1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (C) 2013 Davidlohr Bueso <davidlohr@hp.com> 4 * 5 * futex-requeue: Block a bunch of threads on futex1 and requeue them 6 * on futex2, N at a time. 7 * 8 * This program is particularly useful to measure the latency of nthread 9 * requeues without waking up any tasks -- thus mimicking a regular futex_wait. 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 "bench.h" 24 #include "futex.h" 25 #include "cpumap.h" 26 27 #include <err.h> 28 #include <stdlib.h> 29 #include <sys/time.h> 30 31 static u_int32_t futex1 = 0, futex2 = 0; 32 33 /* 34 * How many tasks to requeue at a time. 35 * Default to 1 in order to make the kernel work more. 36 */ 37 static unsigned int nrequeue = 1; 38 39 static pthread_t *worker; 40 static bool done = false, silent = false, fshared = false; 41 static pthread_mutex_t thread_lock; 42 static pthread_cond_t thread_parent, thread_worker; 43 static struct stats requeuetime_stats, requeued_stats; 44 static unsigned int threads_starting, nthreads = 0; 45 static int futex_flag = 0; 46 47 static const struct option options[] = { 48 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"), 49 OPT_UINTEGER('q', "nrequeue", &nrequeue, "Specify amount of threads to requeue at once"), 50 OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"), 51 OPT_BOOLEAN( 'S', "shared", &fshared, "Use shared futexes instead of private ones"), 52 OPT_END() 53 }; 54 55 static const char * const bench_futex_requeue_usage[] = { 56 "perf bench futex requeue <options>", 57 NULL 58 }; 59 60 static void print_summary(void) 61 { 62 double requeuetime_avg = avg_stats(&requeuetime_stats); 63 double requeuetime_stddev = stddev_stats(&requeuetime_stats); 64 unsigned int requeued_avg = avg_stats(&requeued_stats); 65 66 printf("Requeued %d of %d threads in %.4f ms (+-%.2f%%)\n", 67 requeued_avg, 68 nthreads, 69 requeuetime_avg / USEC_PER_MSEC, 70 rel_stddev_stats(requeuetime_stddev, requeuetime_avg)); 71 } 72 73 static void *workerfn(void *arg __maybe_unused) 74 { 75 pthread_mutex_lock(&thread_lock); 76 threads_starting--; 77 if (!threads_starting) 78 pthread_cond_signal(&thread_parent); 79 pthread_cond_wait(&thread_worker, &thread_lock); 80 pthread_mutex_unlock(&thread_lock); 81 82 futex_wait(&futex1, 0, NULL, futex_flag); 83 return NULL; 84 } 85 86 static void block_threads(pthread_t *w, 87 pthread_attr_t thread_attr, struct cpu_map *cpu) 88 { 89 cpu_set_t cpuset; 90 unsigned int i; 91 92 threads_starting = nthreads; 93 94 /* create and block all threads */ 95 for (i = 0; i < nthreads; i++) { 96 CPU_ZERO(&cpuset); 97 CPU_SET(cpu->map[i % cpu->nr], &cpuset); 98 99 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset)) 100 err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); 101 102 if (pthread_create(&w[i], &thread_attr, workerfn, NULL)) 103 err(EXIT_FAILURE, "pthread_create"); 104 } 105 } 106 107 static void toggle_done(int sig __maybe_unused, 108 siginfo_t *info __maybe_unused, 109 void *uc __maybe_unused) 110 { 111 done = true; 112 } 113 114 int bench_futex_requeue(int argc, const char **argv) 115 { 116 int ret = 0; 117 unsigned int i, j; 118 struct sigaction act; 119 pthread_attr_t thread_attr; 120 struct cpu_map *cpu; 121 122 argc = parse_options(argc, argv, options, bench_futex_requeue_usage, 0); 123 if (argc) 124 goto err; 125 126 cpu = cpu_map__new(NULL); 127 if (!cpu) 128 err(EXIT_FAILURE, "cpu_map__new"); 129 130 sigfillset(&act.sa_mask); 131 act.sa_sigaction = toggle_done; 132 sigaction(SIGINT, &act, NULL); 133 134 if (!nthreads) 135 nthreads = cpu->nr; 136 137 worker = calloc(nthreads, sizeof(*worker)); 138 if (!worker) 139 err(EXIT_FAILURE, "calloc"); 140 141 if (!fshared) 142 futex_flag = FUTEX_PRIVATE_FLAG; 143 144 if (nrequeue > nthreads) 145 nrequeue = nthreads; 146 147 printf("Run summary [PID %d]: Requeuing %d threads (from [%s] %p to %p), " 148 "%d at a time.\n\n", getpid(), nthreads, 149 fshared ? "shared":"private", &futex1, &futex2, nrequeue); 150 151 init_stats(&requeued_stats); 152 init_stats(&requeuetime_stats); 153 pthread_attr_init(&thread_attr); 154 pthread_mutex_init(&thread_lock, NULL); 155 pthread_cond_init(&thread_parent, NULL); 156 pthread_cond_init(&thread_worker, NULL); 157 158 for (j = 0; j < bench_repeat && !done; j++) { 159 unsigned int nrequeued = 0; 160 struct timeval start, end, runtime; 161 162 /* create, launch & block all threads */ 163 block_threads(worker, thread_attr, cpu); 164 165 /* make sure all threads are already blocked */ 166 pthread_mutex_lock(&thread_lock); 167 while (threads_starting) 168 pthread_cond_wait(&thread_parent, &thread_lock); 169 pthread_cond_broadcast(&thread_worker); 170 pthread_mutex_unlock(&thread_lock); 171 172 usleep(100000); 173 174 /* Ok, all threads are patiently blocked, start requeueing */ 175 gettimeofday(&start, NULL); 176 while (nrequeued < nthreads) { 177 /* 178 * Do not wakeup any tasks blocked on futex1, allowing 179 * us to really measure futex_wait functionality. 180 */ 181 nrequeued += futex_cmp_requeue(&futex1, 0, &futex2, 0, 182 nrequeue, futex_flag); 183 } 184 185 gettimeofday(&end, NULL); 186 timersub(&end, &start, &runtime); 187 188 update_stats(&requeued_stats, nrequeued); 189 update_stats(&requeuetime_stats, runtime.tv_usec); 190 191 if (!silent) { 192 printf("[Run %d]: Requeued %d of %d threads in %.4f ms\n", 193 j + 1, nrequeued, nthreads, runtime.tv_usec / (double)USEC_PER_MSEC); 194 } 195 196 /* everybody should be blocked on futex2, wake'em up */ 197 nrequeued = futex_wake(&futex2, nrequeued, futex_flag); 198 if (nthreads != nrequeued) 199 warnx("couldn't wakeup all tasks (%d/%d)", nrequeued, nthreads); 200 201 for (i = 0; i < nthreads; i++) { 202 ret = pthread_join(worker[i], NULL); 203 if (ret) 204 err(EXIT_FAILURE, "pthread_join"); 205 } 206 } 207 208 /* cleanup & report results */ 209 pthread_cond_destroy(&thread_parent); 210 pthread_cond_destroy(&thread_worker); 211 pthread_mutex_destroy(&thread_lock); 212 pthread_attr_destroy(&thread_attr); 213 214 print_summary(); 215 216 free(worker); 217 return ret; 218 err: 219 usage_with_options(bench_futex_requeue_usage, options); 220 exit(EXIT_FAILURE); 221 } 222