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 .nbuckets = -1, 46 /* 47 * How many wakeups to do at a time. 48 * Default to 1 in order to make the kernel work more. 49 */ 50 .nwakes = 1, 51 }; 52 53 static const struct option options[] = { 54 OPT_INTEGER( 'b', "buckets", ¶ms.nbuckets, "Specify amount of hash buckets"), 55 OPT_UINTEGER('t', "threads", ¶ms.nthreads, "Specify amount of threads"), 56 OPT_UINTEGER('w', "nwakes", ¶ms.nwakes, "Specify amount of threads to wake at once"), 57 OPT_BOOLEAN( 's', "silent", ¶ms.silent, "Silent mode: do not display data/details"), 58 OPT_BOOLEAN( 'S', "shared", ¶ms.fshared, "Use shared futexes instead of private ones"), 59 OPT_BOOLEAN( 'm', "mlockall", ¶ms.mlockall, "Lock all current and future memory"), 60 61 OPT_END() 62 }; 63 64 static const char * const bench_futex_wake_usage[] = { 65 "perf bench futex wake <options>", 66 NULL 67 }; 68 69 static void *workerfn(void *arg __maybe_unused) 70 { 71 mutex_lock(&thread_lock); 72 threads_starting--; 73 if (!threads_starting) 74 cond_signal(&thread_parent); 75 cond_wait(&thread_worker, &thread_lock); 76 mutex_unlock(&thread_lock); 77 78 while (1) { 79 if (futex_wait(&futex1, 0, NULL, futex_flag) != EINTR) 80 break; 81 } 82 83 pthread_exit(NULL); 84 return NULL; 85 } 86 87 static void print_summary(void) 88 { 89 double waketime_avg = avg_stats(&waketime_stats); 90 double waketime_stddev = stddev_stats(&waketime_stats); 91 unsigned int wakeup_avg = avg_stats(&wakeup_stats); 92 93 printf("Wokeup %d of %d threads in %.4f ms (+-%.2f%%)\n", 94 wakeup_avg, 95 params.nthreads, 96 waketime_avg / USEC_PER_MSEC, 97 rel_stddev_stats(waketime_stddev, waketime_avg)); 98 futex_print_nbuckets(¶ms); 99 } 100 101 static void block_threads(pthread_t *w, struct perf_cpu_map *cpu) 102 { 103 cpu_set_t *cpuset; 104 unsigned int i; 105 size_t size; 106 int nrcpus = cpu__max_cpu().cpu; 107 threads_starting = params.nthreads; 108 109 cpuset = CPU_ALLOC(nrcpus); 110 BUG_ON(!cpuset); 111 size = CPU_ALLOC_SIZE(nrcpus); 112 113 /* create and block all threads */ 114 for (i = 0; i < params.nthreads; i++) { 115 pthread_attr_t thread_attr; 116 117 pthread_attr_init(&thread_attr); 118 CPU_ZERO_S(size, cpuset); 119 CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset); 120 121 if (pthread_attr_setaffinity_np(&thread_attr, size, cpuset)) { 122 CPU_FREE(cpuset); 123 err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); 124 } 125 126 if (pthread_create(&w[i], &thread_attr, workerfn, NULL)) { 127 CPU_FREE(cpuset); 128 err(EXIT_FAILURE, "pthread_create"); 129 } 130 pthread_attr_destroy(&thread_attr); 131 } 132 CPU_FREE(cpuset); 133 } 134 135 static void toggle_done(int sig __maybe_unused, 136 siginfo_t *info __maybe_unused, 137 void *uc __maybe_unused) 138 { 139 done = true; 140 } 141 142 int bench_futex_wake(int argc, const char **argv) 143 { 144 int ret = 0; 145 unsigned int i, j; 146 struct sigaction act; 147 struct perf_cpu_map *cpu; 148 149 argc = parse_options(argc, argv, options, bench_futex_wake_usage, 0); 150 if (argc) { 151 usage_with_options(bench_futex_wake_usage, options); 152 exit(EXIT_FAILURE); 153 } 154 155 cpu = perf_cpu_map__new_online_cpus(); 156 if (!cpu) 157 err(EXIT_FAILURE, "calloc"); 158 159 memset(&act, 0, sizeof(act)); 160 sigfillset(&act.sa_mask); 161 act.sa_sigaction = toggle_done; 162 sigaction(SIGINT, &act, NULL); 163 164 if (params.mlockall) { 165 if (mlockall(MCL_CURRENT | MCL_FUTURE)) 166 err(EXIT_FAILURE, "mlockall"); 167 } 168 169 if (!params.nthreads) 170 params.nthreads = perf_cpu_map__nr(cpu); 171 172 worker = calloc(params.nthreads, sizeof(*worker)); 173 if (!worker) 174 err(EXIT_FAILURE, "calloc"); 175 176 if (!params.fshared) 177 futex_flag = FUTEX_PRIVATE_FLAG; 178 179 printf("Run summary [PID %d]: blocking on %d threads (at [%s] futex %p), " 180 "waking up %d at a time.\n\n", 181 getpid(), params.nthreads, params.fshared ? "shared":"private", 182 &futex1, params.nwakes); 183 184 init_stats(&wakeup_stats); 185 init_stats(&waketime_stats); 186 mutex_init(&thread_lock); 187 cond_init(&thread_parent); 188 cond_init(&thread_worker); 189 190 for (j = 0; j < bench_repeat && !done; j++) { 191 unsigned int nwoken = 0; 192 struct timeval start, end, runtime; 193 194 /* create, launch & block all threads */ 195 block_threads(worker, cpu); 196 197 /* make sure all threads are already blocked */ 198 mutex_lock(&thread_lock); 199 while (threads_starting) 200 cond_wait(&thread_parent, &thread_lock); 201 cond_broadcast(&thread_worker); 202 mutex_unlock(&thread_lock); 203 204 usleep(100000); 205 206 /* Ok, all threads are patiently blocked, start waking folks up */ 207 gettimeofday(&start, NULL); 208 while (nwoken != params.nthreads) 209 nwoken += futex_wake(&futex1, 210 params.nwakes, futex_flag); 211 gettimeofday(&end, NULL); 212 timersub(&end, &start, &runtime); 213 214 update_stats(&wakeup_stats, nwoken); 215 update_stats(&waketime_stats, runtime.tv_usec); 216 217 if (!params.silent) { 218 printf("[Run %d]: Wokeup %d of %d threads in %.4f ms\n", 219 j + 1, nwoken, params.nthreads, 220 runtime.tv_usec / (double)USEC_PER_MSEC); 221 } 222 223 for (i = 0; i < params.nthreads; i++) { 224 ret = pthread_join(worker[i], NULL); 225 if (ret) 226 err(EXIT_FAILURE, "pthread_join"); 227 } 228 229 } 230 231 /* cleanup & report results */ 232 cond_destroy(&thread_parent); 233 cond_destroy(&thread_worker); 234 mutex_destroy(&thread_lock); 235 236 print_summary(); 237 238 free(worker); 239 perf_cpu_map__put(cpu); 240 return ret; 241 } 242