1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (C) 2015 Davidlohr Bueso. 4 * 5 * Block a bunch of threads and let parallel waker threads wakeup an 6 * equal amount of them. The program output reflects the avg latency 7 * for each individual thread to service its share of work. Ultimately 8 * it can be used to measure futex_wake() changes. 9 */ 10 #include "bench.h" 11 #include <linux/compiler.h> 12 #include "../util/debug.h" 13 14 #ifndef HAVE_PTHREAD_BARRIER 15 int bench_futex_wake_parallel(int argc __maybe_unused, const char **argv __maybe_unused) 16 { 17 pr_err("%s: pthread_barrier_t unavailable, disabling this test...\n", __func__); 18 return 0; 19 } 20 #else /* HAVE_PTHREAD_BARRIER */ 21 /* For the CLR_() macros */ 22 #include <string.h> 23 #include <pthread.h> 24 25 #include <signal.h> 26 #include "../util/stat.h" 27 #include <subcmd/parse-options.h> 28 #include <linux/kernel.h> 29 #include <linux/time64.h> 30 #include <errno.h> 31 #include "futex.h" 32 #include <internal/cpumap.h> 33 #include <perf/cpumap.h> 34 35 #include <err.h> 36 #include <stdlib.h> 37 #include <sys/time.h> 38 39 struct thread_data { 40 pthread_t worker; 41 unsigned int nwoken; 42 struct timeval runtime; 43 }; 44 45 static unsigned int nwakes = 1; 46 47 /* all threads will block on the same futex -- hash bucket chaos ;) */ 48 static u_int32_t futex = 0; 49 50 static pthread_t *blocked_worker; 51 static bool done = false, silent = false, fshared = false; 52 static unsigned int nblocked_threads = 0, nwaking_threads = 0; 53 static pthread_mutex_t thread_lock; 54 static pthread_cond_t thread_parent, thread_worker; 55 static pthread_barrier_t barrier; 56 static struct stats waketime_stats, wakeup_stats; 57 static unsigned int threads_starting; 58 static int futex_flag = 0; 59 60 static const struct option options[] = { 61 OPT_UINTEGER('t', "threads", &nblocked_threads, "Specify amount of threads"), 62 OPT_UINTEGER('w', "nwakers", &nwaking_threads, "Specify amount of waking threads"), 63 OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"), 64 OPT_BOOLEAN( 'S', "shared", &fshared, "Use shared futexes instead of private ones"), 65 OPT_END() 66 }; 67 68 static const char * const bench_futex_wake_parallel_usage[] = { 69 "perf bench futex wake-parallel <options>", 70 NULL 71 }; 72 73 static void *waking_workerfn(void *arg) 74 { 75 struct thread_data *waker = (struct thread_data *) arg; 76 struct timeval start, end; 77 78 pthread_barrier_wait(&barrier); 79 80 gettimeofday(&start, NULL); 81 82 waker->nwoken = futex_wake(&futex, nwakes, futex_flag); 83 if (waker->nwoken != nwakes) 84 warnx("couldn't wakeup all tasks (%d/%d)", 85 waker->nwoken, nwakes); 86 87 gettimeofday(&end, NULL); 88 timersub(&end, &start, &waker->runtime); 89 90 pthread_exit(NULL); 91 return NULL; 92 } 93 94 static void wakeup_threads(struct thread_data *td, pthread_attr_t thread_attr) 95 { 96 unsigned int i; 97 98 pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE); 99 100 pthread_barrier_init(&barrier, NULL, nwaking_threads + 1); 101 102 /* create and block all threads */ 103 for (i = 0; i < nwaking_threads; i++) { 104 /* 105 * Thread creation order will impact per-thread latency 106 * as it will affect the order to acquire the hb spinlock. 107 * For now let the scheduler decide. 108 */ 109 if (pthread_create(&td[i].worker, &thread_attr, 110 waking_workerfn, (void *)&td[i])) 111 err(EXIT_FAILURE, "pthread_create"); 112 } 113 114 pthread_barrier_wait(&barrier); 115 116 for (i = 0; i < nwaking_threads; i++) 117 if (pthread_join(td[i].worker, NULL)) 118 err(EXIT_FAILURE, "pthread_join"); 119 120 pthread_barrier_destroy(&barrier); 121 } 122 123 static void *blocked_workerfn(void *arg __maybe_unused) 124 { 125 pthread_mutex_lock(&thread_lock); 126 threads_starting--; 127 if (!threads_starting) 128 pthread_cond_signal(&thread_parent); 129 pthread_cond_wait(&thread_worker, &thread_lock); 130 pthread_mutex_unlock(&thread_lock); 131 132 while (1) { /* handle spurious wakeups */ 133 if (futex_wait(&futex, 0, NULL, futex_flag) != EINTR) 134 break; 135 } 136 137 pthread_exit(NULL); 138 return NULL; 139 } 140 141 static void block_threads(pthread_t *w, pthread_attr_t thread_attr, 142 struct perf_cpu_map *cpu) 143 { 144 cpu_set_t cpuset; 145 unsigned int i; 146 147 threads_starting = nblocked_threads; 148 149 /* create and block all threads */ 150 for (i = 0; i < nblocked_threads; i++) { 151 CPU_ZERO(&cpuset); 152 CPU_SET(cpu->map[i % cpu->nr], &cpuset); 153 154 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset)) 155 err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); 156 157 if (pthread_create(&w[i], &thread_attr, blocked_workerfn, NULL)) 158 err(EXIT_FAILURE, "pthread_create"); 159 } 160 } 161 162 static void print_run(struct thread_data *waking_worker, unsigned int run_num) 163 { 164 unsigned int i, wakeup_avg; 165 double waketime_avg, waketime_stddev; 166 struct stats __waketime_stats, __wakeup_stats; 167 168 init_stats(&__wakeup_stats); 169 init_stats(&__waketime_stats); 170 171 for (i = 0; i < nwaking_threads; i++) { 172 update_stats(&__waketime_stats, waking_worker[i].runtime.tv_usec); 173 update_stats(&__wakeup_stats, waking_worker[i].nwoken); 174 } 175 176 waketime_avg = avg_stats(&__waketime_stats); 177 waketime_stddev = stddev_stats(&__waketime_stats); 178 wakeup_avg = avg_stats(&__wakeup_stats); 179 180 printf("[Run %d]: Avg per-thread latency (waking %d/%d threads) " 181 "in %.4f ms (+-%.2f%%)\n", run_num + 1, wakeup_avg, 182 nblocked_threads, waketime_avg / USEC_PER_MSEC, 183 rel_stddev_stats(waketime_stddev, waketime_avg)); 184 } 185 186 static void print_summary(void) 187 { 188 unsigned int wakeup_avg; 189 double waketime_avg, waketime_stddev; 190 191 waketime_avg = avg_stats(&waketime_stats); 192 waketime_stddev = stddev_stats(&waketime_stats); 193 wakeup_avg = avg_stats(&wakeup_stats); 194 195 printf("Avg per-thread latency (waking %d/%d threads) in %.4f ms (+-%.2f%%)\n", 196 wakeup_avg, 197 nblocked_threads, 198 waketime_avg / USEC_PER_MSEC, 199 rel_stddev_stats(waketime_stddev, waketime_avg)); 200 } 201 202 203 static void do_run_stats(struct thread_data *waking_worker) 204 { 205 unsigned int i; 206 207 for (i = 0; i < nwaking_threads; i++) { 208 update_stats(&waketime_stats, waking_worker[i].runtime.tv_usec); 209 update_stats(&wakeup_stats, waking_worker[i].nwoken); 210 } 211 212 } 213 214 static void toggle_done(int sig __maybe_unused, 215 siginfo_t *info __maybe_unused, 216 void *uc __maybe_unused) 217 { 218 done = true; 219 } 220 221 int bench_futex_wake_parallel(int argc, const char **argv) 222 { 223 int ret = 0; 224 unsigned int i, j; 225 struct sigaction act; 226 pthread_attr_t thread_attr; 227 struct thread_data *waking_worker; 228 struct perf_cpu_map *cpu; 229 230 argc = parse_options(argc, argv, options, 231 bench_futex_wake_parallel_usage, 0); 232 if (argc) { 233 usage_with_options(bench_futex_wake_parallel_usage, options); 234 exit(EXIT_FAILURE); 235 } 236 237 sigfillset(&act.sa_mask); 238 act.sa_sigaction = toggle_done; 239 sigaction(SIGINT, &act, NULL); 240 241 cpu = perf_cpu_map__new(NULL); 242 if (!cpu) 243 err(EXIT_FAILURE, "calloc"); 244 245 if (!nblocked_threads) 246 nblocked_threads = cpu->nr; 247 248 /* some sanity checks */ 249 if (nwaking_threads > nblocked_threads || !nwaking_threads) 250 nwaking_threads = nblocked_threads; 251 252 if (nblocked_threads % nwaking_threads) 253 errx(EXIT_FAILURE, "Must be perfectly divisible"); 254 /* 255 * Each thread will wakeup nwakes tasks in 256 * a single futex_wait call. 257 */ 258 nwakes = nblocked_threads/nwaking_threads; 259 260 blocked_worker = calloc(nblocked_threads, sizeof(*blocked_worker)); 261 if (!blocked_worker) 262 err(EXIT_FAILURE, "calloc"); 263 264 if (!fshared) 265 futex_flag = FUTEX_PRIVATE_FLAG; 266 267 printf("Run summary [PID %d]: blocking on %d threads (at [%s] " 268 "futex %p), %d threads waking up %d at a time.\n\n", 269 getpid(), nblocked_threads, fshared ? "shared":"private", 270 &futex, nwaking_threads, nwakes); 271 272 init_stats(&wakeup_stats); 273 init_stats(&waketime_stats); 274 275 pthread_attr_init(&thread_attr); 276 pthread_mutex_init(&thread_lock, NULL); 277 pthread_cond_init(&thread_parent, NULL); 278 pthread_cond_init(&thread_worker, NULL); 279 280 for (j = 0; j < bench_repeat && !done; j++) { 281 waking_worker = calloc(nwaking_threads, sizeof(*waking_worker)); 282 if (!waking_worker) 283 err(EXIT_FAILURE, "calloc"); 284 285 /* create, launch & block all threads */ 286 block_threads(blocked_worker, thread_attr, cpu); 287 288 /* make sure all threads are already blocked */ 289 pthread_mutex_lock(&thread_lock); 290 while (threads_starting) 291 pthread_cond_wait(&thread_parent, &thread_lock); 292 pthread_cond_broadcast(&thread_worker); 293 pthread_mutex_unlock(&thread_lock); 294 295 usleep(100000); 296 297 /* Ok, all threads are patiently blocked, start waking folks up */ 298 wakeup_threads(waking_worker, thread_attr); 299 300 for (i = 0; i < nblocked_threads; i++) { 301 ret = pthread_join(blocked_worker[i], NULL); 302 if (ret) 303 err(EXIT_FAILURE, "pthread_join"); 304 } 305 306 do_run_stats(waking_worker); 307 if (!silent) 308 print_run(waking_worker, j); 309 310 free(waking_worker); 311 } 312 313 /* cleanup & report results */ 314 pthread_cond_destroy(&thread_parent); 315 pthread_cond_destroy(&thread_worker); 316 pthread_mutex_destroy(&thread_lock); 317 pthread_attr_destroy(&thread_attr); 318 319 print_summary(); 320 321 free(blocked_worker); 322 return ret; 323 } 324 #endif /* HAVE_PTHREAD_BARRIER */ 325