1 /* 2 * Copyright (C) 2013 Davidlohr Bueso <davidlohr@hp.com> 3 * 4 * futex-wake: Block a bunch of threads on a futex and wake'em up, N at a time. 5 * 6 * This program is particularly useful to measure the latency of nthread wakeups 7 * in non-error situations: all waiters are queued and all wake calls wakeup 8 * one or more tasks, and thus the waitqueue is never empty. 9 */ 10 11 #include "../perf.h" 12 #include "../util/util.h" 13 #include "../util/stat.h" 14 #include <subcmd/parse-options.h> 15 #include "../util/header.h" 16 #include "bench.h" 17 #include "futex.h" 18 19 #include <err.h> 20 #include <stdlib.h> 21 #include <sys/time.h> 22 #include <pthread.h> 23 24 /* all threads will block on the same futex */ 25 static u_int32_t futex1 = 0; 26 27 /* 28 * How many wakeups to do at a time. 29 * Default to 1 in order to make the kernel work more. 30 */ 31 static unsigned int nwakes = 1; 32 33 pthread_t *worker; 34 static bool done = false, silent = false, fshared = false; 35 static pthread_mutex_t thread_lock; 36 static pthread_cond_t thread_parent, thread_worker; 37 static struct stats waketime_stats, wakeup_stats; 38 static unsigned int ncpus, threads_starting, nthreads = 0; 39 static int futex_flag = 0; 40 41 static const struct option options[] = { 42 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"), 43 OPT_UINTEGER('w', "nwakes", &nwakes, "Specify amount of threads to wake at once"), 44 OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"), 45 OPT_BOOLEAN( 'S', "shared", &fshared, "Use shared futexes instead of private ones"), 46 OPT_END() 47 }; 48 49 static const char * const bench_futex_wake_usage[] = { 50 "perf bench futex wake <options>", 51 NULL 52 }; 53 54 static void *workerfn(void *arg __maybe_unused) 55 { 56 pthread_mutex_lock(&thread_lock); 57 threads_starting--; 58 if (!threads_starting) 59 pthread_cond_signal(&thread_parent); 60 pthread_cond_wait(&thread_worker, &thread_lock); 61 pthread_mutex_unlock(&thread_lock); 62 63 while (1) { 64 if (futex_wait(&futex1, 0, NULL, futex_flag) != EINTR) 65 break; 66 } 67 68 pthread_exit(NULL); 69 return NULL; 70 } 71 72 static void print_summary(void) 73 { 74 double waketime_avg = avg_stats(&waketime_stats); 75 double waketime_stddev = stddev_stats(&waketime_stats); 76 unsigned int wakeup_avg = avg_stats(&wakeup_stats); 77 78 printf("Wokeup %d of %d threads in %.4f ms (+-%.2f%%)\n", 79 wakeup_avg, 80 nthreads, 81 waketime_avg/1e3, 82 rel_stddev_stats(waketime_stddev, waketime_avg)); 83 } 84 85 static void block_threads(pthread_t *w, 86 pthread_attr_t thread_attr) 87 { 88 cpu_set_t cpu; 89 unsigned int i; 90 91 threads_starting = nthreads; 92 93 /* create and block all threads */ 94 for (i = 0; i < nthreads; i++) { 95 CPU_ZERO(&cpu); 96 CPU_SET(i % ncpus, &cpu); 97 98 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu)) 99 err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); 100 101 if (pthread_create(&w[i], &thread_attr, workerfn, NULL)) 102 err(EXIT_FAILURE, "pthread_create"); 103 } 104 } 105 106 static void toggle_done(int sig __maybe_unused, 107 siginfo_t *info __maybe_unused, 108 void *uc __maybe_unused) 109 { 110 done = true; 111 } 112 113 int bench_futex_wake(int argc, const char **argv, 114 const char *prefix __maybe_unused) 115 { 116 int ret = 0; 117 unsigned int i, j; 118 struct sigaction act; 119 pthread_attr_t thread_attr; 120 121 argc = parse_options(argc, argv, options, bench_futex_wake_usage, 0); 122 if (argc) { 123 usage_with_options(bench_futex_wake_usage, options); 124 exit(EXIT_FAILURE); 125 } 126 127 ncpus = sysconf(_SC_NPROCESSORS_ONLN); 128 129 sigfillset(&act.sa_mask); 130 act.sa_sigaction = toggle_done; 131 sigaction(SIGINT, &act, NULL); 132 133 if (!nthreads) 134 nthreads = ncpus; 135 136 worker = calloc(nthreads, sizeof(*worker)); 137 if (!worker) 138 err(EXIT_FAILURE, "calloc"); 139 140 if (!fshared) 141 futex_flag = FUTEX_PRIVATE_FLAG; 142 143 printf("Run summary [PID %d]: blocking on %d threads (at [%s] futex %p), " 144 "waking up %d at a time.\n\n", 145 getpid(), nthreads, fshared ? "shared":"private", &futex1, nwakes); 146 147 init_stats(&wakeup_stats); 148 init_stats(&waketime_stats); 149 pthread_attr_init(&thread_attr); 150 pthread_mutex_init(&thread_lock, NULL); 151 pthread_cond_init(&thread_parent, NULL); 152 pthread_cond_init(&thread_worker, NULL); 153 154 for (j = 0; j < bench_repeat && !done; j++) { 155 unsigned int nwoken = 0; 156 struct timeval start, end, runtime; 157 158 /* create, launch & block all threads */ 159 block_threads(worker, thread_attr); 160 161 /* make sure all threads are already blocked */ 162 pthread_mutex_lock(&thread_lock); 163 while (threads_starting) 164 pthread_cond_wait(&thread_parent, &thread_lock); 165 pthread_cond_broadcast(&thread_worker); 166 pthread_mutex_unlock(&thread_lock); 167 168 usleep(100000); 169 170 /* Ok, all threads are patiently blocked, start waking folks up */ 171 gettimeofday(&start, NULL); 172 while (nwoken != nthreads) 173 nwoken += futex_wake(&futex1, nwakes, futex_flag); 174 gettimeofday(&end, NULL); 175 timersub(&end, &start, &runtime); 176 177 update_stats(&wakeup_stats, nwoken); 178 update_stats(&waketime_stats, runtime.tv_usec); 179 180 if (!silent) { 181 printf("[Run %d]: Wokeup %d of %d threads in %.4f ms\n", 182 j + 1, nwoken, nthreads, runtime.tv_usec/1e3); 183 } 184 185 for (i = 0; i < nthreads; i++) { 186 ret = pthread_join(worker[i], NULL); 187 if (ret) 188 err(EXIT_FAILURE, "pthread_join"); 189 } 190 191 } 192 193 /* cleanup & report results */ 194 pthread_cond_destroy(&thread_parent); 195 pthread_cond_destroy(&thread_worker); 196 pthread_mutex_destroy(&thread_lock); 197 pthread_attr_destroy(&thread_attr); 198 199 print_summary(); 200 201 free(worker); 202 return ret; 203 } 204