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))
2451 * Play it safe and assume not safe to re-import and reissue if we're
2452 * not in the original thread group (or in task context).
2454 if (!same_thread_group(req->task, current) || !in_task())
2459 static bool io_resubmit_prep(struct io_kiocb *req)
2463 static bool io_rw_should_reissue(struct io_kiocb *req)
2469 static void io_fallback_req_func(struct work_struct *work)
2471 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
2472 fallback_work.work);
2473 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
2474 struct io_kiocb *req, *tmp;
2476 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
2477 req->io_task_work.func(req);
2480 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2481 unsigned int issue_flags)
2485 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2486 kiocb_end_write(req);
2487 if (res != req->result) {
2488 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2489 io_rw_should_reissue(req)) {
2490 req->flags |= REQ_F_REISSUE;
2495 if (req->flags & REQ_F_BUFFER_SELECTED)
2496 cflags = io_put_rw_kbuf(req);
2497 __io_req_complete(req, issue_flags, res, cflags);
2500 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2502 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2504 __io_complete_rw(req, res, res2, 0);
2507 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2509 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2511 if (kiocb->ki_flags & IOCB_WRITE)
2512 kiocb_end_write(req);
2513 if (unlikely(res != req->result)) {
2514 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2515 io_resubmit_prep(req))) {
2517 req->flags |= REQ_F_DONT_REISSUE;
2521 WRITE_ONCE(req->result, res);
2522 /* order with io_iopoll_complete() checking ->result */
2524 WRITE_ONCE(req->iopoll_completed, 1);
2528 * After the iocb has been issued, it's safe to be found on the poll list.
2529 * Adding the kiocb to the list AFTER submission ensures that we don't
2530 * find it from a io_do_iopoll() thread before the issuer is done
2531 * accessing the kiocb cookie.
2533 static void io_iopoll_req_issued(struct io_kiocb *req)
2535 struct io_ring_ctx *ctx = req->ctx;
2536 const bool in_async = io_wq_current_is_worker();
2538 /* workqueue context doesn't hold uring_lock, grab it now */
2539 if (unlikely(in_async))
2540 mutex_lock(&ctx->uring_lock);
2543 * Track whether we have multiple files in our lists. This will impact
2544 * how we do polling eventually, not spinning if we're on potentially
2545 * different devices.
2547 if (list_empty(&ctx->iopoll_list)) {
2548 ctx->poll_multi_queue = false;
2549 } else if (!ctx->poll_multi_queue) {
2550 struct io_kiocb *list_req;
2551 unsigned int queue_num0, queue_num1;
2553 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2556 if (list_req->file != req->file) {
2557 ctx->poll_multi_queue = true;
2559 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2560 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2561 if (queue_num0 != queue_num1)
2562 ctx->poll_multi_queue = true;
2567 * For fast devices, IO may have already completed. If it has, add
2568 * it to the front so we find it first.
2570 if (READ_ONCE(req->iopoll_completed))
2571 list_add(&req->inflight_entry, &ctx->iopoll_list);
2573 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2575 if (unlikely(in_async)) {
2577 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2578 * in sq thread task context or in io worker task context. If
2579 * current task context is sq thread, we don't need to check
2580 * whether should wake up sq thread.
2582 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2583 wq_has_sleeper(&ctx->sq_data->wait))
2584 wake_up(&ctx->sq_data->wait);
2586 mutex_unlock(&ctx->uring_lock);
2590 static inline void io_state_file_put(struct io_submit_state *state)
2592 if (state->file_refs) {
2593 fput_many(state->file, state->file_refs);
2594 state->file_refs = 0;
2599 * Get as many references to a file as we have IOs left in this submission,
2600 * assuming most submissions are for one file, or at least that each file
2601 * has more than one submission.
2603 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2608 if (state->file_refs) {
2609 if (state->fd == fd) {
2613 io_state_file_put(state);
2615 state->file = fget_many(fd, state->ios_left);
2616 if (unlikely(!state->file))
2620 state->file_refs = state->ios_left - 1;
2624 static bool io_bdev_nowait(struct block_device *bdev)
2626 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2630 * If we tracked the file through the SCM inflight mechanism, we could support
2631 * any file. For now, just ensure that anything potentially problematic is done
2634 static bool __io_file_supports_async(struct file *file, int rw)
2636 umode_t mode = file_inode(file)->i_mode;
2638 if (S_ISBLK(mode)) {
2639 if (IS_ENABLED(CONFIG_BLOCK) &&
2640 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2646 if (S_ISREG(mode)) {
2647 if (IS_ENABLED(CONFIG_BLOCK) &&
2648 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2649 file->f_op != &io_uring_fops)
2654 /* any ->read/write should understand O_NONBLOCK */
2655 if (file->f_flags & O_NONBLOCK)
2658 if (!(file->f_mode & FMODE_NOWAIT))
2662 return file->f_op->read_iter != NULL;
2664 return file->f_op->write_iter != NULL;
2667 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2669 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2671 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2674 return __io_file_supports_async(req->file, rw);
2677 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2679 struct io_ring_ctx *ctx = req->ctx;
2680 struct kiocb *kiocb = &req->rw.kiocb;
2681 struct file *file = req->file;
2685 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2686 req->flags |= REQ_F_ISREG;
2688 kiocb->ki_pos = READ_ONCE(sqe->off);
2689 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2690 req->flags |= REQ_F_CUR_POS;
2691 kiocb->ki_pos = file->f_pos;
2693 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2694 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2695 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2699 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2700 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2701 req->flags |= REQ_F_NOWAIT;
2703 ioprio = READ_ONCE(sqe->ioprio);
2705 ret = ioprio_check_cap(ioprio);
2709 kiocb->ki_ioprio = ioprio;
2711 kiocb->ki_ioprio = get_current_ioprio();
2713 if (ctx->flags & IORING_SETUP_IOPOLL) {
2714 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2715 !kiocb->ki_filp->f_op->iopoll)
2718 kiocb->ki_flags |= IOCB_HIPRI;
2719 kiocb->ki_complete = io_complete_rw_iopoll;
2720 req->iopoll_completed = 0;
2722 if (kiocb->ki_flags & IOCB_HIPRI)
2724 kiocb->ki_complete = io_complete_rw;
2727 if (req->opcode == IORING_OP_READ_FIXED ||
2728 req->opcode == IORING_OP_WRITE_FIXED) {
2730 io_req_set_rsrc_node(req);
2733 req->rw.addr = READ_ONCE(sqe->addr);
2734 req->rw.len = READ_ONCE(sqe->len);
2735 req->buf_index = READ_ONCE(sqe->buf_index);
2739 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2745 case -ERESTARTNOINTR:
2746 case -ERESTARTNOHAND:
2747 case -ERESTART_RESTARTBLOCK:
2749 * We can't just restart the syscall, since previously
2750 * submitted sqes may already be in progress. Just fail this
2756 kiocb->ki_complete(kiocb, ret, 0);
2760 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2761 unsigned int issue_flags)
2763 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2764 struct io_async_rw *io = req->async_data;
2765 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2767 /* add previously done IO, if any */
2768 if (io && io->bytes_done > 0) {
2770 ret = io->bytes_done;
2772 ret += io->bytes_done;
2775 if (req->flags & REQ_F_CUR_POS)
2776 req->file->f_pos = kiocb->ki_pos;
2777 if (ret >= 0 && check_reissue)
2778 __io_complete_rw(req, ret, 0, issue_flags);
2780 io_rw_done(kiocb, ret);
2782 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2783 req->flags &= ~REQ_F_REISSUE;
2784 if (io_resubmit_prep(req)) {
2786 io_req_task_queue_reissue(req);
2791 if (req->flags & REQ_F_BUFFER_SELECTED)
2792 cflags = io_put_rw_kbuf(req);
2793 __io_req_complete(req, issue_flags, ret, cflags);
2798 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2799 struct io_mapped_ubuf *imu)
2801 size_t len = req->rw.len;
2802 u64 buf_end, buf_addr = req->rw.addr;
2805 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2807 /* not inside the mapped region */
2808 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2812 * May not be a start of buffer, set size appropriately
2813 * and advance us to the beginning.
2815 offset = buf_addr - imu->ubuf;
2816 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2820 * Don't use iov_iter_advance() here, as it's really slow for
2821 * using the latter parts of a big fixed buffer - it iterates
2822 * over each segment manually. We can cheat a bit here, because
2825 * 1) it's a BVEC iter, we set it up
2826 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2827 * first and last bvec
2829 * So just find our index, and adjust the iterator afterwards.
2830 * If the offset is within the first bvec (or the whole first
2831 * bvec, just use iov_iter_advance(). This makes it easier
2832 * since we can just skip the first segment, which may not
2833 * be PAGE_SIZE aligned.
2835 const struct bio_vec *bvec = imu->bvec;
2837 if (offset <= bvec->bv_len) {
2838 iov_iter_advance(iter, offset);
2840 unsigned long seg_skip;
2842 /* skip first vec */
2843 offset -= bvec->bv_len;
2844 seg_skip = 1 + (offset >> PAGE_SHIFT);
2846 iter->bvec = bvec + seg_skip;
2847 iter->nr_segs -= seg_skip;
2848 iter->count -= bvec->bv_len + offset;
2849 iter->iov_offset = offset & ~PAGE_MASK;
2856 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2858 struct io_ring_ctx *ctx = req->ctx;
2859 struct io_mapped_ubuf *imu = req->imu;
2860 u16 index, buf_index = req->buf_index;
2863 if (unlikely(buf_index >= ctx->nr_user_bufs))
2865 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2866 imu = READ_ONCE(ctx->user_bufs[index]);
2869 return __io_import_fixed(req, rw, iter, imu);
2872 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2875 mutex_unlock(&ctx->uring_lock);
2878 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2881 * "Normal" inline submissions always hold the uring_lock, since we
2882 * grab it from the system call. Same is true for the SQPOLL offload.
2883 * The only exception is when we've detached the request and issue it
2884 * from an async worker thread, grab the lock for that case.
2887 mutex_lock(&ctx->uring_lock);
2890 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2891 int bgid, struct io_buffer *kbuf,
2894 struct io_buffer *head;
2896 if (req->flags & REQ_F_BUFFER_SELECTED)
2899 io_ring_submit_lock(req->ctx, needs_lock);
2901 lockdep_assert_held(&req->ctx->uring_lock);
2903 head = xa_load(&req->ctx->io_buffers, bgid);
2905 if (!list_empty(&head->list)) {
2906 kbuf = list_last_entry(&head->list, struct io_buffer,
2908 list_del(&kbuf->list);
2911 xa_erase(&req->ctx->io_buffers, bgid);
2913 if (*len > kbuf->len)
2916 kbuf = ERR_PTR(-ENOBUFS);
2919 io_ring_submit_unlock(req->ctx, needs_lock);
2924 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2927 struct io_buffer *kbuf;
2930 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2931 bgid = req->buf_index;
2932 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2935 req->rw.addr = (u64) (unsigned long) kbuf;
2936 req->flags |= REQ_F_BUFFER_SELECTED;
2937 return u64_to_user_ptr(kbuf->addr);
2940 #ifdef CONFIG_COMPAT
2941 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2944 struct compat_iovec __user *uiov;
2945 compat_ssize_t clen;
2949 uiov = u64_to_user_ptr(req->rw.addr);
2950 if (!access_ok(uiov, sizeof(*uiov)))
2952 if (__get_user(clen, &uiov->iov_len))
2958 buf = io_rw_buffer_select(req, &len, needs_lock);
2960 return PTR_ERR(buf);
2961 iov[0].iov_base = buf;
2962 iov[0].iov_len = (compat_size_t) len;
2967 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2970 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2974 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2977 len = iov[0].iov_len;
2980 buf = io_rw_buffer_select(req, &len, needs_lock);
2982 return PTR_ERR(buf);
2983 iov[0].iov_base = buf;
2984 iov[0].iov_len = len;
2988 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2991 if (req->flags & REQ_F_BUFFER_SELECTED) {
2992 struct io_buffer *kbuf;
2994 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2995 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2996 iov[0].iov_len = kbuf->len;
2999 if (req->rw.len != 1)
3002 #ifdef CONFIG_COMPAT
3003 if (req->ctx->compat)
3004 return io_compat_import(req, iov, needs_lock);
3007 return __io_iov_buffer_select(req, iov, needs_lock);
3010 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3011 struct iov_iter *iter, bool needs_lock)
3013 void __user *buf = u64_to_user_ptr(req->rw.addr);
3014 size_t sqe_len = req->rw.len;
3015 u8 opcode = req->opcode;
3018 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3020 return io_import_fixed(req, rw, iter);
3023 /* buffer index only valid with fixed read/write, or buffer select */
3024 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3027 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3028 if (req->flags & REQ_F_BUFFER_SELECT) {
3029 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3031 return PTR_ERR(buf);
3032 req->rw.len = sqe_len;
3035 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3040 if (req->flags & REQ_F_BUFFER_SELECT) {
3041 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3043 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3048 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3052 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3054 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3058 * For files that don't have ->read_iter() and ->write_iter(), handle them
3059 * by looping over ->read() or ->write() manually.
3061 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3063 struct kiocb *kiocb = &req->rw.kiocb;
3064 struct file *file = req->file;
3068 * Don't support polled IO through this interface, and we can't
3069 * support non-blocking either. For the latter, this just causes
3070 * the kiocb to be handled from an async context.
3072 if (kiocb->ki_flags & IOCB_HIPRI)
3074 if (kiocb->ki_flags & IOCB_NOWAIT)
3077 while (iov_iter_count(iter)) {
3081 if (!iov_iter_is_bvec(iter)) {
3082 iovec = iov_iter_iovec(iter);
3084 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3085 iovec.iov_len = req->rw.len;
3089 nr = file->f_op->read(file, iovec.iov_base,
3090 iovec.iov_len, io_kiocb_ppos(kiocb));
3092 nr = file->f_op->write(file, iovec.iov_base,
3093 iovec.iov_len, io_kiocb_ppos(kiocb));
3102 if (nr != iovec.iov_len)
3106 iov_iter_advance(iter, nr);
3112 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3113 const struct iovec *fast_iov, struct iov_iter *iter)
3115 struct io_async_rw *rw = req->async_data;
3117 memcpy(&rw->iter, iter, sizeof(*iter));
3118 rw->free_iovec = iovec;
3120 /* can only be fixed buffers, no need to do anything */
3121 if (iov_iter_is_bvec(iter))
3124 unsigned iov_off = 0;
3126 rw->iter.iov = rw->fast_iov;
3127 if (iter->iov != fast_iov) {
3128 iov_off = iter->iov - fast_iov;
3129 rw->iter.iov += iov_off;
3131 if (rw->fast_iov != fast_iov)
3132 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3133 sizeof(struct iovec) * iter->nr_segs);
3135 req->flags |= REQ_F_NEED_CLEANUP;
3139 static inline int io_alloc_async_data(struct io_kiocb *req)
3141 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3142 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3143 return req->async_data == NULL;
3146 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3147 const struct iovec *fast_iov,
3148 struct iov_iter *iter, bool force)
3150 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3152 if (!req->async_data) {
3153 if (io_alloc_async_data(req)) {
3158 io_req_map_rw(req, iovec, fast_iov, iter);
3163 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3165 struct io_async_rw *iorw = req->async_data;
3166 struct iovec *iov = iorw->fast_iov;
3169 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3170 if (unlikely(ret < 0))
3173 iorw->bytes_done = 0;
3174 iorw->free_iovec = iov;
3176 req->flags |= REQ_F_NEED_CLEANUP;
3180 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3182 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3184 return io_prep_rw(req, sqe);
3188 * This is our waitqueue callback handler, registered through lock_page_async()
3189 * when we initially tried to do the IO with the iocb armed our waitqueue.
3190 * This gets called when the page is unlocked, and we generally expect that to
3191 * happen when the page IO is completed and the page is now uptodate. This will
3192 * queue a task_work based retry of the operation, attempting to copy the data
3193 * again. If the latter fails because the page was NOT uptodate, then we will
3194 * do a thread based blocking retry of the operation. That's the unexpected
3197 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3198 int sync, void *arg)
3200 struct wait_page_queue *wpq;
3201 struct io_kiocb *req = wait->private;
3202 struct wait_page_key *key = arg;
3204 wpq = container_of(wait, struct wait_page_queue, wait);
3206 if (!wake_page_match(wpq, key))
3209 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3210 list_del_init(&wait->entry);
3212 /* submit ref gets dropped, acquire a new one */
3214 io_req_task_queue(req);
3219 * This controls whether a given IO request should be armed for async page
3220 * based retry. If we return false here, the request is handed to the async
3221 * worker threads for retry. If we're doing buffered reads on a regular file,
3222 * we prepare a private wait_page_queue entry and retry the operation. This
3223 * will either succeed because the page is now uptodate and unlocked, or it
3224 * will register a callback when the page is unlocked at IO completion. Through
3225 * that callback, io_uring uses task_work to setup a retry of the operation.
3226 * That retry will attempt the buffered read again. The retry will generally
3227 * succeed, or in rare cases where it fails, we then fall back to using the
3228 * async worker threads for a blocking retry.
3230 static bool io_rw_should_retry(struct io_kiocb *req)
3232 struct io_async_rw *rw = req->async_data;
3233 struct wait_page_queue *wait = &rw->wpq;
3234 struct kiocb *kiocb = &req->rw.kiocb;
3236 /* never retry for NOWAIT, we just complete with -EAGAIN */
3237 if (req->flags & REQ_F_NOWAIT)
3240 /* Only for buffered IO */
3241 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3245 * just use poll if we can, and don't attempt if the fs doesn't
3246 * support callback based unlocks
3248 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3251 wait->wait.func = io_async_buf_func;
3252 wait->wait.private = req;
3253 wait->wait.flags = 0;
3254 INIT_LIST_HEAD(&wait->wait.entry);
3255 kiocb->ki_flags |= IOCB_WAITQ;
3256 kiocb->ki_flags &= ~IOCB_NOWAIT;
3257 kiocb->ki_waitq = wait;
3261 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3263 if (req->file->f_op->read_iter)
3264 return call_read_iter(req->file, &req->rw.kiocb, iter);
3265 else if (req->file->f_op->read)
3266 return loop_rw_iter(READ, req, iter);
3271 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3273 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3274 struct kiocb *kiocb = &req->rw.kiocb;
3275 struct iov_iter __iter, *iter = &__iter;
3276 struct io_async_rw *rw = req->async_data;
3277 ssize_t io_size, ret, ret2;
3278 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3284 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3288 io_size = iov_iter_count(iter);
3289 req->result = io_size;
3291 /* Ensure we clear previously set non-block flag */
3292 if (!force_nonblock)
3293 kiocb->ki_flags &= ~IOCB_NOWAIT;
3295 kiocb->ki_flags |= IOCB_NOWAIT;
3297 /* If the file doesn't support async, just async punt */
3298 if (force_nonblock && !io_file_supports_async(req, READ)) {
3299 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3300 return ret ?: -EAGAIN;
3303 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3304 if (unlikely(ret)) {
3309 ret = io_iter_do_read(req, iter);
3311 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3312 req->flags &= ~REQ_F_REISSUE;
3313 /* IOPOLL retry should happen for io-wq threads */
3314 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3316 /* no retry on NONBLOCK nor RWF_NOWAIT */
3317 if (req->flags & REQ_F_NOWAIT)
3319 /* some cases will consume bytes even on error returns */
3320 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3322 } else if (ret == -EIOCBQUEUED) {
3324 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3325 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3326 /* read all, failed, already did sync or don't want to retry */
3330 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3335 rw = req->async_data;
3336 /* now use our persistent iterator, if we aren't already */
3341 rw->bytes_done += ret;
3342 /* if we can retry, do so with the callbacks armed */
3343 if (!io_rw_should_retry(req)) {
3344 kiocb->ki_flags &= ~IOCB_WAITQ;
3349 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3350 * we get -EIOCBQUEUED, then we'll get a notification when the
3351 * desired page gets unlocked. We can also get a partial read
3352 * here, and if we do, then just retry at the new offset.
3354 ret = io_iter_do_read(req, iter);
3355 if (ret == -EIOCBQUEUED)
3357 /* we got some bytes, but not all. retry. */
3358 kiocb->ki_flags &= ~IOCB_WAITQ;
3359 } while (ret > 0 && ret < io_size);
3361 kiocb_done(kiocb, ret, issue_flags);
3363 /* it's faster to check here then delegate to kfree */
3369 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3371 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3373 return io_prep_rw(req, sqe);
3376 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3378 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3379 struct kiocb *kiocb = &req->rw.kiocb;
3380 struct iov_iter __iter, *iter = &__iter;
3381 struct io_async_rw *rw = req->async_data;
3382 ssize_t ret, ret2, io_size;
3383 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3389 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3393 io_size = iov_iter_count(iter);
3394 req->result = io_size;
3396 /* Ensure we clear previously set non-block flag */
3397 if (!force_nonblock)
3398 kiocb->ki_flags &= ~IOCB_NOWAIT;
3400 kiocb->ki_flags |= IOCB_NOWAIT;
3402 /* If the file doesn't support async, just async punt */
3403 if (force_nonblock && !io_file_supports_async(req, WRITE))
3406 /* file path doesn't support NOWAIT for non-direct_IO */
3407 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3408 (req->flags & REQ_F_ISREG))
3411 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3416 * Open-code file_start_write here to grab freeze protection,
3417 * which will be released by another thread in
3418 * io_complete_rw(). Fool lockdep by telling it the lock got
3419 * released so that it doesn't complain about the held lock when
3420 * we return to userspace.
3422 if (req->flags & REQ_F_ISREG) {
3423 sb_start_write(file_inode(req->file)->i_sb);
3424 __sb_writers_release(file_inode(req->file)->i_sb,
3427 kiocb->ki_flags |= IOCB_WRITE;
3429 if (req->file->f_op->write_iter)
3430 ret2 = call_write_iter(req->file, kiocb, iter);
3431 else if (req->file->f_op->write)
3432 ret2 = loop_rw_iter(WRITE, req, iter);
3436 if (req->flags & REQ_F_REISSUE) {
3437 req->flags &= ~REQ_F_REISSUE;
3442 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3443 * retry them without IOCB_NOWAIT.
3445 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3447 /* no retry on NONBLOCK nor RWF_NOWAIT */
3448 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3450 if (!force_nonblock || ret2 != -EAGAIN) {
3451 /* IOPOLL retry should happen for io-wq threads */
3452 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3455 kiocb_done(kiocb, ret2, issue_flags);
3458 /* some cases will consume bytes even on error returns */
3459 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3460 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3461 return ret ?: -EAGAIN;
3464 /* it's reportedly faster than delegating the null check to kfree() */
3470 static int io_renameat_prep(struct io_kiocb *req,
3471 const struct io_uring_sqe *sqe)
3473 struct io_rename *ren = &req->rename;
3474 const char __user *oldf, *newf;
3476 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3478 if (sqe->ioprio || sqe->buf_index)
3480 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3483 ren->old_dfd = READ_ONCE(sqe->fd);
3484 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3485 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3486 ren->new_dfd = READ_ONCE(sqe->len);
3487 ren->flags = READ_ONCE(sqe->rename_flags);
3489 ren->oldpath = getname(oldf);
3490 if (IS_ERR(ren->oldpath))
3491 return PTR_ERR(ren->oldpath);
3493 ren->newpath = getname(newf);
3494 if (IS_ERR(ren->newpath)) {
3495 putname(ren->oldpath);
3496 return PTR_ERR(ren->newpath);
3499 req->flags |= REQ_F_NEED_CLEANUP;
3503 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3505 struct io_rename *ren = &req->rename;
3508 if (issue_flags & IO_URING_F_NONBLOCK)
3511 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3512 ren->newpath, ren->flags);
3514 req->flags &= ~REQ_F_NEED_CLEANUP;
3517 io_req_complete(req, ret);
3521 static int io_unlinkat_prep(struct io_kiocb *req,
3522 const struct io_uring_sqe *sqe)
3524 struct io_unlink *un = &req->unlink;
3525 const char __user *fname;
3527 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3529 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3531 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3534 un->dfd = READ_ONCE(sqe->fd);
3536 un->flags = READ_ONCE(sqe->unlink_flags);
3537 if (un->flags & ~AT_REMOVEDIR)
3540 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3541 un->filename = getname(fname);
3542 if (IS_ERR(un->filename))
3543 return PTR_ERR(un->filename);
3545 req->flags |= REQ_F_NEED_CLEANUP;
3549 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3551 struct io_unlink *un = &req->unlink;
3554 if (issue_flags & IO_URING_F_NONBLOCK)
3557 if (un->flags & AT_REMOVEDIR)
3558 ret = do_rmdir(un->dfd, un->filename);
3560 ret = do_unlinkat(un->dfd, un->filename);
3562 req->flags &= ~REQ_F_NEED_CLEANUP;
3565 io_req_complete(req, ret);
3569 static int io_shutdown_prep(struct io_kiocb *req,
3570 const struct io_uring_sqe *sqe)
3572 #if defined(CONFIG_NET)
3573 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3575 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3579 req->shutdown.how = READ_ONCE(sqe->len);
3586 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3588 #if defined(CONFIG_NET)
3589 struct socket *sock;
3592 if (issue_flags & IO_URING_F_NONBLOCK)
3595 sock = sock_from_file(req->file);
3596 if (unlikely(!sock))
3599 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3602 io_req_complete(req, ret);
3609 static int __io_splice_prep(struct io_kiocb *req,
3610 const struct io_uring_sqe *sqe)
3612 struct io_splice *sp = &req->splice;
3613 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3615 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3619 sp->len = READ_ONCE(sqe->len);
3620 sp->flags = READ_ONCE(sqe->splice_flags);
3622 if (unlikely(sp->flags & ~valid_flags))
3625 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3626 (sp->flags & SPLICE_F_FD_IN_FIXED));
3629 req->flags |= REQ_F_NEED_CLEANUP;
3633 static int io_tee_prep(struct io_kiocb *req,
3634 const struct io_uring_sqe *sqe)
3636 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3638 return __io_splice_prep(req, sqe);
3641 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3643 struct io_splice *sp = &req->splice;
3644 struct file *in = sp->file_in;
3645 struct file *out = sp->file_out;
3646 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3649 if (issue_flags & IO_URING_F_NONBLOCK)
3652 ret = do_tee(in, out, sp->len, flags);
3654 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3656 req->flags &= ~REQ_F_NEED_CLEANUP;
3660 io_req_complete(req, ret);
3664 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3666 struct io_splice *sp = &req->splice;
3668 sp->off_in = READ_ONCE(sqe->splice_off_in);
3669 sp->off_out = READ_ONCE(sqe->off);
3670 return __io_splice_prep(req, sqe);
3673 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3675 struct io_splice *sp = &req->splice;
3676 struct file *in = sp->file_in;
3677 struct file *out = sp->file_out;
3678 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3679 loff_t *poff_in, *poff_out;
3682 if (issue_flags & IO_URING_F_NONBLOCK)
3685 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3686 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3689 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3691 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3693 req->flags &= ~REQ_F_NEED_CLEANUP;
3697 io_req_complete(req, ret);
3702 * IORING_OP_NOP just posts a completion event, nothing else.
3704 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3706 struct io_ring_ctx *ctx = req->ctx;
3708 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3711 __io_req_complete(req, issue_flags, 0, 0);
3715 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3717 struct io_ring_ctx *ctx = req->ctx;
3722 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3724 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3727 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3728 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3731 req->sync.off = READ_ONCE(sqe->off);
3732 req->sync.len = READ_ONCE(sqe->len);
3736 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3738 loff_t end = req->sync.off + req->sync.len;
3741 /* fsync always requires a blocking context */
3742 if (issue_flags & IO_URING_F_NONBLOCK)
3745 ret = vfs_fsync_range(req->file, req->sync.off,
3746 end > 0 ? end : LLONG_MAX,
3747 req->sync.flags & IORING_FSYNC_DATASYNC);
3750 io_req_complete(req, ret);
3754 static int io_fallocate_prep(struct io_kiocb *req,
3755 const struct io_uring_sqe *sqe)
3757 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3759 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3762 req->sync.off = READ_ONCE(sqe->off);
3763 req->sync.len = READ_ONCE(sqe->addr);
3764 req->sync.mode = READ_ONCE(sqe->len);
3768 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3772 /* fallocate always requiring blocking context */
3773 if (issue_flags & IO_URING_F_NONBLOCK)
3775 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3779 io_req_complete(req, ret);
3783 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3785 const char __user *fname;
3788 if (unlikely(sqe->ioprio || sqe->buf_index))
3790 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3793 /* open.how should be already initialised */
3794 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3795 req->open.how.flags |= O_LARGEFILE;
3797 req->open.dfd = READ_ONCE(sqe->fd);
3798 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3799 req->open.filename = getname(fname);
3800 if (IS_ERR(req->open.filename)) {
3801 ret = PTR_ERR(req->open.filename);
3802 req->open.filename = NULL;
3805 req->open.nofile = rlimit(RLIMIT_NOFILE);
3806 req->flags |= REQ_F_NEED_CLEANUP;
3810 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3814 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3816 mode = READ_ONCE(sqe->len);
3817 flags = READ_ONCE(sqe->open_flags);
3818 req->open.how = build_open_how(flags, mode);
3819 return __io_openat_prep(req, sqe);
3822 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3824 struct open_how __user *how;
3828 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3830 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3831 len = READ_ONCE(sqe->len);
3832 if (len < OPEN_HOW_SIZE_VER0)
3835 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3840 return __io_openat_prep(req, sqe);
3843 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3845 struct open_flags op;
3848 bool resolve_nonblock;
3851 ret = build_open_flags(&req->open.how, &op);
3854 nonblock_set = op.open_flag & O_NONBLOCK;
3855 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3856 if (issue_flags & IO_URING_F_NONBLOCK) {
3858 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3859 * it'll always -EAGAIN
3861 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3863 op.lookup_flags |= LOOKUP_CACHED;
3864 op.open_flag |= O_NONBLOCK;
3867 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3871 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3874 * We could hang on to this 'fd' on retrying, but seems like
3875 * marginal gain for something that is now known to be a slower
3876 * path. So just put it, and we'll get a new one when we retry.
3880 ret = PTR_ERR(file);
3881 /* only retry if RESOLVE_CACHED wasn't already set by application */
3882 if (ret == -EAGAIN &&
3883 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3888 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3889 file->f_flags &= ~O_NONBLOCK;
3890 fsnotify_open(file);
3891 fd_install(ret, file);
3893 putname(req->open.filename);
3894 req->flags &= ~REQ_F_NEED_CLEANUP;
3897 __io_req_complete(req, issue_flags, ret, 0);
3901 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3903 return io_openat2(req, issue_flags);
3906 static int io_remove_buffers_prep(struct io_kiocb *req,
3907 const struct io_uring_sqe *sqe)
3909 struct io_provide_buf *p = &req->pbuf;
3912 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3915 tmp = READ_ONCE(sqe->fd);
3916 if (!tmp || tmp > USHRT_MAX)
3919 memset(p, 0, sizeof(*p));
3921 p->bgid = READ_ONCE(sqe->buf_group);
3925 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3926 int bgid, unsigned nbufs)
3930 /* shouldn't happen */
3934 /* the head kbuf is the list itself */
3935 while (!list_empty(&buf->list)) {
3936 struct io_buffer *nxt;
3938 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3939 list_del(&nxt->list);
3946 xa_erase(&ctx->io_buffers, bgid);
3951 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3953 struct io_provide_buf *p = &req->pbuf;
3954 struct io_ring_ctx *ctx = req->ctx;
3955 struct io_buffer *head;
3957 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3959 io_ring_submit_lock(ctx, !force_nonblock);
3961 lockdep_assert_held(&ctx->uring_lock);
3964 head = xa_load(&ctx->io_buffers, p->bgid);
3966 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3970 /* complete before unlock, IOPOLL may need the lock */
3971 __io_req_complete(req, issue_flags, ret, 0);
3972 io_ring_submit_unlock(ctx, !force_nonblock);
3976 static int io_provide_buffers_prep(struct io_kiocb *req,
3977 const struct io_uring_sqe *sqe)
3979 unsigned long size, tmp_check;
3980 struct io_provide_buf *p = &req->pbuf;
3983 if (sqe->ioprio || sqe->rw_flags)
3986 tmp = READ_ONCE(sqe->fd);
3987 if (!tmp || tmp > USHRT_MAX)
3990 p->addr = READ_ONCE(sqe->addr);
3991 p->len = READ_ONCE(sqe->len);
3993 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3996 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3999 size = (unsigned long)p->len * p->nbufs;
4000 if (!access_ok(u64_to_user_ptr(p->addr), size))
4003 p->bgid = READ_ONCE(sqe->buf_group);
4004 tmp = READ_ONCE(sqe->off);
4005 if (tmp > USHRT_MAX)
4011 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4013 struct io_buffer *buf;
4014 u64 addr = pbuf->addr;
4015 int i, bid = pbuf->bid;
4017 for (i = 0; i < pbuf->nbufs; i++) {
4018 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4023 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4028 INIT_LIST_HEAD(&buf->list);
4031 list_add_tail(&buf->list, &(*head)->list);
4035 return i ? i : -ENOMEM;
4038 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4040 struct io_provide_buf *p = &req->pbuf;
4041 struct io_ring_ctx *ctx = req->ctx;
4042 struct io_buffer *head, *list;
4044 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4046 io_ring_submit_lock(ctx, !force_nonblock);
4048 lockdep_assert_held(&ctx->uring_lock);
4050 list = head = xa_load(&ctx->io_buffers, p->bgid);
4052 ret = io_add_buffers(p, &head);
4053 if (ret >= 0 && !list) {
4054 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4056 __io_remove_buffers(ctx, head, p->bgid, -1U);
4060 /* complete before unlock, IOPOLL may need the lock */
4061 __io_req_complete(req, issue_flags, ret, 0);
4062 io_ring_submit_unlock(ctx, !force_nonblock);
4066 static int io_epoll_ctl_prep(struct io_kiocb *req,
4067 const struct io_uring_sqe *sqe)
4069 #if defined(CONFIG_EPOLL)
4070 if (sqe->ioprio || sqe->buf_index)
4072 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4075 req->epoll.epfd = READ_ONCE(sqe->fd);
4076 req->epoll.op = READ_ONCE(sqe->len);
4077 req->epoll.fd = READ_ONCE(sqe->off);
4079 if (ep_op_has_event(req->epoll.op)) {
4080 struct epoll_event __user *ev;
4082 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4083 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4093 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4095 #if defined(CONFIG_EPOLL)
4096 struct io_epoll *ie = &req->epoll;
4098 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4100 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4101 if (force_nonblock && ret == -EAGAIN)
4106 __io_req_complete(req, issue_flags, ret, 0);
4113 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4115 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4116 if (sqe->ioprio || sqe->buf_index || sqe->off)
4118 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4121 req->madvise.addr = READ_ONCE(sqe->addr);
4122 req->madvise.len = READ_ONCE(sqe->len);
4123 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4130 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4132 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4133 struct io_madvise *ma = &req->madvise;
4136 if (issue_flags & IO_URING_F_NONBLOCK)
4139 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4142 io_req_complete(req, ret);
4149 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4151 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4153 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4156 req->fadvise.offset = READ_ONCE(sqe->off);
4157 req->fadvise.len = READ_ONCE(sqe->len);
4158 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4162 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4164 struct io_fadvise *fa = &req->fadvise;
4167 if (issue_flags & IO_URING_F_NONBLOCK) {
4168 switch (fa->advice) {
4169 case POSIX_FADV_NORMAL:
4170 case POSIX_FADV_RANDOM:
4171 case POSIX_FADV_SEQUENTIAL:
4178 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4181 __io_req_complete(req, issue_flags, ret, 0);
4185 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4187 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4189 if (sqe->ioprio || sqe->buf_index)
4191 if (req->flags & REQ_F_FIXED_FILE)
4194 req->statx.dfd = READ_ONCE(sqe->fd);
4195 req->statx.mask = READ_ONCE(sqe->len);
4196 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4197 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4198 req->statx.flags = READ_ONCE(sqe->statx_flags);
4203 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4205 struct io_statx *ctx = &req->statx;
4208 if (issue_flags & IO_URING_F_NONBLOCK)
4211 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4216 io_req_complete(req, ret);
4220 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4222 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4224 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4225 sqe->rw_flags || sqe->buf_index)
4227 if (req->flags & REQ_F_FIXED_FILE)
4230 req->close.fd = READ_ONCE(sqe->fd);
4234 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4236 struct files_struct *files = current->files;
4237 struct io_close *close = &req->close;
4238 struct fdtable *fdt;
4239 struct file *file = NULL;
4242 spin_lock(&files->file_lock);
4243 fdt = files_fdtable(files);
4244 if (close->fd >= fdt->max_fds) {
4245 spin_unlock(&files->file_lock);
4248 file = fdt->fd[close->fd];
4249 if (!file || file->f_op == &io_uring_fops) {
4250 spin_unlock(&files->file_lock);
4255 /* if the file has a flush method, be safe and punt to async */
4256 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4257 spin_unlock(&files->file_lock);
4261 ret = __close_fd_get_file(close->fd, &file);
4262 spin_unlock(&files->file_lock);
4269 /* No ->flush() or already async, safely close from here */
4270 ret = filp_close(file, current->files);
4276 __io_req_complete(req, issue_flags, ret, 0);
4280 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4282 struct io_ring_ctx *ctx = req->ctx;
4284 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4286 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4289 req->sync.off = READ_ONCE(sqe->off);
4290 req->sync.len = READ_ONCE(sqe->len);
4291 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4295 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4299 /* sync_file_range always requires a blocking context */
4300 if (issue_flags & IO_URING_F_NONBLOCK)
4303 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4307 io_req_complete(req, ret);
4311 #if defined(CONFIG_NET)
4312 static int io_setup_async_msg(struct io_kiocb *req,
4313 struct io_async_msghdr *kmsg)
4315 struct io_async_msghdr *async_msg = req->async_data;
4319 if (io_alloc_async_data(req)) {
4320 kfree(kmsg->free_iov);
4323 async_msg = req->async_data;
4324 req->flags |= REQ_F_NEED_CLEANUP;
4325 memcpy(async_msg, kmsg, sizeof(*kmsg));
4326 async_msg->msg.msg_name = &async_msg->addr;
4327 /* if were using fast_iov, set it to the new one */
4328 if (!async_msg->free_iov)
4329 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4334 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4335 struct io_async_msghdr *iomsg)
4337 iomsg->msg.msg_name = &iomsg->addr;
4338 iomsg->free_iov = iomsg->fast_iov;
4339 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4340 req->sr_msg.msg_flags, &iomsg->free_iov);
4343 static int io_sendmsg_prep_async(struct io_kiocb *req)
4347 ret = io_sendmsg_copy_hdr(req, req->async_data);
4349 req->flags |= REQ_F_NEED_CLEANUP;
4353 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4355 struct io_sr_msg *sr = &req->sr_msg;
4357 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4360 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4361 sr->len = READ_ONCE(sqe->len);
4362 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4363 if (sr->msg_flags & MSG_DONTWAIT)
4364 req->flags |= REQ_F_NOWAIT;
4366 #ifdef CONFIG_COMPAT
4367 if (req->ctx->compat)
4368 sr->msg_flags |= MSG_CMSG_COMPAT;
4373 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4375 struct io_async_msghdr iomsg, *kmsg;
4376 struct socket *sock;
4381 sock = sock_from_file(req->file);
4382 if (unlikely(!sock))
4385 kmsg = req->async_data;
4387 ret = io_sendmsg_copy_hdr(req, &iomsg);
4393 flags = req->sr_msg.msg_flags;
4394 if (issue_flags & IO_URING_F_NONBLOCK)
4395 flags |= MSG_DONTWAIT;
4396 if (flags & MSG_WAITALL)
4397 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4399 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4400 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4401 return io_setup_async_msg(req, kmsg);
4402 if (ret == -ERESTARTSYS)
4405 /* fast path, check for non-NULL to avoid function call */
4407 kfree(kmsg->free_iov);
4408 req->flags &= ~REQ_F_NEED_CLEANUP;
4411 __io_req_complete(req, issue_flags, ret, 0);
4415 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4417 struct io_sr_msg *sr = &req->sr_msg;
4420 struct socket *sock;
4425 sock = sock_from_file(req->file);
4426 if (unlikely(!sock))
4429 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4433 msg.msg_name = NULL;
4434 msg.msg_control = NULL;
4435 msg.msg_controllen = 0;
4436 msg.msg_namelen = 0;
4438 flags = req->sr_msg.msg_flags;
4439 if (issue_flags & IO_URING_F_NONBLOCK)
4440 flags |= MSG_DONTWAIT;
4441 if (flags & MSG_WAITALL)
4442 min_ret = iov_iter_count(&msg.msg_iter);
4444 msg.msg_flags = flags;
4445 ret = sock_sendmsg(sock, &msg);
4446 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4448 if (ret == -ERESTARTSYS)
4453 __io_req_complete(req, issue_flags, ret, 0);
4457 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4458 struct io_async_msghdr *iomsg)
4460 struct io_sr_msg *sr = &req->sr_msg;
4461 struct iovec __user *uiov;
4465 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4466 &iomsg->uaddr, &uiov, &iov_len);
4470 if (req->flags & REQ_F_BUFFER_SELECT) {
4473 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4475 sr->len = iomsg->fast_iov[0].iov_len;
4476 iomsg->free_iov = NULL;
4478 iomsg->free_iov = iomsg->fast_iov;
4479 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4480 &iomsg->free_iov, &iomsg->msg.msg_iter,
4489 #ifdef CONFIG_COMPAT
4490 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4491 struct io_async_msghdr *iomsg)
4493 struct io_sr_msg *sr = &req->sr_msg;
4494 struct compat_iovec __user *uiov;
4499 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4504 uiov = compat_ptr(ptr);
4505 if (req->flags & REQ_F_BUFFER_SELECT) {
4506 compat_ssize_t clen;
4510 if (!access_ok(uiov, sizeof(*uiov)))
4512 if (__get_user(clen, &uiov->iov_len))
4517 iomsg->free_iov = NULL;
4519 iomsg->free_iov = iomsg->fast_iov;
4520 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4521 UIO_FASTIOV, &iomsg->free_iov,
4522 &iomsg->msg.msg_iter, true);
4531 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4532 struct io_async_msghdr *iomsg)
4534 iomsg->msg.msg_name = &iomsg->addr;
4536 #ifdef CONFIG_COMPAT
4537 if (req->ctx->compat)
4538 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4541 return __io_recvmsg_copy_hdr(req, iomsg);
4544 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4547 struct io_sr_msg *sr = &req->sr_msg;
4548 struct io_buffer *kbuf;
4550 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4555 req->flags |= REQ_F_BUFFER_SELECTED;
4559 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4561 return io_put_kbuf(req, req->sr_msg.kbuf);
4564 static int io_recvmsg_prep_async(struct io_kiocb *req)
4568 ret = io_recvmsg_copy_hdr(req, req->async_data);
4570 req->flags |= REQ_F_NEED_CLEANUP;
4574 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4576 struct io_sr_msg *sr = &req->sr_msg;
4578 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4581 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4582 sr->len = READ_ONCE(sqe->len);
4583 sr->bgid = READ_ONCE(sqe->buf_group);
4584 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4585 if (sr->msg_flags & MSG_DONTWAIT)
4586 req->flags |= REQ_F_NOWAIT;
4588 #ifdef CONFIG_COMPAT
4589 if (req->ctx->compat)
4590 sr->msg_flags |= MSG_CMSG_COMPAT;
4595 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4597 struct io_async_msghdr iomsg, *kmsg;
4598 struct socket *sock;
4599 struct io_buffer *kbuf;
4602 int ret, cflags = 0;
4603 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4605 sock = sock_from_file(req->file);
4606 if (unlikely(!sock))
4609 kmsg = req->async_data;
4611 ret = io_recvmsg_copy_hdr(req, &iomsg);
4617 if (req->flags & REQ_F_BUFFER_SELECT) {
4618 kbuf = io_recv_buffer_select(req, !force_nonblock);
4620 return PTR_ERR(kbuf);
4621 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4622 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4623 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4624 1, req->sr_msg.len);
4627 flags = req->sr_msg.msg_flags;
4629 flags |= MSG_DONTWAIT;
4630 if (flags & MSG_WAITALL)
4631 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4633 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4634 kmsg->uaddr, flags);
4635 if (force_nonblock && ret == -EAGAIN)
4636 return io_setup_async_msg(req, kmsg);
4637 if (ret == -ERESTARTSYS)
4640 if (req->flags & REQ_F_BUFFER_SELECTED)
4641 cflags = io_put_recv_kbuf(req);
4642 /* fast path, check for non-NULL to avoid function call */
4644 kfree(kmsg->free_iov);
4645 req->flags &= ~REQ_F_NEED_CLEANUP;
4646 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4648 __io_req_complete(req, issue_flags, ret, cflags);
4652 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4654 struct io_buffer *kbuf;
4655 struct io_sr_msg *sr = &req->sr_msg;
4657 void __user *buf = sr->buf;
4658 struct socket *sock;
4662 int ret, cflags = 0;
4663 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4665 sock = sock_from_file(req->file);
4666 if (unlikely(!sock))
4669 if (req->flags & REQ_F_BUFFER_SELECT) {
4670 kbuf = io_recv_buffer_select(req, !force_nonblock);
4672 return PTR_ERR(kbuf);
4673 buf = u64_to_user_ptr(kbuf->addr);
4676 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4680 msg.msg_name = NULL;
4681 msg.msg_control = NULL;
4682 msg.msg_controllen = 0;
4683 msg.msg_namelen = 0;
4684 msg.msg_iocb = NULL;
4687 flags = req->sr_msg.msg_flags;
4689 flags |= MSG_DONTWAIT;
4690 if (flags & MSG_WAITALL)
4691 min_ret = iov_iter_count(&msg.msg_iter);
4693 ret = sock_recvmsg(sock, &msg, flags);
4694 if (force_nonblock && ret == -EAGAIN)
4696 if (ret == -ERESTARTSYS)
4699 if (req->flags & REQ_F_BUFFER_SELECTED)
4700 cflags = io_put_recv_kbuf(req);
4701 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4703 __io_req_complete(req, issue_flags, ret, cflags);
4707 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4709 struct io_accept *accept = &req->accept;
4711 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4713 if (sqe->ioprio || sqe->len || sqe->buf_index)
4716 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4717 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4718 accept->flags = READ_ONCE(sqe->accept_flags);
4719 accept->nofile = rlimit(RLIMIT_NOFILE);
4723 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4725 struct io_accept *accept = &req->accept;
4726 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4727 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4730 if (req->file->f_flags & O_NONBLOCK)
4731 req->flags |= REQ_F_NOWAIT;
4733 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4734 accept->addr_len, accept->flags,
4736 if (ret == -EAGAIN && force_nonblock)
4739 if (ret == -ERESTARTSYS)
4743 __io_req_complete(req, issue_flags, ret, 0);
4747 static int io_connect_prep_async(struct io_kiocb *req)
4749 struct io_async_connect *io = req->async_data;
4750 struct io_connect *conn = &req->connect;
4752 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4755 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4757 struct io_connect *conn = &req->connect;
4759 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4761 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4764 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4765 conn->addr_len = READ_ONCE(sqe->addr2);
4769 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4771 struct io_async_connect __io, *io;
4772 unsigned file_flags;
4774 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4776 if (req->async_data) {
4777 io = req->async_data;
4779 ret = move_addr_to_kernel(req->connect.addr,
4780 req->connect.addr_len,
4787 file_flags = force_nonblock ? O_NONBLOCK : 0;
4789 ret = __sys_connect_file(req->file, &io->address,
4790 req->connect.addr_len, file_flags);
4791 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4792 if (req->async_data)
4794 if (io_alloc_async_data(req)) {
4798 memcpy(req->async_data, &__io, sizeof(__io));
4801 if (ret == -ERESTARTSYS)
4806 __io_req_complete(req, issue_flags, ret, 0);
4809 #else /* !CONFIG_NET */
4810 #define IO_NETOP_FN(op) \
4811 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4813 return -EOPNOTSUPP; \
4816 #define IO_NETOP_PREP(op) \
4818 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4820 return -EOPNOTSUPP; \
4823 #define IO_NETOP_PREP_ASYNC(op) \
4825 static int io_##op##_prep_async(struct io_kiocb *req) \
4827 return -EOPNOTSUPP; \
4830 IO_NETOP_PREP_ASYNC(sendmsg);
4831 IO_NETOP_PREP_ASYNC(recvmsg);
4832 IO_NETOP_PREP_ASYNC(connect);
4833 IO_NETOP_PREP(accept);
4836 #endif /* CONFIG_NET */
4838 struct io_poll_table {
4839 struct poll_table_struct pt;
4840 struct io_kiocb *req;
4845 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4846 __poll_t mask, io_req_tw_func_t func)
4848 /* for instances that support it check for an event match first: */
4849 if (mask && !(mask & poll->events))
4852 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4854 list_del_init(&poll->wait.entry);
4857 req->io_task_work.func = func;
4860 * If this fails, then the task is exiting. When a task exits, the
4861 * work gets canceled, so just cancel this request as well instead
4862 * of executing it. We can't safely execute it anyway, as we may not
4863 * have the needed state needed for it anyway.
4865 io_req_task_work_add(req);
4869 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4870 __acquires(&req->ctx->completion_lock)
4872 struct io_ring_ctx *ctx = req->ctx;
4874 if (unlikely(req->task->flags & PF_EXITING))
4875 WRITE_ONCE(poll->canceled, true);
4877 if (!req->result && !READ_ONCE(poll->canceled)) {
4878 struct poll_table_struct pt = { ._key = poll->events };
4880 req->result = vfs_poll(req->file, &pt) & poll->events;
4883 spin_lock_irq(&ctx->completion_lock);
4884 if (!req->result && !READ_ONCE(poll->canceled)) {
4885 add_wait_queue(poll->head, &poll->wait);
4892 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4894 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4895 if (req->opcode == IORING_OP_POLL_ADD)
4896 return req->async_data;
4897 return req->apoll->double_poll;
4900 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4902 if (req->opcode == IORING_OP_POLL_ADD)
4904 return &req->apoll->poll;
4907 static void io_poll_remove_double(struct io_kiocb *req)
4908 __must_hold(&req->ctx->completion_lock)
4910 struct io_poll_iocb *poll = io_poll_get_double(req);
4912 lockdep_assert_held(&req->ctx->completion_lock);
4914 if (poll && poll->head) {
4915 struct wait_queue_head *head = poll->head;
4917 spin_lock(&head->lock);
4918 list_del_init(&poll->wait.entry);
4919 if (poll->wait.private)
4922 spin_unlock(&head->lock);
4926 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4927 __must_hold(&req->ctx->completion_lock)
4929 struct io_ring_ctx *ctx = req->ctx;
4930 unsigned flags = IORING_CQE_F_MORE;
4933 if (READ_ONCE(req->poll.canceled)) {
4935 req->poll.events |= EPOLLONESHOT;
4937 error = mangle_poll(mask);
4939 if (req->poll.events & EPOLLONESHOT)
4941 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4942 req->poll.done = true;
4945 if (flags & IORING_CQE_F_MORE)
4948 io_commit_cqring(ctx);
4949 return !(flags & IORING_CQE_F_MORE);
4952 static void io_poll_task_func(struct io_kiocb *req)
4954 struct io_ring_ctx *ctx = req->ctx;
4955 struct io_kiocb *nxt;
4957 if (io_poll_rewait(req, &req->poll)) {
4958 spin_unlock_irq(&ctx->completion_lock);
4962 done = io_poll_complete(req, req->result);
4964 io_poll_remove_double(req);
4965 hash_del(&req->hash_node);
4968 add_wait_queue(req->poll.head, &req->poll.wait);
4970 spin_unlock_irq(&ctx->completion_lock);
4971 io_cqring_ev_posted(ctx);
4974 nxt = io_put_req_find_next(req);
4976 io_req_task_submit(nxt);
4981 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4982 int sync, void *key)
4984 struct io_kiocb *req = wait->private;
4985 struct io_poll_iocb *poll = io_poll_get_single(req);
4986 __poll_t mask = key_to_poll(key);
4988 /* for instances that support it check for an event match first: */
4989 if (mask && !(mask & poll->events))
4991 if (!(poll->events & EPOLLONESHOT))
4992 return poll->wait.func(&poll->wait, mode, sync, key);
4994 list_del_init(&wait->entry);
4999 spin_lock(&poll->head->lock);
5000 done = list_empty(&poll->wait.entry);
5002 list_del_init(&poll->wait.entry);
5003 /* make sure double remove sees this as being gone */
5004 wait->private = NULL;
5005 spin_unlock(&poll->head->lock);
5007 /* use wait func handler, so it matches the rq type */
5008 poll->wait.func(&poll->wait, mode, sync, key);
5015 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5016 wait_queue_func_t wake_func)
5020 poll->canceled = false;
5021 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5022 /* mask in events that we always want/need */
5023 poll->events = events | IO_POLL_UNMASK;
5024 INIT_LIST_HEAD(&poll->wait.entry);
5025 init_waitqueue_func_entry(&poll->wait, wake_func);
5028 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5029 struct wait_queue_head *head,
5030 struct io_poll_iocb **poll_ptr)
5032 struct io_kiocb *req = pt->req;
5035 * The file being polled uses multiple waitqueues for poll handling
5036 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5039 if (unlikely(pt->nr_entries)) {
5040 struct io_poll_iocb *poll_one = poll;
5042 /* already have a 2nd entry, fail a third attempt */
5044 pt->error = -EINVAL;
5048 * Can't handle multishot for double wait for now, turn it
5049 * into one-shot mode.
5051 if (!(poll_one->events & EPOLLONESHOT))
5052 poll_one->events |= EPOLLONESHOT;
5053 /* double add on the same waitqueue head, ignore */
5054 if (poll_one->head == head)
5056 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5058 pt->error = -ENOMEM;
5061 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5063 poll->wait.private = req;
5070 if (poll->events & EPOLLEXCLUSIVE)
5071 add_wait_queue_exclusive(head, &poll->wait);
5073 add_wait_queue(head, &poll->wait);
5076 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5077 struct poll_table_struct *p)
5079 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5080 struct async_poll *apoll = pt->req->apoll;
5082 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5085 static void io_async_task_func(struct io_kiocb *req)
5087 struct async_poll *apoll = req->apoll;
5088 struct io_ring_ctx *ctx = req->ctx;
5090 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5092 if (io_poll_rewait(req, &apoll->poll)) {
5093 spin_unlock_irq(&ctx->completion_lock);
5097 hash_del(&req->hash_node);
5098 io_poll_remove_double(req);
5099 spin_unlock_irq(&ctx->completion_lock);
5101 if (!READ_ONCE(apoll->poll.canceled))
5102 io_req_task_submit(req);
5104 io_req_complete_failed(req, -ECANCELED);
5107 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5110 struct io_kiocb *req = wait->private;
5111 struct io_poll_iocb *poll = &req->apoll->poll;
5113 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5116 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5119 static void io_poll_req_insert(struct io_kiocb *req)
5121 struct io_ring_ctx *ctx = req->ctx;
5122 struct hlist_head *list;
5124 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5125 hlist_add_head(&req->hash_node, list);
5128 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5129 struct io_poll_iocb *poll,
5130 struct io_poll_table *ipt, __poll_t mask,
5131 wait_queue_func_t wake_func)
5132 __acquires(&ctx->completion_lock)
5134 struct io_ring_ctx *ctx = req->ctx;
5135 bool cancel = false;
5137 INIT_HLIST_NODE(&req->hash_node);
5138 io_init_poll_iocb(poll, mask, wake_func);
5139 poll->file = req->file;
5140 poll->wait.private = req;
5142 ipt->pt._key = mask;
5145 ipt->nr_entries = 0;
5147 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5148 if (unlikely(!ipt->nr_entries) && !ipt->error)
5149 ipt->error = -EINVAL;
5151 spin_lock_irq(&ctx->completion_lock);
5152 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5153 io_poll_remove_double(req);
5154 if (likely(poll->head)) {
5155 spin_lock(&poll->head->lock);
5156 if (unlikely(list_empty(&poll->wait.entry))) {
5162 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5163 list_del_init(&poll->wait.entry);
5165 WRITE_ONCE(poll->canceled, true);
5166 else if (!poll->done) /* actually waiting for an event */
5167 io_poll_req_insert(req);
5168 spin_unlock(&poll->head->lock);
5180 static int io_arm_poll_handler(struct io_kiocb *req)
5182 const struct io_op_def *def = &io_op_defs[req->opcode];
5183 struct io_ring_ctx *ctx = req->ctx;
5184 struct async_poll *apoll;
5185 struct io_poll_table ipt;
5186 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5189 if (!req->file || !file_can_poll(req->file))
5190 return IO_APOLL_ABORTED;
5191 if (req->flags & REQ_F_POLLED)
5192 return IO_APOLL_ABORTED;
5193 if (!def->pollin && !def->pollout)
5194 return IO_APOLL_ABORTED;
5198 mask |= POLLIN | POLLRDNORM;
5200 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5201 if ((req->opcode == IORING_OP_RECVMSG) &&
5202 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5206 mask |= POLLOUT | POLLWRNORM;
5209 /* if we can't nonblock try, then no point in arming a poll handler */
5210 if (!io_file_supports_async(req, rw))
5211 return IO_APOLL_ABORTED;
5213 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5214 if (unlikely(!apoll))
5215 return IO_APOLL_ABORTED;
5216 apoll->double_poll = NULL;
5218 req->flags |= REQ_F_POLLED;
5219 ipt.pt._qproc = io_async_queue_proc;
5221 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5223 if (ret || ipt.error) {
5224 spin_unlock_irq(&ctx->completion_lock);
5226 return IO_APOLL_READY;
5227 return IO_APOLL_ABORTED;
5229 spin_unlock_irq(&ctx->completion_lock);
5230 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5231 mask, apoll->poll.events);
5235 static bool __io_poll_remove_one(struct io_kiocb *req,
5236 struct io_poll_iocb *poll, bool do_cancel)
5237 __must_hold(&req->ctx->completion_lock)
5239 bool do_complete = false;
5243 spin_lock(&poll->head->lock);
5245 WRITE_ONCE(poll->canceled, true);
5246 if (!list_empty(&poll->wait.entry)) {
5247 list_del_init(&poll->wait.entry);
5250 spin_unlock(&poll->head->lock);
5251 hash_del(&req->hash_node);
5255 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5256 __must_hold(&req->ctx->completion_lock)
5260 io_poll_remove_double(req);
5261 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5263 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5264 /* non-poll requests have submit ref still */
5270 static bool io_poll_remove_one(struct io_kiocb *req)
5271 __must_hold(&req->ctx->completion_lock)
5275 do_complete = io_poll_remove_waitqs(req);
5277 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5278 io_commit_cqring(req->ctx);
5280 io_put_req_deferred(req, 1);
5287 * Returns true if we found and killed one or more poll requests
5289 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5292 struct hlist_node *tmp;
5293 struct io_kiocb *req;
5296 spin_lock_irq(&ctx->completion_lock);
5297 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5298 struct hlist_head *list;
5300 list = &ctx->cancel_hash[i];
5301 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5302 if (io_match_task(req, tsk, cancel_all))
5303 posted += io_poll_remove_one(req);
5306 spin_unlock_irq(&ctx->completion_lock);
5309 io_cqring_ev_posted(ctx);
5314 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5316 __must_hold(&ctx->completion_lock)
5318 struct hlist_head *list;
5319 struct io_kiocb *req;
5321 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5322 hlist_for_each_entry(req, list, hash_node) {
5323 if (sqe_addr != req->user_data)
5325 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5332 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5334 __must_hold(&ctx->completion_lock)
5336 struct io_kiocb *req;
5338 req = io_poll_find(ctx, sqe_addr, poll_only);
5341 if (io_poll_remove_one(req))
5347 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5352 events = READ_ONCE(sqe->poll32_events);
5354 events = swahw32(events);
5356 if (!(flags & IORING_POLL_ADD_MULTI))
5357 events |= EPOLLONESHOT;
5358 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5361 static int io_poll_update_prep(struct io_kiocb *req,
5362 const struct io_uring_sqe *sqe)
5364 struct io_poll_update *upd = &req->poll_update;
5367 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5369 if (sqe->ioprio || sqe->buf_index)
5371 flags = READ_ONCE(sqe->len);
5372 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5373 IORING_POLL_ADD_MULTI))
5375 /* meaningless without update */
5376 if (flags == IORING_POLL_ADD_MULTI)
5379 upd->old_user_data = READ_ONCE(sqe->addr);
5380 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5381 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5383 upd->new_user_data = READ_ONCE(sqe->off);
5384 if (!upd->update_user_data && upd->new_user_data)
5386 if (upd->update_events)
5387 upd->events = io_poll_parse_events(sqe, flags);
5388 else if (sqe->poll32_events)
5394 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5397 struct io_kiocb *req = wait->private;
5398 struct io_poll_iocb *poll = &req->poll;
5400 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5403 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5404 struct poll_table_struct *p)
5406 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5408 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5411 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5413 struct io_poll_iocb *poll = &req->poll;
5416 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5418 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5420 flags = READ_ONCE(sqe->len);
5421 if (flags & ~IORING_POLL_ADD_MULTI)
5424 poll->events = io_poll_parse_events(sqe, flags);
5428 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5430 struct io_poll_iocb *poll = &req->poll;
5431 struct io_ring_ctx *ctx = req->ctx;
5432 struct io_poll_table ipt;
5435 ipt.pt._qproc = io_poll_queue_proc;
5437 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5440 if (mask) { /* no async, we'd stolen it */
5442 io_poll_complete(req, mask);
5444 spin_unlock_irq(&ctx->completion_lock);
5447 io_cqring_ev_posted(ctx);
5448 if (poll->events & EPOLLONESHOT)
5454 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5456 struct io_ring_ctx *ctx = req->ctx;
5457 struct io_kiocb *preq;
5461 spin_lock_irq(&ctx->completion_lock);
5462 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5468 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5470 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5475 * Don't allow racy completion with singleshot, as we cannot safely
5476 * update those. For multishot, if we're racing with completion, just
5477 * let completion re-add it.
5479 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5480 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5484 /* we now have a detached poll request. reissue. */
5488 spin_unlock_irq(&ctx->completion_lock);
5490 io_req_complete(req, ret);
5493 /* only mask one event flags, keep behavior flags */
5494 if (req->poll_update.update_events) {
5495 preq->poll.events &= ~0xffff;
5496 preq->poll.events |= req->poll_update.events & 0xffff;
5497 preq->poll.events |= IO_POLL_UNMASK;
5499 if (req->poll_update.update_user_data)
5500 preq->user_data = req->poll_update.new_user_data;
5501 spin_unlock_irq(&ctx->completion_lock);
5503 /* complete update request, we're done with it */
5504 io_req_complete(req, ret);
5507 ret = io_poll_add(preq, issue_flags);
5510 io_req_complete(preq, ret);
5516 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5518 struct io_timeout_data *data = container_of(timer,
5519 struct io_timeout_data, timer);
5520 struct io_kiocb *req = data->req;
5521 struct io_ring_ctx *ctx = req->ctx;
5522 unsigned long flags;
5524 spin_lock_irqsave(&ctx->completion_lock, flags);
5525 list_del_init(&req->timeout.list);
5526 atomic_set(&req->ctx->cq_timeouts,
5527 atomic_read(&req->ctx->cq_timeouts) + 1);
5529 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5530 io_commit_cqring(ctx);
5531 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5533 io_cqring_ev_posted(ctx);
5536 return HRTIMER_NORESTART;
5539 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5541 __must_hold(&ctx->completion_lock)
5543 struct io_timeout_data *io;
5544 struct io_kiocb *req;
5547 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5548 found = user_data == req->user_data;
5553 return ERR_PTR(-ENOENT);
5555 io = req->async_data;
5556 if (hrtimer_try_to_cancel(&io->timer) == -1)
5557 return ERR_PTR(-EALREADY);
5558 list_del_init(&req->timeout.list);
5562 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5563 __must_hold(&ctx->completion_lock)
5565 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5568 return PTR_ERR(req);
5571 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5572 io_put_req_deferred(req, 1);
5576 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5577 struct timespec64 *ts, enum hrtimer_mode mode)
5578 __must_hold(&ctx->completion_lock)
5580 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5581 struct io_timeout_data *data;
5584 return PTR_ERR(req);
5586 req->timeout.off = 0; /* noseq */
5587 data = req->async_data;
5588 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5589 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5590 data->timer.function = io_timeout_fn;
5591 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5595 static int io_timeout_remove_prep(struct io_kiocb *req,
5596 const struct io_uring_sqe *sqe)
5598 struct io_timeout_rem *tr = &req->timeout_rem;
5600 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5602 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5604 if (sqe->ioprio || sqe->buf_index || sqe->len)
5607 tr->addr = READ_ONCE(sqe->addr);
5608 tr->flags = READ_ONCE(sqe->timeout_flags);
5609 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5610 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5612 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5614 } else if (tr->flags) {
5615 /* timeout removal doesn't support flags */
5622 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5624 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5629 * Remove or update an existing timeout command
5631 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5633 struct io_timeout_rem *tr = &req->timeout_rem;
5634 struct io_ring_ctx *ctx = req->ctx;
5637 spin_lock_irq(&ctx->completion_lock);
5638 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5639 ret = io_timeout_cancel(ctx, tr->addr);
5641 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5642 io_translate_timeout_mode(tr->flags));
5644 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5645 io_commit_cqring(ctx);
5646 spin_unlock_irq(&ctx->completion_lock);
5647 io_cqring_ev_posted(ctx);
5654 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5655 bool is_timeout_link)
5657 struct io_timeout_data *data;
5659 u32 off = READ_ONCE(sqe->off);
5661 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5663 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5665 if (off && is_timeout_link)
5667 flags = READ_ONCE(sqe->timeout_flags);
5668 if (flags & ~IORING_TIMEOUT_ABS)
5671 req->timeout.off = off;
5672 if (unlikely(off && !req->ctx->off_timeout_used))
5673 req->ctx->off_timeout_used = true;
5675 if (!req->async_data && io_alloc_async_data(req))
5678 data = req->async_data;
5681 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5684 data->mode = io_translate_timeout_mode(flags);
5685 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5686 if (is_timeout_link)
5687 io_req_track_inflight(req);
5691 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5693 struct io_ring_ctx *ctx = req->ctx;
5694 struct io_timeout_data *data = req->async_data;
5695 struct list_head *entry;
5696 u32 tail, off = req->timeout.off;
5698 spin_lock_irq(&ctx->completion_lock);
5701 * sqe->off holds how many events that need to occur for this
5702 * timeout event to be satisfied. If it isn't set, then this is
5703 * a pure timeout request, sequence isn't used.
5705 if (io_is_timeout_noseq(req)) {
5706 entry = ctx->timeout_list.prev;
5710 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5711 req->timeout.target_seq = tail + off;
5713 /* Update the last seq here in case io_flush_timeouts() hasn't.
5714 * This is safe because ->completion_lock is held, and submissions
5715 * and completions are never mixed in the same ->completion_lock section.
5717 ctx->cq_last_tm_flush = tail;
5720 * Insertion sort, ensuring the first entry in the list is always
5721 * the one we need first.
5723 list_for_each_prev(entry, &ctx->timeout_list) {
5724 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5727 if (io_is_timeout_noseq(nxt))
5729 /* nxt.seq is behind @tail, otherwise would've been completed */
5730 if (off >= nxt->timeout.target_seq - tail)
5734 list_add(&req->timeout.list, entry);
5735 data->timer.function = io_timeout_fn;
5736 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5737 spin_unlock_irq(&ctx->completion_lock);
5741 struct io_cancel_data {
5742 struct io_ring_ctx *ctx;
5746 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5748 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5749 struct io_cancel_data *cd = data;
5751 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5754 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5755 struct io_ring_ctx *ctx)
5757 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5758 enum io_wq_cancel cancel_ret;
5761 if (!tctx || !tctx->io_wq)
5764 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5765 switch (cancel_ret) {
5766 case IO_WQ_CANCEL_OK:
5769 case IO_WQ_CANCEL_RUNNING:
5772 case IO_WQ_CANCEL_NOTFOUND:
5780 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5781 struct io_kiocb *req, __u64 sqe_addr,
5784 unsigned long flags;
5787 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5788 spin_lock_irqsave(&ctx->completion_lock, flags);
5791 ret = io_timeout_cancel(ctx, sqe_addr);
5794 ret = io_poll_cancel(ctx, sqe_addr, false);
5798 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5799 io_commit_cqring(ctx);
5800 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5801 io_cqring_ev_posted(ctx);
5807 static int io_async_cancel_prep(struct io_kiocb *req,
5808 const struct io_uring_sqe *sqe)
5810 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5812 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5814 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5817 req->cancel.addr = READ_ONCE(sqe->addr);
5821 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5823 struct io_ring_ctx *ctx = req->ctx;
5824 u64 sqe_addr = req->cancel.addr;
5825 struct io_tctx_node *node;
5828 /* tasks should wait for their io-wq threads, so safe w/o sync */
5829 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5830 spin_lock_irq(&ctx->completion_lock);
5833 ret = io_timeout_cancel(ctx, sqe_addr);
5836 ret = io_poll_cancel(ctx, sqe_addr, false);
5839 spin_unlock_irq(&ctx->completion_lock);
5841 /* slow path, try all io-wq's */
5842 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5844 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5845 struct io_uring_task *tctx = node->task->io_uring;
5847 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5851 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5853 spin_lock_irq(&ctx->completion_lock);
5855 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5856 io_commit_cqring(ctx);
5857 spin_unlock_irq(&ctx->completion_lock);
5858 io_cqring_ev_posted(ctx);
5866 static int io_rsrc_update_prep(struct io_kiocb *req,
5867 const struct io_uring_sqe *sqe)
5869 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5871 if (sqe->ioprio || sqe->rw_flags)
5874 req->rsrc_update.offset = READ_ONCE(sqe->off);
5875 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5876 if (!req->rsrc_update.nr_args)
5878 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5882 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5884 struct io_ring_ctx *ctx = req->ctx;
5885 struct io_uring_rsrc_update2 up;
5888 if (issue_flags & IO_URING_F_NONBLOCK)
5891 up.offset = req->rsrc_update.offset;
5892 up.data = req->rsrc_update.arg;
5897 mutex_lock(&ctx->uring_lock);
5898 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5899 &up, req->rsrc_update.nr_args);
5900 mutex_unlock(&ctx->uring_lock);
5904 __io_req_complete(req, issue_flags, ret, 0);
5908 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5910 switch (req->opcode) {
5913 case IORING_OP_READV:
5914 case IORING_OP_READ_FIXED:
5915 case IORING_OP_READ:
5916 return io_read_prep(req, sqe);
5917 case IORING_OP_WRITEV:
5918 case IORING_OP_WRITE_FIXED:
5919 case IORING_OP_WRITE:
5920 return io_write_prep(req, sqe);
5921 case IORING_OP_POLL_ADD:
5922 return io_poll_add_prep(req, sqe);
5923 case IORING_OP_POLL_REMOVE:
5924 return io_poll_update_prep(req, sqe);
5925 case IORING_OP_FSYNC:
5926 return io_fsync_prep(req, sqe);
5927 case IORING_OP_SYNC_FILE_RANGE:
5928 return io_sfr_prep(req, sqe);
5929 case IORING_OP_SENDMSG:
5930 case IORING_OP_SEND:
5931 return io_sendmsg_prep(req, sqe);
5932 case IORING_OP_RECVMSG:
5933 case IORING_OP_RECV:
5934 return io_recvmsg_prep(req, sqe);
5935 case IORING_OP_CONNECT:
5936 return io_connect_prep(req, sqe);
5937 case IORING_OP_TIMEOUT:
5938 return io_timeout_prep(req, sqe, false);
5939 case IORING_OP_TIMEOUT_REMOVE:
5940 return io_timeout_remove_prep(req, sqe);
5941 case IORING_OP_ASYNC_CANCEL:
5942 return io_async_cancel_prep(req, sqe);
5943 case IORING_OP_LINK_TIMEOUT:
5944 return io_timeout_prep(req, sqe, true);
5945 case IORING_OP_ACCEPT:
5946 return io_accept_prep(req, sqe);
5947 case IORING_OP_FALLOCATE:
5948 return io_fallocate_prep(req, sqe);
5949 case IORING_OP_OPENAT:
5950 return io_openat_prep(req, sqe);
5951 case IORING_OP_CLOSE:
5952 return io_close_prep(req, sqe);
5953 case IORING_OP_FILES_UPDATE:
5954 return io_rsrc_update_prep(req, sqe);
5955 case IORING_OP_STATX:
5956 return io_statx_prep(req, sqe);
5957 case IORING_OP_FADVISE:
5958 return io_fadvise_prep(req, sqe);
5959 case IORING_OP_MADVISE:
5960 return io_madvise_prep(req, sqe);
5961 case IORING_OP_OPENAT2:
5962 return io_openat2_prep(req, sqe);
5963 case IORING_OP_EPOLL_CTL:
5964 return io_epoll_ctl_prep(req, sqe);
5965 case IORING_OP_SPLICE:
5966 return io_splice_prep(req, sqe);
5967 case IORING_OP_PROVIDE_BUFFERS:
5968 return io_provide_buffers_prep(req, sqe);
5969 case IORING_OP_REMOVE_BUFFERS:
5970 return io_remove_buffers_prep(req, sqe);
5972 return io_tee_prep(req, sqe);
5973 case IORING_OP_SHUTDOWN:
5974 return io_shutdown_prep(req, sqe);
5975 case IORING_OP_RENAMEAT:
5976 return io_renameat_prep(req, sqe);
5977 case IORING_OP_UNLINKAT:
5978 return io_unlinkat_prep(req, sqe);
5981 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5986 static int io_req_prep_async(struct io_kiocb *req)
5988 if (!io_op_defs[req->opcode].needs_async_setup)
5990 if (WARN_ON_ONCE(req->async_data))
5992 if (io_alloc_async_data(req))
5995 switch (req->opcode) {
5996 case IORING_OP_READV:
5997 return io_rw_prep_async(req, READ);
5998 case IORING_OP_WRITEV:
5999 return io_rw_prep_async(req, WRITE);
6000 case IORING_OP_SENDMSG:
6001 return io_sendmsg_prep_async(req);
6002 case IORING_OP_RECVMSG:
6003 return io_recvmsg_prep_async(req);
6004 case IORING_OP_CONNECT:
6005 return io_connect_prep_async(req);
6007 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6012 static u32 io_get_sequence(struct io_kiocb *req)
6014 u32 seq = req->ctx->cached_sq_head;
6016 /* need original cached_sq_head, but it was increased for each req */
6017 io_for_each_link(req, req)
6022 static bool io_drain_req(struct io_kiocb *req)
6024 struct io_kiocb *pos;
6025 struct io_ring_ctx *ctx = req->ctx;
6026 struct io_defer_entry *de;
6031 * If we need to drain a request in the middle of a link, drain the
6032 * head request and the next request/link after the current link.
6033 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6034 * maintained for every request of our link.
6036 if (ctx->drain_next) {
6037 req->flags |= REQ_F_IO_DRAIN;
6038 ctx->drain_next = false;
6040 /* not interested in head, start from the first linked */
6041 io_for_each_link(pos, req->link) {
6042 if (pos->flags & REQ_F_IO_DRAIN) {
6043 ctx->drain_next = true;
6044 req->flags |= REQ_F_IO_DRAIN;
6049 /* Still need defer if there is pending req in defer list. */
6050 if (likely(list_empty_careful(&ctx->defer_list) &&
6051 !(req->flags & REQ_F_IO_DRAIN))) {
6052 ctx->drain_active = false;
6056 seq = io_get_sequence(req);
6057 /* Still a chance to pass the sequence check */
6058 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6061 ret = io_req_prep_async(req);
6064 io_prep_async_link(req);
6065 de = kmalloc(sizeof(*de), GFP_KERNEL);
6069 io_req_complete_failed(req, ret);
6073 spin_lock_irq(&ctx->completion_lock);
6074 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6075 spin_unlock_irq(&ctx->completion_lock);
6077 io_queue_async_work(req);
6081 trace_io_uring_defer(ctx, req, req->user_data);
6084 list_add_tail(&de->list, &ctx->defer_list);
6085 spin_unlock_irq(&ctx->completion_lock);
6089 static void io_clean_op(struct io_kiocb *req)
6091 if (req->flags & REQ_F_BUFFER_SELECTED) {
6092 switch (req->opcode) {
6093 case IORING_OP_READV:
6094 case IORING_OP_READ_FIXED:
6095 case IORING_OP_READ:
6096 kfree((void *)(unsigned long)req->rw.addr);
6098 case IORING_OP_RECVMSG:
6099 case IORING_OP_RECV:
6100 kfree(req->sr_msg.kbuf);
6105 if (req->flags & REQ_F_NEED_CLEANUP) {
6106 switch (req->opcode) {
6107 case IORING_OP_READV:
6108 case IORING_OP_READ_FIXED:
6109 case IORING_OP_READ:
6110 case IORING_OP_WRITEV:
6111 case IORING_OP_WRITE_FIXED:
6112 case IORING_OP_WRITE: {
6113 struct io_async_rw *io = req->async_data;
6115 kfree(io->free_iovec);
6118 case IORING_OP_RECVMSG:
6119 case IORING_OP_SENDMSG: {
6120 struct io_async_msghdr *io = req->async_data;
6122 kfree(io->free_iov);
6125 case IORING_OP_SPLICE:
6127 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6128 io_put_file(req->splice.file_in);
6130 case IORING_OP_OPENAT:
6131 case IORING_OP_OPENAT2:
6132 if (req->open.filename)
6133 putname(req->open.filename);
6135 case IORING_OP_RENAMEAT:
6136 putname(req->rename.oldpath);
6137 putname(req->rename.newpath);
6139 case IORING_OP_UNLINKAT:
6140 putname(req->unlink.filename);
6144 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6145 kfree(req->apoll->double_poll);
6149 if (req->flags & REQ_F_INFLIGHT) {
6150 struct io_uring_task *tctx = req->task->io_uring;
6152 atomic_dec(&tctx->inflight_tracked);
6154 if (req->flags & REQ_F_CREDS)
6155 put_cred(req->creds);
6157 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6160 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6162 struct io_ring_ctx *ctx = req->ctx;
6163 const struct cred *creds = NULL;
6166 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6167 creds = override_creds(req->creds);
6169 switch (req->opcode) {
6171 ret = io_nop(req, issue_flags);
6173 case IORING_OP_READV:
6174 case IORING_OP_READ_FIXED:
6175 case IORING_OP_READ:
6176 ret = io_read(req, issue_flags);
6178 case IORING_OP_WRITEV:
6179 case IORING_OP_WRITE_FIXED:
6180 case IORING_OP_WRITE:
6181 ret = io_write(req, issue_flags);
6183 case IORING_OP_FSYNC:
6184 ret = io_fsync(req, issue_flags);
6186 case IORING_OP_POLL_ADD:
6187 ret = io_poll_add(req, issue_flags);
6189 case IORING_OP_POLL_REMOVE:
6190 ret = io_poll_update(req, issue_flags);
6192 case IORING_OP_SYNC_FILE_RANGE:
6193 ret = io_sync_file_range(req, issue_flags);
6195 case IORING_OP_SENDMSG:
6196 ret = io_sendmsg(req, issue_flags);
6198 case IORING_OP_SEND:
6199 ret = io_send(req, issue_flags);
6201 case IORING_OP_RECVMSG:
6202 ret = io_recvmsg(req, issue_flags);
6204 case IORING_OP_RECV:
6205 ret = io_recv(req, issue_flags);
6207 case IORING_OP_TIMEOUT:
6208 ret = io_timeout(req, issue_flags);
6210 case IORING_OP_TIMEOUT_REMOVE:
6211 ret = io_timeout_remove(req, issue_flags);
6213 case IORING_OP_ACCEPT:
6214 ret = io_accept(req, issue_flags);
6216 case IORING_OP_CONNECT:
6217 ret = io_connect(req, issue_flags);
6219 case IORING_OP_ASYNC_CANCEL:
6220 ret = io_async_cancel(req, issue_flags);
6222 case IORING_OP_FALLOCATE:
6223 ret = io_fallocate(req, issue_flags);
6225 case IORING_OP_OPENAT:
6226 ret = io_openat(req, issue_flags);
6228 case IORING_OP_CLOSE:
6229 ret = io_close(req, issue_flags);
6231 case IORING_OP_FILES_UPDATE:
6232 ret = io_files_update(req, issue_flags);
6234 case IORING_OP_STATX:
6235 ret = io_statx(req, issue_flags);
6237 case IORING_OP_FADVISE:
6238 ret = io_fadvise(req, issue_flags);
6240 case IORING_OP_MADVISE:
6241 ret = io_madvise(req, issue_flags);
6243 case IORING_OP_OPENAT2:
6244 ret = io_openat2(req, issue_flags);
6246 case IORING_OP_EPOLL_CTL:
6247 ret = io_epoll_ctl(req, issue_flags);
6249 case IORING_OP_SPLICE:
6250 ret = io_splice(req, issue_flags);
6252 case IORING_OP_PROVIDE_BUFFERS:
6253 ret = io_provide_buffers(req, issue_flags);
6255 case IORING_OP_REMOVE_BUFFERS:
6256 ret = io_remove_buffers(req, issue_flags);
6259 ret = io_tee(req, issue_flags);
6261 case IORING_OP_SHUTDOWN:
6262 ret = io_shutdown(req, issue_flags);
6264 case IORING_OP_RENAMEAT:
6265 ret = io_renameat(req, issue_flags);
6267 case IORING_OP_UNLINKAT:
6268 ret = io_unlinkat(req, issue_flags);
6276 revert_creds(creds);
6279 /* If the op doesn't have a file, we're not polling for it */
6280 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6281 io_iopoll_req_issued(req);
6286 static void io_wq_submit_work(struct io_wq_work *work)
6288 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6289 struct io_kiocb *timeout;
6292 timeout = io_prep_linked_timeout(req);
6294 io_queue_linked_timeout(timeout);
6296 if (work->flags & IO_WQ_WORK_CANCEL)
6301 ret = io_issue_sqe(req, 0);
6303 * We can get EAGAIN for polled IO even though we're
6304 * forcing a sync submission from here, since we can't
6305 * wait for request slots on the block side.
6313 /* avoid locking problems by failing it from a clean context */
6315 /* io-wq is going to take one down */
6317 io_req_task_queue_fail(req, ret);
6321 #define FFS_ASYNC_READ 0x1UL
6322 #define FFS_ASYNC_WRITE 0x2UL
6324 #define FFS_ISREG 0x4UL
6326 #define FFS_ISREG 0x0UL
6328 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6330 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6333 struct io_fixed_file *table_l2;
6335 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6336 return &table_l2[i & IORING_FILE_TABLE_MASK];
6339 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6342 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6344 return (struct file *) (slot->file_ptr & FFS_MASK);
6347 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6349 unsigned long file_ptr = (unsigned long) file;
6351 if (__io_file_supports_async(file, READ))
6352 file_ptr |= FFS_ASYNC_READ;
6353 if (__io_file_supports_async(file, WRITE))
6354 file_ptr |= FFS_ASYNC_WRITE;
6355 if (S_ISREG(file_inode(file)->i_mode))
6356 file_ptr |= FFS_ISREG;
6357 file_slot->file_ptr = file_ptr;
6360 static struct file *io_file_get(struct io_submit_state *state,
6361 struct io_kiocb *req, int fd, bool fixed)
6363 struct io_ring_ctx *ctx = req->ctx;
6367 unsigned long file_ptr;
6369 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6371 fd = array_index_nospec(fd, ctx->nr_user_files);
6372 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6373 file = (struct file *) (file_ptr & FFS_MASK);
6374 file_ptr &= ~FFS_MASK;
6375 /* mask in overlapping REQ_F and FFS bits */
6376 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6377 io_req_set_rsrc_node(req);
6379 trace_io_uring_file_get(ctx, fd);
6380 file = __io_file_get(state, fd);
6382 /* we don't allow fixed io_uring files */
6383 if (file && unlikely(file->f_op == &io_uring_fops))
6384 io_req_track_inflight(req);
6390 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6392 struct io_timeout_data *data = container_of(timer,
6393 struct io_timeout_data, timer);
6394 struct io_kiocb *prev, *req = data->req;
6395 struct io_ring_ctx *ctx = req->ctx;
6396 unsigned long flags;
6398 spin_lock_irqsave(&ctx->completion_lock, flags);
6399 prev = req->timeout.head;
6400 req->timeout.head = NULL;
6403 * We don't expect the list to be empty, that will only happen if we
6404 * race with the completion of the linked work.
6407 io_remove_next_linked(prev);
6408 if (!req_ref_inc_not_zero(prev))
6411 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6414 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6415 io_put_req_deferred(prev, 1);
6416 io_put_req_deferred(req, 1);
6418 io_req_complete_post(req, -ETIME, 0);
6420 return HRTIMER_NORESTART;
6423 static void io_queue_linked_timeout(struct io_kiocb *req)
6425 struct io_ring_ctx *ctx = req->ctx;
6427 spin_lock_irq(&ctx->completion_lock);
6429 * If the back reference is NULL, then our linked request finished
6430 * before we got a chance to setup the timer
6432 if (req->timeout.head) {
6433 struct io_timeout_data *data = req->async_data;
6435 data->timer.function = io_link_timeout_fn;
6436 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6439 spin_unlock_irq(&ctx->completion_lock);
6440 /* drop submission reference */
6444 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6446 struct io_kiocb *nxt = req->link;
6448 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6449 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6452 nxt->timeout.head = req;
6453 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6454 req->flags |= REQ_F_LINK_TIMEOUT;
6458 static void __io_queue_sqe(struct io_kiocb *req)
6460 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6464 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6467 * We async punt it if the file wasn't marked NOWAIT, or if the file
6468 * doesn't support non-blocking read/write attempts
6471 /* drop submission reference */
6472 if (req->flags & REQ_F_COMPLETE_INLINE) {
6473 struct io_ring_ctx *ctx = req->ctx;
6474 struct io_comp_state *cs = &ctx->submit_state.comp;
6476 cs->reqs[cs->nr++] = req;
6477 if (cs->nr == ARRAY_SIZE(cs->reqs))
6478 io_submit_flush_completions(ctx);
6482 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6483 switch (io_arm_poll_handler(req)) {
6484 case IO_APOLL_READY:
6486 case IO_APOLL_ABORTED:
6488 * Queued up for async execution, worker will release
6489 * submit reference when the iocb is actually submitted.
6491 io_queue_async_work(req);
6495 io_req_complete_failed(req, ret);
6498 io_queue_linked_timeout(linked_timeout);
6501 static inline void io_queue_sqe(struct io_kiocb *req)
6503 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6506 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6507 __io_queue_sqe(req);
6509 int ret = io_req_prep_async(req);
6512 io_req_complete_failed(req, ret);
6514 io_queue_async_work(req);
6519 * Check SQE restrictions (opcode and flags).
6521 * Returns 'true' if SQE is allowed, 'false' otherwise.
6523 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6524 struct io_kiocb *req,
6525 unsigned int sqe_flags)
6527 if (likely(!ctx->restricted))
6530 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6533 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6534 ctx->restrictions.sqe_flags_required)
6537 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6538 ctx->restrictions.sqe_flags_required))
6544 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6545 const struct io_uring_sqe *sqe)
6547 struct io_submit_state *state;
6548 unsigned int sqe_flags;
6549 int personality, ret = 0;
6551 req->opcode = READ_ONCE(sqe->opcode);
6552 /* same numerical values with corresponding REQ_F_*, safe to copy */
6553 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6554 req->user_data = READ_ONCE(sqe->user_data);
6556 req->fixed_rsrc_refs = NULL;
6557 /* one is dropped after submission, the other at completion */
6558 atomic_set(&req->refs, 2);
6559 req->task = current;
6561 /* enforce forwards compatibility on users */
6562 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6564 if (unlikely(req->opcode >= IORING_OP_LAST))
6566 if (!io_check_restriction(ctx, req, sqe_flags))
6569 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6570 !io_op_defs[req->opcode].buffer_select)
6572 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6573 ctx->drain_active = true;
6575 personality = READ_ONCE(sqe->personality);
6577 req->creds = xa_load(&ctx->personalities, personality);
6580 get_cred(req->creds);
6581 req->flags |= REQ_F_CREDS;
6583 state = &ctx->submit_state;
6586 * Plug now if we have more than 1 IO left after this, and the target
6587 * is potentially a read/write to block based storage.
6589 if (!state->plug_started && state->ios_left > 1 &&
6590 io_op_defs[req->opcode].plug) {
6591 blk_start_plug(&state->plug);
6592 state->plug_started = true;
6595 if (io_op_defs[req->opcode].needs_file) {
6596 bool fixed = req->flags & REQ_F_FIXED_FILE;
6598 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6599 if (unlikely(!req->file))
6607 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6608 const struct io_uring_sqe *sqe)
6610 struct io_submit_link *link = &ctx->submit_state.link;
6613 ret = io_init_req(ctx, req, sqe);
6614 if (unlikely(ret)) {
6617 /* fail even hard links since we don't submit */
6618 req_set_fail(link->head);
6619 io_req_complete_failed(link->head, -ECANCELED);
6622 io_req_complete_failed(req, ret);
6626 ret = io_req_prep(req, sqe);
6630 /* don't need @sqe from now on */
6631 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6633 ctx->flags & IORING_SETUP_SQPOLL);
6636 * If we already have a head request, queue this one for async
6637 * submittal once the head completes. If we don't have a head but
6638 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6639 * submitted sync once the chain is complete. If none of those
6640 * conditions are true (normal request), then just queue it.
6643 struct io_kiocb *head = link->head;
6645 ret = io_req_prep_async(req);
6648 trace_io_uring_link(ctx, req, head);
6649 link->last->link = req;
6652 /* last request of a link, enqueue the link */
6653 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6658 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6670 * Batched submission is done, ensure local IO is flushed out.
6672 static void io_submit_state_end(struct io_submit_state *state,
6673 struct io_ring_ctx *ctx)
6675 if (state->link.head)
6676 io_queue_sqe(state->link.head);
6678 io_submit_flush_completions(ctx);
6679 if (state->plug_started)
6680 blk_finish_plug(&state->plug);
6681 io_state_file_put(state);
6685 * Start submission side cache.
6687 static void io_submit_state_start(struct io_submit_state *state,
6688 unsigned int max_ios)
6690 state->plug_started = false;
6691 state->ios_left = max_ios;
6692 /* set only head, no need to init link_last in advance */
6693 state->link.head = NULL;
6696 static void io_commit_sqring(struct io_ring_ctx *ctx)
6698 struct io_rings *rings = ctx->rings;
6701 * Ensure any loads from the SQEs are done at this point,
6702 * since once we write the new head, the application could
6703 * write new data to them.
6705 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6709 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6710 * that is mapped by userspace. This means that care needs to be taken to
6711 * ensure that reads are stable, as we cannot rely on userspace always
6712 * being a good citizen. If members of the sqe are validated and then later
6713 * used, it's important that those reads are done through READ_ONCE() to
6714 * prevent a re-load down the line.
6716 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6718 unsigned head, mask = ctx->sq_entries - 1;
6719 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6722 * The cached sq head (or cq tail) serves two purposes:
6724 * 1) allows us to batch the cost of updating the user visible
6726 * 2) allows the kernel side to track the head on its own, even
6727 * though the application is the one updating it.
6729 head = READ_ONCE(ctx->sq_array[sq_idx]);
6730 if (likely(head < ctx->sq_entries))
6731 return &ctx->sq_sqes[head];
6733 /* drop invalid entries */
6735 WRITE_ONCE(ctx->rings->sq_dropped,
6736 READ_ONCE(ctx->rings->sq_dropped) + 1);
6740 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6742 struct io_uring_task *tctx;
6745 /* make sure SQ entry isn't read before tail */
6746 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6747 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6750 tctx = current->io_uring;
6751 tctx->cached_refs -= nr;
6752 if (unlikely(tctx->cached_refs < 0)) {
6753 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6755 percpu_counter_add(&tctx->inflight, refill);
6756 refcount_add(refill, ¤t->usage);
6757 tctx->cached_refs += refill;
6759 io_submit_state_start(&ctx->submit_state, nr);
6761 while (submitted < nr) {
6762 const struct io_uring_sqe *sqe;
6763 struct io_kiocb *req;
6765 req = io_alloc_req(ctx);
6766 if (unlikely(!req)) {
6768 submitted = -EAGAIN;
6771 sqe = io_get_sqe(ctx);
6772 if (unlikely(!sqe)) {
6773 kmem_cache_free(req_cachep, req);
6776 /* will complete beyond this point, count as submitted */
6778 if (io_submit_sqe(ctx, req, sqe))
6782 if (unlikely(submitted != nr)) {
6783 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6784 int unused = nr - ref_used;
6786 current->io_uring->cached_refs += unused;
6787 percpu_ref_put_many(&ctx->refs, unused);
6790 io_submit_state_end(&ctx->submit_state, ctx);
6791 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6792 io_commit_sqring(ctx);
6797 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6799 return READ_ONCE(sqd->state);
6802 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6804 /* Tell userspace we may need a wakeup call */
6805 spin_lock_irq(&ctx->completion_lock);
6806 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6807 spin_unlock_irq(&ctx->completion_lock);
6810 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6812 spin_lock_irq(&ctx->completion_lock);
6813 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6814 spin_unlock_irq(&ctx->completion_lock);
6817 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6819 unsigned int to_submit;
6822 to_submit = io_sqring_entries(ctx);
6823 /* if we're handling multiple rings, cap submit size for fairness */
6824 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6825 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6827 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6828 unsigned nr_events = 0;
6829 const struct cred *creds = NULL;
6831 if (ctx->sq_creds != current_cred())
6832 creds = override_creds(ctx->sq_creds);
6834 mutex_lock(&ctx->uring_lock);
6835 if (!list_empty(&ctx->iopoll_list))
6836 io_do_iopoll(ctx, &nr_events, 0, true);
6839 * Don't submit if refs are dying, good for io_uring_register(),
6840 * but also it is relied upon by io_ring_exit_work()
6842 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6843 !(ctx->flags & IORING_SETUP_R_DISABLED))
6844 ret = io_submit_sqes(ctx, to_submit);
6845 mutex_unlock(&ctx->uring_lock);
6847 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6848 wake_up(&ctx->sqo_sq_wait);
6850 revert_creds(creds);
6856 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6858 struct io_ring_ctx *ctx;
6859 unsigned sq_thread_idle = 0;
6861 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6862 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6863 sqd->sq_thread_idle = sq_thread_idle;
6866 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6868 bool did_sig = false;
6869 struct ksignal ksig;
6871 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6872 signal_pending(current)) {
6873 mutex_unlock(&sqd->lock);
6874 if (signal_pending(current))
6875 did_sig = get_signal(&ksig);
6877 mutex_lock(&sqd->lock);
6879 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6882 static int io_sq_thread(void *data)
6884 struct io_sq_data *sqd = data;
6885 struct io_ring_ctx *ctx;
6886 unsigned long timeout = 0;
6887 char buf[TASK_COMM_LEN];
6890 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6891 set_task_comm(current, buf);
6893 if (sqd->sq_cpu != -1)
6894 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6896 set_cpus_allowed_ptr(current, cpu_online_mask);
6897 current->flags |= PF_NO_SETAFFINITY;
6899 mutex_lock(&sqd->lock);
6901 bool cap_entries, sqt_spin = false;
6903 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6904 if (io_sqd_handle_event(sqd))
6906 timeout = jiffies + sqd->sq_thread_idle;
6909 cap_entries = !list_is_singular(&sqd->ctx_list);
6910 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6911 int ret = __io_sq_thread(ctx, cap_entries);
6913 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6916 if (io_run_task_work())
6919 if (sqt_spin || !time_after(jiffies, timeout)) {
6922 timeout = jiffies + sqd->sq_thread_idle;
6926 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6927 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6928 bool needs_sched = true;
6930 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6931 io_ring_set_wakeup_flag(ctx);
6933 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6934 !list_empty_careful(&ctx->iopoll_list)) {
6935 needs_sched = false;
6938 if (io_sqring_entries(ctx)) {
6939 needs_sched = false;
6945 mutex_unlock(&sqd->lock);
6947 mutex_lock(&sqd->lock);
6949 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6950 io_ring_clear_wakeup_flag(ctx);
6953 finish_wait(&sqd->wait, &wait);
6954 timeout = jiffies + sqd->sq_thread_idle;
6957 io_uring_cancel_generic(true, sqd);
6959 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6960 io_ring_set_wakeup_flag(ctx);
6962 mutex_unlock(&sqd->lock);
6964 complete(&sqd->exited);
6968 struct io_wait_queue {
6969 struct wait_queue_entry wq;
6970 struct io_ring_ctx *ctx;
6972 unsigned nr_timeouts;
6975 static inline bool io_should_wake(struct io_wait_queue *iowq)
6977 struct io_ring_ctx *ctx = iowq->ctx;
6980 * Wake up if we have enough events, or if a timeout occurred since we
6981 * started waiting. For timeouts, we always want to return to userspace,
6982 * regardless of event count.
6984 return io_cqring_events(ctx) >= iowq->to_wait ||
6985 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6988 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6989 int wake_flags, void *key)
6991 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6995 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6996 * the task, and the next invocation will do it.
6998 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
6999 return autoremove_wake_function(curr, mode, wake_flags, key);
7003 static int io_run_task_work_sig(void)
7005 if (io_run_task_work())
7007 if (!signal_pending(current))
7009 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7010 return -ERESTARTSYS;
7014 /* when returns >0, the caller should retry */
7015 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7016 struct io_wait_queue *iowq,
7017 signed long *timeout)
7021 /* make sure we run task_work before checking for signals */
7022 ret = io_run_task_work_sig();
7023 if (ret || io_should_wake(iowq))
7025 /* let the caller flush overflows, retry */
7026 if (test_bit(0, &ctx->check_cq_overflow))
7029 *timeout = schedule_timeout(*timeout);
7030 return !*timeout ? -ETIME : 1;
7034 * Wait until events become available, if we don't already have some. The
7035 * application must reap them itself, as they reside on the shared cq ring.
7037 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7038 const sigset_t __user *sig, size_t sigsz,
7039 struct __kernel_timespec __user *uts)
7041 struct io_wait_queue iowq = {
7044 .func = io_wake_function,
7045 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7048 .to_wait = min_events,
7050 struct io_rings *rings = ctx->rings;
7051 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7055 io_cqring_overflow_flush(ctx, false);
7056 if (io_cqring_events(ctx) >= min_events)
7058 if (!io_run_task_work())
7063 #ifdef CONFIG_COMPAT
7064 if (in_compat_syscall())
7065 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7069 ret = set_user_sigmask(sig, sigsz);
7076 struct timespec64 ts;
7078 if (get_timespec64(&ts, uts))
7080 timeout = timespec64_to_jiffies(&ts);
7083 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7084 trace_io_uring_cqring_wait(ctx, min_events);
7086 /* if we can't even flush overflow, don't wait for more */
7087 if (!io_cqring_overflow_flush(ctx, false)) {
7091 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7092 TASK_INTERRUPTIBLE);
7093 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7094 finish_wait(&ctx->cq_wait, &iowq.wq);
7098 restore_saved_sigmask_unless(ret == -EINTR);
7100 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7103 static void io_free_page_table(void **table, size_t size)
7105 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7107 for (i = 0; i < nr_tables; i++)
7112 static void **io_alloc_page_table(size_t size)
7114 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7115 size_t init_size = size;
7118 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7122 for (i = 0; i < nr_tables; i++) {
7123 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7125 table[i] = kzalloc(this_size, GFP_KERNEL);
7127 io_free_page_table(table, init_size);
7135 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7137 spin_lock_bh(&ctx->rsrc_ref_lock);
7140 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7142 spin_unlock_bh(&ctx->rsrc_ref_lock);
7145 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7147 percpu_ref_exit(&ref_node->refs);
7151 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7152 struct io_rsrc_data *data_to_kill)
7154 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7155 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7158 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7160 rsrc_node->rsrc_data = data_to_kill;
7161 io_rsrc_ref_lock(ctx);
7162 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7163 io_rsrc_ref_unlock(ctx);
7165 atomic_inc(&data_to_kill->refs);
7166 percpu_ref_kill(&rsrc_node->refs);
7167 ctx->rsrc_node = NULL;
7170 if (!ctx->rsrc_node) {
7171 ctx->rsrc_node = ctx->rsrc_backup_node;
7172 ctx->rsrc_backup_node = NULL;
7176 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7178 if (ctx->rsrc_backup_node)
7180 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7181 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7184 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7188 /* As we may drop ->uring_lock, other task may have started quiesce */
7192 data->quiesce = true;
7194 ret = io_rsrc_node_switch_start(ctx);
7197 io_rsrc_node_switch(ctx, data);
7199 /* kill initial ref, already quiesced if zero */
7200 if (atomic_dec_and_test(&data->refs))
7202 flush_delayed_work(&ctx->rsrc_put_work);
7203 ret = wait_for_completion_interruptible(&data->done);
7207 atomic_inc(&data->refs);
7208 /* wait for all works potentially completing data->done */
7209 flush_delayed_work(&ctx->rsrc_put_work);
7210 reinit_completion(&data->done);
7212 mutex_unlock(&ctx->uring_lock);
7213 ret = io_run_task_work_sig();
7214 mutex_lock(&ctx->uring_lock);
7216 data->quiesce = false;
7221 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7223 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7224 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7226 return &data->tags[table_idx][off];
7229 static void io_rsrc_data_free(struct io_rsrc_data *data)
7231 size_t size = data->nr * sizeof(data->tags[0][0]);
7234 io_free_page_table((void **)data->tags, size);
7238 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7239 u64 __user *utags, unsigned nr,
7240 struct io_rsrc_data **pdata)
7242 struct io_rsrc_data *data;
7246 data = kzalloc(sizeof(*data), GFP_KERNEL);
7249 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7257 data->do_put = do_put;
7260 for (i = 0; i < nr; i++) {
7261 u64 *tag_slot = io_get_tag_slot(data, i);
7263 if (copy_from_user(tag_slot, &utags[i],
7269 atomic_set(&data->refs, 1);
7270 init_completion(&data->done);
7274 io_rsrc_data_free(data);
7278 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7280 size_t size = nr_files * sizeof(struct io_fixed_file);
7282 table->files = (struct io_fixed_file **)io_alloc_page_table(size);
7283 return !!table->files;
7286 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7288 size_t size = nr_files * sizeof(struct io_fixed_file);
7290 io_free_page_table((void **)table->files, size);
7291 table->files = NULL;
7294 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7296 #if defined(CONFIG_UNIX)
7297 if (ctx->ring_sock) {
7298 struct sock *sock = ctx->ring_sock->sk;
7299 struct sk_buff *skb;
7301 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7307 for (i = 0; i < ctx->nr_user_files; i++) {
7310 file = io_file_from_index(ctx, i);
7315 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7316 io_rsrc_data_free(ctx->file_data);
7317 ctx->file_data = NULL;
7318 ctx->nr_user_files = 0;
7321 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7325 if (!ctx->file_data)
7327 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7329 __io_sqe_files_unregister(ctx);
7333 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7334 __releases(&sqd->lock)
7336 WARN_ON_ONCE(sqd->thread == current);
7339 * Do the dance but not conditional clear_bit() because it'd race with
7340 * other threads incrementing park_pending and setting the bit.
7342 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7343 if (atomic_dec_return(&sqd->park_pending))
7344 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7345 mutex_unlock(&sqd->lock);
7348 static void io_sq_thread_park(struct io_sq_data *sqd)
7349 __acquires(&sqd->lock)
7351 WARN_ON_ONCE(sqd->thread == current);
7353 atomic_inc(&sqd->park_pending);
7354 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7355 mutex_lock(&sqd->lock);
7357 wake_up_process(sqd->thread);
7360 static void io_sq_thread_stop(struct io_sq_data *sqd)
7362 WARN_ON_ONCE(sqd->thread == current);
7363 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7365 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7366 mutex_lock(&sqd->lock);
7368 wake_up_process(sqd->thread);
7369 mutex_unlock(&sqd->lock);
7370 wait_for_completion(&sqd->exited);
7373 static void io_put_sq_data(struct io_sq_data *sqd)
7375 if (refcount_dec_and_test(&sqd->refs)) {
7376 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7378 io_sq_thread_stop(sqd);
7383 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7385 struct io_sq_data *sqd = ctx->sq_data;
7388 io_sq_thread_park(sqd);
7389 list_del_init(&ctx->sqd_list);
7390 io_sqd_update_thread_idle(sqd);
7391 io_sq_thread_unpark(sqd);
7393 io_put_sq_data(sqd);
7394 ctx->sq_data = NULL;
7398 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7400 struct io_ring_ctx *ctx_attach;
7401 struct io_sq_data *sqd;
7404 f = fdget(p->wq_fd);
7406 return ERR_PTR(-ENXIO);
7407 if (f.file->f_op != &io_uring_fops) {
7409 return ERR_PTR(-EINVAL);
7412 ctx_attach = f.file->private_data;
7413 sqd = ctx_attach->sq_data;
7416 return ERR_PTR(-EINVAL);
7418 if (sqd->task_tgid != current->tgid) {
7420 return ERR_PTR(-EPERM);
7423 refcount_inc(&sqd->refs);
7428 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7431 struct io_sq_data *sqd;
7434 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7435 sqd = io_attach_sq_data(p);
7440 /* fall through for EPERM case, setup new sqd/task */
7441 if (PTR_ERR(sqd) != -EPERM)
7445 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7447 return ERR_PTR(-ENOMEM);
7449 atomic_set(&sqd->park_pending, 0);
7450 refcount_set(&sqd->refs, 1);
7451 INIT_LIST_HEAD(&sqd->ctx_list);
7452 mutex_init(&sqd->lock);
7453 init_waitqueue_head(&sqd->wait);
7454 init_completion(&sqd->exited);
7458 #if defined(CONFIG_UNIX)
7460 * Ensure the UNIX gc is aware of our file set, so we are certain that
7461 * the io_uring can be safely unregistered on process exit, even if we have
7462 * loops in the file referencing.
7464 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7466 struct sock *sk = ctx->ring_sock->sk;
7467 struct scm_fp_list *fpl;
7468 struct sk_buff *skb;
7471 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7475 skb = alloc_skb(0, GFP_KERNEL);
7484 fpl->user = get_uid(current_user());
7485 for (i = 0; i < nr; i++) {
7486 struct file *file = io_file_from_index(ctx, i + offset);
7490 fpl->fp[nr_files] = get_file(file);
7491 unix_inflight(fpl->user, fpl->fp[nr_files]);
7496 fpl->max = SCM_MAX_FD;
7497 fpl->count = nr_files;
7498 UNIXCB(skb).fp = fpl;
7499 skb->destructor = unix_destruct_scm;
7500 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7501 skb_queue_head(&sk->sk_receive_queue, skb);
7503 for (i = 0; i < nr_files; i++)
7514 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7515 * causes regular reference counting to break down. We rely on the UNIX
7516 * garbage collection to take care of this problem for us.
7518 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7520 unsigned left, total;
7524 left = ctx->nr_user_files;
7526 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7528 ret = __io_sqe_files_scm(ctx, this_files, total);
7532 total += this_files;
7538 while (total < ctx->nr_user_files) {
7539 struct file *file = io_file_from_index(ctx, total);
7549 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7555 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7557 struct file *file = prsrc->file;
7558 #if defined(CONFIG_UNIX)
7559 struct sock *sock = ctx->ring_sock->sk;
7560 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7561 struct sk_buff *skb;
7564 __skb_queue_head_init(&list);
7567 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7568 * remove this entry and rearrange the file array.
7570 skb = skb_dequeue(head);
7572 struct scm_fp_list *fp;
7574 fp = UNIXCB(skb).fp;
7575 for (i = 0; i < fp->count; i++) {
7578 if (fp->fp[i] != file)
7581 unix_notinflight(fp->user, fp->fp[i]);
7582 left = fp->count - 1 - i;
7584 memmove(&fp->fp[i], &fp->fp[i + 1],
7585 left * sizeof(struct file *));
7592 __skb_queue_tail(&list, skb);
7602 __skb_queue_tail(&list, skb);
7604 skb = skb_dequeue(head);
7607 if (skb_peek(&list)) {
7608 spin_lock_irq(&head->lock);
7609 while ((skb = __skb_dequeue(&list)) != NULL)
7610 __skb_queue_tail(head, skb);
7611 spin_unlock_irq(&head->lock);
7618 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7620 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7621 struct io_ring_ctx *ctx = rsrc_data->ctx;
7622 struct io_rsrc_put *prsrc, *tmp;
7624 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7625 list_del(&prsrc->list);
7628 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7630 io_ring_submit_lock(ctx, lock_ring);
7631 spin_lock_irq(&ctx->completion_lock);
7632 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7634 io_commit_cqring(ctx);
7635 spin_unlock_irq(&ctx->completion_lock);
7636 io_cqring_ev_posted(ctx);
7637 io_ring_submit_unlock(ctx, lock_ring);
7640 rsrc_data->do_put(ctx, prsrc);
7644 io_rsrc_node_destroy(ref_node);
7645 if (atomic_dec_and_test(&rsrc_data->refs))
7646 complete(&rsrc_data->done);
7649 static void io_rsrc_put_work(struct work_struct *work)
7651 struct io_ring_ctx *ctx;
7652 struct llist_node *node;
7654 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7655 node = llist_del_all(&ctx->rsrc_put_llist);
7658 struct io_rsrc_node *ref_node;
7659 struct llist_node *next = node->next;
7661 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7662 __io_rsrc_put_work(ref_node);
7667 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7669 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7670 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7671 bool first_add = false;
7673 io_rsrc_ref_lock(ctx);
7676 while (!list_empty(&ctx->rsrc_ref_list)) {
7677 node = list_first_entry(&ctx->rsrc_ref_list,
7678 struct io_rsrc_node, node);
7679 /* recycle ref nodes in order */
7682 list_del(&node->node);
7683 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7685 io_rsrc_ref_unlock(ctx);
7688 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7691 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7693 struct io_rsrc_node *ref_node;
7695 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7699 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7704 INIT_LIST_HEAD(&ref_node->node);
7705 INIT_LIST_HEAD(&ref_node->rsrc_list);
7706 ref_node->done = false;
7710 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7711 unsigned nr_args, u64 __user *tags)
7713 __s32 __user *fds = (__s32 __user *) arg;
7722 if (nr_args > IORING_MAX_FIXED_FILES)
7724 ret = io_rsrc_node_switch_start(ctx);
7727 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7733 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7736 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7737 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7741 /* allow sparse sets */
7744 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7751 if (unlikely(!file))
7755 * Don't allow io_uring instances to be registered. If UNIX
7756 * isn't enabled, then this causes a reference cycle and this
7757 * instance can never get freed. If UNIX is enabled we'll
7758 * handle it just fine, but there's still no point in allowing
7759 * a ring fd as it doesn't support regular read/write anyway.
7761 if (file->f_op == &io_uring_fops) {
7765 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7768 ret = io_sqe_files_scm(ctx);
7770 __io_sqe_files_unregister(ctx);
7774 io_rsrc_node_switch(ctx, NULL);
7777 for (i = 0; i < ctx->nr_user_files; i++) {
7778 file = io_file_from_index(ctx, i);
7782 io_free_file_tables(&ctx->file_table, nr_args);
7783 ctx->nr_user_files = 0;
7785 io_rsrc_data_free(ctx->file_data);
7786 ctx->file_data = NULL;
7790 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7793 #if defined(CONFIG_UNIX)
7794 struct sock *sock = ctx->ring_sock->sk;
7795 struct sk_buff_head *head = &sock->sk_receive_queue;
7796 struct sk_buff *skb;
7799 * See if we can merge this file into an existing skb SCM_RIGHTS
7800 * file set. If there's no room, fall back to allocating a new skb
7801 * and filling it in.
7803 spin_lock_irq(&head->lock);
7804 skb = skb_peek(head);
7806 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7808 if (fpl->count < SCM_MAX_FD) {
7809 __skb_unlink(skb, head);
7810 spin_unlock_irq(&head->lock);
7811 fpl->fp[fpl->count] = get_file(file);
7812 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7814 spin_lock_irq(&head->lock);
7815 __skb_queue_head(head, skb);
7820 spin_unlock_irq(&head->lock);
7827 return __io_sqe_files_scm(ctx, 1, index);
7833 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7834 struct io_rsrc_node *node, void *rsrc)
7836 struct io_rsrc_put *prsrc;
7838 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7842 prsrc->tag = *io_get_tag_slot(data, idx);
7844 list_add(&prsrc->list, &node->rsrc_list);
7848 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7849 struct io_uring_rsrc_update2 *up,
7852 u64 __user *tags = u64_to_user_ptr(up->tags);
7853 __s32 __user *fds = u64_to_user_ptr(up->data);
7854 struct io_rsrc_data *data = ctx->file_data;
7855 struct io_fixed_file *file_slot;
7859 bool needs_switch = false;
7861 if (!ctx->file_data)
7863 if (up->offset + nr_args > ctx->nr_user_files)
7866 for (done = 0; done < nr_args; done++) {
7869 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7870 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7874 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7878 if (fd == IORING_REGISTER_FILES_SKIP)
7881 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7882 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7884 if (file_slot->file_ptr) {
7885 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7886 err = io_queue_rsrc_removal(data, up->offset + done,
7887 ctx->rsrc_node, file);
7890 file_slot->file_ptr = 0;
7891 needs_switch = true;
7900 * Don't allow io_uring instances to be registered. If
7901 * UNIX isn't enabled, then this causes a reference
7902 * cycle and this instance can never get freed. If UNIX
7903 * is enabled we'll handle it just fine, but there's
7904 * still no point in allowing a ring fd as it doesn't
7905 * support regular read/write anyway.
7907 if (file->f_op == &io_uring_fops) {
7912 *io_get_tag_slot(data, up->offset + done) = tag;
7913 io_fixed_file_set(file_slot, file);
7914 err = io_sqe_file_register(ctx, file, i);
7916 file_slot->file_ptr = 0;
7924 io_rsrc_node_switch(ctx, data);
7925 return done ? done : err;
7928 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7930 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7932 req = io_put_req_find_next(req);
7933 return req ? &req->work : NULL;
7936 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7937 struct task_struct *task)
7939 struct io_wq_hash *hash;
7940 struct io_wq_data data;
7941 unsigned int concurrency;
7943 mutex_lock(&ctx->uring_lock);
7944 hash = ctx->hash_map;
7946 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7948 mutex_unlock(&ctx->uring_lock);
7949 return ERR_PTR(-ENOMEM);
7951 refcount_set(&hash->refs, 1);
7952 init_waitqueue_head(&hash->wait);
7953 ctx->hash_map = hash;
7955 mutex_unlock(&ctx->uring_lock);
7959 data.free_work = io_free_work;
7960 data.do_work = io_wq_submit_work;
7962 /* Do QD, or 4 * CPUS, whatever is smallest */
7963 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7965 return io_wq_create(concurrency, &data);
7968 static int io_uring_alloc_task_context(struct task_struct *task,
7969 struct io_ring_ctx *ctx)
7971 struct io_uring_task *tctx;
7974 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7975 if (unlikely(!tctx))
7978 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7979 if (unlikely(ret)) {
7984 tctx->io_wq = io_init_wq_offload(ctx, task);
7985 if (IS_ERR(tctx->io_wq)) {
7986 ret = PTR_ERR(tctx->io_wq);
7987 percpu_counter_destroy(&tctx->inflight);
7993 init_waitqueue_head(&tctx->wait);
7994 atomic_set(&tctx->in_idle, 0);
7995 atomic_set(&tctx->inflight_tracked, 0);
7996 task->io_uring = tctx;
7997 spin_lock_init(&tctx->task_lock);
7998 INIT_WQ_LIST(&tctx->task_list);
7999 init_task_work(&tctx->task_work, tctx_task_work);
8003 void __io_uring_free(struct task_struct *tsk)
8005 struct io_uring_task *tctx = tsk->io_uring;
8007 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8008 WARN_ON_ONCE(tctx->io_wq);
8009 WARN_ON_ONCE(tctx->cached_refs);
8011 percpu_counter_destroy(&tctx->inflight);
8013 tsk->io_uring = NULL;
8016 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8017 struct io_uring_params *p)
8021 /* Retain compatibility with failing for an invalid attach attempt */
8022 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8023 IORING_SETUP_ATTACH_WQ) {
8026 f = fdget(p->wq_fd);
8029 if (f.file->f_op != &io_uring_fops) {
8035 if (ctx->flags & IORING_SETUP_SQPOLL) {
8036 struct task_struct *tsk;
8037 struct io_sq_data *sqd;
8040 sqd = io_get_sq_data(p, &attached);
8046 ctx->sq_creds = get_current_cred();
8048 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8049 if (!ctx->sq_thread_idle)
8050 ctx->sq_thread_idle = HZ;
8052 io_sq_thread_park(sqd);
8053 list_add(&ctx->sqd_list, &sqd->ctx_list);
8054 io_sqd_update_thread_idle(sqd);
8055 /* don't attach to a dying SQPOLL thread, would be racy */
8056 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8057 io_sq_thread_unpark(sqd);
8064 if (p->flags & IORING_SETUP_SQ_AFF) {
8065 int cpu = p->sq_thread_cpu;
8068 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8075 sqd->task_pid = current->pid;
8076 sqd->task_tgid = current->tgid;
8077 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8084 ret = io_uring_alloc_task_context(tsk, ctx);
8085 wake_up_new_task(tsk);
8088 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8089 /* Can't have SQ_AFF without SQPOLL */
8096 complete(&ctx->sq_data->exited);
8098 io_sq_thread_finish(ctx);
8102 static inline void __io_unaccount_mem(struct user_struct *user,
8103 unsigned long nr_pages)
8105 atomic_long_sub(nr_pages, &user->locked_vm);
8108 static inline int __io_account_mem(struct user_struct *user,
8109 unsigned long nr_pages)
8111 unsigned long page_limit, cur_pages, new_pages;
8113 /* Don't allow more pages than we can safely lock */
8114 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8117 cur_pages = atomic_long_read(&user->locked_vm);
8118 new_pages = cur_pages + nr_pages;
8119 if (new_pages > page_limit)
8121 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8122 new_pages) != cur_pages);
8127 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8130 __io_unaccount_mem(ctx->user, nr_pages);
8132 if (ctx->mm_account)
8133 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8136 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8141 ret = __io_account_mem(ctx->user, nr_pages);
8146 if (ctx->mm_account)
8147 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8152 static void io_mem_free(void *ptr)
8159 page = virt_to_head_page(ptr);
8160 if (put_page_testzero(page))
8161 free_compound_page(page);
8164 static void *io_mem_alloc(size_t size)
8166 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8167 __GFP_NORETRY | __GFP_ACCOUNT;
8169 return (void *) __get_free_pages(gfp_flags, get_order(size));
8172 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8175 struct io_rings *rings;
8176 size_t off, sq_array_size;
8178 off = struct_size(rings, cqes, cq_entries);
8179 if (off == SIZE_MAX)
8183 off = ALIGN(off, SMP_CACHE_BYTES);
8191 sq_array_size = array_size(sizeof(u32), sq_entries);
8192 if (sq_array_size == SIZE_MAX)
8195 if (check_add_overflow(off, sq_array_size, &off))
8201 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8203 struct io_mapped_ubuf *imu = *slot;
8206 if (imu != ctx->dummy_ubuf) {
8207 for (i = 0; i < imu->nr_bvecs; i++)
8208 unpin_user_page(imu->bvec[i].bv_page);
8209 if (imu->acct_pages)
8210 io_unaccount_mem(ctx, imu->acct_pages);
8216 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8218 io_buffer_unmap(ctx, &prsrc->buf);
8222 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8226 for (i = 0; i < ctx->nr_user_bufs; i++)
8227 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8228 kfree(ctx->user_bufs);
8229 io_rsrc_data_free(ctx->buf_data);
8230 ctx->user_bufs = NULL;
8231 ctx->buf_data = NULL;
8232 ctx->nr_user_bufs = 0;
8235 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8242 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8244 __io_sqe_buffers_unregister(ctx);
8248 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8249 void __user *arg, unsigned index)
8251 struct iovec __user *src;
8253 #ifdef CONFIG_COMPAT
8255 struct compat_iovec __user *ciovs;
8256 struct compat_iovec ciov;
8258 ciovs = (struct compat_iovec __user *) arg;
8259 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8262 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8263 dst->iov_len = ciov.iov_len;
8267 src = (struct iovec __user *) arg;
8268 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8274 * Not super efficient, but this is just a registration time. And we do cache
8275 * the last compound head, so generally we'll only do a full search if we don't
8278 * We check if the given compound head page has already been accounted, to
8279 * avoid double accounting it. This allows us to account the full size of the
8280 * page, not just the constituent pages of a huge page.
8282 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8283 int nr_pages, struct page *hpage)
8287 /* check current page array */
8288 for (i = 0; i < nr_pages; i++) {
8289 if (!PageCompound(pages[i]))
8291 if (compound_head(pages[i]) == hpage)
8295 /* check previously registered pages */
8296 for (i = 0; i < ctx->nr_user_bufs; i++) {
8297 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8299 for (j = 0; j < imu->nr_bvecs; j++) {
8300 if (!PageCompound(imu->bvec[j].bv_page))
8302 if (compound_head(imu->bvec[j].bv_page) == hpage)
8310 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8311 int nr_pages, struct io_mapped_ubuf *imu,
8312 struct page **last_hpage)
8316 imu->acct_pages = 0;
8317 for (i = 0; i < nr_pages; i++) {
8318 if (!PageCompound(pages[i])) {
8323 hpage = compound_head(pages[i]);
8324 if (hpage == *last_hpage)
8326 *last_hpage = hpage;
8327 if (headpage_already_acct(ctx, pages, i, hpage))
8329 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8333 if (!imu->acct_pages)
8336 ret = io_account_mem(ctx, imu->acct_pages);
8338 imu->acct_pages = 0;
8342 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8343 struct io_mapped_ubuf **pimu,
8344 struct page **last_hpage)
8346 struct io_mapped_ubuf *imu = NULL;
8347 struct vm_area_struct **vmas = NULL;
8348 struct page **pages = NULL;
8349 unsigned long off, start, end, ubuf;
8351 int ret, pret, nr_pages, i;
8353 if (!iov->iov_base) {
8354 *pimu = ctx->dummy_ubuf;
8358 ubuf = (unsigned long) iov->iov_base;
8359 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8360 start = ubuf >> PAGE_SHIFT;
8361 nr_pages = end - start;
8366 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8370 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8375 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8380 mmap_read_lock(current->mm);
8381 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8383 if (pret == nr_pages) {
8384 /* don't support file backed memory */
8385 for (i = 0; i < nr_pages; i++) {
8386 struct vm_area_struct *vma = vmas[i];
8388 if (vma_is_shmem(vma))
8391 !is_file_hugepages(vma->vm_file)) {
8397 ret = pret < 0 ? pret : -EFAULT;
8399 mmap_read_unlock(current->mm);
8402 * if we did partial map, or found file backed vmas,
8403 * release any pages we did get
8406 unpin_user_pages(pages, pret);
8410 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8412 unpin_user_pages(pages, pret);
8416 off = ubuf & ~PAGE_MASK;
8417 size = iov->iov_len;
8418 for (i = 0; i < nr_pages; i++) {
8421 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8422 imu->bvec[i].bv_page = pages[i];
8423 imu->bvec[i].bv_len = vec_len;
8424 imu->bvec[i].bv_offset = off;
8428 /* store original address for later verification */
8430 imu->ubuf_end = ubuf + iov->iov_len;
8431 imu->nr_bvecs = nr_pages;
8442 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8444 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8445 return ctx->user_bufs ? 0 : -ENOMEM;
8448 static int io_buffer_validate(struct iovec *iov)
8450 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8453 * Don't impose further limits on the size and buffer
8454 * constraints here, we'll -EINVAL later when IO is
8455 * submitted if they are wrong.
8458 return iov->iov_len ? -EFAULT : 0;
8462 /* arbitrary limit, but we need something */
8463 if (iov->iov_len > SZ_1G)
8466 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8472 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8473 unsigned int nr_args, u64 __user *tags)
8475 struct page *last_hpage = NULL;
8476 struct io_rsrc_data *data;
8482 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8484 ret = io_rsrc_node_switch_start(ctx);
8487 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8490 ret = io_buffers_map_alloc(ctx, nr_args);
8492 io_rsrc_data_free(data);
8496 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8497 ret = io_copy_iov(ctx, &iov, arg, i);
8500 ret = io_buffer_validate(&iov);
8503 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8508 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8514 WARN_ON_ONCE(ctx->buf_data);
8516 ctx->buf_data = data;
8518 __io_sqe_buffers_unregister(ctx);
8520 io_rsrc_node_switch(ctx, NULL);
8524 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8525 struct io_uring_rsrc_update2 *up,
8526 unsigned int nr_args)
8528 u64 __user *tags = u64_to_user_ptr(up->tags);
8529 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8530 struct page *last_hpage = NULL;
8531 bool needs_switch = false;
8537 if (up->offset + nr_args > ctx->nr_user_bufs)
8540 for (done = 0; done < nr_args; done++) {
8541 struct io_mapped_ubuf *imu;
8542 int offset = up->offset + done;
8545 err = io_copy_iov(ctx, &iov, iovs, done);
8548 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8552 err = io_buffer_validate(&iov);
8555 if (!iov.iov_base && tag) {
8559 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8563 i = array_index_nospec(offset, ctx->nr_user_bufs);
8564 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8565 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8566 ctx->rsrc_node, ctx->user_bufs[i]);
8567 if (unlikely(err)) {
8568 io_buffer_unmap(ctx, &imu);
8571 ctx->user_bufs[i] = NULL;
8572 needs_switch = true;
8575 ctx->user_bufs[i] = imu;
8576 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8580 io_rsrc_node_switch(ctx, ctx->buf_data);
8581 return done ? done : err;
8584 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8586 __s32 __user *fds = arg;
8592 if (copy_from_user(&fd, fds, sizeof(*fds)))
8595 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8596 if (IS_ERR(ctx->cq_ev_fd)) {
8597 int ret = PTR_ERR(ctx->cq_ev_fd);
8599 ctx->cq_ev_fd = NULL;
8606 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8608 if (ctx->cq_ev_fd) {
8609 eventfd_ctx_put(ctx->cq_ev_fd);
8610 ctx->cq_ev_fd = NULL;
8617 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8619 struct io_buffer *buf;
8620 unsigned long index;
8622 xa_for_each(&ctx->io_buffers, index, buf)
8623 __io_remove_buffers(ctx, buf, index, -1U);
8626 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8628 struct io_kiocb *req, *nxt;
8630 list_for_each_entry_safe(req, nxt, list, compl.list) {
8631 if (tsk && req->task != tsk)
8633 list_del(&req->compl.list);
8634 kmem_cache_free(req_cachep, req);
8638 static void io_req_caches_free(struct io_ring_ctx *ctx)
8640 struct io_submit_state *submit_state = &ctx->submit_state;
8641 struct io_comp_state *cs = &ctx->submit_state.comp;
8643 mutex_lock(&ctx->uring_lock);
8645 if (submit_state->free_reqs) {
8646 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8647 submit_state->reqs);
8648 submit_state->free_reqs = 0;
8651 io_flush_cached_locked_reqs(ctx, cs);
8652 io_req_cache_free(&cs->free_list, NULL);
8653 mutex_unlock(&ctx->uring_lock);
8656 static bool io_wait_rsrc_data(struct io_rsrc_data *data)
8660 if (!atomic_dec_and_test(&data->refs))
8661 wait_for_completion(&data->done);
8665 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8667 io_sq_thread_finish(ctx);
8669 if (ctx->mm_account) {
8670 mmdrop(ctx->mm_account);
8671 ctx->mm_account = NULL;
8674 mutex_lock(&ctx->uring_lock);
8675 if (io_wait_rsrc_data(ctx->buf_data))
8676 __io_sqe_buffers_unregister(ctx);
8677 if (io_wait_rsrc_data(ctx->file_data))
8678 __io_sqe_files_unregister(ctx);
8680 __io_cqring_overflow_flush(ctx, true);
8681 mutex_unlock(&ctx->uring_lock);
8682 io_eventfd_unregister(ctx);
8683 io_destroy_buffers(ctx);
8685 put_cred(ctx->sq_creds);
8687 /* there are no registered resources left, nobody uses it */
8689 io_rsrc_node_destroy(ctx->rsrc_node);
8690 if (ctx->rsrc_backup_node)
8691 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8692 flush_delayed_work(&ctx->rsrc_put_work);
8694 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8695 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8697 #if defined(CONFIG_UNIX)
8698 if (ctx->ring_sock) {
8699 ctx->ring_sock->file = NULL; /* so that iput() is called */
8700 sock_release(ctx->ring_sock);
8704 io_mem_free(ctx->rings);
8705 io_mem_free(ctx->sq_sqes);
8707 percpu_ref_exit(&ctx->refs);
8708 free_uid(ctx->user);
8709 io_req_caches_free(ctx);
8711 io_wq_put_hash(ctx->hash_map);
8712 kfree(ctx->cancel_hash);
8713 kfree(ctx->dummy_ubuf);
8717 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8719 struct io_ring_ctx *ctx = file->private_data;
8722 poll_wait(file, &ctx->poll_wait, wait);
8724 * synchronizes with barrier from wq_has_sleeper call in
8728 if (!io_sqring_full(ctx))
8729 mask |= EPOLLOUT | EPOLLWRNORM;
8732 * Don't flush cqring overflow list here, just do a simple check.
8733 * Otherwise there could possible be ABBA deadlock:
8736 * lock(&ctx->uring_lock);
8738 * lock(&ctx->uring_lock);
8741 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8742 * pushs them to do the flush.
8744 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8745 mask |= EPOLLIN | EPOLLRDNORM;
8750 static int io_uring_fasync(int fd, struct file *file, int on)
8752 struct io_ring_ctx *ctx = file->private_data;
8754 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8757 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8759 const struct cred *creds;
8761 creds = xa_erase(&ctx->personalities, id);
8770 struct io_tctx_exit {
8771 struct callback_head task_work;
8772 struct completion completion;
8773 struct io_ring_ctx *ctx;
8776 static void io_tctx_exit_cb(struct callback_head *cb)
8778 struct io_uring_task *tctx = current->io_uring;
8779 struct io_tctx_exit *work;
8781 work = container_of(cb, struct io_tctx_exit, task_work);
8783 * When @in_idle, we're in cancellation and it's racy to remove the
8784 * node. It'll be removed by the end of cancellation, just ignore it.
8786 if (!atomic_read(&tctx->in_idle))
8787 io_uring_del_tctx_node((unsigned long)work->ctx);
8788 complete(&work->completion);
8791 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8793 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8795 return req->ctx == data;
8798 static void io_ring_exit_work(struct work_struct *work)
8800 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8801 unsigned long timeout = jiffies + HZ * 60 * 5;
8802 struct io_tctx_exit exit;
8803 struct io_tctx_node *node;
8807 * If we're doing polled IO and end up having requests being
8808 * submitted async (out-of-line), then completions can come in while
8809 * we're waiting for refs to drop. We need to reap these manually,
8810 * as nobody else will be looking for them.
8813 io_uring_try_cancel_requests(ctx, NULL, true);
8815 struct io_sq_data *sqd = ctx->sq_data;
8816 struct task_struct *tsk;
8818 io_sq_thread_park(sqd);
8820 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8821 io_wq_cancel_cb(tsk->io_uring->io_wq,
8822 io_cancel_ctx_cb, ctx, true);
8823 io_sq_thread_unpark(sqd);
8826 WARN_ON_ONCE(time_after(jiffies, timeout));
8827 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8829 init_completion(&exit.completion);
8830 init_task_work(&exit.task_work, io_tctx_exit_cb);
8833 * Some may use context even when all refs and requests have been put,
8834 * and they are free to do so while still holding uring_lock or
8835 * completion_lock, see io_req_task_submit(). Apart from other work,
8836 * this lock/unlock section also waits them to finish.
8838 mutex_lock(&ctx->uring_lock);
8839 while (!list_empty(&ctx->tctx_list)) {
8840 WARN_ON_ONCE(time_after(jiffies, timeout));
8842 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8844 /* don't spin on a single task if cancellation failed */
8845 list_rotate_left(&ctx->tctx_list);
8846 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8847 if (WARN_ON_ONCE(ret))
8849 wake_up_process(node->task);
8851 mutex_unlock(&ctx->uring_lock);
8852 wait_for_completion(&exit.completion);
8853 mutex_lock(&ctx->uring_lock);
8855 mutex_unlock(&ctx->uring_lock);
8856 spin_lock_irq(&ctx->completion_lock);
8857 spin_unlock_irq(&ctx->completion_lock);
8859 io_ring_ctx_free(ctx);
8862 /* Returns true if we found and killed one or more timeouts */
8863 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8866 struct io_kiocb *req, *tmp;
8869 spin_lock_irq(&ctx->completion_lock);
8870 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8871 if (io_match_task(req, tsk, cancel_all)) {
8872 io_kill_timeout(req, -ECANCELED);
8877 io_commit_cqring(ctx);
8878 spin_unlock_irq(&ctx->completion_lock);
8880 io_cqring_ev_posted(ctx);
8881 return canceled != 0;
8884 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8886 unsigned long index;
8887 struct creds *creds;
8889 mutex_lock(&ctx->uring_lock);
8890 percpu_ref_kill(&ctx->refs);
8892 __io_cqring_overflow_flush(ctx, true);
8893 xa_for_each(&ctx->personalities, index, creds)
8894 io_unregister_personality(ctx, index);
8895 mutex_unlock(&ctx->uring_lock);
8897 io_kill_timeouts(ctx, NULL, true);
8898 io_poll_remove_all(ctx, NULL, true);
8900 /* if we failed setting up the ctx, we might not have any rings */
8901 io_iopoll_try_reap_events(ctx);
8903 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8905 * Use system_unbound_wq to avoid spawning tons of event kworkers
8906 * if we're exiting a ton of rings at the same time. It just adds
8907 * noise and overhead, there's no discernable change in runtime
8908 * over using system_wq.
8910 queue_work(system_unbound_wq, &ctx->exit_work);
8913 static int io_uring_release(struct inode *inode, struct file *file)
8915 struct io_ring_ctx *ctx = file->private_data;
8917 file->private_data = NULL;
8918 io_ring_ctx_wait_and_kill(ctx);
8922 struct io_task_cancel {
8923 struct task_struct *task;
8927 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8929 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8930 struct io_task_cancel *cancel = data;
8933 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8934 unsigned long flags;
8935 struct io_ring_ctx *ctx = req->ctx;
8937 /* protect against races with linked timeouts */
8938 spin_lock_irqsave(&ctx->completion_lock, flags);
8939 ret = io_match_task(req, cancel->task, cancel->all);
8940 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8942 ret = io_match_task(req, cancel->task, cancel->all);
8947 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8948 struct task_struct *task, bool cancel_all)
8950 struct io_defer_entry *de;
8953 spin_lock_irq(&ctx->completion_lock);
8954 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8955 if (io_match_task(de->req, task, cancel_all)) {
8956 list_cut_position(&list, &ctx->defer_list, &de->list);
8960 spin_unlock_irq(&ctx->completion_lock);
8961 if (list_empty(&list))
8964 while (!list_empty(&list)) {
8965 de = list_first_entry(&list, struct io_defer_entry, list);
8966 list_del_init(&de->list);
8967 io_req_complete_failed(de->req, -ECANCELED);
8973 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8975 struct io_tctx_node *node;
8976 enum io_wq_cancel cret;
8979 mutex_lock(&ctx->uring_lock);
8980 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8981 struct io_uring_task *tctx = node->task->io_uring;
8984 * io_wq will stay alive while we hold uring_lock, because it's
8985 * killed after ctx nodes, which requires to take the lock.
8987 if (!tctx || !tctx->io_wq)
8989 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8990 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8992 mutex_unlock(&ctx->uring_lock);
8997 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8998 struct task_struct *task,
9001 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9002 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9005 enum io_wq_cancel cret;
9009 ret |= io_uring_try_cancel_iowq(ctx);
9010 } else if (tctx && tctx->io_wq) {
9012 * Cancels requests of all rings, not only @ctx, but
9013 * it's fine as the task is in exit/exec.
9015 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9017 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9020 /* SQPOLL thread does its own polling */
9021 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9022 (ctx->sq_data && ctx->sq_data->thread == current)) {
9023 while (!list_empty_careful(&ctx->iopoll_list)) {
9024 io_iopoll_try_reap_events(ctx);
9029 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9030 ret |= io_poll_remove_all(ctx, task, cancel_all);
9031 ret |= io_kill_timeouts(ctx, task, cancel_all);
9033 ret |= io_run_task_work();
9040 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9042 struct io_uring_task *tctx = current->io_uring;
9043 struct io_tctx_node *node;
9046 if (unlikely(!tctx)) {
9047 ret = io_uring_alloc_task_context(current, ctx);
9050 tctx = current->io_uring;
9052 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9053 node = kmalloc(sizeof(*node), GFP_KERNEL);
9057 node->task = current;
9059 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9066 mutex_lock(&ctx->uring_lock);
9067 list_add(&node->ctx_node, &ctx->tctx_list);
9068 mutex_unlock(&ctx->uring_lock);
9075 * Note that this task has used io_uring. We use it for cancelation purposes.
9077 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9079 struct io_uring_task *tctx = current->io_uring;
9081 if (likely(tctx && tctx->last == ctx))
9083 return __io_uring_add_tctx_node(ctx);
9087 * Remove this io_uring_file -> task mapping.
9089 static void io_uring_del_tctx_node(unsigned long index)
9091 struct io_uring_task *tctx = current->io_uring;
9092 struct io_tctx_node *node;
9096 node = xa_erase(&tctx->xa, index);
9100 WARN_ON_ONCE(current != node->task);
9101 WARN_ON_ONCE(list_empty(&node->ctx_node));
9103 mutex_lock(&node->ctx->uring_lock);
9104 list_del(&node->ctx_node);
9105 mutex_unlock(&node->ctx->uring_lock);
9107 if (tctx->last == node->ctx)
9112 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9114 struct io_wq *wq = tctx->io_wq;
9115 struct io_tctx_node *node;
9116 unsigned long index;
9118 xa_for_each(&tctx->xa, index, node)
9119 io_uring_del_tctx_node(index);
9122 * Must be after io_uring_del_task_file() (removes nodes under
9123 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9126 io_wq_put_and_exit(wq);
9130 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9133 return atomic_read(&tctx->inflight_tracked);
9134 return percpu_counter_sum(&tctx->inflight);
9137 static void io_uring_drop_tctx_refs(struct task_struct *task)
9139 struct io_uring_task *tctx = task->io_uring;
9140 unsigned int refs = tctx->cached_refs;
9142 tctx->cached_refs = 0;
9143 percpu_counter_sub(&tctx->inflight, refs);
9144 put_task_struct_many(task, refs);
9148 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9149 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9151 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9153 struct io_uring_task *tctx = current->io_uring;
9154 struct io_ring_ctx *ctx;
9158 WARN_ON_ONCE(sqd && sqd->thread != current);
9160 if (!current->io_uring)
9163 io_wq_exit_start(tctx->io_wq);
9165 io_uring_drop_tctx_refs(current);
9166 atomic_inc(&tctx->in_idle);
9168 /* read completions before cancelations */
9169 inflight = tctx_inflight(tctx, !cancel_all);
9174 struct io_tctx_node *node;
9175 unsigned long index;
9177 xa_for_each(&tctx->xa, index, node) {
9178 /* sqpoll task will cancel all its requests */
9179 if (node->ctx->sq_data)
9181 io_uring_try_cancel_requests(node->ctx, current,
9185 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9186 io_uring_try_cancel_requests(ctx, current,
9190 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9192 * If we've seen completions, retry without waiting. This
9193 * avoids a race where a completion comes in before we did
9194 * prepare_to_wait().
9196 if (inflight == tctx_inflight(tctx, !cancel_all))
9198 finish_wait(&tctx->wait, &wait);
9200 atomic_dec(&tctx->in_idle);
9202 io_uring_clean_tctx(tctx);
9204 /* for exec all current's requests should be gone, kill tctx */
9205 __io_uring_free(current);
9209 void __io_uring_cancel(struct files_struct *files)
9211 io_uring_cancel_generic(!files, NULL);
9214 static void *io_uring_validate_mmap_request(struct file *file,
9215 loff_t pgoff, size_t sz)
9217 struct io_ring_ctx *ctx = file->private_data;
9218 loff_t offset = pgoff << PAGE_SHIFT;
9223 case IORING_OFF_SQ_RING:
9224 case IORING_OFF_CQ_RING:
9227 case IORING_OFF_SQES:
9231 return ERR_PTR(-EINVAL);
9234 page = virt_to_head_page(ptr);
9235 if (sz > page_size(page))
9236 return ERR_PTR(-EINVAL);
9243 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9245 size_t sz = vma->vm_end - vma->vm_start;
9249 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9251 return PTR_ERR(ptr);
9253 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9254 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9257 #else /* !CONFIG_MMU */
9259 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9261 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9264 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9266 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9269 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9270 unsigned long addr, unsigned long len,
9271 unsigned long pgoff, unsigned long flags)
9275 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9277 return PTR_ERR(ptr);
9279 return (unsigned long) ptr;
9282 #endif /* !CONFIG_MMU */
9284 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9289 if (!io_sqring_full(ctx))
9291 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9293 if (!io_sqring_full(ctx))
9296 } while (!signal_pending(current));
9298 finish_wait(&ctx->sqo_sq_wait, &wait);
9302 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9303 struct __kernel_timespec __user **ts,
9304 const sigset_t __user **sig)
9306 struct io_uring_getevents_arg arg;
9309 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9310 * is just a pointer to the sigset_t.
9312 if (!(flags & IORING_ENTER_EXT_ARG)) {
9313 *sig = (const sigset_t __user *) argp;
9319 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9320 * timespec and sigset_t pointers if good.
9322 if (*argsz != sizeof(arg))
9324 if (copy_from_user(&arg, argp, sizeof(arg)))
9326 *sig = u64_to_user_ptr(arg.sigmask);
9327 *argsz = arg.sigmask_sz;
9328 *ts = u64_to_user_ptr(arg.ts);
9332 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9333 u32, min_complete, u32, flags, const void __user *, argp,
9336 struct io_ring_ctx *ctx;
9343 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9344 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9348 if (unlikely(!f.file))
9352 if (unlikely(f.file->f_op != &io_uring_fops))
9356 ctx = f.file->private_data;
9357 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9361 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9365 * For SQ polling, the thread will do all submissions and completions.
9366 * Just return the requested submit count, and wake the thread if
9370 if (ctx->flags & IORING_SETUP_SQPOLL) {
9371 io_cqring_overflow_flush(ctx, false);
9374 if (unlikely(ctx->sq_data->thread == NULL))
9376 if (flags & IORING_ENTER_SQ_WAKEUP)
9377 wake_up(&ctx->sq_data->wait);
9378 if (flags & IORING_ENTER_SQ_WAIT) {
9379 ret = io_sqpoll_wait_sq(ctx);
9383 submitted = to_submit;
9384 } else if (to_submit) {
9385 ret = io_uring_add_tctx_node(ctx);
9388 mutex_lock(&ctx->uring_lock);
9389 submitted = io_submit_sqes(ctx, to_submit);
9390 mutex_unlock(&ctx->uring_lock);
9392 if (submitted != to_submit)
9395 if (flags & IORING_ENTER_GETEVENTS) {
9396 const sigset_t __user *sig;
9397 struct __kernel_timespec __user *ts;
9399 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9403 min_complete = min(min_complete, ctx->cq_entries);
9406 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9407 * space applications don't need to do io completion events
9408 * polling again, they can rely on io_sq_thread to do polling
9409 * work, which can reduce cpu usage and uring_lock contention.
9411 if (ctx->flags & IORING_SETUP_IOPOLL &&
9412 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9413 ret = io_iopoll_check(ctx, min_complete);
9415 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9420 percpu_ref_put(&ctx->refs);
9423 return submitted ? submitted : ret;
9426 #ifdef CONFIG_PROC_FS
9427 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9428 const struct cred *cred)
9430 struct user_namespace *uns = seq_user_ns(m);
9431 struct group_info *gi;
9436 seq_printf(m, "%5d\n", id);
9437 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9438 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9439 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9440 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9441 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9442 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9443 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9444 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9445 seq_puts(m, "\n\tGroups:\t");
9446 gi = cred->group_info;
9447 for (g = 0; g < gi->ngroups; g++) {
9448 seq_put_decimal_ull(m, g ? " " : "",
9449 from_kgid_munged(uns, gi->gid[g]));
9451 seq_puts(m, "\n\tCapEff:\t");
9452 cap = cred->cap_effective;
9453 CAP_FOR_EACH_U32(__capi)
9454 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9459 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9461 struct io_sq_data *sq = NULL;
9466 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9467 * since fdinfo case grabs it in the opposite direction of normal use
9468 * cases. If we fail to get the lock, we just don't iterate any
9469 * structures that could be going away outside the io_uring mutex.
9471 has_lock = mutex_trylock(&ctx->uring_lock);
9473 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9479 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9480 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9481 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9482 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9483 struct file *f = io_file_from_index(ctx, i);
9486 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9488 seq_printf(m, "%5u: <none>\n", i);
9490 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9491 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9492 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9493 unsigned int len = buf->ubuf_end - buf->ubuf;
9495 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9497 if (has_lock && !xa_empty(&ctx->personalities)) {
9498 unsigned long index;
9499 const struct cred *cred;
9501 seq_printf(m, "Personalities:\n");
9502 xa_for_each(&ctx->personalities, index, cred)
9503 io_uring_show_cred(m, index, cred);
9505 seq_printf(m, "PollList:\n");
9506 spin_lock_irq(&ctx->completion_lock);
9507 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9508 struct hlist_head *list = &ctx->cancel_hash[i];
9509 struct io_kiocb *req;
9511 hlist_for_each_entry(req, list, hash_node)
9512 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9513 req->task->task_works != NULL);
9515 spin_unlock_irq(&ctx->completion_lock);
9517 mutex_unlock(&ctx->uring_lock);
9520 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9522 struct io_ring_ctx *ctx = f->private_data;
9524 if (percpu_ref_tryget(&ctx->refs)) {
9525 __io_uring_show_fdinfo(ctx, m);
9526 percpu_ref_put(&ctx->refs);
9531 static const struct file_operations io_uring_fops = {
9532 .release = io_uring_release,
9533 .mmap = io_uring_mmap,
9535 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9536 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9538 .poll = io_uring_poll,
9539 .fasync = io_uring_fasync,
9540 #ifdef CONFIG_PROC_FS
9541 .show_fdinfo = io_uring_show_fdinfo,
9545 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9546 struct io_uring_params *p)
9548 struct io_rings *rings;
9549 size_t size, sq_array_offset;
9551 /* make sure these are sane, as we already accounted them */
9552 ctx->sq_entries = p->sq_entries;
9553 ctx->cq_entries = p->cq_entries;
9555 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9556 if (size == SIZE_MAX)
9559 rings = io_mem_alloc(size);
9564 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9565 rings->sq_ring_mask = p->sq_entries - 1;
9566 rings->cq_ring_mask = p->cq_entries - 1;
9567 rings->sq_ring_entries = p->sq_entries;
9568 rings->cq_ring_entries = p->cq_entries;
9570 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9571 if (size == SIZE_MAX) {
9572 io_mem_free(ctx->rings);
9577 ctx->sq_sqes = io_mem_alloc(size);
9578 if (!ctx->sq_sqes) {
9579 io_mem_free(ctx->rings);
9587 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9591 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9595 ret = io_uring_add_tctx_node(ctx);
9600 fd_install(fd, file);
9605 * Allocate an anonymous fd, this is what constitutes the application
9606 * visible backing of an io_uring instance. The application mmaps this
9607 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9608 * we have to tie this fd to a socket for file garbage collection purposes.
9610 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9613 #if defined(CONFIG_UNIX)
9616 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9619 return ERR_PTR(ret);
9622 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9623 O_RDWR | O_CLOEXEC);
9624 #if defined(CONFIG_UNIX)
9626 sock_release(ctx->ring_sock);
9627 ctx->ring_sock = NULL;
9629 ctx->ring_sock->file = file;
9635 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9636 struct io_uring_params __user *params)
9638 struct io_ring_ctx *ctx;
9644 if (entries > IORING_MAX_ENTRIES) {
9645 if (!(p->flags & IORING_SETUP_CLAMP))
9647 entries = IORING_MAX_ENTRIES;
9651 * Use twice as many entries for the CQ ring. It's possible for the
9652 * application to drive a higher depth than the size of the SQ ring,
9653 * since the sqes are only used at submission time. This allows for
9654 * some flexibility in overcommitting a bit. If the application has
9655 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9656 * of CQ ring entries manually.
9658 p->sq_entries = roundup_pow_of_two(entries);
9659 if (p->flags & IORING_SETUP_CQSIZE) {
9661 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9662 * to a power-of-two, if it isn't already. We do NOT impose
9663 * any cq vs sq ring sizing.
9667 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9668 if (!(p->flags & IORING_SETUP_CLAMP))
9670 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9672 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9673 if (p->cq_entries < p->sq_entries)
9676 p->cq_entries = 2 * p->sq_entries;
9679 ctx = io_ring_ctx_alloc(p);
9682 ctx->compat = in_compat_syscall();
9683 if (!capable(CAP_IPC_LOCK))
9684 ctx->user = get_uid(current_user());
9687 * This is just grabbed for accounting purposes. When a process exits,
9688 * the mm is exited and dropped before the files, hence we need to hang
9689 * on to this mm purely for the purposes of being able to unaccount
9690 * memory (locked/pinned vm). It's not used for anything else.
9692 mmgrab(current->mm);
9693 ctx->mm_account = current->mm;
9695 ret = io_allocate_scq_urings(ctx, p);
9699 ret = io_sq_offload_create(ctx, p);
9702 /* always set a rsrc node */
9703 ret = io_rsrc_node_switch_start(ctx);
9706 io_rsrc_node_switch(ctx, NULL);
9708 memset(&p->sq_off, 0, sizeof(p->sq_off));
9709 p->sq_off.head = offsetof(struct io_rings, sq.head);
9710 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9711 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9712 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9713 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9714 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9715 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9717 memset(&p->cq_off, 0, sizeof(p->cq_off));
9718 p->cq_off.head = offsetof(struct io_rings, cq.head);
9719 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9720 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9721 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9722 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9723 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9724 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9726 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9727 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9728 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9729 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9730 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9731 IORING_FEAT_RSRC_TAGS;
9733 if (copy_to_user(params, p, sizeof(*p))) {
9738 file = io_uring_get_file(ctx);
9740 ret = PTR_ERR(file);
9745 * Install ring fd as the very last thing, so we don't risk someone
9746 * having closed it before we finish setup
9748 ret = io_uring_install_fd(ctx, file);
9750 /* fput will clean it up */
9755 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9758 io_ring_ctx_wait_and_kill(ctx);
9763 * Sets up an aio uring context, and returns the fd. Applications asks for a
9764 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9765 * params structure passed in.
9767 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9769 struct io_uring_params p;
9772 if (copy_from_user(&p, params, sizeof(p)))
9774 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9779 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9780 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9781 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9782 IORING_SETUP_R_DISABLED))
9785 return io_uring_create(entries, &p, params);
9788 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9789 struct io_uring_params __user *, params)
9791 return io_uring_setup(entries, params);
9794 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9796 struct io_uring_probe *p;
9800 size = struct_size(p, ops, nr_args);
9801 if (size == SIZE_MAX)
9803 p = kzalloc(size, GFP_KERNEL);
9808 if (copy_from_user(p, arg, size))
9811 if (memchr_inv(p, 0, size))
9814 p->last_op = IORING_OP_LAST - 1;
9815 if (nr_args > IORING_OP_LAST)
9816 nr_args = IORING_OP_LAST;
9818 for (i = 0; i < nr_args; i++) {
9820 if (!io_op_defs[i].not_supported)
9821 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9826 if (copy_to_user(arg, p, size))
9833 static int io_register_personality(struct io_ring_ctx *ctx)
9835 const struct cred *creds;
9839 creds = get_current_cred();
9841 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9842 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9849 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9850 unsigned int nr_args)
9852 struct io_uring_restriction *res;
9856 /* Restrictions allowed only if rings started disabled */
9857 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9860 /* We allow only a single restrictions registration */
9861 if (ctx->restrictions.registered)
9864 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9867 size = array_size(nr_args, sizeof(*res));
9868 if (size == SIZE_MAX)
9871 res = memdup_user(arg, size);
9873 return PTR_ERR(res);
9877 for (i = 0; i < nr_args; i++) {
9878 switch (res[i].opcode) {
9879 case IORING_RESTRICTION_REGISTER_OP:
9880 if (res[i].register_op >= IORING_REGISTER_LAST) {
9885 __set_bit(res[i].register_op,
9886 ctx->restrictions.register_op);
9888 case IORING_RESTRICTION_SQE_OP:
9889 if (res[i].sqe_op >= IORING_OP_LAST) {
9894 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9896 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9897 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9899 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9900 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9909 /* Reset all restrictions if an error happened */
9911 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9913 ctx->restrictions.registered = true;
9919 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9921 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9924 if (ctx->restrictions.registered)
9925 ctx->restricted = 1;
9927 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9928 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9929 wake_up(&ctx->sq_data->wait);
9933 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9934 struct io_uring_rsrc_update2 *up,
9942 if (check_add_overflow(up->offset, nr_args, &tmp))
9944 err = io_rsrc_node_switch_start(ctx);
9949 case IORING_RSRC_FILE:
9950 return __io_sqe_files_update(ctx, up, nr_args);
9951 case IORING_RSRC_BUFFER:
9952 return __io_sqe_buffers_update(ctx, up, nr_args);
9957 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9960 struct io_uring_rsrc_update2 up;
9964 memset(&up, 0, sizeof(up));
9965 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9967 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9970 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9971 unsigned size, unsigned type)
9973 struct io_uring_rsrc_update2 up;
9975 if (size != sizeof(up))
9977 if (copy_from_user(&up, arg, sizeof(up)))
9979 if (!up.nr || up.resv)
9981 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9984 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9985 unsigned int size, unsigned int type)
9987 struct io_uring_rsrc_register rr;
9989 /* keep it extendible */
9990 if (size != sizeof(rr))
9993 memset(&rr, 0, sizeof(rr));
9994 if (copy_from_user(&rr, arg, size))
9996 if (!rr.nr || rr.resv || rr.resv2)
10000 case IORING_RSRC_FILE:
10001 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10002 rr.nr, u64_to_user_ptr(rr.tags));
10003 case IORING_RSRC_BUFFER:
10004 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10005 rr.nr, u64_to_user_ptr(rr.tags));
10010 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10013 struct io_uring_task *tctx = current->io_uring;
10014 cpumask_var_t new_mask;
10017 if (!tctx || !tctx->io_wq)
10020 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10023 cpumask_clear(new_mask);
10024 if (len > cpumask_size())
10025 len = cpumask_size();
10027 if (copy_from_user(new_mask, arg, len)) {
10028 free_cpumask_var(new_mask);
10032 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10033 free_cpumask_var(new_mask);
10037 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10039 struct io_uring_task *tctx = current->io_uring;
10041 if (!tctx || !tctx->io_wq)
10044 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10047 static bool io_register_op_must_quiesce(int op)
10050 case IORING_REGISTER_BUFFERS:
10051 case IORING_UNREGISTER_BUFFERS:
10052 case IORING_REGISTER_FILES:
10053 case IORING_UNREGISTER_FILES:
10054 case IORING_REGISTER_FILES_UPDATE:
10055 case IORING_REGISTER_PROBE:
10056 case IORING_REGISTER_PERSONALITY:
10057 case IORING_UNREGISTER_PERSONALITY:
10058 case IORING_REGISTER_FILES2:
10059 case IORING_REGISTER_FILES_UPDATE2:
10060 case IORING_REGISTER_BUFFERS2:
10061 case IORING_REGISTER_BUFFERS_UPDATE:
10062 case IORING_REGISTER_IOWQ_AFF:
10063 case IORING_UNREGISTER_IOWQ_AFF:
10070 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10071 void __user *arg, unsigned nr_args)
10072 __releases(ctx->uring_lock)
10073 __acquires(ctx->uring_lock)
10078 * We're inside the ring mutex, if the ref is already dying, then
10079 * someone else killed the ctx or is already going through
10080 * io_uring_register().
10082 if (percpu_ref_is_dying(&ctx->refs))
10085 if (ctx->restricted) {
10086 if (opcode >= IORING_REGISTER_LAST)
10088 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10089 if (!test_bit(opcode, ctx->restrictions.register_op))
10093 if (io_register_op_must_quiesce(opcode)) {
10094 percpu_ref_kill(&ctx->refs);
10097 * Drop uring mutex before waiting for references to exit. If
10098 * another thread is currently inside io_uring_enter() it might
10099 * need to grab the uring_lock to make progress. If we hold it
10100 * here across the drain wait, then we can deadlock. It's safe
10101 * to drop the mutex here, since no new references will come in
10102 * after we've killed the percpu ref.
10104 mutex_unlock(&ctx->uring_lock);
10106 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10109 ret = io_run_task_work_sig();
10113 mutex_lock(&ctx->uring_lock);
10116 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10122 case IORING_REGISTER_BUFFERS:
10123 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10125 case IORING_UNREGISTER_BUFFERS:
10127 if (arg || nr_args)
10129 ret = io_sqe_buffers_unregister(ctx);
10131 case IORING_REGISTER_FILES:
10132 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10134 case IORING_UNREGISTER_FILES:
10136 if (arg || nr_args)
10138 ret = io_sqe_files_unregister(ctx);
10140 case IORING_REGISTER_FILES_UPDATE:
10141 ret = io_register_files_update(ctx, arg, nr_args);
10143 case IORING_REGISTER_EVENTFD:
10144 case IORING_REGISTER_EVENTFD_ASYNC:
10148 ret = io_eventfd_register(ctx, arg);
10151 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10152 ctx->eventfd_async = 1;
10154 ctx->eventfd_async = 0;
10156 case IORING_UNREGISTER_EVENTFD:
10158 if (arg || nr_args)
10160 ret = io_eventfd_unregister(ctx);
10162 case IORING_REGISTER_PROBE:
10164 if (!arg || nr_args > 256)
10166 ret = io_probe(ctx, arg, nr_args);
10168 case IORING_REGISTER_PERSONALITY:
10170 if (arg || nr_args)
10172 ret = io_register_personality(ctx);
10174 case IORING_UNREGISTER_PERSONALITY:
10178 ret = io_unregister_personality(ctx, nr_args);
10180 case IORING_REGISTER_ENABLE_RINGS:
10182 if (arg || nr_args)
10184 ret = io_register_enable_rings(ctx);
10186 case IORING_REGISTER_RESTRICTIONS:
10187 ret = io_register_restrictions(ctx, arg, nr_args);
10189 case IORING_REGISTER_FILES2:
10190 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10192 case IORING_REGISTER_FILES_UPDATE2:
10193 ret = io_register_rsrc_update(ctx, arg, nr_args,
10196 case IORING_REGISTER_BUFFERS2:
10197 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10199 case IORING_REGISTER_BUFFERS_UPDATE:
10200 ret = io_register_rsrc_update(ctx, arg, nr_args,
10201 IORING_RSRC_BUFFER);
10203 case IORING_REGISTER_IOWQ_AFF:
10205 if (!arg || !nr_args)
10207 ret = io_register_iowq_aff(ctx, arg, nr_args);
10209 case IORING_UNREGISTER_IOWQ_AFF:
10211 if (arg || nr_args)
10213 ret = io_unregister_iowq_aff(ctx);
10220 if (io_register_op_must_quiesce(opcode)) {
10221 /* bring the ctx back to life */
10222 percpu_ref_reinit(&ctx->refs);
10223 reinit_completion(&ctx->ref_comp);
10228 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10229 void __user *, arg, unsigned int, nr_args)
10231 struct io_ring_ctx *ctx;
10240 if (f.file->f_op != &io_uring_fops)
10243 ctx = f.file->private_data;
10245 io_run_task_work();
10247 mutex_lock(&ctx->uring_lock);
10248 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10249 mutex_unlock(&ctx->uring_lock);
10250 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10251 ctx->cq_ev_fd != NULL, ret);
10257 static int __init io_uring_init(void)
10259 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10260 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10261 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10264 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10265 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10266 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10267 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10268 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10269 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10270 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10271 BUILD_BUG_SQE_ELEM(8, __u64, off);
10272 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10273 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10274 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10275 BUILD_BUG_SQE_ELEM(24, __u32, len);
10276 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10277 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10278 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10279 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10280 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10281 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10282 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10283 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10284 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10285 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10286 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10287 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10288 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10289 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10290 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10291 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10292 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10293 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10294 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10295 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10297 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10298 sizeof(struct io_uring_rsrc_update));
10299 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10300 sizeof(struct io_uring_rsrc_update2));
10301 /* should fit into one byte */
10302 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10304 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10305 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10307 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10311 __initcall(io_uring_init);