| rjw | 1f88458 | 2022-01-06 17:20:42 +0800 | [diff] [blame] | 1 | // SPDX-License-Identifier: GPL-2.0 | 
|  | 2 | /* | 
|  | 3 | * | 
|  | 4 | * sched-messaging.c | 
|  | 5 | * | 
|  | 6 | * messaging: Benchmark for scheduler and IPC mechanisms | 
|  | 7 | * | 
|  | 8 | * Based on hackbench by Rusty Russell <rusty@rustcorp.com.au> | 
|  | 9 | * Ported to perf by Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp> | 
|  | 10 | * | 
|  | 11 | */ | 
|  | 12 |  | 
|  | 13 | #include "../perf.h" | 
|  | 14 | #include "../util/util.h" | 
|  | 15 | #include <subcmd/parse-options.h> | 
|  | 16 | #include "../builtin.h" | 
|  | 17 | #include "bench.h" | 
|  | 18 |  | 
|  | 19 | /* Test groups of 20 processes spraying to 20 receivers */ | 
|  | 20 | #include <pthread.h> | 
|  | 21 | #include <stdio.h> | 
|  | 22 | #include <stdlib.h> | 
|  | 23 | #include <string.h> | 
|  | 24 | #include <errno.h> | 
|  | 25 | #include <unistd.h> | 
|  | 26 | #include <sys/types.h> | 
|  | 27 | #include <sys/socket.h> | 
|  | 28 | #include <sys/wait.h> | 
|  | 29 | #include <sys/time.h> | 
|  | 30 | #include <poll.h> | 
|  | 31 | #include <limits.h> | 
|  | 32 | #include <err.h> | 
|  | 33 | #include <linux/time64.h> | 
|  | 34 |  | 
|  | 35 | #define DATASIZE 100 | 
|  | 36 |  | 
|  | 37 | static bool use_pipes = false; | 
|  | 38 | static unsigned int nr_loops = 100; | 
|  | 39 | static bool thread_mode = false; | 
|  | 40 | static unsigned int num_groups = 10; | 
|  | 41 |  | 
|  | 42 | struct sender_context { | 
|  | 43 | unsigned int num_fds; | 
|  | 44 | int ready_out; | 
|  | 45 | int wakefd; | 
|  | 46 | int out_fds[0]; | 
|  | 47 | }; | 
|  | 48 |  | 
|  | 49 | struct receiver_context { | 
|  | 50 | unsigned int num_packets; | 
|  | 51 | int in_fds[2]; | 
|  | 52 | int ready_out; | 
|  | 53 | int wakefd; | 
|  | 54 | }; | 
|  | 55 |  | 
|  | 56 | static void fdpair(int fds[2]) | 
|  | 57 | { | 
|  | 58 | if (use_pipes) { | 
|  | 59 | if (pipe(fds) == 0) | 
|  | 60 | return; | 
|  | 61 | } else { | 
|  | 62 | if (socketpair(AF_UNIX, SOCK_STREAM, 0, fds) == 0) | 
|  | 63 | return; | 
|  | 64 | } | 
|  | 65 |  | 
|  | 66 | err(EXIT_FAILURE, use_pipes ? "pipe()" : "socketpair()"); | 
|  | 67 | } | 
|  | 68 |  | 
|  | 69 | /* Block until we're ready to go */ | 
|  | 70 | static void ready(int ready_out, int wakefd) | 
|  | 71 | { | 
|  | 72 | char dummy; | 
|  | 73 | struct pollfd pollfd = { .fd = wakefd, .events = POLLIN }; | 
|  | 74 |  | 
|  | 75 | /* Tell them we're ready. */ | 
|  | 76 | if (write(ready_out, &dummy, 1) != 1) | 
|  | 77 | err(EXIT_FAILURE, "CLIENT: ready write"); | 
|  | 78 |  | 
|  | 79 | /* Wait for "GO" signal */ | 
|  | 80 | if (poll(&pollfd, 1, -1) != 1) | 
|  | 81 | err(EXIT_FAILURE, "poll"); | 
|  | 82 | } | 
|  | 83 |  | 
|  | 84 | /* Sender sprays nr_loops messages down each file descriptor */ | 
|  | 85 | static void *sender(struct sender_context *ctx) | 
|  | 86 | { | 
|  | 87 | char data[DATASIZE]; | 
|  | 88 | unsigned int i, j; | 
|  | 89 |  | 
|  | 90 | ready(ctx->ready_out, ctx->wakefd); | 
|  | 91 |  | 
|  | 92 | /* Now pump to every receiver. */ | 
|  | 93 | for (i = 0; i < nr_loops; i++) { | 
|  | 94 | for (j = 0; j < ctx->num_fds; j++) { | 
|  | 95 | int ret, done = 0; | 
|  | 96 |  | 
|  | 97 | again: | 
|  | 98 | ret = write(ctx->out_fds[j], data + done, | 
|  | 99 | sizeof(data)-done); | 
|  | 100 | if (ret < 0) | 
|  | 101 | err(EXIT_FAILURE, "SENDER: write"); | 
|  | 102 | done += ret; | 
|  | 103 | if (done < DATASIZE) | 
|  | 104 | goto again; | 
|  | 105 | } | 
|  | 106 | } | 
|  | 107 |  | 
|  | 108 | return NULL; | 
|  | 109 | } | 
|  | 110 |  | 
|  | 111 |  | 
|  | 112 | /* One receiver per fd */ | 
|  | 113 | static void *receiver(struct receiver_context* ctx) | 
|  | 114 | { | 
|  | 115 | unsigned int i; | 
|  | 116 |  | 
|  | 117 | if (!thread_mode) | 
|  | 118 | close(ctx->in_fds[1]); | 
|  | 119 |  | 
|  | 120 | /* Wait for start... */ | 
|  | 121 | ready(ctx->ready_out, ctx->wakefd); | 
|  | 122 |  | 
|  | 123 | /* Receive them all */ | 
|  | 124 | for (i = 0; i < ctx->num_packets; i++) { | 
|  | 125 | char data[DATASIZE]; | 
|  | 126 | int ret, done = 0; | 
|  | 127 |  | 
|  | 128 | again: | 
|  | 129 | ret = read(ctx->in_fds[0], data + done, DATASIZE - done); | 
|  | 130 | if (ret < 0) | 
|  | 131 | err(EXIT_FAILURE, "SERVER: read"); | 
|  | 132 | done += ret; | 
|  | 133 | if (done < DATASIZE) | 
|  | 134 | goto again; | 
|  | 135 | } | 
|  | 136 |  | 
|  | 137 | return NULL; | 
|  | 138 | } | 
|  | 139 |  | 
|  | 140 | static pthread_t create_worker(void *ctx, void *(*func)(void *)) | 
|  | 141 | { | 
|  | 142 | pthread_attr_t attr; | 
|  | 143 | pthread_t childid; | 
|  | 144 | int ret; | 
|  | 145 |  | 
|  | 146 | if (!thread_mode) { | 
|  | 147 | /* process mode */ | 
|  | 148 | /* Fork the receiver. */ | 
|  | 149 | switch (fork()) { | 
|  | 150 | case -1: | 
|  | 151 | err(EXIT_FAILURE, "fork()"); | 
|  | 152 | break; | 
|  | 153 | case 0: | 
|  | 154 | (*func) (ctx); | 
|  | 155 | exit(0); | 
|  | 156 | break; | 
|  | 157 | default: | 
|  | 158 | break; | 
|  | 159 | } | 
|  | 160 |  | 
|  | 161 | return (pthread_t)0; | 
|  | 162 | } | 
|  | 163 |  | 
|  | 164 | if (pthread_attr_init(&attr) != 0) | 
|  | 165 | err(EXIT_FAILURE, "pthread_attr_init:"); | 
|  | 166 |  | 
|  | 167 | #ifndef __ia64__ | 
|  | 168 | if (pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN) != 0) | 
|  | 169 | err(EXIT_FAILURE, "pthread_attr_setstacksize"); | 
|  | 170 | #endif | 
|  | 171 |  | 
|  | 172 | ret = pthread_create(&childid, &attr, func, ctx); | 
|  | 173 | if (ret != 0) | 
|  | 174 | err(EXIT_FAILURE, "pthread_create failed"); | 
|  | 175 |  | 
|  | 176 | return childid; | 
|  | 177 | } | 
|  | 178 |  | 
|  | 179 | static void reap_worker(pthread_t id) | 
|  | 180 | { | 
|  | 181 | int proc_status; | 
|  | 182 | void *thread_status; | 
|  | 183 |  | 
|  | 184 | if (!thread_mode) { | 
|  | 185 | /* process mode */ | 
|  | 186 | wait(&proc_status); | 
|  | 187 | if (!WIFEXITED(proc_status)) | 
|  | 188 | exit(1); | 
|  | 189 | } else { | 
|  | 190 | pthread_join(id, &thread_status); | 
|  | 191 | } | 
|  | 192 | } | 
|  | 193 |  | 
|  | 194 | /* One group of senders and receivers */ | 
|  | 195 | static unsigned int group(pthread_t *pth, | 
|  | 196 | unsigned int num_fds, | 
|  | 197 | int ready_out, | 
|  | 198 | int wakefd) | 
|  | 199 | { | 
|  | 200 | unsigned int i; | 
|  | 201 | struct sender_context *snd_ctx = malloc(sizeof(struct sender_context) | 
|  | 202 | + num_fds * sizeof(int)); | 
|  | 203 |  | 
|  | 204 | if (!snd_ctx) | 
|  | 205 | err(EXIT_FAILURE, "malloc()"); | 
|  | 206 |  | 
|  | 207 | for (i = 0; i < num_fds; i++) { | 
|  | 208 | int fds[2]; | 
|  | 209 | struct receiver_context *ctx = malloc(sizeof(*ctx)); | 
|  | 210 |  | 
|  | 211 | if (!ctx) | 
|  | 212 | err(EXIT_FAILURE, "malloc()"); | 
|  | 213 |  | 
|  | 214 |  | 
|  | 215 | /* Create the pipe between client and server */ | 
|  | 216 | fdpair(fds); | 
|  | 217 |  | 
|  | 218 | ctx->num_packets = num_fds * nr_loops; | 
|  | 219 | ctx->in_fds[0] = fds[0]; | 
|  | 220 | ctx->in_fds[1] = fds[1]; | 
|  | 221 | ctx->ready_out = ready_out; | 
|  | 222 | ctx->wakefd = wakefd; | 
|  | 223 |  | 
|  | 224 | pth[i] = create_worker(ctx, (void *)receiver); | 
|  | 225 |  | 
|  | 226 | snd_ctx->out_fds[i] = fds[1]; | 
|  | 227 | if (!thread_mode) | 
|  | 228 | close(fds[0]); | 
|  | 229 | } | 
|  | 230 |  | 
|  | 231 | /* Now we have all the fds, fork the senders */ | 
|  | 232 | for (i = 0; i < num_fds; i++) { | 
|  | 233 | snd_ctx->ready_out = ready_out; | 
|  | 234 | snd_ctx->wakefd = wakefd; | 
|  | 235 | snd_ctx->num_fds = num_fds; | 
|  | 236 |  | 
|  | 237 | pth[num_fds+i] = create_worker(snd_ctx, (void *)sender); | 
|  | 238 | } | 
|  | 239 |  | 
|  | 240 | /* Close the fds we have left */ | 
|  | 241 | if (!thread_mode) | 
|  | 242 | for (i = 0; i < num_fds; i++) | 
|  | 243 | close(snd_ctx->out_fds[i]); | 
|  | 244 |  | 
|  | 245 | /* Return number of children to reap */ | 
|  | 246 | return num_fds * 2; | 
|  | 247 | } | 
|  | 248 |  | 
|  | 249 | static const struct option options[] = { | 
|  | 250 | OPT_BOOLEAN('p', "pipe", &use_pipes, | 
|  | 251 | "Use pipe() instead of socketpair()"), | 
|  | 252 | OPT_BOOLEAN('t', "thread", &thread_mode, | 
|  | 253 | "Be multi thread instead of multi process"), | 
|  | 254 | OPT_UINTEGER('g', "group", &num_groups, "Specify number of groups"), | 
|  | 255 | OPT_UINTEGER('l', "nr_loops", &nr_loops, "Specify the number of loops to run (default: 100)"), | 
|  | 256 | OPT_END() | 
|  | 257 | }; | 
|  | 258 |  | 
|  | 259 | static const char * const bench_sched_message_usage[] = { | 
|  | 260 | "perf bench sched messaging <options>", | 
|  | 261 | NULL | 
|  | 262 | }; | 
|  | 263 |  | 
|  | 264 | int bench_sched_messaging(int argc, const char **argv) | 
|  | 265 | { | 
|  | 266 | unsigned int i, total_children; | 
|  | 267 | struct timeval start, stop, diff; | 
|  | 268 | unsigned int num_fds = 20; | 
|  | 269 | int readyfds[2], wakefds[2]; | 
|  | 270 | char dummy; | 
|  | 271 | pthread_t *pth_tab; | 
|  | 272 |  | 
|  | 273 | argc = parse_options(argc, argv, options, | 
|  | 274 | bench_sched_message_usage, 0); | 
|  | 275 |  | 
|  | 276 | pth_tab = malloc(num_fds * 2 * num_groups * sizeof(pthread_t)); | 
|  | 277 | if (!pth_tab) | 
|  | 278 | err(EXIT_FAILURE, "main:malloc()"); | 
|  | 279 |  | 
|  | 280 | fdpair(readyfds); | 
|  | 281 | fdpair(wakefds); | 
|  | 282 |  | 
|  | 283 | total_children = 0; | 
|  | 284 | for (i = 0; i < num_groups; i++) | 
|  | 285 | total_children += group(pth_tab+total_children, num_fds, | 
|  | 286 | readyfds[1], wakefds[0]); | 
|  | 287 |  | 
|  | 288 | /* Wait for everyone to be ready */ | 
|  | 289 | for (i = 0; i < total_children; i++) | 
|  | 290 | if (read(readyfds[0], &dummy, 1) != 1) | 
|  | 291 | err(EXIT_FAILURE, "Reading for readyfds"); | 
|  | 292 |  | 
|  | 293 | gettimeofday(&start, NULL); | 
|  | 294 |  | 
|  | 295 | /* Kick them off */ | 
|  | 296 | if (write(wakefds[1], &dummy, 1) != 1) | 
|  | 297 | err(EXIT_FAILURE, "Writing to start them"); | 
|  | 298 |  | 
|  | 299 | /* Reap them all */ | 
|  | 300 | for (i = 0; i < total_children; i++) | 
|  | 301 | reap_worker(pth_tab[i]); | 
|  | 302 |  | 
|  | 303 | gettimeofday(&stop, NULL); | 
|  | 304 |  | 
|  | 305 | timersub(&stop, &start, &diff); | 
|  | 306 |  | 
|  | 307 | switch (bench_format) { | 
|  | 308 | case BENCH_FORMAT_DEFAULT: | 
|  | 309 | printf("# %d sender and receiver %s per group\n", | 
|  | 310 | num_fds, thread_mode ? "threads" : "processes"); | 
|  | 311 | printf("# %d groups == %d %s run\n\n", | 
|  | 312 | num_groups, num_groups * 2 * num_fds, | 
|  | 313 | thread_mode ? "threads" : "processes"); | 
|  | 314 | printf(" %14s: %lu.%03lu [sec]\n", "Total time", | 
|  | 315 | diff.tv_sec, | 
|  | 316 | (unsigned long) (diff.tv_usec / USEC_PER_MSEC)); | 
|  | 317 | break; | 
|  | 318 | case BENCH_FORMAT_SIMPLE: | 
|  | 319 | printf("%lu.%03lu\n", diff.tv_sec, | 
|  | 320 | (unsigned long) (diff.tv_usec / USEC_PER_MSEC)); | 
|  | 321 | break; | 
|  | 322 | default: | 
|  | 323 | /* reaching here is something disaster */ | 
|  | 324 | fprintf(stderr, "Unknown format:%d\n", bench_format); | 
|  | 325 | exit(1); | 
|  | 326 | break; | 
|  | 327 | } | 
|  | 328 |  | 
|  | 329 | free(pth_tab); | 
|  | 330 |  | 
|  | 331 | return 0; | 
|  | 332 | } |