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/mutex.h" 18 #include "../util/stat.h" 19 #include <subcmd/parse-options.h> 20 #include <linux/compiler.h> 21 #include <linux/kernel.h> 22 #include <linux/time64.h> 23 #include <errno.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 #include <sys/mman.h> 32 33 /* all threads will block on the same futex */ 34 static u_int32_t futex1 = 0; 35 36 static pthread_t *worker; 37 static bool done = false; 38 static struct mutex thread_lock; 39 static struct cond thread_parent, thread_worker; 40 static struct stats waketime_stats, wakeup_stats; 41 static unsigned int threads_starting; 42 static int futex_flag = 0; 43 44 static struct bench_futex_parameters params = { 45 /* 46 * How many wakeups to do at a time. 47 * Default to 1 in order to make the kernel work more. 48 */ 49 .nwakes = 1, 50 }; 51 52 static const struct option options[] = { 53 OPT_UINTEGER('t', "threads", ¶ms.nthreads, "Specify amount of threads"), 54 OPT_UINTEGER('w', "nwakes", ¶ms.nwakes, "Specify amount of threads to wake at once"), 55 OPT_BOOLEAN( 's', "silent", ¶ms.silent, "Silent mode: do not display data/details"), 56 OPT_BOOLEAN( 'S', "shared", ¶ms.fshared, "Use shared futexes instead of private ones"), 57 OPT_BOOLEAN( 'm', "mlockall", ¶ms.mlockall, "Lock all current and future memory"), 58 59 OPT_END() 60 }; 61 62 static const char * const bench_futex_wake_usage[] = { 63 "perf bench futex wake <options>", 64 NULL 65 }; 66 67 static void *workerfn(void *arg __maybe_unused) 68 { 69 mutex_lock(&thread_lock); 70 threads_starting--; 71 if (!threads_starting) 72 cond_signal(&thread_parent); 73 cond_wait(&thread_worker, &thread_lock); 74 mutex_unlock(&thread_lock); 75 76 while (1) { 77 if (futex_wait(&futex1, 0, NULL, futex_flag) != EINTR) 78 break; 79 } 80 81 pthread_exit(NULL); 82 return NULL; 83 } 84 85 static void print_summary(void) 86 { 87 double waketime_avg = avg_stats(&waketime_stats); 88 double waketime_stddev = stddev_stats(&waketime_stats); 89 unsigned int wakeup_avg = avg_stats(&wakeup_stats); 90 91 printf("Wokeup %d of %d threads in %.4f ms (+-%.2f%%)\n", 92 wakeup_avg, 93 params.nthreads, 94 waketime_avg / USEC_PER_MSEC, 95 rel_stddev_stats(waketime_stddev, waketime_avg)); 96 } 97 98 static void block_threads(pthread_t *w, 99 pthread_attr_t thread_attr, struct perf_cpu_map *cpu) 100 { 101 cpu_set_t *cpuset; 102 unsigned int i; 103 size_t size; 104 int nrcpus = perf_cpu_map__nr(cpu); 105 threads_starting = params.nthreads; 106 107 cpuset = CPU_ALLOC(nrcpus); 108 BUG_ON(!cpuset); 109 size = CPU_ALLOC_SIZE(nrcpus); 110 111 /* create and block all threads */ 112 for (i = 0; i < params.nthreads; i++) { 113 CPU_ZERO_S(size, cpuset); 114 CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset); 115 116 if (pthread_attr_setaffinity_np(&thread_attr, size, cpuset)) { 117 CPU_FREE(cpuset); 118 err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); 119 } 120 121 if (pthread_create(&w[i], &thread_attr, workerfn, NULL)) { 122 CPU_FREE(cpuset); 123 err(EXIT_FAILURE, "pthread_create"); 124 } 125 } 126 CPU_FREE(cpuset); 127 } 128 129 static void toggle_done(int sig __maybe_unused, 130 siginfo_t *info __maybe_unused, 131 void *uc __maybe_unused) 132 { 133 done = true; 134 } 135 136 int bench_futex_wake(int argc, const char **argv) 137 { 138 int ret = 0; 139 unsigned int i, j; 140 struct sigaction act; 141 pthread_attr_t thread_attr; 142 struct perf_cpu_map *cpu; 143 144 argc = parse_options(argc, argv, options, bench_futex_wake_usage, 0); 145 if (argc) { 146 usage_with_options(bench_futex_wake_usage, options); 147 exit(EXIT_FAILURE); 148 } 149 150 cpu = perf_cpu_map__new(NULL); 151 if (!cpu) 152 err(EXIT_FAILURE, "calloc"); 153 154 memset(&act, 0, sizeof(act)); 155 sigfillset(&act.sa_mask); 156 act.sa_sigaction = toggle_done; 157 sigaction(SIGINT, &act, NULL); 158 159 if (params.mlockall) { 160 if (mlockall(MCL_CURRENT | MCL_FUTURE)) 161 err(EXIT_FAILURE, "mlockall"); 162 } 163 164 if (!params.nthreads) 165 params.nthreads = perf_cpu_map__nr(cpu); 166 167 worker = calloc(params.nthreads, sizeof(*worker)); 168 if (!worker) 169 err(EXIT_FAILURE, "calloc"); 170 171 if (!params.fshared) 172 futex_flag = FUTEX_PRIVATE_FLAG; 173 174 printf("Run summary [PID %d]: blocking on %d threads (at [%s] futex %p), " 175 "waking up %d at a time.\n\n", 176 getpid(), params.nthreads, params.fshared ? "shared":"private", 177 &futex1, params.nwakes); 178 179 init_stats(&wakeup_stats); 180 init_stats(&waketime_stats); 181 pthread_attr_init(&thread_attr); 182 mutex_init(&thread_lock); 183 cond_init(&thread_parent); 184 cond_init(&thread_worker); 185 186 for (j = 0; j < bench_repeat && !done; j++) { 187 unsigned int nwoken = 0; 188 struct timeval start, end, runtime; 189 190 /* create, launch & block all threads */ 191 block_threads(worker, thread_attr, cpu); 192 193 /* make sure all threads are already blocked */ 194 mutex_lock(&thread_lock); 195 while (threads_starting) 196 cond_wait(&thread_parent, &thread_lock); 197 cond_broadcast(&thread_worker); 198 mutex_unlock(&thread_lock); 199 200 usleep(100000); 201 202 /* Ok, all threads are patiently blocked, start waking folks up */ 203 gettimeofday(&start, NULL); 204 while (nwoken != params.nthreads) 205 nwoken += futex_wake(&futex1, 206 params.nwakes, futex_flag); 207 gettimeofday(&end, NULL); 208 timersub(&end, &start, &runtime); 209 210 update_stats(&wakeup_stats, nwoken); 211 update_stats(&waketime_stats, runtime.tv_usec); 212 213 if (!params.silent) { 214 printf("[Run %d]: Wokeup %d of %d threads in %.4f ms\n", 215 j + 1, nwoken, params.nthreads, 216 runtime.tv_usec / (double)USEC_PER_MSEC); 217 } 218 219 for (i = 0; i < params.nthreads; i++) { 220 ret = pthread_join(worker[i], NULL); 221 if (ret) 222 err(EXIT_FAILURE, "pthread_join"); 223 } 224 225 } 226 227 /* cleanup & report results */ 228 cond_destroy(&thread_parent); 229 cond_destroy(&thread_worker); 230 mutex_destroy(&thread_lock); 231 pthread_attr_destroy(&thread_attr); 232 233 print_summary(); 234 235 free(worker); 236 perf_cpu_map__put(cpu); 237 return ret; 238 } 239