonblock);
4620 return PTR_ERR(kbuf);
4621 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4622 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4623 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4624 1, req->sr_msg.len);
4627 flags = req->sr_msg.msg_flags;
4629 flags |= MSG_DONTWAIT;
4630 if (flags & MSG_WAITALL)
4631 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4633 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4634 kmsg->uaddr, flags);
4635 if (force_nonblock && ret == -EAGAIN)
4636 return io_setup_async_msg(req, kmsg);
4637 if (ret == -ERESTARTSYS)
4640 if (req->flags & REQ_F_BUFFER_SELECTED)
4641 cflags = io_put_recv_kbuf(req);
4642 /* fast path, check for non-NULL to avoid function call */
4644 kfree(kmsg->free_iov);
4645 req->flags &= ~REQ_F_NEED_CLEANUP;
4646 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4648 __io_req_complete(req, issue_flags, ret, cflags);
4652 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4654 struct io_buffer *kbuf;
4655 struct io_sr_msg *sr = &req->sr_msg;
4657 void __user *buf = sr->buf;
4658 struct socket *sock;
4662 int ret, cflags = 0;
4663 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4665 sock = sock_from_file(req->file);
4666 if (unlikely(!sock))
4669 if (req->flags & REQ_F_BUFFER_SELECT) {
4670 kbuf = io_recv_buffer_select(req, !force_nonblock);
4672 return PTR_ERR(kbuf);
4673 buf = u64_to_user_ptr(kbuf->addr);
4676 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4680 msg.msg_name = NULL;
4681 msg.msg_control = NULL;
4682 msg.msg_controllen = 0;
4683 msg.msg_namelen = 0;
4684 msg.msg_iocb = NULL;
4687 flags = req->sr_msg.msg_flags;
4689 flags |= MSG_DONTWAIT;
4690 if (flags & MSG_WAITALL)
4691 min_ret = iov_iter_count(&msg.msg_iter);
4693 ret = sock_recvmsg(sock, &msg, flags);
4694 if (force_nonblock && ret == -EAGAIN)
4696 if (ret == -ERESTARTSYS)
4699 if (req->flags & REQ_F_BUFFER_SELECTED)
4700 cflags = io_put_recv_kbuf(req);
4701 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4703 __io_req_complete(req, issue_flags, ret, cflags);
4707 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4709 struct io_accept *accept = &req->accept;
4711 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4713 if (sqe->ioprio || sqe->len || sqe->buf_index)
4716 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4717 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4718 accept->flags = READ_ONCE(sqe->accept_flags);
4719 accept->nofile = rlimit(RLIMIT_NOFILE);
4723 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4725 struct io_accept *accept = &req->accept;
4726 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4727 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4730 if (req->file->f_flags & O_NONBLOCK)
4731 req->flags |= REQ_F_NOWAIT;
4733 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4734 accept->addr_len, accept->flags,
4736 if (ret == -EAGAIN && force_nonblock)
4739 if (ret == -ERESTARTSYS)
4743 __io_req_complete(req, issue_flags, ret, 0);
4747 static int io_connect_prep_async(struct io_kiocb *req)
4749 struct io_async_connect *io = req->async_data;
4750 struct io_connect *conn = &req->connect;
4752 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4755 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4757 struct io_connect *conn = &req->connect;
4759 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4761 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4764 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4765 conn->addr_len = READ_ONCE(sqe->addr2);
4769 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4771 struct io_async_connect __io, *io;
4772 unsigned file_flags;
4774 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4776 if (req->async_data) {
4777 io = req->async_data;
4779 ret = move_addr_to_kernel(req->connect.addr,
4780 req->connect.addr_len,
4787 file_flags = force_nonblock ? O_NONBLOCK : 0;
4789 ret = __sys_connect_file(req->file, &io->address,
4790 req->connect.addr_len, file_flags);
4791 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4792 if (req->async_data)
4794 if (io_alloc_async_data(req)) {
4798 memcpy(req->async_data, &__io, sizeof(__io));
4801 if (ret == -ERESTARTSYS)
4806 __io_req_complete(req, issue_flags, ret, 0);
4809 #else /* !CONFIG_NET */
4810 #define IO_NETOP_FN(op) \
4811 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4813 return -EOPNOTSUPP; \
4816 #define IO_NETOP_PREP(op) \
4818 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4820 return -EOPNOTSUPP; \
4823 #define IO_NETOP_PREP_ASYNC(op) \
4825 static int io_##op##_prep_async(struct io_kiocb *req) \
4827 return -EOPNOTSUPP; \
4830 IO_NETOP_PREP_ASYNC(sendmsg);
4831 IO_NETOP_PREP_ASYNC(recvmsg);
4832 IO_NETOP_PREP_ASYNC(connect);
4833 IO_NETOP_PREP(accept);
4836 #endif /* CONFIG_NET */
4838 struct io_poll_table {
4839 struct poll_table_struct pt;
4840 struct io_kiocb *req;
4845 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4846 __poll_t mask, io_req_tw_func_t func)
4848 /* for instances that support it check for an event match first: */
4849 if (mask && !(mask & poll->events))
4852 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4854 list_del_init(&poll->wait.entry);
4857 req->io_task_work.func = func;
4860 * If this fails, then the task is exiting. When a task exits, the
4861 * work gets canceled, so just cancel this request as well instead
4862 * of executing it. We can't safely execute it anyway, as we may not
4863 * have the needed state needed for it anyway.
4865 io_req_task_work_add(req);
4869 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4870 __acquires(&req->ctx->completion_lock)
4872 struct io_ring_ctx *ctx = req->ctx;
4874 if (unlikely(req->task->flags & PF_EXITING))
4875 WRITE_ONCE(poll->canceled, true);
4877 if (!req->result && !READ_ONCE(poll->canceled)) {
4878 struct poll_table_struct pt = { ._key = poll->events };
4880 req->result = vfs_poll(req->file, &pt) & poll->events;
4883 spin_lock_irq(&ctx->completion_lock);
4884 if (!req->result && !READ_ONCE(poll->canceled)) {
4885 add_wait_queue(poll->head, &poll->wait);
4892 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4894 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4895 if (req->opcode == IORING_OP_POLL_ADD)
4896 return req->async_data;
4897 return req->apoll->double_poll;
4900 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4902 if (req->opcode == IORING_OP_POLL_ADD)
4904 return &req->apoll->poll;
4907 static void io_poll_remove_double(struct io_kiocb *req)
4908 __must_hold(&req->ctx->completion_lock)
4910 struct io_poll_iocb *poll = io_poll_get_double(req);
4912 lockdep_assert_held(&req->ctx->completion_lock);
4914 if (poll && poll->head) {
4915 struct wait_queue_head *head = poll->head;
4917 spin_lock(&head->lock);
4918 list_del_init(&poll->wait.entry);
4919 if (poll->wait.private)
4922 spin_unlock(&head->lock);
4926 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4927 __must_hold(&req->ctx->completion_lock)
4929 struct io_ring_ctx *ctx = req->ctx;
4930 unsigned flags = IORING_CQE_F_MORE;
4933 if (READ_ONCE(req->poll.canceled)) {
4935 req->poll.events |= EPOLLONESHOT;
4937 error = mangle_poll(mask);
4939 if (req->poll.events & EPOLLONESHOT)
4941 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4942 req->poll.done = true;
4945 if (flags & IORING_CQE_F_MORE)
4948 io_commit_cqring(ctx);
4949 return !(flags & IORING_CQE_F_MORE);
4952 static void io_poll_task_func(struct io_kiocb *req)
4954 struct io_ring_ctx *ctx = req->ctx;
4955 struct io_kiocb *nxt;
4957 if (io_poll_rewait(req, &req->poll)) {
4958 spin_unlock_irq(&ctx->completion_lock);
4962 done = io_poll_complete(req, req->result);
4964 io_poll_remove_double(req);
4965 hash_del(&req->hash_node);
4968 add_wait_queue(req->poll.head, &req->poll.wait);
4970 spin_unlock_irq(&ctx->completion_lock);
4971 io_cqring_ev_posted(ctx);
4974 nxt = io_put_req_find_next(req);
4976 io_req_task_submit(nxt);
4981 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4982 int sync, void *key)
4984 struct io_kiocb *req = wait->private;
4985 struct io_poll_iocb *poll = io_poll_get_single(req);
4986 __poll_t mask = key_to_poll(key);
4988 /* for instances that support it check for an event match first: */
4989 if (mask && !(mask & poll->events))
4991 if (!(poll->events & EPOLLONESHOT))
4992 return poll->wait.func(&poll->wait, mode, sync, key);
4994 list_del_init(&wait->entry);
4999 spin_lock(&poll->head->lock);
5000 done = list_empty(&poll->wait.entry);
5002 list_del_init(&poll->wait.entry);
5003 /* make sure double remove sees this as being gone */
5004 wait->private = NULL;
5005 spin_unlock(&poll->head->lock);
5007 /* use wait func handler, so it matches the rq type */
5008 poll->wait.func(&poll->wait, mode, sync, key);
5015 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5016 wait_queue_func_t wake_func)
5020 poll->canceled = false;
5021 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5022 /* mask in events that we always want/need */
5023 poll->events = events | IO_POLL_UNMASK;
5024 INIT_LIST_HEAD(&poll->wait.entry);
5025 init_waitqueue_func_entry(&poll->wait, wake_func);
5028 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5029 struct wait_queue_head *head,
5030 struct io_poll_iocb **poll_ptr)
5032 struct io_kiocb *req = pt->req;
5035 * The file being polled uses multiple waitqueues for poll handling
5036 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5039 if (unlikely(pt->nr_entries)) {
5040 struct io_poll_iocb *poll_one = poll;
5042 /* already have a 2nd entry, fail a third attempt */
5044 pt->error = -EINVAL;
5048 * Can't handle multishot for double wait for now, turn it
5049 * into one-shot mode.
5051 if (!(poll_one->events & EPOLLONESHOT))
5052 poll_one->events |= EPOLLONESHOT;
5053 /* double add on the same waitqueue head, ignore */
5054 if (poll_one->head == head)
5056 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5058 pt->error = -ENOMEM;
5061 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5063 poll->wait.private = req;
5070 if (poll->events & EPOLLEXCLUSIVE)
5071 add_wait_queue_exclusive(head, &poll->wait);
5073 add_wait_queue(head, &poll->wait);
5076 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5077 struct poll_table_struct *p)
5079 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5080 struct async_poll *apoll = pt->req->apoll;
5082 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5085 static void io_async_task_func(struct io_kiocb *req)
5087 struct async_poll *apoll = req->apoll;
5088 struct io_ring_ctx *ctx = req->ctx;
5090 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5092 if (io_poll_rewait(req, &apoll->poll)) {
5093 spin_unlock_irq(&ctx->completion_lock);
5097 hash_del(&req->hash_node);
5098 io_poll_remove_double(req);
5099 spin_unlock_irq(&ctx->completion_lock);
5101 if (!READ_ONCE(apoll->poll.canceled))
5102 io_req_task_submit(req);
5104 io_req_complete_failed(req, -ECANCELED);
5107 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5110 struct io_kiocb *req = wait->private;
5111 struct io_poll_iocb *poll = &req->apoll->poll;
5113 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5116 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5119 static void io_poll_req_insert(struct io_kiocb *req)
5121 struct io_ring_ctx *ctx = req->ctx;
5122 struct hlist_head *list;
5124 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5125 hlist_add_head(&req->hash_node, list);
5128 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5129 struct io_poll_iocb *poll,
5130 struct io_poll_table *ipt, __poll_t mask,
5131 wait_queue_func_t wake_func)
5132 __acquires(&ctx->completion_lock)
5134 struct io_ring_ctx *ctx = req->ctx;
5135 bool cancel = false;
5137 INIT_HLIST_NODE(&req->hash_node);
5138 io_init_poll_iocb(poll, mask, wake_func);
5139 poll->file = req->file;
5140 poll->wait.private = req;
5142 ipt->pt._key = mask;
5145 ipt->nr_entries = 0;
5147 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5148 if (unlikely(!ipt->nr_entries) && !ipt->error)
5149 ipt->error = -EINVAL;
5151 spin_lock_irq(&ctx->completion_lock);
5152 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5153 io_poll_remove_double(req);
5154 if (likely(poll->head)) {
5155 spin_lock(&poll->head->lock);
5156 if (unlikely(list_empty(&poll->wait.entry))) {
5162 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5163 list_del_init(&poll->wait.entry);
5165 WRITE_ONCE(poll->canceled, true);
5166 else if (!poll->done) /* actually waiting for an event */
5167 io_poll_req_insert(req);
5168 spin_unlock(&poll->head->lock);
5180 static int io_arm_poll_handler(struct io_kiocb *req)
5182 const struct io_op_def *def = &io_op_defs[req->opcode];
5183 struct io_ring_ctx *ctx = req->ctx;
5184 struct async_poll *apoll;
5185 struct io_poll_table ipt;
5186 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5189 if (!req->file || !file_can_poll(req->file))
5190 return IO_APOLL_ABORTED;
5191 if (req->flags & REQ_F_POLLED)
5192 return IO_APOLL_ABORTED;
5193 if (!def->pollin && !def->pollout)
5194 return IO_APOLL_ABORTED;
5198 mask |= POLLIN | POLLRDNORM;
5200 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5201 if ((req->opcode == IORING_OP_RECVMSG) &&
5202 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5206 mask |= POLLOUT | POLLWRNORM;
5209 /* if we can't nonblock try, then no point in arming a poll handler */
5210 if (!io_file_supports_nowait(req, rw))
5211 return IO_APOLL_ABORTED;
5213 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5214 if (unlikely(!apoll))
5215 return IO_APOLL_ABORTED;
5216 apoll->double_poll = NULL;
5218 req->flags |= REQ_F_POLLED;
5219 ipt.pt._qproc = io_async_queue_proc;
5221 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5223 if (ret || ipt.error) {
5224 spin_unlock_irq(&ctx->completion_lock);
5226 return IO_APOLL_READY;
5227 return IO_APOLL_ABORTED;
5229 spin_unlock_irq(&ctx->completion_lock);
5230 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5231 mask, apoll->poll.events);
5235 static bool __io_poll_remove_one(struct io_kiocb *req,
5236 struct io_poll_iocb *poll, bool do_cancel)
5237 __must_hold(&req->ctx->completion_lock)
5239 bool do_complete = false;
5243 spin_lock(&poll->head->lock);
5245 WRITE_ONCE(poll->canceled, true);
5246 if (!list_empty(&poll->wait.entry)) {
5247 list_del_init(&poll->wait.entry);
5250 spin_unlock(&poll->head->lock);
5251 hash_del(&req->hash_node);
5255 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5256 __must_hold(&req->ctx->completion_lock)
5260 io_poll_remove_double(req);
5261 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5263 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5264 /* non-poll requests have submit ref still */
5270 static bool io_poll_remove_one(struct io_kiocb *req)
5271 __must_hold(&req->ctx->completion_lock)
5275 do_complete = io_poll_remove_waitqs(req);
5277 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5278 io_commit_cqring(req->ctx);
5280 io_put_req_deferred(req, 1);
5287 * Returns true if we found and killed one or more poll requests
5289 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5292 struct hlist_node *tmp;
5293 struct io_kiocb *req;
5296 spin_lock_irq(&ctx->completion_lock);
5297 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5298 struct hlist_head *list;
5300 list = &ctx->cancel_hash[i];
5301 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5302 if (io_match_task(req, tsk, cancel_all))
5303 posted += io_poll_remove_one(req);
5306 spin_unlock_irq(&ctx->completion_lock);
5309 io_cqring_ev_posted(ctx);
5314 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5316 __must_hold(&ctx->completion_lock)
5318 struct hlist_head *list;
5319 struct io_kiocb *req;
5321 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5322 hlist_for_each_entry(req, list, hash_node) {
5323 if (sqe_addr != req->user_data)
5325 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5332 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5334 __must_hold(&ctx->completion_lock)
5336 struct io_kiocb *req;
5338 req = io_poll_find(ctx, sqe_addr, poll_only);
5341 if (io_poll_remove_one(req))
5347 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5352 events = READ_ONCE(sqe->poll32_events);
5354 events = swahw32(events);
5356 if (!(flags & IORING_POLL_ADD_MULTI))
5357 events |= EPOLLONESHOT;
5358 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5361 static int io_poll_update_prep(struct io_kiocb *req,
5362 const struct io_uring_sqe *sqe)
5364 struct io_poll_update *upd = &req->poll_update;
5367 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5369 if (sqe->ioprio || sqe->buf_index)
5371 flags = READ_ONCE(sqe->len);
5372 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5373 IORING_POLL_ADD_MULTI))
5375 /* meaningless without update */
5376 if (flags == IORING_POLL_ADD_MULTI)
5379 upd->old_user_data = READ_ONCE(sqe->addr);
5380 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5381 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5383 upd->new_user_data = READ_ONCE(sqe->off);
5384 if (!upd->update_user_data && upd->new_user_data)
5386 if (upd->update_events)
5387 upd->events = io_poll_parse_events(sqe, flags);
5388 else if (sqe->poll32_events)
5394 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5397 struct io_kiocb *req = wait->private;
5398 struct io_poll_iocb *poll = &req->poll;
5400 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5403 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5404 struct poll_table_struct *p)
5406 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5408 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5411 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5413 struct io_poll_iocb *poll = &req->poll;
5416 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5418 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5420 flags = READ_ONCE(sqe->len);
5421 if (flags & ~IORING_POLL_ADD_MULTI)
5424 poll->events = io_poll_parse_events(sqe, flags);
5428 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5430 struct io_poll_iocb *poll = &req->poll;
5431 struct io_ring_ctx *ctx = req->ctx;
5432 struct io_poll_table ipt;
5435 ipt.pt._qproc = io_poll_queue_proc;
5437 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5440 if (mask) { /* no async, we'd stolen it */
5442 io_poll_complete(req, mask);
5444 spin_unlock_irq(&ctx->completion_lock);
5447 io_cqring_ev_posted(ctx);
5448 if (poll->events & EPOLLONESHOT)
5454 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5456 struct io_ring_ctx *ctx = req->ctx;
5457 struct io_kiocb *preq;
5461 spin_lock_irq(&ctx->completion_lock);
5462 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5468 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5470 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5475 * Don't allow racy completion with singleshot, as we cannot safely
5476 * update those. For multishot, if we're racing with completion, just
5477 * let completion re-add it.
5479 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5480 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5484 /* we now have a detached poll request. reissue. */
5488 spin_unlock_irq(&ctx->completion_lock);
5490 io_req_complete(req, ret);
5493 /* only mask one event flags, keep behavior flags */
5494 if (req->poll_update.update_events) {
5495 preq->poll.events &= ~0xffff;
5496 preq->poll.events |= req->poll_update.events & 0xffff;
5497 preq->poll.events |= IO_POLL_UNMASK;
5499 if (req->poll_update.update_user_data)
5500 preq->user_data = req->poll_update.new_user_data;
5501 spin_unlock_irq(&ctx->completion_lock);
5503 /* complete update request, we're done with it */
5504 io_req_complete(req, ret);
5507 ret = io_poll_add(preq, issue_flags);
5510 io_req_complete(preq, ret);
5516 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5518 struct io_timeout_data *data = container_of(timer,
5519 struct io_timeout_data, timer);
5520 struct io_kiocb *req = data->req;
5521 struct io_ring_ctx *ctx = req->ctx;
5522 unsigned long flags;
5524 spin_lock_irqsave(&ctx->completion_lock, flags);
5525 list_del_init(&req->timeout.list);
5526 atomic_set(&req->ctx->cq_timeouts,
5527 atomic_read(&req->ctx->cq_timeouts) + 1);
5529 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5530 io_commit_cqring(ctx);
5531 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5533 io_cqring_ev_posted(ctx);
5536 return HRTIMER_NORESTART;
5539 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5541 __must_hold(&ctx->completion_lock)
5543 struct io_timeout_data *io;
5544 struct io_kiocb *req;
5547 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5548 found = user_data == req->user_data;
5553 return ERR_PTR(-ENOENT);
5555 io = req->async_data;
5556 if (hrtimer_try_to_cancel(&io->timer) == -1)
5557 return ERR_PTR(-EALREADY);
5558 list_del_init(&req->timeout.list);
5562 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5563 __must_hold(&ctx->completion_lock)
5565 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5568 return PTR_ERR(req);
5571 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5572 io_put_req_deferred(req, 1);
5576 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5577 struct timespec64 *ts, enum hrtimer_mode mode)
5578 __must_hold(&ctx->completion_lock)
5580 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5581 struct io_timeout_data *data;
5584 return PTR_ERR(req);
5586 req->timeout.off = 0; /* noseq */
5587 data = req->async_data;
5588 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5589 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5590 data->timer.function = io_timeout_fn;
5591 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5595 static int io_timeout_remove_prep(struct io_kiocb *req,
5596 const struct io_uring_sqe *sqe)
5598 struct io_timeout_rem *tr = &req->timeout_rem;
5600 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5602 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5604 if (sqe->ioprio || sqe->buf_index || sqe->len)
5607 tr->addr = READ_ONCE(sqe->addr);
5608 tr->flags = READ_ONCE(sqe->timeout_flags);
5609 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5610 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5612 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5614 } else if (tr->flags) {
5615 /* timeout removal doesn't support flags */
5622 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5624 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5629 * Remove or update an existing timeout command
5631 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5633 struct io_timeout_rem *tr = &req->timeout_rem;
5634 struct io_ring_ctx *ctx = req->ctx;
5637 spin_lock_irq(&ctx->completion_lock);
5638 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5639 ret = io_timeout_cancel(ctx, tr->addr);
5641 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5642 io_translate_timeout_mode(tr->flags));
5644 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5645 io_commit_cqring(ctx);
5646 spin_unlock_irq(&ctx->completion_lock);
5647 io_cqring_ev_posted(ctx);
5654 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5655 bool is_timeout_link)
5657 struct io_timeout_data *data;
5659 u32 off = READ_ONCE(sqe->off);
5661 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5663 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5665 if (off && is_timeout_link)
5667 flags = READ_ONCE(sqe->timeout_flags);
5668 if (flags & ~IORING_TIMEOUT_ABS)
5671 req->timeout.off = off;
5672 if (unlikely(off && !req->ctx->off_timeout_used))
5673 req->ctx->off_timeout_used = true;
5675 if (!req->async_data && io_alloc_async_data(req))
5678 data = req->async_data;
5681 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5684 data->mode = io_translate_timeout_mode(flags);
5685 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5686 if (is_timeout_link)
5687 io_req_track_inflight(req);
5691 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5693 struct io_ring_ctx *ctx = req->ctx;
5694 struct io_timeout_data *data = req->async_data;
5695 struct list_head *entry;
5696 u32 tail, off = req->timeout.off;
5698 spin_lock_irq(&ctx->completion_lock);
5701 * sqe->off holds how many events that need to occur for this
5702 * timeout event to be satisfied. If it isn't set, then this is
5703 * a pure timeout request, sequence isn't used.
5705 if (io_is_timeout_noseq(req)) {
5706 entry = ctx->timeout_list.prev;
5710 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5711 req->timeout.target_seq = tail + off;
5713 /* Update the last seq here in case io_flush_timeouts() hasn't.
5714 * This is safe because ->completion_lock is held, and submissions
5715 * and completions are never mixed in the same ->completion_lock section.
5717 ctx->cq_last_tm_flush = tail;
5720 * Insertion sort, ensuring the first entry in the list is always
5721 * the one we need first.
5723 list_for_each_prev(entry, &ctx->timeout_list) {
5724 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5727 if (io_is_timeout_noseq(nxt))
5729 /* nxt.seq is behind @tail, otherwise would've been completed */
5730 if (off >= nxt->timeout.target_seq - tail)
5734 list_add(&req->timeout.list, entry);
5735 data->timer.function = io_timeout_fn;
5736 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5737 spin_unlock_irq(&ctx->completion_lock);
5741 struct io_cancel_data {
5742 struct io_ring_ctx *ctx;
5746 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5748 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5749 struct io_cancel_data *cd = data;
5751 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5754 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5755 struct io_ring_ctx *ctx)
5757 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5758 enum io_wq_cancel cancel_ret;
5761 if (!tctx || !tctx->io_wq)
5764 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5765 switch (cancel_ret) {
5766 case IO_WQ_CANCEL_OK:
5769 case IO_WQ_CANCEL_RUNNING:
5772 case IO_WQ_CANCEL_NOTFOUND:
5780 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5781 struct io_kiocb *req, __u64 sqe_addr,
5784 unsigned long flags;
5787 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5788 spin_lock_irqsave(&ctx->completion_lock, flags);
5791 ret = io_timeout_cancel(ctx, sqe_addr);
5794 ret = io_poll_cancel(ctx, sqe_addr, false);
5798 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5799 io_commit_cqring(ctx);
5800 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5801 io_cqring_ev_posted(ctx);
5807 static int io_async_cancel_prep(struct io_kiocb *req,
5808 const struct io_uring_sqe *sqe)
5810 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5812 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5814 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5817 req->cancel.addr = READ_ONCE(sqe->addr);
5821 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5823 struct io_ring_ctx *ctx = req->ctx;
5824 u64 sqe_addr = req->cancel.addr;
5825 struct io_tctx_node *node;
5828 /* tasks should wait for their io-wq threads, so safe w/o sync */
5829 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5830 spin_lock_irq(&ctx->completion_lock);
5833 ret = io_timeout_cancel(ctx, sqe_addr);
5836 ret = io_poll_cancel(ctx, sqe_addr, false);
5839 spin_unlock_irq(&ctx->completion_lock);
5841 /* slow path, try all io-wq's */
5842 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5844 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5845 struct io_uring_task *tctx = node->task->io_uring;
5847 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5851 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5853 spin_lock_irq(&ctx->completion_lock);
5855 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5856 io_commit_cqring(ctx);
5857 spin_unlock_irq(&ctx->completion_lock);
5858 io_cqring_ev_posted(ctx);
5866 static int io_rsrc_update_prep(struct io_kiocb *req,
5867 const struct io_uring_sqe *sqe)
5869 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5871 if (sqe->ioprio || sqe->rw_flags)
5874 req->rsrc_update.offset = READ_ONCE(sqe->off);
5875 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5876 if (!req->rsrc_update.nr_args)
5878 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5882 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5884 struct io_ring_ctx *ctx = req->ctx;
5885 struct io_uring_rsrc_update2 up;
5888 if (issue_flags & IO_URING_F_NONBLOCK)
5891 up.offset = req->rsrc_update.offset;
5892 up.data = req->rsrc_update.arg;
5897 mutex_lock(&ctx->uring_lock);
5898 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5899 &up, req->rsrc_update.nr_args);
5900 mutex_unlock(&ctx->uring_lock);
5904 __io_req_complete(req, issue_flags, ret, 0);
5908 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5910 switch (req->opcode) {
5913 case IORING_OP_READV:
5914 case IORING_OP_READ_FIXED:
5915 case IORING_OP_READ:
5916 return io_read_prep(req, sqe);
5917 case IORING_OP_WRITEV:
5918 case IORING_OP_WRITE_FIXED:
5919 case IORING_OP_WRITE:
5920 return io_write_prep(req, sqe);
5921 case IORING_OP_POLL_ADD:
5922 return io_poll_add_prep(req, sqe);
5923 case IORING_OP_POLL_REMOVE:
5924 return io_poll_update_prep(req, sqe);
5925 case IORING_OP_FSYNC:
5926 return io_fsync_prep(req, sqe);
5927 case IORING_OP_SYNC_FILE_RANGE:
5928 return io_sfr_prep(req, sqe);
5929 case IORING_OP_SENDMSG:
5930 case IORING_OP_SEND:
5931 return io_sendmsg_prep(req, sqe);
5932 case IORING_OP_RECVMSG:
5933 case IORING_OP_RECV:
5934 return io_recvmsg_prep(req, sqe);
5935 case IORING_OP_CONNECT:
5936 return io_connect_prep(req, sqe);
5937 case IORING_OP_TIMEOUT:
5938 return io_timeout_prep(req, sqe, false);
5939 case IORING_OP_TIMEOUT_REMOVE:
5940 return io_timeout_remove_prep(req, sqe);
5941 case IORING_OP_ASYNC_CANCEL:
5942 return io_async_cancel_prep(req, sqe);
5943 case IORING_OP_LINK_TIMEOUT:
5944 return io_timeout_prep(req, sqe, true);
5945 case IORING_OP_ACCEPT:
5946 return io_accept_prep(req, sqe);
5947 case IORING_OP_FALLOCATE:
5948 return io_fallocate_prep(req, sqe);
5949 case IORING_OP_OPENAT:
5950 return io_openat_prep(req, sqe);
5951 case IORING_OP_CLOSE:
5952 return io_close_prep(req, sqe);
5953 case IORING_OP_FILES_UPDATE:
5954 return io_rsrc_update_prep(req, sqe);
5955 case IORING_OP_STATX:
5956 return io_statx_prep(req, sqe);
5957 case IORING_OP_FADVISE:
5958 return io_fadvise_prep(req, sqe);
5959 case IORING_OP_MADVISE:
5960 return io_madvise_prep(req, sqe);
5961 case IORING_OP_OPENAT2:
5962 return io_openat2_prep(req, sqe);
5963 case IORING_OP_EPOLL_CTL:
5964 return io_epoll_ctl_prep(req, sqe);
5965 case IORING_OP_SPLICE:
5966 return io_splice_prep(req, sqe);
5967 case IORING_OP_PROVIDE_BUFFERS:
5968 return io_provide_buffers_prep(req, sqe);
5969 case IORING_OP_REMOVE_BUFFERS:
5970 return io_remove_buffers_prep(req, sqe);
5972 return io_tee_prep(req, sqe);
5973 case IORING_OP_SHUTDOWN:
5974 return io_shutdown_prep(req, sqe);
5975 case IORING_OP_RENAMEAT:
5976 return io_renameat_prep(req, sqe);
5977 case IORING_OP_UNLINKAT:
5978 return io_unlinkat_prep(req, sqe);
5981 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5986 static int io_req_prep_async(struct io_kiocb *req)
5988 if (!io_op_defs[req->opcode].needs_async_setup)
5990 if (WARN_ON_ONCE(req->async_data))
5992 if (io_alloc_async_data(req))
5995 switch (req->opcode) {
5996 case IORING_OP_READV:
5997 return io_rw_prep_async(req, READ);
5998 case IORING_OP_WRITEV:
5999 return io_rw_prep_async(req, WRITE);
6000 case IORING_OP_SENDMSG:
6001 return io_sendmsg_prep_async(req);
6002 case IORING_OP_RECVMSG:
6003 return io_recvmsg_prep_async(req);
6004 case IORING_OP_CONNECT:
6005 return io_connect_prep_async(req);
6007 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6012 static u32 io_get_sequence(struct io_kiocb *req)
6014 u32 seq = req->ctx->cached_sq_head;
6016 /* need original cached_sq_head, but it was increased for each req */
6017 io_for_each_link(req, req)
6022 static bool io_drain_req(struct io_kiocb *req)
6024 struct io_kiocb *pos;
6025 struct io_ring_ctx *ctx = req->ctx;
6026 struct io_defer_entry *de;
6031 * If we need to drain a request in the middle of a link, drain the
6032 * head request and the next request/link after the current link.
6033 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6034 * maintained for every request of our link.
6036 if (ctx->drain_next) {
6037 req->flags |= REQ_F_IO_DRAIN;
6038 ctx->drain_next = false;
6040 /* not interested in head, start from the first linked */
6041 io_for_each_link(pos, req->link) {
6042 if (pos->flags & REQ_F_IO_DRAIN) {
6043 ctx->drain_next = true;
6044 req->flags |= REQ_F_IO_DRAIN;
6049 /* Still need defer if there is pending req in defer list. */
6050 if (likely(list_empty_careful(&ctx->defer_list) &&
6051 !(req->flags & REQ_F_IO_DRAIN))) {
6052 ctx->drain_active = false;
6056 seq = io_get_sequence(req);
6057 /* Still a chance to pass the sequence check */
6058 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6061 ret = io_req_prep_async(req);
6064 io_prep_async_link(req);
6065 de = kmalloc(sizeof(*de), GFP_KERNEL);
6069 io_req_complete_failed(req, ret);
6073 spin_lock_irq(&ctx->completion_lock);
6074 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6075 spin_unlock_irq(&ctx->completion_lock);
6077 io_queue_async_work(req);
6081 trace_io_uring_defer(ctx, req, req->user_data);
6084 list_add_tail(&de->list, &ctx->defer_list);
6085 spin_unlock_irq(&ctx->completion_lock);
6089 static void io_clean_op(struct io_kiocb *req)
6091 if (req->flags & REQ_F_BUFFER_SELECTED) {
6092 switch (req->opcode) {
6093 case IORING_OP_READV:
6094 case IORING_OP_READ_FIXED:
6095 case IORING_OP_READ:
6096 kfree((void *)(unsigned long)req->rw.addr);
6098 case IORING_OP_RECVMSG:
6099 case IORING_OP_RECV:
6100 kfree(req->sr_msg.kbuf);
6105 if (req->flags & REQ_F_NEED_CLEANUP) {
6106 switch (req->opcode) {
6107 case IORING_OP_READV:
6108 case IORING_OP_READ_FIXED:
6109 case IORING_OP_READ:
6110 case IORING_OP_WRITEV:
6111 case IORING_OP_WRITE_FIXED:
6112 case IORING_OP_WRITE: {
6113 struct io_async_rw *io = req->async_data;
6115 kfree(io->free_iovec);
6118 case IORING_OP_RECVMSG:
6119 case IORING_OP_SENDMSG: {
6120 struct io_async_msghdr *io = req->async_data;
6122 kfree(io->free_iov);
6125 case IORING_OP_SPLICE:
6127 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6128 io_put_file(req->splice.file_in);
6130 case IORING_OP_OPENAT:
6131 case IORING_OP_OPENAT2:
6132 if (req->open.filename)
6133 putname(req->open.filename);
6135 case IORING_OP_RENAMEAT:
6136 putname(req->rename.oldpath);
6137 putname(req->rename.newpath);
6139 case IORING_OP_UNLINKAT:
6140 putname(req->unlink.filename);
6144 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6145 kfree(req->apoll->double_poll);
6149 if (req->flags & REQ_F_INFLIGHT) {
6150 struct io_uring_task *tctx = req->task->io_uring;
6152 atomic_dec(&tctx->inflight_tracked);
6154 if (req->flags & REQ_F_CREDS)
6155 put_cred(req->creds);
6157 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6160 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6162 struct io_ring_ctx *ctx = req->ctx;
6163 const struct cred *creds = NULL;
6166 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6167 creds = override_creds(req->creds);
6169 switch (req->opcode) {
6171 ret = io_nop(req, issue_flags);
6173 case IORING_OP_READV:
6174 case IORING_OP_READ_FIXED:
6175 case IORING_OP_READ:
6176 ret = io_read(req, issue_flags);
6178 case IORING_OP_WRITEV:
6179 case IORING_OP_WRITE_FIXED:
6180 case IORING_OP_WRITE:
6181 ret = io_write(req, issue_flags);
6183 case IORING_OP_FSYNC:
6184 ret = io_fsync(req, issue_flags);
6186 case IORING_OP_POLL_ADD:
6187 ret = io_poll_add(req, issue_flags);
6189 case IORING_OP_POLL_REMOVE:
6190 ret = io_poll_update(req, issue_flags);
6192 case IORING_OP_SYNC_FILE_RANGE:
6193 ret = io_sync_file_range(req, issue_flags);
6195 case IORING_OP_SENDMSG:
6196 ret = io_sendmsg(req, issue_flags);
6198 case IORING_OP_SEND:
6199 ret = io_send(req, issue_flags);
6201 case IORING_OP_RECVMSG:
6202 ret = io_recvmsg(req, issue_flags);
6204 case IORING_OP_RECV:
6205 ret = io_recv(req, issue_flags);
6207 case IORING_OP_TIMEOUT:
6208 ret = io_timeout(req, issue_flags);
6210 case IORING_OP_TIMEOUT_REMOVE:
6211 ret = io_timeout_remove(req, issue_flags);
6213 case IORING_OP_ACCEPT:
6214 ret = io_accept(req, issue_flags);
6216 case IORING_OP_CONNECT:
6217 ret = io_connect(req, issue_flags);
6219 case IORING_OP_ASYNC_CANCEL:
6220 ret = io_async_cancel(req, issue_flags);
6222 case IORING_OP_FALLOCATE:
6223 ret = io_fallocate(req, issue_flags);
6225 case IORING_OP_OPENAT:
6226 ret = io_openat(req, issue_flags);
6228 case IORING_OP_CLOSE:
6229 ret = io_close(req, issue_flags);
6231 case IORING_OP_FILES_UPDATE:
6232 ret = io_files_update(req, issue_flags);
6234 case IORING_OP_STATX:
6235 ret = io_statx(req, issue_flags);
6237 case IORING_OP_FADVISE:
6238 ret = io_fadvise(req, issue_flags);
6240 case IORING_OP_MADVISE:
6241 ret = io_madvise(req, issue_flags);
6243 case IORING_OP_OPENAT2:
6244 ret = io_openat2(req, issue_flags);
6246 case IORING_OP_EPOLL_CTL:
6247 ret = io_epoll_ctl(req, issue_flags);
6249 case IORING_OP_SPLICE:
6250 ret = io_splice(req, issue_flags);
6252 case IORING_OP_PROVIDE_BUFFERS:
6253 ret = io_provide_buffers(req, issue_flags);
6255 case IORING_OP_REMOVE_BUFFERS:
6256 ret = io_remove_buffers(req, issue_flags);
6259 ret = io_tee(req, issue_flags);
6261 case IORING_OP_SHUTDOWN:
6262 ret = io_shutdown(req, issue_flags);
6264 case IORING_OP_RENAMEAT:
6265 ret = io_renameat(req, issue_flags);
6267 case IORING_OP_UNLINKAT:
6268 ret = io_unlinkat(req, issue_flags);
6276 revert_creds(creds);
6279 /* If the op doesn't have a file, we're not polling for it */
6280 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6281 io_iopoll_req_issued(req);
6286 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6288 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6290 req = io_put_req_find_next(req);
6291 return req ? &req->work : NULL;
6294 static void io_wq_submit_work(struct io_wq_work *work)
6296 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6297 struct io_kiocb *timeout;
6300 timeout = io_prep_linked_timeout(req);
6302 io_queue_linked_timeout(timeout);
6304 if (work->flags & IO_WQ_WORK_CANCEL)
6309 ret = io_issue_sqe(req, 0);
6311 * We can get EAGAIN for polled IO even though we're
6312 * forcing a sync submission from here, since we can't
6313 * wait for request slots on the block side.
6321 /* avoid locking problems by failing it from a clean context */
6323 /* io-wq is going to take one down */
6325 io_req_task_queue_fail(req, ret);
6329 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6332 return &table->files[i];
6335 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6338 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6340 return (struct file *) (slot->file_ptr & FFS_MASK);
6343 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6345 unsigned long file_ptr = (unsigned long) file;
6347 if (__io_file_supports_nowait(file, READ))
6348 file_ptr |= FFS_ASYNC_READ;
6349 if (__io_file_supports_nowait(file, WRITE))
6350 file_ptr |= FFS_ASYNC_WRITE;
6351 if (S_ISREG(file_inode(file)->i_mode))
6352 file_ptr |= FFS_ISREG;
6353 file_slot->file_ptr = file_ptr;
6356 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6357 struct io_kiocb *req, int fd)
6360 unsigned long file_ptr;
6362 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6364 fd = array_index_nospec(fd, ctx->nr_user_files);
6365 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6366 file = (struct file *) (file_ptr & FFS_MASK);
6367 file_ptr &= ~FFS_MASK;
6368 /* mask in overlapping REQ_F and FFS bits */
6369 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6370 io_req_set_rsrc_node(req);
6374 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6375 struct io_submit_state *state,
6376 struct io_kiocb *req, int fd)
6378 struct file *file = __io_file_get(state, fd);
6380 trace_io_uring_file_get(ctx, fd);
6382 /* we don't allow fixed io_uring files */
6383 if (file && unlikely(file->f_op == &io_uring_fops))
6384 io_req_track_inflight(req);
6388 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6389 struct io_submit_state *state,
6390 struct io_kiocb *req, int fd, bool fixed)
6393 return io_file_get_fixed(ctx, req, fd);
6395 return io_file_get_normal(ctx, state, req, fd);
6398 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6400 struct io_timeout_data *data = container_of(timer,
6401 struct io_timeout_data, timer);
6402 struct io_kiocb *prev, *req = data->req;
6403 struct io_ring_ctx *ctx = req->ctx;
6404 unsigned long flags;
6406 spin_lock_irqsave(&ctx->completion_lock, flags);
6407 prev = req->timeout.head;
6408 req->timeout.head = NULL;
6411 * We don't expect the list to be empty, that will only happen if we
6412 * race with the completion of the linked work.
6415 io_remove_next_linked(prev);
6416 if (!req_ref_inc_not_zero(prev))
6419 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6422 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6423 io_put_req_deferred(prev, 1);
6424 io_put_req_deferred(req, 1);
6426 io_req_complete_post(req, -ETIME, 0);
6428 return HRTIMER_NORESTART;
6431 static void io_queue_linked_timeout(struct io_kiocb *req)
6433 struct io_ring_ctx *ctx = req->ctx;
6435 spin_lock_irq(&ctx->completion_lock);
6437 * If the back reference is NULL, then our linked request finished
6438 * before we got a chance to setup the timer
6440 if (req->timeout.head) {
6441 struct io_timeout_data *data = req->async_data;
6443 data->timer.function = io_link_timeout_fn;
6444 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6447 spin_unlock_irq(&ctx->completion_lock);
6448 /* drop submission reference */
6452 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6454 struct io_kiocb *nxt = req->link;
6456 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6457 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6460 nxt->timeout.head = req;
6461 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6462 req->flags |= REQ_F_LINK_TIMEOUT;
6466 static void __io_queue_sqe(struct io_kiocb *req)
6467 __must_hold(&req->ctx->uring_lock)
6469 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6473 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6476 * We async punt it if the file wasn't marked NOWAIT, or if the file
6477 * doesn't support non-blocking read/write attempts
6480 /* drop submission reference */
6481 if (req->flags & REQ_F_COMPLETE_INLINE) {
6482 struct io_ring_ctx *ctx = req->ctx;
6483 struct io_comp_state *cs = &ctx->submit_state.comp;
6485 cs->reqs[cs->nr++] = req;
6486 if (cs->nr == ARRAY_SIZE(cs->reqs))
6487 io_submit_flush_completions(ctx);
6491 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6492 switch (io_arm_poll_handler(req)) {
6493 case IO_APOLL_READY:
6495 case IO_APOLL_ABORTED:
6497 * Queued up for async execution, worker will release
6498 * submit reference when the iocb is actually submitted.
6500 io_queue_async_work(req);
6504 io_req_complete_failed(req, ret);
6507 io_queue_linked_timeout(linked_timeout);
6510 static inline void io_queue_sqe(struct io_kiocb *req)
6511 __must_hold(&req->ctx->uring_lock)
6513 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6516 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6517 __io_queue_sqe(req);
6519 int ret = io_req_prep_async(req);
6522 io_req_complete_failed(req, ret);
6524 io_queue_async_work(req);
6529 * Check SQE restrictions (opcode and flags).
6531 * Returns 'true' if SQE is allowed, 'false' otherwise.
6533 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6534 struct io_kiocb *req,
6535 unsigned int sqe_flags)
6537 if (likely(!ctx->restricted))
6540 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6543 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6544 ctx->restrictions.sqe_flags_required)
6547 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6548 ctx->restrictions.sqe_flags_required))
6554 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6555 const struct io_uring_sqe *sqe)
6556 __must_hold(&ctx->uring_lock)
6558 struct io_submit_state *state;
6559 unsigned int sqe_flags;
6560 int personality, ret = 0;
6562 /* req is partially pre-initialised, see io_preinit_req() */
6563 req->opcode = READ_ONCE(sqe->opcode);
6564 /* same numerical values with corresponding REQ_F_*, safe to copy */
6565 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6566 req->user_data = READ_ONCE(sqe->user_data);
6568 req->fixed_rsrc_refs = NULL;
6569 /* one is dropped after submission, the other at completion */
6570 atomic_set(&req->refs, 2);
6571 req->task = current;
6573 /* enforce forwards compatibility on users */
6574 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6576 if (unlikely(req->opcode >= IORING_OP_LAST))
6578 if (!io_check_restriction(ctx, req, sqe_flags))
6581 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6582 !io_op_defs[req->opcode].buffer_select)
6584 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6585 ctx->drain_active = true;
6587 personality = READ_ONCE(sqe->personality);
6589 req->creds = xa_load(&ctx->personalities, personality);
6592 get_cred(req->creds);
6593 req->flags |= REQ_F_CREDS;
6595 state = &ctx->submit_state;
6598 * Plug now if we have more than 1 IO left after this, and the target
6599 * is potentially a read/write to block based storage.
6601 if (!state->plug_started && state->ios_left > 1 &&
6602 io_op_defs[req->opcode].plug) {
6603 blk_start_plug(&state->plug);
6604 state->plug_started = true;
6607 if (io_op_defs[req->opcode].needs_file) {
6608 req->file = io_file_get(ctx, state, req, READ_ONCE(sqe->fd),
6609 (sqe_flags & IOSQE_FIXED_FILE));
6610 if (unlikely(!req->file))
6618 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6619 const struct io_uring_sqe *sqe)
6620 __must_hold(&ctx->uring_lock)
6622 struct io_submit_link *link = &ctx->submit_state.link;
6625 ret = io_init_req(ctx, req, sqe);
6626 if (unlikely(ret)) {
6629 /* fail even hard links since we don't submit */
6630 req_set_fail(link->head);
6631 io_req_complete_failed(link->head, -ECANCELED);
6634 io_req_complete_failed(req, ret);
6638 ret = io_req_prep(req, sqe);
6642 /* don't need @sqe from now on */
6643 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6645 ctx->flags & IORING_SETUP_SQPOLL);
6648 * If we already have a head request, queue this one for async
6649 * submittal once the head completes. If we don't have a head but
6650 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6651 * submitted sync once the chain is complete. If none of those
6652 * conditions are true (normal request), then just queue it.
6655 struct io_kiocb *head = link->head;
6657 ret = io_req_prep_async(req);
6660 trace_io_uring_link(ctx, req, head);
6661 link->last->link = req;
6664 /* last request of a link, enqueue the link */
6665 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6670 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6682 * Batched submission is done, ensure local IO is flushed out.
6684 static void io_submit_state_end(struct io_submit_state *state,
6685 struct io_ring_ctx *ctx)
6687 if (state->link.head)
6688 io_queue_sqe(state->link.head);
6690 io_submit_flush_completions(ctx);
6691 if (state->plug_started)
6692 blk_finish_plug(&state->plug);
6693 io_state_file_put(state);
6697 * Start submission side cache.
6699 static void io_submit_state_start(struct io_submit_state *state,
6700 unsigned int max_ios)
6702 state->plug_started = false;
6703 state->ios_left = max_ios;
6704 /* set only head, no need to init link_last in advance */
6705 state->link.head = NULL;
6708 static void io_commit_sqring(struct io_ring_ctx *ctx)
6710 struct io_rings *rings = ctx->rings;
6713 * Ensure any loads from the SQEs are done at this point,
6714 * since once we write the new head, the application could
6715 * write new data to them.
6717 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6721 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6722 * that is mapped by userspace. This means that care needs to be taken to
6723 * ensure that reads are stable, as we cannot rely on userspace always
6724 * being a good citizen. If members of the sqe are validated and then later
6725 * used, it's important that those reads are done through READ_ONCE() to
6726 * prevent a re-load down the line.
6728 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6730 unsigned head, mask = ctx->sq_entries - 1;
6731 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6734 * The cached sq head (or cq tail) serves two purposes:
6736 * 1) allows us to batch the cost of updating the user visible
6738 * 2) allows the kernel side to track the head on its own, even
6739 * though the application is the one updating it.
6741 head = READ_ONCE(ctx->sq_array[sq_idx]);
6742 if (likely(head < ctx->sq_entries))
6743 return &ctx->sq_sqes[head];
6745 /* drop invalid entries */
6747 WRITE_ONCE(ctx->rings->sq_dropped,
6748 READ_ONCE(ctx->rings->sq_dropped) + 1);
6752 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6753 __must_hold(&ctx->uring_lock)
6755 struct io_uring_task *tctx;
6758 /* make sure SQ entry isn't read before tail */
6759 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6760 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6763 tctx = current->io_uring;
6764 tctx->cached_refs -= nr;
6765 if (unlikely(tctx->cached_refs < 0)) {
6766 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6768 percpu_counter_add(&tctx->inflight, refill);
6769 refcount_add(refill, ¤t->usage);
6770 tctx->cached_refs += refill;
6772 io_submit_state_start(&ctx->submit_state, nr);
6774 while (submitted < nr) {
6775 const struct io_uring_sqe *sqe;
6776 struct io_kiocb *req;
6778 req = io_alloc_req(ctx);
6779 if (unlikely(!req)) {
6781 submitted = -EAGAIN;
6784 sqe = io_get_sqe(ctx);
6785 if (unlikely(!sqe)) {
6786 kmem_cache_free(req_cachep, req);
6789 /* will complete beyond this point, count as submitted */
6791 if (io_submit_sqe(ctx, req, sqe))
6795 if (unlikely(submitted != nr)) {
6796 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6797 int unused = nr - ref_used;
6799 current->io_uring->cached_refs += unused;
6800 percpu_ref_put_many(&ctx->refs, unused);
6803 io_submit_state_end(&ctx->submit_state, ctx);
6804 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6805 io_commit_sqring(ctx);
6810 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6812 return READ_ONCE(sqd->state);
6815 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6817 /* Tell userspace we may need a wakeup call */
6818 spin_lock_irq(&ctx->completion_lock);
6819 WRITE_ONCE(ctx->rings->sq_flags,
6820 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
6821 spin_unlock_irq(&ctx->completion_lock);
6824 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6826 spin_lock_irq(&ctx->completion_lock);
6827 WRITE_ONCE(ctx->rings->sq_flags,
6828 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
6829 spin_unlock_irq(&ctx->completion_lock);
6832 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6834 unsigned int to_submit;
6837 to_submit = io_sqring_entries(ctx);
6838 /* if we're handling multiple rings, cap submit size for fairness */
6839 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6840 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6842 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6843 unsigned nr_events = 0;
6844 const struct cred *creds = NULL;
6846 if (ctx->sq_creds != current_cred())
6847 creds = override_creds(ctx->sq_creds);
6849 mutex_lock(&ctx->uring_lock);
6850 if (!list_empty(&ctx->iopoll_list))
6851 io_do_iopoll(ctx, &nr_events, 0, true);
6854 * Don't submit if refs are dying, good for io_uring_register(),
6855 * but also it is relied upon by io_ring_exit_work()
6857 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6858 !(ctx->flags & IORING_SETUP_R_DISABLED))
6859 ret = io_submit_sqes(ctx, to_submit);
6860 mutex_unlock(&ctx->uring_lock);
6862 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6863 wake_up(&ctx->sqo_sq_wait);
6865 revert_creds(creds);
6871 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6873 struct io_ring_ctx *ctx;
6874 unsigned sq_thread_idle = 0;
6876 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6877 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6878 sqd->sq_thread_idle = sq_thread_idle;
6881 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6883 bool did_sig = false;
6884 struct ksignal ksig;
6886 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6887 signal_pending(current)) {
6888 mutex_unlock(&sqd->lock);
6889 if (signal_pending(current))
6890 did_sig = get_signal(&ksig);
6892 mutex_lock(&sqd->lock);
6894 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6897 static int io_sq_thread(void *data)
6899 struct io_sq_data *sqd = data;
6900 struct io_ring_ctx *ctx;
6901 unsigned long timeout = 0;
6902 char buf[TASK_COMM_LEN];
6905 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6906 set_task_comm(current, buf);
6908 if (sqd->sq_cpu != -1)
6909 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6911 set_cpus_allowed_ptr(current, cpu_online_mask);
6912 current->flags |= PF_NO_SETAFFINITY;
6914 mutex_lock(&sqd->lock);
6916 bool cap_entries, sqt_spin = false;
6918 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6919 if (io_sqd_handle_event(sqd))
6921 timeout = jiffies + sqd->sq_thread_idle;
6924 cap_entries = !list_is_singular(&sqd->ctx_list);
6925 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6926 int ret = __io_sq_thread(ctx, cap_entries);
6928 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6931 if (io_run_task_work())
6934 if (sqt_spin || !time_after(jiffies, timeout)) {
6937 timeout = jiffies + sqd->sq_thread_idle;
6941 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6942 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6943 bool needs_sched = true;
6945 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6946 io_ring_set_wakeup_flag(ctx);
6948 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6949 !list_empty_careful(&ctx->iopoll_list)) {
6950 needs_sched = false;
6953 if (io_sqring_entries(ctx)) {
6954 needs_sched = false;
6960 mutex_unlock(&sqd->lock);
6962 mutex_lock(&sqd->lock);
6964 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6965 io_ring_clear_wakeup_flag(ctx);
6968 finish_wait(&sqd->wait, &wait);
6969 timeout = jiffies + sqd->sq_thread_idle;
6972 io_uring_cancel_generic(true, sqd);
6974 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6975 io_ring_set_wakeup_flag(ctx);
6977 mutex_unlock(&sqd->lock);
6979 complete(&sqd->exited);
6983 struct io_wait_queue {
6984 struct wait_queue_entry wq;
6985 struct io_ring_ctx *ctx;
6987 unsigned nr_timeouts;
6990 static inline bool io_should_wake(struct io_wait_queue *iowq)
6992 struct io_ring_ctx *ctx = iowq->ctx;
6993 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
6996 * Wake up if we have enough events, or if a timeout occurred since we
6997 * started waiting. For timeouts, we always want to return to userspace,
6998 * regardless of event count.
7000 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7003 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7004 int wake_flags, void *key)
7006 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7010 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7011 * the task, and the next invocation will do it.
7013 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7014 return autoremove_wake_function(curr, mode, wake_flags, key);
7018 static int io_run_task_work_sig(void)
7020 if (io_run_task_work())
7022 if (!signal_pending(current))
7024 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7025 return -ERESTARTSYS;
7029 /* when returns >0, the caller should retry */
7030 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7031 struct io_wait_queue *iowq,
7032 signed long *timeout)
7036 /* make sure we run task_work before checking for signals */
7037 ret = io_run_task_work_sig();
7038 if (ret || io_should_wake(iowq))
7040 /* let the caller flush overflows, retry */
7041 if (test_bit(0, &ctx->check_cq_overflow))
7044 *timeout = schedule_timeout(*timeout);
7045 return !*timeout ? -ETIME : 1;
7049 * Wait until events become available, if we don't already have some. The
7050 * application must reap them itself, as they reside on the shared cq ring.
7052 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7053 const sigset_t __user *sig, size_t sigsz,
7054 struct __kernel_timespec __user *uts)
7056 struct io_wait_queue iowq;
7057 struct io_rings *rings = ctx->rings;
7058 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7062 io_cqring_overflow_flush(ctx, false);
7063 if (io_cqring_events(ctx) >= min_events)
7065 if (!io_run_task_work())
7070 #ifdef CONFIG_COMPAT
7071 if (in_compat_syscall())
7072 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7076 ret = set_user_sigmask(sig, sigsz);
7083 struct timespec64 ts;
7085 if (get_timespec64(&ts, uts))
7087 timeout = timespec64_to_jiffies(&ts);
7090 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7091 iowq.wq.private = current;
7092 INIT_LIST_HEAD(&iowq.wq.entry);
7094 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7095 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7097 trace_io_uring_cqring_wait(ctx, min_events);
7099 /* if we can't even flush overflow, don't wait for more */
7100 if (!io_cqring_overflow_flush(ctx, false)) {
7104 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7105 TASK_INTERRUPTIBLE);
7106 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7107 finish_wait(&ctx->cq_wait, &iowq.wq);
7111 restore_saved_sigmask_unless(ret == -EINTR);
7113 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7116 static void io_free_page_table(void **table, size_t size)
7118 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7120 for (i = 0; i < nr_tables; i++)
7125 static void **io_alloc_page_table(size_t size)
7127 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7128 size_t init_size = size;
7131 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7135 for (i = 0; i < nr_tables; i++) {
7136 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7138 table[i] = kzalloc(this_size, GFP_KERNEL);
7140 io_free_page_table(table, init_size);
7148 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7150 percpu_ref_exit(&ref_node->refs);
7154 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7156 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7157 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7158 unsigned long flags;
7159 bool first_add = false;
7161 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7164 while (!list_empty(&ctx->rsrc_ref_list)) {
7165 node = list_first_entry(&ctx->rsrc_ref_list,
7166 struct io_rsrc_node, node);
7167 /* recycle ref nodes in order */
7170 list_del(&node->node);
7171 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7173 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7176 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7179 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7181 struct io_rsrc_node *ref_node;
7183 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7187 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7192 INIT_LIST_HEAD(&ref_node->node);
7193 INIT_LIST_HEAD(&ref_node->rsrc_list);
7194 ref_node->done = false;
7198 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7199 struct io_rsrc_data *data_to_kill)
7201 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7202 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7205 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7207 rsrc_node->rsrc_data = data_to_kill;
7208 spin_lock_irq(&ctx->rsrc_ref_lock);
7209 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7210 spin_unlock_irq(&ctx->rsrc_ref_lock);
7212 atomic_inc(&data_to_kill->refs);
7213 percpu_ref_kill(&rsrc_node->refs);
7214 ctx->rsrc_node = NULL;
7217 if (!ctx->rsrc_node) {
7218 ctx->rsrc_node = ctx->rsrc_backup_node;
7219 ctx->rsrc_backup_node = NULL;
7223 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7225 if (ctx->rsrc_backup_node)
7227 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7228 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7231 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7235 /* As we may drop ->uring_lock, other task may have started quiesce */
7239 data->quiesce = true;
7241 ret = io_rsrc_node_switch_start(ctx);
7244 io_rsrc_node_switch(ctx, data);
7246 /* kill initial ref, already quiesced if zero */
7247 if (atomic_dec_and_test(&data->refs))
7249 mutex_unlock(&ctx->uring_lock);
7250 flush_delayed_work(&ctx->rsrc_put_work);
7251 ret = wait_for_completion_interruptible(&data->done);
7253 mutex_lock(&ctx->uring_lock);
7257 atomic_inc(&data->refs);
7258 /* wait for all works potentially completing data->done */
7259 flush_delayed_work(&ctx->rsrc_put_work);
7260 reinit_completion(&data->done);
7262 ret = io_run_task_work_sig();
7263 mutex_lock(&ctx->uring_lock);
7265 data->quiesce = false;
7270 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7272 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7273 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7275 return &data->tags[table_idx][off];
7278 static void io_rsrc_data_free(struct io_rsrc_data *data)
7280 size_t size = data->nr * sizeof(data->tags[0][0]);
7283 io_free_page_table((void **)data->tags, size);
7287 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7288 u64 __user *utags, unsigned nr,
7289 struct io_rsrc_data **pdata)
7291 struct io_rsrc_data *data;
7295 data = kzalloc(sizeof(*data), GFP_KERNEL);
7298 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7306 data->do_put = do_put;
7309 for (i = 0; i < nr; i++) {
7310 u64 *tag_slot = io_get_tag_slot(data, i);
7312 if (copy_from_user(tag_slot, &utags[i],
7318 atomic_set(&data->refs, 1);
7319 init_completion(&data->done);
7323 io_rsrc_data_free(data);
7327 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7329 table->files = kvcalloc(nr_files, sizeof(table->files[0]), GFP_KERNEL);
7330 return !!table->files;
7333 static void io_free_file_tables(struct io_file_table *table)
7335 kvfree(table->files);
7336 table->files = NULL;
7339 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7341 #if defined(CONFIG_UNIX)
7342 if (ctx->ring_sock) {
7343 struct sock *sock = ctx->ring_sock->sk;
7344 struct sk_buff *skb;
7346 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7352 for (i = 0; i < ctx->nr_user_files; i++) {
7355 file = io_file_from_index(ctx, i);
7360 io_free_file_tables(&ctx->file_table);
7361 io_rsrc_data_free(ctx->file_data);
7362 ctx->file_data = NULL;
7363 ctx->nr_user_files = 0;
7366 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7370 if (!ctx->file_data)
7372 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7374 __io_sqe_files_unregister(ctx);
7378 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7379 __releases(&sqd->lock)
7381 WARN_ON_ONCE(sqd->thread == current);
7384 * Do the dance but not conditional clear_bit() because it'd race with
7385 * other threads incrementing park_pending and setting the bit.
7387 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7388 if (atomic_dec_return(&sqd->park_pending))
7389 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7390 mutex_unlock(&sqd->lock);
7393 static void io_sq_thread_park(struct io_sq_data *sqd)
7394 __acquires(&sqd->lock)
7396 WARN_ON_ONCE(sqd->thread == current);
7398 atomic_inc(&sqd->park_pending);
7399 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7400 mutex_lock(&sqd->lock);
7402 wake_up_process(sqd->thread);
7405 static void io_sq_thread_stop(struct io_sq_data *sqd)
7407 WARN_ON_ONCE(sqd->thread == current);
7408 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7410 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7411 mutex_lock(&sqd->lock);
7413 wake_up_process(sqd->thread);
7414 mutex_unlock(&sqd->lock);
7415 wait_for_completion(&sqd->exited);
7418 static void io_put_sq_data(struct io_sq_data *sqd)
7420 if (refcount_dec_and_test(&sqd->refs)) {
7421 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7423 io_sq_thread_stop(sqd);
7428 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7430 struct io_sq_data *sqd = ctx->sq_data;
7433 io_sq_thread_park(sqd);
7434 list_del_init(&ctx->sqd_list);
7435 io_sqd_update_thread_idle(sqd);
7436 io_sq_thread_unpark(sqd);
7438 io_put_sq_data(sqd);
7439 ctx->sq_data = NULL;
7443 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7445 struct io_ring_ctx *ctx_attach;
7446 struct io_sq_data *sqd;
7449 f = fdget(p->wq_fd);
7451 return ERR_PTR(-ENXIO);
7452 if (f.file->f_op != &io_uring_fops) {
7454 return ERR_PTR(-EINVAL);
7457 ctx_attach = f.file->private_data;
7458 sqd = ctx_attach->sq_data;
7461 return ERR_PTR(-EINVAL);
7463 if (sqd->task_tgid != current->tgid) {
7465 return ERR_PTR(-EPERM);
7468 refcount_inc(&sqd->refs);
7473 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7476 struct io_sq_data *sqd;
7479 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7480 sqd = io_attach_sq_data(p);
7485 /* fall through for EPERM case, setup new sqd/task */
7486 if (PTR_ERR(sqd) != -EPERM)
7490 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7492 return ERR_PTR(-ENOMEM);
7494 atomic_set(&sqd->park_pending, 0);
7495 refcount_set(&sqd->refs, 1);
7496 INIT_LIST_HEAD(&sqd->ctx_list);
7497 mutex_init(&sqd->lock);
7498 init_waitqueue_head(&sqd->wait);
7499 init_completion(&sqd->exited);
7503 #if defined(CONFIG_UNIX)
7505 * Ensure the UNIX gc is aware of our file set, so we are certain that
7506 * the io_uring can be safely unregistered on process exit, even if we have
7507 * loops in the file referencing.
7509 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7511 struct sock *sk = ctx->ring_sock->sk;
7512 struct scm_fp_list *fpl;
7513 struct sk_buff *skb;
7516 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7520 skb = alloc_skb(0, GFP_KERNEL);
7529 fpl->user = get_uid(current_user());
7530 for (i = 0; i < nr; i++) {
7531 struct file *file = io_file_from_index(ctx, i + offset);
7535 fpl->fp[nr_files] = get_file(file);
7536 unix_inflight(fpl->user, fpl->fp[nr_files]);
7541 fpl->max = SCM_MAX_FD;
7542 fpl->count = nr_files;
7543 UNIXCB(skb).fp = fpl;
7544 skb->destructor = unix_destruct_scm;
7545 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7546 skb_queue_head(&sk->sk_receive_queue, skb);
7548 for (i = 0; i < nr_files; i++)
7559 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7560 * causes regular reference counting to break down. We rely on the UNIX
7561 * garbage collection to take care of this problem for us.
7563 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7565 unsigned left, total;
7569 left = ctx->nr_user_files;
7571 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7573 ret = __io_sqe_files_scm(ctx, this_files, total);
7577 total += this_files;
7583 while (total < ctx->nr_user_files) {
7584 struct file *file = io_file_from_index(ctx, total);
7594 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7600 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7602 struct file *file = prsrc->file;
7603 #if defined(CONFIG_UNIX)
7604 struct sock *sock = ctx->ring_sock->sk;
7605 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7606 struct sk_buff *skb;
7609 __skb_queue_head_init(&list);
7612 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7613 * remove this entry and rearrange the file array.
7615 skb = skb_dequeue(head);
7617 struct scm_fp_list *fp;
7619 fp = UNIXCB(skb).fp;
7620 for (i = 0; i < fp->count; i++) {
7623 if (fp->fp[i] != file)
7626 unix_notinflight(fp->user, fp->fp[i]);
7627 left = fp->count - 1 - i;
7629 memmove(&fp->fp[i], &fp->fp[i + 1],
7630 left * sizeof(struct file *));
7637 __skb_queue_tail(&list, skb);
7647 __skb_queue_tail(&list, skb);
7649 skb = skb_dequeue(head);
7652 if (skb_peek(&list)) {
7653 spin_lock_irq(&head->lock);
7654 while ((skb = __skb_dequeue(&list)) != NULL)
7655 __skb_queue_tail(head, skb);
7656 spin_unlock_irq(&head->lock);
7663 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7665 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7666 struct io_ring_ctx *ctx = rsrc_data->ctx;
7667 struct io_rsrc_put *prsrc, *tmp;
7669 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7670 list_del(&prsrc->list);
7673 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7675 io_ring_submit_lock(ctx, lock_ring);
7676 spin_lock_irq(&ctx->completion_lock);
7677 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7679 io_commit_cqring(ctx);
7680 spin_unlock_irq(&ctx->completion_lock);
7681 io_cqring_ev_posted(ctx);
7682 io_ring_submit_unlock(ctx, lock_ring);
7685 rsrc_data->do_put(ctx, prsrc);
7689 io_rsrc_node_destroy(ref_node);
7690 if (atomic_dec_and_test(&rsrc_data->refs))
7691 complete(&rsrc_data->done);
7694 static void io_rsrc_put_work(struct work_struct *work)
7696 struct io_ring_ctx *ctx;
7697 struct llist_node *node;
7699 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7700 node = llist_del_all(&ctx->rsrc_put_llist);
7703 struct io_rsrc_node *ref_node;
7704 struct llist_node *next = node->next;
7706 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7707 __io_rsrc_put_work(ref_node);
7712 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7713 unsigned nr_args, u64 __user *tags)
7715 __s32 __user *fds = (__s32 __user *) arg;
7724 if (nr_args > IORING_MAX_FIXED_FILES)
7726 ret = io_rsrc_node_switch_start(ctx);
7729 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7735 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7738 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7739 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7743 /* allow sparse sets */
7746 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7753 if (unlikely(!file))
7757 * Don't allow io_uring instances to be registered. If UNIX
7758 * isn't enabled, then this causes a reference cycle and this
7759 * instance can never get freed. If UNIX is enabled we'll
7760 * handle it just fine, but there's still no point in allowing
7761 * a ring fd as it doesn't support regular read/write anyway.
7763 if (file->f_op == &io_uring_fops) {
7767 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7770 ret = io_sqe_files_scm(ctx);
7772 __io_sqe_files_unregister(ctx);
7776 io_rsrc_node_switch(ctx, NULL);
7779 for (i = 0; i < ctx->nr_user_files; i++) {
7780 file = io_file_from_index(ctx, i);
7784 io_free_file_tables(&ctx->file_table);
7785 ctx->nr_user_files = 0;
7787 io_rsrc_data_free(ctx->file_data);
7788 ctx->file_data = NULL;
7792 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7795 #if defined(CONFIG_UNIX)
7796 struct sock *sock = ctx->ring_sock->sk;
7797 struct sk_buff_head *head = &sock->sk_receive_queue;
7798 struct sk_buff *skb;
7801 * See if we can merge this file into an existing skb SCM_RIGHTS
7802 * file set. If there's no room, fall back to allocating a new skb
7803 * and filling it in.
7805 spin_lock_irq(&head->lock);
7806 skb = skb_peek(head);
7808 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7810 if (fpl->count < SCM_MAX_FD) {
7811 __skb_unlink(skb, head);
7812 spin_unlock_irq(&head->lock);
7813 fpl->fp[fpl->count] = get_file(file);
7814 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7816 spin_lock_irq(&head->lock);
7817 __skb_queue_head(head, skb);
7822 spin_unlock_irq(&head->lock);
7829 return __io_sqe_files_scm(ctx, 1, index);
7835 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7836 struct io_rsrc_node *node, void *rsrc)
7838 struct io_rsrc_put *prsrc;
7840 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7844 prsrc->tag = *io_get_tag_slot(data, idx);
7846 list_add(&prsrc->list, &node->rsrc_list);
7850 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7851 struct io_uring_rsrc_update2 *up,
7854 u64 __user *tags = u64_to_user_ptr(up->tags);
7855 __s32 __user *fds = u64_to_user_ptr(up->data);
7856 struct io_rsrc_data *data = ctx->file_data;
7857 struct io_fixed_file *file_slot;
7861 bool needs_switch = false;
7863 if (!ctx->file_data)
7865 if (up->offset + nr_args > ctx->nr_user_files)
7868 for (done = 0; done < nr_args; done++) {
7871 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7872 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7876 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7880 if (fd == IORING_REGISTER_FILES_SKIP)
7883 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7884 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7886 if (file_slot->file_ptr) {
7887 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7888 err = io_queue_rsrc_removal(data, up->offset + done,
7889 ctx->rsrc_node, file);
7892 file_slot->file_ptr = 0;
7893 needs_switch = true;
7902 * Don't allow io_uring instances to be registered. If
7903 * UNIX isn't enabled, then this causes a reference
7904 * cycle and this instance can never get freed. If UNIX
7905 * is enabled we'll handle it just fine, but there's
7906 * still no point in allowing a ring fd as it doesn't
7907 * support regular read/write anyway.
7909 if (file->f_op == &io_uring_fops) {
7914 *io_get_tag_slot(data, up->offset + done) = tag;
7915 io_fixed_file_set(file_slot, file);
7916 err = io_sqe_file_register(ctx, file, i);
7918 file_slot->file_ptr = 0;
7926 io_rsrc_node_switch(ctx, data);
7927 return done ? done : err;
7930 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7931 struct task_struct *task)
7933 struct io_wq_hash *hash;
7934 struct io_wq_data data;
7935 unsigned int concurrency;
7937 mutex_lock(&ctx->uring_lock);
7938 hash = ctx->hash_map;
7940 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7942 mutex_unlock(&ctx->uring_lock);
7943 return ERR_PTR(-ENOMEM);
7945 refcount_set(&hash->refs, 1);
7946 init_waitqueue_head(&hash->wait);
7947 ctx->hash_map = hash;
7949 mutex_unlock(&ctx->uring_lock);
7953 data.free_work = io_wq_free_work;
7954 data.do_work = io_wq_submit_work;
7956 /* Do QD, or 4 * CPUS, whatever is smallest */
7957 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7959 return io_wq_create(concurrency, &data);
7962 static int io_uring_alloc_task_context(struct task_struct *task,
7963 struct io_ring_ctx *ctx)
7965 struct io_uring_task *tctx;
7968 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7969 if (unlikely(!tctx))
7972 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7973 if (unlikely(ret)) {
7978 tctx->io_wq = io_init_wq_offload(ctx, task);
7979 if (IS_ERR(tctx->io_wq)) {
7980 ret = PTR_ERR(tctx->io_wq);
7981 percpu_counter_destroy(&tctx->inflight);
7987 init_waitqueue_head(&tctx->wait);
7988 atomic_set(&tctx->in_idle, 0);
7989 atomic_set(&tctx->inflight_tracked, 0);
7990 task->io_uring = tctx;
7991 spin_lock_init(&tctx->task_lock);
7992 INIT_WQ_LIST(&tctx->task_list);
7993 init_task_work(&tctx->task_work, tctx_task_work);
7997 void __io_uring_free(struct task_struct *tsk)
7999 struct io_uring_task *tctx = tsk->io_uring;
8001 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8002 WARN_ON_ONCE(tctx->io_wq);
8003 WARN_ON_ONCE(tctx->cached_refs);
8005 percpu_counter_destroy(&tctx->inflight);
8007 tsk->io_uring = NULL;
8010 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8011 struct io_uring_params *p)
8015 /* Retain compatibility with failing for an invalid attach attempt */
8016 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8017 IORING_SETUP_ATTACH_WQ) {
8020 f = fdget(p->wq_fd);
8023 if (f.file->f_op != &io_uring_fops) {
8029 if (ctx->flags & IORING_SETUP_SQPOLL) {
8030 struct task_struct *tsk;
8031 struct io_sq_data *sqd;
8034 sqd = io_get_sq_data(p, &attached);
8040 ctx->sq_creds = get_current_cred();
8042 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8043 if (!ctx->sq_thread_idle)
8044 ctx->sq_thread_idle = HZ;
8046 io_sq_thread_park(sqd);
8047 list_add(&ctx->sqd_list, &sqd->ctx_list);
8048 io_sqd_update_thread_idle(sqd);
8049 /* don't attach to a dying SQPOLL thread, would be racy */
8050 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8051 io_sq_thread_unpark(sqd);
8058 if (p->flags & IORING_SETUP_SQ_AFF) {
8059 int cpu = p->sq_thread_cpu;
8062 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8069 sqd->task_pid = current->pid;
8070 sqd->task_tgid = current->tgid;
8071 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8078 ret = io_uring_alloc_task_context(tsk, ctx);
8079 wake_up_new_task(tsk);
8082 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8083 /* Can't have SQ_AFF without SQPOLL */
8090 complete(&ctx->sq_data->exited);
8092 io_sq_thread_finish(ctx);
8096 static inline void __io_unaccount_mem(struct user_struct *user,
8097 unsigned long nr_pages)
8099 atomic_long_sub(nr_pages, &user->locked_vm);
8102 static inline int __io_account_mem(struct user_struct *user,
8103 unsigned long nr_pages)
8105 unsigned long page_limit, cur_pages, new_pages;
8107 /* Don't allow more pages than we can safely lock */
8108 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8111 cur_pages = atomic_long_read(&user->locked_vm);
8112 new_pages = cur_pages + nr_pages;
8113 if (new_pages > page_limit)
8115 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8116 new_pages) != cur_pages);
8121 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8124 __io_unaccount_mem(ctx->user, nr_pages);
8126 if (ctx->mm_account)
8127 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8130 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8135 ret = __io_account_mem(ctx->user, nr_pages);
8140 if (ctx->mm_account)
8141 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8146 static void io_mem_free(void *ptr)
8153 page = virt_to_head_page(ptr);
8154 if (put_page_testzero(page))
8155 free_compound_page(page);
8158 static void *io_mem_alloc(size_t size)
8160 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8161 __GFP_NORETRY | __GFP_ACCOUNT;
8163 return (void *) __get_free_pages(gfp_flags, get_order(size));
8166 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8169 struct io_rings *rings;
8170 size_t off, sq_array_size;
8172 off = struct_size(rings, cqes, cq_entries);
8173 if (off == SIZE_MAX)
8177 off = ALIGN(off, SMP_CACHE_BYTES);
8185 sq_array_size = array_size(sizeof(u32), sq_entries);
8186 if (sq_array_size == SIZE_MAX)
8189 if (check_add_overflow(off, sq_array_size, &off))
8195 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8197 struct io_mapped_ubuf *imu = *slot;
8200 if (imu != ctx->dummy_ubuf) {
8201 for (i = 0; i < imu->nr_bvecs; i++)
8202 unpin_user_page(imu->bvec[i].bv_page);
8203 if (imu->acct_pages)
8204 io_unaccount_mem(ctx, imu->acct_pages);
8210 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8212 io_buffer_unmap(ctx, &prsrc->buf);
8216 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8220 for (i = 0; i < ctx->nr_user_bufs; i++)
8221 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8222 kfree(ctx->user_bufs);
8223 io_rsrc_data_free(ctx->buf_data);
8224 ctx->user_bufs = NULL;
8225 ctx->buf_data = NULL;
8226 ctx->nr_user_bufs = 0;
8229 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8236 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8238 __io_sqe_buffers_unregister(ctx);
8242 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8243 void __user *arg, unsigned index)
8245 struct iovec __user *src;
8247 #ifdef CONFIG_COMPAT
8249 struct compat_iovec __user *ciovs;
8250 struct compat_iovec ciov;
8252 ciovs = (struct compat_iovec __user *) arg;
8253 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8256 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8257 dst->iov_len = ciov.iov_len;
8261 src = (struct iovec __user *) arg;
8262 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8268 * Not super efficient, but this is just a registration time. And we do cache
8269 * the last compound head, so generally we'll only do a full search if we don't
8272 * We check if the given compound head page has already been accounted, to
8273 * avoid double accounting it. This allows us to account the full size of the
8274 * page, not just the constituent pages of a huge page.
8276 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8277 int nr_pages, struct page *hpage)
8281 /* check current page array */
8282 for (i = 0; i < nr_pages; i++) {
8283 if (!PageCompound(pages[i]))
8285 if (compound_head(pages[i]) == hpage)
8289 /* check previously registered pages */
8290 for (i = 0; i < ctx->nr_user_bufs; i++) {
8291 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8293 for (j = 0; j < imu->nr_bvecs; j++) {
8294 if (!PageCompound(imu->bvec[j].bv_page))
8296 if (compound_head(imu->bvec[j].bv_page) == hpage)
8304 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8305 int nr_pages, struct io_mapped_ubuf *imu,
8306 struct page **last_hpage)
8310 imu->acct_pages = 0;
8311 for (i = 0; i < nr_pages; i++) {
8312 if (!PageCompound(pages[i])) {
8317 hpage = compound_head(pages[i]);
8318 if (hpage == *last_hpage)
8320 *last_hpage = hpage;
8321 if (headpage_already_acct(ctx, pages, i, hpage))
8323 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8327 if (!imu->acct_pages)
8330 ret = io_account_mem(ctx, imu->acct_pages);
8332 imu->acct_pages = 0;
8336 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8337 struct io_mapped_ubuf **pimu,
8338 struct page **last_hpage)
8340 struct io_mapped_ubuf *imu = NULL;
8341 struct vm_area_struct **vmas = NULL;
8342 struct page **pages = NULL;
8343 unsigned long off, start, end, ubuf;
8345 int ret, pret, nr_pages, i;
8347 if (!iov->iov_base) {
8348 *pimu = ctx->dummy_ubuf;
8352 ubuf = (unsigned long) iov->iov_base;
8353 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8354 start = ubuf >> PAGE_SHIFT;
8355 nr_pages = end - start;
8360 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8364 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8369 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8374 mmap_read_lock(current->mm);
8375 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8377 if (pret == nr_pages) {
8378 /* don't support file backed memory */
8379 for (i = 0; i < nr_pages; i++) {
8380 struct vm_area_struct *vma = vmas[i];
8382 if (vma_is_shmem(vma))
8385 !is_file_hugepages(vma->vm_file)) {
8391 ret = pret < 0 ? pret : -EFAULT;
8393 mmap_read_unlock(current->mm);
8396 * if we did partial map, or found file backed vmas,
8397 * release any pages we did get
8400 unpin_user_pages(pages, pret);
8404 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8406 unpin_user_pages(pages, pret);
8410 off = ubuf & ~PAGE_MASK;
8411 size = iov->iov_len;
8412 for (i = 0; i < nr_pages; i++) {
8415 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8416 imu->bvec[i].bv_page = pages[i];
8417 imu->bvec[i].bv_len = vec_len;
8418 imu->bvec[i].bv_offset = off;
8422 /* store original address for later verification */
8424 imu->ubuf_end = ubuf + iov->iov_len;
8425 imu->nr_bvecs = nr_pages;
8436 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8438 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8439 return ctx->user_bufs ? 0 : -ENOMEM;
8442 static int io_buffer_validate(struct iovec *iov)
8444 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8447 * Don't impose further limits on the size and buffer
8448 * constraints here, we'll -EINVAL later when IO is
8449 * submitted if they are wrong.
8452 return iov->iov_len ? -EFAULT : 0;
8456 /* arbitrary limit, but we need something */
8457 if (iov->iov_len > SZ_1G)
8460 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8466 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8467 unsigned int nr_args, u64 __user *tags)
8469 struct page *last_hpage = NULL;
8470 struct io_rsrc_data *data;
8476 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8478 ret = io_rsrc_node_switch_start(ctx);
8481 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8484 ret = io_buffers_map_alloc(ctx, nr_args);
8486 io_rsrc_data_free(data);
8490 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8491 ret = io_copy_iov(ctx, &iov, arg, i);
8494 ret = io_buffer_validate(&iov);
8497 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8502 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8508 WARN_ON_ONCE(ctx->buf_data);
8510 ctx->buf_data = data;
8512 __io_sqe_buffers_unregister(ctx);
8514 io_rsrc_node_switch(ctx, NULL);
8518 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8519 struct io_uring_rsrc_update2 *up,
8520 unsigned int nr_args)
8522 u64 __user *tags = u64_to_user_ptr(up->tags);
8523 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8524 struct page *last_hpage = NULL;
8525 bool needs_switch = false;
8531 if (up->offset + nr_args > ctx->nr_user_bufs)
8534 for (done = 0; done < nr_args; done++) {
8535 struct io_mapped_ubuf *imu;
8536 int offset = up->offset + done;
8539 err = io_copy_iov(ctx, &iov, iovs, done);
8542 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8546 err = io_buffer_validate(&iov);
8549 if (!iov.iov_base && tag) {
8553 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8557 i = array_index_nospec(offset, ctx->nr_user_bufs);
8558 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8559 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8560 ctx->rsrc_node, ctx->user_bufs[i]);
8561 if (unlikely(err)) {
8562 io_buffer_unmap(ctx, &imu);
8565 ctx->user_bufs[i] = NULL;
8566 needs_switch = true;
8569 ctx->user_bufs[i] = imu;
8570 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8574 io_rsrc_node_switch(ctx, ctx->buf_data);
8575 return done ? done : err;
8578 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8580 __s32 __user *fds = arg;
8586 if (copy_from_user(&fd, fds, sizeof(*fds)))
8589 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8590 if (IS_ERR(ctx->cq_ev_fd)) {
8591 int ret = PTR_ERR(ctx->cq_ev_fd);
8593 ctx->cq_ev_fd = NULL;
8600 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8602 if (ctx->cq_ev_fd) {
8603 eventfd_ctx_put(ctx->cq_ev_fd);
8604 ctx->cq_ev_fd = NULL;
8611 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8613 struct io_buffer *buf;
8614 unsigned long index;
8616 xa_for_each(&ctx->io_buffers, index, buf)
8617 __io_remove_buffers(ctx, buf, index, -1U);
8620 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8622 struct io_kiocb *req, *nxt;
8624 list_for_each_entry_safe(req, nxt, list, compl.list) {
8625 if (tsk && req->task != tsk)
8627 list_del(&req->compl.list);
8628 kmem_cache_free(req_cachep, req);
8632 static void io_req_caches_free(struct io_ring_ctx *ctx)
8634 struct io_submit_state *submit_state = &ctx->submit_state;
8635 struct io_comp_state *cs = &ctx->submit_state.comp;
8637 mutex_lock(&ctx->uring_lock);
8639 if (submit_state->free_reqs) {
8640 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8641 submit_state->reqs);
8642 submit_state->free_reqs = 0;
8645 io_flush_cached_locked_reqs(ctx, cs);
8646 io_req_cache_free(&cs->free_list, NULL);
8647 mutex_unlock(&ctx->uring_lock);
8650 static void io_wait_rsrc_data(struct io_rsrc_data *data)
8652 if (data && !atomic_dec_and_test(&data->refs))
8653 wait_for_completion(&data->done);
8656 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8658 io_sq_thread_finish(ctx);
8660 if (ctx->mm_account) {
8661 mmdrop(ctx->mm_account);
8662 ctx->mm_account = NULL;
8665 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
8666 io_wait_rsrc_data(ctx->buf_data);
8667 io_wait_rsrc_data(ctx->file_data);
8669 mutex_lock(&ctx->uring_lock);
8671 __io_sqe_buffers_unregister(ctx);
8673 __io_sqe_files_unregister(ctx);
8675 __io_cqring_overflow_flush(ctx, true);
8676 mutex_unlock(&ctx->uring_lock);
8677 io_eventfd_unregister(ctx);
8678 io_destroy_buffers(ctx);
8680 put_cred(ctx->sq_creds);
8682 /* there are no registered resources left, nobody uses it */
8684 io_rsrc_node_destroy(ctx->rsrc_node);
8685 if (ctx->rsrc_backup_node)
8686 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8687 flush_delayed_work(&ctx->rsrc_put_work);
8689 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8690 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8692 #if defined(CONFIG_UNIX)
8693 if (ctx->ring_sock) {
8694 ctx->ring_sock->file = NULL; /* so that iput() is called */
8695 sock_release(ctx->ring_sock);
8699 io_mem_free(ctx->rings);
8700 io_mem_free(ctx->sq_sqes);
8702 percpu_ref_exit(&ctx->refs);
8703 free_uid(ctx->user);
8704 io_req_caches_free(ctx);
8706 io_wq_put_hash(ctx->hash_map);
8707 kfree(ctx->cancel_hash);
8708 kfree(ctx->dummy_ubuf);
8712 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8714 struct io_ring_ctx *ctx = file->private_data;
8717 poll_wait(file, &ctx->poll_wait, wait);
8719 * synchronizes with barrier from wq_has_sleeper call in
8723 if (!io_sqring_full(ctx))
8724 mask |= EPOLLOUT | EPOLLWRNORM;
8727 * Don't flush cqring overflow list here, just do a simple check.
8728 * Otherwise there could possible be ABBA deadlock:
8731 * lock(&ctx->uring_lock);
8733 * lock(&ctx->uring_lock);
8736 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8737 * pushs them to do the flush.
8739 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8740 mask |= EPOLLIN | EPOLLRDNORM;
8745 static int io_uring_fasync(int fd, struct file *file, int on)
8747 struct io_ring_ctx *ctx = file->private_data;
8749 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8752 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8754 const struct cred *creds;
8756 creds = xa_erase(&ctx->personalities, id);
8765 struct io_tctx_exit {
8766 struct callback_head task_work;
8767 struct completion completion;
8768 struct io_ring_ctx *ctx;
8771 static void io_tctx_exit_cb(struct callback_head *cb)
8773 struct io_uring_task *tctx = current->io_uring;
8774 struct io_tctx_exit *work;
8776 work = container_of(cb, struct io_tctx_exit, task_work);
8778 * When @in_idle, we're in cancellation and it's racy to remove the
8779 * node. It'll be removed by the end of cancellation, just ignore it.
8781 if (!atomic_read(&tctx->in_idle))
8782 io_uring_del_tctx_node((unsigned long)work->ctx);
8783 complete(&work->completion);
8786 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8788 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8790 return req->ctx == data;
8793 static void io_ring_exit_work(struct work_struct *work)
8795 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8796 unsigned long timeout = jiffies + HZ * 60 * 5;
8797 unsigned long interval = HZ / 20;
8798 struct io_tctx_exit exit;
8799 struct io_tctx_node *node;
8803 * If we're doing polled IO and end up having requests being
8804 * submitted async (out-of-line), then completions can come in while
8805 * we're waiting for refs to drop. We need to reap these manually,
8806 * as nobody else will be looking for them.
8809 io_uring_try_cancel_requests(ctx, NULL, true);
8811 struct io_sq_data *sqd = ctx->sq_data;
8812 struct task_struct *tsk;
8814 io_sq_thread_park(sqd);
8816 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8817 io_wq_cancel_cb(tsk->io_uring->io_wq,
8818 io_cancel_ctx_cb, ctx, true);
8819 io_sq_thread_unpark(sqd);
8822 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
8823 /* there is little hope left, don't run it too often */
8826 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
8828 init_completion(&exit.completion);
8829 init_task_work(&exit.task_work, io_tctx_exit_cb);
8832 * Some may use context even when all refs and requests have been put,
8833 * and they are free to do so while still holding uring_lock or
8834 * completion_lock, see io_req_task_submit(). Apart from other work,
8835 * this lock/unlock section also waits them to finish.
8837 mutex_lock(&ctx->uring_lock);
8838 while (!list_empty(&ctx->tctx_list)) {
8839 WARN_ON_ONCE(time_after(jiffies, timeout));
8841 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8843 /* don't spin on a single task if cancellation failed */
8844 list_rotate_left(&ctx->tctx_list);
8845 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8846 if (WARN_ON_ONCE(ret))
8848 wake_up_process(node->task);
8850 mutex_unlock(&ctx->uring_lock);
8851 wait_for_completion(&exit.completion);
8852 mutex_lock(&ctx->uring_lock);
8854 mutex_unlock(&ctx->uring_lock);
8855 spin_lock_irq(&ctx->completion_lock);
8856 spin_unlock_irq(&ctx->completion_lock);
8858 io_ring_ctx_free(ctx);
8861 /* Returns true if we found and killed one or more timeouts */
8862 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8865 struct io_kiocb *req, *tmp;
8868 spin_lock_irq(&ctx->completion_lock);
8869 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8870 if (io_match_task(req, tsk, cancel_all)) {
8871 io_kill_timeout(req, -ECANCELED);
8876 io_commit_cqring(ctx);
8877 spin_unlock_irq(&ctx->completion_lock);
8879 io_cqring_ev_posted(ctx);
8880 return canceled != 0;
8883 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8885 unsigned long index;
8886 struct creds *creds;
8888 mutex_lock(&ctx->uring_lock);
8889 percpu_ref_kill(&ctx->refs);
8891 __io_cqring_overflow_flush(ctx, true);
8892 xa_for_each(&ctx->personalities, index, creds)
8893 io_unregister_personality(ctx, index);
8894 mutex_unlock(&ctx->uring_lock);
8896 io_kill_timeouts(ctx, NULL, true);
8897 io_poll_remove_all(ctx, NULL, true);
8899 /* if we failed setting up the ctx, we might not have any rings */
8900 io_iopoll_try_reap_events(ctx);
8902 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8904 * Use system_unbound_wq to avoid spawning tons of event kworkers
8905 * if we're exiting a ton of rings at the same time. It just adds
8906 * noise and overhead, there's no discernable change in runtime
8907 * over using system_wq.
8909 queue_work(system_unbound_wq, &ctx->exit_work);
8912 static int io_uring_release(struct inode *inode, struct file *file)
8914 struct io_ring_ctx *ctx = file->private_data;
8916 file->private_data = NULL;
8917 io_ring_ctx_wait_and_kill(ctx);
8921 struct io_task_cancel {
8922 struct task_struct *task;
8926 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8928 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8929 struct io_task_cancel *cancel = data;
8932 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8933 unsigned long flags;
8934 struct io_ring_ctx *ctx = req->ctx;
8936 /* protect against races with linked timeouts */
8937 spin_lock_irqsave(&ctx->completion_lock, flags);
8938 ret = io_match_task(req, cancel->task, cancel->all);
8939 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8941 ret = io_match_task(req, cancel->task, cancel->all);
8946 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8947 struct task_struct *task, bool cancel_all)
8949 struct io_defer_entry *de;
8952 spin_lock_irq(&ctx->completion_lock);
8953 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8954 if (io_match_task(de->req, task, cancel_all)) {
8955 list_cut_position(&list, &ctx->defer_list, &de->list);
8959 spin_unlock_irq(&ctx->completion_lock);
8960 if (list_empty(&list))
8963 while (!list_empty(&list)) {
8964 de = list_first_entry(&list, struct io_defer_entry, list);
8965 list_del_init(&de->list);
8966 io_req_complete_failed(de->req, -ECANCELED);
8972 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8974 struct io_tctx_node *node;
8975 enum io_wq_cancel cret;
8978 mutex_lock(&ctx->uring_lock);
8979 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8980 struct io_uring_task *tctx = node->task->io_uring;
8983 * io_wq will stay alive while we hold uring_lock, because it's
8984 * killed after ctx nodes, which requires to take the lock.
8986 if (!tctx || !tctx->io_wq)
8988 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8989 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8991 mutex_unlock(&ctx->uring_lock);
8996 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8997 struct task_struct *task,
9000 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9001 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9004 enum io_wq_cancel cret;
9008 ret |= io_uring_try_cancel_iowq(ctx);
9009 } else if (tctx && tctx->io_wq) {
9011 * Cancels requests of all rings, not only @ctx, but
9012 * it's fine as the task is in exit/exec.
9014 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9016 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9019 /* SQPOLL thread does its own polling */
9020 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9021 (ctx->sq_data && ctx->sq_data->thread == current)) {
9022 while (!list_empty_careful(&ctx->iopoll_list)) {
9023 io_iopoll_try_reap_events(ctx);
9028 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9029 ret |= io_poll_remove_all(ctx, task, cancel_all);
9030 ret |= io_kill_timeouts(ctx, task, cancel_all);
9032 ret |= io_run_task_work();
9039 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9041 struct io_uring_task *tctx = current->io_uring;
9042 struct io_tctx_node *node;
9045 if (unlikely(!tctx)) {
9046 ret = io_uring_alloc_task_context(current, ctx);
9049 tctx = current->io_uring;
9051 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9052 node = kmalloc(sizeof(*node), GFP_KERNEL);
9056 node->task = current;
9058 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9065 mutex_lock(&ctx->uring_lock);
9066 list_add(&node->ctx_node, &ctx->tctx_list);
9067 mutex_unlock(&ctx->uring_lock);
9074 * Note that this task has used io_uring. We use it for cancelation purposes.
9076 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9078 struct io_uring_task *tctx = current->io_uring;
9080 if (likely(tctx && tctx->last == ctx))
9082 return __io_uring_add_tctx_node(ctx);
9086 * Remove this io_uring_file -> task mapping.
9088 static void io_uring_del_tctx_node(unsigned long index)
9090 struct io_uring_task *tctx = current->io_uring;
9091 struct io_tctx_node *node;
9095 node = xa_erase(&tctx->xa, index);
9099 WARN_ON_ONCE(current != node->task);
9100 WARN_ON_ONCE(list_empty(&node->ctx_node));
9102 mutex_lock(&node->ctx->uring_lock);
9103 list_del(&node->ctx_node);
9104 mutex_unlock(&node->ctx->uring_lock);
9106 if (tctx->last == node->ctx)
9111 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9113 struct io_wq *wq = tctx->io_wq;
9114 struct io_tctx_node *node;
9115 unsigned long index;
9117 xa_for_each(&tctx->xa, index, node)
9118 io_uring_del_tctx_node(index);
9121 * Must be after io_uring_del_task_file() (removes nodes under
9122 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9125 io_wq_put_and_exit(wq);
9129 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9132 return atomic_read(&tctx->inflight_tracked);
9133 return percpu_counter_sum(&tctx->inflight);
9136 static void io_uring_drop_tctx_refs(struct task_struct *task)
9138 struct io_uring_task *tctx = task->io_uring;
9139 unsigned int refs = tctx->cached_refs;
9142 tctx->cached_refs = 0;
9143 percpu_counter_sub(&tctx->inflight, refs);
9144 put_task_struct_many(task, refs);
9149 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9150 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9152 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9154 struct io_uring_task *tctx = current->io_uring;
9155 struct io_ring_ctx *ctx;
9159 WARN_ON_ONCE(sqd && sqd->thread != current);
9161 if (!current->io_uring)
9164 io_wq_exit_start(tctx->io_wq);
9166 atomic_inc(&tctx->in_idle);
9168 io_uring_drop_tctx_refs(current);
9169 /* read completions before cancelations */
9170 inflight = tctx_inflight(tctx, !cancel_all);
9175 struct io_tctx_node *node;
9176 unsigned long index;
9178 xa_for_each(&tctx->xa, index, node) {
9179 /* sqpoll task will cancel all its requests */
9180 if (node->ctx->sq_data)
9182 io_uring_try_cancel_requests(node->ctx, current,
9186 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9187 io_uring_try_cancel_requests(ctx, current,
9191 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9192 io_uring_drop_tctx_refs(current);
9194 * If we've seen completions, retry without waiting. This
9195 * avoids a race where a completion comes in before we did
9196 * prepare_to_wait().
9198 if (inflight == tctx_inflight(tctx, !cancel_all))
9200 finish_wait(&tctx->wait, &wait);
9202 atomic_dec(&tctx->in_idle);
9204 io_uring_clean_tctx(tctx);
9206 /* for exec all current's requests should be gone, kill tctx */
9207 __io_uring_free(current);
9211 void __io_uring_cancel(struct files_struct *files)
9213 io_uring_cancel_generic(!files, NULL);
9216 static void *io_uring_validate_mmap_request(struct file *file,
9217 loff_t pgoff, size_t sz)
9219 struct io_ring_ctx *ctx = file->private_data;
9220 loff_t offset = pgoff << PAGE_SHIFT;
9225 case IORING_OFF_SQ_RING:
9226 case IORING_OFF_CQ_RING:
9229 case IORING_OFF_SQES:
9233 return ERR_PTR(-EINVAL);
9236 page = virt_to_head_page(ptr);
9237 if (sz > page_size(page))
9238 return ERR_PTR(-EINVAL);
9245 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9247 size_t sz = vma->vm_end - vma->vm_start;
9251 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9253 return PTR_ERR(ptr);
9255 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9256 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9259 #else /* !CONFIG_MMU */
9261 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9263 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9266 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9268 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9271 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9272 unsigned long addr, unsigned long len,
9273 unsigned long pgoff, unsigned long flags)
9277 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9279 return PTR_ERR(ptr);
9281 return (unsigned long) ptr;
9284 #endif /* !CONFIG_MMU */
9286 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9291 if (!io_sqring_full(ctx))
9293 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9295 if (!io_sqring_full(ctx))
9298 } while (!signal_pending(current));
9300 finish_wait(&ctx->sqo_sq_wait, &wait);
9304 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9305 struct __kernel_timespec __user **ts,
9306 const sigset_t __user **sig)
9308 struct io_uring_getevents_arg arg;
9311 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9312 * is just a pointer to the sigset_t.
9314 if (!(flags & IORING_ENTER_EXT_ARG)) {
9315 *sig = (const sigset_t __user *) argp;
9321 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9322 * timespec and sigset_t pointers if good.
9324 if (*argsz != sizeof(arg))
9326 if (copy_from_user(&arg, argp, sizeof(arg)))
9328 *sig = u64_to_user_ptr(arg.sigmask);
9329 *argsz = arg.sigmask_sz;
9330 *ts = u64_to_user_ptr(arg.ts);
9334 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9335 u32, min_complete, u32, flags, const void __user *, argp,
9338 struct io_ring_ctx *ctx;
9345 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9346 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9350 if (unlikely(!f.file))
9354 if (unlikely(f.file->f_op != &io_uring_fops))
9358 ctx = f.file->private_data;
9359 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9363 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9367 * For SQ polling, the thread will do all submissions and completions.
9368 * Just return the requested submit count, and wake the thread if
9372 if (ctx->flags & IORING_SETUP_SQPOLL) {
9373 io_cqring_overflow_flush(ctx, false);
9375 if (unlikely(ctx->sq_data->thread == NULL)) {
9379 if (flags & IORING_ENTER_SQ_WAKEUP)
9380 wake_up(&ctx->sq_data->wait);
9381 if (flags & IORING_ENTER_SQ_WAIT) {
9382 ret = io_sqpoll_wait_sq(ctx);
9386 submitted = to_submit;
9387 } else if (to_submit) {
9388 ret = io_uring_add_tctx_node(ctx);
9391 mutex_lock(&ctx->uring_lock);
9392 submitted = io_submit_sqes(ctx, to_submit);
9393 mutex_unlock(&ctx->uring_lock);
9395 if (submitted != to_submit)
9398 if (flags & IORING_ENTER_GETEVENTS) {
9399 const sigset_t __user *sig;
9400 struct __kernel_timespec __user *ts;
9402 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9406 min_complete = min(min_complete, ctx->cq_entries);
9409 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9410 * space applications don't need to do io completion events
9411 * polling again, they can rely on io_sq_thread to do polling
9412 * work, which can reduce cpu usage and uring_lock contention.
9414 if (ctx->flags & IORING_SETUP_IOPOLL &&
9415 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9416 ret = io_iopoll_check(ctx, min_complete);
9418 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9423 percpu_ref_put(&ctx->refs);
9426 return submitted ? submitted : ret;
9429 #ifdef CONFIG_PROC_FS
9430 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9431 const struct cred *cred)
9433 struct user_namespace *uns = seq_user_ns(m);
9434 struct group_info *gi;
9439 seq_printf(m, "%5d\n", id);
9440 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9441 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9442 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9443 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9444 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9445 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9446 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9447 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9448 seq_puts(m, "\n\tGroups:\t");
9449 gi = cred->group_info;
9450 for (g = 0; g < gi->ngroups; g++) {
9451 seq_put_decimal_ull(m, g ? " " : "",
9452 from_kgid_munged(uns, gi->gid[g]));
9454 seq_puts(m, "\n\tCapEff:\t");
9455 cap = cred->cap_effective;
9456 CAP_FOR_EACH_U32(__capi)
9457 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9462 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9464 struct io_sq_data *sq = NULL;
9469 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9470 * since fdinfo case grabs it in the opposite direction of normal use
9471 * cases. If we fail to get the lock, we just don't iterate any
9472 * structures that could be going away outside the io_uring mutex.
9474 has_lock = mutex_trylock(&ctx->uring_lock);
9476 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9482 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9483 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9484 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9485 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9486 struct file *f = io_file_from_index(ctx, i);
9489 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9491 seq_printf(m, "%5u: <none>\n", i);
9493 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9494 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9495 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9496 unsigned int len = buf->ubuf_end - buf->ubuf;
9498 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9500 if (has_lock && !xa_empty(&ctx->personalities)) {
9501 unsigned long index;
9502 const struct cred *cred;
9504 seq_printf(m, "Personalities:\n");
9505 xa_for_each(&ctx->personalities, index, cred)
9506 io_uring_show_cred(m, index, cred);
9508 seq_printf(m, "PollList:\n");
9509 spin_lock_irq(&ctx->completion_lock);
9510 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9511 struct hlist_head *list = &ctx->cancel_hash[i];
9512 struct io_kiocb *req;
9514 hlist_for_each_entry(req, list, hash_node)
9515 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9516 req->task->task_works != NULL);
9518 spin_unlock_irq(&ctx->completion_lock);
9520 mutex_unlock(&ctx->uring_lock);
9523 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9525 struct io_ring_ctx *ctx = f->private_data;
9527 if (percpu_ref_tryget(&ctx->refs)) {
9528 __io_uring_show_fdinfo(ctx, m);
9529 percpu_ref_put(&ctx->refs);
9534 static const struct file_operations io_uring_fops = {
9535 .release = io_uring_release,
9536 .mmap = io_uring_mmap,
9538 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9539 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9541 .poll = io_uring_poll,
9542 .fasync = io_uring_fasync,
9543 #ifdef CONFIG_PROC_FS
9544 .show_fdinfo = io_uring_show_fdinfo,
9548 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9549 struct io_uring_params *p)
9551 struct io_rings *rings;
9552 size_t size, sq_array_offset;
9554 /* make sure these are sane, as we already accounted them */
9555 ctx->sq_entries = p->sq_entries;
9556 ctx->cq_entries = p->cq_entries;
9558 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9559 if (size == SIZE_MAX)
9562 rings = io_mem_alloc(size);
9567 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9568 rings->sq_ring_mask = p->sq_entries - 1;
9569 rings->cq_ring_mask = p->cq_entries - 1;
9570 rings->sq_ring_entries = p->sq_entries;
9571 rings->cq_ring_entries = p->cq_entries;
9573 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9574 if (size == SIZE_MAX) {
9575 io_mem_free(ctx->rings);
9580 ctx->sq_sqes = io_mem_alloc(size);
9581 if (!ctx->sq_sqes) {
9582 io_mem_free(ctx->rings);
9590 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9594 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9598 ret = io_uring_add_tctx_node(ctx);
9603 fd_install(fd, file);
9608 * Allocate an anonymous fd, this is what constitutes the application
9609 * visible backing of an io_uring instance. The application mmaps this
9610 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9611 * we have to tie this fd to a socket for file garbage collection purposes.
9613 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9616 #if defined(CONFIG_UNIX)
9619 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9622 return ERR_PTR(ret);
9625 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9626 O_RDWR | O_CLOEXEC);
9627 #if defined(CONFIG_UNIX)
9629 sock_release(ctx->ring_sock);
9630 ctx->ring_sock = NULL;
9632 ctx->ring_sock->file = file;
9638 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9639 struct io_uring_params __user *params)
9641 struct io_ring_ctx *ctx;
9647 if (entries > IORING_MAX_ENTRIES) {
9648 if (!(p->flags & IORING_SETUP_CLAMP))
9650 entries = IORING_MAX_ENTRIES;
9654 * Use twice as many entries for the CQ ring. It's possible for the
9655 * application to drive a higher depth than the size of the SQ ring,
9656 * since the sqes are only used at submission time. This allows for
9657 * some flexibility in overcommitting a bit. If the application has
9658 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9659 * of CQ ring entries manually.
9661 p->sq_entries = roundup_pow_of_two(entries);
9662 if (p->flags & IORING_SETUP_CQSIZE) {
9664 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9665 * to a power-of-two, if it isn't already. We do NOT impose
9666 * any cq vs sq ring sizing.
9670 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9671 if (!(p->flags & IORING_SETUP_CLAMP))
9673 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9675 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9676 if (p->cq_entries < p->sq_entries)
9679 p->cq_entries = 2 * p->sq_entries;
9682 ctx = io_ring_ctx_alloc(p);
9685 ctx->compat = in_compat_syscall();
9686 if (!capable(CAP_IPC_LOCK))
9687 ctx->user = get_uid(current_user());
9690 * This is just grabbed for accounting purposes. When a process exits,
9691 * the mm is exited and dropped before the files, hence we need to hang
9692 * on to this mm purely for the purposes of being able to unaccount
9693 * memory (locked/pinned vm). It's not used for anything else.
9695 mmgrab(current->mm);
9696 ctx->mm_account = current->mm;
9698 ret = io_allocate_scq_urings(ctx, p);
9702 ret = io_sq_offload_create(ctx, p);
9705 /* always set a rsrc node */
9706 ret = io_rsrc_node_switch_start(ctx);
9709 io_rsrc_node_switch(ctx, NULL);
9711 memset(&p->sq_off, 0, sizeof(p->sq_off));
9712 p->sq_off.head = offsetof(struct io_rings, sq.head);
9713 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9714 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9715 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9716 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9717 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9718 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9720 memset(&p->cq_off, 0, sizeof(p->cq_off));
9721 p->cq_off.head = offsetof(struct io_rings, cq.head);
9722 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9723 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9724 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9725 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9726 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9727 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9729 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9730 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9731 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9732 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9733 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9734 IORING_FEAT_RSRC_TAGS;
9736 if (copy_to_user(params, p, sizeof(*p))) {
9741 file = io_uring_get_file(ctx);
9743 ret = PTR_ERR(file);
9748 * Install ring fd as the very last thing, so we don't risk someone
9749 * having closed it before we finish setup
9751 ret = io_uring_install_fd(ctx, file);
9753 /* fput will clean it up */
9758 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9761 io_ring_ctx_wait_and_kill(ctx);
9766 * Sets up an aio uring context, and returns the fd. Applications asks for a
9767 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9768 * params structure passed in.
9770 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9772 struct io_uring_params p;
9775 if (copy_from_user(&p, params, sizeof(p)))
9777 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9782 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9783 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9784 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9785 IORING_SETUP_R_DISABLED))
9788 return io_uring_create(entries, &p, params);
9791 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9792 struct io_uring_params __user *, params)
9794 return io_uring_setup(entries, params);
9797 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9799 struct io_uring_probe *p;
9803 size = struct_size(p, ops, nr_args);
9804 if (size == SIZE_MAX)
9806 p = kzalloc(size, GFP_KERNEL);
9811 if (copy_from_user(p, arg, size))
9814 if (memchr_inv(p, 0, size))
9817 p->last_op = IORING_OP_LAST - 1;
9818 if (nr_args > IORING_OP_LAST)
9819 nr_args = IORING_OP_LAST;
9821 for (i = 0; i < nr_args; i++) {
9823 if (!io_op_defs[i].not_supported)
9824 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9829 if (copy_to_user(arg, p, size))
9836 static int io_register_personality(struct io_ring_ctx *ctx)
9838 const struct cred *creds;
9842 creds = get_current_cred();
9844 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9845 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9853 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9854 unsigned int nr_args)
9856 struct io_uring_restriction *res;
9860 /* Restrictions allowed only if rings started disabled */
9861 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9864 /* We allow only a single restrictions registration */
9865 if (ctx->restrictions.registered)
9868 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9871 size = array_size(nr_args, sizeof(*res));
9872 if (size == SIZE_MAX)
9875 res = memdup_user(arg, size);
9877 return PTR_ERR(res);
9881 for (i = 0; i < nr_args; i++) {
9882 switch (res[i].opcode) {
9883 case IORING_RESTRICTION_REGISTER_OP:
9884 if (res[i].register_op >= IORING_REGISTER_LAST) {
9889 __set_bit(res[i].register_op,
9890 ctx->restrictions.register_op);
9892 case IORING_RESTRICTION_SQE_OP:
9893 if (res[i].sqe_op >= IORING_OP_LAST) {
9898 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9900 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9901 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9903 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9904 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9913 /* Reset all restrictions if an error happened */
9915 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9917 ctx->restrictions.registered = true;
9923 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9925 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9928 if (ctx->restrictions.registered)
9929 ctx->restricted = 1;
9931 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9932 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9933 wake_up(&ctx->sq_data->wait);
9937 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9938 struct io_uring_rsrc_update2 *up,
9946 if (check_add_overflow(up->offset, nr_args, &tmp))
9948 err = io_rsrc_node_switch_start(ctx);
9953 case IORING_RSRC_FILE:
9954 return __io_sqe_files_update(ctx, up, nr_args);
9955 case IORING_RSRC_BUFFER:
9956 return __io_sqe_buffers_update(ctx, up, nr_args);
9961 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9964 struct io_uring_rsrc_update2 up;
9968 memset(&up, 0, sizeof(up));
9969 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9971 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9974 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9975 unsigned size, unsigned type)
9977 struct io_uring_rsrc_update2 up;
9979 if (size != sizeof(up))
9981 if (copy_from_user(&up, arg, sizeof(up)))
9983 if (!up.nr || up.resv)
9985 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9988 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9989 unsigned int size, unsigned int type)
9991 struct io_uring_rsrc_register rr;
9993 /* keep it extendible */
9994 if (size != sizeof(rr))
9997 memset(&rr, 0, sizeof(rr));
9998 if (copy_from_user(&rr, arg, size))
10000 if (!rr.nr || rr.resv || rr.resv2)
10004 case IORING_RSRC_FILE:
10005 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10006 rr.nr, u64_to_user_ptr(rr.tags));
10007 case IORING_RSRC_BUFFER:
10008 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10009 rr.nr, u64_to_user_ptr(rr.tags));
10014 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10017 struct io_uring_task *tctx = current->io_uring;
10018 cpumask_var_t new_mask;
10021 if (!tctx || !tctx->io_wq)
10024 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10027 cpumask_clear(new_mask);
10028 if (len > cpumask_size())
10029 len = cpumask_size();
10031 if (copy_from_user(new_mask, arg, len)) {
10032 free_cpumask_var(new_mask);
10036 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10037 free_cpumask_var(new_mask);
10041 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10043 struct io_uring_task *tctx = current->io_uring;
10045 if (!tctx || !tctx->io_wq)
10048 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10051 static bool io_register_op_must_quiesce(int op)
10054 case IORING_REGISTER_BUFFERS:
10055 case IORING_UNREGISTER_BUFFERS:
10056 case IORING_REGISTER_FILES:
10057 case IORING_UNREGISTER_FILES:
10058 case IORING_REGISTER_FILES_UPDATE:
10059 case IORING_REGISTER_PROBE:
10060 case IORING_REGISTER_PERSONALITY:
10061 case IORING_UNREGISTER_PERSONALITY:
10062 case IORING_REGISTER_FILES2:
10063 case IORING_REGISTER_FILES_UPDATE2:
10064 case IORING_REGISTER_BUFFERS2:
10065 case IORING_REGISTER_BUFFERS_UPDATE:
10066 case IORING_REGISTER_IOWQ_AFF:
10067 case IORING_UNREGISTER_IOWQ_AFF:
10074 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10078 percpu_ref_kill(&ctx->refs);
10081 * Drop uring mutex before waiting for references to exit. If another
10082 * thread is currently inside io_uring_enter() it might need to grab the
10083 * uring_lock to make progress. If we hold it here across the drain
10084 * wait, then we can deadlock. It's safe to drop the mutex here, since
10085 * no new references will come in after we've killed the percpu ref.
10087 mutex_unlock(&ctx->uring_lock);
10089 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10092 ret = io_run_task_work_sig();
10093 } while (ret >= 0);
10094 mutex_lock(&ctx->uring_lock);
10097 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10101 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10102 void __user *arg, unsigned nr_args)
10103 __releases(ctx->uring_lock)
10104 __acquires(ctx->uring_lock)
10109 * We're inside the ring mutex, if the ref is already dying, then
10110 * someone else killed the ctx or is already going through
10111 * io_uring_register().
10113 if (percpu_ref_is_dying(&ctx->refs))
10116 if (ctx->restricted) {
10117 if (opcode >= IORING_REGISTER_LAST)
10119 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10120 if (!test_bit(opcode, ctx->restrictions.register_op))
10124 if (io_register_op_must_quiesce(opcode)) {
10125 ret = io_ctx_quiesce(ctx);
10131 case IORING_REGISTER_BUFFERS:
10132 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10134 case IORING_UNREGISTER_BUFFERS:
10136 if (arg || nr_args)
10138 ret = io_sqe_buffers_unregister(ctx);
10140 case IORING_REGISTER_FILES:
10141 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10143 case IORING_UNREGISTER_FILES:
10145 if (arg || nr_args)
10147 ret = io_sqe_files_unregister(ctx);
10149 case IORING_REGISTER_FILES_UPDATE:
10150 ret = io_register_files_update(ctx, arg, nr_args);
10152 case IORING_REGISTER_EVENTFD:
10153 case IORING_REGISTER_EVENTFD_ASYNC:
10157 ret = io_eventfd_register(ctx, arg);
10160 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10161 ctx->eventfd_async = 1;
10163 ctx->eventfd_async = 0;
10165 case IORING_UNREGISTER_EVENTFD:
10167 if (arg || nr_args)
10169 ret = io_eventfd_unregister(ctx);
10171 case IORING_REGISTER_PROBE:
10173 if (!arg || nr_args > 256)
10175 ret = io_probe(ctx, arg, nr_args);
10177 case IORING_REGISTER_PERSONALITY:
10179 if (arg || nr_args)
10181 ret = io_register_personality(ctx);
10183 case IORING_UNREGISTER_PERSONALITY:
10187 ret = io_unregister_personality(ctx, nr_args);
10189 case IORING_REGISTER_ENABLE_RINGS:
10191 if (arg || nr_args)
10193 ret = io_register_enable_rings(ctx);
10195 case IORING_REGISTER_RESTRICTIONS:
10196 ret = io_register_restrictions(ctx, arg, nr_args);
10198 case IORING_REGISTER_FILES2:
10199 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10201 case IORING_REGISTER_FILES_UPDATE2:
10202 ret = io_register_rsrc_update(ctx, arg, nr_args,
10205 case IORING_REGISTER_BUFFERS2:
10206 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10208 case IORING_REGISTER_BUFFERS_UPDATE:
10209 ret = io_register_rsrc_update(ctx, arg, nr_args,
10210 IORING_RSRC_BUFFER);
10212 case IORING_REGISTER_IOWQ_AFF:
10214 if (!arg || !nr_args)
10216 ret = io_register_iowq_aff(ctx, arg, nr_args);
10218 case IORING_UNREGISTER_IOWQ_AFF:
10220 if (arg || nr_args)
10222 ret = io_unregister_iowq_aff(ctx);
10229 if (io_register_op_must_quiesce(opcode)) {
10230 /* bring the ctx back to life */
10231 percpu_ref_reinit(&ctx->refs);
10232 reinit_completion(&ctx->ref_comp);
10237 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10238 void __user *, arg, unsigned int, nr_args)
10240 struct io_ring_ctx *ctx;
10249 if (f.file->f_op != &io_uring_fops)
10252 ctx = f.file->private_data;
10254 io_run_task_work();
10256 mutex_lock(&ctx->uring_lock);
10257 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10258 mutex_unlock(&ctx->uring_lock);
10259 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10260 ctx->cq_ev_fd != NULL, ret);
10266 static int __init io_uring_init(void)
10268 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10269 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10270 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10273 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10274 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10275 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10276 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10277 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10278 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10279 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10280 BUILD_BUG_SQE_ELEM(8, __u64, off);
10281 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10282 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10283 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10284 BUILD_BUG_SQE_ELEM(24, __u32, len);
10285 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10286 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10287 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10288 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10289 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10290 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10291 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10292 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10293 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10294 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10295 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10296 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10297 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10298 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10299 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10300 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10301 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10302 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10303 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10304 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10306 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10307 sizeof(struct io_uring_rsrc_update));
10308 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10309 sizeof(struct io_uring_rsrc_update2));
10310 /* should fit into one byte */
10311 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10313 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10314 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10316 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10320 __initcall(io_uring_init);