1 | // Program to test the optimial batchsize in a single threaded process |
---|
2 | extern "C" { |
---|
3 | #ifndef _GNU_SOURCE /* See feature_test_macros(7) */ |
---|
4 | #define _GNU_SOURCE /* See feature_test_macros(7) */ |
---|
5 | #endif |
---|
6 | #include <errno.h> |
---|
7 | #include <stdio.h> |
---|
8 | #include <stdint.h> |
---|
9 | #include <stdlib.h> |
---|
10 | #include <string.h> |
---|
11 | #include <locale.h> |
---|
12 | #include <getopt.h> |
---|
13 | #include <unistd.h> |
---|
14 | #include <sys/mman.h> |
---|
15 | #include <sys/syscall.h> |
---|
16 | #include <sys/uio.h> |
---|
17 | #include <fcntl.h> |
---|
18 | #include <time.h> // timespec |
---|
19 | #include <sys/time.h> // timeval |
---|
20 | |
---|
21 | #include <linux/io_uring.h> |
---|
22 | } |
---|
23 | |
---|
24 | |
---|
25 | enum { TIMEGRAN = 1000000000LL }; // nanosecond granularity, except for timeval |
---|
26 | |
---|
27 | #include <omp.h> |
---|
28 | |
---|
29 | # ifndef __NR_io_uring_setup |
---|
30 | # define __NR_io_uring_setup 425 |
---|
31 | # endif |
---|
32 | # ifndef __NR_io_uring_enter |
---|
33 | # define __NR_io_uring_enter 426 |
---|
34 | # endif |
---|
35 | # ifndef __NR_io_uring_register |
---|
36 | # define __NR_io_uring_register 427 |
---|
37 | # endif |
---|
38 | |
---|
39 | struct io_uring_sq { |
---|
40 | // Head and tail of the ring (associated with array) |
---|
41 | volatile uint32_t * head; |
---|
42 | volatile uint32_t * tail; |
---|
43 | |
---|
44 | // The actual kernel ring which uses head/tail |
---|
45 | // indexes into the sqes arrays |
---|
46 | uint32_t * array; |
---|
47 | |
---|
48 | // number of entries and mask to go with it |
---|
49 | const uint32_t * num; |
---|
50 | const uint32_t * mask; |
---|
51 | |
---|
52 | // Submission flags (Not sure what for) |
---|
53 | uint32_t * flags; |
---|
54 | |
---|
55 | // number of sqes not submitted (whatever that means) |
---|
56 | uint32_t * dropped; |
---|
57 | |
---|
58 | // Like head/tail but not seen by the kernel |
---|
59 | volatile uint32_t alloc; |
---|
60 | |
---|
61 | // A buffer of sqes (not the actual ring) |
---|
62 | struct io_uring_sqe * sqes; |
---|
63 | |
---|
64 | // The location and size of the mmaped area |
---|
65 | void * ring_ptr; |
---|
66 | size_t ring_sz; |
---|
67 | }; |
---|
68 | |
---|
69 | struct io_uring_cq { |
---|
70 | // Head and tail of the ring |
---|
71 | volatile uint32_t * head; |
---|
72 | volatile uint32_t * tail; |
---|
73 | |
---|
74 | // number of entries and mask to go with it |
---|
75 | const uint32_t * mask; |
---|
76 | const uint32_t * num; |
---|
77 | |
---|
78 | // number of cqes not submitted (whatever that means) |
---|
79 | uint32_t * overflow; |
---|
80 | |
---|
81 | // the kernel ring |
---|
82 | struct io_uring_cqe * cqes; |
---|
83 | |
---|
84 | // The location and size of the mmaped area |
---|
85 | void * ring_ptr; |
---|
86 | size_t ring_sz; |
---|
87 | }; |
---|
88 | |
---|
89 | struct io_ring { |
---|
90 | struct io_uring_sq submit_q; |
---|
91 | struct io_uring_cq completion_q; |
---|
92 | uint32_t flags; |
---|
93 | int fd; |
---|
94 | }; |
---|
95 | |
---|
96 | struct fred { |
---|
97 | io_ring io; |
---|
98 | }; |
---|
99 | |
---|
100 | fred self; |
---|
101 | int myfd; |
---|
102 | |
---|
103 | long long unsigned submits = 0; |
---|
104 | long long unsigned completes = 0; |
---|
105 | |
---|
106 | void submit_and_drain(struct iovec * iov, int n) { |
---|
107 | for(int i = 0; i < n; i++) { |
---|
108 | struct io_uring_sqe * sqe = &self.io.submit_q.sqes[ 0 ]; |
---|
109 | |
---|
110 | sqe->opcode = IORING_OP_READV; |
---|
111 | #if !defined(IOSQE_ASYNC) |
---|
112 | sqe->flags = 0; |
---|
113 | #else |
---|
114 | sqe->flags = IOSQE_ASYNC; |
---|
115 | #endif |
---|
116 | sqe->ioprio = 0; |
---|
117 | sqe->fd = myfd; |
---|
118 | sqe->off = 0; |
---|
119 | sqe->addr = (__u64)iov; |
---|
120 | sqe->len = 1; |
---|
121 | sqe->rw_flags = 0; |
---|
122 | sqe->__pad2[0] = sqe->__pad2[1] = sqe->__pad2[2] = 0; |
---|
123 | } |
---|
124 | |
---|
125 | volatile uint32_t * tail = self.io.submit_q.tail; |
---|
126 | __atomic_fetch_add(tail, n, __ATOMIC_SEQ_CST); |
---|
127 | |
---|
128 | int ret = syscall( __NR_io_uring_enter, self.io.fd, n, n, IORING_ENTER_GETEVENTS, nullptr, 0); |
---|
129 | if( ret < 0 ) { |
---|
130 | switch((int)errno) { |
---|
131 | case EAGAIN: |
---|
132 | case EINTR: |
---|
133 | default: |
---|
134 | fprintf(stderr, "KERNEL ERROR: IO_URING WAIT - %s\n", strerror(errno) ); |
---|
135 | abort(); |
---|
136 | } |
---|
137 | } |
---|
138 | |
---|
139 | submits += ret; |
---|
140 | |
---|
141 | uint32_t chead = *self.io.completion_q.head; |
---|
142 | uint32_t ctail = *self.io.completion_q.tail; |
---|
143 | const uint32_t mask = *self.io.completion_q.mask; |
---|
144 | |
---|
145 | // Memory barrier |
---|
146 | __atomic_thread_fence( __ATOMIC_SEQ_CST ); |
---|
147 | |
---|
148 | uint32_t count = ctail - chead; |
---|
149 | __atomic_fetch_add( self.io.completion_q.head, count, __ATOMIC_RELAXED ); |
---|
150 | completes += count; |
---|
151 | } |
---|
152 | |
---|
153 | uint64_t getTimeNsec() { |
---|
154 | timespec curr; |
---|
155 | clock_gettime( CLOCK_REALTIME, &curr ); |
---|
156 | return (int64_t)curr.tv_sec * TIMEGRAN + curr.tv_nsec; |
---|
157 | } |
---|
158 | |
---|
159 | uint64_t to_miliseconds( uint64_t durtn ) { return durtn / (TIMEGRAN / 1000LL); } |
---|
160 | double to_fseconds(uint64_t durtn ) { return durtn / (double)TIMEGRAN; } |
---|
161 | uint64_t from_fseconds(double sec) { return sec * TIMEGRAN; } |
---|
162 | |
---|
163 | int main(int argc, char * argv[]) { |
---|
164 | int buflen = 50; |
---|
165 | int batch = 1; |
---|
166 | double duration = 5; |
---|
167 | |
---|
168 | setlocale(LC_ALL, ""); |
---|
169 | |
---|
170 | for(;;) { |
---|
171 | static struct option options[] = { |
---|
172 | {"duration", required_argument, 0, 'd'}, |
---|
173 | {"batchsize", required_argument, 0, 'b'}, |
---|
174 | {"buflen", required_argument, 0, 'l'}, |
---|
175 | {0, 0, 0, 0} |
---|
176 | }; |
---|
177 | |
---|
178 | int idx = 0; |
---|
179 | int opt = getopt_long(argc, argv, "d:l:b:", options, &idx); |
---|
180 | |
---|
181 | const char * arg = optarg ? optarg : ""; |
---|
182 | char * end; |
---|
183 | switch(opt) { |
---|
184 | // Exit Case |
---|
185 | case -1: |
---|
186 | goto arg_loop; |
---|
187 | case 'd': \ |
---|
188 | duration = strtod(arg, &end); \ |
---|
189 | if(*end != '\0') { \ |
---|
190 | fprintf(stderr, "Duration must be a valid double, was %s\n", arg); \ |
---|
191 | goto usage; \ |
---|
192 | } \ |
---|
193 | break; |
---|
194 | case 'l': |
---|
195 | buflen = strtoul(arg, &end, 10); |
---|
196 | if(*end != '\0' && buflen < 10) { |
---|
197 | fprintf(stderr, "Buffer size must be at least 10, was %s\n", arg); |
---|
198 | goto usage; |
---|
199 | } |
---|
200 | case 'b': |
---|
201 | batch = strtoul(arg, &end, 10); |
---|
202 | if(*end != '\0' && batch < 0) { |
---|
203 | fprintf(stderr, "Batch size must be at least 1, was %s\n", arg); |
---|
204 | goto usage; |
---|
205 | } |
---|
206 | break; |
---|
207 | default: /* ? */ |
---|
208 | fprintf(stderr, "%d\n", opt); |
---|
209 | usage: |
---|
210 | fprintf( stderr, " -l, --buflen=SIZE Number of bytes to read per request\n" ); |
---|
211 | fprintf( stderr, " -b, --batchsize=COUNT Number of request to batch together\n" ); |
---|
212 | exit(EXIT_FAILURE); |
---|
213 | } |
---|
214 | } |
---|
215 | arg_loop: |
---|
216 | |
---|
217 | myfd = open(__FILE__, 0); |
---|
218 | |
---|
219 | // Step 1 : call to setup |
---|
220 | struct io_uring_params params; |
---|
221 | memset(¶ms, 0, sizeof(params)); |
---|
222 | |
---|
223 | uint32_t nentries = 2048; |
---|
224 | |
---|
225 | int fd = syscall(__NR_io_uring_setup, nentries, ¶ms ); |
---|
226 | if(fd < 0) { |
---|
227 | fprintf(stderr, "KERNEL ERROR: IO_URING SETUP - %s\n", strerror(errno)); |
---|
228 | abort(); |
---|
229 | } |
---|
230 | |
---|
231 | // Step 2 : mmap result |
---|
232 | memset(&self.io, 0, sizeof(struct io_ring)); |
---|
233 | struct io_uring_sq & sq = self.io.submit_q; |
---|
234 | struct io_uring_cq & cq = self.io.completion_q; |
---|
235 | |
---|
236 | // calculate the right ring size |
---|
237 | sq.ring_sz = params.sq_off.array + (params.sq_entries * sizeof(unsigned) ); |
---|
238 | cq.ring_sz = params.cq_off.cqes + (params.cq_entries * sizeof(struct io_uring_cqe)); |
---|
239 | |
---|
240 | // Requires features |
---|
241 | // // adjust the size according to the parameters |
---|
242 | // if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) { |
---|
243 | // cq->ring_sz = sq->ring_sz = max(cq->ring_sz, sq->ring_sz); |
---|
244 | // } |
---|
245 | |
---|
246 | // mmap the Submit Queue into existence |
---|
247 | sq.ring_ptr = mmap(0, sq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQ_RING); |
---|
248 | if (sq.ring_ptr == (void*)MAP_FAILED) { |
---|
249 | fprintf(stderr, "KERNEL ERROR: IO_URING MMAP1 - %s\n", strerror(errno)); |
---|
250 | abort(); |
---|
251 | } |
---|
252 | |
---|
253 | // mmap the Completion Queue into existence (may or may not be needed) |
---|
254 | // Requires features |
---|
255 | // if ((params.features & IORING_FEAT_SINGLE_MMAP) != 0) { |
---|
256 | // cq->ring_ptr = sq->ring_ptr; |
---|
257 | // } |
---|
258 | // else { |
---|
259 | // We need multiple call to MMAP |
---|
260 | cq.ring_ptr = mmap(0, cq.ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_CQ_RING); |
---|
261 | if (cq.ring_ptr == (void*)MAP_FAILED) { |
---|
262 | munmap(sq.ring_ptr, sq.ring_sz); |
---|
263 | fprintf(stderr, "KERNEL ERROR: IO_URING MMAP2 - %s\n", strerror(errno)); |
---|
264 | abort(); |
---|
265 | } |
---|
266 | // } |
---|
267 | |
---|
268 | // mmap the submit queue entries |
---|
269 | size_t size = params.sq_entries * sizeof(struct io_uring_sqe); |
---|
270 | sq.sqes = (struct io_uring_sqe *)mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_SQES); |
---|
271 | if (sq.sqes == (struct io_uring_sqe *)MAP_FAILED) { |
---|
272 | munmap(sq.ring_ptr, sq.ring_sz); |
---|
273 | if (cq.ring_ptr != sq.ring_ptr) munmap(cq.ring_ptr, cq.ring_sz); |
---|
274 | fprintf(stderr, "KERNEL ERROR: IO_URING MMAP3 - %s\n", strerror(errno)); |
---|
275 | abort(); |
---|
276 | } |
---|
277 | |
---|
278 | // Get the pointers from the kernel to fill the structure |
---|
279 | // submit queue |
---|
280 | sq.head = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.head); |
---|
281 | sq.tail = (volatile uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.tail); |
---|
282 | sq.mask = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_mask); |
---|
283 | sq.num = ( const uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.ring_entries); |
---|
284 | sq.flags = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.flags); |
---|
285 | sq.dropped = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.dropped); |
---|
286 | sq.array = ( uint32_t *)(((intptr_t)sq.ring_ptr) + params.sq_off.array); |
---|
287 | sq.alloc = *sq.tail; |
---|
288 | |
---|
289 | // completion queue |
---|
290 | cq.head = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.head); |
---|
291 | cq.tail = (volatile uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.tail); |
---|
292 | cq.mask = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_mask); |
---|
293 | cq.num = ( const uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.ring_entries); |
---|
294 | cq.overflow = ( uint32_t *)(((intptr_t)cq.ring_ptr) + params.cq_off.overflow); |
---|
295 | cq.cqes = (struct io_uring_cqe *)(((intptr_t)cq.ring_ptr) + params.cq_off.cqes); |
---|
296 | |
---|
297 | self.io.fd = fd; |
---|
298 | |
---|
299 | // Allocate the sqe |
---|
300 | uint32_t idx = 0; |
---|
301 | |
---|
302 | // Return the sqe |
---|
303 | struct io_uring_sqe * sqe = &self.io.submit_q.sqes[ idx & (*self.io.submit_q.mask)]; |
---|
304 | |
---|
305 | char data[buflen]; |
---|
306 | struct iovec iov = { data, (size_t)buflen }; |
---|
307 | |
---|
308 | sqe->opcode = IORING_OP_READV; |
---|
309 | #if !defined(IOSQE_ASYNC) |
---|
310 | sqe->flags = 0; |
---|
311 | #else |
---|
312 | sqe->flags = IOSQE_ASYNC; |
---|
313 | #endif |
---|
314 | sqe->ioprio = 0; |
---|
315 | sqe->fd = myfd; |
---|
316 | sqe->off = 0; |
---|
317 | sqe->addr = (__u64)&iov; |
---|
318 | sqe->len = 1; |
---|
319 | sqe->rw_flags = 0; |
---|
320 | sqe->__pad2[0] = sqe->__pad2[1] = sqe->__pad2[2] = 0; |
---|
321 | |
---|
322 | // Append to the list of ready entries |
---|
323 | for(unsigned i = 0; i < *self.io.submit_q.num; i++) { |
---|
324 | self.io.submit_q.array[ i ] = 0; |
---|
325 | } |
---|
326 | |
---|
327 | printf("Running for %f second, reading %d bytes in batches of %d\n", duration, buflen, batch); |
---|
328 | uint64_t start = getTimeNsec(); |
---|
329 | uint64_t end = getTimeNsec(); |
---|
330 | uint64_t prev = getTimeNsec(); |
---|
331 | for(;;) { |
---|
332 | submit_and_drain(&iov, batch); |
---|
333 | end = getTimeNsec(); |
---|
334 | uint64_t delta = end - start; |
---|
335 | if( to_fseconds(end - prev) > 0.1 ) { |
---|
336 | printf(" %.1f\r", to_fseconds(delta)); |
---|
337 | fflush(stdout); |
---|
338 | prev = end; |
---|
339 | } |
---|
340 | if( delta >= from_fseconds(duration) ) { |
---|
341 | break; |
---|
342 | } |
---|
343 | } |
---|
344 | |
---|
345 | printf("Took %'ld ms\n", to_miliseconds(end - start)); |
---|
346 | printf("Submitted %'llu\n", submits); |
---|
347 | printf("Completed %'llu\n", completes); |
---|
348 | printf("Submitted / sec %'.f\n", submits / to_fseconds(end - start)); |
---|
349 | printf("Completed / sec %'.f\n", completes / to_fseconds(end - start)); |
---|
350 | } |
---|