1 /* 2 * Copyright (C) 2013 Davidlohr Bueso <davidlohr@hp.com> 3 * 4 * futex-requeue: Block a bunch of threads on futex1 and requeue them 5 * on futex2, N at a time. 6 * 7 * This program is particularly useful to measure the latency of nthread 8 * requeues without waking up any tasks -- thus mimicking a regular futex_wait. 9 */ 10 11 /* For the CLR_() macros */ 12 #include <pthread.h> 13 14 #include <signal.h> 15 #include "../util/stat.h" 16 #include <subcmd/parse-options.h> 17 #include <linux/compiler.h> 18 #include <linux/kernel.h> 19 #include <linux/time64.h> 20 #include <errno.h> 21 #include "bench.h" 22 #include "futex.h" 23 24 #include <err.h> 25 #include <stdlib.h> 26 #include <sys/time.h> 27 28 static u_int32_t futex1 = 0, futex2 = 0; 29 30 /* 31 * How many tasks to requeue at a time. 32 * Default to 1 in order to make the kernel work more. 33 */ 34 static unsigned int nrequeue = 1; 35 36 static pthread_t *worker; 37 static bool done = false, silent = false, fshared = false; 38 static pthread_mutex_t thread_lock; 39 static pthread_cond_t thread_parent, thread_worker; 40 static struct stats requeuetime_stats, requeued_stats; 41 static unsigned int ncpus, threads_starting, nthreads = 0; 42 static int futex_flag = 0; 43 44 static const struct option options[] = { 45 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"), 46 OPT_UINTEGER('q', "nrequeue", &nrequeue, "Specify amount of threads to requeue at once"), 47 OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"), 48 OPT_BOOLEAN( 'S', "shared", &fshared, "Use shared futexes instead of private ones"), 49 OPT_END() 50 }; 51 52 static const char * const bench_futex_requeue_usage[] = { 53 "perf bench futex requeue <options>", 54 NULL 55 }; 56 57 static void print_summary(void) 58 { 59 double requeuetime_avg = avg_stats(&requeuetime_stats); 60 double requeuetime_stddev = stddev_stats(&requeuetime_stats); 61 unsigned int requeued_avg = avg_stats(&requeued_stats); 62 63 printf("Requeued %d of %d threads in %.4f ms (+-%.2f%%)\n", 64 requeued_avg, 65 nthreads, 66 requeuetime_avg / USEC_PER_MSEC, 67 rel_stddev_stats(requeuetime_stddev, requeuetime_avg)); 68 } 69 70 static void *workerfn(void *arg __maybe_unused) 71 { 72 pthread_mutex_lock(&thread_lock); 73 threads_starting--; 74 if (!threads_starting) 75 pthread_cond_signal(&thread_parent); 76 pthread_cond_wait(&thread_worker, &thread_lock); 77 pthread_mutex_unlock(&thread_lock); 78 79 futex_wait(&futex1, 0, NULL, futex_flag); 80 return NULL; 81 } 82 83 static void block_threads(pthread_t *w, 84 pthread_attr_t thread_attr) 85 { 86 cpu_set_t cpu; 87 unsigned int i; 88 89 threads_starting = nthreads; 90 91 /* create and block all threads */ 92 for (i = 0; i < nthreads; i++) { 93 CPU_ZERO(&cpu); 94 CPU_SET(i % ncpus, &cpu); 95 96 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu)) 97 err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); 98 99 if (pthread_create(&w[i], &thread_attr, workerfn, NULL)) 100 err(EXIT_FAILURE, "pthread_create"); 101 } 102 } 103 104 static void toggle_done(int sig __maybe_unused, 105 siginfo_t *info __maybe_unused, 106 void *uc __maybe_unused) 107 { 108 done = true; 109 } 110 111 int bench_futex_requeue(int argc, const char **argv, 112 const char *prefix __maybe_unused) 113 { 114 int ret = 0; 115 unsigned int i, j; 116 struct sigaction act; 117 pthread_attr_t thread_attr; 118 119 argc = parse_options(argc, argv, options, bench_futex_requeue_usage, 0); 120 if (argc) 121 goto err; 122 123 ncpus = sysconf(_SC_NPROCESSORS_ONLN); 124 125 sigfillset(&act.sa_mask); 126 act.sa_sigaction = toggle_done; 127 sigaction(SIGINT, &act, NULL); 128 129 if (!nthreads) 130 nthreads = ncpus; 131 else 132 nthreads = futexbench_sanitize_numeric(nthreads); 133 134 worker = calloc(nthreads, sizeof(*worker)); 135 if (!worker) 136 err(EXIT_FAILURE, "calloc"); 137 138 if (!fshared) 139 futex_flag = FUTEX_PRIVATE_FLAG; 140 141 if (nrequeue > nthreads) 142 nrequeue = nthreads; 143 144 printf("Run summary [PID %d]: Requeuing %d threads (from [%s] %p to %p), " 145 "%d at a time.\n\n", getpid(), nthreads, 146 fshared ? "shared":"private", &futex1, &futex2, nrequeue); 147 148 init_stats(&requeued_stats); 149 init_stats(&requeuetime_stats); 150 pthread_attr_init(&thread_attr); 151 pthread_mutex_init(&thread_lock, NULL); 152 pthread_cond_init(&thread_parent, NULL); 153 pthread_cond_init(&thread_worker, NULL); 154 155 for (j = 0; j < bench_repeat && !done; j++) { 156 unsigned int nrequeued = 0; 157 struct timeval start, end, runtime; 158 159 /* create, launch & block all threads */ 160 block_threads(worker, thread_attr); 161 162 /* make sure all threads are already blocked */ 163 pthread_mutex_lock(&thread_lock); 164 while (threads_starting) 165 pthread_cond_wait(&thread_parent, &thread_lock); 166 pthread_cond_broadcast(&thread_worker); 167 pthread_mutex_unlock(&thread_lock); 168 169 usleep(100000); 170 171 /* Ok, all threads are patiently blocked, start requeueing */ 172 gettimeofday(&start, NULL); 173 while (nrequeued < nthreads) { 174 /* 175 * Do not wakeup any tasks blocked on futex1, allowing 176 * us to really measure futex_wait functionality. 177 */ 178 nrequeued += futex_cmp_requeue(&futex1, 0, &futex2, 0, 179 nrequeue, futex_flag); 180 } 181 182 gettimeofday(&end, NULL); 183 timersub(&end, &start, &runtime); 184 185 update_stats(&requeued_stats, nrequeued); 186 update_stats(&requeuetime_stats, runtime.tv_usec); 187 188 if (!silent) { 189 printf("[Run %d]: Requeued %d of %d threads in %.4f ms\n", 190 j + 1, nrequeued, nthreads, runtime.tv_usec / (double)USEC_PER_MSEC); 191 } 192 193 /* everybody should be blocked on futex2, wake'em up */ 194 nrequeued = futex_wake(&futex2, nrequeued, futex_flag); 195 if (nthreads != nrequeued) 196 warnx("couldn't wakeup all tasks (%d/%d)", nrequeued, nthreads); 197 198 for (i = 0; i < nthreads; i++) { 199 ret = pthread_join(worker[i], NULL); 200 if (ret) 201 err(EXIT_FAILURE, "pthread_join"); 202 } 203 } 204 205 /* cleanup & report results */ 206 pthread_cond_destroy(&thread_parent); 207 pthread_cond_destroy(&thread_worker); 208 pthread_mutex_destroy(&thread_lock); 209 pthread_attr_destroy(&thread_attr); 210 211 print_summary(); 212 213 free(worker); 214 return ret; 215 err: 216 usage_with_options(bench_futex_requeue_usage, options); 217 exit(EXIT_FAILURE); 218 } 219