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>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
92 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
101 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
102 IORING_REGISTER_LAST + IORING_OP_LAST)
104 #define IO_RSRC_TAG_TABLE_SHIFT 9
105 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
106 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
108 #define IORING_MAX_REG_BUFFERS (1U << 14)
110 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
111 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
113 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
114 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
116 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
119 u32 head ____cacheline_aligned_in_smp;
120 u32 tail ____cacheline_aligned_in_smp;
124 * This data is shared with the application through the mmap at offsets
125 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
127 * The offsets to the member fields are published through struct
128 * io_sqring_offsets when calling io_uring_setup.
132 * Head and tail offsets into the ring; the offsets need to be
133 * masked to get valid indices.
135 * The kernel controls head of the sq ring and the tail of the cq ring,
136 * and the application controls tail of the sq ring and the head of the
139 struct io_uring sq, cq;
141 * Bitmasks to apply to head and tail offsets (constant, equals
144 u32 sq_ring_mask, cq_ring_mask;
145 /* Ring sizes (constant, power of 2) */
146 u32 sq_ring_entries, cq_ring_entries;
148 * Number of invalid entries dropped by the kernel due to
149 * invalid index stored in array
151 * Written by the kernel, shouldn't be modified by the
152 * application (i.e. get number of "new events" by comparing to
155 * After a new SQ head value was read by the application this
156 * counter includes all submissions that were dropped reaching
157 * the new SQ head (and possibly more).
163 * Written by the kernel, shouldn't be modified by the
166 * The application needs a full memory barrier before checking
167 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
173 * Written by the application, shouldn't be modified by the
178 * Number of completion events lost because the queue was full;
179 * this should be avoided by the application by making sure
180 * there are not more requests pending than there is space in
181 * the completion queue.
183 * Written by the kernel, shouldn't be modified by the
184 * application (i.e. get number of "new events" by comparing to
187 * As completion events come in out of order this counter is not
188 * ordered with any other data.
192 * Ring buffer of completion events.
194 * The kernel writes completion events fresh every time they are
195 * produced, so the application is allowed to modify pending
198 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
201 enum io_uring_cmd_flags {
202 IO_URING_F_NONBLOCK = 1,
203 IO_URING_F_COMPLETE_DEFER = 2,
206 struct io_mapped_ubuf {
209 unsigned int nr_bvecs;
210 unsigned long acct_pages;
211 struct bio_vec bvec[];
216 struct io_overflow_cqe {
217 struct io_uring_cqe cqe;
218 struct list_head list;
221 struct io_fixed_file {
222 /* file * with additional FFS_* flags */
223 unsigned long file_ptr;
227 struct list_head list;
232 struct io_mapped_ubuf *buf;
236 struct io_file_table {
237 /* two level table */
238 struct io_fixed_file **files;
241 struct io_rsrc_node {
242 struct percpu_ref refs;
243 struct list_head node;
244 struct list_head rsrc_list;
245 struct io_rsrc_data *rsrc_data;
246 struct llist_node llist;
250 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
252 struct io_rsrc_data {
253 struct io_ring_ctx *ctx;
259 struct completion done;
264 struct list_head list;
270 struct io_restriction {
271 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
272 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
273 u8 sqe_flags_allowed;
274 u8 sqe_flags_required;
279 IO_SQ_THREAD_SHOULD_STOP = 0,
280 IO_SQ_THREAD_SHOULD_PARK,
285 atomic_t park_pending;
288 /* ctx's that are using this sqd */
289 struct list_head ctx_list;
291 struct task_struct *thread;
292 struct wait_queue_head wait;
294 unsigned sq_thread_idle;
300 struct completion exited;
303 #define IO_IOPOLL_BATCH 8
304 #define IO_COMPL_BATCH 32
305 #define IO_REQ_CACHE_SIZE 32
306 #define IO_REQ_ALLOC_BATCH 8
308 struct io_comp_state {
309 struct io_kiocb *reqs[IO_COMPL_BATCH];
311 /* inline/task_work completion list, under ->uring_lock */
312 struct list_head free_list;
315 struct io_submit_link {
316 struct io_kiocb *head;
317 struct io_kiocb *last;
320 struct io_submit_state {
321 struct blk_plug plug;
322 struct io_submit_link link;
325 * io_kiocb alloc cache
327 void *reqs[IO_REQ_CACHE_SIZE];
328 unsigned int free_reqs;
333 * Batch completion logic
335 struct io_comp_state comp;
338 * File reference cache
342 unsigned int file_refs;
343 unsigned int ios_left;
347 /* const or read-mostly hot data */
349 struct percpu_ref refs;
351 struct io_rings *rings;
353 unsigned int compat: 1;
354 unsigned int drain_next: 1;
355 unsigned int eventfd_async: 1;
356 unsigned int restricted: 1;
357 unsigned int off_timeout_used: 1;
358 unsigned int drain_active: 1;
359 } ____cacheline_aligned_in_smp;
361 /* submission data */
363 struct mutex uring_lock;
366 * Ring buffer of indices into array of io_uring_sqe, which is
367 * mmapped by the application using the IORING_OFF_SQES offset.
369 * This indirection could e.g. be used to assign fixed
370 * io_uring_sqe entries to operations and only submit them to
371 * the queue when needed.
373 * The kernel modifies neither the indices array nor the entries
377 struct io_uring_sqe *sq_sqes;
378 unsigned cached_sq_head;
380 struct list_head defer_list;
383 * Fixed resources fast path, should be accessed only under
384 * uring_lock, and updated through io_uring_register(2)
386 struct io_rsrc_node *rsrc_node;
387 struct io_file_table file_table;
388 unsigned nr_user_files;
389 unsigned nr_user_bufs;
390 struct io_mapped_ubuf **user_bufs;
392 struct io_submit_state submit_state;
393 struct list_head timeout_list;
394 struct list_head cq_overflow_list;
395 struct xarray io_buffers;
396 struct xarray personalities;
398 unsigned sq_thread_idle;
399 } ____cacheline_aligned_in_smp;
401 /* IRQ completion list, under ->completion_lock */
402 struct list_head locked_free_list;
403 unsigned int locked_free_nr;
405 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
406 struct io_sq_data *sq_data; /* if using sq thread polling */
408 struct wait_queue_head sqo_sq_wait;
409 struct list_head sqd_list;
411 unsigned long check_cq_overflow;
414 unsigned cached_cq_tail;
416 struct eventfd_ctx *cq_ev_fd;
417 struct wait_queue_head poll_wait;
418 struct wait_queue_head cq_wait;
420 atomic_t cq_timeouts;
421 struct fasync_struct *cq_fasync;
422 unsigned cq_last_tm_flush;
423 } ____cacheline_aligned_in_smp;
426 spinlock_t completion_lock;
429 * ->iopoll_list is protected by the ctx->uring_lock for
430 * io_uring instances that don't use IORING_SETUP_SQPOLL.
431 * For SQPOLL, only the single threaded io_sq_thread() will
432 * manipulate the list, hence no extra locking is needed there.
434 struct list_head iopoll_list;
435 struct hlist_head *cancel_hash;
436 unsigned cancel_hash_bits;
437 bool poll_multi_queue;
438 } ____cacheline_aligned_in_smp;
440 struct io_restriction restrictions;
442 /* slow path rsrc auxilary data, used by update/register */
444 struct io_rsrc_node *rsrc_backup_node;
445 struct io_mapped_ubuf *dummy_ubuf;
446 struct io_rsrc_data *file_data;
447 struct io_rsrc_data *buf_data;
449 struct delayed_work rsrc_put_work;
450 struct llist_head rsrc_put_llist;
451 struct list_head rsrc_ref_list;
452 spinlock_t rsrc_ref_lock;
455 /* Keep this last, we don't need it for the fast path */
457 #if defined(CONFIG_UNIX)
458 struct socket *ring_sock;
460 /* hashed buffered write serialization */
461 struct io_wq_hash *hash_map;
463 /* Only used for accounting purposes */
464 struct user_struct *user;
465 struct mm_struct *mm_account;
467 /* ctx exit and cancelation */
468 struct llist_head fallback_llist;
469 struct delayed_work fallback_work;
470 struct work_struct exit_work;
471 struct list_head tctx_list;
472 struct completion ref_comp;
476 struct io_uring_task {
477 /* submission side */
480 struct wait_queue_head wait;
481 const struct io_ring_ctx *last;
483 struct percpu_counter inflight;
484 atomic_t inflight_tracked;
487 spinlock_t task_lock;
488 struct io_wq_work_list task_list;
489 unsigned long task_state;
490 struct callback_head task_work;
494 * First field must be the file pointer in all the
495 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
497 struct io_poll_iocb {
499 struct wait_queue_head *head;
503 struct wait_queue_entry wait;
506 struct io_poll_update {
512 bool update_user_data;
520 struct io_timeout_data {
521 struct io_kiocb *req;
522 struct hrtimer timer;
523 struct timespec64 ts;
524 enum hrtimer_mode mode;
529 struct sockaddr __user *addr;
530 int __user *addr_len;
532 unsigned long nofile;
552 struct list_head list;
553 /* head of the link, used by linked timeouts only */
554 struct io_kiocb *head;
557 struct io_timeout_rem {
562 struct timespec64 ts;
567 /* NOTE: kiocb has the file as the first member, so don't do it here */
575 struct sockaddr __user *addr;
582 struct compat_msghdr __user *umsg_compat;
583 struct user_msghdr __user *umsg;
589 struct io_buffer *kbuf;
595 struct filename *filename;
597 unsigned long nofile;
600 struct io_rsrc_update {
626 struct epoll_event event;
630 struct file *file_out;
631 struct file *file_in;
638 struct io_provide_buf {
652 const char __user *filename;
653 struct statx __user *buffer;
665 struct filename *oldpath;
666 struct filename *newpath;
674 struct filename *filename;
677 struct io_completion {
679 struct list_head list;
683 struct io_async_connect {
684 struct sockaddr_storage address;
687 struct io_async_msghdr {
688 struct iovec fast_iov[UIO_FASTIOV];
689 /* points to an allocated iov, if NULL we use fast_iov instead */
690 struct iovec *free_iov;
691 struct sockaddr __user *uaddr;
693 struct sockaddr_storage addr;
697 struct iovec fast_iov[UIO_FASTIOV];
698 const struct iovec *free_iovec;
699 struct iov_iter iter;
701 struct wait_page_queue wpq;
705 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
706 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
707 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
708 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
709 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
710 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
712 /* first byte is taken by user flags, shift it to not overlap */
717 REQ_F_LINK_TIMEOUT_BIT,
718 REQ_F_NEED_CLEANUP_BIT,
720 REQ_F_BUFFER_SELECTED_BIT,
721 REQ_F_LTIMEOUT_ACTIVE_BIT,
722 REQ_F_COMPLETE_INLINE_BIT,
724 REQ_F_DONT_REISSUE_BIT,
726 /* keep async read/write and isreg together and in order */
727 REQ_F_ASYNC_READ_BIT,
728 REQ_F_ASYNC_WRITE_BIT,
731 /* not a real bit, just to check we're not overflowing the space */
737 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
738 /* drain existing IO first */
739 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
741 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
742 /* doesn't sever on completion < 0 */
743 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
745 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
746 /* IOSQE_BUFFER_SELECT */
747 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
749 /* fail rest of links */
750 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
751 /* on inflight list, should be cancelled and waited on exit reliably */
752 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
753 /* read/write uses file position */
754 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
755 /* must not punt to workers */
756 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
757 /* has or had linked timeout */
758 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
760 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
761 /* already went through poll handler */
762 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
763 /* buffer already selected */
764 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
765 /* linked timeout is active, i.e. prepared by link's head */
766 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
767 /* completion is deferred through io_comp_state */
768 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
769 /* caller should reissue async */
770 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
771 /* don't attempt request reissue, see io_rw_reissue() */
772 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
773 /* supports async reads */
774 REQ_F_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
775 /* supports async writes */
776 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
778 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
779 /* has creds assigned */
780 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
784 struct io_poll_iocb poll;
785 struct io_poll_iocb *double_poll;
788 typedef void (*io_req_tw_func_t)(struct io_kiocb *req);
790 struct io_task_work {
792 struct io_wq_work_node node;
793 struct llist_node fallback_node;
795 io_req_tw_func_t func;
799 IORING_RSRC_FILE = 0,
800 IORING_RSRC_BUFFER = 1,
804 * NOTE! Each of the iocb union members has the file pointer
805 * as the first entry in their struct definition. So you can
806 * access the file pointer through any of the sub-structs,
807 * or directly as just 'ki_filp' in this struct.
813 struct io_poll_iocb poll;
814 struct io_poll_update poll_update;
815 struct io_accept accept;
817 struct io_cancel cancel;
818 struct io_timeout timeout;
819 struct io_timeout_rem timeout_rem;
820 struct io_connect connect;
821 struct io_sr_msg sr_msg;
823 struct io_close close;
824 struct io_rsrc_update rsrc_update;
825 struct io_fadvise fadvise;
826 struct io_madvise madvise;
827 struct io_epoll epoll;
828 struct io_splice splice;
829 struct io_provide_buf pbuf;
830 struct io_statx statx;
831 struct io_shutdown shutdown;
832 struct io_rename rename;
833 struct io_unlink unlink;
834 /* use only after cleaning per-op data, see io_clean_op() */
835 struct io_completion compl;
838 /* opcode allocated if it needs to store data for async defer */
841 /* polled IO has completed */
847 struct io_ring_ctx *ctx;
850 struct task_struct *task;
853 struct io_kiocb *link;
854 struct percpu_ref *fixed_rsrc_refs;
856 /* used with ctx->iopoll_list with reads/writes */
857 struct list_head inflight_entry;
858 struct io_task_work io_task_work;
859 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
860 struct hlist_node hash_node;
861 struct async_poll *apoll;
862 struct io_wq_work work;
863 const struct cred *creds;
865 /* store used ubuf, so we can prevent reloading */
866 struct io_mapped_ubuf *imu;
869 struct io_tctx_node {
870 struct list_head ctx_node;
871 struct task_struct *task;
872 struct io_ring_ctx *ctx;
875 struct io_defer_entry {
876 struct list_head list;
877 struct io_kiocb *req;
882 /* needs req->file assigned */
883 unsigned needs_file : 1;
884 /* hash wq insertion if file is a regular file */
885 unsigned hash_reg_file : 1;
886 /* unbound wq insertion if file is a non-regular file */
887 unsigned unbound_nonreg_file : 1;
888 /* opcode is not supported by this kernel */
889 unsigned not_supported : 1;
890 /* set if opcode supports polled "wait" */
892 unsigned pollout : 1;
893 /* op supports buffer selection */
894 unsigned buffer_select : 1;
895 /* do prep async if is going to be punted */
896 unsigned needs_async_setup : 1;
897 /* should block plug */
899 /* size of async data needed, if any */
900 unsigned short async_size;
903 static const struct io_op_def io_op_defs[] = {
904 [IORING_OP_NOP] = {},
905 [IORING_OP_READV] = {
907 .unbound_nonreg_file = 1,
910 .needs_async_setup = 1,
912 .async_size = sizeof(struct io_async_rw),
914 [IORING_OP_WRITEV] = {
917 .unbound_nonreg_file = 1,
919 .needs_async_setup = 1,
921 .async_size = sizeof(struct io_async_rw),
923 [IORING_OP_FSYNC] = {
926 [IORING_OP_READ_FIXED] = {
928 .unbound_nonreg_file = 1,
931 .async_size = sizeof(struct io_async_rw),
933 [IORING_OP_WRITE_FIXED] = {
936 .unbound_nonreg_file = 1,
939 .async_size = sizeof(struct io_async_rw),
941 [IORING_OP_POLL_ADD] = {
943 .unbound_nonreg_file = 1,
945 [IORING_OP_POLL_REMOVE] = {},
946 [IORING_OP_SYNC_FILE_RANGE] = {
949 [IORING_OP_SENDMSG] = {
951 .unbound_nonreg_file = 1,
953 .needs_async_setup = 1,
954 .async_size = sizeof(struct io_async_msghdr),
956 [IORING_OP_RECVMSG] = {
958 .unbound_nonreg_file = 1,
961 .needs_async_setup = 1,
962 .async_size = sizeof(struct io_async_msghdr),
964 [IORING_OP_TIMEOUT] = {
965 .async_size = sizeof(struct io_timeout_data),
967 [IORING_OP_TIMEOUT_REMOVE] = {
968 /* used by timeout updates' prep() */
970 [IORING_OP_ACCEPT] = {
972 .unbound_nonreg_file = 1,
975 [IORING_OP_ASYNC_CANCEL] = {},
976 [IORING_OP_LINK_TIMEOUT] = {
977 .async_size = sizeof(struct io_timeout_data),
979 [IORING_OP_CONNECT] = {
981 .unbound_nonreg_file = 1,
983 .needs_async_setup = 1,
984 .async_size = sizeof(struct io_async_connect),
986 [IORING_OP_FALLOCATE] = {
989 [IORING_OP_OPENAT] = {},
990 [IORING_OP_CLOSE] = {},
991 [IORING_OP_FILES_UPDATE] = {},
992 [IORING_OP_STATX] = {},
995 .unbound_nonreg_file = 1,
999 .async_size = sizeof(struct io_async_rw),
1001 [IORING_OP_WRITE] = {
1003 .unbound_nonreg_file = 1,
1006 .async_size = sizeof(struct io_async_rw),
1008 [IORING_OP_FADVISE] = {
1011 [IORING_OP_MADVISE] = {},
1012 [IORING_OP_SEND] = {
1014 .unbound_nonreg_file = 1,
1017 [IORING_OP_RECV] = {
1019 .unbound_nonreg_file = 1,
1023 [IORING_OP_OPENAT2] = {
1025 [IORING_OP_EPOLL_CTL] = {
1026 .unbound_nonreg_file = 1,
1028 [IORING_OP_SPLICE] = {
1031 .unbound_nonreg_file = 1,
1033 [IORING_OP_PROVIDE_BUFFERS] = {},
1034 [IORING_OP_REMOVE_BUFFERS] = {},
1038 .unbound_nonreg_file = 1,
1040 [IORING_OP_SHUTDOWN] = {
1043 [IORING_OP_RENAMEAT] = {},
1044 [IORING_OP_UNLINKAT] = {},
1047 static bool io_disarm_next(struct io_kiocb *req);
1048 static void io_uring_del_tctx_node(unsigned long index);
1049 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1050 struct task_struct *task,
1052 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1053 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1055 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1056 long res, unsigned int cflags);
1057 static void io_put_req(struct io_kiocb *req);
1058 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1059 static void io_dismantle_req(struct io_kiocb *req);
1060 static void io_put_task(struct task_struct *task, int nr);
1061 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1062 static void io_queue_linked_timeout(struct io_kiocb *req);
1063 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1064 struct io_uring_rsrc_update2 *up,
1066 static void io_clean_op(struct io_kiocb *req);
1067 static struct file *io_file_get(struct io_submit_state *state,
1068 struct io_kiocb *req, int fd, bool fixed);
1069 static void __io_queue_sqe(struct io_kiocb *req);
1070 static void io_rsrc_put_work(struct work_struct *work);
1072 static void io_req_task_queue(struct io_kiocb *req);
1073 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1074 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1075 static int io_req_prep_async(struct io_kiocb *req);
1077 static void io_fallback_req_func(struct work_struct *unused);
1079 static struct kmem_cache *req_cachep;
1081 static const struct file_operations io_uring_fops;
1083 struct sock *io_uring_get_socket(struct file *file)
1085 #if defined(CONFIG_UNIX)
1086 if (file->f_op == &io_uring_fops) {
1087 struct io_ring_ctx *ctx = file->private_data;
1089 return ctx->ring_sock->sk;
1094 EXPORT_SYMBOL(io_uring_get_socket);
1096 #define io_for_each_link(pos, head) \
1097 for (pos = (head); pos; pos = pos->link)
1099 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1101 struct io_ring_ctx *ctx = req->ctx;
1103 if (!req->fixed_rsrc_refs) {
1104 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1105 percpu_ref_get(req->fixed_rsrc_refs);
1109 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1111 bool got = percpu_ref_tryget(ref);
1113 /* already at zero, wait for ->release() */
1115 wait_for_completion(compl);
1116 percpu_ref_resurrect(ref);
1118 percpu_ref_put(ref);
1121 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1124 struct io_kiocb *req;
1126 if (task && head->task != task)
1131 io_for_each_link(req, head) {
1132 if (req->flags & REQ_F_INFLIGHT)
1138 static inline void req_set_fail(struct io_kiocb *req)
1140 req->flags |= REQ_F_FAIL;
1143 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1145 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1147 complete(&ctx->ref_comp);
1150 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1152 return !req->timeout.off;
1155 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1157 struct io_ring_ctx *ctx;
1160 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1165 * Use 5 bits less than the max cq entries, that should give us around
1166 * 32 entries per hash list if totally full and uniformly spread.
1168 hash_bits = ilog2(p->cq_entries);
1172 ctx->cancel_hash_bits = hash_bits;
1173 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1175 if (!ctx->cancel_hash)
1177 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1179 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1180 if (!ctx->dummy_ubuf)
1182 /* set invalid range, so io_import_fixed() fails meeting it */
1183 ctx->dummy_ubuf->ubuf = -1UL;
1185 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1186 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1189 ctx->flags = p->flags;
1190 init_waitqueue_head(&ctx->sqo_sq_wait);
1191 INIT_LIST_HEAD(&ctx->sqd_list);
1192 init_waitqueue_head(&ctx->poll_wait);
1193 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1194 init_completion(&ctx->ref_comp);
1195 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1196 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1197 mutex_init(&ctx->uring_lock);
1198 init_waitqueue_head(&ctx->cq_wait);
1199 spin_lock_init(&ctx->completion_lock);
1200 INIT_LIST_HEAD(&ctx->iopoll_list);
1201 INIT_LIST_HEAD(&ctx->defer_list);
1202 INIT_LIST_HEAD(&ctx->timeout_list);
1203 spin_lock_init(&ctx->rsrc_ref_lock);
1204 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1205 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1206 init_llist_head(&ctx->rsrc_put_llist);
1207 INIT_LIST_HEAD(&ctx->tctx_list);
1208 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1209 INIT_LIST_HEAD(&ctx->locked_free_list);
1210 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1213 kfree(ctx->dummy_ubuf);
1214 kfree(ctx->cancel_hash);
1219 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1221 struct io_rings *r = ctx->rings;
1223 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1227 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1229 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1230 struct io_ring_ctx *ctx = req->ctx;
1232 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1238 static void io_req_track_inflight(struct io_kiocb *req)
1240 if (!(req->flags & REQ_F_INFLIGHT)) {
1241 req->flags |= REQ_F_INFLIGHT;
1242 atomic_inc(¤t->io_uring->inflight_tracked);
1246 static void io_prep_async_work(struct io_kiocb *req)
1248 const struct io_op_def *def = &io_op_defs[req->opcode];
1249 struct io_ring_ctx *ctx = req->ctx;
1251 if (!(req->flags & REQ_F_CREDS)) {
1252 req->flags |= REQ_F_CREDS;
1253 req->creds = get_current_cred();
1256 req->work.list.next = NULL;
1257 req->work.flags = 0;
1258 if (req->flags & REQ_F_FORCE_ASYNC)
1259 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1261 if (req->flags & REQ_F_ISREG) {
1262 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1263 io_wq_hash_work(&req->work, file_inode(req->file));
1264 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1265 if (def->unbound_nonreg_file)
1266 req->work.flags |= IO_WQ_WORK_UNBOUND;
1269 switch (req->opcode) {
1270 case IORING_OP_SPLICE:
1272 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1273 req->work.flags |= IO_WQ_WORK_UNBOUND;
1278 static void io_prep_async_link(struct io_kiocb *req)
1280 struct io_kiocb *cur;
1282 io_for_each_link(cur, req)
1283 io_prep_async_work(cur);
1286 static void io_queue_async_work(struct io_kiocb *req)
1288 struct io_ring_ctx *ctx = req->ctx;
1289 struct io_kiocb *link = io_prep_linked_timeout(req);
1290 struct io_uring_task *tctx = req->task->io_uring;
1293 BUG_ON(!tctx->io_wq);
1295 /* init ->work of the whole link before punting */
1296 io_prep_async_link(req);
1297 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1298 &req->work, req->flags);
1299 io_wq_enqueue(tctx->io_wq, &req->work);
1301 io_queue_linked_timeout(link);
1304 static void io_kill_timeout(struct io_kiocb *req, int status)
1305 __must_hold(&req->ctx->completion_lock)
1307 struct io_timeout_data *io = req->async_data;
1309 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1310 atomic_set(&req->ctx->cq_timeouts,
1311 atomic_read(&req->ctx->cq_timeouts) + 1);
1312 list_del_init(&req->timeout.list);
1313 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1314 io_put_req_deferred(req, 1);
1318 static void io_queue_deferred(struct io_ring_ctx *ctx)
1320 while (!list_empty(&ctx->defer_list)) {
1321 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1322 struct io_defer_entry, list);
1324 if (req_need_defer(de->req, de->seq))
1326 list_del_init(&de->list);
1327 io_req_task_queue(de->req);
1332 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1334 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1336 while (!list_empty(&ctx->timeout_list)) {
1337 u32 events_needed, events_got;
1338 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1339 struct io_kiocb, timeout.list);
1341 if (io_is_timeout_noseq(req))
1345 * Since seq can easily wrap around over time, subtract
1346 * the last seq at which timeouts were flushed before comparing.
1347 * Assuming not more than 2^31-1 events have happened since,
1348 * these subtractions won't have wrapped, so we can check if
1349 * target is in [last_seq, current_seq] by comparing the two.
1351 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1352 events_got = seq - ctx->cq_last_tm_flush;
1353 if (events_got < events_needed)
1356 list_del_init(&req->timeout.list);
1357 io_kill_timeout(req, 0);
1359 ctx->cq_last_tm_flush = seq;
1362 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1364 if (ctx->off_timeout_used)
1365 io_flush_timeouts(ctx);
1366 if (ctx->drain_active)
1367 io_queue_deferred(ctx);
1370 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1372 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1373 __io_commit_cqring_flush(ctx);
1374 /* order cqe stores with ring update */
1375 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1378 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1380 struct io_rings *r = ctx->rings;
1382 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1385 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1387 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1390 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1392 struct io_rings *rings = ctx->rings;
1393 unsigned tail, mask = ctx->cq_entries - 1;
1396 * writes to the cq entry need to come after reading head; the
1397 * control dependency is enough as we're using WRITE_ONCE to
1400 if (__io_cqring_events(ctx) == ctx->cq_entries)
1403 tail = ctx->cached_cq_tail++;
1404 return &rings->cqes[tail & mask];
1407 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1409 if (likely(!ctx->cq_ev_fd))
1411 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1413 return !ctx->eventfd_async || io_wq_current_is_worker();
1416 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1418 /* see waitqueue_active() comment */
1421 if (waitqueue_active(&ctx->cq_wait))
1422 wake_up(&ctx->cq_wait);
1423 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1424 wake_up(&ctx->sq_data->wait);
1425 if (io_should_trigger_evfd(ctx))
1426 eventfd_signal(ctx->cq_ev_fd, 1);
1427 if (waitqueue_active(&ctx->poll_wait)) {
1428 wake_up_interruptible(&ctx->poll_wait);
1429 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1433 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1435 /* see waitqueue_active() comment */
1438 if (ctx->flags & IORING_SETUP_SQPOLL) {
1439 if (waitqueue_active(&ctx->cq_wait))
1440 wake_up(&ctx->cq_wait);
1442 if (io_should_trigger_evfd(ctx))
1443 eventfd_signal(ctx->cq_ev_fd, 1);
1444 if (waitqueue_active(&ctx->poll_wait)) {
1445 wake_up_interruptible(&ctx->poll_wait);
1446 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1450 /* Returns true if there are no backlogged entries after the flush */
1451 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1453 unsigned long flags;
1454 bool all_flushed, posted;
1456 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1460 spin_lock_irqsave(&ctx->completion_lock, flags);
1461 while (!list_empty(&ctx->cq_overflow_list)) {
1462 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1463 struct io_overflow_cqe *ocqe;
1467 ocqe = list_first_entry(&ctx->cq_overflow_list,
1468 struct io_overflow_cqe, list);
1470 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1472 io_account_cq_overflow(ctx);
1475 list_del(&ocqe->list);
1479 all_flushed = list_empty(&ctx->cq_overflow_list);
1481 clear_bit(0, &ctx->check_cq_overflow);
1482 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1486 io_commit_cqring(ctx);
1487 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1489 io_cqring_ev_posted(ctx);
1493 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1497 if (test_bit(0, &ctx->check_cq_overflow)) {
1498 /* iopoll syncs against uring_lock, not completion_lock */
1499 if (ctx->flags & IORING_SETUP_IOPOLL)
1500 mutex_lock(&ctx->uring_lock);
1501 ret = __io_cqring_overflow_flush(ctx, force);
1502 if (ctx->flags & IORING_SETUP_IOPOLL)
1503 mutex_unlock(&ctx->uring_lock);
1510 * Shamelessly stolen from the mm implementation of page reference checking,
1511 * see commit f958d7b528b1 for details.
1513 #define req_ref_zero_or_close_to_overflow(req) \
1514 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1516 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1518 return atomic_inc_not_zero(&req->refs);
1521 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1523 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1524 return atomic_sub_and_test(refs, &req->refs);
1527 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1529 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1530 return atomic_dec_and_test(&req->refs);
1533 static inline void req_ref_put(struct io_kiocb *req)
1535 WARN_ON_ONCE(req_ref_put_and_test(req));
1538 static inline void req_ref_get(struct io_kiocb *req)
1540 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1541 atomic_inc(&req->refs);
1544 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1545 long res, unsigned int cflags)
1547 struct io_overflow_cqe *ocqe;
1549 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1552 * If we're in ring overflow flush mode, or in task cancel mode,
1553 * or cannot allocate an overflow entry, then we need to drop it
1556 io_account_cq_overflow(ctx);
1559 if (list_empty(&ctx->cq_overflow_list)) {
1560 set_bit(0, &ctx->check_cq_overflow);
1561 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1563 ocqe->cqe.user_data = user_data;
1564 ocqe->cqe.res = res;
1565 ocqe->cqe.flags = cflags;
1566 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1570 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1571 long res, unsigned int cflags)
1573 struct io_uring_cqe *cqe;
1575 trace_io_uring_complete(ctx, user_data, res, cflags);
1578 * If we can't get a cq entry, userspace overflowed the
1579 * submission (by quite a lot). Increment the overflow count in
1582 cqe = io_get_cqe(ctx);
1584 WRITE_ONCE(cqe->user_data, user_data);
1585 WRITE_ONCE(cqe->res, res);
1586 WRITE_ONCE(cqe->flags, cflags);
1589 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1592 /* not as hot to bloat with inlining */
1593 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1594 long res, unsigned int cflags)
1596 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1599 static void io_req_complete_post(struct io_kiocb *req, long res,
1600 unsigned int cflags)
1602 struct io_ring_ctx *ctx = req->ctx;
1603 unsigned long flags;
1605 spin_lock_irqsave(&ctx->completion_lock, flags);
1606 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1608 * If we're the last reference to this request, add to our locked
1611 if (req_ref_put_and_test(req)) {
1612 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1613 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1614 io_disarm_next(req);
1616 io_req_task_queue(req->link);
1620 io_dismantle_req(req);
1621 io_put_task(req->task, 1);
1622 list_add(&req->compl.list, &ctx->locked_free_list);
1623 ctx->locked_free_nr++;
1625 if (!percpu_ref_tryget(&ctx->refs))
1628 io_commit_cqring(ctx);
1629 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1632 io_cqring_ev_posted(ctx);
1633 percpu_ref_put(&ctx->refs);
1637 static inline bool io_req_needs_clean(struct io_kiocb *req)
1639 return req->flags & IO_REQ_CLEAN_FLAGS;
1642 static void io_req_complete_state(struct io_kiocb *req, long res,
1643 unsigned int cflags)
1645 if (io_req_needs_clean(req))
1648 req->compl.cflags = cflags;
1649 req->flags |= REQ_F_COMPLETE_INLINE;
1652 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1653 long res, unsigned cflags)
1655 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1656 io_req_complete_state(req, res, cflags);
1658 io_req_complete_post(req, res, cflags);
1661 static inline void io_req_complete(struct io_kiocb *req, long res)
1663 __io_req_complete(req, 0, res, 0);
1666 static void io_req_complete_failed(struct io_kiocb *req, long res)
1670 io_req_complete_post(req, res, 0);
1673 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1674 struct io_comp_state *cs)
1676 spin_lock_irq(&ctx->completion_lock);
1677 list_splice_init(&ctx->locked_free_list, &cs->free_list);
1678 ctx->locked_free_nr = 0;
1679 spin_unlock_irq(&ctx->completion_lock);
1682 /* Returns true IFF there are requests in the cache */
1683 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1685 struct io_submit_state *state = &ctx->submit_state;
1686 struct io_comp_state *cs = &state->comp;
1690 * If we have more than a batch's worth of requests in our IRQ side
1691 * locked cache, grab the lock and move them over to our submission
1694 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1695 io_flush_cached_locked_reqs(ctx, cs);
1697 nr = state->free_reqs;
1698 while (!list_empty(&cs->free_list)) {
1699 struct io_kiocb *req = list_first_entry(&cs->free_list,
1700 struct io_kiocb, compl.list);
1702 list_del(&req->compl.list);
1703 state->reqs[nr++] = req;
1704 if (nr == ARRAY_SIZE(state->reqs))
1708 state->free_reqs = nr;
1712 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1714 struct io_submit_state *state = &ctx->submit_state;
1716 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1718 if (!state->free_reqs) {
1719 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1722 if (io_flush_cached_reqs(ctx))
1725 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1729 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1730 * retry single alloc to be on the safe side.
1732 if (unlikely(ret <= 0)) {
1733 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1734 if (!state->reqs[0])
1740 * Don't initialise the fields below on every allocation, but
1741 * do that in advance and keep valid on free.
1743 for (i = 0; i < ret; i++) {
1744 struct io_kiocb *req = state->reqs[i];
1748 req->async_data = NULL;
1749 /* not necessary, but safer to zero */
1752 state->free_reqs = ret;
1756 return state->reqs[state->free_reqs];
1759 static inline void io_put_file(struct file *file)
1765 static void io_dismantle_req(struct io_kiocb *req)
1767 unsigned int flags = req->flags;
1769 if (io_req_needs_clean(req))
1771 if (!(flags & REQ_F_FIXED_FILE))
1772 io_put_file(req->file);
1773 if (req->fixed_rsrc_refs)
1774 percpu_ref_put(req->fixed_rsrc_refs);
1775 if (req->async_data) {
1776 kfree(req->async_data);
1777 req->async_data = NULL;
1781 /* must to be called somewhat shortly after putting a request */
1782 static inline void io_put_task(struct task_struct *task, int nr)
1784 struct io_uring_task *tctx = task->io_uring;
1786 percpu_counter_sub(&tctx->inflight, nr);
1787 if (unlikely(atomic_read(&tctx->in_idle)))
1788 wake_up(&tctx->wait);
1789 put_task_struct_many(task, nr);
1792 static void __io_free_req(struct io_kiocb *req)
1794 struct io_ring_ctx *ctx = req->ctx;
1796 io_dismantle_req(req);
1797 io_put_task(req->task, 1);
1799 kmem_cache_free(req_cachep, req);
1800 percpu_ref_put(&ctx->refs);
1803 static inline void io_remove_next_linked(struct io_kiocb *req)
1805 struct io_kiocb *nxt = req->link;
1807 req->link = nxt->link;
1811 static bool io_kill_linked_timeout(struct io_kiocb *req)
1812 __must_hold(&req->ctx->completion_lock)
1814 struct io_kiocb *link = req->link;
1817 * Can happen if a linked timeout fired and link had been like
1818 * req -> link t-out -> link t-out [-> ...]
1820 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1821 struct io_timeout_data *io = link->async_data;
1823 io_remove_next_linked(req);
1824 link->timeout.head = NULL;
1825 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1826 io_cqring_fill_event(link->ctx, link->user_data,
1828 io_put_req_deferred(link, 1);
1835 static void io_fail_links(struct io_kiocb *req)
1836 __must_hold(&req->ctx->completion_lock)
1838 struct io_kiocb *nxt, *link = req->link;
1845 trace_io_uring_fail_link(req, link);
1846 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1847 io_put_req_deferred(link, 2);
1852 static bool io_disarm_next(struct io_kiocb *req)
1853 __must_hold(&req->ctx->completion_lock)
1855 bool posted = false;
1857 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1858 posted = io_kill_linked_timeout(req);
1859 if (unlikely((req->flags & REQ_F_FAIL) &&
1860 !(req->flags & REQ_F_HARDLINK))) {
1861 posted |= (req->link != NULL);
1867 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1869 struct io_kiocb *nxt;
1872 * If LINK is set, we have dependent requests in this chain. If we
1873 * didn't fail this request, queue the first one up, moving any other
1874 * dependencies to the next request. In case of failure, fail the rest
1877 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1878 struct io_ring_ctx *ctx = req->ctx;
1879 unsigned long flags;
1882 spin_lock_irqsave(&ctx->completion_lock, flags);
1883 posted = io_disarm_next(req);
1885 io_commit_cqring(req->ctx);
1886 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1888 io_cqring_ev_posted(ctx);
1895 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1897 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1899 return __io_req_find_next(req);
1902 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1906 if (ctx->submit_state.comp.nr) {
1907 mutex_lock(&ctx->uring_lock);
1908 io_submit_flush_completions(ctx);
1909 mutex_unlock(&ctx->uring_lock);
1911 percpu_ref_put(&ctx->refs);
1914 static void tctx_task_work(struct callback_head *cb)
1916 struct io_ring_ctx *ctx = NULL;
1917 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1921 struct io_wq_work_node *node;
1923 spin_lock_irq(&tctx->task_lock);
1924 node = tctx->task_list.first;
1925 INIT_WQ_LIST(&tctx->task_list);
1926 spin_unlock_irq(&tctx->task_lock);
1929 struct io_wq_work_node *next = node->next;
1930 struct io_kiocb *req = container_of(node, struct io_kiocb,
1933 if (req->ctx != ctx) {
1934 ctx_flush_and_put(ctx);
1936 percpu_ref_get(&ctx->refs);
1938 req->io_task_work.func(req);
1941 if (wq_list_empty(&tctx->task_list)) {
1942 clear_bit(0, &tctx->task_state);
1943 if (wq_list_empty(&tctx->task_list))
1945 /* another tctx_task_work() is enqueued, yield */
1946 if (test_and_set_bit(0, &tctx->task_state))
1952 ctx_flush_and_put(ctx);
1955 static void io_req_task_work_add(struct io_kiocb *req)
1957 struct task_struct *tsk = req->task;
1958 struct io_uring_task *tctx = tsk->io_uring;
1959 enum task_work_notify_mode notify;
1960 struct io_wq_work_node *node;
1961 unsigned long flags;
1963 WARN_ON_ONCE(!tctx);
1965 spin_lock_irqsave(&tctx->task_lock, flags);
1966 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1967 spin_unlock_irqrestore(&tctx->task_lock, flags);
1969 /* task_work already pending, we're done */
1970 if (test_bit(0, &tctx->task_state) ||
1971 test_and_set_bit(0, &tctx->task_state))
1973 if (unlikely(tsk->flags & PF_EXITING))
1977 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1978 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1979 * processing task_work. There's no reliable way to tell if TWA_RESUME
1982 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1983 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1984 wake_up_process(tsk);
1988 clear_bit(0, &tctx->task_state);
1989 spin_lock_irqsave(&tctx->task_lock, flags);
1990 node = tctx->task_list.first;
1991 INIT_WQ_LIST(&tctx->task_list);
1992 spin_unlock_irqrestore(&tctx->task_lock, flags);
1995 req = container_of(node, struct io_kiocb, io_task_work.node);
1997 if (llist_add(&req->io_task_work.fallback_node,
1998 &req->ctx->fallback_llist))
1999 schedule_delayed_work(&req->ctx->fallback_work, 1);
2003 static void io_req_task_cancel(struct io_kiocb *req)
2005 struct io_ring_ctx *ctx = req->ctx;
2007 /* ctx is guaranteed to stay alive while we hold uring_lock */
2008 mutex_lock(&ctx->uring_lock);
2009 io_req_complete_failed(req, req->result);
2010 mutex_unlock(&ctx->uring_lock);
2013 static void io_req_task_submit(struct io_kiocb *req)
2015 struct io_ring_ctx *ctx = req->ctx;
2017 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2018 mutex_lock(&ctx->uring_lock);
2019 if (!(req->task->flags & PF_EXITING) && !req->task->in_execve)
2020 __io_queue_sqe(req);
2022 io_req_complete_failed(req, -EFAULT);
2023 mutex_unlock(&ctx->uring_lock);
2026 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2029 req->io_task_work.func = io_req_task_cancel;
2030 io_req_task_work_add(req);
2033 static void io_req_task_queue(struct io_kiocb *req)
2035 req->io_task_work.func = io_req_task_submit;
2036 io_req_task_work_add(req);
2039 static inline void io_queue_next(struct io_kiocb *req)
2041 struct io_kiocb *nxt = io_req_find_next(req);
2044 io_req_task_queue(nxt);
2047 static void io_free_req(struct io_kiocb *req)
2054 struct task_struct *task;
2059 static inline void io_init_req_batch(struct req_batch *rb)
2066 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2067 struct req_batch *rb)
2070 io_put_task(rb->task, rb->task_refs);
2072 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2075 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2076 struct io_submit_state *state)
2079 io_dismantle_req(req);
2081 if (req->task != rb->task) {
2083 io_put_task(rb->task, rb->task_refs);
2084 rb->task = req->task;
2090 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2091 state->reqs[state->free_reqs++] = req;
2093 list_add(&req->compl.list, &state->comp.free_list);
2096 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2098 struct io_comp_state *cs = &ctx->submit_state.comp;
2100 struct req_batch rb;
2102 spin_lock_irq(&ctx->completion_lock);
2103 for (i = 0; i < nr; i++) {
2104 struct io_kiocb *req = cs->reqs[i];
2106 __io_cqring_fill_event(ctx, req->user_data, req->result,
2109 io_commit_cqring(ctx);
2110 spin_unlock_irq(&ctx->completion_lock);
2111 io_cqring_ev_posted(ctx);
2113 io_init_req_batch(&rb);
2114 for (i = 0; i < nr; i++) {
2115 struct io_kiocb *req = cs->reqs[i];
2117 /* submission and completion refs */
2118 if (req_ref_sub_and_test(req, 2))
2119 io_req_free_batch(&rb, req, &ctx->submit_state);
2122 io_req_free_batch_finish(ctx, &rb);
2127 * Drop reference to request, return next in chain (if there is one) if this
2128 * was the last reference to this request.
2130 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2132 struct io_kiocb *nxt = NULL;
2134 if (req_ref_put_and_test(req)) {
2135 nxt = io_req_find_next(req);
2141 static inline void io_put_req(struct io_kiocb *req)
2143 if (req_ref_put_and_test(req))
2147 static void io_free_req_deferred(struct io_kiocb *req)
2149 req->io_task_work.func = io_free_req;
2150 io_req_task_work_add(req);
2153 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2155 if (req_ref_sub_and_test(req, refs))
2156 io_free_req_deferred(req);
2159 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2161 /* See comment at the top of this file */
2163 return __io_cqring_events(ctx);
2166 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2168 struct io_rings *rings = ctx->rings;
2170 /* make sure SQ entry isn't read before tail */
2171 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2174 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2176 unsigned int cflags;
2178 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2179 cflags |= IORING_CQE_F_BUFFER;
2180 req->flags &= ~REQ_F_BUFFER_SELECTED;
2185 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2187 struct io_buffer *kbuf;
2189 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2190 return io_put_kbuf(req, kbuf);
2193 static inline bool io_run_task_work(void)
2195 if (current->task_works) {
2196 __set_current_state(TASK_RUNNING);
2205 * Find and free completed poll iocbs
2207 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2208 struct list_head *done)
2210 struct req_batch rb;
2211 struct io_kiocb *req;
2213 /* order with ->result store in io_complete_rw_iopoll() */
2216 io_init_req_batch(&rb);
2217 while (!list_empty(done)) {
2220 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2221 list_del(&req->inflight_entry);
2223 if (READ_ONCE(req->result) == -EAGAIN &&
2224 !(req->flags & REQ_F_DONT_REISSUE)) {
2225 req->iopoll_completed = 0;
2227 io_queue_async_work(req);
2231 if (req->flags & REQ_F_BUFFER_SELECTED)
2232 cflags = io_put_rw_kbuf(req);
2234 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2237 if (req_ref_put_and_test(req))
2238 io_req_free_batch(&rb, req, &ctx->submit_state);
2241 io_commit_cqring(ctx);
2242 io_cqring_ev_posted_iopoll(ctx);
2243 io_req_free_batch_finish(ctx, &rb);
2246 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2249 struct io_kiocb *req, *tmp;
2255 * Only spin for completions if we don't have multiple devices hanging
2256 * off our complete list, and we're under the requested amount.
2258 spin = !ctx->poll_multi_queue && *nr_events < min;
2261 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2262 struct kiocb *kiocb = &req->rw.kiocb;
2265 * Move completed and retryable entries to our local lists.
2266 * If we find a request that requires polling, break out
2267 * and complete those lists first, if we have entries there.
2269 if (READ_ONCE(req->iopoll_completed)) {
2270 list_move_tail(&req->inflight_entry, &done);
2273 if (!list_empty(&done))
2276 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2280 /* iopoll may have completed current req */
2281 if (READ_ONCE(req->iopoll_completed))
2282 list_move_tail(&req->inflight_entry, &done);
2289 if (!list_empty(&done))
2290 io_iopoll_complete(ctx, nr_events, &done);
2296 * We can't just wait for polled events to come to us, we have to actively
2297 * find and complete them.
2299 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2301 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2304 mutex_lock(&ctx->uring_lock);
2305 while (!list_empty(&ctx->iopoll_list)) {
2306 unsigned int nr_events = 0;
2308 io_do_iopoll(ctx, &nr_events, 0);
2310 /* let it sleep and repeat later if can't complete a request */
2314 * Ensure we allow local-to-the-cpu processing to take place,
2315 * in this case we need to ensure that we reap all events.
2316 * Also let task_work, etc. to progress by releasing the mutex
2318 if (need_resched()) {
2319 mutex_unlock(&ctx->uring_lock);
2321 mutex_lock(&ctx->uring_lock);
2324 mutex_unlock(&ctx->uring_lock);
2327 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2329 unsigned int nr_events = 0;
2333 * We disallow the app entering submit/complete with polling, but we
2334 * still need to lock the ring to prevent racing with polled issue
2335 * that got punted to a workqueue.
2337 mutex_lock(&ctx->uring_lock);
2339 * Don't enter poll loop if we already have events pending.
2340 * If we do, we can potentially be spinning for commands that
2341 * already triggered a CQE (eg in error).
2343 if (test_bit(0, &ctx->check_cq_overflow))
2344 __io_cqring_overflow_flush(ctx, false);
2345 if (io_cqring_events(ctx))
2349 * If a submit got punted to a workqueue, we can have the
2350 * application entering polling for a command before it gets
2351 * issued. That app will hold the uring_lock for the duration
2352 * of the poll right here, so we need to take a breather every
2353 * now and then to ensure that the issue has a chance to add
2354 * the poll to the issued list. Otherwise we can spin here
2355 * forever, while the workqueue is stuck trying to acquire the
2358 if (list_empty(&ctx->iopoll_list)) {
2359 u32 tail = ctx->cached_cq_tail;
2361 mutex_unlock(&ctx->uring_lock);
2363 mutex_lock(&ctx->uring_lock);
2365 /* some requests don't go through iopoll_list */
2366 if (tail != ctx->cached_cq_tail ||
2367 list_empty(&ctx->iopoll_list))
2370 ret = io_do_iopoll(ctx, &nr_events, min);
2371 } while (!ret && nr_events < min && !need_resched());
2373 mutex_unlock(&ctx->uring_lock);
2377 static void kiocb_end_write(struct io_kiocb *req)
2380 * Tell lockdep we inherited freeze protection from submission
2383 if (req->flags & REQ_F_ISREG) {
2384 struct super_block *sb = file_inode(req->file)->i_sb;
2386 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2392 static bool io_resubmit_prep(struct io_kiocb *req)
2394 struct io_async_rw *rw = req->async_data;
2397 return !io_req_prep_async(req);
2398 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2399 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2403 static bool io_rw_should_reissue(struct io_kiocb *req)
2405 umode_t mode = file_inode(req->file)->i_mode;
2406 struct io_ring_ctx *ctx = req->ctx;
2408 if (!S_ISBLK(mode) && !S_ISREG(mode))
2410 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2411 !(ctx->flags & IORING_SETUP_IOPOLL)))
2414 * If ref is dying, we might be running poll reap from the exit work.
2415 * Don't attempt to reissue from that path, just let it fail with
2418 if (percpu_ref_is_dying(&ctx->refs))
2423 static bool io_resubmit_prep(struct io_kiocb *req)
2427 static bool io_rw_should_reissue(struct io_kiocb *req)
2433 static void io_fallback_req_func(struct work_struct *work)
2435 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
2436 fallback_work.work);
2437 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
2438 struct io_kiocb *req, *tmp;
2440 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
2441 req->io_task_work.func(req);
2444 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2445 unsigned int issue_flags)
2449 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2450 kiocb_end_write(req);
2451 if (res != req->result) {
2452 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2453 io_rw_should_reissue(req)) {
2454 req->flags |= REQ_F_REISSUE;
2459 if (req->flags & REQ_F_BUFFER_SELECTED)
2460 cflags = io_put_rw_kbuf(req);
2461 __io_req_complete(req, issue_flags, res, cflags);
2464 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2466 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2468 __io_complete_rw(req, res, res2, 0);
2471 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2473 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2475 if (kiocb->ki_flags & IOCB_WRITE)
2476 kiocb_end_write(req);
2477 if (unlikely(res != req->result)) {
2478 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2479 io_resubmit_prep(req))) {
2481 req->flags |= REQ_F_DONT_REISSUE;
2485 WRITE_ONCE(req->result, res);
2486 /* order with io_iopoll_complete() checking ->result */
2488 WRITE_ONCE(req->iopoll_completed, 1);
2492 * After the iocb has been issued, it's safe to be found on the poll list.
2493 * Adding the kiocb to the list AFTER submission ensures that we don't
2494 * find it from a io_do_iopoll() thread before the issuer is done
2495 * accessing the kiocb cookie.
2497 static void io_iopoll_req_issued(struct io_kiocb *req)
2499 struct io_ring_ctx *ctx = req->ctx;
2500 const bool in_async = io_wq_current_is_worker();
2502 /* workqueue context doesn't hold uring_lock, grab it now */
2503 if (unlikely(in_async))
2504 mutex_lock(&ctx->uring_lock);
2507 * Track whether we have multiple files in our lists. This will impact
2508 * how we do polling eventually, not spinning if we're on potentially
2509 * different devices.
2511 if (list_empty(&ctx->iopoll_list)) {
2512 ctx->poll_multi_queue = false;
2513 } else if (!ctx->poll_multi_queue) {
2514 struct io_kiocb *list_req;
2515 unsigned int queue_num0, queue_num1;
2517 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2520 if (list_req->file != req->file) {
2521 ctx->poll_multi_queue = true;
2523 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2524 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2525 if (queue_num0 != queue_num1)
2526 ctx->poll_multi_queue = true;
2531 * For fast devices, IO may have already completed. If it has, add
2532 * it to the front so we find it first.
2534 if (READ_ONCE(req->iopoll_completed))
2535 list_add(&req->inflight_entry, &ctx->iopoll_list);
2537 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2539 if (unlikely(in_async)) {
2541 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2542 * in sq thread task context or in io worker task context. If
2543 * current task context is sq thread, we don't need to check
2544 * whether should wake up sq thread.
2546 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2547 wq_has_sleeper(&ctx->sq_data->wait))
2548 wake_up(&ctx->sq_data->wait);
2550 mutex_unlock(&ctx->uring_lock);
2554 static inline void io_state_file_put(struct io_submit_state *state)
2556 if (state->file_refs) {
2557 fput_many(state->file, state->file_refs);
2558 state->file_refs = 0;
2563 * Get as many references to a file as we have IOs left in this submission,
2564 * assuming most submissions are for one file, or at least that each file
2565 * has more than one submission.
2567 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2572 if (state->file_refs) {
2573 if (state->fd == fd) {
2577 io_state_file_put(state);
2579 state->file = fget_many(fd, state->ios_left);
2580 if (unlikely(!state->file))
2584 state->file_refs = state->ios_left - 1;
2588 static bool io_bdev_nowait(struct block_device *bdev)
2590 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2594 * If we tracked the file through the SCM inflight mechanism, we could support
2595 * any file. For now, just ensure that anything potentially problematic is done
2598 static bool __io_file_supports_async(struct file *file, int rw)
2600 umode_t mode = file_inode(file)->i_mode;
2602 if (S_ISBLK(mode)) {
2603 if (IS_ENABLED(CONFIG_BLOCK) &&
2604 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2610 if (S_ISREG(mode)) {
2611 if (IS_ENABLED(CONFIG_BLOCK) &&
2612 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2613 file->f_op != &io_uring_fops)
2618 /* any ->read/write should understand O_NONBLOCK */
2619 if (file->f_flags & O_NONBLOCK)
2622 if (!(file->f_mode & FMODE_NOWAIT))
2626 return file->f_op->read_iter != NULL;
2628 return file->f_op->write_iter != NULL;
2631 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2633 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2635 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2638 return __io_file_supports_async(req->file, rw);
2641 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2643 struct io_ring_ctx *ctx = req->ctx;
2644 struct kiocb *kiocb = &req->rw.kiocb;
2645 struct file *file = req->file;
2649 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2650 req->flags |= REQ_F_ISREG;
2652 kiocb->ki_pos = READ_ONCE(sqe->off);
2653 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2654 req->flags |= REQ_F_CUR_POS;
2655 kiocb->ki_pos = file->f_pos;
2657 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2658 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2659 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2663 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2664 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2665 req->flags |= REQ_F_NOWAIT;
2667 ioprio = READ_ONCE(sqe->ioprio);
2669 ret = ioprio_check_cap(ioprio);
2673 kiocb->ki_ioprio = ioprio;
2675 kiocb->ki_ioprio = get_current_ioprio();
2677 if (ctx->flags & IORING_SETUP_IOPOLL) {
2678 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2679 !kiocb->ki_filp->f_op->iopoll)
2682 kiocb->ki_flags |= IOCB_HIPRI;
2683 kiocb->ki_complete = io_complete_rw_iopoll;
2684 req->iopoll_completed = 0;
2686 if (kiocb->ki_flags & IOCB_HIPRI)
2688 kiocb->ki_complete = io_complete_rw;
2691 if (req->opcode == IORING_OP_READ_FIXED ||
2692 req->opcode == IORING_OP_WRITE_FIXED) {
2694 io_req_set_rsrc_node(req);
2697 req->rw.addr = READ_ONCE(sqe->addr);
2698 req->rw.len = READ_ONCE(sqe->len);
2699 req->buf_index = READ_ONCE(sqe->buf_index);
2703 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2709 case -ERESTARTNOINTR:
2710 case -ERESTARTNOHAND:
2711 case -ERESTART_RESTARTBLOCK:
2713 * We can't just restart the syscall, since previously
2714 * submitted sqes may already be in progress. Just fail this
2720 kiocb->ki_complete(kiocb, ret, 0);
2724 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2725 unsigned int issue_flags)
2727 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2728 struct io_async_rw *io = req->async_data;
2729 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2731 /* add previously done IO, if any */
2732 if (io && io->bytes_done > 0) {
2734 ret = io->bytes_done;
2736 ret += io->bytes_done;
2739 if (req->flags & REQ_F_CUR_POS)
2740 req->file->f_pos = kiocb->ki_pos;
2741 if (ret >= 0 && check_reissue)
2742 __io_complete_rw(req, ret, 0, issue_flags);
2744 io_rw_done(kiocb, ret);
2746 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2747 req->flags &= ~REQ_F_REISSUE;
2748 if (io_resubmit_prep(req)) {
2750 io_queue_async_work(req);
2755 if (req->flags & REQ_F_BUFFER_SELECTED)
2756 cflags = io_put_rw_kbuf(req);
2757 __io_req_complete(req, issue_flags, ret, cflags);
2762 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2763 struct io_mapped_ubuf *imu)
2765 size_t len = req->rw.len;
2766 u64 buf_end, buf_addr = req->rw.addr;
2769 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2771 /* not inside the mapped region */
2772 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2776 * May not be a start of buffer, set size appropriately
2777 * and advance us to the beginning.
2779 offset = buf_addr - imu->ubuf;
2780 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2784 * Don't use iov_iter_advance() here, as it's really slow for
2785 * using the latter parts of a big fixed buffer - it iterates
2786 * over each segment manually. We can cheat a bit here, because
2789 * 1) it's a BVEC iter, we set it up
2790 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2791 * first and last bvec
2793 * So just find our index, and adjust the iterator afterwards.
2794 * If the offset is within the first bvec (or the whole first
2795 * bvec, just use iov_iter_advance(). This makes it easier
2796 * since we can just skip the first segment, which may not
2797 * be PAGE_SIZE aligned.
2799 const struct bio_vec *bvec = imu->bvec;
2801 if (offset <= bvec->bv_len) {
2802 iov_iter_advance(iter, offset);
2804 unsigned long seg_skip;
2806 /* skip first vec */
2807 offset -= bvec->bv_len;
2808 seg_skip = 1 + (offset >> PAGE_SHIFT);
2810 iter->bvec = bvec + seg_skip;
2811 iter->nr_segs -= seg_skip;
2812 iter->count -= bvec->bv_len + offset;
2813 iter->iov_offset = offset & ~PAGE_MASK;
2820 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2822 struct io_ring_ctx *ctx = req->ctx;
2823 struct io_mapped_ubuf *imu = req->imu;
2824 u16 index, buf_index = req->buf_index;
2827 if (unlikely(buf_index >= ctx->nr_user_bufs))
2829 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2830 imu = READ_ONCE(ctx->user_bufs[index]);
2833 return __io_import_fixed(req, rw, iter, imu);
2836 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2839 mutex_unlock(&ctx->uring_lock);
2842 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2845 * "Normal" inline submissions always hold the uring_lock, since we
2846 * grab it from the system call. Same is true for the SQPOLL offload.
2847 * The only exception is when we've detached the request and issue it
2848 * from an async worker thread, grab the lock for that case.
2851 mutex_lock(&ctx->uring_lock);
2854 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2855 int bgid, struct io_buffer *kbuf,
2858 struct io_buffer *head;
2860 if (req->flags & REQ_F_BUFFER_SELECTED)
2863 io_ring_submit_lock(req->ctx, needs_lock);
2865 lockdep_assert_held(&req->ctx->uring_lock);
2867 head = xa_load(&req->ctx->io_buffers, bgid);
2869 if (!list_empty(&head->list)) {
2870 kbuf = list_last_entry(&head->list, struct io_buffer,
2872 list_del(&kbuf->list);
2875 xa_erase(&req->ctx->io_buffers, bgid);
2877 if (*len > kbuf->len)
2880 kbuf = ERR_PTR(-ENOBUFS);
2883 io_ring_submit_unlock(req->ctx, needs_lock);
2888 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2891 struct io_buffer *kbuf;
2894 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2895 bgid = req->buf_index;
2896 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2899 req->rw.addr = (u64) (unsigned long) kbuf;
2900 req->flags |= REQ_F_BUFFER_SELECTED;
2901 return u64_to_user_ptr(kbuf->addr);
2904 #ifdef CONFIG_COMPAT
2905 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2908 struct compat_iovec __user *uiov;
2909 compat_ssize_t clen;
2913 uiov = u64_to_user_ptr(req->rw.addr);
2914 if (!access_ok(uiov, sizeof(*uiov)))
2916 if (__get_user(clen, &uiov->iov_len))
2922 buf = io_rw_buffer_select(req, &len, needs_lock);
2924 return PTR_ERR(buf);
2925 iov[0].iov_base = buf;
2926 iov[0].iov_len = (compat_size_t) len;
2931 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2934 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2938 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2941 len = iov[0].iov_len;
2944 buf = io_rw_buffer_select(req, &len, needs_lock);
2946 return PTR_ERR(buf);
2947 iov[0].iov_base = buf;
2948 iov[0].iov_len = len;
2952 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2955 if (req->flags & REQ_F_BUFFER_SELECTED) {
2956 struct io_buffer *kbuf;
2958 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2959 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2960 iov[0].iov_len = kbuf->len;
2963 if (req->rw.len != 1)
2966 #ifdef CONFIG_COMPAT
2967 if (req->ctx->compat)
2968 return io_compat_import(req, iov, needs_lock);
2971 return __io_iov_buffer_select(req, iov, needs_lock);
2974 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2975 struct iov_iter *iter, bool needs_lock)
2977 void __user *buf = u64_to_user_ptr(req->rw.addr);
2978 size_t sqe_len = req->rw.len;
2979 u8 opcode = req->opcode;
2982 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2984 return io_import_fixed(req, rw, iter);
2987 /* buffer index only valid with fixed read/write, or buffer select */
2988 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2991 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2992 if (req->flags & REQ_F_BUFFER_SELECT) {
2993 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2995 return PTR_ERR(buf);
2996 req->rw.len = sqe_len;
2999 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3004 if (req->flags & REQ_F_BUFFER_SELECT) {
3005 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3007 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3012 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3016 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3018 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3022 * For files that don't have ->read_iter() and ->write_iter(), handle them
3023 * by looping over ->read() or ->write() manually.
3025 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3027 struct kiocb *kiocb = &req->rw.kiocb;
3028 struct file *file = req->file;
3032 * Don't support polled IO through this interface, and we can't
3033 * support non-blocking either. For the latter, this just causes
3034 * the kiocb to be handled from an async context.
3036 if (kiocb->ki_flags & IOCB_HIPRI)
3038 if (kiocb->ki_flags & IOCB_NOWAIT)
3041 while (iov_iter_count(iter)) {
3045 if (!iov_iter_is_bvec(iter)) {
3046 iovec = iov_iter_iovec(iter);
3048 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3049 iovec.iov_len = req->rw.len;
3053 nr = file->f_op->read(file, iovec.iov_base,
3054 iovec.iov_len, io_kiocb_ppos(kiocb));
3056 nr = file->f_op->write(file, iovec.iov_base,
3057 iovec.iov_len, io_kiocb_ppos(kiocb));
3066 if (nr != iovec.iov_len)
3070 iov_iter_advance(iter, nr);
3076 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3077 const struct iovec *fast_iov, struct iov_iter *iter)
3079 struct io_async_rw *rw = req->async_data;
3081 memcpy(&rw->iter, iter, sizeof(*iter));
3082 rw->free_iovec = iovec;
3084 /* can only be fixed buffers, no need to do anything */
3085 if (iov_iter_is_bvec(iter))
3088 unsigned iov_off = 0;
3090 rw->iter.iov = rw->fast_iov;
3091 if (iter->iov != fast_iov) {
3092 iov_off = iter->iov - fast_iov;
3093 rw->iter.iov += iov_off;
3095 if (rw->fast_iov != fast_iov)
3096 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3097 sizeof(struct iovec) * iter->nr_segs);
3099 req->flags |= REQ_F_NEED_CLEANUP;
3103 static inline int io_alloc_async_data(struct io_kiocb *req)
3105 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3106 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3107 return req->async_data == NULL;
3110 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3111 const struct iovec *fast_iov,
3112 struct iov_iter *iter, bool force)
3114 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3116 if (!req->async_data) {
3117 if (io_alloc_async_data(req)) {
3122 io_req_map_rw(req, iovec, fast_iov, iter);
3127 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3129 struct io_async_rw *iorw = req->async_data;
3130 struct iovec *iov = iorw->fast_iov;
3133 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3134 if (unlikely(ret < 0))
3137 iorw->bytes_done = 0;
3138 iorw->free_iovec = iov;
3140 req->flags |= REQ_F_NEED_CLEANUP;
3144 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3146 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3148 return io_prep_rw(req, sqe);
3152 * This is our waitqueue callback handler, registered through lock_page_async()
3153 * when we initially tried to do the IO with the iocb armed our waitqueue.
3154 * This gets called when the page is unlocked, and we generally expect that to
3155 * happen when the page IO is completed and the page is now uptodate. This will
3156 * queue a task_work based retry of the operation, attempting to copy the data
3157 * again. If the latter fails because the page was NOT uptodate, then we will
3158 * do a thread based blocking retry of the operation. That's the unexpected
3161 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3162 int sync, void *arg)
3164 struct wait_page_queue *wpq;
3165 struct io_kiocb *req = wait->private;
3166 struct wait_page_key *key = arg;
3168 wpq = container_of(wait, struct wait_page_queue, wait);
3170 if (!wake_page_match(wpq, key))
3173 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3174 list_del_init(&wait->entry);
3176 /* submit ref gets dropped, acquire a new one */
3178 io_req_task_queue(req);
3183 * This controls whether a given IO request should be armed for async page
3184 * based retry. If we return false here, the request is handed to the async
3185 * worker threads for retry. If we're doing buffered reads on a regular file,
3186 * we prepare a private wait_page_queue entry and retry the operation. This
3187 * will either succeed because the page is now uptodate and unlocked, or it
3188 * will register a callback when the page is unlocked at IO completion. Through
3189 * that callback, io_uring uses task_work to setup a retry of the operation.
3190 * That retry will attempt the buffered read again. The retry will generally
3191 * succeed, or in rare cases where it fails, we then fall back to using the
3192 * async worker threads for a blocking retry.
3194 static bool io_rw_should_retry(struct io_kiocb *req)
3196 struct io_async_rw *rw = req->async_data;
3197 struct wait_page_queue *wait = &rw->wpq;
3198 struct kiocb *kiocb = &req->rw.kiocb;
3200 /* never retry for NOWAIT, we just complete with -EAGAIN */
3201 if (req->flags & REQ_F_NOWAIT)
3204 /* Only for buffered IO */
3205 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3209 * just use poll if we can, and don't attempt if the fs doesn't
3210 * support callback based unlocks
3212 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3215 wait->wait.func = io_async_buf_func;
3216 wait->wait.private = req;
3217 wait->wait.flags = 0;
3218 INIT_LIST_HEAD(&wait->wait.entry);
3219 kiocb->ki_flags |= IOCB_WAITQ;
3220 kiocb->ki_flags &= ~IOCB_NOWAIT;
3221 kiocb->ki_waitq = wait;
3225 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3227 if (req->file->f_op->read_iter)
3228 return call_read_iter(req->file, &req->rw.kiocb, iter);
3229 else if (req->file->f_op->read)
3230 return loop_rw_iter(READ, req, iter);
3235 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3237 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3238 struct kiocb *kiocb = &req->rw.kiocb;
3239 struct iov_iter __iter, *iter = &__iter;
3240 struct io_async_rw *rw = req->async_data;
3241 ssize_t io_size, ret, ret2;
3242 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3248 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3252 io_size = iov_iter_count(iter);
3253 req->result = io_size;
3255 /* Ensure we clear previously set non-block flag */
3256 if (!force_nonblock)
3257 kiocb->ki_flags &= ~IOCB_NOWAIT;
3259 kiocb->ki_flags |= IOCB_NOWAIT;
3261 /* If the file doesn't support async, just async punt */
3262 if (force_nonblock && !io_file_supports_async(req, READ)) {
3263 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3264 return ret ?: -EAGAIN;
3267 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3268 if (unlikely(ret)) {
3273 ret = io_iter_do_read(req, iter);
3275 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3276 req->flags &= ~REQ_F_REISSUE;
3277 /* IOPOLL retry should happen for io-wq threads */
3278 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3280 /* no retry on NONBLOCK nor RWF_NOWAIT */
3281 if (req->flags & REQ_F_NOWAIT)
3283 /* some cases will consume bytes even on error returns */
3284 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3286 } else if (ret == -EIOCBQUEUED) {
3288 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3289 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3290 /* read all, failed, already did sync or don't want to retry */
3294 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3299 rw = req->async_data;
3300 /* now use our persistent iterator, if we aren't already */
3305 rw->bytes_done += ret;
3306 /* if we can retry, do so with the callbacks armed */
3307 if (!io_rw_should_retry(req)) {
3308 kiocb->ki_flags &= ~IOCB_WAITQ;
3313 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3314 * we get -EIOCBQUEUED, then we'll get a notification when the
3315 * desired page gets unlocked. We can also get a partial read
3316 * here, and if we do, then just retry at the new offset.
3318 ret = io_iter_do_read(req, iter);
3319 if (ret == -EIOCBQUEUED)
3321 /* we got some bytes, but not all. retry. */
3322 kiocb->ki_flags &= ~IOCB_WAITQ;
3323 } while (ret > 0 && ret < io_size);
3325 kiocb_done(kiocb, ret, issue_flags);
3327 /* it's faster to check here then delegate to kfree */
3333 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3335 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3337 return io_prep_rw(req, sqe);
3340 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3342 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3343 struct kiocb *kiocb = &req->rw.kiocb;
3344 struct iov_iter __iter, *iter = &__iter;
3345 struct io_async_rw *rw = req->async_data;
3346 ssize_t ret, ret2, io_size;
3347 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3353 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3357 io_size = iov_iter_count(iter);
3358 req->result = io_size;
3360 /* Ensure we clear previously set non-block flag */
3361 if (!force_nonblock)
3362 kiocb->ki_flags &= ~IOCB_NOWAIT;
3364 kiocb->ki_flags |= IOCB_NOWAIT;
3366 /* If the file doesn't support async, just async punt */
3367 if (force_nonblock && !io_file_supports_async(req, WRITE))
3370 /* file path doesn't support NOWAIT for non-direct_IO */
3371 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3372 (req->flags & REQ_F_ISREG))
3375 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3380 * Open-code file_start_write here to grab freeze protection,
3381 * which will be released by another thread in
3382 * io_complete_rw(). Fool lockdep by telling it the lock got
3383 * released so that it doesn't complain about the held lock when
3384 * we return to userspace.
3386 if (req->flags & REQ_F_ISREG) {
3387 sb_start_write(file_inode(req->file)->i_sb);
3388 __sb_writers_release(file_inode(req->file)->i_sb,
3391 kiocb->ki_flags |= IOCB_WRITE;
3393 if (req->file->f_op->write_iter)
3394 ret2 = call_write_iter(req->file, kiocb, iter);
3395 else if (req->file->f_op->write)
3396 ret2 = loop_rw_iter(WRITE, req, iter);
3400 if (req->flags & REQ_F_REISSUE) {
3401 req->flags &= ~REQ_F_REISSUE;
3406 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3407 * retry them without IOCB_NOWAIT.
3409 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3411 /* no retry on NONBLOCK nor RWF_NOWAIT */
3412 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3414 if (!force_nonblock || ret2 != -EAGAIN) {
3415 /* IOPOLL retry should happen for io-wq threads */
3416 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3419 kiocb_done(kiocb, ret2, issue_flags);
3422 /* some cases will consume bytes even on error returns */
3423 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3424 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3425 return ret ?: -EAGAIN;
3428 /* it's reportedly faster than delegating the null check to kfree() */
3434 static int io_renameat_prep(struct io_kiocb *req,
3435 const struct io_uring_sqe *sqe)
3437 struct io_rename *ren = &req->rename;
3438 const char __user *oldf, *newf;
3440 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3442 if (sqe->ioprio || sqe->buf_index)
3444 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3447 ren->old_dfd = READ_ONCE(sqe->fd);
3448 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3449 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3450 ren->new_dfd = READ_ONCE(sqe->len);
3451 ren->flags = READ_ONCE(sqe->rename_flags);
3453 ren->oldpath = getname(oldf);
3454 if (IS_ERR(ren->oldpath))
3455 return PTR_ERR(ren->oldpath);
3457 ren->newpath = getname(newf);
3458 if (IS_ERR(ren->newpath)) {
3459 putname(ren->oldpath);
3460 return PTR_ERR(ren->newpath);
3463 req->flags |= REQ_F_NEED_CLEANUP;
3467 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3469 struct io_rename *ren = &req->rename;
3472 if (issue_flags & IO_URING_F_NONBLOCK)
3475 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3476 ren->newpath, ren->flags);
3478 req->flags &= ~REQ_F_NEED_CLEANUP;
3481 io_req_complete(req, ret);
3485 static int io_unlinkat_prep(struct io_kiocb *req,
3486 const struct io_uring_sqe *sqe)
3488 struct io_unlink *un = &req->unlink;
3489 const char __user *fname;
3491 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3493 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3495 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3498 un->dfd = READ_ONCE(sqe->fd);
3500 un->flags = READ_ONCE(sqe->unlink_flags);
3501 if (un->flags & ~AT_REMOVEDIR)
3504 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3505 un->filename = getname(fname);
3506 if (IS_ERR(un->filename))
3507 return PTR_ERR(un->filename);
3509 req->flags |= REQ_F_NEED_CLEANUP;
3513 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3515 struct io_unlink *un = &req->unlink;
3518 if (issue_flags & IO_URING_F_NONBLOCK)
3521 if (un->flags & AT_REMOVEDIR)
3522 ret = do_rmdir(un->dfd, un->filename);
3524 ret = do_unlinkat(un->dfd, un->filename);
3526 req->flags &= ~REQ_F_NEED_CLEANUP;
3529 io_req_complete(req, ret);
3533 static int io_shutdown_prep(struct io_kiocb *req,
3534 const struct io_uring_sqe *sqe)
3536 #if defined(CONFIG_NET)
3537 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3539 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3543 req->shutdown.how = READ_ONCE(sqe->len);
3550 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3552 #if defined(CONFIG_NET)
3553 struct socket *sock;
3556 if (issue_flags & IO_URING_F_NONBLOCK)
3559 sock = sock_from_file(req->file);
3560 if (unlikely(!sock))
3563 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3566 io_req_complete(req, ret);
3573 static int __io_splice_prep(struct io_kiocb *req,
3574 const struct io_uring_sqe *sqe)
3576 struct io_splice *sp = &req->splice;
3577 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3579 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3583 sp->len = READ_ONCE(sqe->len);
3584 sp->flags = READ_ONCE(sqe->splice_flags);
3586 if (unlikely(sp->flags & ~valid_flags))
3589 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3590 (sp->flags & SPLICE_F_FD_IN_FIXED));
3593 req->flags |= REQ_F_NEED_CLEANUP;
3597 static int io_tee_prep(struct io_kiocb *req,
3598 const struct io_uring_sqe *sqe)
3600 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3602 return __io_splice_prep(req, sqe);
3605 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3607 struct io_splice *sp = &req->splice;
3608 struct file *in = sp->file_in;
3609 struct file *out = sp->file_out;
3610 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3613 if (issue_flags & IO_URING_F_NONBLOCK)
3616 ret = do_tee(in, out, sp->len, flags);
3618 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3620 req->flags &= ~REQ_F_NEED_CLEANUP;
3624 io_req_complete(req, ret);
3628 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3630 struct io_splice *sp = &req->splice;
3632 sp->off_in = READ_ONCE(sqe->splice_off_in);
3633 sp->off_out = READ_ONCE(sqe->off);
3634 return __io_splice_prep(req, sqe);
3637 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3639 struct io_splice *sp = &req->splice;
3640 struct file *in = sp->file_in;
3641 struct file *out = sp->file_out;
3642 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3643 loff_t *poff_in, *poff_out;
3646 if (issue_flags & IO_URING_F_NONBLOCK)
3649 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3650 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3653 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3655 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3657 req->flags &= ~REQ_F_NEED_CLEANUP;
3661 io_req_complete(req, ret);
3666 * IORING_OP_NOP just posts a completion event, nothing else.
3668 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3670 struct io_ring_ctx *ctx = req->ctx;
3672 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3675 __io_req_complete(req, issue_flags, 0, 0);
3679 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3681 struct io_ring_ctx *ctx = req->ctx;
3686 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3688 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3691 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3692 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3695 req->sync.off = READ_ONCE(sqe->off);
3696 req->sync.len = READ_ONCE(sqe->len);
3700 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3702 loff_t end = req->sync.off + req->sync.len;
3705 /* fsync always requires a blocking context */
3706 if (issue_flags & IO_URING_F_NONBLOCK)
3709 ret = vfs_fsync_range(req->file, req->sync.off,
3710 end > 0 ? end : LLONG_MAX,
3711 req->sync.flags & IORING_FSYNC_DATASYNC);
3714 io_req_complete(req, ret);
3718 static int io_fallocate_prep(struct io_kiocb *req,
3719 const struct io_uring_sqe *sqe)
3721 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3723 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3726 req->sync.off = READ_ONCE(sqe->off);
3727 req->sync.len = READ_ONCE(sqe->addr);
3728 req->sync.mode = READ_ONCE(sqe->len);
3732 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3736 /* fallocate always requiring blocking context */
3737 if (issue_flags & IO_URING_F_NONBLOCK)
3739 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3743 io_req_complete(req, ret);
3747 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3749 const char __user *fname;
3752 if (unlikely(sqe->ioprio || sqe->buf_index))
3754 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3757 /* open.how should be already initialised */
3758 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3759 req->open.how.flags |= O_LARGEFILE;
3761 req->open.dfd = READ_ONCE(sqe->fd);
3762 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3763 req->open.filename = getname(fname);
3764 if (IS_ERR(req->open.filename)) {
3765 ret = PTR_ERR(req->open.filename);
3766 req->open.filename = NULL;
3769 req->open.nofile = rlimit(RLIMIT_NOFILE);
3770 req->flags |= REQ_F_NEED_CLEANUP;
3774 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3778 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3780 mode = READ_ONCE(sqe->len);
3781 flags = READ_ONCE(sqe->open_flags);
3782 req->open.how = build_open_how(flags, mode);
3783 return __io_openat_prep(req, sqe);
3786 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3788 struct open_how __user *how;
3792 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3794 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3795 len = READ_ONCE(sqe->len);
3796 if (len < OPEN_HOW_SIZE_VER0)
3799 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3804 return __io_openat_prep(req, sqe);
3807 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3809 struct open_flags op;
3812 bool resolve_nonblock;
3815 ret = build_open_flags(&req->open.how, &op);
3818 nonblock_set = op.open_flag & O_NONBLOCK;
3819 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3820 if (issue_flags & IO_URING_F_NONBLOCK) {
3822 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3823 * it'll always -EAGAIN
3825 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3827 op.lookup_flags |= LOOKUP_CACHED;
3828 op.open_flag |= O_NONBLOCK;
3831 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3835 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3838 * We could hang on to this 'fd' on retrying, but seems like
3839 * marginal gain for something that is now known to be a slower
3840 * path. So just put it, and we'll get a new one when we retry.
3844 ret = PTR_ERR(file);
3845 /* only retry if RESOLVE_CACHED wasn't already set by application */
3846 if (ret == -EAGAIN &&
3847 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3852 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3853 file->f_flags &= ~O_NONBLOCK;
3854 fsnotify_open(file);
3855 fd_install(ret, file);
3857 putname(req->open.filename);
3858 req->flags &= ~REQ_F_NEED_CLEANUP;
3861 __io_req_complete(req, issue_flags, ret, 0);
3865 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3867 return io_openat2(req, issue_flags);
3870 static int io_remove_buffers_prep(struct io_kiocb *req,
3871 const struct io_uring_sqe *sqe)
3873 struct io_provide_buf *p = &req->pbuf;
3876 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3879 tmp = READ_ONCE(sqe->fd);
3880 if (!tmp || tmp > USHRT_MAX)
3883 memset(p, 0, sizeof(*p));
3885 p->bgid = READ_ONCE(sqe->buf_group);
3889 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3890 int bgid, unsigned nbufs)
3894 /* shouldn't happen */
3898 /* the head kbuf is the list itself */
3899 while (!list_empty(&buf->list)) {
3900 struct io_buffer *nxt;
3902 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3903 list_del(&nxt->list);
3910 xa_erase(&ctx->io_buffers, bgid);
3915 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3917 struct io_provide_buf *p = &req->pbuf;
3918 struct io_ring_ctx *ctx = req->ctx;
3919 struct io_buffer *head;
3921 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3923 io_ring_submit_lock(ctx, !force_nonblock);
3925 lockdep_assert_held(&ctx->uring_lock);
3928 head = xa_load(&ctx->io_buffers, p->bgid);
3930 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3934 /* complete before unlock, IOPOLL may need the lock */
3935 __io_req_complete(req, issue_flags, ret, 0);
3936 io_ring_submit_unlock(ctx, !force_nonblock);
3940 static int io_provide_buffers_prep(struct io_kiocb *req,
3941 const struct io_uring_sqe *sqe)
3943 unsigned long size, tmp_check;
3944 struct io_provide_buf *p = &req->pbuf;
3947 if (sqe->ioprio || sqe->rw_flags)
3950 tmp = READ_ONCE(sqe->fd);
3951 if (!tmp || tmp > USHRT_MAX)
3954 p->addr = READ_ONCE(sqe->addr);
3955 p->len = READ_ONCE(sqe->len);
3957 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3960 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3963 size = (unsigned long)p->len * p->nbufs;
3964 if (!access_ok(u64_to_user_ptr(p->addr), size))
3967 p->bgid = READ_ONCE(sqe->buf_group);
3968 tmp = READ_ONCE(sqe->off);
3969 if (tmp > USHRT_MAX)
3975 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3977 struct io_buffer *buf;
3978 u64 addr = pbuf->addr;
3979 int i, bid = pbuf->bid;
3981 for (i = 0; i < pbuf->nbufs; i++) {
3982 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3987 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
3992 INIT_LIST_HEAD(&buf->list);
3995 list_add_tail(&buf->list, &(*head)->list);
3999 return i ? i : -ENOMEM;
4002 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4004 struct io_provide_buf *p = &req->pbuf;
4005 struct io_ring_ctx *ctx = req->ctx;
4006 struct io_buffer *head, *list;
4008 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4010 io_ring_submit_lock(ctx, !force_nonblock);
4012 lockdep_assert_held(&ctx->uring_lock);
4014 list = head = xa_load(&ctx->io_buffers, p->bgid);
4016 ret = io_add_buffers(p, &head);
4017 if (ret >= 0 && !list) {
4018 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4020 __io_remove_buffers(ctx, head, p->bgid, -1U);
4024 /* complete before unlock, IOPOLL may need the lock */
4025 __io_req_complete(req, issue_flags, ret, 0);
4026 io_ring_submit_unlock(ctx, !force_nonblock);
4030 static int io_epoll_ctl_prep(struct io_kiocb *req,
4031 const struct io_uring_sqe *sqe)
4033 #if defined(CONFIG_EPOLL)
4034 if (sqe->ioprio || sqe->buf_index)
4036 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4039 req->epoll.epfd = READ_ONCE(sqe->fd);
4040 req->epoll.op = READ_ONCE(sqe->len);
4041 req->epoll.fd = READ_ONCE(sqe->off);
4043 if (ep_op_has_event(req->epoll.op)) {
4044 struct epoll_event __user *ev;
4046 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4047 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4057 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4059 #if defined(CONFIG_EPOLL)
4060 struct io_epoll *ie = &req->epoll;
4062 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4064 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4065 if (force_nonblock && ret == -EAGAIN)
4070 __io_req_complete(req, issue_flags, ret, 0);
4077 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4079 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4080 if (sqe->ioprio || sqe->buf_index || sqe->off)
4082 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4085 req->madvise.addr = READ_ONCE(sqe->addr);
4086 req->madvise.len = READ_ONCE(sqe->len);
4087 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4094 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4096 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4097 struct io_madvise *ma = &req->madvise;
4100 if (issue_flags & IO_URING_F_NONBLOCK)
4103 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4106 io_req_complete(req, ret);
4113 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4115 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4117 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4120 req->fadvise.offset = READ_ONCE(sqe->off);
4121 req->fadvise.len = READ_ONCE(sqe->len);
4122 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4126 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4128 struct io_fadvise *fa = &req->fadvise;
4131 if (issue_flags & IO_URING_F_NONBLOCK) {
4132 switch (fa->advice) {
4133 case POSIX_FADV_NORMAL:
4134 case POSIX_FADV_RANDOM:
4135 case POSIX_FADV_SEQUENTIAL:
4142 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4145 __io_req_complete(req, issue_flags, ret, 0);
4149 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4151 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4153 if (sqe->ioprio || sqe->buf_index)
4155 if (req->flags & REQ_F_FIXED_FILE)
4158 req->statx.dfd = READ_ONCE(sqe->fd);
4159 req->statx.mask = READ_ONCE(sqe->len);
4160 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4161 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4162 req->statx.flags = READ_ONCE(sqe->statx_flags);
4167 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4169 struct io_statx *ctx = &req->statx;
4172 if (issue_flags & IO_URING_F_NONBLOCK)
4175 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4180 io_req_complete(req, ret);
4184 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4186 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4188 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4189 sqe->rw_flags || sqe->buf_index)
4191 if (req->flags & REQ_F_FIXED_FILE)
4194 req->close.fd = READ_ONCE(sqe->fd);
4198 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4200 struct files_struct *files = current->files;
4201 struct io_close *close = &req->close;
4202 struct fdtable *fdt;
4203 struct file *file = NULL;
4206 spin_lock(&files->file_lock);
4207 fdt = files_fdtable(files);
4208 if (close->fd >= fdt->max_fds) {
4209 spin_unlock(&files->file_lock);
4212 file = fdt->fd[close->fd];
4213 if (!file || file->f_op == &io_uring_fops) {
4214 spin_unlock(&files->file_lock);
4219 /* if the file has a flush method, be safe and punt to async */
4220 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4221 spin_unlock(&files->file_lock);
4225 ret = __close_fd_get_file(close->fd, &file);
4226 spin_unlock(&files->file_lock);
4233 /* No ->flush() or already async, safely close from here */
4234 ret = filp_close(file, current->files);
4240 __io_req_complete(req, issue_flags, ret, 0);
4244 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4246 struct io_ring_ctx *ctx = req->ctx;
4248 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4250 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4253 req->sync.off = READ_ONCE(sqe->off);
4254 req->sync.len = READ_ONCE(sqe->len);
4255 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4259 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4263 /* sync_file_range always requires a blocking context */
4264 if (issue_flags & IO_URING_F_NONBLOCK)
4267 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4271 io_req_complete(req, ret);
4275 #if defined(CONFIG_NET)
4276 static int io_setup_async_msg(struct io_kiocb *req,
4277 struct io_async_msghdr *kmsg)
4279 struct io_async_msghdr *async_msg = req->async_data;
4283 if (io_alloc_async_data(req)) {
4284 kfree(kmsg->free_iov);
4287 async_msg = req->async_data;
4288 req->flags |= REQ_F_NEED_CLEANUP;
4289 memcpy(async_msg, kmsg, sizeof(*kmsg));
4290 async_msg->msg.msg_name = &async_msg->addr;
4291 /* if were using fast_iov, set it to the new one */
4292 if (!async_msg->free_iov)
4293 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4298 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4299 struct io_async_msghdr *iomsg)
4301 iomsg->msg.msg_name = &iomsg->addr;
4302 iomsg->free_iov = iomsg->fast_iov;
4303 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4304 req->sr_msg.msg_flags, &iomsg->free_iov);
4307 static int io_sendmsg_prep_async(struct io_kiocb *req)
4311 ret = io_sendmsg_copy_hdr(req, req->async_data);
4313 req->flags |= REQ_F_NEED_CLEANUP;
4317 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4319 struct io_sr_msg *sr = &req->sr_msg;
4321 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4324 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4325 sr->len = READ_ONCE(sqe->len);
4326 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4327 if (sr->msg_flags & MSG_DONTWAIT)
4328 req->flags |= REQ_F_NOWAIT;
4330 #ifdef CONFIG_COMPAT
4331 if (req->ctx->compat)
4332 sr->msg_flags |= MSG_CMSG_COMPAT;
4337 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4339 struct io_async_msghdr iomsg, *kmsg;
4340 struct socket *sock;
4345 sock = sock_from_file(req->file);
4346 if (unlikely(!sock))
4349 kmsg = req->async_data;
4351 ret = io_sendmsg_copy_hdr(req, &iomsg);
4357 flags = req->sr_msg.msg_flags;
4358 if (issue_flags & IO_URING_F_NONBLOCK)
4359 flags |= MSG_DONTWAIT;
4360 if (flags & MSG_WAITALL)
4361 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4363 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4364 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4365 return io_setup_async_msg(req, kmsg);
4366 if (ret == -ERESTARTSYS)
4369 /* fast path, check for non-NULL to avoid function call */
4371 kfree(kmsg->free_iov);
4372 req->flags &= ~REQ_F_NEED_CLEANUP;
4375 __io_req_complete(req, issue_flags, ret, 0);
4379 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4381 struct io_sr_msg *sr = &req->sr_msg;
4384 struct socket *sock;
4389 sock = sock_from_file(req->file);
4390 if (unlikely(!sock))
4393 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4397 msg.msg_name = NULL;
4398 msg.msg_control = NULL;
4399 msg.msg_controllen = 0;
4400 msg.msg_namelen = 0;
4402 flags = req->sr_msg.msg_flags;
4403 if (issue_flags & IO_URING_F_NONBLOCK)
4404 flags |= MSG_DONTWAIT;
4405 if (flags & MSG_WAITALL)
4406 min_ret = iov_iter_count(&msg.msg_iter);
4408 msg.msg_flags = flags;
4409 ret = sock_sendmsg(sock, &msg);
4410 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4412 if (ret == -ERESTARTSYS)
4417 __io_req_complete(req, issue_flags, ret, 0);
4421 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4422 struct io_async_msghdr *iomsg)
4424 struct io_sr_msg *sr = &req->sr_msg;
4425 struct iovec __user *uiov;
4429 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4430 &iomsg->uaddr, &uiov, &iov_len);
4434 if (req->flags & REQ_F_BUFFER_SELECT) {
4437 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4439 sr->len = iomsg->fast_iov[0].iov_len;
4440 iomsg->free_iov = NULL;
4442 iomsg->free_iov = iomsg->fast_iov;
4443 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4444 &iomsg->free_iov, &iomsg->msg.msg_iter,
4453 #ifdef CONFIG_COMPAT
4454 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4455 struct io_async_msghdr *iomsg)
4457 struct io_sr_msg *sr = &req->sr_msg;
4458 struct compat_iovec __user *uiov;
4463 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4468 uiov = compat_ptr(ptr);
4469 if (req->flags & REQ_F_BUFFER_SELECT) {
4470 compat_ssize_t clen;
4474 if (!access_ok(uiov, sizeof(*uiov)))
4476 if (__get_user(clen, &uiov->iov_len))
4481 iomsg->free_iov = NULL;
4483 iomsg->free_iov = iomsg->fast_iov;
4484 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4485 UIO_FASTIOV, &iomsg->free_iov,
4486 &iomsg->msg.msg_iter, true);
4495 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4496 struct io_async_msghdr *iomsg)
4498 iomsg->msg.msg_name = &iomsg->addr;
4500 #ifdef CONFIG_COMPAT
4501 if (req->ctx->compat)
4502 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4505 return __io_recvmsg_copy_hdr(req, iomsg);
4508 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4511 struct io_sr_msg *sr = &req->sr_msg;
4512 struct io_buffer *kbuf;
4514 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4519 req->flags |= REQ_F_BUFFER_SELECTED;
4523 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4525 return io_put_kbuf(req, req->sr_msg.kbuf);
4528 static int io_recvmsg_prep_async(struct io_kiocb *req)
4532 ret = io_recvmsg_copy_hdr(req, req->async_data);
4534 req->flags |= REQ_F_NEED_CLEANUP;
4538 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4540 struct io_sr_msg *sr = &req->sr_msg;
4542 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4545 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4546 sr->len = READ_ONCE(sqe->len);
4547 sr->bgid = READ_ONCE(sqe->buf_group);
4548 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4549 if (sr->msg_flags & MSG_DONTWAIT)
4550 req->flags |= REQ_F_NOWAIT;
4552 #ifdef CONFIG_COMPAT
4553 if (req->ctx->compat)
4554 sr->msg_flags |= MSG_CMSG_COMPAT;
4559 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4561 struct io_async_msghdr iomsg, *kmsg;
4562 struct socket *sock;
4563 struct io_buffer *kbuf;
4566 int ret, cflags = 0;
4567 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4569 sock = sock_from_file(req->file);
4570 if (unlikely(!sock))
4573 kmsg = req->async_data;
4575 ret = io_recvmsg_copy_hdr(req, &iomsg);
4581 if (req->flags & REQ_F_BUFFER_SELECT) {
4582 kbuf = io_recv_buffer_select(req, !force_nonblock);
4584 return PTR_ERR(kbuf);
4585 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4586 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4587 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4588 1, req->sr_msg.len);
4591 flags = req->sr_msg.msg_flags;
4593 flags |= MSG_DONTWAIT;
4594 if (flags & MSG_WAITALL)
4595 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4597 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4598 kmsg->uaddr, flags);
4599 if (force_nonblock && ret == -EAGAIN)
4600 return io_setup_async_msg(req, kmsg);
4601 if (ret == -ERESTARTSYS)
4604 if (req->flags & REQ_F_BUFFER_SELECTED)
4605 cflags = io_put_recv_kbuf(req);
4606 /* fast path, check for non-NULL to avoid function call */
4608 kfree(kmsg->free_iov);
4609 req->flags &= ~REQ_F_NEED_CLEANUP;
4610 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4612 __io_req_complete(req, issue_flags, ret, cflags);
4616 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4618 struct io_buffer *kbuf;
4619 struct io_sr_msg *sr = &req->sr_msg;
4621 void __user *buf = sr->buf;
4622 struct socket *sock;
4626 int ret, cflags = 0;
4627 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4629 sock = sock_from_file(req->file);
4630 if (unlikely(!sock))
4633 if (req->flags & REQ_F_BUFFER_SELECT) {
4634 kbuf = io_recv_buffer_select(req, !force_nonblock);
4636 return PTR_ERR(kbuf);
4637 buf = u64_to_user_ptr(kbuf->addr);
4640 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4644 msg.msg_name = NULL;
4645 msg.msg_control = NULL;
4646 msg.msg_controllen = 0;
4647 msg.msg_namelen = 0;
4648 msg.msg_iocb = NULL;
4651 flags = req->sr_msg.msg_flags;
4653 flags |= MSG_DONTWAIT;
4654 if (flags & MSG_WAITALL)
4655 min_ret = iov_iter_count(&msg.msg_iter);
4657 ret = sock_recvmsg(sock, &msg, flags);
4658 if (force_nonblock && ret == -EAGAIN)
4660 if (ret == -ERESTARTSYS)
4663 if (req->flags & REQ_F_BUFFER_SELECTED)
4664 cflags = io_put_recv_kbuf(req);
4665 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4667 __io_req_complete(req, issue_flags, ret, cflags);
4671 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4673 struct io_accept *accept = &req->accept;
4675 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4677 if (sqe->ioprio || sqe->len || sqe->buf_index)
4680 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4681 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4682 accept->flags = READ_ONCE(sqe->accept_flags);
4683 accept->nofile = rlimit(RLIMIT_NOFILE);
4687 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4689 struct io_accept *accept = &req->accept;
4690 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4691 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4694 if (req->file->f_flags & O_NONBLOCK)
4695 req->flags |= REQ_F_NOWAIT;
4697 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4698 accept->addr_len, accept->flags,
4700 if (ret == -EAGAIN && force_nonblock)
4703 if (ret == -ERESTARTSYS)
4707 __io_req_complete(req, issue_flags, ret, 0);
4711 static int io_connect_prep_async(struct io_kiocb *req)
4713 struct io_async_connect *io = req->async_data;
4714 struct io_connect *conn = &req->connect;
4716 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4719 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4721 struct io_connect *conn = &req->connect;
4723 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4725 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4728 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4729 conn->addr_len = READ_ONCE(sqe->addr2);
4733 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4735 struct io_async_connect __io, *io;
4736 unsigned file_flags;
4738 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4740 if (req->async_data) {
4741 io = req->async_data;
4743 ret = move_addr_to_kernel(req->connect.addr,
4744 req->connect.addr_len,
4751 file_flags = force_nonblock ? O_NONBLOCK : 0;
4753 ret = __sys_connect_file(req->file, &io->address,
4754 req->connect.addr_len, file_flags);
4755 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4756 if (req->async_data)
4758 if (io_alloc_async_data(req)) {
4762 memcpy(req->async_data, &__io, sizeof(__io));
4765 if (ret == -ERESTARTSYS)
4770 __io_req_complete(req, issue_flags, ret, 0);
4773 #else /* !CONFIG_NET */
4774 #define IO_NETOP_FN(op) \
4775 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4777 return -EOPNOTSUPP; \
4780 #define IO_NETOP_PREP(op) \
4782 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4784 return -EOPNOTSUPP; \
4787 #define IO_NETOP_PREP_ASYNC(op) \
4789 static int io_##op##_prep_async(struct io_kiocb *req) \
4791 return -EOPNOTSUPP; \
4794 IO_NETOP_PREP_ASYNC(sendmsg);
4795 IO_NETOP_PREP_ASYNC(recvmsg);
4796 IO_NETOP_PREP_ASYNC(connect);
4797 IO_NETOP_PREP(accept);
4800 #endif /* CONFIG_NET */
4802 struct io_poll_table {
4803 struct poll_table_struct pt;
4804 struct io_kiocb *req;
4809 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4810 __poll_t mask, io_req_tw_func_t func)
4812 /* for instances that support it check for an event match first: */
4813 if (mask && !(mask & poll->events))
4816 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4818 list_del_init(&poll->wait.entry);
4821 req->io_task_work.func = func;
4824 * If this fails, then the task is exiting. When a task exits, the
4825 * work gets canceled, so just cancel this request as well instead
4826 * of executing it. We can't safely execute it anyway, as we may not
4827 * have the needed state needed for it anyway.
4829 io_req_task_work_add(req);
4833 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4834 __acquires(&req->ctx->completion_lock)
4836 struct io_ring_ctx *ctx = req->ctx;
4838 if (unlikely(req->task->flags & PF_EXITING))
4839 WRITE_ONCE(poll->canceled, true);
4841 if (!req->result && !READ_ONCE(poll->canceled)) {
4842 struct poll_table_struct pt = { ._key = poll->events };
4844 req->result = vfs_poll(req->file, &pt) & poll->events;
4847 spin_lock_irq(&ctx->completion_lock);
4848 if (!req->result && !READ_ONCE(poll->canceled)) {
4849 add_wait_queue(poll->head, &poll->wait);
4856 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4858 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4859 if (req->opcode == IORING_OP_POLL_ADD)
4860 return req->async_data;
4861 return req->apoll->double_poll;
4864 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4866 if (req->opcode == IORING_OP_POLL_ADD)
4868 return &req->apoll->poll;
4871 static void io_poll_remove_double(struct io_kiocb *req)
4872 __must_hold(&req->ctx->completion_lock)
4874 struct io_poll_iocb *poll = io_poll_get_double(req);
4876 lockdep_assert_held(&req->ctx->completion_lock);
4878 if (poll && poll->head) {
4879 struct wait_queue_head *head = poll->head;
4881 spin_lock(&head->lock);
4882 list_del_init(&poll->wait.entry);
4883 if (poll->wait.private)
4886 spin_unlock(&head->lock);
4890 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4891 __must_hold(&req->ctx->completion_lock)
4893 struct io_ring_ctx *ctx = req->ctx;
4894 unsigned flags = IORING_CQE_F_MORE;
4897 if (READ_ONCE(req->poll.canceled)) {
4899 req->poll.events |= EPOLLONESHOT;
4901 error = mangle_poll(mask);
4903 if (req->poll.events & EPOLLONESHOT)
4905 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4906 io_poll_remove_waitqs(req);
4907 req->poll.done = true;
4910 if (flags & IORING_CQE_F_MORE)
4913 io_commit_cqring(ctx);
4914 return !(flags & IORING_CQE_F_MORE);
4917 static void io_poll_task_func(struct io_kiocb *req)
4919 struct io_ring_ctx *ctx = req->ctx;
4920 struct io_kiocb *nxt;
4922 if (io_poll_rewait(req, &req->poll)) {
4923 spin_unlock_irq(&ctx->completion_lock);
4927 done = io_poll_complete(req, req->result);
4929 hash_del(&req->hash_node);
4932 add_wait_queue(req->poll.head, &req->poll.wait);
4934 spin_unlock_irq(&ctx->completion_lock);
4935 io_cqring_ev_posted(ctx);
4938 nxt = io_put_req_find_next(req);
4940 io_req_task_submit(nxt);
4945 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4946 int sync, void *key)
4948 struct io_kiocb *req = wait->private;
4949 struct io_poll_iocb *poll = io_poll_get_single(req);
4950 __poll_t mask = key_to_poll(key);
4952 /* for instances that support it check for an event match first: */
4953 if (mask && !(mask & poll->events))
4955 if (!(poll->events & EPOLLONESHOT))
4956 return poll->wait.func(&poll->wait, mode, sync, key);
4958 list_del_init(&wait->entry);
4963 spin_lock(&poll->head->lock);
4964 done = list_empty(&poll->wait.entry);
4966 list_del_init(&poll->wait.entry);
4967 /* make sure double remove sees this as being gone */
4968 wait->private = NULL;
4969 spin_unlock(&poll->head->lock);
4971 /* use wait func handler, so it matches the rq type */
4972 poll->wait.func(&poll->wait, mode, sync, key);
4979 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4980 wait_queue_func_t wake_func)
4984 poll->canceled = false;
4985 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
4986 /* mask in events that we always want/need */
4987 poll->events = events | IO_POLL_UNMASK;
4988 INIT_LIST_HEAD(&poll->wait.entry);
4989 init_waitqueue_func_entry(&poll->wait, wake_func);
4992 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4993 struct wait_queue_head *head,
4994 struct io_poll_iocb **poll_ptr)
4996 struct io_kiocb *req = pt->req;
4999 * The file being polled uses multiple waitqueues for poll handling
5000 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5003 if (unlikely(pt->nr_entries)) {
5004 struct io_poll_iocb *poll_one = poll;
5006 /* already have a 2nd entry, fail a third attempt */
5008 pt->error = -EINVAL;
5012 * Can't handle multishot for double wait for now, turn it
5013 * into one-shot mode.
5015 if (!(poll_one->events & EPOLLONESHOT))
5016 poll_one->events |= EPOLLONESHOT;
5017 /* double add on the same waitqueue head, ignore */
5018 if (poll_one->head == head)
5020 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5022 pt->error = -ENOMEM;
5025 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5027 poll->wait.private = req;
5034 if (poll->events & EPOLLEXCLUSIVE)
5035 add_wait_queue_exclusive(head, &poll->wait);
5037 add_wait_queue(head, &poll->wait);
5040 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5041 struct poll_table_struct *p)
5043 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5044 struct async_poll *apoll = pt->req->apoll;
5046 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5049 static void io_async_task_func(struct io_kiocb *req)
5051 struct async_poll *apoll = req->apoll;
5052 struct io_ring_ctx *ctx = req->ctx;
5054 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5056 if (io_poll_rewait(req, &apoll->poll)) {
5057 spin_unlock_irq(&ctx->completion_lock);
5061 hash_del(&req->hash_node);
5062 io_poll_remove_double(req);
5063 spin_unlock_irq(&ctx->completion_lock);
5065 if (!READ_ONCE(apoll->poll.canceled))
5066 io_req_task_submit(req);
5068 io_req_complete_failed(req, -ECANCELED);
5071 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5074 struct io_kiocb *req = wait->private;
5075 struct io_poll_iocb *poll = &req->apoll->poll;
5077 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5080 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5083 static void io_poll_req_insert(struct io_kiocb *req)
5085 struct io_ring_ctx *ctx = req->ctx;
5086 struct hlist_head *list;
5088 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5089 hlist_add_head(&req->hash_node, list);
5092 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5093 struct io_poll_iocb *poll,
5094 struct io_poll_table *ipt, __poll_t mask,
5095 wait_queue_func_t wake_func)
5096 __acquires(&ctx->completion_lock)
5098 struct io_ring_ctx *ctx = req->ctx;
5099 bool cancel = false;
5101 INIT_HLIST_NODE(&req->hash_node);
5102 io_init_poll_iocb(poll, mask, wake_func);
5103 poll->file = req->file;
5104 poll->wait.private = req;
5106 ipt->pt._key = mask;
5109 ipt->nr_entries = 0;
5111 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5112 if (unlikely(!ipt->nr_entries) && !ipt->error)
5113 ipt->error = -EINVAL;
5115 spin_lock_irq(&ctx->completion_lock);
5116 if (likely(poll->head)) {
5117 spin_lock(&poll->head->lock);
5118 if (unlikely(list_empty(&poll->wait.entry))) {
5124 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5125 list_del_init(&poll->wait.entry);
5127 WRITE_ONCE(poll->canceled, true);
5128 else if (!poll->done) /* actually waiting for an event */
5129 io_poll_req_insert(req);
5130 spin_unlock(&poll->head->lock);
5142 static int io_arm_poll_handler(struct io_kiocb *req)
5144 const struct io_op_def *def = &io_op_defs[req->opcode];
5145 struct io_ring_ctx *ctx = req->ctx;
5146 struct async_poll *apoll;
5147 struct io_poll_table ipt;
5148 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5151 if (!req->file || !file_can_poll(req->file))
5152 return IO_APOLL_ABORTED;
5153 if (req->flags & REQ_F_POLLED)
5154 return IO_APOLL_ABORTED;
5155 if (!def->pollin && !def->pollout)
5156 return IO_APOLL_ABORTED;
5160 mask |= POLLIN | POLLRDNORM;
5162 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5163 if ((req->opcode == IORING_OP_RECVMSG) &&
5164 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5168 mask |= POLLOUT | POLLWRNORM;
5171 /* if we can't nonblock try, then no point in arming a poll handler */
5172 if (!io_file_supports_async(req, rw))
5173 return IO_APOLL_ABORTED;
5175 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5176 if (unlikely(!apoll))
5177 return IO_APOLL_ABORTED;
5178 apoll->double_poll = NULL;
5180 req->flags |= REQ_F_POLLED;
5181 ipt.pt._qproc = io_async_queue_proc;
5183 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5185 if (ret || ipt.error) {
5186 io_poll_remove_double(req);
5187 spin_unlock_irq(&ctx->completion_lock);
5189 return IO_APOLL_READY;
5190 return IO_APOLL_ABORTED;
5192 spin_unlock_irq(&ctx->completion_lock);
5193 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5194 mask, apoll->poll.events);
5198 static bool __io_poll_remove_one(struct io_kiocb *req,
5199 struct io_poll_iocb *poll, bool do_cancel)
5200 __must_hold(&req->ctx->completion_lock)
5202 bool do_complete = false;
5206 spin_lock(&poll->head->lock);
5208 WRITE_ONCE(poll->canceled, true);
5209 if (!list_empty(&poll->wait.entry)) {
5210 list_del_init(&poll->wait.entry);
5213 spin_unlock(&poll->head->lock);
5214 hash_del(&req->hash_node);
5218 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5219 __must_hold(&req->ctx->completion_lock)
5223 io_poll_remove_double(req);
5224 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5226 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5227 /* non-poll requests have submit ref still */
5233 static bool io_poll_remove_one(struct io_kiocb *req)
5234 __must_hold(&req->ctx->completion_lock)
5238 do_complete = io_poll_remove_waitqs(req);
5240 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5241 io_commit_cqring(req->ctx);
5243 io_put_req_deferred(req, 1);
5250 * Returns true if we found and killed one or more poll requests
5252 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5255 struct hlist_node *tmp;
5256 struct io_kiocb *req;
5259 spin_lock_irq(&ctx->completion_lock);
5260 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5261 struct hlist_head *list;
5263 list = &ctx->cancel_hash[i];
5264 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5265 if (io_match_task(req, tsk, cancel_all))
5266 posted += io_poll_remove_one(req);
5269 spin_unlock_irq(&ctx->completion_lock);
5272 io_cqring_ev_posted(ctx);
5277 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5279 __must_hold(&ctx->completion_lock)
5281 struct hlist_head *list;
5282 struct io_kiocb *req;
5284 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5285 hlist_for_each_entry(req, list, hash_node) {
5286 if (sqe_addr != req->user_data)
5288 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5295 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5297 __must_hold(&ctx->completion_lock)
5299 struct io_kiocb *req;
5301 req = io_poll_find(ctx, sqe_addr, poll_only);
5304 if (io_poll_remove_one(req))
5310 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5315 events = READ_ONCE(sqe->poll32_events);
5317 events = swahw32(events);
5319 if (!(flags & IORING_POLL_ADD_MULTI))
5320 events |= EPOLLONESHOT;
5321 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5324 static int io_poll_update_prep(struct io_kiocb *req,
5325 const struct io_uring_sqe *sqe)
5327 struct io_poll_update *upd = &req->poll_update;
5330 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5332 if (sqe->ioprio || sqe->buf_index)
5334 flags = READ_ONCE(sqe->len);
5335 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5336 IORING_POLL_ADD_MULTI))
5338 /* meaningless without update */
5339 if (flags == IORING_POLL_ADD_MULTI)
5342 upd->old_user_data = READ_ONCE(sqe->addr);
5343 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5344 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5346 upd->new_user_data = READ_ONCE(sqe->off);
5347 if (!upd->update_user_data && upd->new_user_data)
5349 if (upd->update_events)
5350 upd->events = io_poll_parse_events(sqe, flags);
5351 else if (sqe->poll32_events)
5357 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5360 struct io_kiocb *req = wait->private;
5361 struct io_poll_iocb *poll = &req->poll;
5363 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5366 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5367 struct poll_table_struct *p)
5369 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5371 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5374 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5376 struct io_poll_iocb *poll = &req->poll;
5379 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5381 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5383 flags = READ_ONCE(sqe->len);
5384 if (flags & ~IORING_POLL_ADD_MULTI)
5387 poll->events = io_poll_parse_events(sqe, flags);
5391 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5393 struct io_poll_iocb *poll = &req->poll;
5394 struct io_ring_ctx *ctx = req->ctx;
5395 struct io_poll_table ipt;
5398 ipt.pt._qproc = io_poll_queue_proc;
5400 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5403 if (mask) { /* no async, we'd stolen it */
5405 io_poll_complete(req, mask);
5407 spin_unlock_irq(&ctx->completion_lock);
5410 io_cqring_ev_posted(ctx);
5411 if (poll->events & EPOLLONESHOT)
5417 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5419 struct io_ring_ctx *ctx = req->ctx;
5420 struct io_kiocb *preq;
5424 spin_lock_irq(&ctx->completion_lock);
5425 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5431 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5433 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5438 * Don't allow racy completion with singleshot, as we cannot safely
5439 * update those. For multishot, if we're racing with completion, just
5440 * let completion re-add it.
5442 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5443 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5447 /* we now have a detached poll request. reissue. */
5451 spin_unlock_irq(&ctx->completion_lock);
5453 io_req_complete(req, ret);
5456 /* only mask one event flags, keep behavior flags */
5457 if (req->poll_update.update_events) {
5458 preq->poll.events &= ~0xffff;
5459 preq->poll.events |= req->poll_update.events & 0xffff;
5460 preq->poll.events |= IO_POLL_UNMASK;
5462 if (req->poll_update.update_user_data)
5463 preq->user_data = req->poll_update.new_user_data;
5464 spin_unlock_irq(&ctx->completion_lock);
5466 /* complete update request, we're done with it */
5467 io_req_complete(req, ret);
5470 ret = io_poll_add(preq, issue_flags);
5473 io_req_complete(preq, ret);
5479 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5481 struct io_timeout_data *data = container_of(timer,
5482 struct io_timeout_data, timer);
5483 struct io_kiocb *req = data->req;
5484 struct io_ring_ctx *ctx = req->ctx;
5485 unsigned long flags;
5487 spin_lock_irqsave(&ctx->completion_lock, flags);
5488 list_del_init(&req->timeout.list);
5489 atomic_set(&req->ctx->cq_timeouts,
5490 atomic_read(&req->ctx->cq_timeouts) + 1);
5492 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5493 io_commit_cqring(ctx);
5494 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5496 io_cqring_ev_posted(ctx);
5499 return HRTIMER_NORESTART;
5502 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5504 __must_hold(&ctx->completion_lock)
5506 struct io_timeout_data *io;
5507 struct io_kiocb *req;
5510 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5511 found = user_data == req->user_data;
5516 return ERR_PTR(-ENOENT);
5518 io = req->async_data;
5519 if (hrtimer_try_to_cancel(&io->timer) == -1)
5520 return ERR_PTR(-EALREADY);
5521 list_del_init(&req->timeout.list);
5525 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5526 __must_hold(&ctx->completion_lock)
5528 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5531 return PTR_ERR(req);
5534 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5535 io_put_req_deferred(req, 1);
5539 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5540 struct timespec64 *ts, enum hrtimer_mode mode)
5541 __must_hold(&ctx->completion_lock)
5543 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5544 struct io_timeout_data *data;
5547 return PTR_ERR(req);
5549 req->timeout.off = 0; /* noseq */
5550 data = req->async_data;
5551 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5552 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5553 data->timer.function = io_timeout_fn;
5554 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5558 static int io_timeout_remove_prep(struct io_kiocb *req,
5559 const struct io_uring_sqe *sqe)
5561 struct io_timeout_rem *tr = &req->timeout_rem;
5563 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5565 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5567 if (sqe->ioprio || sqe->buf_index || sqe->len)
5570 tr->addr = READ_ONCE(sqe->addr);
5571 tr->flags = READ_ONCE(sqe->timeout_flags);
5572 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5573 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5575 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5577 } else if (tr->flags) {
5578 /* timeout removal doesn't support flags */
5585 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5587 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5592 * Remove or update an existing timeout command
5594 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5596 struct io_timeout_rem *tr = &req->timeout_rem;
5597 struct io_ring_ctx *ctx = req->ctx;
5600 spin_lock_irq(&ctx->completion_lock);
5601 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5602 ret = io_timeout_cancel(ctx, tr->addr);
5604 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5605 io_translate_timeout_mode(tr->flags));
5607 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5608 io_commit_cqring(ctx);
5609 spin_unlock_irq(&ctx->completion_lock);
5610 io_cqring_ev_posted(ctx);
5617 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5618 bool is_timeout_link)
5620 struct io_timeout_data *data;
5622 u32 off = READ_ONCE(sqe->off);
5624 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5626 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5628 if (off && is_timeout_link)
5630 flags = READ_ONCE(sqe->timeout_flags);
5631 if (flags & ~IORING_TIMEOUT_ABS)
5634 req->timeout.off = off;
5635 if (unlikely(off && !req->ctx->off_timeout_used))
5636 req->ctx->off_timeout_used = true;
5638 if (!req->async_data && io_alloc_async_data(req))
5641 data = req->async_data;
5644 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5647 data->mode = io_translate_timeout_mode(flags);
5648 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5649 if (is_timeout_link)
5650 io_req_track_inflight(req);
5654 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5656 struct io_ring_ctx *ctx = req->ctx;
5657 struct io_timeout_data *data = req->async_data;
5658 struct list_head *entry;
5659 u32 tail, off = req->timeout.off;
5661 spin_lock_irq(&ctx->completion_lock);
5664 * sqe->off holds how many events that need to occur for this
5665 * timeout event to be satisfied. If it isn't set, then this is
5666 * a pure timeout request, sequence isn't used.
5668 if (io_is_timeout_noseq(req)) {
5669 entry = ctx->timeout_list.prev;
5673 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5674 req->timeout.target_seq = tail + off;
5676 /* Update the last seq here in case io_flush_timeouts() hasn't.
5677 * This is safe because ->completion_lock is held, and submissions
5678 * and completions are never mixed in the same ->completion_lock section.
5680 ctx->cq_last_tm_flush = tail;
5683 * Insertion sort, ensuring the first entry in the list is always
5684 * the one we need first.
5686 list_for_each_prev(entry, &ctx->timeout_list) {
5687 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5690 if (io_is_timeout_noseq(nxt))
5692 /* nxt.seq is behind @tail, otherwise would've been completed */
5693 if (off >= nxt->timeout.target_seq - tail)
5697 list_add(&req->timeout.list, entry);
5698 data->timer.function = io_timeout_fn;
5699 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5700 spin_unlock_irq(&ctx->completion_lock);
5704 struct io_cancel_data {
5705 struct io_ring_ctx *ctx;
5709 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5711 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5712 struct io_cancel_data *cd = data;
5714 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5717 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5718 struct io_ring_ctx *ctx)
5720 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5721 enum io_wq_cancel cancel_ret;
5724 if (!tctx || !tctx->io_wq)
5727 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5728 switch (cancel_ret) {
5729 case IO_WQ_CANCEL_OK:
5732 case IO_WQ_CANCEL_RUNNING:
5735 case IO_WQ_CANCEL_NOTFOUND:
5743 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5744 struct io_kiocb *req, __u64 sqe_addr,
5747 unsigned long flags;
5750 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5751 spin_lock_irqsave(&ctx->completion_lock, flags);
5754 ret = io_timeout_cancel(ctx, sqe_addr);
5757 ret = io_poll_cancel(ctx, sqe_addr, false);
5761 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5762 io_commit_cqring(ctx);
5763 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5764 io_cqring_ev_posted(ctx);
5770 static int io_async_cancel_prep(struct io_kiocb *req,
5771 const struct io_uring_sqe *sqe)
5773 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5775 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5777 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5780 req->cancel.addr = READ_ONCE(sqe->addr);
5784 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5786 struct io_ring_ctx *ctx = req->ctx;
5787 u64 sqe_addr = req->cancel.addr;
5788 struct io_tctx_node *node;
5791 /* tasks should wait for their io-wq threads, so safe w/o sync */
5792 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5793 spin_lock_irq(&ctx->completion_lock);
5796 ret = io_timeout_cancel(ctx, sqe_addr);
5799 ret = io_poll_cancel(ctx, sqe_addr, false);
5802 spin_unlock_irq(&ctx->completion_lock);
5804 /* slow path, try all io-wq's */
5805 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5807 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5808 struct io_uring_task *tctx = node->task->io_uring;
5810 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5814 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5816 spin_lock_irq(&ctx->completion_lock);
5818 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5819 io_commit_cqring(ctx);
5820 spin_unlock_irq(&ctx->completion_lock);
5821 io_cqring_ev_posted(ctx);
5829 static int io_rsrc_update_prep(struct io_kiocb *req,
5830 const struct io_uring_sqe *sqe)
5832 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5834 if (sqe->ioprio || sqe->rw_flags)
5837 req->rsrc_update.offset = READ_ONCE(sqe->off);
5838 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5839 if (!req->rsrc_update.nr_args)
5841 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5845 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5847 struct io_ring_ctx *ctx = req->ctx;
5848 struct io_uring_rsrc_update2 up;
5851 if (issue_flags & IO_URING_F_NONBLOCK)
5854 up.offset = req->rsrc_update.offset;
5855 up.data = req->rsrc_update.arg;
5860 mutex_lock(&ctx->uring_lock);
5861 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5862 &up, req->rsrc_update.nr_args);
5863 mutex_unlock(&ctx->uring_lock);
5867 __io_req_complete(req, issue_flags, ret, 0);
5871 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5873 switch (req->opcode) {
5876 case IORING_OP_READV:
5877 case IORING_OP_READ_FIXED:
5878 case IORING_OP_READ:
5879 return io_read_prep(req, sqe);
5880 case IORING_OP_WRITEV:
5881 case IORING_OP_WRITE_FIXED:
5882 case IORING_OP_WRITE:
5883 return io_write_prep(req, sqe);
5884 case IORING_OP_POLL_ADD:
5885 return io_poll_add_prep(req, sqe);
5886 case IORING_OP_POLL_REMOVE:
5887 return io_poll_update_prep(req, sqe);
5888 case IORING_OP_FSYNC:
5889 return io_fsync_prep(req, sqe);
5890 case IORING_OP_SYNC_FILE_RANGE:
5891 return io_sfr_prep(req, sqe);
5892 case IORING_OP_SENDMSG:
5893 case IORING_OP_SEND:
5894 return io_sendmsg_prep(req, sqe);
5895 case IORING_OP_RECVMSG:
5896 case IORING_OP_RECV:
5897 return io_recvmsg_prep(req, sqe);
5898 case IORING_OP_CONNECT:
5899 return io_connect_prep(req, sqe);
5900 case IORING_OP_TIMEOUT:
5901 return io_timeout_prep(req, sqe, false);
5902 case IORING_OP_TIMEOUT_REMOVE:
5903 return io_timeout_remove_prep(req, sqe);
5904 case IORING_OP_ASYNC_CANCEL:
5905 return io_async_cancel_prep(req, sqe);
5906 case IORING_OP_LINK_TIMEOUT:
5907 return io_timeout_prep(req, sqe, true);
5908 case IORING_OP_ACCEPT:
5909 return io_accept_prep(req, sqe);
5910 case IORING_OP_FALLOCATE:
5911 return io_fallocate_prep(req, sqe);
5912 case IORING_OP_OPENAT:
5913 return io_openat_prep(req, sqe);
5914 case IORING_OP_CLOSE:
5915 return io_close_prep(req, sqe);
5916 case IORING_OP_FILES_UPDATE:
5917 return io_rsrc_update_prep(req, sqe);
5918 case IORING_OP_STATX:
5919 return io_statx_prep(req, sqe);
5920 case IORING_OP_FADVISE:
5921 return io_fadvise_prep(req, sqe);
5922 case IORING_OP_MADVISE:
5923 return io_madvise_prep(req, sqe);
5924 case IORING_OP_OPENAT2:
5925 return io_openat2_prep(req, sqe);
5926 case IORING_OP_EPOLL_CTL:
5927 return io_epoll_ctl_prep(req, sqe);
5928 case IORING_OP_SPLICE:
5929 return io_splice_prep(req, sqe);
5930 case IORING_OP_PROVIDE_BUFFERS:
5931 return io_provide_buffers_prep(req, sqe);
5932 case IORING_OP_REMOVE_BUFFERS:
5933 return io_remove_buffers_prep(req, sqe);
5935 return io_tee_prep(req, sqe);
5936 case IORING_OP_SHUTDOWN:
5937 return io_shutdown_prep(req, sqe);
5938 case IORING_OP_RENAMEAT:
5939 return io_renameat_prep(req, sqe);
5940 case IORING_OP_UNLINKAT:
5941 return io_unlinkat_prep(req, sqe);
5944 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5949 static int io_req_prep_async(struct io_kiocb *req)
5951 if (!io_op_defs[req->opcode].needs_async_setup)
5953 if (WARN_ON_ONCE(req->async_data))
5955 if (io_alloc_async_data(req))
5958 switch (req->opcode) {
5959 case IORING_OP_READV:
5960 return io_rw_prep_async(req, READ);
5961 case IORING_OP_WRITEV:
5962 return io_rw_prep_async(req, WRITE);
5963 case IORING_OP_SENDMSG:
5964 return io_sendmsg_prep_async(req);
5965 case IORING_OP_RECVMSG:
5966 return io_recvmsg_prep_async(req);
5967 case IORING_OP_CONNECT:
5968 return io_connect_prep_async(req);
5970 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5975 static u32 io_get_sequence(struct io_kiocb *req)
5977 u32 seq = req->ctx->cached_sq_head;
5979 /* need original cached_sq_head, but it was increased for each req */
5980 io_for_each_link(req, req)
5985 static bool io_drain_req(struct io_kiocb *req)
5987 struct io_kiocb *pos;
5988 struct io_ring_ctx *ctx = req->ctx;
5989 struct io_defer_entry *de;
5994 * If we need to drain a request in the middle of a link, drain the
5995 * head request and the next request/link after the current link.
5996 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
5997 * maintained for every request of our link.
5999 if (ctx->drain_next) {
6000 req->flags |= REQ_F_IO_DRAIN;
6001 ctx->drain_next = false;
6003 /* not interested in head, start from the first linked */
6004 io_for_each_link(pos, req->link) {
6005 if (pos->flags & REQ_F_IO_DRAIN) {
6006 ctx->drain_next = true;
6007 req->flags |= REQ_F_IO_DRAIN;
6012 /* Still need defer if there is pending req in defer list. */
6013 if (likely(list_empty_careful(&ctx->defer_list) &&
6014 !(req->flags & REQ_F_IO_DRAIN))) {
6015 ctx->drain_active = false;
6019 seq = io_get_sequence(req);
6020 /* Still a chance to pass the sequence check */
6021 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6024 ret = io_req_prep_async(req);
6027 io_prep_async_link(req);
6028 de = kmalloc(sizeof(*de), GFP_KERNEL);
6032 io_req_complete_failed(req, ret);
6036 spin_lock_irq(&ctx->completion_lock);
6037 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6038 spin_unlock_irq(&ctx->completion_lock);
6040 io_queue_async_work(req);
6044 trace_io_uring_defer(ctx, req, req->user_data);
6047 list_add_tail(&de->list, &ctx->defer_list);
6048 spin_unlock_irq(&ctx->completion_lock);
6052 static void io_clean_op(struct io_kiocb *req)
6054 if (req->flags & REQ_F_BUFFER_SELECTED) {
6055 switch (req->opcode) {
6056 case IORING_OP_READV:
6057 case IORING_OP_READ_FIXED:
6058 case IORING_OP_READ:
6059 kfree((void *)(unsigned long)req->rw.addr);
6061 case IORING_OP_RECVMSG:
6062 case IORING_OP_RECV:
6063 kfree(req->sr_msg.kbuf);
6068 if (req->flags & REQ_F_NEED_CLEANUP) {
6069 switch (req->opcode) {
6070 case IORING_OP_READV:
6071 case IORING_OP_READ_FIXED:
6072 case IORING_OP_READ:
6073 case IORING_OP_WRITEV:
6074 case IORING_OP_WRITE_FIXED:
6075 case IORING_OP_WRITE: {
6076 struct io_async_rw *io = req->async_data;
6078 kfree(io->free_iovec);
6081 case IORING_OP_RECVMSG:
6082 case IORING_OP_SENDMSG: {
6083 struct io_async_msghdr *io = req->async_data;
6085 kfree(io->free_iov);
6088 case IORING_OP_SPLICE:
6090 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6091 io_put_file(req->splice.file_in);
6093 case IORING_OP_OPENAT:
6094 case IORING_OP_OPENAT2:
6095 if (req->open.filename)
6096 putname(req->open.filename);
6098 case IORING_OP_RENAMEAT:
6099 putname(req->rename.oldpath);
6100 putname(req->rename.newpath);
6102 case IORING_OP_UNLINKAT:
6103 putname(req->unlink.filename);
6107 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6108 kfree(req->apoll->double_poll);
6112 if (req->flags & REQ_F_INFLIGHT) {
6113 struct io_uring_task *tctx = req->task->io_uring;
6115 atomic_dec(&tctx->inflight_tracked);
6117 if (req->flags & REQ_F_CREDS)
6118 put_cred(req->creds);
6120 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6123 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6125 struct io_ring_ctx *ctx = req->ctx;
6126 const struct cred *creds = NULL;
6129 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6130 creds = override_creds(req->creds);
6132 switch (req->opcode) {
6134 ret = io_nop(req, issue_flags);
6136 case IORING_OP_READV:
6137 case IORING_OP_READ_FIXED:
6138 case IORING_OP_READ:
6139 ret = io_read(req, issue_flags);
6141 case IORING_OP_WRITEV:
6142 case IORING_OP_WRITE_FIXED:
6143 case IORING_OP_WRITE:
6144 ret = io_write(req, issue_flags);
6146 case IORING_OP_FSYNC:
6147 ret = io_fsync(req, issue_flags);
6149 case IORING_OP_POLL_ADD:
6150 ret = io_poll_add(req, issue_flags);
6152 case IORING_OP_POLL_REMOVE:
6153 ret = io_poll_update(req, issue_flags);
6155 case IORING_OP_SYNC_FILE_RANGE:
6156 ret = io_sync_file_range(req, issue_flags);
6158 case IORING_OP_SENDMSG:
6159 ret = io_sendmsg(req, issue_flags);
6161 case IORING_OP_SEND:
6162 ret = io_send(req, issue_flags);
6164 case IORING_OP_RECVMSG:
6165 ret = io_recvmsg(req, issue_flags);
6167 case IORING_OP_RECV:
6168 ret = io_recv(req, issue_flags);
6170 case IORING_OP_TIMEOUT:
6171 ret = io_timeout(req, issue_flags);
6173 case IORING_OP_TIMEOUT_REMOVE:
6174 ret = io_timeout_remove(req, issue_flags);
6176 case IORING_OP_ACCEPT:
6177 ret = io_accept(req, issue_flags);
6179 case IORING_OP_CONNECT:
6180 ret = io_connect(req, issue_flags);
6182 case IORING_OP_ASYNC_CANCEL:
6183 ret = io_async_cancel(req, issue_flags);
6185 case IORING_OP_FALLOCATE:
6186 ret = io_fallocate(req, issue_flags);
6188 case IORING_OP_OPENAT:
6189 ret = io_openat(req, issue_flags);
6191 case IORING_OP_CLOSE:
6192 ret = io_close(req, issue_flags);
6194 case IORING_OP_FILES_UPDATE:
6195 ret = io_files_update(req, issue_flags);
6197 case IORING_OP_STATX:
6198 ret = io_statx(req, issue_flags);
6200 case IORING_OP_FADVISE:
6201 ret = io_fadvise(req, issue_flags);
6203 case IORING_OP_MADVISE:
6204 ret = io_madvise(req, issue_flags);
6206 case IORING_OP_OPENAT2:
6207 ret = io_openat2(req, issue_flags);
6209 case IORING_OP_EPOLL_CTL:
6210 ret = io_epoll_ctl(req, issue_flags);
6212 case IORING_OP_SPLICE:
6213 ret = io_splice(req, issue_flags);
6215 case IORING_OP_PROVIDE_BUFFERS:
6216 ret = io_provide_buffers(req, issue_flags);
6218 case IORING_OP_REMOVE_BUFFERS:
6219 ret = io_remove_buffers(req, issue_flags);
6222 ret = io_tee(req, issue_flags);
6224 case IORING_OP_SHUTDOWN:
6225 ret = io_shutdown(req, issue_flags);
6227 case IORING_OP_RENAMEAT:
6228 ret = io_renameat(req, issue_flags);
6230 case IORING_OP_UNLINKAT:
6231 ret = io_unlinkat(req, issue_flags);
6239 revert_creds(creds);
6242 /* If the op doesn't have a file, we're not polling for it */
6243 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6244 io_iopoll_req_issued(req);
6249 static void io_wq_submit_work(struct io_wq_work *work)
6251 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6252 struct io_kiocb *timeout;
6255 timeout = io_prep_linked_timeout(req);
6257 io_queue_linked_timeout(timeout);
6259 if (work->flags & IO_WQ_WORK_CANCEL)
6264 ret = io_issue_sqe(req, 0);
6266 * We can get EAGAIN for polled IO even though we're
6267 * forcing a sync submission from here, since we can't
6268 * wait for request slots on the block side.
6276 /* avoid locking problems by failing it from a clean context */
6278 /* io-wq is going to take one down */
6280 io_req_task_queue_fail(req, ret);
6284 #define FFS_ASYNC_READ 0x1UL
6285 #define FFS_ASYNC_WRITE 0x2UL
6287 #define FFS_ISREG 0x4UL
6289 #define FFS_ISREG 0x0UL
6291 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6293 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6296 struct io_fixed_file *table_l2;
6298 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6299 return &table_l2[i & IORING_FILE_TABLE_MASK];
6302 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6305 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6307 return (struct file *) (slot->file_ptr & FFS_MASK);
6310 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6312 unsigned long file_ptr = (unsigned long) file;
6314 if (__io_file_supports_async(file, READ))
6315 file_ptr |= FFS_ASYNC_READ;
6316 if (__io_file_supports_async(file, WRITE))
6317 file_ptr |= FFS_ASYNC_WRITE;
6318 if (S_ISREG(file_inode(file)->i_mode))
6319 file_ptr |= FFS_ISREG;
6320 file_slot->file_ptr = file_ptr;
6323 static struct file *io_file_get(struct io_submit_state *state,
6324 struct io_kiocb *req, int fd, bool fixed)
6326 struct io_ring_ctx *ctx = req->ctx;
6330 unsigned long file_ptr;
6332 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6334 fd = array_index_nospec(fd, ctx->nr_user_files);
6335 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6336 file = (struct file *) (file_ptr & FFS_MASK);
6337 file_ptr &= ~FFS_MASK;
6338 /* mask in overlapping REQ_F and FFS bits */
6339 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6340 io_req_set_rsrc_node(req);
6342 trace_io_uring_file_get(ctx, fd);
6343 file = __io_file_get(state, fd);
6345 /* we don't allow fixed io_uring files */
6346 if (file && unlikely(file->f_op == &io_uring_fops))
6347 io_req_track_inflight(req);
6353 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6355 struct io_timeout_data *data = container_of(timer,
6356 struct io_timeout_data, timer);
6357 struct io_kiocb *prev, *req = data->req;
6358 struct io_ring_ctx *ctx = req->ctx;
6359 unsigned long flags;
6361 spin_lock_irqsave(&ctx->completion_lock, flags);
6362 prev = req->timeout.head;
6363 req->timeout.head = NULL;
6366 * We don't expect the list to be empty, that will only happen if we
6367 * race with the completion of the linked work.
6370 io_remove_next_linked(prev);
6371 if (!req_ref_inc_not_zero(prev))
6374 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6377 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6378 io_put_req_deferred(prev, 1);
6379 io_put_req_deferred(req, 1);
6381 io_req_complete_post(req, -ETIME, 0);
6383 return HRTIMER_NORESTART;
6386 static void io_queue_linked_timeout(struct io_kiocb *req)
6388 struct io_ring_ctx *ctx = req->ctx;
6390 spin_lock_irq(&ctx->completion_lock);
6392 * If the back reference is NULL, then our linked request finished
6393 * before we got a chance to setup the timer
6395 if (req->timeout.head) {
6396 struct io_timeout_data *data = req->async_data;
6398 data->timer.function = io_link_timeout_fn;
6399 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6402 spin_unlock_irq(&ctx->completion_lock);
6403 /* drop submission reference */
6407 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6409 struct io_kiocb *nxt = req->link;
6411 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6412 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6415 nxt->timeout.head = req;
6416 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6417 req->flags |= REQ_F_LINK_TIMEOUT;
6421 static void __io_queue_sqe(struct io_kiocb *req)
6423 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6427 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6430 * We async punt it if the file wasn't marked NOWAIT, or if the file
6431 * doesn't support non-blocking read/write attempts
6434 /* drop submission reference */
6435 if (req->flags & REQ_F_COMPLETE_INLINE) {
6436 struct io_ring_ctx *ctx = req->ctx;
6437 struct io_comp_state *cs = &ctx->submit_state.comp;
6439 cs->reqs[cs->nr++] = req;
6440 if (cs->nr == ARRAY_SIZE(cs->reqs))
6441 io_submit_flush_completions(ctx);
6445 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6446 switch (io_arm_poll_handler(req)) {
6447 case IO_APOLL_READY:
6449 case IO_APOLL_ABORTED:
6451 * Queued up for async execution, worker will release
6452 * submit reference when the iocb is actually submitted.
6454 io_queue_async_work(req);
6458 io_req_complete_failed(req, ret);
6461 io_queue_linked_timeout(linked_timeout);
6464 static inline void io_queue_sqe(struct io_kiocb *req)
6466 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6469 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6470 __io_queue_sqe(req);
6472 int ret = io_req_prep_async(req);
6475 io_req_complete_failed(req, ret);
6477 io_queue_async_work(req);
6482 * Check SQE restrictions (opcode and flags).
6484 * Returns 'true' if SQE is allowed, 'false' otherwise.
6486 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6487 struct io_kiocb *req,
6488 unsigned int sqe_flags)
6490 if (likely(!ctx->restricted))
6493 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6496 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6497 ctx->restrictions.sqe_flags_required)
6500 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6501 ctx->restrictions.sqe_flags_required))
6507 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6508 const struct io_uring_sqe *sqe)
6510 struct io_submit_state *state;
6511 unsigned int sqe_flags;
6512 int personality, ret = 0;
6514 req->opcode = READ_ONCE(sqe->opcode);
6515 /* same numerical values with corresponding REQ_F_*, safe to copy */
6516 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6517 req->user_data = READ_ONCE(sqe->user_data);
6519 req->fixed_rsrc_refs = NULL;
6520 /* one is dropped after submission, the other at completion */
6521 atomic_set(&req->refs, 2);
6522 req->task = current;
6524 /* enforce forwards compatibility on users */
6525 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6527 if (unlikely(req->opcode >= IORING_OP_LAST))
6529 if (!io_check_restriction(ctx, req, sqe_flags))
6532 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6533 !io_op_defs[req->opcode].buffer_select)
6535 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6536 ctx->drain_active = true;
6538 personality = READ_ONCE(sqe->personality);
6540 req->creds = xa_load(&ctx->personalities, personality);
6543 get_cred(req->creds);
6544 req->flags |= REQ_F_CREDS;
6546 state = &ctx->submit_state;
6549 * Plug now if we have more than 1 IO left after this, and the target
6550 * is potentially a read/write to block based storage.
6552 if (!state->plug_started && state->ios_left > 1 &&
6553 io_op_defs[req->opcode].plug) {
6554 blk_start_plug(&state->plug);
6555 state->plug_started = true;
6558 if (io_op_defs[req->opcode].needs_file) {
6559 bool fixed = req->flags & REQ_F_FIXED_FILE;
6561 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6562 if (unlikely(!req->file))
6570 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6571 const struct io_uring_sqe *sqe)
6573 struct io_submit_link *link = &ctx->submit_state.link;
6576 ret = io_init_req(ctx, req, sqe);
6577 if (unlikely(ret)) {
6580 /* fail even hard links since we don't submit */
6581 req_set_fail(link->head);
6582 io_req_complete_failed(link->head, -ECANCELED);
6585 io_req_complete_failed(req, ret);
6589 ret = io_req_prep(req, sqe);
6593 /* don't need @sqe from now on */
6594 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6596 ctx->flags & IORING_SETUP_SQPOLL);
6599 * If we already have a head request, queue this one for async
6600 * submittal once the head completes. If we don't have a head but
6601 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6602 * submitted sync once the chain is complete. If none of those
6603 * conditions are true (normal request), then just queue it.
6606 struct io_kiocb *head = link->head;
6608 ret = io_req_prep_async(req);
6611 trace_io_uring_link(ctx, req, head);
6612 link->last->link = req;
6615 /* last request of a link, enqueue the link */
6616 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6621 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6633 * Batched submission is done, ensure local IO is flushed out.
6635 static void io_submit_state_end(struct io_submit_state *state,
6636 struct io_ring_ctx *ctx)
6638 if (state->link.head)
6639 io_queue_sqe(state->link.head);
6641 io_submit_flush_completions(ctx);
6642 if (state->plug_started)
6643 blk_finish_plug(&state->plug);
6644 io_state_file_put(state);
6648 * Start submission side cache.
6650 static void io_submit_state_start(struct io_submit_state *state,
6651 unsigned int max_ios)
6653 state->plug_started = false;
6654 state->ios_left = max_ios;
6655 /* set only head, no need to init link_last in advance */
6656 state->link.head = NULL;
6659 static void io_commit_sqring(struct io_ring_ctx *ctx)
6661 struct io_rings *rings = ctx->rings;
6664 * Ensure any loads from the SQEs are done at this point,
6665 * since once we write the new head, the application could
6666 * write new data to them.
6668 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6672 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6673 * that is mapped by userspace. This means that care needs to be taken to
6674 * ensure that reads are stable, as we cannot rely on userspace always
6675 * being a good citizen. If members of the sqe are validated and then later
6676 * used, it's important that those reads are done through READ_ONCE() to
6677 * prevent a re-load down the line.
6679 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6681 unsigned head, mask = ctx->sq_entries - 1;
6682 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6685 * The cached sq head (or cq tail) serves two purposes:
6687 * 1) allows us to batch the cost of updating the user visible
6689 * 2) allows the kernel side to track the head on its own, even
6690 * though the application is the one updating it.
6692 head = READ_ONCE(ctx->sq_array[sq_idx]);
6693 if (likely(head < ctx->sq_entries))
6694 return &ctx->sq_sqes[head];
6696 /* drop invalid entries */
6698 WRITE_ONCE(ctx->rings->sq_dropped,
6699 READ_ONCE(ctx->rings->sq_dropped) + 1);
6703 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6705 struct io_uring_task *tctx;
6708 /* make sure SQ entry isn't read before tail */
6709 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6710 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6713 tctx = current->io_uring;
6714 tctx->cached_refs -= nr;
6715 if (unlikely(tctx->cached_refs < 0)) {
6716 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6718 percpu_counter_add(&tctx->inflight, refill);
6719 refcount_add(refill, ¤t->usage);
6720 tctx->cached_refs += refill;
6722 io_submit_state_start(&ctx->submit_state, nr);
6724 while (submitted < nr) {
6725 const struct io_uring_sqe *sqe;
6726 struct io_kiocb *req;
6728 req = io_alloc_req(ctx);
6729 if (unlikely(!req)) {
6731 submitted = -EAGAIN;
6734 sqe = io_get_sqe(ctx);
6735 if (unlikely(!sqe)) {
6736 kmem_cache_free(req_cachep, req);
6739 /* will complete beyond this point, count as submitted */
6741 if (io_submit_sqe(ctx, req, sqe))
6745 if (unlikely(submitted != nr)) {
6746 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6747 int unused = nr - ref_used;
6749 current->io_uring->cached_refs += unused;
6750 percpu_ref_put_many(&ctx->refs, unused);
6753 io_submit_state_end(&ctx->submit_state, ctx);
6754 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6755 io_commit_sqring(ctx);
6760 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6762 return READ_ONCE(sqd->state);
6765 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6767 /* Tell userspace we may need a wakeup call */
6768 spin_lock_irq(&ctx->completion_lock);
6769 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6770 spin_unlock_irq(&ctx->completion_lock);
6773 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6775 spin_lock_irq(&ctx->completion_lock);
6776 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6777 spin_unlock_irq(&ctx->completion_lock);
6780 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6782 unsigned int to_submit;
6785 to_submit = io_sqring_entries(ctx);
6786 /* if we're handling multiple rings, cap submit size for fairness */
6787 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6788 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6790 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6791 unsigned nr_events = 0;
6792 const struct cred *creds = NULL;
6794 if (ctx->sq_creds != current_cred())
6795 creds = override_creds(ctx->sq_creds);
6797 mutex_lock(&ctx->uring_lock);
6798 if (!list_empty(&ctx->iopoll_list))
6799 io_do_iopoll(ctx, &nr_events, 0);
6802 * Don't submit if refs are dying, good for io_uring_register(),
6803 * but also it is relied upon by io_ring_exit_work()
6805 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6806 !(ctx->flags & IORING_SETUP_R_DISABLED))
6807 ret = io_submit_sqes(ctx, to_submit);
6808 mutex_unlock(&ctx->uring_lock);
6810 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6811 wake_up(&ctx->sqo_sq_wait);
6813 revert_creds(creds);
6819 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6821 struct io_ring_ctx *ctx;
6822 unsigned sq_thread_idle = 0;
6824 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6825 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6826 sqd->sq_thread_idle = sq_thread_idle;
6829 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6831 bool did_sig = false;
6832 struct ksignal ksig;
6834 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6835 signal_pending(current)) {
6836 mutex_unlock(&sqd->lock);
6837 if (signal_pending(current))
6838 did_sig = get_signal(&ksig);
6840 mutex_lock(&sqd->lock);
6842 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6845 static int io_sq_thread(void *data)
6847 struct io_sq_data *sqd = data;
6848 struct io_ring_ctx *ctx;
6849 unsigned long timeout = 0;
6850 char buf[TASK_COMM_LEN];
6853 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6854 set_task_comm(current, buf);
6856 if (sqd->sq_cpu != -1)
6857 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6859 set_cpus_allowed_ptr(current, cpu_online_mask);
6860 current->flags |= PF_NO_SETAFFINITY;
6862 mutex_lock(&sqd->lock);
6864 bool cap_entries, sqt_spin = false;
6866 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6867 if (io_sqd_handle_event(sqd))
6869 timeout = jiffies + sqd->sq_thread_idle;
6872 cap_entries = !list_is_singular(&sqd->ctx_list);
6873 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6874 int ret = __io_sq_thread(ctx, cap_entries);
6876 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6879 if (io_run_task_work())
6882 if (sqt_spin || !time_after(jiffies, timeout)) {
6885 timeout = jiffies + sqd->sq_thread_idle;
6889 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6890 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6891 bool needs_sched = true;
6893 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6894 io_ring_set_wakeup_flag(ctx);
6896 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6897 !list_empty_careful(&ctx->iopoll_list)) {
6898 needs_sched = false;
6901 if (io_sqring_entries(ctx)) {
6902 needs_sched = false;
6908 mutex_unlock(&sqd->lock);
6910 mutex_lock(&sqd->lock);
6912 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6913 io_ring_clear_wakeup_flag(ctx);
6916 finish_wait(&sqd->wait, &wait);
6917 timeout = jiffies + sqd->sq_thread_idle;
6920 io_uring_cancel_generic(true, sqd);
6922 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6923 io_ring_set_wakeup_flag(ctx);
6925 mutex_unlock(&sqd->lock);
6927 complete(&sqd->exited);
6931 struct io_wait_queue {
6932 struct wait_queue_entry wq;
6933 struct io_ring_ctx *ctx;
6935 unsigned nr_timeouts;
6938 static inline bool io_should_wake(struct io_wait_queue *iowq)
6940 struct io_ring_ctx *ctx = iowq->ctx;
6943 * Wake up if we have enough events, or if a timeout occurred since we
6944 * started waiting. For timeouts, we always want to return to userspace,
6945 * regardless of event count.
6947 return io_cqring_events(ctx) >= iowq->to_wait ||
6948 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6951 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6952 int wake_flags, void *key)
6954 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6958 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6959 * the task, and the next invocation will do it.
6961 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
6962 return autoremove_wake_function(curr, mode, wake_flags, key);
6966 static int io_run_task_work_sig(void)
6968 if (io_run_task_work())
6970 if (!signal_pending(current))
6972 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6973 return -ERESTARTSYS;
6977 /* when returns >0, the caller should retry */
6978 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6979 struct io_wait_queue *iowq,
6980 signed long *timeout)
6984 /* make sure we run task_work before checking for signals */
6985 ret = io_run_task_work_sig();
6986 if (ret || io_should_wake(iowq))
6988 /* let the caller flush overflows, retry */
6989 if (test_bit(0, &ctx->check_cq_overflow))
6992 *timeout = schedule_timeout(*timeout);
6993 return !*timeout ? -ETIME : 1;
6997 * Wait until events become available, if we don't already have some. The
6998 * application must reap them itself, as they reside on the shared cq ring.
7000 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7001 const sigset_t __user *sig, size_t sigsz,
7002 struct __kernel_timespec __user *uts)
7004 struct io_wait_queue iowq = {
7007 .func = io_wake_function,
7008 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7011 .to_wait = min_events,
7013 struct io_rings *rings = ctx->rings;
7014 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7018 io_cqring_overflow_flush(ctx, false);
7019 if (io_cqring_events(ctx) >= min_events)
7021 if (!io_run_task_work())
7026 #ifdef CONFIG_COMPAT
7027 if (in_compat_syscall())
7028 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7032 ret = set_user_sigmask(sig, sigsz);
7039 struct timespec64 ts;
7041 if (get_timespec64(&ts, uts))
7043 timeout = timespec64_to_jiffies(&ts);
7046 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7047 trace_io_uring_cqring_wait(ctx, min_events);
7049 /* if we can't even flush overflow, don't wait for more */
7050 if (!io_cqring_overflow_flush(ctx, false)) {
7054 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7055 TASK_INTERRUPTIBLE);
7056 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7057 finish_wait(&ctx->cq_wait, &iowq.wq);
7061 restore_saved_sigmask_unless(ret == -EINTR);
7063 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7066 static void io_free_page_table(void **table, size_t size)
7068 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7070 for (i = 0; i < nr_tables; i++)
7075 static void **io_alloc_page_table(size_t size)
7077 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7078 size_t init_size = size;
7081 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7085 for (i = 0; i < nr_tables; i++) {
7086 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7088 table[i] = kzalloc(this_size, GFP_KERNEL);
7090 io_free_page_table(table, init_size);
7098 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7100 spin_lock_bh(&ctx->rsrc_ref_lock);
7103 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7105 spin_unlock_bh(&ctx->rsrc_ref_lock);
7108 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7110 percpu_ref_exit(&ref_node->refs);
7114 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7115 struct io_rsrc_data *data_to_kill)
7117 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7118 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7121 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7123 rsrc_node->rsrc_data = data_to_kill;
7124 io_rsrc_ref_lock(ctx);
7125 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7126 io_rsrc_ref_unlock(ctx);
7128 atomic_inc(&data_to_kill->refs);
7129 percpu_ref_kill(&rsrc_node->refs);
7130 ctx->rsrc_node = NULL;
7133 if (!ctx->rsrc_node) {
7134 ctx->rsrc_node = ctx->rsrc_backup_node;
7135 ctx->rsrc_backup_node = NULL;
7139 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7141 if (ctx->rsrc_backup_node)
7143 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7144 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7147 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7151 /* As we may drop ->uring_lock, other task may have started quiesce */
7155 data->quiesce = true;
7157 ret = io_rsrc_node_switch_start(ctx);
7160 io_rsrc_node_switch(ctx, data);
7162 /* kill initial ref, already quiesced if zero */
7163 if (atomic_dec_and_test(&data->refs))
7165 flush_delayed_work(&ctx->rsrc_put_work);
7166 ret = wait_for_completion_interruptible(&data->done);
7170 atomic_inc(&data->refs);
7171 /* wait for all works potentially completing data->done */
7172 flush_delayed_work(&ctx->rsrc_put_work);
7173 reinit_completion(&data->done);
7175 mutex_unlock(&ctx->uring_lock);
7176 ret = io_run_task_work_sig();
7177 mutex_lock(&ctx->uring_lock);
7179 data->quiesce = false;
7184 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7186 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7187 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7189 return &data->tags[table_idx][off];
7192 static void io_rsrc_data_free(struct io_rsrc_data *data)
7194 size_t size = data->nr * sizeof(data->tags[0][0]);
7197 io_free_page_table((void **)data->tags, size);
7201 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7202 u64 __user *utags, unsigned nr,
7203 struct io_rsrc_data **pdata)
7205 struct io_rsrc_data *data;
7209 data = kzalloc(sizeof(*data), GFP_KERNEL);
7212 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7220 data->do_put = do_put;
7223 for (i = 0; i < nr; i++) {
7224 u64 *tag_slot = io_get_tag_slot(data, i);
7226 if (copy_from_user(tag_slot, &utags[i],
7232 atomic_set(&data->refs, 1);
7233 init_completion(&data->done);
7237 io_rsrc_data_free(data);
7241 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7243 size_t size = nr_files * sizeof(struct io_fixed_file);
7245 table->files = (struct io_fixed_file **)io_alloc_page_table(size);
7246 return !!table->files;
7249 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7251 size_t size = nr_files * sizeof(struct io_fixed_file);
7253 io_free_page_table((void **)table->files, size);
7254 table->files = NULL;
7257 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7259 #if defined(CONFIG_UNIX)
7260 if (ctx->ring_sock) {
7261 struct sock *sock = ctx->ring_sock->sk;
7262 struct sk_buff *skb;
7264 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7270 for (i = 0; i < ctx->nr_user_files; i++) {
7273 file = io_file_from_index(ctx, i);
7278 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7279 io_rsrc_data_free(ctx->file_data);
7280 ctx->file_data = NULL;
7281 ctx->nr_user_files = 0;
7284 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7288 if (!ctx->file_data)
7290 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7292 __io_sqe_files_unregister(ctx);
7296 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7297 __releases(&sqd->lock)
7299 WARN_ON_ONCE(sqd->thread == current);
7302 * Do the dance but not conditional clear_bit() because it'd race with
7303 * other threads incrementing park_pending and setting the bit.
7305 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7306 if (atomic_dec_return(&sqd->park_pending))
7307 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7308 mutex_unlock(&sqd->lock);
7311 static void io_sq_thread_park(struct io_sq_data *sqd)
7312 __acquires(&sqd->lock)
7314 WARN_ON_ONCE(sqd->thread == current);
7316 atomic_inc(&sqd->park_pending);
7317 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7318 mutex_lock(&sqd->lock);
7320 wake_up_process(sqd->thread);
7323 static void io_sq_thread_stop(struct io_sq_data *sqd)
7325 WARN_ON_ONCE(sqd->thread == current);
7326 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7328 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7329 mutex_lock(&sqd->lock);
7331 wake_up_process(sqd->thread);
7332 mutex_unlock(&sqd->lock);
7333 wait_for_completion(&sqd->exited);
7336 static void io_put_sq_data(struct io_sq_data *sqd)
7338 if (refcount_dec_and_test(&sqd->refs)) {
7339 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7341 io_sq_thread_stop(sqd);
7346 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7348 struct io_sq_data *sqd = ctx->sq_data;
7351 io_sq_thread_park(sqd);
7352 list_del_init(&ctx->sqd_list);
7353 io_sqd_update_thread_idle(sqd);
7354 io_sq_thread_unpark(sqd);
7356 io_put_sq_data(sqd);
7357 ctx->sq_data = NULL;
7361 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7363 struct io_ring_ctx *ctx_attach;
7364 struct io_sq_data *sqd;
7367 f = fdget(p->wq_fd);
7369 return ERR_PTR(-ENXIO);
7370 if (f.file->f_op != &io_uring_fops) {
7372 return ERR_PTR(-EINVAL);
7375 ctx_attach = f.file->private_data;
7376 sqd = ctx_attach->sq_data;
7379 return ERR_PTR(-EINVAL);
7381 if (sqd->task_tgid != current->tgid) {
7383 return ERR_PTR(-EPERM);
7386 refcount_inc(&sqd->refs);
7391 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7394 struct io_sq_data *sqd;
7397 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7398 sqd = io_attach_sq_data(p);
7403 /* fall through for EPERM case, setup new sqd/task */
7404 if (PTR_ERR(sqd) != -EPERM)
7408 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7410 return ERR_PTR(-ENOMEM);
7412 atomic_set(&sqd->park_pending, 0);
7413 refcount_set(&sqd->refs, 1);
7414 INIT_LIST_HEAD(&sqd->ctx_list);
7415 mutex_init(&sqd->lock);
7416 init_waitqueue_head(&sqd->wait);
7417 init_completion(&sqd->exited);
7421 #if defined(CONFIG_UNIX)
7423 * Ensure the UNIX gc is aware of our file set, so we are certain that
7424 * the io_uring can be safely unregistered on process exit, even if we have
7425 * loops in the file referencing.
7427 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7429 struct sock *sk = ctx->ring_sock->sk;
7430 struct scm_fp_list *fpl;
7431 struct sk_buff *skb;
7434 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7438 skb = alloc_skb(0, GFP_KERNEL);
7447 fpl->user = get_uid(current_user());
7448 for (i = 0; i < nr; i++) {
7449 struct file *file = io_file_from_index(ctx, i + offset);
7453 fpl->fp[nr_files] = get_file(file);
7454 unix_inflight(fpl->user, fpl->fp[nr_files]);
7459 fpl->max = SCM_MAX_FD;
7460 fpl->count = nr_files;
7461 UNIXCB(skb).fp = fpl;
7462 skb->destructor = unix_destruct_scm;
7463 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7464 skb_queue_head(&sk->sk_receive_queue, skb);
7466 for (i = 0; i < nr_files; i++)
7477 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7478 * causes regular reference counting to break down. We rely on the UNIX
7479 * garbage collection to take care of this problem for us.
7481 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7483 unsigned left, total;
7487 left = ctx->nr_user_files;
7489 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7491 ret = __io_sqe_files_scm(ctx, this_files, total);
7495 total += this_files;
7501 while (total < ctx->nr_user_files) {
7502 struct file *file = io_file_from_index(ctx, total);
7512 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7518 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7520 struct file *file = prsrc->file;
7521 #if defined(CONFIG_UNIX)
7522 struct sock *sock = ctx->ring_sock->sk;
7523 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7524 struct sk_buff *skb;
7527 __skb_queue_head_init(&list);
7530 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7531 * remove this entry and rearrange the file array.
7533 skb = skb_dequeue(head);
7535 struct scm_fp_list *fp;
7537 fp = UNIXCB(skb).fp;
7538 for (i = 0; i < fp->count; i++) {
7541 if (fp->fp[i] != file)
7544 unix_notinflight(fp->user, fp->fp[i]);
7545 left = fp->count - 1 - i;
7547 memmove(&fp->fp[i], &fp->fp[i + 1],
7548 left * sizeof(struct file *));
7555 __skb_queue_tail(&list, skb);
7565 __skb_queue_tail(&list, skb);
7567 skb = skb_dequeue(head);
7570 if (skb_peek(&list)) {
7571 spin_lock_irq(&head->lock);
7572 while ((skb = __skb_dequeue(&list)) != NULL)
7573 __skb_queue_tail(head, skb);
7574 spin_unlock_irq(&head->lock);
7581 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7583 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7584 struct io_ring_ctx *ctx = rsrc_data->ctx;
7585 struct io_rsrc_put *prsrc, *tmp;
7587 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7588 list_del(&prsrc->list);
7591 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7593 io_ring_submit_lock(ctx, lock_ring);
7594 spin_lock_irq(&ctx->completion_lock);
7595 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7597 io_commit_cqring(ctx);
7598 spin_unlock_irq(&ctx->completion_lock);
7599 io_cqring_ev_posted(ctx);
7600 io_ring_submit_unlock(ctx, lock_ring);
7603 rsrc_data->do_put(ctx, prsrc);
7607 io_rsrc_node_destroy(ref_node);
7608 if (atomic_dec_and_test(&rsrc_data->refs))
7609 complete(&rsrc_data->done);
7612 static void io_rsrc_put_work(struct work_struct *work)
7614 struct io_ring_ctx *ctx;
7615 struct llist_node *node;
7617 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7618 node = llist_del_all(&ctx->rsrc_put_llist);
7621 struct io_rsrc_node *ref_node;
7622 struct llist_node *next = node->next;
7624 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7625 __io_rsrc_put_work(ref_node);
7630 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7632 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7633 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7634 bool first_add = false;
7636 io_rsrc_ref_lock(ctx);
7639 while (!list_empty(&ctx->rsrc_ref_list)) {
7640 node = list_first_entry(&ctx->rsrc_ref_list,
7641 struct io_rsrc_node, node);
7642 /* recycle ref nodes in order */
7645 list_del(&node->node);
7646 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7648 io_rsrc_ref_unlock(ctx);
7651 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7654 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7656 struct io_rsrc_node *ref_node;
7658 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7662 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7667 INIT_LIST_HEAD(&ref_node->node);
7668 INIT_LIST_HEAD(&ref_node->rsrc_list);
7669 ref_node->done = false;
7673 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7674 unsigned nr_args, u64 __user *tags)
7676 __s32 __user *fds = (__s32 __user *) arg;
7685 if (nr_args > IORING_MAX_FIXED_FILES)
7687 ret = io_rsrc_node_switch_start(ctx);
7690 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7696 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7699 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7700 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7704 /* allow sparse sets */
7707 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7714 if (unlikely(!file))
7718 * Don't allow io_uring instances to be registered. If UNIX
7719 * isn't enabled, then this causes a reference cycle and this
7720 * instance can never get freed. If UNIX is enabled we'll
7721 * handle it just fine, but there's still no point in allowing
7722 * a ring fd as it doesn't support regular read/write anyway.
7724 if (file->f_op == &io_uring_fops) {
7728 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7731 ret = io_sqe_files_scm(ctx);
7733 __io_sqe_files_unregister(ctx);
7737 io_rsrc_node_switch(ctx, NULL);
7740 for (i = 0; i < ctx->nr_user_files; i++) {
7741 file = io_file_from_index(ctx, i);
7745 io_free_file_tables(&ctx->file_table, nr_args);
7746 ctx->nr_user_files = 0;
7748 io_rsrc_data_free(ctx->file_data);
7749 ctx->file_data = NULL;
7753 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7756 #if defined(CONFIG_UNIX)
7757 struct sock *sock = ctx->ring_sock->sk;
7758 struct sk_buff_head *head = &sock->sk_receive_queue;
7759 struct sk_buff *skb;
7762 * See if we can merge this file into an existing skb SCM_RIGHTS
7763 * file set. If there's no room, fall back to allocating a new skb
7764 * and filling it in.
7766 spin_lock_irq(&head->lock);
7767 skb = skb_peek(head);
7769 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7771 if (fpl->count < SCM_MAX_FD) {
7772 __skb_unlink(skb, head);
7773 spin_unlock_irq(&head->lock);
7774 fpl->fp[fpl->count] = get_file(file);
7775 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7777 spin_lock_irq(&head->lock);
7778 __skb_queue_head(head, skb);
7783 spin_unlock_irq(&head->lock);
7790 return __io_sqe_files_scm(ctx, 1, index);
7796 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7797 struct io_rsrc_node *node, void *rsrc)
7799 struct io_rsrc_put *prsrc;
7801 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7805 prsrc->tag = *io_get_tag_slot(data, idx);
7807 list_add(&prsrc->list, &node->rsrc_list);
7811 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7812 struct io_uring_rsrc_update2 *up,
7815 u64 __user *tags = u64_to_user_ptr(up->tags);
7816 __s32 __user *fds = u64_to_user_ptr(up->data);
7817 struct io_rsrc_data *data = ctx->file_data;
7818 struct io_fixed_file *file_slot;
7822 bool needs_switch = false;
7824 if (!ctx->file_data)
7826 if (up->offset + nr_args > ctx->nr_user_files)
7829 for (done = 0; done < nr_args; done++) {
7832 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7833 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7837 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7841 if (fd == IORING_REGISTER_FILES_SKIP)
7844 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7845 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7847 if (file_slot->file_ptr) {
7848 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7849 err = io_queue_rsrc_removal(data, up->offset + done,
7850 ctx->rsrc_node, file);
7853 file_slot->file_ptr = 0;
7854 needs_switch = true;
7863 * Don't allow io_uring instances to be registered. If
7864 * UNIX isn't enabled, then this causes a reference
7865 * cycle and this instance can never get freed. If UNIX
7866 * is enabled we'll handle it just fine, but there's
7867 * still no point in allowing a ring fd as it doesn't
7868 * support regular read/write anyway.
7870 if (file->f_op == &io_uring_fops) {
7875 *io_get_tag_slot(data, up->offset + done) = tag;
7876 io_fixed_file_set(file_slot, file);
7877 err = io_sqe_file_register(ctx, file, i);
7879 file_slot->file_ptr = 0;
7887 io_rsrc_node_switch(ctx, data);
7888 return done ? done : err;
7891 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7893 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7895 req = io_put_req_find_next(req);
7896 return req ? &req->work : NULL;
7899 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7900 struct task_struct *task)
7902 struct io_wq_hash *hash;
7903 struct io_wq_data data;
7904 unsigned int concurrency;
7906 hash = ctx->hash_map;
7908 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7910 return ERR_PTR(-ENOMEM);
7911 refcount_set(&hash->refs, 1);
7912 init_waitqueue_head(&hash->wait);
7913 ctx->hash_map = hash;
7918 data.free_work = io_free_work;
7919 data.do_work = io_wq_submit_work;
7921 /* Do QD, or 4 * CPUS, whatever is smallest */
7922 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7924 return io_wq_create(concurrency, &data);
7927 static int io_uring_alloc_task_context(struct task_struct *task,
7928 struct io_ring_ctx *ctx)
7930 struct io_uring_task *tctx;
7933 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7934 if (unlikely(!tctx))
7937 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7938 if (unlikely(ret)) {
7943 tctx->io_wq = io_init_wq_offload(ctx, task);
7944 if (IS_ERR(tctx->io_wq)) {
7945 ret = PTR_ERR(tctx->io_wq);
7946 percpu_counter_destroy(&tctx->inflight);
7952 init_waitqueue_head(&tctx->wait);
7953 atomic_set(&tctx->in_idle, 0);
7954 atomic_set(&tctx->inflight_tracked, 0);
7955 task->io_uring = tctx;
7956 spin_lock_init(&tctx->task_lock);
7957 INIT_WQ_LIST(&tctx->task_list);
7958 init_task_work(&tctx->task_work, tctx_task_work);
7962 void __io_uring_free(struct task_struct *tsk)
7964 struct io_uring_task *tctx = tsk->io_uring;
7966 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7967 WARN_ON_ONCE(tctx->io_wq);
7968 WARN_ON_ONCE(tctx->cached_refs);
7970 percpu_counter_destroy(&tctx->inflight);
7972 tsk->io_uring = NULL;
7975 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7976 struct io_uring_params *p)
7980 /* Retain compatibility with failing for an invalid attach attempt */
7981 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7982 IORING_SETUP_ATTACH_WQ) {
7985 f = fdget(p->wq_fd);
7989 if (f.file->f_op != &io_uring_fops)
7992 if (ctx->flags & IORING_SETUP_SQPOLL) {
7993 struct task_struct *tsk;
7994 struct io_sq_data *sqd;
7997 sqd = io_get_sq_data(p, &attached);
8003 ctx->sq_creds = get_current_cred();
8005 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8006 if (!ctx->sq_thread_idle)
8007 ctx->sq_thread_idle = HZ;
8009 io_sq_thread_park(sqd);
8010 list_add(&ctx->sqd_list, &sqd->ctx_list);
8011 io_sqd_update_thread_idle(sqd);
8012 /* don't attach to a dying SQPOLL thread, would be racy */
8013 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8014 io_sq_thread_unpark(sqd);
8021 if (p->flags & IORING_SETUP_SQ_AFF) {
8022 int cpu = p->sq_thread_cpu;
8025 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8032 sqd->task_pid = current->pid;
8033 sqd->task_tgid = current->tgid;
8034 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8041 ret = io_uring_alloc_task_context(tsk, ctx);
8042 wake_up_new_task(tsk);
8045 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8046 /* Can't have SQ_AFF without SQPOLL */
8053 complete(&ctx->sq_data->exited);
8055 io_sq_thread_finish(ctx);
8059 static inline void __io_unaccount_mem(struct user_struct *user,
8060 unsigned long nr_pages)
8062 atomic_long_sub(nr_pages, &user->locked_vm);
8065 static inline int __io_account_mem(struct user_struct *user,
8066 unsigned long nr_pages)
8068 unsigned long page_limit, cur_pages, new_pages;
8070 /* Don't allow more pages than we can safely lock */
8071 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8074 cur_pages = atomic_long_read(&user->locked_vm);
8075 new_pages = cur_pages + nr_pages;
8076 if (new_pages > page_limit)
8078 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8079 new_pages) != cur_pages);
8084 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8087 __io_unaccount_mem(ctx->user, nr_pages);
8089 if (ctx->mm_account)
8090 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8093 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8098 ret = __io_account_mem(ctx->user, nr_pages);
8103 if (ctx->mm_account)
8104 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8109 static void io_mem_free(void *ptr)
8116 page = virt_to_head_page(ptr);
8117 if (put_page_testzero(page))
8118 free_compound_page(page);
8121 static void *io_mem_alloc(size_t size)
8123 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8124 __GFP_NORETRY | __GFP_ACCOUNT;
8126 return (void *) __get_free_pages(gfp_flags, get_order(size));
8129 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8132 struct io_rings *rings;
8133 size_t off, sq_array_size;
8135 off = struct_size(rings, cqes, cq_entries);
8136 if (off == SIZE_MAX)
8140 off = ALIGN(off, SMP_CACHE_BYTES);
8148 sq_array_size = array_size(sizeof(u32), sq_entries);
8149 if (sq_array_size == SIZE_MAX)
8152 if (check_add_overflow(off, sq_array_size, &off))
8158 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8160 struct io_mapped_ubuf *imu = *slot;
8163 if (imu != ctx->dummy_ubuf) {
8164 for (i = 0; i < imu->nr_bvecs; i++)
8165 unpin_user_page(imu->bvec[i].bv_page);
8166 if (imu->acct_pages)
8167 io_unaccount_mem(ctx, imu->acct_pages);
8173 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8175 io_buffer_unmap(ctx, &prsrc->buf);
8179 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8183 for (i = 0; i < ctx->nr_user_bufs; i++)
8184 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8185 kfree(ctx->user_bufs);
8186 io_rsrc_data_free(ctx->buf_data);
8187 ctx->user_bufs = NULL;
8188 ctx->buf_data = NULL;
8189 ctx->nr_user_bufs = 0;
8192 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8199 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8201 __io_sqe_buffers_unregister(ctx);
8205 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8206 void __user *arg, unsigned index)
8208 struct iovec __user *src;
8210 #ifdef CONFIG_COMPAT
8212 struct compat_iovec __user *ciovs;
8213 struct compat_iovec ciov;
8215 ciovs = (struct compat_iovec __user *) arg;
8216 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8219 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8220 dst->iov_len = ciov.iov_len;
8224 src = (struct iovec __user *) arg;
8225 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8231 * Not super efficient, but this is just a registration time. And we do cache
8232 * the last compound head, so generally we'll only do a full search if we don't
8235 * We check if the given compound head page has already been accounted, to
8236 * avoid double accounting it. This allows us to account the full size of the
8237 * page, not just the constituent pages of a huge page.
8239 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8240 int nr_pages, struct page *hpage)
8244 /* check current page array */
8245 for (i = 0; i < nr_pages; i++) {
8246 if (!PageCompound(pages[i]))
8248 if (compound_head(pages[i]) == hpage)
8252 /* check previously registered pages */
8253 for (i = 0; i < ctx->nr_user_bufs; i++) {
8254 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8256 for (j = 0; j < imu->nr_bvecs; j++) {
8257 if (!PageCompound(imu->bvec[j].bv_page))
8259 if (compound_head(imu->bvec[j].bv_page) == hpage)
8267 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8268 int nr_pages, struct io_mapped_ubuf *imu,
8269 struct page **last_hpage)
8273 imu->acct_pages = 0;
8274 for (i = 0; i < nr_pages; i++) {
8275 if (!PageCompound(pages[i])) {
8280 hpage = compound_head(pages[i]);
8281 if (hpage == *last_hpage)
8283 *last_hpage = hpage;
8284 if (headpage_already_acct(ctx, pages, i, hpage))
8286 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8290 if (!imu->acct_pages)
8293 ret = io_account_mem(ctx, imu->acct_pages);
8295 imu->acct_pages = 0;
8299 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8300 struct io_mapped_ubuf **pimu,
8301 struct page **last_hpage)
8303 struct io_mapped_ubuf *imu = NULL;
8304 struct vm_area_struct **vmas = NULL;
8305 struct page **pages = NULL;
8306 unsigned long off, start, end, ubuf;
8308 int ret, pret, nr_pages, i;
8310 if (!iov->iov_base) {
8311 *pimu = ctx->dummy_ubuf;
8315 ubuf = (unsigned long) iov->iov_base;
8316 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8317 start = ubuf >> PAGE_SHIFT;
8318 nr_pages = end - start;
8323 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8327 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8332 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8337 mmap_read_lock(current->mm);
8338 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8340 if (pret == nr_pages) {
8341 /* don't support file backed memory */
8342 for (i = 0; i < nr_pages; i++) {
8343 struct vm_area_struct *vma = vmas[i];
8345 if (vma_is_shmem(vma))
8348 !is_file_hugepages(vma->vm_file)) {
8354 ret = pret < 0 ? pret : -EFAULT;
8356 mmap_read_unlock(current->mm);
8359 * if we did partial map, or found file backed vmas,
8360 * release any pages we did get
8363 unpin_user_pages(pages, pret);
8367 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8369 unpin_user_pages(pages, pret);
8373 off = ubuf & ~PAGE_MASK;
8374 size = iov->iov_len;
8375 for (i = 0; i < nr_pages; i++) {
8378 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8379 imu->bvec[i].bv_page = pages[i];
8380 imu->bvec[i].bv_len = vec_len;
8381 imu->bvec[i].bv_offset = off;
8385 /* store original address for later verification */
8387 imu->ubuf_end = ubuf + iov->iov_len;
8388 imu->nr_bvecs = nr_pages;
8399 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8401 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8402 return ctx->user_bufs ? 0 : -ENOMEM;
8405 static int io_buffer_validate(struct iovec *iov)
8407 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8410 * Don't impose further limits on the size and buffer
8411 * constraints here, we'll -EINVAL later when IO is
8412 * submitted if they are wrong.
8415 return iov->iov_len ? -EFAULT : 0;
8419 /* arbitrary limit, but we need something */
8420 if (iov->iov_len > SZ_1G)
8423 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8429 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8430 unsigned int nr_args, u64 __user *tags)
8432 struct page *last_hpage = NULL;
8433 struct io_rsrc_data *data;
8439 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8441 ret = io_rsrc_node_switch_start(ctx);
8444 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8447 ret = io_buffers_map_alloc(ctx, nr_args);
8449 io_rsrc_data_free(data);
8453 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8454 ret = io_copy_iov(ctx, &iov, arg, i);
8457 ret = io_buffer_validate(&iov);
8460 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8465 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8471 WARN_ON_ONCE(ctx->buf_data);
8473 ctx->buf_data = data;
8475 __io_sqe_buffers_unregister(ctx);
8477 io_rsrc_node_switch(ctx, NULL);
8481 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8482 struct io_uring_rsrc_update2 *up,
8483 unsigned int nr_args)
8485 u64 __user *tags = u64_to_user_ptr(up->tags);
8486 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8487 struct page *last_hpage = NULL;
8488 bool needs_switch = false;
8494 if (up->offset + nr_args > ctx->nr_user_bufs)
8497 for (done = 0; done < nr_args; done++) {
8498 struct io_mapped_ubuf *imu;
8499 int offset = up->offset + done;
8502 err = io_copy_iov(ctx, &iov, iovs, done);
8505 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8509 err = io_buffer_validate(&iov);
8512 if (!iov.iov_base && tag) {
8516 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8520 i = array_index_nospec(offset, ctx->nr_user_bufs);
8521 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8522 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8523 ctx->rsrc_node, ctx->user_bufs[i]);
8524 if (unlikely(err)) {
8525 io_buffer_unmap(ctx, &imu);
8528 ctx->user_bufs[i] = NULL;
8529 needs_switch = true;
8532 ctx->user_bufs[i] = imu;
8533 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8537 io_rsrc_node_switch(ctx, ctx->buf_data);
8538 return done ? done : err;
8541 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8543 __s32 __user *fds = arg;
8549 if (copy_from_user(&fd, fds, sizeof(*fds)))
8552 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8553 if (IS_ERR(ctx->cq_ev_fd)) {
8554 int ret = PTR_ERR(ctx->cq_ev_fd);
8556 ctx->cq_ev_fd = NULL;
8563 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8565 if (ctx->cq_ev_fd) {
8566 eventfd_ctx_put(ctx->cq_ev_fd);
8567 ctx->cq_ev_fd = NULL;
8574 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8576 struct io_buffer *buf;
8577 unsigned long index;
8579 xa_for_each(&ctx->io_buffers, index, buf)
8580 __io_remove_buffers(ctx, buf, index, -1U);
8583 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8585 struct io_kiocb *req, *nxt;
8587 list_for_each_entry_safe(req, nxt, list, compl.list) {
8588 if (tsk && req->task != tsk)
8590 list_del(&req->compl.list);
8591 kmem_cache_free(req_cachep, req);
8595 static void io_req_caches_free(struct io_ring_ctx *ctx)
8597 struct io_submit_state *submit_state = &ctx->submit_state;
8598 struct io_comp_state *cs = &ctx->submit_state.comp;
8600 mutex_lock(&ctx->uring_lock);
8602 if (submit_state->free_reqs) {
8603 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8604 submit_state->reqs);
8605 submit_state->free_reqs = 0;
8608 io_flush_cached_locked_reqs(ctx, cs);
8609 io_req_cache_free(&cs->free_list, NULL);
8610 mutex_unlock(&ctx->uring_lock);
8613 static bool io_wait_rsrc_data(struct io_rsrc_data *data)
8617 if (!atomic_dec_and_test(&data->refs))
8618 wait_for_completion(&data->done);
8622 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8624 io_sq_thread_finish(ctx);
8626 if (ctx->mm_account) {
8627 mmdrop(ctx->mm_account);
8628 ctx->mm_account = NULL;
8631 mutex_lock(&ctx->uring_lock);
8632 if (io_wait_rsrc_data(ctx->buf_data))
8633 __io_sqe_buffers_unregister(ctx);
8634 if (io_wait_rsrc_data(ctx->file_data))
8635 __io_sqe_files_unregister(ctx);
8637 __io_cqring_overflow_flush(ctx, true);
8638 mutex_unlock(&ctx->uring_lock);
8639 io_eventfd_unregister(ctx);
8640 io_destroy_buffers(ctx);
8642 put_cred(ctx->sq_creds);
8644 /* there are no registered resources left, nobody uses it */
8646 io_rsrc_node_destroy(ctx->rsrc_node);
8647 if (ctx->rsrc_backup_node)
8648 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8649 flush_delayed_work(&ctx->rsrc_put_work);
8651 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8652 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8654 #if defined(CONFIG_UNIX)
8655 if (ctx->ring_sock) {
8656 ctx->ring_sock->file = NULL; /* so that iput() is called */
8657 sock_release(ctx->ring_sock);
8661 io_mem_free(ctx->rings);
8662 io_mem_free(ctx->sq_sqes);
8664 percpu_ref_exit(&ctx->refs);
8665 free_uid(ctx->user);
8666 io_req_caches_free(ctx);
8668 io_wq_put_hash(ctx->hash_map);
8669 kfree(ctx->cancel_hash);
8670 kfree(ctx->dummy_ubuf);
8674 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8676 struct io_ring_ctx *ctx = file->private_data;
8679 poll_wait(file, &ctx->poll_wait, wait);
8681 * synchronizes with barrier from wq_has_sleeper call in
8685 if (!io_sqring_full(ctx))
8686 mask |= EPOLLOUT | EPOLLWRNORM;
8689 * Don't flush cqring overflow list here, just do a simple check.
8690 * Otherwise there could possible be ABBA deadlock:
8693 * lock(&ctx->uring_lock);
8695 * lock(&ctx->uring_lock);
8698 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8699 * pushs them to do the flush.
8701 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8702 mask |= EPOLLIN | EPOLLRDNORM;
8707 static int io_uring_fasync(int fd, struct file *file, int on)
8709 struct io_ring_ctx *ctx = file->private_data;
8711 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8714 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8716 const struct cred *creds;
8718 creds = xa_erase(&ctx->personalities, id);
8727 struct io_tctx_exit {
8728 struct callback_head task_work;
8729 struct completion completion;
8730 struct io_ring_ctx *ctx;
8733 static void io_tctx_exit_cb(struct callback_head *cb)
8735 struct io_uring_task *tctx = current->io_uring;
8736 struct io_tctx_exit *work;
8738 work = container_of(cb, struct io_tctx_exit, task_work);
8740 * When @in_idle, we're in cancellation and it's racy to remove the
8741 * node. It'll be removed by the end of cancellation, just ignore it.
8743 if (!atomic_read(&tctx->in_idle))
8744 io_uring_del_tctx_node((unsigned long)work->ctx);
8745 complete(&work->completion);
8748 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8750 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8752 return req->ctx == data;
8755 static void io_ring_exit_work(struct work_struct *work)
8757 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8758 unsigned long timeout = jiffies + HZ * 60 * 5;
8759 struct io_tctx_exit exit;
8760 struct io_tctx_node *node;
8764 * If we're doing polled IO and end up having requests being
8765 * submitted async (out-of-line), then completions can come in while
8766 * we're waiting for refs to drop. We need to reap these manually,
8767 * as nobody else will be looking for them.
8770 io_uring_try_cancel_requests(ctx, NULL, true);
8772 struct io_sq_data *sqd = ctx->sq_data;
8773 struct task_struct *tsk;
8775 io_sq_thread_park(sqd);
8777 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8778 io_wq_cancel_cb(tsk->io_uring->io_wq,
8779 io_cancel_ctx_cb, ctx, true);
8780 io_sq_thread_unpark(sqd);
8783 WARN_ON_ONCE(time_after(jiffies, timeout));
8784 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8786 init_completion(&exit.completion);
8787 init_task_work(&exit.task_work, io_tctx_exit_cb);
8790 * Some may use context even when all refs and requests have been put,
8791 * and they are free to do so while still holding uring_lock or
8792 * completion_lock, see io_req_task_submit(). Apart from other work,
8793 * this lock/unlock section also waits them to finish.
8795 mutex_lock(&ctx->uring_lock);
8796 while (!list_empty(&ctx->tctx_list)) {
8797 WARN_ON_ONCE(time_after(jiffies, timeout));
8799 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8801 /* don't spin on a single task if cancellation failed */
8802 list_rotate_left(&ctx->tctx_list);
8803 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8804 if (WARN_ON_ONCE(ret))
8806 wake_up_process(node->task);
8808 mutex_unlock(&ctx->uring_lock);
8809 wait_for_completion(&exit.completion);
8810 mutex_lock(&ctx->uring_lock);
8812 mutex_unlock(&ctx->uring_lock);
8813 spin_lock_irq(&ctx->completion_lock);
8814 spin_unlock_irq(&ctx->completion_lock);
8816 io_ring_ctx_free(ctx);
8819 /* Returns true if we found and killed one or more timeouts */
8820 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8823 struct io_kiocb *req, *tmp;
8826 spin_lock_irq(&ctx->completion_lock);
8827 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8828 if (io_match_task(req, tsk, cancel_all)) {
8829 io_kill_timeout(req, -ECANCELED);
8834 io_commit_cqring(ctx);
8835 spin_unlock_irq(&ctx->completion_lock);
8837 io_cqring_ev_posted(ctx);
8838 return canceled != 0;
8841 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8843 unsigned long index;
8844 struct creds *creds;
8846 mutex_lock(&ctx->uring_lock);
8847 percpu_ref_kill(&ctx->refs);
8849 __io_cqring_overflow_flush(ctx, true);
8850 xa_for_each(&ctx->personalities, index, creds)
8851 io_unregister_personality(ctx, index);
8852 mutex_unlock(&ctx->uring_lock);
8854 io_kill_timeouts(ctx, NULL, true);
8855 io_poll_remove_all(ctx, NULL, true);
8857 /* if we failed setting up the ctx, we might not have any rings */
8858 io_iopoll_try_reap_events(ctx);
8860 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8862 * Use system_unbound_wq to avoid spawning tons of event kworkers
8863 * if we're exiting a ton of rings at the same time. It just adds
8864 * noise and overhead, there's no discernable change in runtime
8865 * over using system_wq.
8867 queue_work(system_unbound_wq, &ctx->exit_work);
8870 static int io_uring_release(struct inode *inode, struct file *file)
8872 struct io_ring_ctx *ctx = file->private_data;
8874 file->private_data = NULL;
8875 io_ring_ctx_wait_and_kill(ctx);
8879 struct io_task_cancel {
8880 struct task_struct *task;
8884 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8886 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8887 struct io_task_cancel *cancel = data;
8890 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8891 unsigned long flags;
8892 struct io_ring_ctx *ctx = req->ctx;
8894 /* protect against races with linked timeouts */
8895 spin_lock_irqsave(&ctx->completion_lock, flags);
8896 ret = io_match_task(req, cancel->task, cancel->all);
8897 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8899 ret = io_match_task(req, cancel->task, cancel->all);
8904 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8905 struct task_struct *task, bool cancel_all)
8907 struct io_defer_entry *de;
8910 spin_lock_irq(&ctx->completion_lock);
8911 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8912 if (io_match_task(de->req, task, cancel_all)) {
8913 list_cut_position(&list, &ctx->defer_list, &de->list);
8917 spin_unlock_irq(&ctx->completion_lock);
8918 if (list_empty(&list))
8921 while (!list_empty(&list)) {
8922 de = list_first_entry(&list, struct io_defer_entry, list);
8923 list_del_init(&de->list);
8924 io_req_complete_failed(de->req, -ECANCELED);
8930 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8932 struct io_tctx_node *node;
8933 enum io_wq_cancel cret;
8936 mutex_lock(&ctx->uring_lock);
8937 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8938 struct io_uring_task *tctx = node->task->io_uring;
8941 * io_wq will stay alive while we hold uring_lock, because it's
8942 * killed after ctx nodes, which requires to take the lock.
8944 if (!tctx || !tctx->io_wq)
8946 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8947 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8949 mutex_unlock(&ctx->uring_lock);
8954 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8955 struct task_struct *task,
8958 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
8959 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8962 enum io_wq_cancel cret;
8966 ret |= io_uring_try_cancel_iowq(ctx);
8967 } else if (tctx && tctx->io_wq) {
8969 * Cancels requests of all rings, not only @ctx, but
8970 * it's fine as the task is in exit/exec.
8972 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8974 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8977 /* SQPOLL thread does its own polling */
8978 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
8979 (ctx->sq_data && ctx->sq_data->thread == current)) {
8980 while (!list_empty_careful(&ctx->iopoll_list)) {
8981 io_iopoll_try_reap_events(ctx);
8986 ret |= io_cancel_defer_files(ctx, task, cancel_all);
8987 ret |= io_poll_remove_all(ctx, task, cancel_all);
8988 ret |= io_kill_timeouts(ctx, task, cancel_all);
8990 ret |= io_run_task_work();
8997 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
8999 struct io_uring_task *tctx = current->io_uring;
9000 struct io_tctx_node *node;
9003 if (unlikely(!tctx)) {
9004 ret = io_uring_alloc_task_context(current, ctx);
9007 tctx = current->io_uring;
9009 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9010 node = kmalloc(sizeof(*node), GFP_KERNEL);
9014 node->task = current;
9016 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9023 mutex_lock(&ctx->uring_lock);
9024 list_add(&node->ctx_node, &ctx->tctx_list);
9025 mutex_unlock(&ctx->uring_lock);
9032 * Note that this task has used io_uring. We use it for cancelation purposes.
9034 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9036 struct io_uring_task *tctx = current->io_uring;
9038 if (likely(tctx && tctx->last == ctx))
9040 return __io_uring_add_tctx_node(ctx);
9044 * Remove this io_uring_file -> task mapping.
9046 static void io_uring_del_tctx_node(unsigned long index)
9048 struct io_uring_task *tctx = current->io_uring;
9049 struct io_tctx_node *node;
9053 node = xa_erase(&tctx->xa, index);
9057 WARN_ON_ONCE(current != node->task);
9058 WARN_ON_ONCE(list_empty(&node->ctx_node));
9060 mutex_lock(&node->ctx->uring_lock);
9061 list_del(&node->ctx_node);
9062 mutex_unlock(&node->ctx->uring_lock);
9064 if (tctx->last == node->ctx)
9069 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9071 struct io_wq *wq = tctx->io_wq;
9072 struct io_tctx_node *node;
9073 unsigned long index;
9075 xa_for_each(&tctx->xa, index, node)
9076 io_uring_del_tctx_node(index);
9079 * Must be after io_uring_del_task_file() (removes nodes under
9080 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9083 io_wq_put_and_exit(wq);
9087 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9090 return atomic_read(&tctx->inflight_tracked);
9091 return percpu_counter_sum(&tctx->inflight);
9094 static void io_uring_drop_tctx_refs(struct task_struct *task)
9096 struct io_uring_task *tctx = task->io_uring;
9097 unsigned int refs = tctx->cached_refs;
9099 tctx->cached_refs = 0;
9100 percpu_counter_sub(&tctx->inflight, refs);
9101 put_task_struct_many(task, refs);
9105 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9106 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9108 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9110 struct io_uring_task *tctx = current->io_uring;
9111 struct io_ring_ctx *ctx;
9115 WARN_ON_ONCE(sqd && sqd->thread != current);
9117 if (!current->io_uring)
9120 io_wq_exit_start(tctx->io_wq);
9122 io_uring_drop_tctx_refs(current);
9123 atomic_inc(&tctx->in_idle);
9125 /* read completions before cancelations */
9126 inflight = tctx_inflight(tctx, !cancel_all);
9131 struct io_tctx_node *node;
9132 unsigned long index;
9134 xa_for_each(&tctx->xa, index, node) {
9135 /* sqpoll task will cancel all its requests */
9136 if (node->ctx->sq_data)
9138 io_uring_try_cancel_requests(node->ctx, current,
9142 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9143 io_uring_try_cancel_requests(ctx, current,
9147 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9149 * If we've seen completions, retry without waiting. This
9150 * avoids a race where a completion comes in before we did
9151 * prepare_to_wait().
9153 if (inflight == tctx_inflight(tctx, !cancel_all))
9155 finish_wait(&tctx->wait, &wait);
9157 atomic_dec(&tctx->in_idle);
9159 io_uring_clean_tctx(tctx);
9161 /* for exec all current's requests should be gone, kill tctx */
9162 __io_uring_free(current);
9166 void __io_uring_cancel(struct files_struct *files)
9168 io_uring_cancel_generic(!files, NULL);
9171 static void *io_uring_validate_mmap_request(struct file *file,
9172 loff_t pgoff, size_t sz)
9174 struct io_ring_ctx *ctx = file->private_data;
9175 loff_t offset = pgoff << PAGE_SHIFT;
9180 case IORING_OFF_SQ_RING:
9181 case IORING_OFF_CQ_RING:
9184 case IORING_OFF_SQES:
9188 return ERR_PTR(-EINVAL);
9191 page = virt_to_head_page(ptr);
9192 if (sz > page_size(page))
9193 return ERR_PTR(-EINVAL);
9200 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9202 size_t sz = vma->vm_end - vma->vm_start;
9206 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9208 return PTR_ERR(ptr);
9210 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9211 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9214 #else /* !CONFIG_MMU */
9216 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9218 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9221 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9223 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9226 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9227 unsigned long addr, unsigned long len,
9228 unsigned long pgoff, unsigned long flags)
9232 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9234 return PTR_ERR(ptr);
9236 return (unsigned long) ptr;
9239 #endif /* !CONFIG_MMU */
9241 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9246 if (!io_sqring_full(ctx))
9248 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9250 if (!io_sqring_full(ctx))
9253 } while (!signal_pending(current));
9255 finish_wait(&ctx->sqo_sq_wait, &wait);
9259 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9260 struct __kernel_timespec __user **ts,
9261 const sigset_t __user **sig)
9263 struct io_uring_getevents_arg arg;
9266 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9267 * is just a pointer to the sigset_t.
9269 if (!(flags & IORING_ENTER_EXT_ARG)) {
9270 *sig = (const sigset_t __user *) argp;
9276 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9277 * timespec and sigset_t pointers if good.
9279 if (*argsz != sizeof(arg))
9281 if (copy_from_user(&arg, argp, sizeof(arg)))
9283 *sig = u64_to_user_ptr(arg.sigmask);
9284 *argsz = arg.sigmask_sz;
9285 *ts = u64_to_user_ptr(arg.ts);
9289 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9290 u32, min_complete, u32, flags, const void __user *, argp,
9293 struct io_ring_ctx *ctx;
9300 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9301 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9305 if (unlikely(!f.file))
9309 if (unlikely(f.file->f_op != &io_uring_fops))
9313 ctx = f.file->private_data;
9314 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9318 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9322 * For SQ polling, the thread will do all submissions and completions.
9323 * Just return the requested submit count, and wake the thread if
9327 if (ctx->flags & IORING_SETUP_SQPOLL) {
9328 io_cqring_overflow_flush(ctx, false);
9331 if (unlikely(ctx->sq_data->thread == NULL))
9333 if (flags & IORING_ENTER_SQ_WAKEUP)
9334 wake_up(&ctx->sq_data->wait);
9335 if (flags & IORING_ENTER_SQ_WAIT) {
9336 ret = io_sqpoll_wait_sq(ctx);
9340 submitted = to_submit;
9341 } else if (to_submit) {
9342 ret = io_uring_add_tctx_node(ctx);
9345 mutex_lock(&ctx->uring_lock);
9346 submitted = io_submit_sqes(ctx, to_submit);
9347 mutex_unlock(&ctx->uring_lock);
9349 if (submitted != to_submit)
9352 if (flags & IORING_ENTER_GETEVENTS) {
9353 const sigset_t __user *sig;
9354 struct __kernel_timespec __user *ts;
9356 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9360 min_complete = min(min_complete, ctx->cq_entries);
9363 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9364 * space applications don't need to do io completion events
9365 * polling again, they can rely on io_sq_thread to do polling
9366 * work, which can reduce cpu usage and uring_lock contention.
9368 if (ctx->flags & IORING_SETUP_IOPOLL &&
9369 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9370 ret = io_iopoll_check(ctx, min_complete);
9372 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9377 percpu_ref_put(&ctx->refs);
9380 return submitted ? submitted : ret;
9383 #ifdef CONFIG_PROC_FS
9384 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9385 const struct cred *cred)
9387 struct user_namespace *uns = seq_user_ns(m);
9388 struct group_info *gi;
9393 seq_printf(m, "%5d\n", id);
9394 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9395 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9396 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9397 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9398 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9399 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9400 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9401 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9402 seq_puts(m, "\n\tGroups:\t");
9403 gi = cred->group_info;
9404 for (g = 0; g < gi->ngroups; g++) {
9405 seq_put_decimal_ull(m, g ? " " : "",
9406 from_kgid_munged(uns, gi->gid[g]));
9408 seq_puts(m, "\n\tCapEff:\t");
9409 cap = cred->cap_effective;
9410 CAP_FOR_EACH_U32(__capi)
9411 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9416 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9418 struct io_sq_data *sq = NULL;
9423 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9424 * since fdinfo case grabs it in the opposite direction of normal use
9425 * cases. If we fail to get the lock, we just don't iterate any
9426 * structures that could be going away outside the io_uring mutex.
9428 has_lock = mutex_trylock(&ctx->uring_lock);
9430 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9436 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9437 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9438 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9439 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9440 struct file *f = io_file_from_index(ctx, i);
9443 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9445 seq_printf(m, "%5u: <none>\n", i);
9447 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9448 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9449 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9450 unsigned int len = buf->ubuf_end - buf->ubuf;
9452 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9454 if (has_lock && !xa_empty(&ctx->personalities)) {
9455 unsigned long index;
9456 const struct cred *cred;
9458 seq_printf(m, "Personalities:\n");
9459 xa_for_each(&ctx->personalities, index, cred)
9460 io_uring_show_cred(m, index, cred);
9462 seq_printf(m, "PollList:\n");
9463 spin_lock_irq(&ctx->completion_lock);
9464 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9465 struct hlist_head *list = &ctx->cancel_hash[i];
9466 struct io_kiocb *req;
9468 hlist_for_each_entry(req, list, hash_node)
9469 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9470 req->task->task_works != NULL);
9472 spin_unlock_irq(&ctx->completion_lock);
9474 mutex_unlock(&ctx->uring_lock);
9477 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9479 struct io_ring_ctx *ctx = f->private_data;
9481 if (percpu_ref_tryget(&ctx->refs)) {
9482 __io_uring_show_fdinfo(ctx, m);
9483 percpu_ref_put(&ctx->refs);
9488 static const struct file_operations io_uring_fops = {
9489 .release = io_uring_release,
9490 .mmap = io_uring_mmap,
9492 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9493 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9495 .poll = io_uring_poll,
9496 .fasync = io_uring_fasync,
9497 #ifdef CONFIG_PROC_FS
9498 .show_fdinfo = io_uring_show_fdinfo,
9502 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9503 struct io_uring_params *p)
9505 struct io_rings *rings;
9506 size_t size, sq_array_offset;
9508 /* make sure these are sane, as we already accounted them */
9509 ctx->sq_entries = p->sq_entries;
9510 ctx->cq_entries = p->cq_entries;
9512 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9513 if (size == SIZE_MAX)
9516 rings = io_mem_alloc(size);
9521 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9522 rings->sq_ring_mask = p->sq_entries - 1;
9523 rings->cq_ring_mask = p->cq_entries - 1;
9524 rings->sq_ring_entries = p->sq_entries;
9525 rings->cq_ring_entries = p->cq_entries;
9527 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9528 if (size == SIZE_MAX) {
9529 io_mem_free(ctx->rings);
9534 ctx->sq_sqes = io_mem_alloc(size);
9535 if (!ctx->sq_sqes) {
9536 io_mem_free(ctx->rings);
9544 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9548 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9552 ret = io_uring_add_tctx_node(ctx);
9557 fd_install(fd, file);
9562 * Allocate an anonymous fd, this is what constitutes the application
9563 * visible backing of an io_uring instance. The application mmaps this
9564 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9565 * we have to tie this fd to a socket for file garbage collection purposes.
9567 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9570 #if defined(CONFIG_UNIX)
9573 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9576 return ERR_PTR(ret);
9579 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9580 O_RDWR | O_CLOEXEC);
9581 #if defined(CONFIG_UNIX)
9583 sock_release(ctx->ring_sock);
9584 ctx->ring_sock = NULL;
9586 ctx->ring_sock->file = file;
9592 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9593 struct io_uring_params __user *params)
9595 struct io_ring_ctx *ctx;
9601 if (entries > IORING_MAX_ENTRIES) {
9602 if (!(p->flags & IORING_SETUP_CLAMP))
9604 entries = IORING_MAX_ENTRIES;
9608 * Use twice as many entries for the CQ ring. It's possible for the
9609 * application to drive a higher depth than the size of the SQ ring,
9610 * since the sqes are only used at submission time. This allows for
9611 * some flexibility in overcommitting a bit. If the application has
9612 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9613 * of CQ ring entries manually.
9615 p->sq_entries = roundup_pow_of_two(entries);
9616 if (p->flags & IORING_SETUP_CQSIZE) {
9618 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9619 * to a power-of-two, if it isn't already. We do NOT impose
9620 * any cq vs sq ring sizing.
9624 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9625 if (!(p->flags & IORING_SETUP_CLAMP))
9627 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9629 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9630 if (p->cq_entries < p->sq_entries)
9633 p->cq_entries = 2 * p->sq_entries;
9636 ctx = io_ring_ctx_alloc(p);
9639 ctx->compat = in_compat_syscall();
9640 if (!capable(CAP_IPC_LOCK))
9641 ctx->user = get_uid(current_user());
9644 * This is just grabbed for accounting purposes. When a process exits,
9645 * the mm is exited and dropped before the files, hence we need to hang
9646 * on to this mm purely for the purposes of being able to unaccount
9647 * memory (locked/pinned vm). It's not used for anything else.
9649 mmgrab(current->mm);
9650 ctx->mm_account = current->mm;
9652 ret = io_allocate_scq_urings(ctx, p);
9656 ret = io_sq_offload_create(ctx, p);
9659 /* always set a rsrc node */
9660 ret = io_rsrc_node_switch_start(ctx);
9663 io_rsrc_node_switch(ctx, NULL);
9665 memset(&p->sq_off, 0, sizeof(p->sq_off));
9666 p->sq_off.head = offsetof(struct io_rings, sq.head);
9667 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9668 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9669 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9670 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9671 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9672 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9674 memset(&p->cq_off, 0, sizeof(p->cq_off));
9675 p->cq_off.head = offsetof(struct io_rings, cq.head);
9676 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9677 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9678 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9679 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9680 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9681 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9683 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9684 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9685 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9686 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9687 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9688 IORING_FEAT_RSRC_TAGS;
9690 if (copy_to_user(params, p, sizeof(*p))) {
9695 file = io_uring_get_file(ctx);
9697 ret = PTR_ERR(file);
9702 * Install ring fd as the very last thing, so we don't risk someone
9703 * having closed it before we finish setup
9705 ret = io_uring_install_fd(ctx, file);
9707 /* fput will clean it up */
9712 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9715 io_ring_ctx_wait_and_kill(ctx);
9720 * Sets up an aio uring context, and returns the fd. Applications asks for a
9721 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9722 * params structure passed in.
9724 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9726 struct io_uring_params p;
9729 if (copy_from_user(&p, params, sizeof(p)))
9731 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9736 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9737 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9738 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9739 IORING_SETUP_R_DISABLED))
9742 return io_uring_create(entries, &p, params);
9745 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9746 struct io_uring_params __user *, params)
9748 return io_uring_setup(entries, params);
9751 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9753 struct io_uring_probe *p;
9757 size = struct_size(p, ops, nr_args);
9758 if (size == SIZE_MAX)
9760 p = kzalloc(size, GFP_KERNEL);
9765 if (copy_from_user(p, arg, size))
9768 if (memchr_inv(p, 0, size))
9771 p->last_op = IORING_OP_LAST - 1;
9772 if (nr_args > IORING_OP_LAST)
9773 nr_args = IORING_OP_LAST;
9775 for (i = 0; i < nr_args; i++) {
9777 if (!io_op_defs[i].not_supported)
9778 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9783 if (copy_to_user(arg, p, size))
9790 static int io_register_personality(struct io_ring_ctx *ctx)
9792 const struct cred *creds;
9796 creds = get_current_cred();
9798 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9799 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9806 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9807 unsigned int nr_args)
9809 struct io_uring_restriction *res;
9813 /* Restrictions allowed only if rings started disabled */
9814 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9817 /* We allow only a single restrictions registration */
9818 if (ctx->restrictions.registered)
9821 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9824 size = array_size(nr_args, sizeof(*res));
9825 if (size == SIZE_MAX)
9828 res = memdup_user(arg, size);
9830 return PTR_ERR(res);
9834 for (i = 0; i < nr_args; i++) {
9835 switch (res[i].opcode) {
9836 case IORING_RESTRICTION_REGISTER_OP:
9837 if (res[i].register_op >= IORING_REGISTER_LAST) {
9842 __set_bit(res[i].register_op,
9843 ctx->restrictions.register_op);
9845 case IORING_RESTRICTION_SQE_OP:
9846 if (res[i].sqe_op >= IORING_OP_LAST) {
9851 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9853 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9854 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9856 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9857 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9866 /* Reset all restrictions if an error happened */
9868 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9870 ctx->restrictions.registered = true;
9876 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9878 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9881 if (ctx->restrictions.registered)
9882 ctx->restricted = 1;
9884 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9885 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9886 wake_up(&ctx->sq_data->wait);
9890 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9891 struct io_uring_rsrc_update2 *up,
9899 if (check_add_overflow(up->offset, nr_args, &tmp))
9901 err = io_rsrc_node_switch_start(ctx);
9906 case IORING_RSRC_FILE:
9907 return __io_sqe_files_update(ctx, up, nr_args);
9908 case IORING_RSRC_BUFFER:
9909 return __io_sqe_buffers_update(ctx, up, nr_args);
9914 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9917 struct io_uring_rsrc_update2 up;
9921 memset(&up, 0, sizeof(up));
9922 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9924 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9927 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9928 unsigned size, unsigned type)
9930 struct io_uring_rsrc_update2 up;
9932 if (size != sizeof(up))
9934 if (copy_from_user(&up, arg, sizeof(up)))
9936 if (!up.nr || up.resv)
9938 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9941 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9942 unsigned int size, unsigned int type)
9944 struct io_uring_rsrc_register rr;
9946 /* keep it extendible */
9947 if (size != sizeof(rr))
9950 memset(&rr, 0, sizeof(rr));
9951 if (copy_from_user(&rr, arg, size))
9953 if (!rr.nr || rr.resv || rr.resv2)
9957 case IORING_RSRC_FILE:
9958 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
9959 rr.nr, u64_to_user_ptr(rr.tags));
9960 case IORING_RSRC_BUFFER:
9961 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
9962 rr.nr, u64_to_user_ptr(rr.tags));
9967 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
9970 struct io_uring_task *tctx = current->io_uring;
9971 cpumask_var_t new_mask;
9974 if (!tctx || !tctx->io_wq)
9977 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
9980 cpumask_clear(new_mask);
9981 if (len > cpumask_size())
9982 len = cpumask_size();
9984 if (copy_from_user(new_mask, arg, len)) {
9985 free_cpumask_var(new_mask);
9989 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
9990 free_cpumask_var(new_mask);
9994 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
9996 struct io_uring_task *tctx = current->io_uring;
9998 if (!tctx || !tctx->io_wq)
10001 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10004 static bool io_register_op_must_quiesce(int op)
10007 case IORING_REGISTER_BUFFERS:
10008 case IORING_UNREGISTER_BUFFERS:
10009 case IORING_REGISTER_FILES:
10010 case IORING_UNREGISTER_FILES:
10011 case IORING_REGISTER_FILES_UPDATE:
10012 case IORING_REGISTER_PROBE:
10013 case IORING_REGISTER_PERSONALITY:
10014 case IORING_UNREGISTER_PERSONALITY:
10015 case IORING_REGISTER_FILES2:
10016 case IORING_REGISTER_FILES_UPDATE2:
10017 case IORING_REGISTER_BUFFERS2:
10018 case IORING_REGISTER_BUFFERS_UPDATE:
10019 case IORING_REGISTER_IOWQ_AFF:
10020 case IORING_UNREGISTER_IOWQ_AFF:
10027 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10028 void __user *arg, unsigned nr_args)
10029 __releases(ctx->uring_lock)
10030 __acquires(ctx->uring_lock)
10035 * We're inside the ring mutex, if the ref is already dying, then
10036 * someone else killed the ctx or is already going through
10037 * io_uring_register().
10039 if (percpu_ref_is_dying(&ctx->refs))
10042 if (ctx->restricted) {
10043 if (opcode >= IORING_REGISTER_LAST)
10045 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10046 if (!test_bit(opcode, ctx->restrictions.register_op))
10050 if (io_register_op_must_quiesce(opcode)) {
10051 percpu_ref_kill(&ctx->refs);
10054 * Drop uring mutex before waiting for references to exit. If
10055 * another thread is currently inside io_uring_enter() it might
10056 * need to grab the uring_lock to make progress. If we hold it
10057 * here across the drain wait, then we can deadlock. It's safe
10058 * to drop the mutex here, since no new references will come in
10059 * after we've killed the percpu ref.
10061 mutex_unlock(&ctx->uring_lock);
10063 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10066 ret = io_run_task_work_sig();
10070 mutex_lock(&ctx->uring_lock);
10073 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10079 case IORING_REGISTER_BUFFERS:
10080 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10082 case IORING_UNREGISTER_BUFFERS:
10084 if (arg || nr_args)
10086 ret = io_sqe_buffers_unregister(ctx);
10088 case IORING_REGISTER_FILES:
10089 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10091 case IORING_UNREGISTER_FILES:
10093 if (arg || nr_args)
10095 ret = io_sqe_files_unregister(ctx);
10097 case IORING_REGISTER_FILES_UPDATE:
10098 ret = io_register_files_update(ctx, arg, nr_args);
10100 case IORING_REGISTER_EVENTFD:
10101 case IORING_REGISTER_EVENTFD_ASYNC:
10105 ret = io_eventfd_register(ctx, arg);
10108 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10109 ctx->eventfd_async = 1;
10111 ctx->eventfd_async = 0;
10113 case IORING_UNREGISTER_EVENTFD:
10115 if (arg || nr_args)
10117 ret = io_eventfd_unregister(ctx);
10119 case IORING_REGISTER_PROBE:
10121 if (!arg || nr_args > 256)
10123 ret = io_probe(ctx, arg, nr_args);
10125 case IORING_REGISTER_PERSONALITY:
10127 if (arg || nr_args)
10129 ret = io_register_personality(ctx);
10131 case IORING_UNREGISTER_PERSONALITY:
10135 ret = io_unregister_personality(ctx, nr_args);
10137 case IORING_REGISTER_ENABLE_RINGS:
10139 if (arg || nr_args)
10141 ret = io_register_enable_rings(ctx);
10143 case IORING_REGISTER_RESTRICTIONS:
10144 ret = io_register_restrictions(ctx, arg, nr_args);
10146 case IORING_REGISTER_FILES2:
10147 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10149 case IORING_REGISTER_FILES_UPDATE2:
10150 ret = io_register_rsrc_update(ctx, arg, nr_args,
10153 case IORING_REGISTER_BUFFERS2:
10154 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10156 case IORING_REGISTER_BUFFERS_UPDATE:
10157 ret = io_register_rsrc_update(ctx, arg, nr_args,
10158 IORING_RSRC_BUFFER);
10160 case IORING_REGISTER_IOWQ_AFF:
10162 if (!arg || !nr_args)
10164 ret = io_register_iowq_aff(ctx, arg, nr_args);
10166 case IORING_UNREGISTER_IOWQ_AFF:
10168 if (arg || nr_args)
10170 ret = io_unregister_iowq_aff(ctx);
10177 if (io_register_op_must_quiesce(opcode)) {
10178 /* bring the ctx back to life */
10179 percpu_ref_reinit(&ctx->refs);
10180 reinit_completion(&ctx->ref_comp);
10185 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10186 void __user *, arg, unsigned int, nr_args)
10188 struct io_ring_ctx *ctx;
10197 if (f.file->f_op != &io_uring_fops)
10200 ctx = f.file->private_data;
10202 io_run_task_work();
10204 mutex_lock(&ctx->uring_lock);
10205 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10206 mutex_unlock(&ctx->uring_lock);
10207 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10208 ctx->cq_ev_fd != NULL, ret);
10214 static int __init io_uring_init(void)
10216 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10217 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10218 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10221 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10222 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10223 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10224 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10225 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10226 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10227 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10228 BUILD_BUG_SQE_ELEM(8, __u64, off);
10229 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10230 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10231 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10232 BUILD_BUG_SQE_ELEM(24, __u32, len);
10233 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10234 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10235 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10236 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10237 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10238 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10239 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10240 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10241 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10242 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10243 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10244 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10245 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10246 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10247 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10248 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10249 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10250 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10251 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10252 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10254 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10255 sizeof(struct io_uring_rsrc_update));
10256 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10257 sizeof(struct io_uring_rsrc_update2));
10258 /* should fit into one byte */
10259 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10261 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10262 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10264 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10268 __initcall(io_uring_init);