1 /* 2 * Context switch microbenchmark. 3 * 4 * Copyright (C) 2015 Anton Blanchard <anton@au.ibm.com>, IBM 5 * 6 * This program is free software; you can redistribute it and/or 7 * modify it under the terms of the GNU General Public License 8 * as published by the Free Software Foundation; either version 9 * 2 of the License, or (at your option) any later version. 10 */ 11 12 #define _GNU_SOURCE 13 #include <sched.h> 14 #include <string.h> 15 #include <stdio.h> 16 #include <unistd.h> 17 #include <stdlib.h> 18 #include <getopt.h> 19 #include <signal.h> 20 #include <assert.h> 21 #include <pthread.h> 22 #include <limits.h> 23 #include <sys/time.h> 24 #include <sys/syscall.h> 25 #include <sys/types.h> 26 #include <sys/shm.h> 27 #include <linux/futex.h> 28 #ifdef __powerpc__ 29 #include <altivec.h> 30 #endif 31 #include "utils.h" 32 33 static unsigned int timeout = 30; 34 35 static int touch_vdso; 36 struct timeval tv; 37 38 static int touch_fp = 1; 39 double fp; 40 41 static int touch_vector = 1; 42 vector int a, b, c; 43 44 #ifdef __powerpc__ 45 static int touch_altivec = 1; 46 47 /* 48 * Note: LTO (Link Time Optimisation) doesn't play well with this function 49 * attribute. Be very careful enabling LTO for this test. 50 */ 51 static void __attribute__((__target__("no-vsx"))) altivec_touch_fn(void) 52 { 53 c = a + b; 54 } 55 #endif 56 57 static void touch(void) 58 { 59 if (touch_vdso) 60 gettimeofday(&tv, NULL); 61 62 if (touch_fp) 63 fp += 0.1; 64 65 #ifdef __powerpc__ 66 if (touch_altivec) 67 altivec_touch_fn(); 68 #endif 69 70 if (touch_vector) 71 c = a + b; 72 73 asm volatile("# %0 %1 %2": : "r"(&tv), "r"(&fp), "r"(&c)); 74 } 75 76 static void start_thread_on(void *(*fn)(void *), void *arg, unsigned long cpu) 77 { 78 pthread_t tid; 79 cpu_set_t cpuset; 80 pthread_attr_t attr; 81 82 CPU_ZERO(&cpuset); 83 CPU_SET(cpu, &cpuset); 84 85 pthread_attr_init(&attr); 86 87 if (pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpuset)) { 88 perror("pthread_attr_setaffinity_np"); 89 exit(1); 90 } 91 92 if (pthread_create(&tid, &attr, fn, arg)) { 93 perror("pthread_create"); 94 exit(1); 95 } 96 } 97 98 static void start_process_on(void *(*fn)(void *), void *arg, unsigned long cpu) 99 { 100 int pid; 101 cpu_set_t cpuset; 102 103 pid = fork(); 104 if (pid == -1) { 105 perror("fork"); 106 exit(1); 107 } 108 109 if (pid) 110 return; 111 112 CPU_ZERO(&cpuset); 113 CPU_SET(cpu, &cpuset); 114 115 if (sched_setaffinity(0, sizeof(cpuset), &cpuset)) { 116 perror("sched_setaffinity"); 117 exit(1); 118 } 119 120 fn(arg); 121 122 exit(0); 123 } 124 125 static unsigned long iterations; 126 static unsigned long iterations_prev; 127 128 static void sigalrm_handler(int junk) 129 { 130 unsigned long i = iterations; 131 132 printf("%ld\n", i - iterations_prev); 133 iterations_prev = i; 134 135 if (--timeout == 0) 136 kill(0, SIGUSR1); 137 138 alarm(1); 139 } 140 141 static void sigusr1_handler(int junk) 142 { 143 exit(0); 144 } 145 146 struct actions { 147 void (*setup)(int, int); 148 void *(*thread1)(void *); 149 void *(*thread2)(void *); 150 }; 151 152 #define READ 0 153 #define WRITE 1 154 155 static int pipe_fd1[2]; 156 static int pipe_fd2[2]; 157 158 static void pipe_setup(int cpu1, int cpu2) 159 { 160 if (pipe(pipe_fd1) || pipe(pipe_fd2)) 161 exit(1); 162 } 163 164 static void *pipe_thread1(void *arg) 165 { 166 signal(SIGALRM, sigalrm_handler); 167 alarm(1); 168 169 while (1) { 170 assert(read(pipe_fd1[READ], &c, 1) == 1); 171 touch(); 172 173 assert(write(pipe_fd2[WRITE], &c, 1) == 1); 174 touch(); 175 176 iterations += 2; 177 } 178 179 return NULL; 180 } 181 182 static void *pipe_thread2(void *arg) 183 { 184 while (1) { 185 assert(write(pipe_fd1[WRITE], &c, 1) == 1); 186 touch(); 187 188 assert(read(pipe_fd2[READ], &c, 1) == 1); 189 touch(); 190 } 191 192 return NULL; 193 } 194 195 static struct actions pipe_actions = { 196 .setup = pipe_setup, 197 .thread1 = pipe_thread1, 198 .thread2 = pipe_thread2, 199 }; 200 201 static void yield_setup(int cpu1, int cpu2) 202 { 203 if (cpu1 != cpu2) { 204 fprintf(stderr, "Both threads must be on the same CPU for yield test\n"); 205 exit(1); 206 } 207 } 208 209 static void *yield_thread1(void *arg) 210 { 211 signal(SIGALRM, sigalrm_handler); 212 alarm(1); 213 214 while (1) { 215 sched_yield(); 216 touch(); 217 218 iterations += 2; 219 } 220 221 return NULL; 222 } 223 224 static void *yield_thread2(void *arg) 225 { 226 while (1) { 227 sched_yield(); 228 touch(); 229 } 230 231 return NULL; 232 } 233 234 static struct actions yield_actions = { 235 .setup = yield_setup, 236 .thread1 = yield_thread1, 237 .thread2 = yield_thread2, 238 }; 239 240 static long sys_futex(void *addr1, int op, int val1, struct timespec *timeout, 241 void *addr2, int val3) 242 { 243 return syscall(SYS_futex, addr1, op, val1, timeout, addr2, val3); 244 } 245 246 static unsigned long cmpxchg(unsigned long *p, unsigned long expected, 247 unsigned long desired) 248 { 249 unsigned long exp = expected; 250 251 __atomic_compare_exchange_n(p, &exp, desired, 0, 252 __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST); 253 return exp; 254 } 255 256 static unsigned long xchg(unsigned long *p, unsigned long val) 257 { 258 return __atomic_exchange_n(p, val, __ATOMIC_SEQ_CST); 259 } 260 261 static int processes; 262 263 static int mutex_lock(unsigned long *m) 264 { 265 int c; 266 int flags = FUTEX_WAIT; 267 if (!processes) 268 flags |= FUTEX_PRIVATE_FLAG; 269 270 c = cmpxchg(m, 0, 1); 271 if (!c) 272 return 0; 273 274 if (c == 1) 275 c = xchg(m, 2); 276 277 while (c) { 278 sys_futex(m, flags, 2, NULL, NULL, 0); 279 c = xchg(m, 2); 280 } 281 282 return 0; 283 } 284 285 static int mutex_unlock(unsigned long *m) 286 { 287 int flags = FUTEX_WAKE; 288 if (!processes) 289 flags |= FUTEX_PRIVATE_FLAG; 290 291 if (*m == 2) 292 *m = 0; 293 else if (xchg(m, 0) == 1) 294 return 0; 295 296 sys_futex(m, flags, 1, NULL, NULL, 0); 297 298 return 0; 299 } 300 301 static unsigned long *m1, *m2; 302 303 static void futex_setup(int cpu1, int cpu2) 304 { 305 if (!processes) { 306 static unsigned long _m1, _m2; 307 m1 = &_m1; 308 m2 = &_m2; 309 } else { 310 int shmid; 311 void *shmaddr; 312 313 shmid = shmget(IPC_PRIVATE, getpagesize(), SHM_R | SHM_W); 314 if (shmid < 0) { 315 perror("shmget"); 316 exit(1); 317 } 318 319 shmaddr = shmat(shmid, NULL, 0); 320 if (shmaddr == (char *)-1) { 321 perror("shmat"); 322 shmctl(shmid, IPC_RMID, NULL); 323 exit(1); 324 } 325 326 shmctl(shmid, IPC_RMID, NULL); 327 328 m1 = shmaddr; 329 m2 = shmaddr + sizeof(*m1); 330 } 331 332 *m1 = 0; 333 *m2 = 0; 334 335 mutex_lock(m1); 336 mutex_lock(m2); 337 } 338 339 static void *futex_thread1(void *arg) 340 { 341 signal(SIGALRM, sigalrm_handler); 342 alarm(1); 343 344 while (1) { 345 mutex_lock(m2); 346 mutex_unlock(m1); 347 348 iterations += 2; 349 } 350 351 return NULL; 352 } 353 354 static void *futex_thread2(void *arg) 355 { 356 while (1) { 357 mutex_unlock(m2); 358 mutex_lock(m1); 359 } 360 361 return NULL; 362 } 363 364 static struct actions futex_actions = { 365 .setup = futex_setup, 366 .thread1 = futex_thread1, 367 .thread2 = futex_thread2, 368 }; 369 370 static struct option options[] = { 371 { "test", required_argument, 0, 't' }, 372 { "process", no_argument, &processes, 1 }, 373 { "timeout", required_argument, 0, 's' }, 374 { "vdso", no_argument, &touch_vdso, 1 }, 375 { "no-fp", no_argument, &touch_fp, 0 }, 376 #ifdef __powerpc__ 377 { "no-altivec", no_argument, &touch_altivec, 0 }, 378 #endif 379 { "no-vector", no_argument, &touch_vector, 0 }, 380 { 0, }, 381 }; 382 383 static void usage(void) 384 { 385 fprintf(stderr, "Usage: context_switch2 <options> CPU1 CPU2\n\n"); 386 fprintf(stderr, "\t\t--test=X\tpipe, futex or yield (default)\n"); 387 fprintf(stderr, "\t\t--process\tUse processes (default threads)\n"); 388 fprintf(stderr, "\t\t--timeout=X\tDuration in seconds to run (default 30)\n"); 389 fprintf(stderr, "\t\t--vdso\t\ttouch VDSO\n"); 390 fprintf(stderr, "\t\t--no-fp\t\tDon't touch FP\n"); 391 #ifdef __powerpc__ 392 fprintf(stderr, "\t\t--no-altivec\tDon't touch altivec\n"); 393 #endif 394 fprintf(stderr, "\t\t--no-vector\tDon't touch vector\n"); 395 } 396 397 int main(int argc, char *argv[]) 398 { 399 signed char c; 400 struct actions *actions = &yield_actions; 401 int cpu1; 402 int cpu2; 403 static void (*start_fn)(void *(*fn)(void *), void *arg, unsigned long cpu); 404 405 while (1) { 406 int option_index = 0; 407 408 c = getopt_long(argc, argv, "", options, &option_index); 409 410 if (c == -1) 411 break; 412 413 switch (c) { 414 case 0: 415 if (options[option_index].flag != 0) 416 break; 417 418 usage(); 419 exit(1); 420 break; 421 422 case 't': 423 if (!strcmp(optarg, "pipe")) { 424 actions = &pipe_actions; 425 } else if (!strcmp(optarg, "yield")) { 426 actions = &yield_actions; 427 } else if (!strcmp(optarg, "futex")) { 428 actions = &futex_actions; 429 } else { 430 usage(); 431 exit(1); 432 } 433 break; 434 435 case 's': 436 timeout = atoi(optarg); 437 break; 438 439 default: 440 usage(); 441 exit(1); 442 } 443 } 444 445 if (processes) 446 start_fn = start_process_on; 447 else 448 start_fn = start_thread_on; 449 450 if (((argc - optind) != 2)) { 451 cpu1 = cpu2 = pick_online_cpu(); 452 } else { 453 cpu1 = atoi(argv[optind++]); 454 cpu2 = atoi(argv[optind++]); 455 } 456 457 printf("Using %s with ", processes ? "processes" : "threads"); 458 459 if (actions == &pipe_actions) 460 printf("pipe"); 461 else if (actions == &yield_actions) 462 printf("yield"); 463 else 464 printf("futex"); 465 466 printf(" on cpus %d/%d touching FP:%s altivec:%s vector:%s vdso:%s\n", 467 cpu1, cpu2, touch_fp ? "yes" : "no", touch_altivec ? "yes" : "no", 468 touch_vector ? "yes" : "no", touch_vdso ? "yes" : "no"); 469 470 /* Create a new process group so we can signal everyone for exit */ 471 setpgid(getpid(), getpid()); 472 473 signal(SIGUSR1, sigusr1_handler); 474 475 actions->setup(cpu1, cpu2); 476 477 start_fn(actions->thread1, NULL, cpu1); 478 start_fn(actions->thread2, NULL, cpu2); 479 480 while (1) 481 sleep(3600); 482 483 return 0; 484 } 485