1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (C) 2013 Davidlohr Bueso <davidlohr@hp.com> 4 * 5 * futex-wake: Block a bunch of threads on a futex and wake'em up, N at a time. 6 * 7 * This program is particularly useful to measure the latency of nthread wakeups 8 * in non-error situations: all waiters are queued and all wake calls wakeup 9 * one or more tasks, and thus the waitqueue is never empty. 10 */ 11 12 /* For the CLR_() macros */ 13 #include <string.h> 14 #include <pthread.h> 15 16 #include <signal.h> 17 #include "../util/stat.h" 18 #include <subcmd/parse-options.h> 19 #include <linux/compiler.h> 20 #include <linux/kernel.h> 21 #include <linux/time64.h> 22 #include <errno.h> 23 #include <internal/cpumap.h> 24 #include <perf/cpumap.h> 25 #include "bench.h" 26 #include "futex.h" 27 28 #include <err.h> 29 #include <stdlib.h> 30 #include <sys/time.h> 31 32 /* all threads will block on the same futex */ 33 static u_int32_t futex1 = 0; 34 35 /* 36 * How many wakeups to do at a time. 37 * Default to 1 in order to make the kernel work more. 38 */ 39 static unsigned int nwakes = 1; 40 41 pthread_t *worker; 42 static bool done = false, silent = false, fshared = false; 43 static pthread_mutex_t thread_lock; 44 static pthread_cond_t thread_parent, thread_worker; 45 static struct stats waketime_stats, wakeup_stats; 46 static unsigned int threads_starting, nthreads = 0; 47 static int futex_flag = 0; 48 49 static const struct option options[] = { 50 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"), 51 OPT_UINTEGER('w', "nwakes", &nwakes, "Specify amount of threads to wake at once"), 52 OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"), 53 OPT_BOOLEAN( 'S', "shared", &fshared, "Use shared futexes instead of private ones"), 54 OPT_END() 55 }; 56 57 static const char * const bench_futex_wake_usage[] = { 58 "perf bench futex wake <options>", 59 NULL 60 }; 61 62 static void *workerfn(void *arg __maybe_unused) 63 { 64 pthread_mutex_lock(&thread_lock); 65 threads_starting--; 66 if (!threads_starting) 67 pthread_cond_signal(&thread_parent); 68 pthread_cond_wait(&thread_worker, &thread_lock); 69 pthread_mutex_unlock(&thread_lock); 70 71 while (1) { 72 if (futex_wait(&futex1, 0, NULL, futex_flag) != EINTR) 73 break; 74 } 75 76 pthread_exit(NULL); 77 return NULL; 78 } 79 80 static void print_summary(void) 81 { 82 double waketime_avg = avg_stats(&waketime_stats); 83 double waketime_stddev = stddev_stats(&waketime_stats); 84 unsigned int wakeup_avg = avg_stats(&wakeup_stats); 85 86 printf("Wokeup %d of %d threads in %.4f ms (+-%.2f%%)\n", 87 wakeup_avg, 88 nthreads, 89 waketime_avg / USEC_PER_MSEC, 90 rel_stddev_stats(waketime_stddev, waketime_avg)); 91 } 92 93 static void block_threads(pthread_t *w, 94 pthread_attr_t thread_attr, struct perf_cpu_map *cpu) 95 { 96 cpu_set_t cpuset; 97 unsigned int i; 98 99 threads_starting = nthreads; 100 101 /* create and block all threads */ 102 for (i = 0; i < nthreads; i++) { 103 CPU_ZERO(&cpuset); 104 CPU_SET(cpu->map[i % cpu->nr], &cpuset); 105 106 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset)) 107 err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); 108 109 if (pthread_create(&w[i], &thread_attr, workerfn, NULL)) 110 err(EXIT_FAILURE, "pthread_create"); 111 } 112 } 113 114 static void toggle_done(int sig __maybe_unused, 115 siginfo_t *info __maybe_unused, 116 void *uc __maybe_unused) 117 { 118 done = true; 119 } 120 121 int bench_futex_wake(int argc, const char **argv) 122 { 123 int ret = 0; 124 unsigned int i, j; 125 struct sigaction act; 126 pthread_attr_t thread_attr; 127 struct perf_cpu_map *cpu; 128 129 argc = parse_options(argc, argv, options, bench_futex_wake_usage, 0); 130 if (argc) { 131 usage_with_options(bench_futex_wake_usage, options); 132 exit(EXIT_FAILURE); 133 } 134 135 cpu = perf_cpu_map__new(NULL); 136 if (!cpu) 137 err(EXIT_FAILURE, "calloc"); 138 139 memset(&act, 0, sizeof(act)); 140 sigfillset(&act.sa_mask); 141 act.sa_sigaction = toggle_done; 142 sigaction(SIGINT, &act, NULL); 143 144 if (!nthreads) 145 nthreads = cpu->nr; 146 147 worker = calloc(nthreads, sizeof(*worker)); 148 if (!worker) 149 err(EXIT_FAILURE, "calloc"); 150 151 if (!fshared) 152 futex_flag = FUTEX_PRIVATE_FLAG; 153 154 printf("Run summary [PID %d]: blocking on %d threads (at [%s] futex %p), " 155 "waking up %d at a time.\n\n", 156 getpid(), nthreads, fshared ? "shared":"private", &futex1, nwakes); 157 158 init_stats(&wakeup_stats); 159 init_stats(&waketime_stats); 160 pthread_attr_init(&thread_attr); 161 pthread_mutex_init(&thread_lock, NULL); 162 pthread_cond_init(&thread_parent, NULL); 163 pthread_cond_init(&thread_worker, NULL); 164 165 for (j = 0; j < bench_repeat && !done; j++) { 166 unsigned int nwoken = 0; 167 struct timeval start, end, runtime; 168 169 /* create, launch & block all threads */ 170 block_threads(worker, thread_attr, cpu); 171 172 /* make sure all threads are already blocked */ 173 pthread_mutex_lock(&thread_lock); 174 while (threads_starting) 175 pthread_cond_wait(&thread_parent, &thread_lock); 176 pthread_cond_broadcast(&thread_worker); 177 pthread_mutex_unlock(&thread_lock); 178 179 usleep(100000); 180 181 /* Ok, all threads are patiently blocked, start waking folks up */ 182 gettimeofday(&start, NULL); 183 while (nwoken != nthreads) 184 nwoken += futex_wake(&futex1, nwakes, futex_flag); 185 gettimeofday(&end, NULL); 186 timersub(&end, &start, &runtime); 187 188 update_stats(&wakeup_stats, nwoken); 189 update_stats(&waketime_stats, runtime.tv_usec); 190 191 if (!silent) { 192 printf("[Run %d]: Wokeup %d of %d threads in %.4f ms\n", 193 j + 1, nwoken, nthreads, runtime.tv_usec / (double)USEC_PER_MSEC); 194 } 195 196 for (i = 0; i < nthreads; i++) { 197 ret = pthread_join(worker[i], NULL); 198 if (ret) 199 err(EXIT_FAILURE, "pthread_join"); 200 } 201 202 } 203 204 /* cleanup & report results */ 205 pthread_cond_destroy(&thread_parent); 206 pthread_cond_destroy(&thread_worker); 207 pthread_mutex_destroy(&thread_lock); 208 pthread_attr_destroy(&thread_attr); 209 210 print_summary(); 211 212 free(worker); 213 return ret; 214 } 215