1 /* SPDX-License-Identifier: MIT */
2
3 #include <linux/io_uring.h>
4 #include <sys/mman.h>
5 #include <sys/syscall.h>
6 #include <stdio.h>
7 #include <string.h>
8 #include <unistd.h>
9
10 struct io_sq_ring {
11 unsigned int *head;
12 unsigned int *tail;
13 unsigned int *ring_mask;
14 unsigned int *ring_entries;
15 unsigned int *flags;
16 unsigned int *array;
17 };
18
19 struct io_cq_ring {
20 unsigned int *head;
21 unsigned int *tail;
22 unsigned int *ring_mask;
23 unsigned int *ring_entries;
24 struct io_uring_cqe *cqes;
25 };
26
27 struct io_uring_sq {
28 unsigned int *khead;
29 unsigned int *ktail;
30 unsigned int *kring_mask;
31 unsigned int *kring_entries;
32 unsigned int *kflags;
33 unsigned int *kdropped;
34 unsigned int *array;
35 struct io_uring_sqe *sqes;
36
37 unsigned int sqe_head;
38 unsigned int sqe_tail;
39
40 size_t ring_sz;
41 };
42
43 struct io_uring_cq {
44 unsigned int *khead;
45 unsigned int *ktail;
46 unsigned int *kring_mask;
47 unsigned int *kring_entries;
48 unsigned int *koverflow;
49 struct io_uring_cqe *cqes;
50
51 size_t ring_sz;
52 };
53
54 struct io_uring {
55 struct io_uring_sq sq;
56 struct io_uring_cq cq;
57 int ring_fd;
58 };
59
60 #if defined(__x86_64) || defined(__i386__)
61 #define read_barrier() __asm__ __volatile__("":::"memory")
62 #define write_barrier() __asm__ __volatile__("":::"memory")
63 #else
64 #define read_barrier() __sync_synchronize()
65 #define write_barrier() __sync_synchronize()
66 #endif
67
68 static inline int io_uring_mmap(int fd, struct io_uring_params *p,
69 struct io_uring_sq *sq, struct io_uring_cq *cq)
70 {
71 size_t size;
72 void *ptr;
73 int ret;
74
75 sq->ring_sz = p->sq_off.array + p->sq_entries * sizeof(unsigned int);
76 ptr = mmap(0, sq->ring_sz, PROT_READ | PROT_WRITE,
77 MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQ_RING);
78 if (ptr == MAP_FAILED)
79 return -errno;
80 sq->khead = ptr + p->sq_off.head;
81 sq->ktail = ptr + p->sq_off.tail;
82 sq->kring_mask = ptr + p->sq_off.ring_mask;
83 sq->kring_entries = ptr + p->sq_off.ring_entries;
84 sq->kflags = ptr + p->sq_off.flags;
85 sq->kdropped = ptr + p->sq_off.dropped;
86 sq->array = ptr + p->sq_off.array;
87
88 size = p->sq_entries * sizeof(struct io_uring_sqe);
89 sq->sqes = mmap(0, size, PROT_READ | PROT_WRITE,
90 MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQES);
91 if (sq->sqes == MAP_FAILED) {
92 ret = -errno;
93 err:
94 munmap(sq->khead, sq->ring_sz);
95 return ret;
96 }
97
98 cq->ring_sz = p->cq_off.cqes + p->cq_entries * sizeof(struct io_uring_cqe);
99 ptr = mmap(0, cq->ring_sz, PROT_READ | PROT_WRITE,
100 MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_CQ_RING);
101 if (ptr == MAP_FAILED) {
102 ret = -errno;
103 munmap(sq->sqes, p->sq_entries * sizeof(struct io_uring_sqe));
104 goto err;
105 }
106 cq->khead = ptr + p->cq_off.head;
107 cq->ktail = ptr + p->cq_off.tail;
108 cq->kring_mask = ptr + p->cq_off.ring_mask;
109 cq->kring_entries = ptr + p->cq_off.ring_entries;
110 cq->koverflow = ptr + p->cq_off.overflow;
111 cq->cqes = ptr + p->cq_off.cqes;
112 return 0;
113 }
114
115 static inline int io_uring_setup(unsigned int entries,
116 struct io_uring_params *p)
117 {
118 return syscall(__NR_io_uring_setup, entries, p);
119 }
120
121 static inline int io_uring_enter(int fd, unsigned int to_submit,
122 unsigned int min_complete,
123 unsigned int flags, sigset_t *sig)
124 {
125 return syscall(__NR_io_uring_enter, fd, to_submit, min_complete,
126 flags, sig, _NSIG / 8);
127 }
128
129 static inline int io_uring_queue_init(unsigned int entries,
130 struct io_uring *ring,
131 unsigned int flags)
132 {
133 struct io_uring_params p;
134 int fd, ret;
135
136 memset(ring, 0, sizeof(*ring));
137 memset(&p, 0, sizeof(p));
138 p.flags = flags;
139
140 fd = io_uring_setup(entries, &p);
141 if (fd < 0)
142 return fd;
143 ret = io_uring_mmap(fd, &p, &ring->sq, &ring->cq);
144 if (!ret)
145 ring->ring_fd = fd;
146 else
147 close(fd);
148 return ret;
149 }
150
151 /* Get a sqe */
152 static inline struct io_uring_sqe *io_uring_get_sqe(struct io_uring *ring)
153 {
154 struct io_uring_sq *sq = &ring->sq;
155
156 if (sq->sqe_tail + 1 - sq->sqe_head > *sq->kring_entries)
157 return NULL;
158 return &sq->sqes[sq->sqe_tail++ & *sq->kring_mask];
159 }
160
161 static inline int io_uring_wait_cqe(struct io_uring *ring,
162 struct io_uring_cqe **cqe_ptr)
163 {
164 struct io_uring_cq *cq = &ring->cq;
165 const unsigned int mask = *cq->kring_mask;
166 unsigned int head = *cq->khead;
167 int ret;
168
169 *cqe_ptr = NULL;
170 do {
171 read_barrier();
172 if (head != *cq->ktail) {
173 *cqe_ptr = &cq->cqes[head & mask];
174 break;
175 }
176 ret = io_uring_enter(ring->ring_fd, 0, 1,
177 IORING_ENTER_GETEVENTS, NULL);
178 if (ret < 0)
179 return -errno;
180 } while (1);
181
182 return 0;
183 }
184
185 static inline int io_uring_submit(struct io_uring *ring)
186 {
187 struct io_uring_sq *sq = &ring->sq;
188 const unsigned int mask = *sq->kring_mask;
189 unsigned int ktail, submitted, to_submit;
190 int ret;
191
192 read_barrier();
193 if (*sq->khead != *sq->ktail) {
194 submitted = *sq->kring_entries;
195 goto submit;
196 }
197 if (sq->sqe_head == sq->sqe_tail)
198 return 0;
199
200 ktail = *sq->ktail;
201 to_submit = sq->sqe_tail - sq->sqe_head;
202 for (submitted = 0; submitted < to_submit; submitted++) {
203 read_barrier();
204 sq->array[ktail++ & mask] = sq->sqe_head++ & mask;
205 }
206 if (!submitted)
207 return 0;
208
209 if (*sq->ktail != ktail) {
210 write_barrier();
211 *sq->ktail = ktail;
212 write_barrier();
213 }
214 submit:
215 ret = io_uring_enter(ring->ring_fd, submitted, 0,
216 IORING_ENTER_GETEVENTS, NULL);
217 return ret < 0 ? -errno : ret;
218 }
219
220 static inline void io_uring_queue_exit(struct io_uring *ring)
221 {
222 struct io_uring_sq *sq = &ring->sq;
223
224 munmap(sq->sqes, *sq->kring_entries * sizeof(struct io_uring_sqe));
225 munmap(sq->khead, sq->ring_sz);
226 close(ring->ring_fd);
227 }
228
229 /* Prepare and send the SQE */
230 static inline void io_uring_prep_cmd(struct io_uring_sqe *sqe, int op,
231 int sockfd,
232 int level, int optname,
233 const void *optval,
234 int optlen)
235 {
236 memset(sqe, 0, sizeof(*sqe));
237 sqe->opcode = (__u8)IORING_OP_URING_CMD;
238 sqe->fd = sockfd;
239 sqe->cmd_op = op;
240
241 sqe->level = level;
242 sqe->optname = optname;
243 sqe->optval = (unsigned long long)optval;
244 sqe->optlen = optlen;
245 }
246
247 static inline int io_uring_register_buffers(struct io_uring *ring,
248 const struct iovec *iovecs,
249 unsigned int nr_iovecs)
250 {
251 int ret;
252
253 ret = syscall(__NR_io_uring_register, ring->ring_fd,
254 IORING_REGISTER_BUFFERS, iovecs, nr_iovecs);
255 return (ret < 0) ? -errno : ret;
256 }
257
258 static inline void io_uring_prep_send(struct io_uring_sqe *sqe, int sockfd,
259 const void *buf, size_t len, int flags)
260 {
261 memset(sqe, 0, sizeof(*sqe));
262 sqe->opcode = (__u8)IORING_OP_SEND;
263 sqe->fd = sockfd;
264 sqe->addr = (unsigned long)buf;
265 sqe->len = len;
266 sqe->msg_flags = (__u32)flags;
267 }
268
269 static inline void io_uring_prep_sendzc(struct io_uring_sqe *sqe, int sockfd,
270 const void *buf, size_t len, int flags,
271 unsigned int zc_flags)
272 {
273 io_uring_prep_send(sqe, sockfd, buf, len, flags);
274 sqe->opcode = (__u8)IORING_OP_SEND_ZC;
275 sqe->ioprio = zc_flags;
276 }
277
278 static inline void io_uring_cqe_seen(struct io_uring *ring)
279 {
280 *(&ring->cq)->khead += 1;
281 write_barrier();
282 }
283
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.