1 /* 2 * Copyright (C) 2013 Davidlohr Bueso <davidlohr@hp.com> 3 * 4 * futex-hash: Stress the hell out of the Linux kernel futex uaddr hashing. 5 * 6 * This program is particularly useful for measuring the kernel's futex hash 7 * table/function implementation. In order for it to make sense, use with as 8 * many threads and futexes as possible. 9 */ 10 11 #include "../perf.h" 12 #include "../util/util.h" 13 #include "../util/stat.h" 14 #include "../util/parse-options.h" 15 #include "../util/header.h" 16 #include "bench.h" 17 #include "futex.h" 18 19 #include <err.h> 20 #include <stdlib.h> 21 #include <sys/time.h> 22 #include <pthread.h> 23 24 static unsigned int nthreads = 0; 25 static unsigned int nsecs = 10; 26 /* amount of futexes per thread */ 27 static unsigned int nfutexes = 1024; 28 static bool fshared = false, done = false, silent = false; 29 30 struct timeval start, end, runtime; 31 static pthread_mutex_t thread_lock; 32 static unsigned int threads_starting; 33 static struct stats throughput_stats; 34 static pthread_cond_t thread_parent, thread_worker; 35 36 struct worker { 37 int tid; 38 u_int32_t *futex; 39 pthread_t thread; 40 unsigned long ops; 41 }; 42 43 static const struct option options[] = { 44 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"), 45 OPT_UINTEGER('r', "runtime", &nsecs, "Specify runtime (in seconds)"), 46 OPT_UINTEGER('f', "futexes", &nfutexes, "Specify amount of futexes per threads"), 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_hash_usage[] = { 53 "perf bench futex hash <options>", 54 NULL 55 }; 56 57 static void *workerfn(void *arg) 58 { 59 int ret; 60 unsigned int i; 61 struct worker *w = (struct worker *) arg; 62 63 pthread_mutex_lock(&thread_lock); 64 threads_starting--; 65 if (!threads_starting) 66 pthread_cond_signal(&thread_parent); 67 pthread_cond_wait(&thread_worker, &thread_lock); 68 pthread_mutex_unlock(&thread_lock); 69 70 do { 71 for (i = 0; i < nfutexes; i++, w->ops++) { 72 /* 73 * We want the futex calls to fail in order to stress 74 * the hashing of uaddr and not measure other steps, 75 * such as internal waitqueue handling, thus enlarging 76 * the critical region protected by hb->lock. 77 */ 78 ret = futex_wait(&w->futex[i], 1234, NULL, 79 fshared ? 0 : FUTEX_PRIVATE_FLAG); 80 if (!silent && 81 (!ret || errno != EAGAIN || errno != EWOULDBLOCK)) 82 warn("Non-expected futex return call"); 83 } 84 } while (!done); 85 86 return NULL; 87 } 88 89 static void toggle_done(int sig __maybe_unused, 90 siginfo_t *info __maybe_unused, 91 void *uc __maybe_unused) 92 { 93 /* inform all threads that we're done for the day */ 94 done = true; 95 gettimeofday(&end, NULL); 96 timersub(&end, &start, &runtime); 97 } 98 99 static void print_summary(void) 100 { 101 unsigned long avg = avg_stats(&throughput_stats); 102 double stddev = stddev_stats(&throughput_stats); 103 104 printf("%sAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n", 105 !silent ? "\n" : "", avg, rel_stddev_stats(stddev, avg), 106 (int) runtime.tv_sec); 107 } 108 109 int bench_futex_hash(int argc, const char **argv, 110 const char *prefix __maybe_unused) 111 { 112 int ret = 0; 113 cpu_set_t cpu; 114 struct sigaction act; 115 unsigned int i, ncpus; 116 pthread_attr_t thread_attr; 117 struct worker *worker = NULL; 118 119 argc = parse_options(argc, argv, options, bench_futex_hash_usage, 0); 120 if (argc) { 121 usage_with_options(bench_futex_hash_usage, options); 122 exit(EXIT_FAILURE); 123 } 124 125 ncpus = sysconf(_SC_NPROCESSORS_ONLN); 126 127 sigfillset(&act.sa_mask); 128 act.sa_sigaction = toggle_done; 129 sigaction(SIGINT, &act, NULL); 130 131 if (!nthreads) /* default to the number of CPUs */ 132 nthreads = ncpus; 133 134 worker = calloc(nthreads, sizeof(*worker)); 135 if (!worker) 136 goto errmem; 137 138 printf("Run summary [PID %d]: %d threads, each operating on %d [%s] futexes for %d secs.\n\n", 139 getpid(), nthreads, nfutexes, fshared ? "shared":"private", nsecs); 140 141 init_stats(&throughput_stats); 142 pthread_mutex_init(&thread_lock, NULL); 143 pthread_cond_init(&thread_parent, NULL); 144 pthread_cond_init(&thread_worker, NULL); 145 146 threads_starting = nthreads; 147 pthread_attr_init(&thread_attr); 148 gettimeofday(&start, NULL); 149 for (i = 0; i < nthreads; i++) { 150 worker[i].tid = i; 151 worker[i].futex = calloc(nfutexes, sizeof(*worker[i].futex)); 152 if (!worker[i].futex) 153 goto errmem; 154 155 CPU_ZERO(&cpu); 156 CPU_SET(i % ncpus, &cpu); 157 158 ret = pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu); 159 if (ret) 160 err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); 161 162 ret = pthread_create(&worker[i].thread, &thread_attr, workerfn, 163 (void *)(struct worker *) &worker[i]); 164 if (ret) 165 err(EXIT_FAILURE, "pthread_create"); 166 167 } 168 pthread_attr_destroy(&thread_attr); 169 170 pthread_mutex_lock(&thread_lock); 171 while (threads_starting) 172 pthread_cond_wait(&thread_parent, &thread_lock); 173 pthread_cond_broadcast(&thread_worker); 174 pthread_mutex_unlock(&thread_lock); 175 176 sleep(nsecs); 177 toggle_done(0, NULL, NULL); 178 179 for (i = 0; i < nthreads; i++) { 180 ret = pthread_join(worker[i].thread, NULL); 181 if (ret) 182 err(EXIT_FAILURE, "pthread_join"); 183 } 184 185 /* cleanup & report results */ 186 pthread_cond_destroy(&thread_parent); 187 pthread_cond_destroy(&thread_worker); 188 pthread_mutex_destroy(&thread_lock); 189 190 for (i = 0; i < nthreads; i++) { 191 unsigned long t = worker[i].ops/runtime.tv_sec; 192 update_stats(&throughput_stats, t); 193 if (!silent) { 194 if (nfutexes == 1) 195 printf("[thread %2d] futex: %p [ %ld ops/sec ]\n", 196 worker[i].tid, &worker[i].futex[0], t); 197 else 198 printf("[thread %2d] futexes: %p ... %p [ %ld ops/sec ]\n", 199 worker[i].tid, &worker[i].futex[0], 200 &worker[i].futex[nfutexes-1], t); 201 } 202 203 free(worker[i].futex); 204 } 205 206 print_summary(); 207 208 free(worker); 209 return ret; 210 errmem: 211 err(EXIT_FAILURE, "calloc"); 212 } 213