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 if (req->flags & REQ_F_LINK_TIMEOUT) {
1283 struct io_ring_ctx *ctx = req->ctx;
1285 spin_lock_irq(&ctx->completion_lock);
1286 io_for_each_link(cur, req)
1287 io_prep_async_work(cur);
1288 spin_unlock_irq(&ctx->completion_lock);
1290 io_for_each_link(cur, req)
1291 io_prep_async_work(cur);
1295 static void io_queue_async_work(struct io_kiocb *req)
1297 struct io_ring_ctx *ctx = req->ctx;
1298 struct io_kiocb *link = io_prep_linked_timeout(req);
1299 struct io_uring_task *tctx = req->task->io_uring;
1302 BUG_ON(!tctx->io_wq);
1304 /* init ->work of the whole link before punting */
1305 io_prep_async_link(req);
1308 * Not expected to happen, but if we do have a bug where this _can_
1309 * happen, catch it here and ensure the request is marked as
1310 * canceled. That will make io-wq go through the usual work cancel
1311 * procedure rather than attempt to run this request (or create a new
1314 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1315 req->work.flags |= IO_WQ_WORK_CANCEL;
1317 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1318 &req->work, req->flags);
1319 io_wq_enqueue(tctx->io_wq, &req->work);
1321 io_queue_linked_timeout(link);
1324 static void io_kill_timeout(struct io_kiocb *req, int status)
1325 __must_hold(&req->ctx->completion_lock)
1327 struct io_timeout_data *io = req->async_data;
1329 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1330 atomic_set(&req->ctx->cq_timeouts,
1331 atomic_read(&req->ctx->cq_timeouts) + 1);
1332 list_del_init(&req->timeout.list);
1333 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1334 io_put_req_deferred(req, 1);
1338 static void io_queue_deferred(struct io_ring_ctx *ctx)
1340 while (!list_empty(&ctx->defer_list)) {
1341 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1342 struct io_defer_entry, list);
1344 if (req_need_defer(de->req, de->seq))
1346 list_del_init(&de->list);
1347 io_req_task_queue(de->req);
1352 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1354 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1356 while (!list_empty(&ctx->timeout_list)) {
1357 u32 events_needed, events_got;
1358 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1359 struct io_kiocb, timeout.list);
1361 if (io_is_timeout_noseq(req))
1365 * Since seq can easily wrap around over time, subtract
1366 * the last seq at which timeouts were flushed before comparing.
1367 * Assuming not more than 2^31-1 events have happened since,
1368 * these subtractions won't have wrapped, so we can check if
1369 * target is in [last_seq, current_seq] by comparing the two.
1371 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1372 events_got = seq - ctx->cq_last_tm_flush;
1373 if (events_got < events_needed)
1376 list_del_init(&req->timeout.list);
1377 io_kill_timeout(req, 0);
1379 ctx->cq_last_tm_flush = seq;
1382 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1384 if (ctx->off_timeout_used)
1385 io_flush_timeouts(ctx);
1386 if (ctx->drain_active)
1387 io_queue_deferred(ctx);
1390 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1392 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1393 __io_commit_cqring_flush(ctx);
1394 /* order cqe stores with ring update */
1395 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1398 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1400 struct io_rings *r = ctx->rings;
1402 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1405 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1407 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1410 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1412 struct io_rings *rings = ctx->rings;
1413 unsigned tail, mask = ctx->cq_entries - 1;
1416 * writes to the cq entry need to come after reading head; the
1417 * control dependency is enough as we're using WRITE_ONCE to
1420 if (__io_cqring_events(ctx) == ctx->cq_entries)
1423 tail = ctx->cached_cq_tail++;
1424 return &rings->cqes[tail & mask];
1427 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1429 if (likely(!ctx->cq_ev_fd))
1431 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1433 return !ctx->eventfd_async || io_wq_current_is_worker();
1436 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1438 /* see waitqueue_active() comment */
1441 if (waitqueue_active(&ctx->cq_wait))
1442 wake_up(&ctx->cq_wait);
1443 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1444 wake_up(&ctx->sq_data->wait);
1445 if (io_should_trigger_evfd(ctx))
1446 eventfd_signal(ctx->cq_ev_fd, 1);
1447 if (waitqueue_active(&ctx->poll_wait)) {
1448 wake_up_interruptible(&ctx->poll_wait);
1449 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1453 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1455 /* see waitqueue_active() comment */
1458 if (ctx->flags & IORING_SETUP_SQPOLL) {
1459 if (waitqueue_active(&ctx->cq_wait))
1460 wake_up(&ctx->cq_wait);
1462 if (io_should_trigger_evfd(ctx))
1463 eventfd_signal(ctx->cq_ev_fd, 1);
1464 if (waitqueue_active(&ctx->poll_wait)) {
1465 wake_up_interruptible(&ctx->poll_wait);
1466 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1470 /* Returns true if there are no backlogged entries after the flush */
1471 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1473 unsigned long flags;
1474 bool all_flushed, posted;
1476 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1480 spin_lock_irqsave(&ctx->completion_lock, flags);
1481 while (!list_empty(&ctx->cq_overflow_list)) {
1482 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1483 struct io_overflow_cqe *ocqe;
1487 ocqe = list_first_entry(&ctx->cq_overflow_list,
1488 struct io_overflow_cqe, list);
1490 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1492 io_account_cq_overflow(ctx);
1495 list_del(&ocqe->list);
1499 all_flushed = list_empty(&ctx->cq_overflow_list);
1501 clear_bit(0, &ctx->check_cq_overflow);
1502 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1506 io_commit_cqring(ctx);
1507 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1509 io_cqring_ev_posted(ctx);
1513 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1517 if (test_bit(0, &ctx->check_cq_overflow)) {
1518 /* iopoll syncs against uring_lock, not completion_lock */
1519 if (ctx->flags & IORING_SETUP_IOPOLL)
1520 mutex_lock(&ctx->uring_lock);
1521 ret = __io_cqring_overflow_flush(ctx, force);
1522 if (ctx->flags & IORING_SETUP_IOPOLL)
1523 mutex_unlock(&ctx->uring_lock);
1530 * Shamelessly stolen from the mm implementation of page reference checking,
1531 * see commit f958d7b528b1 for details.
1533 #define req_ref_zero_or_close_to_overflow(req) \
1534 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1536 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1538 return atomic_inc_not_zero(&req->refs);
1541 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1543 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1544 return atomic_sub_and_test(refs, &req->refs);
1547 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1549 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1550 return atomic_dec_and_test(&req->refs);
1553 static inline void req_ref_put(struct io_kiocb *req)
1555 WARN_ON_ONCE(req_ref_put_and_test(req));
1558 static inline void req_ref_get(struct io_kiocb *req)
1560 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1561 atomic_inc(&req->refs);
1564 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1565 long res, unsigned int cflags)
1567 struct io_overflow_cqe *ocqe;
1569 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1572 * If we're in ring overflow flush mode, or in task cancel mode,
1573 * or cannot allocate an overflow entry, then we need to drop it
1576 io_account_cq_overflow(ctx);
1579 if (list_empty(&ctx->cq_overflow_list)) {
1580 set_bit(0, &ctx->check_cq_overflow);
1581 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1583 ocqe->cqe.user_data = user_data;
1584 ocqe->cqe.res = res;
1585 ocqe->cqe.flags = cflags;
1586 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1590 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1591 long res, unsigned int cflags)
1593 struct io_uring_cqe *cqe;
1595 trace_io_uring_complete(ctx, user_data, res, cflags);
1598 * If we can't get a cq entry, userspace overflowed the
1599 * submission (by quite a lot). Increment the overflow count in
1602 cqe = io_get_cqe(ctx);
1604 WRITE_ONCE(cqe->user_data, user_data);
1605 WRITE_ONCE(cqe->res, res);
1606 WRITE_ONCE(cqe->flags, cflags);
1609 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1612 /* not as hot to bloat with inlining */
1613 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1614 long res, unsigned int cflags)
1616 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1619 static void io_req_complete_post(struct io_kiocb *req, long res,
1620 unsigned int cflags)
1622 struct io_ring_ctx *ctx = req->ctx;
1623 unsigned long flags;
1625 spin_lock_irqsave(&ctx->completion_lock, flags);
1626 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1628 * If we're the last reference to this request, add to our locked
1631 if (req_ref_put_and_test(req)) {
1632 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1633 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1634 io_disarm_next(req);
1636 io_req_task_queue(req->link);
1640 io_dismantle_req(req);
1641 io_put_task(req->task, 1);
1642 list_add(&req->compl.list, &ctx->locked_free_list);
1643 ctx->locked_free_nr++;
1645 if (!percpu_ref_tryget(&ctx->refs))
1648 io_commit_cqring(ctx);
1649 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1652 io_cqring_ev_posted(ctx);
1653 percpu_ref_put(&ctx->refs);
1657 static inline bool io_req_needs_clean(struct io_kiocb *req)
1659 return req->flags & IO_REQ_CLEAN_FLAGS;
1662 static void io_req_complete_state(struct io_kiocb *req, long res,
1663 unsigned int cflags)
1665 if (io_req_needs_clean(req))
1668 req->compl.cflags = cflags;
1669 req->flags |= REQ_F_COMPLETE_INLINE;
1672 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1673 long res, unsigned cflags)
1675 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1676 io_req_complete_state(req, res, cflags);
1678 io_req_complete_post(req, res, cflags);
1681 static inline void io_req_complete(struct io_kiocb *req, long res)
1683 __io_req_complete(req, 0, res, 0);
1686 static void io_req_complete_failed(struct io_kiocb *req, long res)
1690 io_req_complete_post(req, res, 0);
1693 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1694 struct io_comp_state *cs)
1696 spin_lock_irq(&ctx->completion_lock);
1697 list_splice_init(&ctx->locked_free_list, &cs->free_list);
1698 ctx->locked_free_nr = 0;
1699 spin_unlock_irq(&ctx->completion_lock);
1702 /* Returns true IFF there are requests in the cache */
1703 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1705 struct io_submit_state *state = &ctx->submit_state;
1706 struct io_comp_state *cs = &state->comp;
1710 * If we have more than a batch's worth of requests in our IRQ side
1711 * locked cache, grab the lock and move them over to our submission
1714 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1715 io_flush_cached_locked_reqs(ctx, cs);
1717 nr = state->free_reqs;
1718 while (!list_empty(&cs->free_list)) {
1719 struct io_kiocb *req = list_first_entry(&cs->free_list,
1720 struct io_kiocb, compl.list);
1722 list_del(&req->compl.list);
1723 state->reqs[nr++] = req;
1724 if (nr == ARRAY_SIZE(state->reqs))
1728 state->free_reqs = nr;
1732 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1734 struct io_submit_state *state = &ctx->submit_state;
1736 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1738 if (!state->free_reqs) {
1739 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1742 if (io_flush_cached_reqs(ctx))
1745 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1749 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1750 * retry single alloc to be on the safe side.
1752 if (unlikely(ret <= 0)) {
1753 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1754 if (!state->reqs[0])
1760 * Don't initialise the fields below on every allocation, but
1761 * do that in advance and keep valid on free.
1763 for (i = 0; i < ret; i++) {
1764 struct io_kiocb *req = state->reqs[i];
1768 req->async_data = NULL;
1769 /* not necessary, but safer to zero */
1772 state->free_reqs = ret;
1776 return state->reqs[state->free_reqs];
1779 static inline void io_put_file(struct file *file)
1785 static void io_dismantle_req(struct io_kiocb *req)
1787 unsigned int flags = req->flags;
1789 if (io_req_needs_clean(req))
1791 if (!(flags & REQ_F_FIXED_FILE))
1792 io_put_file(req->file);
1793 if (req->fixed_rsrc_refs)
1794 percpu_ref_put(req->fixed_rsrc_refs);
1795 if (req->async_data) {
1796 kfree(req->async_data);
1797 req->async_data = NULL;
1801 /* must to be called somewhat shortly after putting a request */
1802 static inline void io_put_task(struct task_struct *task, int nr)
1804 struct io_uring_task *tctx = task->io_uring;
1806 percpu_counter_sub(&tctx->inflight, nr);
1807 if (unlikely(atomic_read(&tctx->in_idle)))
1808 wake_up(&tctx->wait);
1809 put_task_struct_many(task, nr);
1812 static void __io_free_req(struct io_kiocb *req)
1814 struct io_ring_ctx *ctx = req->ctx;
1816 io_dismantle_req(req);
1817 io_put_task(req->task, 1);
1819 kmem_cache_free(req_cachep, req);
1820 percpu_ref_put(&ctx->refs);
1823 static inline void io_remove_next_linked(struct io_kiocb *req)
1825 struct io_kiocb *nxt = req->link;
1827 req->link = nxt->link;
1831 static bool io_kill_linked_timeout(struct io_kiocb *req)
1832 __must_hold(&req->ctx->completion_lock)
1834 struct io_kiocb *link = req->link;
1837 * Can happen if a linked timeout fired and link had been like
1838 * req -> link t-out -> link t-out [-> ...]
1840 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1841 struct io_timeout_data *io = link->async_data;
1843 io_remove_next_linked(req);
1844 link->timeout.head = NULL;
1845 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1846 io_cqring_fill_event(link->ctx, link->user_data,
1848 io_put_req_deferred(link, 1);
1855 static void io_fail_links(struct io_kiocb *req)
1856 __must_hold(&req->ctx->completion_lock)
1858 struct io_kiocb *nxt, *link = req->link;
1865 trace_io_uring_fail_link(req, link);
1866 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1867 io_put_req_deferred(link, 2);
1872 static bool io_disarm_next(struct io_kiocb *req)
1873 __must_hold(&req->ctx->completion_lock)
1875 bool posted = false;
1877 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1878 posted = io_kill_linked_timeout(req);
1879 if (unlikely((req->flags & REQ_F_FAIL) &&
1880 !(req->flags & REQ_F_HARDLINK))) {
1881 posted |= (req->link != NULL);
1887 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1889 struct io_kiocb *nxt;
1892 * If LINK is set, we have dependent requests in this chain. If we
1893 * didn't fail this request, queue the first one up, moving any other
1894 * dependencies to the next request. In case of failure, fail the rest
1897 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1898 struct io_ring_ctx *ctx = req->ctx;
1899 unsigned long flags;
1902 spin_lock_irqsave(&ctx->completion_lock, flags);
1903 posted = io_disarm_next(req);
1905 io_commit_cqring(req->ctx);
1906 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1908 io_cqring_ev_posted(ctx);
1915 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1917 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1919 return __io_req_find_next(req);
1922 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1926 if (ctx->submit_state.comp.nr) {
1927 mutex_lock(&ctx->uring_lock);
1928 io_submit_flush_completions(ctx);
1929 mutex_unlock(&ctx->uring_lock);
1931 percpu_ref_put(&ctx->refs);
1934 static void tctx_task_work(struct callback_head *cb)
1936 struct io_ring_ctx *ctx = NULL;
1937 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1941 struct io_wq_work_node *node;
1943 spin_lock_irq(&tctx->task_lock);
1944 node = tctx->task_list.first;
1945 INIT_WQ_LIST(&tctx->task_list);
1946 spin_unlock_irq(&tctx->task_lock);
1949 struct io_wq_work_node *next = node->next;
1950 struct io_kiocb *req = container_of(node, struct io_kiocb,
1953 if (req->ctx != ctx) {
1954 ctx_flush_and_put(ctx);
1956 percpu_ref_get(&ctx->refs);
1958 req->io_task_work.func(req);
1961 if (wq_list_empty(&tctx->task_list)) {
1962 spin_lock_irq(&tctx->task_lock);
1963 clear_bit(0, &tctx->task_state);
1964 if (wq_list_empty(&tctx->task_list)) {
1965 spin_unlock_irq(&tctx->task_lock);
1968 spin_unlock_irq(&tctx->task_lock);
1969 /* another tctx_task_work() is enqueued, yield */
1970 if (test_and_set_bit(0, &tctx->task_state))
1976 ctx_flush_and_put(ctx);
1979 static void io_req_task_work_add(struct io_kiocb *req)
1981 struct task_struct *tsk = req->task;
1982 struct io_uring_task *tctx = tsk->io_uring;
1983 enum task_work_notify_mode notify;
1984 struct io_wq_work_node *node;
1985 unsigned long flags;
1987 WARN_ON_ONCE(!tctx);
1989 spin_lock_irqsave(&tctx->task_lock, flags);
1990 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1991 spin_unlock_irqrestore(&tctx->task_lock, flags);
1993 /* task_work already pending, we're done */
1994 if (test_bit(0, &tctx->task_state) ||
1995 test_and_set_bit(0, &tctx->task_state))
1997 if (unlikely(tsk->flags & PF_EXITING))
2001 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2002 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2003 * processing task_work. There's no reliable way to tell if TWA_RESUME
2006 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2007 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2008 wake_up_process(tsk);
2012 clear_bit(0, &tctx->task_state);
2013 spin_lock_irqsave(&tctx->task_lock, flags);
2014 node = tctx->task_list.first;
2015 INIT_WQ_LIST(&tctx->task_list);
2016 spin_unlock_irqrestore(&tctx->task_lock, flags);
2019 req = container_of(node, struct io_kiocb, io_task_work.node);
2021 if (llist_add(&req->io_task_work.fallback_node,
2022 &req->ctx->fallback_llist))
2023 schedule_delayed_work(&req->ctx->fallback_work, 1);
2027 static void io_req_task_cancel(struct io_kiocb *req)
2029 struct io_ring_ctx *ctx = req->ctx;
2031 /* ctx is guaranteed to stay alive while we hold uring_lock */
2032 mutex_lock(&ctx->uring_lock);
2033 io_req_complete_failed(req, req->result);
2034 mutex_unlock(&ctx->uring_lock);
2037 static void io_req_task_submit(struct io_kiocb *req)
2039 struct io_ring_ctx *ctx = req->ctx;
2041 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2042 mutex_lock(&ctx->uring_lock);
2043 if (!(req->task->flags & PF_EXITING) && !req->task->in_execve)
2044 __io_queue_sqe(req);
2046 io_req_complete_failed(req, -EFAULT);
2047 mutex_unlock(&ctx->uring_lock);
2050 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2053 req->io_task_work.func = io_req_task_cancel;
2054 io_req_task_work_add(req);
2057 static void io_req_task_queue(struct io_kiocb *req)
2059 req->io_task_work.func = io_req_task_submit;
2060 io_req_task_work_add(req);
2063 static void io_req_task_queue_reissue(struct io_kiocb *req)
2065 req->io_task_work.func = io_queue_async_work;
2066 io_req_task_work_add(req);
2069 static inline void io_queue_next(struct io_kiocb *req)
2071 struct io_kiocb *nxt = io_req_find_next(req);
2074 io_req_task_queue(nxt);
2077 static void io_free_req(struct io_kiocb *req)
2084 struct task_struct *task;
2089 static inline void io_init_req_batch(struct req_batch *rb)
2096 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2097 struct req_batch *rb)
2100 io_put_task(rb->task, rb->task_refs);
2102 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2105 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2106 struct io_submit_state *state)
2109 io_dismantle_req(req);
2111 if (req->task != rb->task) {
2113 io_put_task(rb->task, rb->task_refs);
2114 rb->task = req->task;
2120 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2121 state->reqs[state->free_reqs++] = req;
2123 list_add(&req->compl.list, &state->comp.free_list);
2126 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2128 struct io_comp_state *cs = &ctx->submit_state.comp;
2130 struct req_batch rb;
2132 spin_lock_irq(&ctx->completion_lock);
2133 for (i = 0; i < nr; i++) {
2134 struct io_kiocb *req = cs->reqs[i];
2136 __io_cqring_fill_event(ctx, req->user_data, req->result,
2139 io_commit_cqring(ctx);
2140 spin_unlock_irq(&ctx->completion_lock);
2141 io_cqring_ev_posted(ctx);
2143 io_init_req_batch(&rb);
2144 for (i = 0; i < nr; i++) {
2145 struct io_kiocb *req = cs->reqs[i];
2147 /* submission and completion refs */
2148 if (req_ref_sub_and_test(req, 2))
2149 io_req_free_batch(&rb, req, &ctx->submit_state);
2152 io_req_free_batch_finish(ctx, &rb);
2157 * Drop reference to request, return next in chain (if there is one) if this
2158 * was the last reference to this request.
2160 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2162 struct io_kiocb *nxt = NULL;
2164 if (req_ref_put_and_test(req)) {
2165 nxt = io_req_find_next(req);
2171 static inline void io_put_req(struct io_kiocb *req)
2173 if (req_ref_put_and_test(req))
2177 static void io_free_req_deferred(struct io_kiocb *req)
2179 req->io_task_work.func = io_free_req;
2180 io_req_task_work_add(req);
2183 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2185 if (req_ref_sub_and_test(req, refs))
2186 io_free_req_deferred(req);
2189 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2191 /* See comment at the top of this file */
2193 return __io_cqring_events(ctx);
2196 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2198 struct io_rings *rings = ctx->rings;
2200 /* make sure SQ entry isn't read before tail */
2201 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2204 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2206 unsigned int cflags;
2208 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2209 cflags |= IORING_CQE_F_BUFFER;
2210 req->flags &= ~REQ_F_BUFFER_SELECTED;
2215 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2217 struct io_buffer *kbuf;
2219 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2220 return io_put_kbuf(req, kbuf);
2223 static inline bool io_run_task_work(void)
2225 if (current->task_works) {
2226 __set_current_state(TASK_RUNNING);
2235 * Find and free completed poll iocbs
2237 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2238 struct list_head *done, bool resubmit)
2240 struct req_batch rb;
2241 struct io_kiocb *req;
2243 /* order with ->result store in io_complete_rw_iopoll() */
2246 io_init_req_batch(&rb);
2247 while (!list_empty(done)) {
2250 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2251 list_del(&req->inflight_entry);
2253 if (READ_ONCE(req->result) == -EAGAIN && resubmit &&
2254 !(req->flags & REQ_F_DONT_REISSUE)) {
2255 req->iopoll_completed = 0;
2257 io_req_task_queue_reissue(req);
2261 if (req->flags & REQ_F_BUFFER_SELECTED)
2262 cflags = io_put_rw_kbuf(req);
2264 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2267 if (req_ref_put_and_test(req))
2268 io_req_free_batch(&rb, req, &ctx->submit_state);
2271 io_commit_cqring(ctx);
2272 io_cqring_ev_posted_iopoll(ctx);
2273 io_req_free_batch_finish(ctx, &rb);
2276 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2277 long min, bool resubmit)
2279 struct io_kiocb *req, *tmp;
2285 * Only spin for completions if we don't have multiple devices hanging
2286 * off our complete list, and we're under the requested amount.
2288 spin = !ctx->poll_multi_queue && *nr_events < min;
2291 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2292 struct kiocb *kiocb = &req->rw.kiocb;
2295 * Move completed and retryable entries to our local lists.
2296 * If we find a request that requires polling, break out
2297 * and complete those lists first, if we have entries there.
2299 if (READ_ONCE(req->iopoll_completed)) {
2300 list_move_tail(&req->inflight_entry, &done);
2303 if (!list_empty(&done))
2306 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2310 /* iopoll may have completed current req */
2311 if (READ_ONCE(req->iopoll_completed))
2312 list_move_tail(&req->inflight_entry, &done);
2319 if (!list_empty(&done))
2320 io_iopoll_complete(ctx, nr_events, &done, resubmit);
2326 * We can't just wait for polled events to come to us, we have to actively
2327 * find and complete them.
2329 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2331 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2334 mutex_lock(&ctx->uring_lock);
2335 while (!list_empty(&ctx->iopoll_list)) {
2336 unsigned int nr_events = 0;
2338 io_do_iopoll(ctx, &nr_events, 0, false);
2340 /* let it sleep and repeat later if can't complete a request */
2344 * Ensure we allow local-to-the-cpu processing to take place,
2345 * in this case we need to ensure that we reap all events.
2346 * Also let task_work, etc. to progress by releasing the mutex
2348 if (need_resched()) {
2349 mutex_unlock(&ctx->uring_lock);
2351 mutex_lock(&ctx->uring_lock);
2354 mutex_unlock(&ctx->uring_lock);
2357 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2359 unsigned int nr_events = 0;
2363 * We disallow the app entering submit/complete with polling, but we
2364 * still need to lock the ring to prevent racing with polled issue
2365 * that got punted to a workqueue.
2367 mutex_lock(&ctx->uring_lock);
2369 * Don't enter poll loop if we already have events pending.
2370 * If we do, we can potentially be spinning for commands that
2371 * already triggered a CQE (eg in error).
2373 if (test_bit(0, &ctx->check_cq_overflow))
2374 __io_cqring_overflow_flush(ctx, false);
2375 if (io_cqring_events(ctx))
2379 * If a submit got punted to a workqueue, we can have the
2380 * application entering polling for a command before it gets
2381 * issued. That app will hold the uring_lock for the duration
2382 * of the poll right here, so we need to take a breather every
2383 * now and then to ensure that the issue has a chance to add
2384 * the poll to the issued list. Otherwise we can spin here
2385 * forever, while the workqueue is stuck trying to acquire the
2388 if (list_empty(&ctx->iopoll_list)) {
2389 u32 tail = ctx->cached_cq_tail;
2391 mutex_unlock(&ctx->uring_lock);
2393 mutex_lock(&ctx->uring_lock);
2395 /* some requests don't go through iopoll_list */
2396 if (tail != ctx->cached_cq_tail ||
2397 list_empty(&ctx->iopoll_list))
2400 ret = io_do_iopoll(ctx, &nr_events, min, true);
2401 } while (!ret && nr_events < min && !need_resched());
2403 mutex_unlock(&ctx->uring_lock);
2407 static void kiocb_end_write(struct io_kiocb *req)
2410 * Tell lockdep we inherited freeze protection from submission
2413 if (req->flags & REQ_F_ISREG) {
2414 struct super_block *sb = file_inode(req->file)->i_sb;
2416 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2422 static bool io_resubmit_prep(struct io_kiocb *req)
2424 struct io_async_rw *rw = req->async_data;
2427 return !io_req_prep_async(req);
2428 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2429 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2433 static bool io_rw_should_reissue(struct io_kiocb *req)
2435 umode_t mode = file_inode(req->file)->i_mode;
2436 struct io_ring_ctx *ctx = req->ctx;
2438 if (!S_ISBLK(mode) && !S_ISREG(mode))
2440 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2441 !(ctx->flags & IORING_SETUP_IOPOLL)))
2444 * If ref is dying, we might be running poll reap from the exit work.
2445 * Don't attempt to reissue from that path, just let it fail with
2448 if (percpu_ref_is_dying(&ctx->refs))
2453 static bool io_resubmit_prep(struct io_kiocb *req)
2457 static bool io_rw_should_reissue(struct io_kiocb *req)
2463 static void io_fallback_req_func(struct work_struct *work)
2465 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
2466 fallback_work.work);
2467 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
2468 struct io_kiocb *req, *tmp;
2470 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
2471 req->io_task_work.func(req);
2474 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2475 unsigned int issue_flags)
2479 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2480 kiocb_end_write(req);
2481 if (res != req->result) {
2482 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2483 io_rw_should_reissue(req)) {
2484 req->flags |= REQ_F_REISSUE;
2489 if (req->flags & REQ_F_BUFFER_SELECTED)
2490 cflags = io_put_rw_kbuf(req);
2491 __io_req_complete(req, issue_flags, res, cflags);
2494 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2496 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2498 __io_complete_rw(req, res, res2, 0);
2501 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2503 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2505 if (kiocb->ki_flags & IOCB_WRITE)
2506 kiocb_end_write(req);
2507 if (unlikely(res != req->result)) {
2508 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2509 io_resubmit_prep(req))) {
2511 req->flags |= REQ_F_DONT_REISSUE;
2515 WRITE_ONCE(req->result, res);
2516 /* order with io_iopoll_complete() checking ->result */
2518 WRITE_ONCE(req->iopoll_completed, 1);
2522 * After the iocb has been issued, it's safe to be found on the poll list.
2523 * Adding the kiocb to the list AFTER submission ensures that we don't
2524 * find it from a io_do_iopoll() thread before the issuer is done
2525 * accessing the kiocb cookie.
2527 static void io_iopoll_req_issued(struct io_kiocb *req)
2529 struct io_ring_ctx *ctx = req->ctx;
2530 const bool in_async = io_wq_current_is_worker();
2532 /* workqueue context doesn't hold uring_lock, grab it now */
2533 if (unlikely(in_async))
2534 mutex_lock(&ctx->uring_lock);
2537 * Track whether we have multiple files in our lists. This will impact
2538 * how we do polling eventually, not spinning if we're on potentially
2539 * different devices.
2541 if (list_empty(&ctx->iopoll_list)) {
2542 ctx->poll_multi_queue = false;
2543 } else if (!ctx->poll_multi_queue) {
2544 struct io_kiocb *list_req;
2545 unsigned int queue_num0, queue_num1;
2547 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2550 if (list_req->file != req->file) {
2551 ctx->poll_multi_queue = true;
2553 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2554 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2555 if (queue_num0 != queue_num1)
2556 ctx->poll_multi_queue = true;
2561 * For fast devices, IO may have already completed. If it has, add
2562 * it to the front so we find it first.
2564 if (READ_ONCE(req->iopoll_completed))
2565 list_add(&req->inflight_entry, &ctx->iopoll_list);
2567 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2569 if (unlikely(in_async)) {
2571 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2572 * in sq thread task context or in io worker task context. If
2573 * current task context is sq thread, we don't need to check
2574 * whether should wake up sq thread.
2576 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2577 wq_has_sleeper(&ctx->sq_data->wait))
2578 wake_up(&ctx->sq_data->wait);
2580 mutex_unlock(&ctx->uring_lock);
2584 static inline void io_state_file_put(struct io_submit_state *state)
2586 if (state->file_refs) {
2587 fput_many(state->file, state->file_refs);
2588 state->file_refs = 0;
2593 * Get as many references to a file as we have IOs left in this submission,
2594 * assuming most submissions are for one file, or at least that each file
2595 * has more than one submission.
2597 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2602 if (state->file_refs) {
2603 if (state->fd == fd) {
2607 io_state_file_put(state);
2609 state->file = fget_many(fd, state->ios_left);
2610 if (unlikely(!state->file))
2614 state->file_refs = state->ios_left - 1;
2618 static bool io_bdev_nowait(struct block_device *bdev)
2620 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2624 * If we tracked the file through the SCM inflight mechanism, we could support
2625 * any file. For now, just ensure that anything potentially problematic is done
2628 static bool __io_file_supports_async(struct file *file, int rw)
2630 umode_t mode = file_inode(file)->i_mode;
2632 if (S_ISBLK(mode)) {
2633 if (IS_ENABLED(CONFIG_BLOCK) &&
2634 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2640 if (S_ISREG(mode)) {
2641 if (IS_ENABLED(CONFIG_BLOCK) &&
2642 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2643 file->f_op != &io_uring_fops)
2648 /* any ->read/write should understand O_NONBLOCK */
2649 if (file->f_flags & O_NONBLOCK)
2652 if (!(file->f_mode & FMODE_NOWAIT))
2656 return file->f_op->read_iter != NULL;
2658 return file->f_op->write_iter != NULL;
2661 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2663 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2665 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2668 return __io_file_supports_async(req->file, rw);
2671 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2673 struct io_ring_ctx *ctx = req->ctx;
2674 struct kiocb *kiocb = &req->rw.kiocb;
2675 struct file *file = req->file;
2679 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2680 req->flags |= REQ_F_ISREG;
2682 kiocb->ki_pos = READ_ONCE(sqe->off);
2683 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2684 req->flags |= REQ_F_CUR_POS;
2685 kiocb->ki_pos = file->f_pos;
2687 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2688 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2689 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2693 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2694 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2695 req->flags |= REQ_F_NOWAIT;
2697 ioprio = READ_ONCE(sqe->ioprio);
2699 ret = ioprio_check_cap(ioprio);
2703 kiocb->ki_ioprio = ioprio;
2705 kiocb->ki_ioprio = get_current_ioprio();
2707 if (ctx->flags & IORING_SETUP_IOPOLL) {
2708 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2709 !kiocb->ki_filp->f_op->iopoll)
2712 kiocb->ki_flags |= IOCB_HIPRI;
2713 kiocb->ki_complete = io_complete_rw_iopoll;
2714 req->iopoll_completed = 0;
2716 if (kiocb->ki_flags & IOCB_HIPRI)
2718 kiocb->ki_complete = io_complete_rw;
2721 if (req->opcode == IORING_OP_READ_FIXED ||
2722 req->opcode == IORING_OP_WRITE_FIXED) {
2724 io_req_set_rsrc_node(req);
2727 req->rw.addr = READ_ONCE(sqe->addr);
2728 req->rw.len = READ_ONCE(sqe->len);
2729 req->buf_index = READ_ONCE(sqe->buf_index);
2733 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2739 case -ERESTARTNOINTR:
2740 case -ERESTARTNOHAND:
2741 case -ERESTART_RESTARTBLOCK:
2743 * We can't just restart the syscall, since previously
2744 * submitted sqes may already be in progress. Just fail this
2750 kiocb->ki_complete(kiocb, ret, 0);
2754 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2755 unsigned int issue_flags)
2757 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2758 struct io_async_rw *io = req->async_data;
2759 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2761 /* add previously done IO, if any */
2762 if (io && io->bytes_done > 0) {
2764 ret = io->bytes_done;
2766 ret += io->bytes_done;
2769 if (req->flags & REQ_F_CUR_POS)
2770 req->file->f_pos = kiocb->ki_pos;
2771 if (ret >= 0 && check_reissue)
2772 __io_complete_rw(req, ret, 0, issue_flags);
2774 io_rw_done(kiocb, ret);
2776 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2777 req->flags &= ~REQ_F_REISSUE;
2778 if (io_resubmit_prep(req)) {
2780 io_req_task_queue_reissue(req);
2785 if (req->flags & REQ_F_BUFFER_SELECTED)
2786 cflags = io_put_rw_kbuf(req);
2787 __io_req_complete(req, issue_flags, ret, cflags);
2792 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2793 struct io_mapped_ubuf *imu)
2795 size_t len = req->rw.len;
2796 u64 buf_end, buf_addr = req->rw.addr;
2799 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2801 /* not inside the mapped region */
2802 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2806 * May not be a start of buffer, set size appropriately
2807 * and advance us to the beginning.
2809 offset = buf_addr - imu->ubuf;
2810 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2814 * Don't use iov_iter_advance() here, as it's really slow for
2815 * using the latter parts of a big fixed buffer - it iterates
2816 * over each segment manually. We can cheat a bit here, because
2819 * 1) it's a BVEC iter, we set it up
2820 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2821 * first and last bvec
2823 * So just find our index, and adjust the iterator afterwards.
2824 * If the offset is within the first bvec (or the whole first
2825 * bvec, just use iov_iter_advance(). This makes it easier
2826 * since we can just skip the first segment, which may not
2827 * be PAGE_SIZE aligned.
2829 const struct bio_vec *bvec = imu->bvec;
2831 if (offset <= bvec->bv_len) {
2832 iov_iter_advance(iter, offset);
2834 unsigned long seg_skip;
2836 /* skip first vec */
2837 offset -= bvec->bv_len;
2838 seg_skip = 1 + (offset >> PAGE_SHIFT);
2840 iter->bvec = bvec + seg_skip;
2841 iter->nr_segs -= seg_skip;
2842 iter->count -= bvec->bv_len + offset;
2843 iter->iov_offset = offset & ~PAGE_MASK;
2850 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2852 struct io_ring_ctx *ctx = req->ctx;
2853 struct io_mapped_ubuf *imu = req->imu;
2854 u16 index, buf_index = req->buf_index;
2857 if (unlikely(buf_index >= ctx->nr_user_bufs))
2859 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2860 imu = READ_ONCE(ctx->user_bufs[index]);
2863 return __io_import_fixed(req, rw, iter, imu);
2866 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2869 mutex_unlock(&ctx->uring_lock);
2872 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2875 * "Normal" inline submissions always hold the uring_lock, since we
2876 * grab it from the system call. Same is true for the SQPOLL offload.
2877 * The only exception is when we've detached the request and issue it
2878 * from an async worker thread, grab the lock for that case.
2881 mutex_lock(&ctx->uring_lock);
2884 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2885 int bgid, struct io_buffer *kbuf,
2888 struct io_buffer *head;
2890 if (req->flags & REQ_F_BUFFER_SELECTED)
2893 io_ring_submit_lock(req->ctx, needs_lock);
2895 lockdep_assert_held(&req->ctx->uring_lock);
2897 head = xa_load(&req->ctx->io_buffers, bgid);
2899 if (!list_empty(&head->list)) {
2900 kbuf = list_last_entry(&head->list, struct io_buffer,
2902 list_del(&kbuf->list);
2905 xa_erase(&req->ctx->io_buffers, bgid);
2907 if (*len > kbuf->len)
2910 kbuf = ERR_PTR(-ENOBUFS);
2913 io_ring_submit_unlock(req->ctx, needs_lock);
2918 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2921 struct io_buffer *kbuf;
2924 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2925 bgid = req->buf_index;
2926 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2929 req->rw.addr = (u64) (unsigned long) kbuf;
2930 req->flags |= REQ_F_BUFFER_SELECTED;
2931 return u64_to_user_ptr(kbuf->addr);
2934 #ifdef CONFIG_COMPAT
2935 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2938 struct compat_iovec __user *uiov;
2939 compat_ssize_t clen;
2943 uiov = u64_to_user_ptr(req->rw.addr);
2944 if (!access_ok(uiov, sizeof(*uiov)))
2946 if (__get_user(clen, &uiov->iov_len))
2952 buf = io_rw_buffer_select(req, &len, needs_lock);
2954 return PTR_ERR(buf);
2955 iov[0].iov_base = buf;
2956 iov[0].iov_len = (compat_size_t) len;
2961 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2964 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2968 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2971 len = iov[0].iov_len;
2974 buf = io_rw_buffer_select(req, &len, needs_lock);
2976 return PTR_ERR(buf);
2977 iov[0].iov_base = buf;
2978 iov[0].iov_len = len;
2982 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2985 if (req->flags & REQ_F_BUFFER_SELECTED) {
2986 struct io_buffer *kbuf;
2988 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2989 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2990 iov[0].iov_len = kbuf->len;
2993 if (req->rw.len != 1)
2996 #ifdef CONFIG_COMPAT
2997 if (req->ctx->compat)
2998 return io_compat_import(req, iov, needs_lock);
3001 return __io_iov_buffer_select(req, iov, needs_lock);
3004 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3005 struct iov_iter *iter, bool needs_lock)
3007 void __user *buf = u64_to_user_ptr(req->rw.addr);
3008 size_t sqe_len = req->rw.len;
3009 u8 opcode = req->opcode;
3012 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3014 return io_import_fixed(req, rw, iter);
3017 /* buffer index only valid with fixed read/write, or buffer select */
3018 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3021 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3022 if (req->flags & REQ_F_BUFFER_SELECT) {
3023 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3025 return PTR_ERR(buf);
3026 req->rw.len = sqe_len;
3029 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3034 if (req->flags & REQ_F_BUFFER_SELECT) {
3035 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3037 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3042 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3046 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3048 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3052 * For files that don't have ->read_iter() and ->write_iter(), handle them
3053 * by looping over ->read() or ->write() manually.
3055 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3057 struct kiocb *kiocb = &req->rw.kiocb;
3058 struct file *file = req->file;
3062 * Don't support polled IO through this interface, and we can't
3063 * support non-blocking either. For the latter, this just causes
3064 * the kiocb to be handled from an async context.
3066 if (kiocb->ki_flags & IOCB_HIPRI)
3068 if (kiocb->ki_flags & IOCB_NOWAIT)
3071 while (iov_iter_count(iter)) {
3075 if (!iov_iter_is_bvec(iter)) {
3076 iovec = iov_iter_iovec(iter);
3078 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3079 iovec.iov_len = req->rw.len;
3083 nr = file->f_op->read(file, iovec.iov_base,
3084 iovec.iov_len, io_kiocb_ppos(kiocb));
3086 nr = file->f_op->write(file, iovec.iov_base,
3087 iovec.iov_len, io_kiocb_ppos(kiocb));
3096 if (nr != iovec.iov_len)
3100 iov_iter_advance(iter, nr);
3106 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3107 const struct iovec *fast_iov, struct iov_iter *iter)
3109 struct io_async_rw *rw = req->async_data;
3111 memcpy(&rw->iter, iter, sizeof(*iter));
3112 rw->free_iovec = iovec;
3114 /* can only be fixed buffers, no need to do anything */
3115 if (iov_iter_is_bvec(iter))
3118 unsigned iov_off = 0;
3120 rw->iter.iov = rw->fast_iov;
3121 if (iter->iov != fast_iov) {
3122 iov_off = iter->iov - fast_iov;
3123 rw->iter.iov += iov_off;
3125 if (rw->fast_iov != fast_iov)
3126 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3127 sizeof(struct iovec) * iter->nr_segs);
3129 req->flags |= REQ_F_NEED_CLEANUP;
3133 static inline int io_alloc_async_data(struct io_kiocb *req)
3135 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3136 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3137 return req->async_data == NULL;
3140 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3141 const struct iovec *fast_iov,
3142 struct iov_iter *iter, bool force)
3144 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3146 if (!req->async_data) {
3147 if (io_alloc_async_data(req)) {
3152 io_req_map_rw(req, iovec, fast_iov, iter);
3157 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3159 struct io_async_rw *iorw = req->async_data;
3160 struct iovec *iov = iorw->fast_iov;
3163 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3164 if (unlikely(ret < 0))
3167 iorw->bytes_done = 0;
3168 iorw->free_iovec = iov;
3170 req->flags |= REQ_F_NEED_CLEANUP;
3174 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3176 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3178 return io_prep_rw(req, sqe);
3182 * This is our waitqueue callback handler, registered through lock_page_async()
3183 * when we initially tried to do the IO with the iocb armed our waitqueue.
3184 * This gets called when the page is unlocked, and we generally expect that to
3185 * happen when the page IO is completed and the page is now uptodate. This will
3186 * queue a task_work based retry of the operation, attempting to copy the data
3187 * again. If the latter fails because the page was NOT uptodate, then we will
3188 * do a thread based blocking retry of the operation. That's the unexpected
3191 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3192 int sync, void *arg)
3194 struct wait_page_queue *wpq;
3195 struct io_kiocb *req = wait->private;
3196 struct wait_page_key *key = arg;
3198 wpq = container_of(wait, struct wait_page_queue, wait);
3200 if (!wake_page_match(wpq, key))
3203 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3204 list_del_init(&wait->entry);
3206 /* submit ref gets dropped, acquire a new one */
3208 io_req_task_queue(req);
3213 * This controls whether a given IO request should be armed for async page
3214 * based retry. If we return false here, the request is handed to the async
3215 * worker threads for retry. If we're doing buffered reads on a regular file,
3216 * we prepare a private wait_page_queue entry and retry the operation. This
3217 * will either succeed because the page is now uptodate and unlocked, or it
3218 * will register a callback when the page is unlocked at IO completion. Through
3219 * that callback, io_uring uses task_work to setup a retry of the operation.
3220 * That retry will attempt the buffered read again. The retry will generally
3221 * succeed, or in rare cases where it fails, we then fall back to using the
3222 * async worker threads for a blocking retry.
3224 static bool io_rw_should_retry(struct io_kiocb *req)
3226 struct io_async_rw *rw = req->async_data;
3227 struct wait_page_queue *wait = &rw->wpq;
3228 struct kiocb *kiocb = &req->rw.kiocb;
3230 /* never retry for NOWAIT, we just complete with -EAGAIN */
3231 if (req->flags & REQ_F_NOWAIT)
3234 /* Only for buffered IO */
3235 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3239 * just use poll if we can, and don't attempt if the fs doesn't
3240 * support callback based unlocks
3242 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3245 wait->wait.func = io_async_buf_func;
3246 wait->wait.private = req;
3247 wait->wait.flags = 0;
3248 INIT_LIST_HEAD(&wait->wait.entry);
3249 kiocb->ki_flags |= IOCB_WAITQ;
3250 kiocb->ki_flags &= ~IOCB_NOWAIT;
3251 kiocb->ki_waitq = wait;
3255 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3257 if (req->file->f_op->read_iter)
3258 return call_read_iter(req->file, &req->rw.kiocb, iter);
3259 else if (req->file->f_op->read)
3260 return loop_rw_iter(READ, req, iter);
3265 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3267 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3268 struct kiocb *kiocb = &req->rw.kiocb;
3269 struct iov_iter __iter, *iter = &__iter;
3270 struct io_async_rw *rw = req->async_data;
3271 ssize_t io_size, ret, ret2;
3272 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3278 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3282 io_size = iov_iter_count(iter);
3283 req->result = io_size;
3285 /* Ensure we clear previously set non-block flag */
3286 if (!force_nonblock)
3287 kiocb->ki_flags &= ~IOCB_NOWAIT;
3289 kiocb->ki_flags |= IOCB_NOWAIT;
3291 /* If the file doesn't support async, just async punt */
3292 if (force_nonblock && !io_file_supports_async(req, READ)) {
3293 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3294 return ret ?: -EAGAIN;
3297 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3298 if (unlikely(ret)) {
3303 ret = io_iter_do_read(req, iter);
3305 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3306 req->flags &= ~REQ_F_REISSUE;
3307 /* IOPOLL retry should happen for io-wq threads */
3308 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3310 /* no retry on NONBLOCK nor RWF_NOWAIT */
3311 if (req->flags & REQ_F_NOWAIT)
3313 /* some cases will consume bytes even on error returns */
3314 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3316 } else if (ret == -EIOCBQUEUED) {
3318 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3319 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3320 /* read all, failed, already did sync or don't want to retry */
3324 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3329 rw = req->async_data;
3330 /* now use our persistent iterator, if we aren't already */
3335 rw->bytes_done += ret;
3336 /* if we can retry, do so with the callbacks armed */
3337 if (!io_rw_should_retry(req)) {
3338 kiocb->ki_flags &= ~IOCB_WAITQ;
3343 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3344 * we get -EIOCBQUEUED, then we'll get a notification when the
3345 * desired page gets unlocked. We can also get a partial read
3346 * here, and if we do, then just retry at the new offset.
3348 ret = io_iter_do_read(req, iter);
3349 if (ret == -EIOCBQUEUED)
3351 /* we got some bytes, but not all. retry. */
3352 kiocb->ki_flags &= ~IOCB_WAITQ;
3353 } while (ret > 0 && ret < io_size);
3355 kiocb_done(kiocb, ret, issue_flags);
3357 /* it's faster to check here then delegate to kfree */
3363 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3365 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3367 return io_prep_rw(req, sqe);
3370 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3372 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3373 struct kiocb *kiocb = &req->rw.kiocb;
3374 struct iov_iter __iter, *iter = &__iter;
3375 struct io_async_rw *rw = req->async_data;
3376 ssize_t ret, ret2, io_size;
3377 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3383 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3387 io_size = iov_iter_count(iter);
3388 req->result = io_size;
3390 /* Ensure we clear previously set non-block flag */
3391 if (!force_nonblock)
3392 kiocb->ki_flags &= ~IOCB_NOWAIT;
3394 kiocb->ki_flags |= IOCB_NOWAIT;
3396 /* If the file doesn't support async, just async punt */
3397 if (force_nonblock && !io_file_supports_async(req, WRITE))
3400 /* file path doesn't support NOWAIT for non-direct_IO */
3401 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3402 (req->flags & REQ_F_ISREG))
3405 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3410 * Open-code file_start_write here to grab freeze protection,
3411 * which will be released by another thread in
3412 * io_complete_rw(). Fool lockdep by telling it the lock got
3413 * released so that it doesn't complain about the held lock when
3414 * we return to userspace.
3416 if (req->flags & REQ_F_ISREG) {
3417 sb_start_write(file_inode(req->file)->i_sb);
3418 __sb_writers_release(file_inode(req->file)->i_sb,
3421 kiocb->ki_flags |= IOCB_WRITE;
3423 if (req->file->f_op->write_iter)
3424 ret2 = call_write_iter(req->file, kiocb, iter);
3425 else if (req->file->f_op->write)
3426 ret2 = loop_rw_iter(WRITE, req, iter);
3430 if (req->flags & REQ_F_REISSUE) {
3431 req->flags &= ~REQ_F_REISSUE;
3436 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3437 * retry them without IOCB_NOWAIT.
3439 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3441 /* no retry on NONBLOCK nor RWF_NOWAIT */
3442 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3444 if (!force_nonblock || ret2 != -EAGAIN) {
3445 /* IOPOLL retry should happen for io-wq threads */
3446 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3449 kiocb_done(kiocb, ret2, issue_flags);
3452 /* some cases will consume bytes even on error returns */
3453 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3454 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3455 return ret ?: -EAGAIN;
3458 /* it's reportedly faster than delegating the null check to kfree() */
3464 static int io_renameat_prep(struct io_kiocb *req,
3465 const struct io_uring_sqe *sqe)
3467 struct io_rename *ren = &req->rename;
3468 const char __user *oldf, *newf;
3470 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3472 if (sqe->ioprio || sqe->buf_index)
3474 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3477 ren->old_dfd = READ_ONCE(sqe->fd);
3478 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3479 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3480 ren->new_dfd = READ_ONCE(sqe->len);
3481 ren->flags = READ_ONCE(sqe->rename_flags);
3483 ren->oldpath = getname(oldf);
3484 if (IS_ERR(ren->oldpath))
3485 return PTR_ERR(ren->oldpath);
3487 ren->newpath = getname(newf);
3488 if (IS_ERR(ren->newpath)) {
3489 putname(ren->oldpath);
3490 return PTR_ERR(ren->newpath);
3493 req->flags |= REQ_F_NEED_CLEANUP;
3497 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3499 struct io_rename *ren = &req->rename;
3502 if (issue_flags & IO_URING_F_NONBLOCK)
3505 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3506 ren->newpath, ren->flags);
3508 req->flags &= ~REQ_F_NEED_CLEANUP;
3511 io_req_complete(req, ret);
3515 static int io_unlinkat_prep(struct io_kiocb *req,
3516 const struct io_uring_sqe *sqe)
3518 struct io_unlink *un = &req->unlink;
3519 const char __user *fname;
3521 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3523 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3525 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3528 un->dfd = READ_ONCE(sqe->fd);
3530 un->flags = READ_ONCE(sqe->unlink_flags);
3531 if (un->flags & ~AT_REMOVEDIR)
3534 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3535 un->filename = getname(fname);
3536 if (IS_ERR(un->filename))
3537 return PTR_ERR(un->filename);
3539 req->flags |= REQ_F_NEED_CLEANUP;
3543 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3545 struct io_unlink *un = &req->unlink;
3548 if (issue_flags & IO_URING_F_NONBLOCK)
3551 if (un->flags & AT_REMOVEDIR)
3552 ret = do_rmdir(un->dfd, un->filename);
3554 ret = do_unlinkat(un->dfd, un->filename);
3556 req->flags &= ~REQ_F_NEED_CLEANUP;
3559 io_req_complete(req, ret);
3563 static int io_shutdown_prep(struct io_kiocb *req,
3564 const struct io_uring_sqe *sqe)
3566 #if defined(CONFIG_NET)
3567 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3569 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3573 req->shutdown.how = READ_ONCE(sqe->len);
3580 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3582 #if defined(CONFIG_NET)
3583 struct socket *sock;
3586 if (issue_flags & IO_URING_F_NONBLOCK)
3589 sock = sock_from_file(req->file);
3590 if (unlikely(!sock))
3593 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3596 io_req_complete(req, ret);
3603 static int __io_splice_prep(struct io_kiocb *req,
3604 const struct io_uring_sqe *sqe)
3606 struct io_splice *sp = &req->splice;
3607 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3609 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3613 sp->len = READ_ONCE(sqe->len);
3614 sp->flags = READ_ONCE(sqe->splice_flags);
3616 if (unlikely(sp->flags & ~valid_flags))
3619 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3620 (sp->flags & SPLICE_F_FD_IN_FIXED));
3623 req->flags |= REQ_F_NEED_CLEANUP;
3627 static int io_tee_prep(struct io_kiocb *req,
3628 const struct io_uring_sqe *sqe)
3630 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3632 return __io_splice_prep(req, sqe);
3635 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3637 struct io_splice *sp = &req->splice;
3638 struct file *in = sp->file_in;
3639 struct file *out = sp->file_out;
3640 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3643 if (issue_flags & IO_URING_F_NONBLOCK)
3646 ret = do_tee(in, out, sp->len, flags);
3648 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3650 req->flags &= ~REQ_F_NEED_CLEANUP;
3654 io_req_complete(req, ret);
3658 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3660 struct io_splice *sp = &req->splice;
3662 sp->off_in = READ_ONCE(sqe->splice_off_in);
3663 sp->off_out = READ_ONCE(sqe->off);
3664 return __io_splice_prep(req, sqe);
3667 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3669 struct io_splice *sp = &req->splice;
3670 struct file *in = sp->file_in;
3671 struct file *out = sp->file_out;
3672 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3673 loff_t *poff_in, *poff_out;
3676 if (issue_flags & IO_URING_F_NONBLOCK)
3679 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3680 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3683 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3685 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3687 req->flags &= ~REQ_F_NEED_CLEANUP;
3691 io_req_complete(req, ret);
3696 * IORING_OP_NOP just posts a completion event, nothing else.
3698 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3700 struct io_ring_ctx *ctx = req->ctx;
3702 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3705 __io_req_complete(req, issue_flags, 0, 0);
3709 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3711 struct io_ring_ctx *ctx = req->ctx;
3716 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3718 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3721 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3722 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3725 req->sync.off = READ_ONCE(sqe->off);
3726 req->sync.len = READ_ONCE(sqe->len);
3730 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3732 loff_t end = req->sync.off + req->sync.len;
3735 /* fsync always requires a blocking context */
3736 if (issue_flags & IO_URING_F_NONBLOCK)
3739 ret = vfs_fsync_range(req->file, req->sync.off,
3740 end > 0 ? end : LLONG_MAX,
3741 req->sync.flags & IORING_FSYNC_DATASYNC);
3744 io_req_complete(req, ret);
3748 static int io_fallocate_prep(struct io_kiocb *req,
3749 const struct io_uring_sqe *sqe)
3751 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3753 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3756 req->sync.off = READ_ONCE(sqe->off);
3757 req->sync.len = READ_ONCE(sqe->addr);
3758 req->sync.mode = READ_ONCE(sqe->len);
3762 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3766 /* fallocate always requiring blocking context */
3767 if (issue_flags & IO_URING_F_NONBLOCK)
3769 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3773 io_req_complete(req, ret);
3777 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3779 const char __user *fname;
3782 if (unlikely(sqe->ioprio || sqe->buf_index))
3784 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3787 /* open.how should be already initialised */
3788 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3789 req->open.how.flags |= O_LARGEFILE;
3791 req->open.dfd = READ_ONCE(sqe->fd);
3792 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3793 req->open.filename = getname(fname);
3794 if (IS_ERR(req->open.filename)) {
3795 ret = PTR_ERR(req->open.filename);
3796 req->open.filename = NULL;
3799 req->open.nofile = rlimit(RLIMIT_NOFILE);
3800 req->flags |= REQ_F_NEED_CLEANUP;
3804 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3808 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3810 mode = READ_ONCE(sqe->len);
3811 flags = READ_ONCE(sqe->open_flags);
3812 req->open.how = build_open_how(flags, mode);
3813 return __io_openat_prep(req, sqe);
3816 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3818 struct open_how __user *how;
3822 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3824 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3825 len = READ_ONCE(sqe->len);
3826 if (len < OPEN_HOW_SIZE_VER0)
3829 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3834 return __io_openat_prep(req, sqe);
3837 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3839 struct open_flags op;
3842 bool resolve_nonblock;
3845 ret = build_open_flags(&req->open.how, &op);
3848 nonblock_set = op.open_flag & O_NONBLOCK;
3849 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3850 if (issue_flags & IO_URING_F_NONBLOCK) {
3852 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3853 * it'll always -EAGAIN
3855 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3857 op.lookup_flags |= LOOKUP_CACHED;
3858 op.open_flag |= O_NONBLOCK;
3861 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3865 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3868 * We could hang on to this 'fd' on retrying, but seems like
3869 * marginal gain for something that is now known to be a slower
3870 * path. So just put it, and we'll get a new one when we retry.
3874 ret = PTR_ERR(file);
3875 /* only retry if RESOLVE_CACHED wasn't already set by application */
3876 if (ret == -EAGAIN &&
3877 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3882 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3883 file->f_flags &= ~O_NONBLOCK;
3884 fsnotify_open(file);
3885 fd_install(ret, file);
3887 putname(req->open.filename);
3888 req->flags &= ~REQ_F_NEED_CLEANUP;
3891 __io_req_complete(req, issue_flags, ret, 0);
3895 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3897 return io_openat2(req, issue_flags);
3900 static int io_remove_buffers_prep(struct io_kiocb *req,
3901 const struct io_uring_sqe *sqe)
3903 struct io_provide_buf *p = &req->pbuf;
3906 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3909 tmp = READ_ONCE(sqe->fd);
3910 if (!tmp || tmp > USHRT_MAX)
3913 memset(p, 0, sizeof(*p));
3915 p->bgid = READ_ONCE(sqe->buf_group);
3919 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3920 int bgid, unsigned nbufs)
3924 /* shouldn't happen */
3928 /* the head kbuf is the list itself */
3929 while (!list_empty(&buf->list)) {
3930 struct io_buffer *nxt;
3932 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3933 list_del(&nxt->list);
3940 xa_erase(&ctx->io_buffers, bgid);
3945 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3947 struct io_provide_buf *p = &req->pbuf;
3948 struct io_ring_ctx *ctx = req->ctx;
3949 struct io_buffer *head;
3951 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3953 io_ring_submit_lock(ctx, !force_nonblock);
3955 lockdep_assert_held(&ctx->uring_lock);
3958 head = xa_load(&ctx->io_buffers, p->bgid);
3960 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3964 /* complete before unlock, IOPOLL may need the lock */
3965 __io_req_complete(req, issue_flags, ret, 0);
3966 io_ring_submit_unlock(ctx, !force_nonblock);
3970 static int io_provide_buffers_prep(struct io_kiocb *req,
3971 const struct io_uring_sqe *sqe)
3973 unsigned long size, tmp_check;
3974 struct io_provide_buf *p = &req->pbuf;
3977 if (sqe->ioprio || sqe->rw_flags)
3980 tmp = READ_ONCE(sqe->fd);
3981 if (!tmp || tmp > USHRT_MAX)
3984 p->addr = READ_ONCE(sqe->addr);
3985 p->len = READ_ONCE(sqe->len);
3987 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3990 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3993 size = (unsigned long)p->len * p->nbufs;
3994 if (!access_ok(u64_to_user_ptr(p->addr), size))
3997 p->bgid = READ_ONCE(sqe->buf_group);
3998 tmp = READ_ONCE(sqe->off);
3999 if (tmp > USHRT_MAX)
4005 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4007 struct io_buffer *buf;
4008 u64 addr = pbuf->addr;
4009 int i, bid = pbuf->bid;
4011 for (i = 0; i < pbuf->nbufs; i++) {
4012 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4017 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4022 INIT_LIST_HEAD(&buf->list);
4025 list_add_tail(&buf->list, &(*head)->list);
4029 return i ? i : -ENOMEM;
4032 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4034 struct io_provide_buf *p = &req->pbuf;
4035 struct io_ring_ctx *ctx = req->ctx;
4036 struct io_buffer *head, *list;
4038 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4040 io_ring_submit_lock(ctx, !force_nonblock);
4042 lockdep_assert_held(&ctx->uring_lock);
4044 list = head = xa_load(&ctx->io_buffers, p->bgid);
4046 ret = io_add_buffers(p, &head);
4047 if (ret >= 0 && !list) {
4048 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4050 __io_remove_buffers(ctx, head, p->bgid, -1U);
4054 /* complete before unlock, IOPOLL may need the lock */
4055 __io_req_complete(req, issue_flags, ret, 0);
4056 io_ring_submit_unlock(ctx, !force_nonblock);
4060 static int io_epoll_ctl_prep(struct io_kiocb *req,
4061 const struct io_uring_sqe *sqe)
4063 #if defined(CONFIG_EPOLL)
4064 if (sqe->ioprio || sqe->buf_index)
4066 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4069 req->epoll.epfd = READ_ONCE(sqe->fd);
4070 req->epoll.op = READ_ONCE(sqe->len);
4071 req->epoll.fd = READ_ONCE(sqe->off);
4073 if (ep_op_has_event(req->epoll.op)) {
4074 struct epoll_event __user *ev;
4076 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4077 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4087 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4089 #if defined(CONFIG_EPOLL)
4090 struct io_epoll *ie = &req->epoll;
4092 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4094 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4095 if (force_nonblock && ret == -EAGAIN)
4100 __io_req_complete(req, issue_flags, ret, 0);
4107 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4109 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4110 if (sqe->ioprio || sqe->buf_index || sqe->off)
4112 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4115 req->madvise.addr = READ_ONCE(sqe->addr);
4116 req->madvise.len = READ_ONCE(sqe->len);
4117 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4124 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4126 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4127 struct io_madvise *ma = &req->madvise;
4130 if (issue_flags & IO_URING_F_NONBLOCK)
4133 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4136 io_req_complete(req, ret);
4143 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4145 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4147 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4150 req->fadvise.offset = READ_ONCE(sqe->off);
4151 req->fadvise.len = READ_ONCE(sqe->len);
4152 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4156 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4158 struct io_fadvise *fa = &req->fadvise;
4161 if (issue_flags & IO_URING_F_NONBLOCK) {
4162 switch (fa->advice) {
4163 case POSIX_FADV_NORMAL:
4164 case POSIX_FADV_RANDOM:
4165 case POSIX_FADV_SEQUENTIAL:
4172 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4175 __io_req_complete(req, issue_flags, ret, 0);
4179 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4181 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4183 if (sqe->ioprio || sqe->buf_index)
4185 if (req->flags & REQ_F_FIXED_FILE)
4188 req->statx.dfd = READ_ONCE(sqe->fd);
4189 req->statx.mask = READ_ONCE(sqe->len);
4190 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4191 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4192 req->statx.flags = READ_ONCE(sqe->statx_flags);
4197 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4199 struct io_statx *ctx = &req->statx;
4202 if (issue_flags & IO_URING_F_NONBLOCK)
4205 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4210 io_req_complete(req, ret);
4214 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4216 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4218 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4219 sqe->rw_flags || sqe->buf_index)
4221 if (req->flags & REQ_F_FIXED_FILE)
4224 req->close.fd = READ_ONCE(sqe->fd);
4228 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4230 struct files_struct *files = current->files;
4231 struct io_close *close = &req->close;
4232 struct fdtable *fdt;
4233 struct file *file = NULL;
4236 spin_lock(&files->file_lock);
4237 fdt = files_fdtable(files);
4238 if (close->fd >= fdt->max_fds) {
4239 spin_unlock(&files->file_lock);
4242 file = fdt->fd[close->fd];
4243 if (!file || file->f_op == &io_uring_fops) {
4244 spin_unlock(&files->file_lock);
4249 /* if the file has a flush method, be safe and punt to async */
4250 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4251 spin_unlock(&files->file_lock);
4255 ret = __close_fd_get_file(close->fd, &file);
4256 spin_unlock(&files->file_lock);
4263 /* No ->flush() or already async, safely close from here */
4264 ret = filp_close(file, current->files);
4270 __io_req_complete(req, issue_flags, ret, 0);
4274 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4276 struct io_ring_ctx *ctx = req->ctx;
4278 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4280 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4283 req->sync.off = READ_ONCE(sqe->off);
4284 req->sync.len = READ_ONCE(sqe->len);
4285 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4289 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4293 /* sync_file_range always requires a blocking context */
4294 if (issue_flags & IO_URING_F_NONBLOCK)
4297 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4301 io_req_complete(req, ret);
4305 #if defined(CONFIG_NET)
4306 static int io_setup_async_msg(struct io_kiocb *req,
4307 struct io_async_msghdr *kmsg)
4309 struct io_async_msghdr *async_msg = req->async_data;
4313 if (io_alloc_async_data(req)) {
4314 kfree(kmsg->free_iov);
4317 async_msg = req->async_data;
4318 req->flags |= REQ_F_NEED_CLEANUP;
4319 memcpy(async_msg, kmsg, sizeof(*kmsg));
4320 async_msg->msg.msg_name = &async_msg->addr;
4321 /* if were using fast_iov, set it to the new one */
4322 if (!async_msg->free_iov)
4323 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4328 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4329 struct io_async_msghdr *iomsg)
4331 iomsg->msg.msg_name = &iomsg->addr;
4332 iomsg->free_iov = iomsg->fast_iov;
4333 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4334 req->sr_msg.msg_flags, &iomsg->free_iov);
4337 static int io_sendmsg_prep_async(struct io_kiocb *req)
4341 ret = io_sendmsg_copy_hdr(req, req->async_data);
4343 req->flags |= REQ_F_NEED_CLEANUP;
4347 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4349 struct io_sr_msg *sr = &req->sr_msg;
4351 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4354 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4355 sr->len = READ_ONCE(sqe->len);
4356 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4357 if (sr->msg_flags & MSG_DONTWAIT)
4358 req->flags |= REQ_F_NOWAIT;
4360 #ifdef CONFIG_COMPAT
4361 if (req->ctx->compat)
4362 sr->msg_flags |= MSG_CMSG_COMPAT;
4367 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4369 struct io_async_msghdr iomsg, *kmsg;
4370 struct socket *sock;
4375 sock = sock_from_file(req->file);
4376 if (unlikely(!sock))
4379 kmsg = req->async_data;
4381 ret = io_sendmsg_copy_hdr(req, &iomsg);
4387 flags = req->sr_msg.msg_flags;
4388 if (issue_flags & IO_URING_F_NONBLOCK)
4389 flags |= MSG_DONTWAIT;
4390 if (flags & MSG_WAITALL)
4391 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4393 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4394 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4395 return io_setup_async_msg(req, kmsg);
4396 if (ret == -ERESTARTSYS)
4399 /* fast path, check for non-NULL to avoid function call */
4401 kfree(kmsg->free_iov);
4402 req->flags &= ~REQ_F_NEED_CLEANUP;
4405 __io_req_complete(req, issue_flags, ret, 0);
4409 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4411 struct io_sr_msg *sr = &req->sr_msg;
4414 struct socket *sock;
4419 sock = sock_from_file(req->file);
4420 if (unlikely(!sock))
4423 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4427 msg.msg_name = NULL;
4428 msg.msg_control = NULL;
4429 msg.msg_controllen = 0;
4430 msg.msg_namelen = 0;
4432 flags = req->sr_msg.msg_flags;
4433 if (issue_flags & IO_URING_F_NONBLOCK)
4434 flags |= MSG_DONTWAIT;
4435 if (flags & MSG_WAITALL)
4436 min_ret = iov_iter_count(&msg.msg_iter);
4438 msg.msg_flags = flags;
4439 ret = sock_sendmsg(sock, &msg);
4440 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4442 if (ret == -ERESTARTSYS)
4447 __io_req_complete(req, issue_flags, ret, 0);
4451 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4452 struct io_async_msghdr *iomsg)
4454 struct io_sr_msg *sr = &req->sr_msg;
4455 struct iovec __user *uiov;
4459 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4460 &iomsg->uaddr, &uiov, &iov_len);
4464 if (req->flags & REQ_F_BUFFER_SELECT) {
4467 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4469 sr->len = iomsg->fast_iov[0].iov_len;
4470 iomsg->free_iov = NULL;
4472 iomsg->free_iov = iomsg->fast_iov;
4473 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4474 &iomsg->free_iov, &iomsg->msg.msg_iter,
4483 #ifdef CONFIG_COMPAT
4484 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4485 struct io_async_msghdr *iomsg)
4487 struct io_sr_msg *sr = &req->sr_msg;
4488 struct compat_iovec __user *uiov;
4493 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4498 uiov = compat_ptr(ptr);
4499 if (req->flags & REQ_F_BUFFER_SELECT) {
4500 compat_ssize_t clen;
4504 if (!access_ok(uiov, sizeof(*uiov)))
4506 if (__get_user(clen, &uiov->iov_len))
4511 iomsg->free_iov = NULL;
4513 iomsg->free_iov = iomsg->fast_iov;
4514 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4515 UIO_FASTIOV, &iomsg->free_iov,
4516 &iomsg->msg.msg_iter, true);
4525 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4526 struct io_async_msghdr *iomsg)
4528 iomsg->msg.msg_name = &iomsg->addr;
4530 #ifdef CONFIG_COMPAT
4531 if (req->ctx->compat)
4532 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4535 return __io_recvmsg_copy_hdr(req, iomsg);
4538 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4541 struct io_sr_msg *sr = &req->sr_msg;
4542 struct io_buffer *kbuf;
4544 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4549 req->flags |= REQ_F_BUFFER_SELECTED;
4553 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4555 return io_put_kbuf(req, req->sr_msg.kbuf);
4558 static int io_recvmsg_prep_async(struct io_kiocb *req)
4562 ret = io_recvmsg_copy_hdr(req, req->async_data);
4564 req->flags |= REQ_F_NEED_CLEANUP;
4568 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4570 struct io_sr_msg *sr = &req->sr_msg;
4572 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4575 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4576 sr->len = READ_ONCE(sqe->len);
4577 sr->bgid = READ_ONCE(sqe->buf_group);
4578 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4579 if (sr->msg_flags & MSG_DONTWAIT)
4580 req->flags |= REQ_F_NOWAIT;
4582 #ifdef CONFIG_COMPAT
4583 if (req->ctx->compat)
4584 sr->msg_flags |= MSG_CMSG_COMPAT;
4589 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4591 struct io_async_msghdr iomsg, *kmsg;
4592 struct socket *sock;
4593 struct io_buffer *kbuf;
4596 int ret, cflags = 0;
4597 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4599 sock = sock_from_file(req->file);
4600 if (unlikely(!sock))
4603 kmsg = req->async_data;
4605 ret = io_recvmsg_copy_hdr(req, &iomsg);
4611 if (req->flags & REQ_F_BUFFER_SELECT) {
4612 kbuf = io_recv_buffer_select(req, !force_nonblock);
4614 return PTR_ERR(kbuf);
4615 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4616 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4617 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4618 1, req->sr_msg.len);
4621 flags = req->sr_msg.msg_flags;
4623 flags |= MSG_DONTWAIT;
4624 if (flags & MSG_WAITALL)
4625 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4627 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4628 kmsg->uaddr, flags);
4629 if (force_nonblock && ret == -EAGAIN)
4630 return io_setup_async_msg(req, kmsg);
4631 if (ret == -ERESTARTSYS)
4634 if (req->flags & REQ_F_BUFFER_SELECTED)
4635 cflags = io_put_recv_kbuf(req);
4636 /* fast path, check for non-NULL to avoid function call */
4638 kfree(kmsg->free_iov);
4639 req->flags &= ~REQ_F_NEED_CLEANUP;
4640 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4642 __io_req_complete(req, issue_flags, ret, cflags);
4646 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4648 struct io_buffer *kbuf;
4649 struct io_sr_msg *sr = &req->sr_msg;
4651 void __user *buf = sr->buf;
4652 struct socket *sock;
4656 int ret, cflags = 0;
4657 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4659 sock = sock_from_file(req->file);
4660 if (unlikely(!sock))
4663 if (req->flags & REQ_F_BUFFER_SELECT) {
4664 kbuf = io_recv_buffer_select(req, !force_nonblock);
4666 return PTR_ERR(kbuf);
4667 buf = u64_to_user_ptr(kbuf->addr);
4670 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4674 msg.msg_name = NULL;
4675 msg.msg_control = NULL;
4676 msg.msg_controllen = 0;
4677 msg.msg_namelen = 0;
4678 msg.msg_iocb = NULL;
4681 flags = req->sr_msg.msg_flags;
4683 flags |= MSG_DONTWAIT;
4684 if (flags & MSG_WAITALL)
4685 min_ret = iov_iter_count(&msg.msg_iter);
4687 ret = sock_recvmsg(sock, &msg, flags);
4688 if (force_nonblock && ret == -EAGAIN)
4690 if (ret == -ERESTARTSYS)
4693 if (req->flags & REQ_F_BUFFER_SELECTED)
4694 cflags = io_put_recv_kbuf(req);
4695 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4697 __io_req_complete(req, issue_flags, ret, cflags);
4701 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4703 struct io_accept *accept = &req->accept;
4705 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4707 if (sqe->ioprio || sqe->len || sqe->buf_index)
4710 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4711 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4712 accept->flags = READ_ONCE(sqe->accept_flags);
4713 accept->nofile = rlimit(RLIMIT_NOFILE);
4717 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4719 struct io_accept *accept = &req->accept;
4720 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4721 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4724 if (req->file->f_flags & O_NONBLOCK)
4725 req->flags |= REQ_F_NOWAIT;
4727 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4728 accept->addr_len, accept->flags,
4730 if (ret == -EAGAIN && force_nonblock)
4733 if (ret == -ERESTARTSYS)
4737 __io_req_complete(req, issue_flags, ret, 0);
4741 static int io_connect_prep_async(struct io_kiocb *req)
4743 struct io_async_connect *io = req->async_data;
4744 struct io_connect *conn = &req->connect;
4746 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4749 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4751 struct io_connect *conn = &req->connect;
4753 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4755 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4758 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4759 conn->addr_len = READ_ONCE(sqe->addr2);
4763 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4765 struct io_async_connect __io, *io;
4766 unsigned file_flags;
4768 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4770 if (req->async_data) {
4771 io = req->async_data;
4773 ret = move_addr_to_kernel(req->connect.addr,
4774 req->connect.addr_len,
4781 file_flags = force_nonblock ? O_NONBLOCK : 0;
4783 ret = __sys_connect_file(req->file, &io->address,
4784 req->connect.addr_len, file_flags);
4785 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4786 if (req->async_data)
4788 if (io_alloc_async_data(req)) {
4792 memcpy(req->async_data, &__io, sizeof(__io));
4795 if (ret == -ERESTARTSYS)
4800 __io_req_complete(req, issue_flags, ret, 0);
4803 #else /* !CONFIG_NET */
4804 #define IO_NETOP_FN(op) \
4805 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4807 return -EOPNOTSUPP; \
4810 #define IO_NETOP_PREP(op) \
4812 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4814 return -EOPNOTSUPP; \
4817 #define IO_NETOP_PREP_ASYNC(op) \
4819 static int io_##op##_prep_async(struct io_kiocb *req) \
4821 return -EOPNOTSUPP; \
4824 IO_NETOP_PREP_ASYNC(sendmsg);
4825 IO_NETOP_PREP_ASYNC(recvmsg);
4826 IO_NETOP_PREP_ASYNC(connect);
4827 IO_NETOP_PREP(accept);
4830 #endif /* CONFIG_NET */
4832 struct io_poll_table {
4833 struct poll_table_struct pt;
4834 struct io_kiocb *req;
4839 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4840 __poll_t mask, io_req_tw_func_t func)
4842 /* for instances that support it check for an event match first: */
4843 if (mask && !(mask & poll->events))
4846 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4848 list_del_init(&poll->wait.entry);
4851 req->io_task_work.func = func;
4854 * If this fails, then the task is exiting. When a task exits, the
4855 * work gets canceled, so just cancel this request as well instead
4856 * of executing it. We can't safely execute it anyway, as we may not
4857 * have the needed state needed for it anyway.
4859 io_req_task_work_add(req);
4863 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4864 __acquires(&req->ctx->completion_lock)
4866 struct io_ring_ctx *ctx = req->ctx;
4868 if (unlikely(req->task->flags & PF_EXITING))
4869 WRITE_ONCE(poll->canceled, true);
4871 if (!req->result && !READ_ONCE(poll->canceled)) {
4872 struct poll_table_struct pt = { ._key = poll->events };
4874 req->result = vfs_poll(req->file, &pt) & poll->events;
4877 spin_lock_irq(&ctx->completion_lock);
4878 if (!req->result && !READ_ONCE(poll->canceled)) {
4879 add_wait_queue(poll->head, &poll->wait);
4886 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4888 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4889 if (req->opcode == IORING_OP_POLL_ADD)
4890 return req->async_data;
4891 return req->apoll->double_poll;
4894 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4896 if (req->opcode == IORING_OP_POLL_ADD)
4898 return &req->apoll->poll;
4901 static void io_poll_remove_double(struct io_kiocb *req)
4902 __must_hold(&req->ctx->completion_lock)
4904 struct io_poll_iocb *poll = io_poll_get_double(req);
4906 lockdep_assert_held(&req->ctx->completion_lock);
4908 if (poll && poll->head) {
4909 struct wait_queue_head *head = poll->head;
4911 spin_lock(&head->lock);
4912 list_del_init(&poll->wait.entry);
4913 if (poll->wait.private)
4916 spin_unlock(&head->lock);
4920 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4921 __must_hold(&req->ctx->completion_lock)
4923 struct io_ring_ctx *ctx = req->ctx;
4924 unsigned flags = IORING_CQE_F_MORE;
4927 if (READ_ONCE(req->poll.canceled)) {
4929 req->poll.events |= EPOLLONESHOT;
4931 error = mangle_poll(mask);
4933 if (req->poll.events & EPOLLONESHOT)
4935 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4936 io_poll_remove_waitqs(req);
4937 req->poll.done = true;
4940 if (flags & IORING_CQE_F_MORE)
4943 io_commit_cqring(ctx);
4944 return !(flags & IORING_CQE_F_MORE);
4947 static void io_poll_task_func(struct io_kiocb *req)
4949 struct io_ring_ctx *ctx = req->ctx;
4950 struct io_kiocb *nxt;
4952 if (io_poll_rewait(req, &req->poll)) {
4953 spin_unlock_irq(&ctx->completion_lock);
4957 done = io_poll_complete(req, req->result);
4959 hash_del(&req->hash_node);
4962 add_wait_queue(req->poll.head, &req->poll.wait);
4964 spin_unlock_irq(&ctx->completion_lock);
4965 io_cqring_ev_posted(ctx);
4968 nxt = io_put_req_find_next(req);
4970 io_req_task_submit(nxt);
4975 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4976 int sync, void *key)
4978 struct io_kiocb *req = wait->private;
4979 struct io_poll_iocb *poll = io_poll_get_single(req);
4980 __poll_t mask = key_to_poll(key);
4982 /* for instances that support it check for an event match first: */
4983 if (mask && !(mask & poll->events))
4985 if (!(poll->events & EPOLLONESHOT))
4986 return poll->wait.func(&poll->wait, mode, sync, key);
4988 list_del_init(&wait->entry);
4993 spin_lock(&poll->head->lock);
4994 done = list_empty(&poll->wait.entry);
4996 list_del_init(&poll->wait.entry);
4997 /* make sure double remove sees this as being gone */
4998 wait->private = NULL;
4999 spin_unlock(&poll->head->lock);
5001 /* use wait func handler, so it matches the rq type */
5002 poll->wait.func(&poll->wait, mode, sync, key);
5009 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5010 wait_queue_func_t wake_func)
5014 poll->canceled = false;
5015 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5016 /* mask in events that we always want/need */
5017 poll->events = events | IO_POLL_UNMASK;
5018 INIT_LIST_HEAD(&poll->wait.entry);
5019 init_waitqueue_func_entry(&poll->wait, wake_func);
5022 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5023 struct wait_queue_head *head,
5024 struct io_poll_iocb **poll_ptr)
5026 struct io_kiocb *req = pt->req;
5029 * The file being polled uses multiple waitqueues for poll handling
5030 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5033 if (unlikely(pt->nr_entries)) {
5034 struct io_poll_iocb *poll_one = poll;
5036 /* already have a 2nd entry, fail a third attempt */
5038 pt->error = -EINVAL;
5042 * Can't handle multishot for double wait for now, turn it
5043 * into one-shot mode.
5045 if (!(poll_one->events & EPOLLONESHOT))
5046 poll_one->events |= EPOLLONESHOT;
5047 /* double add on the same waitqueue head, ignore */
5048 if (poll_one->head == head)
5050 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5052 pt->error = -ENOMEM;
5055 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5057 poll->wait.private = req;
5064 if (poll->events & EPOLLEXCLUSIVE)
5065 add_wait_queue_exclusive(head, &poll->wait);
5067 add_wait_queue(head, &poll->wait);
5070 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5071 struct poll_table_struct *p)
5073 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5074 struct async_poll *apoll = pt->req->apoll;
5076 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5079 static void io_async_task_func(struct io_kiocb *req)
5081 struct async_poll *apoll = req->apoll;
5082 struct io_ring_ctx *ctx = req->ctx;
5084 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5086 if (io_poll_rewait(req, &apoll->poll)) {
5087 spin_unlock_irq(&ctx->completion_lock);
5091 hash_del(&req->hash_node);
5092 io_poll_remove_double(req);
5093 spin_unlock_irq(&ctx->completion_lock);
5095 if (!READ_ONCE(apoll->poll.canceled))
5096 io_req_task_submit(req);
5098 io_req_complete_failed(req, -ECANCELED);
5101 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5104 struct io_kiocb *req = wait->private;
5105 struct io_poll_iocb *poll = &req->apoll->poll;
5107 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5110 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5113 static void io_poll_req_insert(struct io_kiocb *req)
5115 struct io_ring_ctx *ctx = req->ctx;
5116 struct hlist_head *list;
5118 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5119 hlist_add_head(&req->hash_node, list);
5122 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5123 struct io_poll_iocb *poll,
5124 struct io_poll_table *ipt, __poll_t mask,
5125 wait_queue_func_t wake_func)
5126 __acquires(&ctx->completion_lock)
5128 struct io_ring_ctx *ctx = req->ctx;
5129 bool cancel = false;
5131 INIT_HLIST_NODE(&req->hash_node);
5132 io_init_poll_iocb(poll, mask, wake_func);
5133 poll->file = req->file;
5134 poll->wait.private = req;
5136 ipt->pt._key = mask;
5139 ipt->nr_entries = 0;
5141 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5142 if (unlikely(!ipt->nr_entries) && !ipt->error)
5143 ipt->error = -EINVAL;
5145 spin_lock_irq(&ctx->completion_lock);
5147 io_poll_remove_double(req);
5148 if (likely(poll->head)) {
5149 spin_lock(&poll->head->lock);
5150 if (unlikely(list_empty(&poll->wait.entry))) {
5156 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5157 list_del_init(&poll->wait.entry);
5159 WRITE_ONCE(poll->canceled, true);
5160 else if (!poll->done) /* actually waiting for an event */
5161 io_poll_req_insert(req);
5162 spin_unlock(&poll->head->lock);
5174 static int io_arm_poll_handler(struct io_kiocb *req)
5176 const struct io_op_def *def = &io_op_defs[req->opcode];
5177 struct io_ring_ctx *ctx = req->ctx;
5178 struct async_poll *apoll;
5179 struct io_poll_table ipt;
5180 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5183 if (!req->file || !file_can_poll(req->file))
5184 return IO_APOLL_ABORTED;
5185 if (req->flags & REQ_F_POLLED)
5186 return IO_APOLL_ABORTED;
5187 if (!def->pollin && !def->pollout)
5188 return IO_APOLL_ABORTED;
5192 mask |= POLLIN | POLLRDNORM;
5194 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5195 if ((req->opcode == IORING_OP_RECVMSG) &&
5196 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5200 mask |= POLLOUT | POLLWRNORM;
5203 /* if we can't nonblock try, then no point in arming a poll handler */
5204 if (!io_file_supports_async(req, rw))
5205 return IO_APOLL_ABORTED;
5207 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5208 if (unlikely(!apoll))
5209 return IO_APOLL_ABORTED;
5210 apoll->double_poll = NULL;
5212 req->flags |= REQ_F_POLLED;
5213 ipt.pt._qproc = io_async_queue_proc;
5215 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5217 if (ret || ipt.error) {
5218 io_poll_remove_double(req);
5219 spin_unlock_irq(&ctx->completion_lock);
5221 return IO_APOLL_READY;
5222 return IO_APOLL_ABORTED;
5224 spin_unlock_irq(&ctx->completion_lock);
5225 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5226 mask, apoll->poll.events);
5230 static bool __io_poll_remove_one(struct io_kiocb *req,
5231 struct io_poll_iocb *poll, bool do_cancel)
5232 __must_hold(&req->ctx->completion_lock)
5234 bool do_complete = false;
5238 spin_lock(&poll->head->lock);
5240 WRITE_ONCE(poll->canceled, true);
5241 if (!list_empty(&poll->wait.entry)) {
5242 list_del_init(&poll->wait.entry);
5245 spin_unlock(&poll->head->lock);
5246 hash_del(&req->hash_node);
5250 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5251 __must_hold(&req->ctx->completion_lock)
5255 io_poll_remove_double(req);
5256 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5258 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5259 /* non-poll requests have submit ref still */
5265 static bool io_poll_remove_one(struct io_kiocb *req)
5266 __must_hold(&req->ctx->completion_lock)
5270 do_complete = io_poll_remove_waitqs(req);
5272 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5273 io_commit_cqring(req->ctx);
5275 io_put_req_deferred(req, 1);
5282 * Returns true if we found and killed one or more poll requests
5284 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5287 struct hlist_node *tmp;
5288 struct io_kiocb *req;
5291 spin_lock_irq(&ctx->completion_lock);
5292 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5293 struct hlist_head *list;
5295 list = &ctx->cancel_hash[i];
5296 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5297 if (io_match_task(req, tsk, cancel_all))
5298 posted += io_poll_remove_one(req);
5301 spin_unlock_irq(&ctx->completion_lock);
5304 io_cqring_ev_posted(ctx);
5309 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5311 __must_hold(&ctx->completion_lock)
5313 struct hlist_head *list;
5314 struct io_kiocb *req;
5316 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5317 hlist_for_each_entry(req, list, hash_node) {
5318 if (sqe_addr != req->user_data)
5320 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5327 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5329 __must_hold(&ctx->completion_lock)
5331 struct io_kiocb *req;
5333 req = io_poll_find(ctx, sqe_addr, poll_only);
5336 if (io_poll_remove_one(req))
5342 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5347 events = READ_ONCE(sqe->poll32_events);
5349 events = swahw32(events);
5351 if (!(flags & IORING_POLL_ADD_MULTI))
5352 events |= EPOLLONESHOT;
5353 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5356 static int io_poll_update_prep(struct io_kiocb *req,
5357 const struct io_uring_sqe *sqe)
5359 struct io_poll_update *upd = &req->poll_update;
5362 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5364 if (sqe->ioprio || sqe->buf_index)
5366 flags = READ_ONCE(sqe->len);
5367 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5368 IORING_POLL_ADD_MULTI))
5370 /* meaningless without update */
5371 if (flags == IORING_POLL_ADD_MULTI)
5374 upd->old_user_data = READ_ONCE(sqe->addr);
5375 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5376 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5378 upd->new_user_data = READ_ONCE(sqe->off);
5379 if (!upd->update_user_data && upd->new_user_data)
5381 if (upd->update_events)
5382 upd->events = io_poll_parse_events(sqe, flags);
5383 else if (sqe->poll32_events)
5389 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5392 struct io_kiocb *req = wait->private;
5393 struct io_poll_iocb *poll = &req->poll;
5395 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5398 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5399 struct poll_table_struct *p)
5401 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5403 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5406 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5408 struct io_poll_iocb *poll = &req->poll;
5411 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5413 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5415 flags = READ_ONCE(sqe->len);
5416 if (flags & ~IORING_POLL_ADD_MULTI)
5419 poll->events = io_poll_parse_events(sqe, flags);
5423 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5425 struct io_poll_iocb *poll = &req->poll;
5426 struct io_ring_ctx *ctx = req->ctx;
5427 struct io_poll_table ipt;
5430 ipt.pt._qproc = io_poll_queue_proc;
5432 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5435 if (mask) { /* no async, we'd stolen it */
5437 io_poll_complete(req, mask);
5439 spin_unlock_irq(&ctx->completion_lock);
5442 io_cqring_ev_posted(ctx);
5443 if (poll->events & EPOLLONESHOT)
5449 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5451 struct io_ring_ctx *ctx = req->ctx;
5452 struct io_kiocb *preq;
5456 spin_lock_irq(&ctx->completion_lock);
5457 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5463 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5465 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5470 * Don't allow racy completion with singleshot, as we cannot safely
5471 * update those. For multishot, if we're racing with completion, just
5472 * let completion re-add it.
5474 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5475 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5479 /* we now have a detached poll request. reissue. */
5483 spin_unlock_irq(&ctx->completion_lock);
5485 io_req_complete(req, ret);
5488 /* only mask one event flags, keep behavior flags */
5489 if (req->poll_update.update_events) {
5490 preq->poll.events &= ~0xffff;
5491 preq->poll.events |= req->poll_update.events & 0xffff;
5492 preq->poll.events |= IO_POLL_UNMASK;
5494 if (req->poll_update.update_user_data)
5495 preq->user_data = req->poll_update.new_user_data;
5496 spin_unlock_irq(&ctx->completion_lock);
5498 /* complete update request, we're done with it */
5499 io_req_complete(req, ret);
5502 ret = io_poll_add(preq, issue_flags);
5505 io_req_complete(preq, ret);
5511 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5513 struct io_timeout_data *data = container_of(timer,
5514 struct io_timeout_data, timer);
5515 struct io_kiocb *req = data->req;
5516 struct io_ring_ctx *ctx = req->ctx;
5517 unsigned long flags;
5519 spin_lock_irqsave(&ctx->completion_lock, flags);
5520 list_del_init(&req->timeout.list);
5521 atomic_set(&req->ctx->cq_timeouts,
5522 atomic_read(&req->ctx->cq_timeouts) + 1);
5524 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5525 io_commit_cqring(ctx);
5526 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5528 io_cqring_ev_posted(ctx);
5531 return HRTIMER_NORESTART;
5534 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5536 __must_hold(&ctx->completion_lock)
5538 struct io_timeout_data *io;
5539 struct io_kiocb *req;
5542 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5543 found = user_data == req->user_data;
5548 return ERR_PTR(-ENOENT);
5550 io = req->async_data;
5551 if (hrtimer_try_to_cancel(&io->timer) == -1)
5552 return ERR_PTR(-EALREADY);
5553 list_del_init(&req->timeout.list);
5557 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5558 __must_hold(&ctx->completion_lock)
5560 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5563 return PTR_ERR(req);
5566 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5567 io_put_req_deferred(req, 1);
5571 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5572 struct timespec64 *ts, enum hrtimer_mode mode)
5573 __must_hold(&ctx->completion_lock)
5575 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5576 struct io_timeout_data *data;
5579 return PTR_ERR(req);
5581 req->timeout.off = 0; /* noseq */
5582 data = req->async_data;
5583 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5584 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5585 data->timer.function = io_timeout_fn;
5586 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5590 static int io_timeout_remove_prep(struct io_kiocb *req,
5591 const struct io_uring_sqe *sqe)
5593 struct io_timeout_rem *tr = &req->timeout_rem;
5595 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5597 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5599 if (sqe->ioprio || sqe->buf_index || sqe->len)
5602 tr->addr = READ_ONCE(sqe->addr);
5603 tr->flags = READ_ONCE(sqe->timeout_flags);
5604 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5605 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5607 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5609 } else if (tr->flags) {
5610 /* timeout removal doesn't support flags */
5617 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5619 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5624 * Remove or update an existing timeout command
5626 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5628 struct io_timeout_rem *tr = &req->timeout_rem;
5629 struct io_ring_ctx *ctx = req->ctx;
5632 spin_lock_irq(&ctx->completion_lock);
5633 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5634 ret = io_timeout_cancel(ctx, tr->addr);
5636 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5637 io_translate_timeout_mode(tr->flags));
5639 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5640 io_commit_cqring(ctx);
5641 spin_unlock_irq(&ctx->completion_lock);
5642 io_cqring_ev_posted(ctx);
5649 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5650 bool is_timeout_link)
5652 struct io_timeout_data *data;
5654 u32 off = READ_ONCE(sqe->off);
5656 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5658 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5660 if (off && is_timeout_link)
5662 flags = READ_ONCE(sqe->timeout_flags);
5663 if (flags & ~IORING_TIMEOUT_ABS)
5666 req->timeout.off = off;
5667 if (unlikely(off && !req->ctx->off_timeout_used))
5668 req->ctx->off_timeout_used = true;
5670 if (!req->async_data && io_alloc_async_data(req))
5673 data = req->async_data;
5676 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5679 data->mode = io_translate_timeout_mode(flags);
5680 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5681 if (is_timeout_link)
5682 io_req_track_inflight(req);
5686 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5688 struct io_ring_ctx *ctx = req->ctx;
5689 struct io_timeout_data *data = req->async_data;
5690 struct list_head *entry;
5691 u32 tail, off = req->timeout.off;
5693 spin_lock_irq(&ctx->completion_lock);
5696 * sqe->off holds how many events that need to occur for this
5697 * timeout event to be satisfied. If it isn't set, then this is
5698 * a pure timeout request, sequence isn't used.
5700 if (io_is_timeout_noseq(req)) {
5701 entry = ctx->timeout_list.prev;
5705 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5706 req->timeout.target_seq = tail + off;
5708 /* Update the last seq here in case io_flush_timeouts() hasn't.
5709 * This is safe because ->completion_lock is held, and submissions
5710 * and completions are never mixed in the same ->completion_lock section.
5712 ctx->cq_last_tm_flush = tail;
5715 * Insertion sort, ensuring the first entry in the list is always
5716 * the one we need first.
5718 list_for_each_prev(entry, &ctx->timeout_list) {
5719 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5722 if (io_is_timeout_noseq(nxt))
5724 /* nxt.seq is behind @tail, otherwise would've been completed */
5725 if (off >= nxt->timeout.target_seq - tail)
5729 list_add(&req->timeout.list, entry);
5730 data->timer.function = io_timeout_fn;
5731 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5732 spin_unlock_irq(&ctx->completion_lock);
5736 struct io_cancel_data {
5737 struct io_ring_ctx *ctx;
5741 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5743 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5744 struct io_cancel_data *cd = data;
5746 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5749 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5750 struct io_ring_ctx *ctx)
5752 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5753 enum io_wq_cancel cancel_ret;
5756 if (!tctx || !tctx->io_wq)
5759 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5760 switch (cancel_ret) {
5761 case IO_WQ_CANCEL_OK:
5764 case IO_WQ_CANCEL_RUNNING:
5767 case IO_WQ_CANCEL_NOTFOUND:
5775 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5776 struct io_kiocb *req, __u64 sqe_addr,
5779 unsigned long flags;
5782 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5783 spin_lock_irqsave(&ctx->completion_lock, flags);
5786 ret = io_timeout_cancel(ctx, sqe_addr);
5789 ret = io_poll_cancel(ctx, sqe_addr, false);
5793 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5794 io_commit_cqring(ctx);
5795 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5796 io_cqring_ev_posted(ctx);
5802 static int io_async_cancel_prep(struct io_kiocb *req,
5803 const struct io_uring_sqe *sqe)
5805 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5807 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5809 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5812 req->cancel.addr = READ_ONCE(sqe->addr);
5816 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5818 struct io_ring_ctx *ctx = req->ctx;
5819 u64 sqe_addr = req->cancel.addr;
5820 struct io_tctx_node *node;
5823 /* tasks should wait for their io-wq threads, so safe w/o sync */
5824 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5825 spin_lock_irq(&ctx->completion_lock);
5828 ret = io_timeout_cancel(ctx, sqe_addr);
5831 ret = io_poll_cancel(ctx, sqe_addr, false);
5834 spin_unlock_irq(&ctx->completion_lock);
5836 /* slow path, try all io-wq's */
5837 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5839 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5840 struct io_uring_task *tctx = node->task->io_uring;
5842 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5846 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5848 spin_lock_irq(&ctx->completion_lock);
5850 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5851 io_commit_cqring(ctx);
5852 spin_unlock_irq(&ctx->completion_lock);
5853 io_cqring_ev_posted(ctx);
5861 static int io_rsrc_update_prep(struct io_kiocb *req,
5862 const struct io_uring_sqe *sqe)
5864 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5866 if (sqe->ioprio || sqe->rw_flags)
5869 req->rsrc_update.offset = READ_ONCE(sqe->off);
5870 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5871 if (!req->rsrc_update.nr_args)
5873 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5877 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5879 struct io_ring_ctx *ctx = req->ctx;
5880 struct io_uring_rsrc_update2 up;
5883 if (issue_flags & IO_URING_F_NONBLOCK)
5886 up.offset = req->rsrc_update.offset;
5887 up.data = req->rsrc_update.arg;
5892 mutex_lock(&ctx->uring_lock);
5893 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5894 &up, req->rsrc_update.nr_args);
5895 mutex_unlock(&ctx->uring_lock);
5899 __io_req_complete(req, issue_flags, ret, 0);
5903 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5905 switch (req->opcode) {
5908 case IORING_OP_READV:
5909 case IORING_OP_READ_FIXED:
5910 case IORING_OP_READ:
5911 return io_read_prep(req, sqe);
5912 case IORING_OP_WRITEV:
5913 case IORING_OP_WRITE_FIXED:
5914 case IORING_OP_WRITE:
5915 return io_write_prep(req, sqe);
5916 case IORING_OP_POLL_ADD:
5917 return io_poll_add_prep(req, sqe);
5918 case IORING_OP_POLL_REMOVE:
5919 return io_poll_update_prep(req, sqe);
5920 case IORING_OP_FSYNC:
5921 return io_fsync_prep(req, sqe);
5922 case IORING_OP_SYNC_FILE_RANGE:
5923 return io_sfr_prep(req, sqe);
5924 case IORING_OP_SENDMSG:
5925 case IORING_OP_SEND:
5926 return io_sendmsg_prep(req, sqe);
5927 case IORING_OP_RECVMSG:
5928 case IORING_OP_RECV:
5929 return io_recvmsg_prep(req, sqe);
5930 case IORING_OP_CONNECT:
5931 return io_connect_prep(req, sqe);
5932 case IORING_OP_TIMEOUT:
5933 return io_timeout_prep(req, sqe, false);
5934 case IORING_OP_TIMEOUT_REMOVE:
5935 return io_timeout_remove_prep(req, sqe);
5936 case IORING_OP_ASYNC_CANCEL:
5937 return io_async_cancel_prep(req, sqe);
5938 case IORING_OP_LINK_TIMEOUT:
5939 return io_timeout_prep(req, sqe, true);
5940 case IORING_OP_ACCEPT:
5941 return io_accept_prep(req, sqe);
5942 case IORING_OP_FALLOCATE:
5943 return io_fallocate_prep(req, sqe);
5944 case IORING_OP_OPENAT:
5945 return io_openat_prep(req, sqe);
5946 case IORING_OP_CLOSE:
5947 return io_close_prep(req, sqe);
5948 case IORING_OP_FILES_UPDATE:
5949 return io_rsrc_update_prep(req, sqe);
5950 case IORING_OP_STATX:
5951 return io_statx_prep(req, sqe);
5952 case IORING_OP_FADVISE:
5953 return io_fadvise_prep(req, sqe);
5954 case IORING_OP_MADVISE:
5955 return io_madvise_prep(req, sqe);
5956 case IORING_OP_OPENAT2:
5957 return io_openat2_prep(req, sqe);
5958 case IORING_OP_EPOLL_CTL:
5959 return io_epoll_ctl_prep(req, sqe);
5960 case IORING_OP_SPLICE:
5961 return io_splice_prep(req, sqe);
5962 case IORING_OP_PROVIDE_BUFFERS:
5963 return io_provide_buffers_prep(req, sqe);
5964 case IORING_OP_REMOVE_BUFFERS:
5965 return io_remove_buffers_prep(req, sqe);
5967 return io_tee_prep(req, sqe);
5968 case IORING_OP_SHUTDOWN:
5969 return io_shutdown_prep(req, sqe);
5970 case IORING_OP_RENAMEAT:
5971 return io_renameat_prep(req, sqe);
5972 case IORING_OP_UNLINKAT:
5973 return io_unlinkat_prep(req, sqe);
5976 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5981 static int io_req_prep_async(struct io_kiocb *req)
5983 if (!io_op_defs[req->opcode].needs_async_setup)
5985 if (WARN_ON_ONCE(req->async_data))
5987 if (io_alloc_async_data(req))
5990 switch (req->opcode) {
5991 case IORING_OP_READV:
5992 return io_rw_prep_async(req, READ);
5993 case IORING_OP_WRITEV:
5994 return io_rw_prep_async(req, WRITE);
5995 case IORING_OP_SENDMSG:
5996 return io_sendmsg_prep_async(req);
5997 case IORING_OP_RECVMSG:
5998 return io_recvmsg_prep_async(req);
5999 case IORING_OP_CONNECT:
6000 return io_connect_prep_async(req);
6002 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6007 static u32 io_get_sequence(struct io_kiocb *req)
6009 u32 seq = req->ctx->cached_sq_head;
6011 /* need original cached_sq_head, but it was increased for each req */
6012 io_for_each_link(req, req)
6017 static bool io_drain_req(struct io_kiocb *req)
6019 struct io_kiocb *pos;
6020 struct io_ring_ctx *ctx = req->ctx;
6021 struct io_defer_entry *de;
6026 * If we need to drain a request in the middle of a link, drain the
6027 * head request and the next request/link after the current link.
6028 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6029 * maintained for every request of our link.
6031 if (ctx->drain_next) {
6032 req->flags |= REQ_F_IO_DRAIN;
6033 ctx->drain_next = false;
6035 /* not interested in head, start from the first linked */
6036 io_for_each_link(pos, req->link) {
6037 if (pos->flags & REQ_F_IO_DRAIN) {
6038 ctx->drain_next = true;
6039 req->flags |= REQ_F_IO_DRAIN;
6044 /* Still need defer if there is pending req in defer list. */
6045 if (likely(list_empty_careful(&ctx->defer_list) &&
6046 !(req->flags & REQ_F_IO_DRAIN))) {
6047 ctx->drain_active = false;
6051 seq = io_get_sequence(req);
6052 /* Still a chance to pass the sequence check */
6053 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6056 ret = io_req_prep_async(req);
6059 io_prep_async_link(req);
6060 de = kmalloc(sizeof(*de), GFP_KERNEL);
6064 io_req_complete_failed(req, ret);
6068 spin_lock_irq(&ctx->completion_lock);
6069 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6070 spin_unlock_irq(&ctx->completion_lock);
6072 io_queue_async_work(req);
6076 trace_io_uring_defer(ctx, req, req->user_data);
6079 list_add_tail(&de->list, &ctx->defer_list);
6080 spin_unlock_irq(&ctx->completion_lock);
6084 static void io_clean_op(struct io_kiocb *req)
6086 if (req->flags & REQ_F_BUFFER_SELECTED) {
6087 switch (req->opcode) {
6088 case IORING_OP_READV:
6089 case IORING_OP_READ_FIXED:
6090 case IORING_OP_READ:
6091 kfree((void *)(unsigned long)req->rw.addr);
6093 case IORING_OP_RECVMSG:
6094 case IORING_OP_RECV:
6095 kfree(req->sr_msg.kbuf);
6100 if (req->flags & REQ_F_NEED_CLEANUP) {
6101 switch (req->opcode) {
6102 case IORING_OP_READV:
6103 case IORING_OP_READ_FIXED:
6104 case IORING_OP_READ:
6105 case IORING_OP_WRITEV:
6106 case IORING_OP_WRITE_FIXED:
6107 case IORING_OP_WRITE: {
6108 struct io_async_rw *io = req->async_data;
6110 kfree(io->free_iovec);
6113 case IORING_OP_RECVMSG:
6114 case IORING_OP_SENDMSG: {
6115 struct io_async_msghdr *io = req->async_data;
6117 kfree(io->free_iov);
6120 case IORING_OP_SPLICE:
6122 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6123 io_put_file(req->splice.file_in);
6125 case IORING_OP_OPENAT:
6126 case IORING_OP_OPENAT2:
6127 if (req->open.filename)
6128 putname(req->open.filename);
6130 case IORING_OP_RENAMEAT:
6131 putname(req->rename.oldpath);
6132 putname(req->rename.newpath);
6134 case IORING_OP_UNLINKAT:
6135 putname(req->unlink.filename);
6139 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6140 kfree(req->apoll->double_poll);
6144 if (req->flags & REQ_F_INFLIGHT) {
6145 struct io_uring_task *tctx = req->task->io_uring;
6147 atomic_dec(&tctx->inflight_tracked);
6149 if (req->flags & REQ_F_CREDS)
6150 put_cred(req->creds);
6152 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6155 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6157 struct io_ring_ctx *ctx = req->ctx;
6158 const struct cred *creds = NULL;
6161 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6162 creds = override_creds(req->creds);
6164 switch (req->opcode) {
6166 ret = io_nop(req, issue_flags);
6168 case IORING_OP_READV:
6169 case IORING_OP_READ_FIXED:
6170 case IORING_OP_READ:
6171 ret = io_read(req, issue_flags);
6173 case IORING_OP_WRITEV:
6174 case IORING_OP_WRITE_FIXED:
6175 case IORING_OP_WRITE:
6176 ret = io_write(req, issue_flags);
6178 case IORING_OP_FSYNC:
6179 ret = io_fsync(req, issue_flags);
6181 case IORING_OP_POLL_ADD:
6182 ret = io_poll_add(req, issue_flags);
6184 case IORING_OP_POLL_REMOVE:
6185 ret = io_poll_update(req, issue_flags);
6187 case IORING_OP_SYNC_FILE_RANGE:
6188 ret = io_sync_file_range(req, issue_flags);
6190 case IORING_OP_SENDMSG:
6191 ret = io_sendmsg(req, issue_flags);
6193 case IORING_OP_SEND:
6194 ret = io_send(req, issue_flags);
6196 case IORING_OP_RECVMSG:
6197 ret = io_recvmsg(req, issue_flags);
6199 case IORING_OP_RECV:
6200 ret = io_recv(req, issue_flags);
6202 case IORING_OP_TIMEOUT:
6203 ret = io_timeout(req, issue_flags);
6205 case IORING_OP_TIMEOUT_REMOVE:
6206 ret = io_timeout_remove(req, issue_flags);
6208 case IORING_OP_ACCEPT:
6209 ret = io_accept(req, issue_flags);
6211 case IORING_OP_CONNECT:
6212 ret = io_connect(req, issue_flags);
6214 case IORING_OP_ASYNC_CANCEL:
6215 ret = io_async_cancel(req, issue_flags);
6217 case IORING_OP_FALLOCATE:
6218 ret = io_fallocate(req, issue_flags);
6220 case IORING_OP_OPENAT:
6221 ret = io_openat(req, issue_flags);
6223 case IORING_OP_CLOSE:
6224 ret = io_close(req, issue_flags);
6226 case IORING_OP_FILES_UPDATE:
6227 ret = io_files_update(req, issue_flags);
6229 case IORING_OP_STATX:
6230 ret = io_statx(req, issue_flags);
6232 case IORING_OP_FADVISE:
6233 ret = io_fadvise(req, issue_flags);
6235 case IORING_OP_MADVISE:
6236 ret = io_madvise(req, issue_flags);
6238 case IORING_OP_OPENAT2:
6239 ret = io_openat2(req, issue_flags);
6241 case IORING_OP_EPOLL_CTL:
6242 ret = io_epoll_ctl(req, issue_flags);
6244 case IORING_OP_SPLICE:
6245 ret = io_splice(req, issue_flags);
6247 case IORING_OP_PROVIDE_BUFFERS:
6248 ret = io_provide_buffers(req, issue_flags);
6250 case IORING_OP_REMOVE_BUFFERS:
6251 ret = io_remove_buffers(req, issue_flags);
6254 ret = io_tee(req, issue_flags);
6256 case IORING_OP_SHUTDOWN:
6257 ret = io_shutdown(req, issue_flags);
6259 case IORING_OP_RENAMEAT:
6260 ret = io_renameat(req, issue_flags);
6262 case IORING_OP_UNLINKAT:
6263 ret = io_unlinkat(req, issue_flags);
6271 revert_creds(creds);
6274 /* If the op doesn't have a file, we're not polling for it */
6275 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6276 io_iopoll_req_issued(req);
6281 static void io_wq_submit_work(struct io_wq_work *work)
6283 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6284 struct io_kiocb *timeout;
6287 timeout = io_prep_linked_timeout(req);
6289 io_queue_linked_timeout(timeout);
6291 if (work->flags & IO_WQ_WORK_CANCEL)
6296 ret = io_issue_sqe(req, 0);
6298 * We can get EAGAIN for polled IO even though we're
6299 * forcing a sync submission from here, since we can't
6300 * wait for request slots on the block side.
6308 /* avoid locking problems by failing it from a clean context */
6310 /* io-wq is going to take one down */
6312 io_req_task_queue_fail(req, ret);
6316 #define FFS_ASYNC_READ 0x1UL
6317 #define FFS_ASYNC_WRITE 0x2UL
6319 #define FFS_ISREG 0x4UL
6321 #define FFS_ISREG 0x0UL
6323 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6325 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6328 struct io_fixed_file *table_l2;
6330 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6331 return &table_l2[i & IORING_FILE_TABLE_MASK];
6334 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6337 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6339 return (struct file *) (slot->file_ptr & FFS_MASK);
6342 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6344 unsigned long file_ptr = (unsigned long) file;
6346 if (__io_file_supports_async(file, READ))
6347 file_ptr |= FFS_ASYNC_READ;
6348 if (__io_file_supports_async(file, WRITE))
6349 file_ptr |= FFS_ASYNC_WRITE;
6350 if (S_ISREG(file_inode(file)->i_mode))
6351 file_ptr |= FFS_ISREG;
6352 file_slot->file_ptr = file_ptr;
6355 static struct file *io_file_get(struct io_submit_state *state,
6356 struct io_kiocb *req, int fd, bool fixed)
6358 struct io_ring_ctx *ctx = req->ctx;
6362 unsigned long file_ptr;
6364 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6366 fd = array_index_nospec(fd, ctx->nr_user_files);
6367 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6368 file = (struct file *) (file_ptr & FFS_MASK);
6369 file_ptr &= ~FFS_MASK;
6370 /* mask in overlapping REQ_F and FFS bits */
6371 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6372 io_req_set_rsrc_node(req);
6374 trace_io_uring_file_get(ctx, fd);
6375 file = __io_file_get(state, fd);
6377 /* we don't allow fixed io_uring files */
6378 if (file && unlikely(file->f_op == &io_uring_fops))
6379 io_req_track_inflight(req);
6385 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6387 struct io_timeout_data *data = container_of(timer,
6388 struct io_timeout_data, timer);
6389 struct io_kiocb *prev, *req = data->req;
6390 struct io_ring_ctx *ctx = req->ctx;
6391 unsigned long flags;
6393 spin_lock_irqsave(&ctx->completion_lock, flags);
6394 prev = req->timeout.head;
6395 req->timeout.head = NULL;
6398 * We don't expect the list to be empty, that will only happen if we
6399 * race with the completion of the linked work.
6402 io_remove_next_linked(prev);
6403 if (!req_ref_inc_not_zero(prev))
6406 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6409 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6410 io_put_req_deferred(prev, 1);
6411 io_put_req_deferred(req, 1);
6413 io_req_complete_post(req, -ETIME, 0);
6415 return HRTIMER_NORESTART;
6418 static void io_queue_linked_timeout(struct io_kiocb *req)
6420 struct io_ring_ctx *ctx = req->ctx;
6422 spin_lock_irq(&ctx->completion_lock);
6424 * If the back reference is NULL, then our linked request finished
6425 * before we got a chance to setup the timer
6427 if (req->timeout.head) {
6428 struct io_timeout_data *data = req->async_data;
6430 data->timer.function = io_link_timeout_fn;
6431 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6434 spin_unlock_irq(&ctx->completion_lock);
6435 /* drop submission reference */
6439 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6441 struct io_kiocb *nxt = req->link;
6443 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6444 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6447 nxt->timeout.head = req;
6448 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6449 req->flags |= REQ_F_LINK_TIMEOUT;
6453 static void __io_queue_sqe(struct io_kiocb *req)
6455 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6459 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6462 * We async punt it if the file wasn't marked NOWAIT, or if the file
6463 * doesn't support non-blocking read/write attempts
6466 /* drop submission reference */
6467 if (req->flags & REQ_F_COMPLETE_INLINE) {
6468 struct io_ring_ctx *ctx = req->ctx;
6469 struct io_comp_state *cs = &ctx->submit_state.comp;
6471 cs->reqs[cs->nr++] = req;
6472 if (cs->nr == ARRAY_SIZE(cs->reqs))
6473 io_submit_flush_completions(ctx);
6477 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6478 switch (io_arm_poll_handler(req)) {
6479 case IO_APOLL_READY:
6481 case IO_APOLL_ABORTED:
6483 * Queued up for async execution, worker will release
6484 * submit reference when the iocb is actually submitted.
6486 io_queue_async_work(req);
6490 io_req_complete_failed(req, ret);
6493 io_queue_linked_timeout(linked_timeout);
6496 static inline void io_queue_sqe(struct io_kiocb *req)
6498 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6501 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6502 __io_queue_sqe(req);
6504 int ret = io_req_prep_async(req);
6507 io_req_complete_failed(req, ret);
6509 io_queue_async_work(req);
6514 * Check SQE restrictions (opcode and flags).
6516 * Returns 'true' if SQE is allowed, 'false' otherwise.
6518 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6519 struct io_kiocb *req,
6520 unsigned int sqe_flags)
6522 if (likely(!ctx->restricted))
6525 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6528 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6529 ctx->restrictions.sqe_flags_required)
6532 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6533 ctx->restrictions.sqe_flags_required))
6539 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6540 const struct io_uring_sqe *sqe)
6542 struct io_submit_state *state;
6543 unsigned int sqe_flags;
6544 int personality, ret = 0;
6546 req->opcode = READ_ONCE(sqe->opcode);
6547 /* same numerical values with corresponding REQ_F_*, safe to copy */
6548 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6549 req->user_data = READ_ONCE(sqe->user_data);
6551 req->fixed_rsrc_refs = NULL;
6552 /* one is dropped after submission, the other at completion */
6553 atomic_set(&req->refs, 2);
6554 req->task = current;
6556 /* enforce forwards compatibility on users */
6557 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6559 if (unlikely(req->opcode >= IORING_OP_LAST))
6561 if (!io_check_restriction(ctx, req, sqe_flags))
6564 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6565 !io_op_defs[req->opcode].buffer_select)
6567 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6568 ctx->drain_active = true;
6570 personality = READ_ONCE(sqe->personality);
6572 req->creds = xa_load(&ctx->personalities, personality);
6575 get_cred(req->creds);
6576 req->flags |= REQ_F_CREDS;
6578 state = &ctx->submit_state;
6581 * Plug now if we have more than 1 IO left after this, and the target
6582 * is potentially a read/write to block based storage.
6584 if (!state->plug_started && state->ios_left > 1 &&
6585 io_op_defs[req->opcode].plug) {
6586 blk_start_plug(&state->plug);
6587 state->plug_started = true;
6590 if (io_op_defs[req->opcode].needs_file) {
6591 bool fixed = req->flags & REQ_F_FIXED_FILE;
6593 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6594 if (unlikely(!req->file))
6602 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6603 const struct io_uring_sqe *sqe)
6605 struct io_submit_link *link = &ctx->submit_state.link;
6608 ret = io_init_req(ctx, req, sqe);
6609 if (unlikely(ret)) {
6612 /* fail even hard links since we don't submit */
6613 req_set_fail(link->head);
6614 io_req_complete_failed(link->head, -ECANCELED);
6617 io_req_complete_failed(req, ret);
6621 ret = io_req_prep(req, sqe);
6625 /* don't need @sqe from now on */
6626 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6628 ctx->flags & IORING_SETUP_SQPOLL);
6631 * If we already have a head request, queue this one for async
6632 * submittal once the head completes. If we don't have a head but
6633 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6634 * submitted sync once the chain is complete. If none of those
6635 * conditions are true (normal request), then just queue it.
6638 struct io_kiocb *head = link->head;
6640 ret = io_req_prep_async(req);
6643 trace_io_uring_link(ctx, req, head);
6644 link->last->link = req;
6647 /* last request of a link, enqueue the link */
6648 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6653 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6665 * Batched submission is done, ensure local IO is flushed out.
6667 static void io_submit_state_end(struct io_submit_state *state,
6668 struct io_ring_ctx *ctx)
6670 if (state->link.head)
6671 io_queue_sqe(state->link.head);
6673 io_submit_flush_completions(ctx);
6674 if (state->plug_started)
6675 blk_finish_plug(&state->plug);
6676 io_state_file_put(state);
6680 * Start submission side cache.
6682 static void io_submit_state_start(struct io_submit_state *state,
6683 unsigned int max_ios)
6685 state->plug_started = false;
6686 state->ios_left = max_ios;
6687 /* set only head, no need to init link_last in advance */
6688 state->link.head = NULL;
6691 static void io_commit_sqring(struct io_ring_ctx *ctx)
6693 struct io_rings *rings = ctx->rings;
6696 * Ensure any loads from the SQEs are done at this point,
6697 * since once we write the new head, the application could
6698 * write new data to them.
6700 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6704 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6705 * that is mapped by userspace. This means that care needs to be taken to
6706 * ensure that reads are stable, as we cannot rely on userspace always
6707 * being a good citizen. If members of the sqe are validated and then later
6708 * used, it's important that those reads are done through READ_ONCE() to
6709 * prevent a re-load down the line.
6711 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6713 unsigned head, mask = ctx->sq_entries - 1;
6714 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6717 * The cached sq head (or cq tail) serves two purposes:
6719 * 1) allows us to batch the cost of updating the user visible
6721 * 2) allows the kernel side to track the head on its own, even
6722 * though the application is the one updating it.
6724 head = READ_ONCE(ctx->sq_array[sq_idx]);
6725 if (likely(head < ctx->sq_entries))
6726 return &ctx->sq_sqes[head];
6728 /* drop invalid entries */
6730 WRITE_ONCE(ctx->rings->sq_dropped,
6731 READ_ONCE(ctx->rings->sq_dropped) + 1);
6735 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6737 struct io_uring_task *tctx;
6740 /* make sure SQ entry isn't read before tail */
6741 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6742 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6745 tctx = current->io_uring;
6746 tctx->cached_refs -= nr;
6747 if (unlikely(tctx->cached_refs < 0)) {
6748 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6750 percpu_counter_add(&tctx->inflight, refill);
6751 refcount_add(refill, ¤t->usage);
6752 tctx->cached_refs += refill;
6754 io_submit_state_start(&ctx->submit_state, nr);
6756 while (submitted < nr) {
6757 const struct io_uring_sqe *sqe;
6758 struct io_kiocb *req;
6760 req = io_alloc_req(ctx);
6761 if (unlikely(!req)) {
6763 submitted = -EAGAIN;
6766 sqe = io_get_sqe(ctx);
6767 if (unlikely(!sqe)) {
6768 kmem_cache_free(req_cachep, req);
6771 /* will complete beyond this point, count as submitted */
6773 if (io_submit_sqe(ctx, req, sqe))
6777 if (unlikely(submitted != nr)) {
6778 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6779 int unused = nr - ref_used;
6781 current->io_uring->cached_refs += unused;
6782 percpu_ref_put_many(&ctx->refs, unused);
6785 io_submit_state_end(&ctx->submit_state, ctx);
6786 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6787 io_commit_sqring(ctx);
6792 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6794 return READ_ONCE(sqd->state);
6797 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6799 /* Tell userspace we may need a wakeup call */
6800 spin_lock_irq(&ctx->completion_lock);
6801 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6802 spin_unlock_irq(&ctx->completion_lock);
6805 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6807 spin_lock_irq(&ctx->completion_lock);
6808 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6809 spin_unlock_irq(&ctx->completion_lock);
6812 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6814 unsigned int to_submit;
6817 to_submit = io_sqring_entries(ctx);
6818 /* if we're handling multiple rings, cap submit size for fairness */
6819 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6820 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6822 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6823 unsigned nr_events = 0;
6824 const struct cred *creds = NULL;
6826 if (ctx->sq_creds != current_cred())
6827 creds = override_creds(ctx->sq_creds);
6829 mutex_lock(&ctx->uring_lock);
6830 if (!list_empty(&ctx->iopoll_list))
6831 io_do_iopoll(ctx, &nr_events, 0, true);
6834 * Don't submit if refs are dying, good for io_uring_register(),
6835 * but also it is relied upon by io_ring_exit_work()
6837 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6838 !(ctx->flags & IORING_SETUP_R_DISABLED))
6839 ret = io_submit_sqes(ctx, to_submit);
6840 mutex_unlock(&ctx->uring_lock);
6842 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6843 wake_up(&ctx->sqo_sq_wait);
6845 revert_creds(creds);
6851 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6853 struct io_ring_ctx *ctx;
6854 unsigned sq_thread_idle = 0;
6856 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6857 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6858 sqd->sq_thread_idle = sq_thread_idle;
6861 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6863 bool did_sig = false;
6864 struct ksignal ksig;
6866 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6867 signal_pending(current)) {
6868 mutex_unlock(&sqd->lock);
6869 if (signal_pending(current))
6870 did_sig = get_signal(&ksig);
6872 mutex_lock(&sqd->lock);
6874 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6877 static int io_sq_thread(void *data)
6879 struct io_sq_data *sqd = data;
6880 struct io_ring_ctx *ctx;
6881 unsigned long timeout = 0;
6882 char buf[TASK_COMM_LEN];
6885 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6886 set_task_comm(current, buf);
6888 if (sqd->sq_cpu != -1)
6889 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6891 set_cpus_allowed_ptr(current, cpu_online_mask);
6892 current->flags |= PF_NO_SETAFFINITY;
6894 mutex_lock(&sqd->lock);
6896 bool cap_entries, sqt_spin = false;
6898 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6899 if (io_sqd_handle_event(sqd))
6901 timeout = jiffies + sqd->sq_thread_idle;
6904 cap_entries = !list_is_singular(&sqd->ctx_list);
6905 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6906 int ret = __io_sq_thread(ctx, cap_entries);
6908 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6911 if (io_run_task_work())
6914 if (sqt_spin || !time_after(jiffies, timeout)) {
6917 timeout = jiffies + sqd->sq_thread_idle;
6921 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6922 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6923 bool needs_sched = true;
6925 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6926 io_ring_set_wakeup_flag(ctx);
6928 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6929 !list_empty_careful(&ctx->iopoll_list)) {
6930 needs_sched = false;
6933 if (io_sqring_entries(ctx)) {
6934 needs_sched = false;
6940 mutex_unlock(&sqd->lock);
6942 mutex_lock(&sqd->lock);
6944 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6945 io_ring_clear_wakeup_flag(ctx);
6948 finish_wait(&sqd->wait, &wait);
6949 timeout = jiffies + sqd->sq_thread_idle;
6952 io_uring_cancel_generic(true, sqd);
6954 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6955 io_ring_set_wakeup_flag(ctx);
6957 mutex_unlock(&sqd->lock);
6959 complete(&sqd->exited);
6963 struct io_wait_queue {
6964 struct wait_queue_entry wq;
6965 struct io_ring_ctx *ctx;
6967 unsigned nr_timeouts;
6970 static inline bool io_should_wake(struct io_wait_queue *iowq)
6972 struct io_ring_ctx *ctx = iowq->ctx;
6975 * Wake up if we have enough events, or if a timeout occurred since we
6976 * started waiting. For timeouts, we always want to return to userspace,
6977 * regardless of event count.
6979 return io_cqring_events(ctx) >= iowq->to_wait ||
6980 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6983 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6984 int wake_flags, void *key)
6986 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6990 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6991 * the task, and the next invocation will do it.
6993 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
6994 return autoremove_wake_function(curr, mode, wake_flags, key);
6998 static int io_run_task_work_sig(void)
7000 if (io_run_task_work())
7002 if (!signal_pending(current))
7004 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7005 return -ERESTARTSYS;
7009 /* when returns >0, the caller should retry */
7010 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7011 struct io_wait_queue *iowq,
7012 signed long *timeout)
7016 /* make sure we run task_work before checking for signals */
7017 ret = io_run_task_work_sig();
7018 if (ret || io_should_wake(iowq))
7020 /* let the caller flush overflows, retry */
7021 if (test_bit(0, &ctx->check_cq_overflow))
7024 *timeout = schedule_timeout(*timeout);
7025 return !*timeout ? -ETIME : 1;
7029 * Wait until events become available, if we don't already have some. The
7030 * application must reap them itself, as they reside on the shared cq ring.
7032 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7033 const sigset_t __user *sig, size_t sigsz,
7034 struct __kernel_timespec __user *uts)
7036 struct io_wait_queue iowq = {
7039 .func = io_wake_function,
7040 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7043 .to_wait = min_events,
7045 struct io_rings *rings = ctx->rings;
7046 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7050 io_cqring_overflow_flush(ctx, false);
7051 if (io_cqring_events(ctx) >= min_events)
7053 if (!io_run_task_work())
7058 #ifdef CONFIG_COMPAT
7059 if (in_compat_syscall())
7060 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7064 ret = set_user_sigmask(sig, sigsz);
7071 struct timespec64 ts;
7073 if (get_timespec64(&ts, uts))
7075 timeout = timespec64_to_jiffies(&ts);
7078 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7079 trace_io_uring_cqring_wait(ctx, min_events);
7081 /* if we can't even flush overflow, don't wait for more */
7082 if (!io_cqring_overflow_flush(ctx, false)) {
7086 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7087 TASK_INTERRUPTIBLE);
7088 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7089 finish_wait(&ctx->cq_wait, &iowq.wq);
7093 restore_saved_sigmask_unless(ret == -EINTR);
7095 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7098 static void io_free_page_table(void **table, size_t size)
7100 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7102 for (i = 0; i < nr_tables; i++)
7107 static void **io_alloc_page_table(size_t size)
7109 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7110 size_t init_size = size;
7113 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7117 for (i = 0; i < nr_tables; i++) {
7118 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7120 table[i] = kzalloc(this_size, GFP_KERNEL);
7122 io_free_page_table(table, init_size);
7130 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7132 spin_lock_bh(&ctx->rsrc_ref_lock);
7135 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7137 spin_unlock_bh(&ctx->rsrc_ref_lock);
7140 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7142 percpu_ref_exit(&ref_node->refs);
7146 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7147 struct io_rsrc_data *data_to_kill)
7149 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7150 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7153 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7155 rsrc_node->rsrc_data = data_to_kill;
7156 io_rsrc_ref_lock(ctx);
7157 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7158 io_rsrc_ref_unlock(ctx);
7160 atomic_inc(&data_to_kill->refs);
7161 percpu_ref_kill(&rsrc_node->refs);
7162 ctx->rsrc_node = NULL;
7165 if (!ctx->rsrc_node) {
7166 ctx->rsrc_node = ctx->rsrc_backup_node;
7167 ctx->rsrc_backup_node = NULL;
7171 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7173 if (ctx->rsrc_backup_node)
7175 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7176 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7179 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7183 /* As we may drop ->uring_lock, other task may have started quiesce */
7187 data->quiesce = true;
7189 ret = io_rsrc_node_switch_start(ctx);
7192 io_rsrc_node_switch(ctx, data);
7194 /* kill initial ref, already quiesced if zero */
7195 if (atomic_dec_and_test(&data->refs))
7197 flush_delayed_work(&ctx->rsrc_put_work);
7198 ret = wait_for_completion_interruptible(&data->done);
7202 atomic_inc(&data->refs);
7203 /* wait for all works potentially completing data->done */
7204 flush_delayed_work(&ctx->rsrc_put_work);
7205 reinit_completion(&data->done);
7207 mutex_unlock(&ctx->uring_lock);
7208 ret = io_run_task_work_sig();
7209 mutex_lock(&ctx->uring_lock);
7211 data->quiesce = false;
7216 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7218 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7219 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7221 return &data->tags[table_idx][off];
7224 static void io_rsrc_data_free(struct io_rsrc_data *data)
7226 size_t size = data->nr * sizeof(data->tags[0][0]);
7229 io_free_page_table((void **)data->tags, size);
7233 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7234 u64 __user *utags, unsigned nr,
7235 struct io_rsrc_data **pdata)
7237 struct io_rsrc_data *data;
7241 data = kzalloc(sizeof(*data), GFP_KERNEL);
7244 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7252 data->do_put = do_put;
7255 for (i = 0; i < nr; i++) {
7256 u64 *tag_slot = io_get_tag_slot(data, i);
7258 if (copy_from_user(tag_slot, &utags[i],
7264 atomic_set(&data->refs, 1);
7265 init_completion(&data->done);
7269 io_rsrc_data_free(data);
7273 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7275 size_t size = nr_files * sizeof(struct io_fixed_file);
7277 table->files = (struct io_fixed_file **)io_alloc_page_table(size);
7278 return !!table->files;
7281 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7283 size_t size = nr_files * sizeof(struct io_fixed_file);
7285 io_free_page_table((void **)table->files, size);
7286 table->files = NULL;
7289 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7291 #if defined(CONFIG_UNIX)
7292 if (ctx->ring_sock) {
7293 struct sock *sock = ctx->ring_sock->sk;
7294 struct sk_buff *skb;
7296 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7302 for (i = 0; i < ctx->nr_user_files; i++) {
7305 file = io_file_from_index(ctx, i);
7310 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7311 io_rsrc_data_free(ctx->file_data);
7312 ctx->file_data = NULL;
7313 ctx->nr_user_files = 0;
7316 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7320 if (!ctx->file_data)
7322 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7324 __io_sqe_files_unregister(ctx);
7328 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7329 __releases(&sqd->lock)
7331 WARN_ON_ONCE(sqd->thread == current);
7334 * Do the dance but not conditional clear_bit() because it'd race with
7335 * other threads incrementing park_pending and setting the bit.
7337 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7338 if (atomic_dec_return(&sqd->park_pending))
7339 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7340 mutex_unlock(&sqd->lock);
7343 static void io_sq_thread_park(struct io_sq_data *sqd)
7344 __acquires(&sqd->lock)
7346 WARN_ON_ONCE(sqd->thread == current);
7348 atomic_inc(&sqd->park_pending);
7349 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7350 mutex_lock(&sqd->lock);
7352 wake_up_process(sqd->thread);
7355 static void io_sq_thread_stop(struct io_sq_data *sqd)
7357 WARN_ON_ONCE(sqd->thread == current);
7358 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7360 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7361 mutex_lock(&sqd->lock);
7363 wake_up_process(sqd->thread);
7364 mutex_unlock(&sqd->lock);
7365 wait_for_completion(&sqd->exited);
7368 static void io_put_sq_data(struct io_sq_data *sqd)
7370 if (refcount_dec_and_test(&sqd->refs)) {
7371 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7373 io_sq_thread_stop(sqd);
7378 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7380 struct io_sq_data *sqd = ctx->sq_data;
7383 io_sq_thread_park(sqd);
7384 list_del_init(&ctx->sqd_list);
7385 io_sqd_update_thread_idle(sqd);
7386 io_sq_thread_unpark(sqd);
7388 io_put_sq_data(sqd);
7389 ctx->sq_data = NULL;
7393 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7395 struct io_ring_ctx *ctx_attach;
7396 struct io_sq_data *sqd;
7399 f = fdget(p->wq_fd);
7401 return ERR_PTR(-ENXIO);
7402 if (f.file->f_op != &io_uring_fops) {
7404 return ERR_PTR(-EINVAL);
7407 ctx_attach = f.file->private_data;
7408 sqd = ctx_attach->sq_data;
7411 return ERR_PTR(-EINVAL);
7413 if (sqd->task_tgid != current->tgid) {
7415 return ERR_PTR(-EPERM);
7418 refcount_inc(&sqd->refs);
7423 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7426 struct io_sq_data *sqd;
7429 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7430 sqd = io_attach_sq_data(p);
7435 /* fall through for EPERM case, setup new sqd/task */
7436 if (PTR_ERR(sqd) != -EPERM)
7440 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7442 return ERR_PTR(-ENOMEM);
7444 atomic_set(&sqd->park_pending, 0);
7445 refcount_set(&sqd->refs, 1);
7446 INIT_LIST_HEAD(&sqd->ctx_list);
7447 mutex_init(&sqd->lock);
7448 init_waitqueue_head(&sqd->wait);
7449 init_completion(&sqd->exited);
7453 #if defined(CONFIG_UNIX)
7455 * Ensure the UNIX gc is aware of our file set, so we are certain that
7456 * the io_uring can be safely unregistered on process exit, even if we have
7457 * loops in the file referencing.
7459 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7461 struct sock *sk = ctx->ring_sock->sk;
7462 struct scm_fp_list *fpl;
7463 struct sk_buff *skb;
7466 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7470 skb = alloc_skb(0, GFP_KERNEL);
7479 fpl->user = get_uid(current_user());
7480 for (i = 0; i < nr; i++) {
7481 struct file *file = io_file_from_index(ctx, i + offset);
7485 fpl->fp[nr_files] = get_file(file);
7486 unix_inflight(fpl->user, fpl->fp[nr_files]);
7491 fpl->max = SCM_MAX_FD;
7492 fpl->count = nr_files;
7493 UNIXCB(skb).fp = fpl;
7494 skb->destructor = unix_destruct_scm;
7495 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7496 skb_queue_head(&sk->sk_receive_queue, skb);
7498 for (i = 0; i < nr_files; i++)
7509 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7510 * causes regular reference counting to break down. We rely on the UNIX
7511 * garbage collection to take care of this problem for us.
7513 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7515 unsigned left, total;
7519 left = ctx->nr_user_files;
7521 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7523 ret = __io_sqe_files_scm(ctx, this_files, total);
7527 total += this_files;
7533 while (total < ctx->nr_user_files) {
7534 struct file *file = io_file_from_index(ctx, total);
7544 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7550 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7552 struct file *file = prsrc->file;
7553 #if defined(CONFIG_UNIX)
7554 struct sock *sock = ctx->ring_sock->sk;
7555 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7556 struct sk_buff *skb;
7559 __skb_queue_head_init(&list);
7562 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7563 * remove this entry and rearrange the file array.
7565 skb = skb_dequeue(head);
7567 struct scm_fp_list *fp;
7569 fp = UNIXCB(skb).fp;
7570 for (i = 0; i < fp->count; i++) {
7573 if (fp->fp[i] != file)
7576 unix_notinflight(fp->user, fp->fp[i]);
7577 left = fp->count - 1 - i;
7579 memmove(&fp->fp[i], &fp->fp[i + 1],
7580 left * sizeof(struct file *));
7587 __skb_queue_tail(&list, skb);
7597 __skb_queue_tail(&list, skb);
7599 skb = skb_dequeue(head);
7602 if (skb_peek(&list)) {
7603 spin_lock_irq(&head->lock);
7604 while ((skb = __skb_dequeue(&list)) != NULL)
7605 __skb_queue_tail(head, skb);
7606 spin_unlock_irq(&head->lock);
7613 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7615 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7616 struct io_ring_ctx *ctx = rsrc_data->ctx;
7617 struct io_rsrc_put *prsrc, *tmp;
7619 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7620 list_del(&prsrc->list);
7623 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7625 io_ring_submit_lock(ctx, lock_ring);
7626 spin_lock_irq(&ctx->completion_lock);
7627 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7629 io_commit_cqring(ctx);
7630 spin_unlock_irq(&ctx->completion_lock);
7631 io_cqring_ev_posted(ctx);
7632 io_ring_submit_unlock(ctx, lock_ring);
7635 rsrc_data->do_put(ctx, prsrc);
7639 io_rsrc_node_destroy(ref_node);
7640 if (atomic_dec_and_test(&rsrc_data->refs))
7641 complete(&rsrc_data->done);
7644 static void io_rsrc_put_work(struct work_struct *work)
7646 struct io_ring_ctx *ctx;
7647 struct llist_node *node;
7649 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7650 node = llist_del_all(&ctx->rsrc_put_llist);
7653 struct io_rsrc_node *ref_node;
7654 struct llist_node *next = node->next;
7656 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7657 __io_rsrc_put_work(ref_node);
7662 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7664 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7665 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7666 bool first_add = false;
7668 io_rsrc_ref_lock(ctx);
7671 while (!list_empty(&ctx->rsrc_ref_list)) {
7672 node = list_first_entry(&ctx->rsrc_ref_list,
7673 struct io_rsrc_node, node);
7674 /* recycle ref nodes in order */
7677 list_del(&node->node);
7678 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7680 io_rsrc_ref_unlock(ctx);
7683 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7686 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7688 struct io_rsrc_node *ref_node;
7690 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7694 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7699 INIT_LIST_HEAD(&ref_node->node);
7700 INIT_LIST_HEAD(&ref_node->rsrc_list);
7701 ref_node->done = false;
7705 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7706 unsigned nr_args, u64 __user *tags)
7708 __s32 __user *fds = (__s32 __user *) arg;
7717 if (nr_args > IORING_MAX_FIXED_FILES)
7719 ret = io_rsrc_node_switch_start(ctx);
7722 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7728 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7731 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7732 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7736 /* allow sparse sets */
7739 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7746 if (unlikely(!file))
7750 * Don't allow io_uring instances to be registered. If UNIX
7751 * isn't enabled, then this causes a reference cycle and this
7752 * instance can never get freed. If UNIX is enabled we'll
7753 * handle it just fine, but there's still no point in allowing
7754 * a ring fd as it doesn't support regular read/write anyway.
7756 if (file->f_op == &io_uring_fops) {
7760 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7763 ret = io_sqe_files_scm(ctx);
7765 __io_sqe_files_unregister(ctx);
7769 io_rsrc_node_switch(ctx, NULL);
7772 for (i = 0; i < ctx->nr_user_files; i++) {
7773 file = io_file_from_index(ctx, i);
7777 io_free_file_tables(&ctx->file_table, nr_args);
7778 ctx->nr_user_files = 0;
7780 io_rsrc_data_free(ctx->file_data);
7781 ctx->file_data = NULL;
7785 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7788 #if defined(CONFIG_UNIX)
7789 struct sock *sock = ctx->ring_sock->sk;
7790 struct sk_buff_head *head = &sock->sk_receive_queue;
7791 struct sk_buff *skb;
7794 * See if we can merge this file into an existing skb SCM_RIGHTS
7795 * file set. If there's no room, fall back to allocating a new skb
7796 * and filling it in.
7798 spin_lock_irq(&head->lock);
7799 skb = skb_peek(head);
7801 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7803 if (fpl->count < SCM_MAX_FD) {
7804 __skb_unlink(skb, head);
7805 spin_unlock_irq(&head->lock);
7806 fpl->fp[fpl->count] = get_file(file);
7807 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7809 spin_lock_irq(&head->lock);
7810 __skb_queue_head(head, skb);
7815 spin_unlock_irq(&head->lock);
7822 return __io_sqe_files_scm(ctx, 1, index);
7828 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7829 struct io_rsrc_node *node, void *rsrc)
7831 struct io_rsrc_put *prsrc;
7833 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7837 prsrc->tag = *io_get_tag_slot(data, idx);
7839 list_add(&prsrc->list, &node->rsrc_list);
7843 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7844 struct io_uring_rsrc_update2 *up,
7847 u64 __user *tags = u64_to_user_ptr(up->tags);
7848 __s32 __user *fds = u64_to_user_ptr(up->data);
7849 struct io_rsrc_data *data = ctx->file_data;
7850 struct io_fixed_file *file_slot;
7854 bool needs_switch = false;
7856 if (!ctx->file_data)
7858 if (up->offset + nr_args > ctx->nr_user_files)
7861 for (done = 0; done < nr_args; done++) {
7864 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7865 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7869 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7873 if (fd == IORING_REGISTER_FILES_SKIP)
7876 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7877 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7879 if (file_slot->file_ptr) {
7880 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7881 err = io_queue_rsrc_removal(data, up->offset + done,
7882 ctx->rsrc_node, file);
7885 file_slot->file_ptr = 0;
7886 needs_switch = true;
7895 * Don't allow io_uring instances to be registered. If
7896 * UNIX isn't enabled, then this causes a reference
7897 * cycle and this instance can never get freed. If UNIX
7898 * is enabled we'll handle it just fine, but there's
7899 * still no point in allowing a ring fd as it doesn't
7900 * support regular read/write anyway.
7902 if (file->f_op == &io_uring_fops) {
7907 *io_get_tag_slot(data, up->offset + done) = tag;
7908 io_fixed_file_set(file_slot, file);
7909 err = io_sqe_file_register(ctx, file, i);
7911 file_slot->file_ptr = 0;
7919 io_rsrc_node_switch(ctx, data);
7920 return done ? done : err;
7923 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7925 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7927 req = io_put_req_find_next(req);
7928 return req ? &req->work : NULL;
7931 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7932 struct task_struct *task)
7934 struct io_wq_hash *hash;
7935 struct io_wq_data data;
7936 unsigned int concurrency;
7938 mutex_lock(&ctx->uring_lock);
7939 hash = ctx->hash_map;
7941 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7943 mutex_unlock(&ctx->uring_lock);
7944 return ERR_PTR(-ENOMEM);
7946 refcount_set(&hash->refs, 1);
7947 init_waitqueue_head(&hash->wait);
7948 ctx->hash_map = hash;
7950 mutex_unlock(&ctx->uring_lock);
7954 data.free_work = io_free_work;
7955 data.do_work = io_wq_submit_work;
7957 /* Do QD, or 4 * CPUS, whatever is smallest */
7958 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7960 return io_wq_create(concurrency, &data);
7963 static int io_uring_alloc_task_context(struct task_struct *task,
7964 struct io_ring_ctx *ctx)
7966 struct io_uring_task *tctx;
7969 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7970 if (unlikely(!tctx))
7973 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7974 if (unlikely(ret)) {
7979 tctx->io_wq = io_init_wq_offload(ctx, task);
7980 if (IS_ERR(tctx->io_wq)) {
7981 ret = PTR_ERR(tctx->io_wq);
7982 percpu_counter_destroy(&tctx->inflight);
7988 init_waitqueue_head(&tctx->wait);
7989 atomic_set(&tctx->in_idle, 0);
7990 atomic_set(&tctx->inflight_tracked, 0);
7991 task->io_uring = tctx;
7992 spin_lock_init(&tctx->task_lock);
7993 INIT_WQ_LIST(&tctx->task_list);
7994 init_task_work(&tctx->task_work, tctx_task_work);
7998 void __io_uring_free(struct task_struct *tsk)
8000 struct io_uring_task *tctx = tsk->io_uring;
8002 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8003 WARN_ON_ONCE(tctx->io_wq);
8004 WARN_ON_ONCE(tctx->cached_refs);
8006 percpu_counter_destroy(&tctx->inflight);
8008 tsk->io_uring = NULL;
8011 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8012 struct io_uring_params *p)
8016 /* Retain compatibility with failing for an invalid attach attempt */
8017 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8018 IORING_SETUP_ATTACH_WQ) {
8021 f = fdget(p->wq_fd);
8024 if (f.file->f_op != &io_uring_fops) {
8030 if (ctx->flags & IORING_SETUP_SQPOLL) {
8031 struct task_struct *tsk;
8032 struct io_sq_data *sqd;
8035 sqd = io_get_sq_data(p, &attached);
8041 ctx->sq_creds = get_current_cred();
8043 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8044 if (!ctx->sq_thread_idle)
8045 ctx->sq_thread_idle = HZ;
8047 io_sq_thread_park(sqd);
8048 list_add(&ctx->sqd_list, &sqd->ctx_list);
8049 io_sqd_update_thread_idle(sqd);
8050 /* don't attach to a dying SQPOLL thread, would be racy */
8051 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8052 io_sq_thread_unpark(sqd);
8059 if (p->flags & IORING_SETUP_SQ_AFF) {
8060 int cpu = p->sq_thread_cpu;
8063 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8070 sqd->task_pid = current->pid;
8071 sqd->task_tgid = current->tgid;
8072 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8079 ret = io_uring_alloc_task_context(tsk, ctx);
8080 wake_up_new_task(tsk);
8083 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8084 /* Can't have SQ_AFF without SQPOLL */
8091 complete(&ctx->sq_data->exited);
8093 io_sq_thread_finish(ctx);
8097 static inline void __io_unaccount_mem(struct user_struct *user,
8098 unsigned long nr_pages)
8100 atomic_long_sub(nr_pages, &user->locked_vm);
8103 static inline int __io_account_mem(struct user_struct *user,
8104 unsigned long nr_pages)
8106 unsigned long page_limit, cur_pages, new_pages;
8108 /* Don't allow more pages than we can safely lock */
8109 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8112 cur_pages = atomic_long_read(&user->locked_vm);
8113 new_pages = cur_pages + nr_pages;
8114 if (new_pages > page_limit)
8116 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8117 new_pages) != cur_pages);
8122 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8125 __io_unaccount_mem(ctx->user, nr_pages);
8127 if (ctx->mm_account)
8128 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8131 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8136 ret = __io_account_mem(ctx->user, nr_pages);
8141 if (ctx->mm_account)
8142 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8147 static void io_mem_free(void *ptr)
8154 page = virt_to_head_page(ptr);
8155 if (put_page_testzero(page))
8156 free_compound_page(page);
8159 static void *io_mem_alloc(size_t size)
8161 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8162 __GFP_NORETRY | __GFP_ACCOUNT;
8164 return (void *) __get_free_pages(gfp_flags, get_order(size));
8167 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8170 struct io_rings *rings;
8171 size_t off, sq_array_size;
8173 off = struct_size(rings, cqes, cq_entries);
8174 if (off == SIZE_MAX)
8178 off = ALIGN(off, SMP_CACHE_BYTES);
8186 sq_array_size = array_size(sizeof(u32), sq_entries);
8187 if (sq_array_size == SIZE_MAX)
8190 if (check_add_overflow(off, sq_array_size, &off))
8196 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8198 struct io_mapped_ubuf *imu = *slot;
8201 if (imu != ctx->dummy_ubuf) {
8202 for (i = 0; i < imu->nr_bvecs; i++)
8203 unpin_user_page(imu->bvec[i].bv_page);
8204 if (imu->acct_pages)
8205 io_unaccount_mem(ctx, imu->acct_pages);
8211 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8213 io_buffer_unmap(ctx, &prsrc->buf);
8217 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8221 for (i = 0; i < ctx->nr_user_bufs; i++)
8222 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8223 kfree(ctx->user_bufs);
8224 io_rsrc_data_free(ctx->buf_data);
8225 ctx->user_bufs = NULL;
8226 ctx->buf_data = NULL;
8227 ctx->nr_user_bufs = 0;
8230 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8237 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8239 __io_sqe_buffers_unregister(ctx);
8243 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8244 void __user *arg, unsigned index)
8246 struct iovec __user *src;
8248 #ifdef CONFIG_COMPAT
8250 struct compat_iovec __user *ciovs;
8251 struct compat_iovec ciov;
8253 ciovs = (struct compat_iovec __user *) arg;
8254 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8257 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8258 dst->iov_len = ciov.iov_len;
8262 src = (struct iovec __user *) arg;
8263 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8269 * Not super efficient, but this is just a registration time. And we do cache
8270 * the last compound head, so generally we'll only do a full search if we don't
8273 * We check if the given compound head page has already been accounted, to
8274 * avoid double accounting it. This allows us to account the full size of the
8275 * page, not just the constituent pages of a huge page.
8277 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8278 int nr_pages, struct page *hpage)
8282 /* check current page array */
8283 for (i = 0; i < nr_pages; i++) {
8284 if (!PageCompound(pages[i]))
8286 if (compound_head(pages[i]) == hpage)
8290 /* check previously registered pages */
8291 for (i = 0; i < ctx->nr_user_bufs; i++) {
8292 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8294 for (j = 0; j < imu->nr_bvecs; j++) {
8295 if (!PageCompound(imu->bvec[j].bv_page))
8297 if (compound_head(imu->bvec[j].bv_page) == hpage)
8305 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8306 int nr_pages, struct io_mapped_ubuf *imu,
8307 struct page **last_hpage)
8311 imu->acct_pages = 0;
8312 for (i = 0; i < nr_pages; i++) {
8313 if (!PageCompound(pages[i])) {
8318 hpage = compound_head(pages[i]);
8319 if (hpage == *last_hpage)
8321 *last_hpage = hpage;
8322 if (headpage_already_acct(ctx, pages, i, hpage))
8324 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8328 if (!imu->acct_pages)
8331 ret = io_account_mem(ctx, imu->acct_pages);
8333 imu->acct_pages = 0;
8337 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8338 struct io_mapped_ubuf **pimu,
8339 struct page **last_hpage)
8341 struct io_mapped_ubuf *imu = NULL;
8342 struct vm_area_struct **vmas = NULL;
8343 struct page **pages = NULL;
8344 unsigned long off, start, end, ubuf;
8346 int ret, pret, nr_pages, i;
8348 if (!iov->iov_base) {
8349 *pimu = ctx->dummy_ubuf;
8353 ubuf = (unsigned long) iov->iov_base;
8354 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8355 start = ubuf >> PAGE_SHIFT;
8356 nr_pages = end - start;
8361 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8365 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8370 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8375 mmap_read_lock(current->mm);
8376 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8378 if (pret == nr_pages) {
8379 /* don't support file backed memory */
8380 for (i = 0; i < nr_pages; i++) {
8381 struct vm_area_struct *vma = vmas[i];
8383 if (vma_is_shmem(vma))
8386 !is_file_hugepages(vma->vm_file)) {
8392 ret = pret < 0 ? pret : -EFAULT;
8394 mmap_read_unlock(current->mm);
8397 * if we did partial map, or found file backed vmas,
8398 * release any pages we did get
8401 unpin_user_pages(pages, pret);
8405 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8407 unpin_user_pages(pages, pret);
8411 off = ubuf & ~PAGE_MASK;
8412 size = iov->iov_len;
8413 for (i = 0; i < nr_pages; i++) {
8416 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8417 imu->bvec[i].bv_page = pages[i];
8418 imu->bvec[i].bv_len = vec_len;
8419 imu->bvec[i].bv_offset = off;
8423 /* store original address for later verification */
8425 imu->ubuf_end = ubuf + iov->iov_len;
8426 imu->nr_bvecs = nr_pages;
8437 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8439 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8440 return ctx->user_bufs ? 0 : -ENOMEM;
8443 static int io_buffer_validate(struct iovec *iov)
8445 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8448 * Don't impose further limits on the size and buffer
8449 * constraints here, we'll -EINVAL later when IO is
8450 * submitted if they are wrong.
8453 return iov->iov_len ? -EFAULT : 0;
8457 /* arbitrary limit, but we need something */
8458 if (iov->iov_len > SZ_1G)
8461 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8467 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8468 unsigned int nr_args, u64 __user *tags)
8470 struct page *last_hpage = NULL;
8471 struct io_rsrc_data *data;
8477 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8479 ret = io_rsrc_node_switch_start(ctx);
8482 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8485 ret = io_buffers_map_alloc(ctx, nr_args);
8487 io_rsrc_data_free(data);
8491 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8492 ret = io_copy_iov(ctx, &iov, arg, i);
8495 ret = io_buffer_validate(&iov);
8498 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8503 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8509 WARN_ON_ONCE(ctx->buf_data);
8511 ctx->buf_data = data;
8513 __io_sqe_buffers_unregister(ctx);
8515 io_rsrc_node_switch(ctx, NULL);
8519 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8520 struct io_uring_rsrc_update2 *up,
8521 unsigned int nr_args)
8523 u64 __user *tags = u64_to_user_ptr(up->tags);
8524 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8525 struct page *last_hpage = NULL;
8526 bool needs_switch = false;
8532 if (up->offset + nr_args > ctx->nr_user_bufs)
8535 for (done = 0; done < nr_args; done++) {
8536 struct io_mapped_ubuf *imu;
8537 int offset = up->offset + done;
8540 err = io_copy_iov(ctx, &iov, iovs, done);
8543 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8547 err = io_buffer_validate(&iov);
8550 if (!iov.iov_base && tag) {
8554 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8558 i = array_index_nospec(offset, ctx->nr_user_bufs);
8559 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8560 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8561 ctx->rsrc_node, ctx->user_bufs[i]);
8562 if (unlikely(err)) {
8563 io_buffer_unmap(ctx, &imu);
8566 ctx->user_bufs[i] = NULL;
8567 needs_switch = true;
8570 ctx->user_bufs[i] = imu;
8571 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8575 io_rsrc_node_switch(ctx, ctx->buf_data);
8576 return done ? done : err;
8579 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8581 __s32 __user *fds = arg;
8587 if (copy_from_user(&fd, fds, sizeof(*fds)))
8590 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8591 if (IS_ERR(ctx->cq_ev_fd)) {
8592 int ret = PTR_ERR(ctx->cq_ev_fd);
8594 ctx->cq_ev_fd = NULL;
8601 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8603 if (ctx->cq_ev_fd) {
8604 eventfd_ctx_put(ctx->cq_ev_fd);
8605 ctx->cq_ev_fd = NULL;
8612 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8614 struct io_buffer *buf;
8615 unsigned long index;
8617 xa_for_each(&ctx->io_buffers, index, buf)
8618 __io_remove_buffers(ctx, buf, index, -1U);
8621 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8623 struct io_kiocb *req, *nxt;
8625 list_for_each_entry_safe(req, nxt, list, compl.list) {
8626 if (tsk && req->task != tsk)
8628 list_del(&req->compl.list);
8629 kmem_cache_free(req_cachep, req);
8633 static void io_req_caches_free(struct io_ring_ctx *ctx)
8635 struct io_submit_state *submit_state = &ctx->submit_state;
8636 struct io_comp_state *cs = &ctx->submit_state.comp;
8638 mutex_lock(&ctx->uring_lock);
8640 if (submit_state->free_reqs) {
8641 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8642 submit_state->reqs);
8643 submit_state->free_reqs = 0;
8646 io_flush_cached_locked_reqs(ctx, cs);
8647 io_req_cache_free(&cs->free_list, NULL);
8648 mutex_unlock(&ctx->uring_lock);
8651 static bool io_wait_rsrc_data(struct io_rsrc_data *data)
8655 if (!atomic_dec_and_test(&data->refs))
8656 wait_for_completion(&data->done);
8660 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8662 io_sq_thread_finish(ctx);
8664 if (ctx->mm_account) {
8665 mmdrop(ctx->mm_account);
8666 ctx->mm_account = NULL;
8669 mutex_lock(&ctx->uring_lock);
8670 if (io_wait_rsrc_data(ctx->buf_data))
8671 __io_sqe_buffers_unregister(ctx);
8672 if (io_wait_rsrc_data(ctx->file_data))
8673 __io_sqe_files_unregister(ctx);
8675 __io_cqring_overflow_flush(ctx, true);
8676 mutex_unlock(&ctx->uring_lock);
8677 io_eventfd_unregister(ctx);
8678 io_destroy_buffers(ctx);
8680 put_cred(ctx->sq_creds);
8682 /* there are no registered resources left, nobody uses it */
8684 io_rsrc_node_destroy(ctx->rsrc_node);
8685 if (ctx->rsrc_backup_node)
8686 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8687 flush_delayed_work(&ctx->rsrc_put_work);
8689 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8690 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8692 #if defined(CONFIG_UNIX)
8693 if (ctx->ring_sock) {
8694 ctx->ring_sock->file = NULL; /* so that iput() is called */
8695 sock_release(ctx->ring_sock);
8699 io_mem_free(ctx->rings);
8700 io_mem_free(ctx->sq_sqes);
8702 percpu_ref_exit(&ctx->refs);
8703 free_uid(ctx->user);
8704 io_req_caches_free(ctx);
8706 io_wq_put_hash(ctx->hash_map);
8707 kfree(ctx->cancel_hash);
8708 kfree(ctx->dummy_ubuf);
8712 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8714 struct io_ring_ctx *ctx = file->private_data;
8717 poll_wait(file, &ctx->poll_wait, wait);
8719 * synchronizes with barrier from wq_has_sleeper call in
8723 if (!io_sqring_full(ctx))
8724 mask |= EPOLLOUT | EPOLLWRNORM;
8727 * Don't flush cqring overflow list here, just do a simple check.
8728 * Otherwise there could possible be ABBA deadlock:
8731 * lock(&ctx->uring_lock);
8733 * lock(&ctx->uring_lock);
8736 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8737 * pushs them to do the flush.
8739 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8740 mask |= EPOLLIN | EPOLLRDNORM;
8745 static int io_uring_fasync(int fd, struct file *file, int on)
8747 struct io_ring_ctx *ctx = file->private_data;
8749 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8752 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8754 const struct cred *creds;
8756 creds = xa_erase(&ctx->personalities, id);
8765 struct io_tctx_exit {
8766 struct callback_head task_work;
8767 struct completion completion;
8768 struct io_ring_ctx *ctx;
8771 static void io_tctx_exit_cb(struct callback_head *cb)
8773 struct io_uring_task *tctx = current->io_uring;
8774 struct io_tctx_exit *work;
8776 work = container_of(cb, struct io_tctx_exit, task_work);
8778 * When @in_idle, we're in cancellation and it's racy to remove the
8779 * node. It'll be removed by the end of cancellation, just ignore it.
8781 if (!atomic_read(&tctx->in_idle))
8782 io_uring_del_tctx_node((unsigned long)work->ctx);
8783 complete(&work->completion);
8786 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8788 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8790 return req->ctx == data;
8793 static void io_ring_exit_work(struct work_struct *work)
8795 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8796 unsigned long timeout = jiffies + HZ * 60 * 5;
8797 struct io_tctx_exit exit;
8798 struct io_tctx_node *node;
8802 * If we're doing polled IO and end up having requests being
8803 * submitted async (out-of-line), then completions can come in while
8804 * we're waiting for refs to drop. We need to reap these manually,
8805 * as nobody else will be looking for them.
8808 io_uring_try_cancel_requests(ctx, NULL, true);
8810 struct io_sq_data *sqd = ctx->sq_data;
8811 struct task_struct *tsk;
8813 io_sq_thread_park(sqd);
8815 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8816 io_wq_cancel_cb(tsk->io_uring->io_wq,
8817 io_cancel_ctx_cb, ctx, true);
8818 io_sq_thread_unpark(sqd);
8821 WARN_ON_ONCE(time_after(jiffies, timeout));
8822 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8824 init_completion(&exit.completion);
8825 init_task_work(&exit.task_work, io_tctx_exit_cb);
8828 * Some may use context even when all refs and requests have been put,
8829 * and they are free to do so while still holding uring_lock or
8830 * completion_lock, see io_req_task_submit(). Apart from other work,
8831 * this lock/unlock section also waits them to finish.
8833 mutex_lock(&ctx->uring_lock);
8834 while (!list_empty(&ctx->tctx_list)) {
8835 WARN_ON_ONCE(time_after(jiffies, timeout));
8837 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8839 /* don't spin on a single task if cancellation failed */
8840 list_rotate_left(&ctx->tctx_list);
8841 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8842 if (WARN_ON_ONCE(ret))
8844 wake_up_process(node->task);
8846 mutex_unlock(&ctx->uring_lock);
8847 wait_for_completion(&exit.completion);
8848 mutex_lock(&ctx->uring_lock);
8850 mutex_unlock(&ctx->uring_lock);
8851 spin_lock_irq(&ctx->completion_lock);
8852 spin_unlock_irq(&ctx->completion_lock);
8854 io_ring_ctx_free(ctx);
8857 /* Returns true if we found and killed one or more timeouts */
8858 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8861 struct io_kiocb *req, *tmp;
8864 spin_lock_irq(&ctx->completion_lock);
8865 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8866 if (io_match_task(req, tsk, cancel_all)) {
8867 io_kill_timeout(req, -ECANCELED);
8872 io_commit_cqring(ctx);
8873 spin_unlock_irq(&ctx->completion_lock);
8875 io_cqring_ev_posted(ctx);
8876 return canceled != 0;
8879 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8881 unsigned long index;
8882 struct creds *creds;
8884 mutex_lock(&ctx->uring_lock);
8885 percpu_ref_kill(&ctx->refs);
8887 __io_cqring_overflow_flush(ctx, true);
8888 xa_for_each(&ctx->personalities, index, creds)
8889 io_unregister_personality(ctx, index);
8890 mutex_unlock(&ctx->uring_lock);
8892 io_kill_timeouts(ctx, NULL, true);
8893 io_poll_remove_all(ctx, NULL, true);
8895 /* if we failed setting up the ctx, we might not have any rings */
8896 io_iopoll_try_reap_events(ctx);
8898 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8900 * Use system_unbound_wq to avoid spawning tons of event kworkers
8901 * if we're exiting a ton of rings at the same time. It just adds
8902 * noise and overhead, there's no discernable change in runtime
8903 * over using system_wq.
8905 queue_work(system_unbound_wq, &ctx->exit_work);
8908 static int io_uring_release(struct inode *inode, struct file *file)
8910 struct io_ring_ctx *ctx = file->private_data;
8912 file->private_data = NULL;
8913 io_ring_ctx_wait_and_kill(ctx);
8917 struct io_task_cancel {
8918 struct task_struct *task;
8922 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8924 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8925 struct io_task_cancel *cancel = data;
8928 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8929 unsigned long flags;
8930 struct io_ring_ctx *ctx = req->ctx;
8932 /* protect against races with linked timeouts */
8933 spin_lock_irqsave(&ctx->completion_lock, flags);
8934 ret = io_match_task(req, cancel->task, cancel->all);
8935 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8937 ret = io_match_task(req, cancel->task, cancel->all);
8942 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8943 struct task_struct *task, bool cancel_all)
8945 struct io_defer_entry *de;
8948 spin_lock_irq(&ctx->completion_lock);
8949 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8950 if (io_match_task(de->req, task, cancel_all)) {
8951 list_cut_position(&list, &ctx->defer_list, &de->list);
8955 spin_unlock_irq(&ctx->completion_lock);
8956 if (list_empty(&list))
8959 while (!list_empty(&list)) {
8960 de = list_first_entry(&list, struct io_defer_entry, list);
8961 list_del_init(&de->list);
8962 io_req_complete_failed(de->req, -ECANCELED);
8968 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8970 struct io_tctx_node *node;
8971 enum io_wq_cancel cret;
8974 mutex_lock(&ctx->uring_lock);
8975 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8976 struct io_uring_task *tctx = node->task->io_uring;
8979 * io_wq will stay alive while we hold uring_lock, because it's
8980 * killed after ctx nodes, which requires to take the lock.
8982 if (!tctx || !tctx->io_wq)
8984 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8985 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8987 mutex_unlock(&ctx->uring_lock);
8992 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8993 struct task_struct *task,
8996 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
8997 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9000 enum io_wq_cancel cret;
9004 ret |= io_uring_try_cancel_iowq(ctx);
9005 } else if (tctx && tctx->io_wq) {
9007 * Cancels requests of all rings, not only @ctx, but
9008 * it's fine as the task is in exit/exec.
9010 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9012 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9015 /* SQPOLL thread does its own polling */
9016 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9017 (ctx->sq_data && ctx->sq_data->thread == current)) {
9018 while (!list_empty_careful(&ctx->iopoll_list)) {
9019 io_iopoll_try_reap_events(ctx);
9024 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9025 ret |= io_poll_remove_all(ctx, task, cancel_all);
9026 ret |= io_kill_timeouts(ctx, task, cancel_all);
9028 ret |= io_run_task_work();
9035 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9037 struct io_uring_task *tctx = current->io_uring;
9038 struct io_tctx_node *node;
9041 if (unlikely(!tctx)) {
9042 ret = io_uring_alloc_task_context(current, ctx);
9045 tctx = current->io_uring;
9047 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9048 node = kmalloc(sizeof(*node), GFP_KERNEL);
9052 node->task = current;
9054 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9061 mutex_lock(&ctx->uring_lock);
9062 list_add(&node->ctx_node, &ctx->tctx_list);
9063 mutex_unlock(&ctx->uring_lock);
9070 * Note that this task has used io_uring. We use it for cancelation purposes.
9072 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9074 struct io_uring_task *tctx = current->io_uring;
9076 if (likely(tctx && tctx->last == ctx))
9078 return __io_uring_add_tctx_node(ctx);
9082 * Remove this io_uring_file -> task mapping.
9084 static void io_uring_del_tctx_node(unsigned long index)
9086 struct io_uring_task *tctx = current->io_uring;
9087 struct io_tctx_node *node;
9091 node = xa_erase(&tctx->xa, index);
9095 WARN_ON_ONCE(current != node->task);
9096 WARN_ON_ONCE(list_empty(&node->ctx_node));
9098 mutex_lock(&node->ctx->uring_lock);
9099 list_del(&node->ctx_node);
9100 mutex_unlock(&node->ctx->uring_lock);
9102 if (tctx->last == node->ctx)
9107 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9109 struct io_wq *wq = tctx->io_wq;
9110 struct io_tctx_node *node;
9111 unsigned long index;
9113 xa_for_each(&tctx->xa, index, node)
9114 io_uring_del_tctx_node(index);
9117 * Must be after io_uring_del_task_file() (removes nodes under
9118 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9121 io_wq_put_and_exit(wq);
9125 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9128 return atomic_read(&tctx->inflight_tracked);
9129 return percpu_counter_sum(&tctx->inflight);
9132 static void io_uring_drop_tctx_refs(struct task_struct *task)
9134 struct io_uring_task *tctx = task->io_uring;
9135 unsigned int refs = tctx->cached_refs;
9137 tctx->cached_refs = 0;
9138 percpu_counter_sub(&tctx->inflight, refs);
9139 put_task_struct_many(task, refs);
9143 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9144 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9146 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9148 struct io_uring_task *tctx = current->io_uring;
9149 struct io_ring_ctx *ctx;
9153 WARN_ON_ONCE(sqd && sqd->thread != current);
9155 if (!current->io_uring)
9158 io_wq_exit_start(tctx->io_wq);
9160 io_uring_drop_tctx_refs(current);
9161 atomic_inc(&tctx->in_idle);
9163 /* read completions before cancelations */
9164 inflight = tctx_inflight(tctx, !cancel_all);
9169 struct io_tctx_node *node;
9170 unsigned long index;
9172 xa_for_each(&tctx->xa, index, node) {
9173 /* sqpoll task will cancel all its requests */
9174 if (node->ctx->sq_data)
9176 io_uring_try_cancel_requests(node->ctx, current,
9180 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9181 io_uring_try_cancel_requests(ctx, current,
9185 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9187 * If we've seen completions, retry without waiting. This
9188 * avoids a race where a completion comes in before we did
9189 * prepare_to_wait().
9191 if (inflight == tctx_inflight(tctx, !cancel_all))
9193 finish_wait(&tctx->wait, &wait);
9195 atomic_dec(&tctx->in_idle);
9197 io_uring_clean_tctx(tctx);
9199 /* for exec all current's requests should be gone, kill tctx */
9200 __io_uring_free(current);
9204 void __io_uring_cancel(struct files_struct *files)
9206 io_uring_cancel_generic(!files, NULL);
9209 static void *io_uring_validate_mmap_request(struct file *file,
9210 loff_t pgoff, size_t sz)
9212 struct io_ring_ctx *ctx = file->private_data;
9213 loff_t offset = pgoff << PAGE_SHIFT;
9218 case IORING_OFF_SQ_RING:
9219 case IORING_OFF_CQ_RING:
9222 case IORING_OFF_SQES:
9226 return ERR_PTR(-EINVAL);
9229 page = virt_to_head_page(ptr);
9230 if (sz > page_size(page))
9231 return ERR_PTR(-EINVAL);
9238 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9240 size_t sz = vma->vm_end - vma->vm_start;
9244 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9246 return PTR_ERR(ptr);
9248 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9249 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9252 #else /* !CONFIG_MMU */
9254 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9256 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9259 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9261 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9264 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9265 unsigned long addr, unsigned long len,
9266 unsigned long pgoff, unsigned long flags)
9270 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9272 return PTR_ERR(ptr);
9274 return (unsigned long) ptr;
9277 #endif /* !CONFIG_MMU */
9279 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9284 if (!io_sqring_full(ctx))
9286 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9288 if (!io_sqring_full(ctx))
9291 } while (!signal_pending(current));
9293 finish_wait(&ctx->sqo_sq_wait, &wait);
9297 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9298 struct __kernel_timespec __user **ts,
9299 const sigset_t __user **sig)
9301 struct io_uring_getevents_arg arg;
9304 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9305 * is just a pointer to the sigset_t.
9307 if (!(flags & IORING_ENTER_EXT_ARG)) {
9308 *sig = (const sigset_t __user *) argp;
9314 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9315 * timespec and sigset_t pointers if good.
9317 if (*argsz != sizeof(arg))
9319 if (copy_from_user(&arg, argp, sizeof(arg)))
9321 *sig = u64_to_user_ptr(arg.sigmask);
9322 *argsz = arg.sigmask_sz;
9323 *ts = u64_to_user_ptr(arg.ts);
9327 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9328 u32, min_complete, u32, flags, const void __user *, argp,
9331 struct io_ring_ctx *ctx;
9338 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9339 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9343 if (unlikely(!f.file))
9347 if (unlikely(f.file->f_op != &io_uring_fops))
9351 ctx = f.file->private_data;
9352 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9356 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9360 * For SQ polling, the thread will do all submissions and completions.
9361 * Just return the requested submit count, and wake the thread if
9365 if (ctx->flags & IORING_SETUP_SQPOLL) {
9366 io_cqring_overflow_flush(ctx, false);
9369 if (unlikely(ctx->sq_data->thread == NULL))
9371 if (flags & IORING_ENTER_SQ_WAKEUP)
9372 wake_up(&ctx->sq_data->wait);
9373 if (flags & IORING_ENTER_SQ_WAIT) {
9374 ret = io_sqpoll_wait_sq(ctx);
9378 submitted = to_submit;
9379 } else if (to_submit) {
9380 ret = io_uring_add_tctx_node(ctx);
9383 mutex_lock(&ctx->uring_lock);
9384 submitted = io_submit_sqes(ctx, to_submit);
9385 mutex_unlock(&ctx->uring_lock);
9387 if (submitted != to_submit)
9390 if (flags & IORING_ENTER_GETEVENTS) {
9391 const sigset_t __user *sig;
9392 struct __kernel_timespec __user *ts;
9394 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9398 min_complete = min(min_complete, ctx->cq_entries);
9401 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9402 * space applications don't need to do io completion events
9403 * polling again, they can rely on io_sq_thread to do polling
9404 * work, which can reduce cpu usage and uring_lock contention.
9406 if (ctx->flags & IORING_SETUP_IOPOLL &&
9407 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9408 ret = io_iopoll_check(ctx, min_complete);
9410 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9415 percpu_ref_put(&ctx->refs);
9418 return submitted ? submitted : ret;
9421 #ifdef CONFIG_PROC_FS
9422 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9423 const struct cred *cred)
9425 struct user_namespace *uns = seq_user_ns(m);
9426 struct group_info *gi;
9431 seq_printf(m, "%5d\n", id);
9432 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9433 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9434 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9435 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9436 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9437 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9438 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9439 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9440 seq_puts(m, "\n\tGroups:\t");
9441 gi = cred->group_info;
9442 for (g = 0; g < gi->ngroups; g++) {
9443 seq_put_decimal_ull(m, g ? " " : "",
9444 from_kgid_munged(uns, gi->gid[g]));
9446 seq_puts(m, "\n\tCapEff:\t");
9447 cap = cred->cap_effective;
9448 CAP_FOR_EACH_U32(__capi)
9449 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9454 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9456 struct io_sq_data *sq = NULL;
9461 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9462 * since fdinfo case grabs it in the opposite direction of normal use
9463 * cases. If we fail to get the lock, we just don't iterate any
9464 * structures that could be going away outside the io_uring mutex.
9466 has_lock = mutex_trylock(&ctx->uring_lock);
9468 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9474 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9475 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9476 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9477 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9478 struct file *f = io_file_from_index(ctx, i);
9481 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9483 seq_printf(m, "%5u: <none>\n", i);
9485 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9486 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9487 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9488 unsigned int len = buf->ubuf_end - buf->ubuf;
9490 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9492 if (has_lock && !xa_empty(&ctx->personalities)) {
9493 unsigned long index;
9494 const struct cred *cred;
9496 seq_printf(m, "Personalities:\n");
9497 xa_for_each(&ctx->personalities, index, cred)
9498 io_uring_show_cred(m, index, cred);
9500 seq_printf(m, "PollList:\n");
9501 spin_lock_irq(&ctx->completion_lock);
9502 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9503 struct hlist_head *list = &ctx->cancel_hash[i];
9504 struct io_kiocb *req;
9506 hlist_for_each_entry(req, list, hash_node)
9507 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9508 req->task->task_works != NULL);
9510 spin_unlock_irq(&ctx->completion_lock);
9512 mutex_unlock(&ctx->uring_lock);
9515 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9517 struct io_ring_ctx *ctx = f->private_data;
9519 if (percpu_ref_tryget(&ctx->refs)) {
9520 __io_uring_show_fdinfo(ctx, m);
9521 percpu_ref_put(&ctx->refs);
9526 static const struct file_operations io_uring_fops = {
9527 .release = io_uring_release,
9528 .mmap = io_uring_mmap,
9530 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9531 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9533 .poll = io_uring_poll,
9534 .fasync = io_uring_fasync,
9535 #ifdef CONFIG_PROC_FS
9536 .show_fdinfo = io_uring_show_fdinfo,
9540 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9541 struct io_uring_params *p)
9543 struct io_rings *rings;
9544 size_t size, sq_array_offset;
9546 /* make sure these are sane, as we already accounted them */
9547 ctx->sq_entries = p->sq_entries;
9548 ctx->cq_entries = p->cq_entries;
9550 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9551 if (size == SIZE_MAX)
9554 rings = io_mem_alloc(size);
9559 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9560 rings->sq_ring_mask = p->sq_entries - 1;
9561 rings->cq_ring_mask = p->cq_entries - 1;
9562 rings->sq_ring_entries = p->sq_entries;
9563 rings->cq_ring_entries = p->cq_entries;
9565 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9566 if (size == SIZE_MAX) {
9567 io_mem_free(ctx->rings);
9572 ctx->sq_sqes = io_mem_alloc(size);
9573 if (!ctx->sq_sqes) {
9574 io_mem_free(ctx->rings);
9582 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9586 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9590 ret = io_uring_add_tctx_node(ctx);
9595 fd_install(fd, file);
9600 * Allocate an anonymous fd, this is what constitutes the application
9601 * visible backing of an io_uring instance. The application mmaps this
9602 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9603 * we have to tie this fd to a socket for file garbage collection purposes.
9605 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9608 #if defined(CONFIG_UNIX)
9611 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9614 return ERR_PTR(ret);
9617 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9618 O_RDWR | O_CLOEXEC);
9619 #if defined(CONFIG_UNIX)
9621 sock_release(ctx->ring_sock);
9622 ctx->ring_sock = NULL;
9624 ctx->ring_sock->file = file;
9630 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9631 struct io_uring_params __user *params)
9633 struct io_ring_ctx *ctx;
9639 if (entries > IORING_MAX_ENTRIES) {
9640 if (!(p->flags & IORING_SETUP_CLAMP))
9642 entries = IORING_MAX_ENTRIES;
9646 * Use twice as many entries for the CQ ring. It's possible for the
9647 * application to drive a higher depth than the size of the SQ ring,
9648 * since the sqes are only used at submission time. This allows for
9649 * some flexibility in overcommitting a bit. If the application has
9650 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9651 * of CQ ring entries manually.
9653 p->sq_entries = roundup_pow_of_two(entries);
9654 if (p->flags & IORING_SETUP_CQSIZE) {
9656 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9657 * to a power-of-two, if it isn't already. We do NOT impose
9658 * any cq vs sq ring sizing.
9662 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9663 if (!(p->flags & IORING_SETUP_CLAMP))
9665 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9667 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9668 if (p->cq_entries < p->sq_entries)
9671 p->cq_entries = 2 * p->sq_entries;
9674 ctx = io_ring_ctx_alloc(p);
9677 ctx->compat = in_compat_syscall();
9678 if (!capable(CAP_IPC_LOCK))
9679 ctx->user = get_uid(current_user());
9682 * This is just grabbed for accounting purposes. When a process exits,
9683 * the mm is exited and dropped before the files, hence we need to hang
9684 * on to this mm purely for the purposes of being able to unaccount
9685 * memory (locked/pinned vm). It's not used for anything else.
9687 mmgrab(current->mm);
9688 ctx->mm_account = current->mm;
9690 ret = io_allocate_scq_urings(ctx, p);
9694 ret = io_sq_offload_create(ctx, p);
9697 /* always set a rsrc node */
9698 ret = io_rsrc_node_switch_start(ctx);
9701 io_rsrc_node_switch(ctx, NULL);
9703 memset(&p->sq_off, 0, sizeof(p->sq_off));
9704 p->sq_off.head = offsetof(struct io_rings, sq.head);
9705 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9706 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9707 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9708 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9709 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9710 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9712 memset(&p->cq_off, 0, sizeof(p->cq_off));
9713 p->cq_off.head = offsetof(struct io_rings, cq.head);
9714 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9715 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9716 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9717 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9718 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9719 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9721 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9722 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9723 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9724 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9725 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9726 IORING_FEAT_RSRC_TAGS;
9728 if (copy_to_user(params, p, sizeof(*p))) {
9733 file = io_uring_get_file(ctx);
9735 ret = PTR_ERR(file);
9740 * Install ring fd as the very last thing, so we don't risk someone
9741 * having closed it before we finish setup
9743 ret = io_uring_install_fd(ctx, file);
9745 /* fput will clean it up */
9750 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9753 io_ring_ctx_wait_and_kill(ctx);
9758 * Sets up an aio uring context, and returns the fd. Applications asks for a
9759 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9760 * params structure passed in.
9762 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9764 struct io_uring_params p;
9767 if (copy_from_user(&p, params, sizeof(p)))
9769 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9774 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9775 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9776 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9777 IORING_SETUP_R_DISABLED))
9780 return io_uring_create(entries, &p, params);
9783 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9784 struct io_uring_params __user *, params)
9786 return io_uring_setup(entries, params);
9789 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9791 struct io_uring_probe *p;
9795 size = struct_size(p, ops, nr_args);
9796 if (size == SIZE_MAX)
9798 p = kzalloc(size, GFP_KERNEL);
9803 if (copy_from_user(p, arg, size))
9806 if (memchr_inv(p, 0, size))
9809 p->last_op = IORING_OP_LAST - 1;
9810 if (nr_args > IORING_OP_LAST)
9811 nr_args = IORING_OP_LAST;
9813 for (i = 0; i < nr_args; i++) {
9815 if (!io_op_defs[i].not_supported)
9816 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9821 if (copy_to_user(arg, p, size))
9828 static int io_register_personality(struct io_ring_ctx *ctx)
9830 const struct cred *creds;
9834 creds = get_current_cred();
9836 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9837 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9844 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9845 unsigned int nr_args)
9847 struct io_uring_restriction *res;
9851 /* Restrictions allowed only if rings started disabled */
9852 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9855 /* We allow only a single restrictions registration */
9856 if (ctx->restrictions.registered)
9859 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9862 size = array_size(nr_args, sizeof(*res));
9863 if (size == SIZE_MAX)
9866 res = memdup_user(arg, size);
9868 return PTR_ERR(res);
9872 for (i = 0; i < nr_args; i++) {
9873 switch (res[i].opcode) {
9874 case IORING_RESTRICTION_REGISTER_OP:
9875 if (res[i].register_op >= IORING_REGISTER_LAST) {
9880 __set_bit(res[i].register_op,
9881 ctx->restrictions.register_op);
9883 case IORING_RESTRICTION_SQE_OP:
9884 if (res[i].sqe_op >= IORING_OP_LAST) {
9889 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9891 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9892 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9894 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9895 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9904 /* Reset all restrictions if an error happened */
9906 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9908 ctx->restrictions.registered = true;
9914 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9916 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9919 if (ctx->restrictions.registered)
9920 ctx->restricted = 1;
9922 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9923 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9924 wake_up(&ctx->sq_data->wait);
9928 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9929 struct io_uring_rsrc_update2 *up,
9937 if (check_add_overflow(up->offset, nr_args, &tmp))
9939 err = io_rsrc_node_switch_start(ctx);
9944 case IORING_RSRC_FILE:
9945 return __io_sqe_files_update(ctx, up, nr_args);
9946 case IORING_RSRC_BUFFER:
9947 return __io_sqe_buffers_update(ctx, up, nr_args);
9952 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9955 struct io_uring_rsrc_update2 up;
9959 memset(&up, 0, sizeof(up));
9960 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9962 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9965 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9966 unsigned size, unsigned type)
9968 struct io_uring_rsrc_update2 up;
9970 if (size != sizeof(up))
9972 if (copy_from_user(&up, arg, sizeof(up)))
9974 if (!up.nr || up.resv)
9976 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9979 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9980 unsigned int size, unsigned int type)
9982 struct io_uring_rsrc_register rr;
9984 /* keep it extendible */
9985 if (size != sizeof(rr))
9988 memset(&rr, 0, sizeof(rr));
9989 if (copy_from_user(&rr, arg, size))
9991 if (!rr.nr || rr.resv || rr.resv2)
9995 case IORING_RSRC_FILE:
9996 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
9997 rr.nr, u64_to_user_ptr(rr.tags));
9998 case IORING_RSRC_BUFFER:
9999 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10000 rr.nr, u64_to_user_ptr(rr.tags));
10005 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10008 struct io_uring_task *tctx = current->io_uring;
10009 cpumask_var_t new_mask;
10012 if (!tctx || !tctx->io_wq)
10015 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10018 cpumask_clear(new_mask);
10019 if (len > cpumask_size())
10020 len = cpumask_size();
10022 if (copy_from_user(new_mask, arg, len)) {
10023 free_cpumask_var(new_mask);
10027 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10028 free_cpumask_var(new_mask);
10032 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10034 struct io_uring_task *tctx = current->io_uring;
10036 if (!tctx || !tctx->io_wq)
10039 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10042 static bool io_register_op_must_quiesce(int op)
10045 case IORING_REGISTER_BUFFERS:
10046 case IORING_UNREGISTER_BUFFERS:
10047 case IORING_REGISTER_FILES:
10048 case IORING_UNREGISTER_FILES:
10049 case IORING_REGISTER_FILES_UPDATE:
10050 case IORING_REGISTER_PROBE:
10051 case IORING_REGISTER_PERSONALITY:
10052 case IORING_UNREGISTER_PERSONALITY:
10053 case IORING_REGISTER_FILES2:
10054 case IORING_REGISTER_FILES_UPDATE2:
10055 case IORING_REGISTER_BUFFERS2:
10056 case IORING_REGISTER_BUFFERS_UPDATE:
10057 case IORING_REGISTER_IOWQ_AFF:
10058 case IORING_UNREGISTER_IOWQ_AFF:
10065 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10066 void __user *arg, unsigned nr_args)
10067 __releases(ctx->uring_lock)
10068 __acquires(ctx->uring_lock)
10073 * We're inside the ring mutex, if the ref is already dying, then
10074 * someone else killed the ctx or is already going through
10075 * io_uring_register().
10077 if (percpu_ref_is_dying(&ctx->refs))
10080 if (ctx->restricted) {
10081 if (opcode >= IORING_REGISTER_LAST)
10083 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10084 if (!test_bit(opcode, ctx->restrictions.register_op))
10088 if (io_register_op_must_quiesce(opcode)) {
10089 percpu_ref_kill(&ctx->refs);
10092 * Drop uring mutex before waiting for references to exit. If
10093 * another thread is currently inside io_uring_enter() it might
10094 * need to grab the uring_lock to make progress. If we hold it
10095 * here across the drain wait, then we can deadlock. It's safe
10096 * to drop the mutex here, since no new references will come in
10097 * after we've killed the percpu ref.
10099 mutex_unlock(&ctx->uring_lock);
10101 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10104 ret = io_run_task_work_sig();
10108 mutex_lock(&ctx->uring_lock);
10111 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10117 case IORING_REGISTER_BUFFERS:
10118 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10120 case IORING_UNREGISTER_BUFFERS:
10122 if (arg || nr_args)
10124 ret = io_sqe_buffers_unregister(ctx);
10126 case IORING_REGISTER_FILES:
10127 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10129 case IORING_UNREGISTER_FILES:
10131 if (arg || nr_args)
10133 ret = io_sqe_files_unregister(ctx);
10135 case IORING_REGISTER_FILES_UPDATE:
10136 ret = io_register_files_update(ctx, arg, nr_args);
10138 case IORING_REGISTER_EVENTFD:
10139 case IORING_REGISTER_EVENTFD_ASYNC:
10143 ret = io_eventfd_register(ctx, arg);
10146 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10147 ctx->eventfd_async = 1;
10149 ctx->eventfd_async = 0;
10151 case IORING_UNREGISTER_EVENTFD:
10153 if (arg || nr_args)
10155 ret = io_eventfd_unregister(ctx);
10157 case IORING_REGISTER_PROBE:
10159 if (!arg || nr_args > 256)
10161 ret = io_probe(ctx, arg, nr_args);
10163 case IORING_REGISTER_PERSONALITY:
10165 if (arg || nr_args)
10167 ret = io_register_personality(ctx);
10169 case IORING_UNREGISTER_PERSONALITY:
10173 ret = io_unregister_personality(ctx, nr_args);
10175 case IORING_REGISTER_ENABLE_RINGS:
10177 if (arg || nr_args)
10179 ret = io_register_enable_rings(ctx);
10181 case IORING_REGISTER_RESTRICTIONS:
10182 ret = io_register_restrictions(ctx, arg, nr_args);
10184 case IORING_REGISTER_FILES2:
10185 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10187 case IORING_REGISTER_FILES_UPDATE2:
10188 ret = io_register_rsrc_update(ctx, arg, nr_args,
10191 case IORING_REGISTER_BUFFERS2:
10192 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10194 case IORING_REGISTER_BUFFERS_UPDATE:
10195 ret = io_register_rsrc_update(ctx, arg, nr_args,
10196 IORING_RSRC_BUFFER);
10198 case IORING_REGISTER_IOWQ_AFF:
10200 if (!arg || !nr_args)
10202 ret = io_register_iowq_aff(ctx, arg, nr_args);
10204 case IORING_UNREGISTER_IOWQ_AFF:
10206 if (arg || nr_args)
10208 ret = io_unregister_iowq_aff(ctx);
10215 if (io_register_op_must_quiesce(opcode)) {
10216 /* bring the ctx back to life */
10217 percpu_ref_reinit(&ctx->refs);
10218 reinit_completion(&ctx->ref_comp);
10223 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10224 void __user *, arg, unsigned int, nr_args)
10226 struct io_ring_ctx *ctx;
10235 if (f.file->f_op != &io_uring_fops)
10238 ctx = f.file->private_data;
10240 io_run_task_work();
10242 mutex_lock(&ctx->uring_lock);
10243 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10244 mutex_unlock(&ctx->uring_lock);
10245 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10246 ctx->cq_ev_fd != NULL, ret);
10252 static int __init io_uring_init(void)
10254 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10255 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10256 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10259 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10260 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10261 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10262 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10263 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10264 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10265 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10266 BUILD_BUG_SQE_ELEM(8, __u64, off);
10267 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10268 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10269 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10270 BUILD_BUG_SQE_ELEM(24, __u32, len);
10271 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10272 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10273 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10274 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10275 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10276 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10277 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10278 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10279 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10280 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10281 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10282 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10283 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10284 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10285 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10286 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10287 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10288 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10289 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10290 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10292 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10293 sizeof(struct io_uring_rsrc_update));
10294 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10295 sizeof(struct io_uring_rsrc_update2));
10296 /* should fit into one byte */
10297 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10299 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10300 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10302 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10306 __initcall(io_uring_init);