1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (C) 2015 Davidlohr Bueso. 4 */ 5 6 /* For the CLR_() macros */ 7 #include <string.h> 8 #include <pthread.h> 9 10 #include <signal.h> 11 #include "../util/stat.h" 12 #include <subcmd/parse-options.h> 13 #include <linux/compiler.h> 14 #include <linux/kernel.h> 15 #include <linux/zalloc.h> 16 #include <errno.h> 17 #include <perf/cpumap.h> 18 #include "bench.h" 19 #include "futex.h" 20 21 #include <err.h> 22 #include <stdlib.h> 23 #include <sys/time.h> 24 #include <sys/mman.h> 25 26 struct worker { 27 int tid; 28 u_int32_t *futex; 29 pthread_t thread; 30 unsigned long ops; 31 }; 32 33 static u_int32_t global_futex = 0; 34 static struct worker *worker; 35 static bool done = false; 36 static int futex_flag = 0; 37 static pthread_mutex_t thread_lock; 38 static unsigned int threads_starting; 39 static struct stats throughput_stats; 40 static pthread_cond_t thread_parent, thread_worker; 41 42 static struct bench_futex_parameters params = { 43 .runtime = 10, 44 }; 45 46 static const struct option options[] = { 47 OPT_UINTEGER('t', "threads", ¶ms.nthreads, "Specify amount of threads"), 48 OPT_UINTEGER('r', "runtime", ¶ms.runtime, "Specify runtime (in seconds)"), 49 OPT_BOOLEAN( 'M', "multi", ¶ms.multi, "Use multiple futexes"), 50 OPT_BOOLEAN( 's', "silent", ¶ms.silent, "Silent mode: do not display data/details"), 51 OPT_BOOLEAN( 'S', "shared", ¶ms.fshared, "Use shared futexes instead of private ones"), 52 OPT_BOOLEAN( 'm', "mlockall", ¶ms.mlockall, "Lock all current and future memory"), 53 OPT_END() 54 }; 55 56 static const char * const bench_futex_lock_pi_usage[] = { 57 "perf bench futex lock-pi <options>", 58 NULL 59 }; 60 61 static void print_summary(void) 62 { 63 unsigned long avg = avg_stats(&throughput_stats); 64 double stddev = stddev_stats(&throughput_stats); 65 66 printf("%sAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n", 67 !params.silent ? "\n" : "", avg, rel_stddev_stats(stddev, avg), 68 (int)bench__runtime.tv_sec); 69 } 70 71 static void toggle_done(int sig __maybe_unused, 72 siginfo_t *info __maybe_unused, 73 void *uc __maybe_unused) 74 { 75 /* inform all threads that we're done for the day */ 76 done = true; 77 gettimeofday(&bench__end, NULL); 78 timersub(&bench__end, &bench__start, &bench__runtime); 79 } 80 81 static void *workerfn(void *arg) 82 { 83 struct worker *w = (struct worker *) arg; 84 unsigned long ops = w->ops; 85 86 pthread_mutex_lock(&thread_lock); 87 threads_starting--; 88 if (!threads_starting) 89 pthread_cond_signal(&thread_parent); 90 pthread_cond_wait(&thread_worker, &thread_lock); 91 pthread_mutex_unlock(&thread_lock); 92 93 do { 94 int ret; 95 again: 96 ret = futex_lock_pi(w->futex, NULL, futex_flag); 97 98 if (ret) { /* handle lock acquisition */ 99 if (!params.silent) 100 warn("thread %d: Could not lock pi-lock for %p (%d)", 101 w->tid, w->futex, ret); 102 if (done) 103 break; 104 105 goto again; 106 } 107 108 usleep(1); 109 ret = futex_unlock_pi(w->futex, futex_flag); 110 if (ret && !params.silent) 111 warn("thread %d: Could not unlock pi-lock for %p (%d)", 112 w->tid, w->futex, ret); 113 ops++; /* account for thread's share of work */ 114 } while (!done); 115 116 w->ops = ops; 117 return NULL; 118 } 119 120 static void create_threads(struct worker *w, pthread_attr_t thread_attr, 121 struct perf_cpu_map *cpu) 122 { 123 cpu_set_t cpuset; 124 unsigned int i; 125 126 threads_starting = params.nthreads; 127 128 for (i = 0; i < params.nthreads; i++) { 129 worker[i].tid = i; 130 131 if (params.multi) { 132 worker[i].futex = calloc(1, sizeof(u_int32_t)); 133 if (!worker[i].futex) 134 err(EXIT_FAILURE, "calloc"); 135 } else 136 worker[i].futex = &global_futex; 137 138 CPU_ZERO(&cpuset); 139 CPU_SET(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, &cpuset); 140 141 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset)) 142 err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); 143 144 if (pthread_create(&w[i].thread, &thread_attr, workerfn, &worker[i])) 145 err(EXIT_FAILURE, "pthread_create"); 146 } 147 } 148 149 int bench_futex_lock_pi(int argc, const char **argv) 150 { 151 int ret = 0; 152 unsigned int i; 153 struct sigaction act; 154 pthread_attr_t thread_attr; 155 struct perf_cpu_map *cpu; 156 157 argc = parse_options(argc, argv, options, bench_futex_lock_pi_usage, 0); 158 if (argc) 159 goto err; 160 161 cpu = perf_cpu_map__new(NULL); 162 if (!cpu) 163 err(EXIT_FAILURE, "calloc"); 164 165 memset(&act, 0, sizeof(act)); 166 sigfillset(&act.sa_mask); 167 act.sa_sigaction = toggle_done; 168 sigaction(SIGINT, &act, NULL); 169 170 if (params.mlockall) { 171 if (mlockall(MCL_CURRENT | MCL_FUTURE)) 172 err(EXIT_FAILURE, "mlockall"); 173 } 174 175 if (!params.nthreads) 176 params.nthreads = perf_cpu_map__nr(cpu); 177 178 worker = calloc(params.nthreads, sizeof(*worker)); 179 if (!worker) 180 err(EXIT_FAILURE, "calloc"); 181 182 if (!params.fshared) 183 futex_flag = FUTEX_PRIVATE_FLAG; 184 185 printf("Run summary [PID %d]: %d threads doing pi lock/unlock pairing for %d secs.\n\n", 186 getpid(), params.nthreads, params.runtime); 187 188 init_stats(&throughput_stats); 189 pthread_mutex_init(&thread_lock, NULL); 190 pthread_cond_init(&thread_parent, NULL); 191 pthread_cond_init(&thread_worker, NULL); 192 193 threads_starting = params.nthreads; 194 pthread_attr_init(&thread_attr); 195 gettimeofday(&bench__start, NULL); 196 197 create_threads(worker, thread_attr, cpu); 198 pthread_attr_destroy(&thread_attr); 199 200 pthread_mutex_lock(&thread_lock); 201 while (threads_starting) 202 pthread_cond_wait(&thread_parent, &thread_lock); 203 pthread_cond_broadcast(&thread_worker); 204 pthread_mutex_unlock(&thread_lock); 205 206 sleep(params.runtime); 207 toggle_done(0, NULL, NULL); 208 209 for (i = 0; i < params.nthreads; i++) { 210 ret = pthread_join(worker[i].thread, NULL); 211 if (ret) 212 err(EXIT_FAILURE, "pthread_join"); 213 } 214 215 /* cleanup & report results */ 216 pthread_cond_destroy(&thread_parent); 217 pthread_cond_destroy(&thread_worker); 218 pthread_mutex_destroy(&thread_lock); 219 220 for (i = 0; i < params.nthreads; i++) { 221 unsigned long t = bench__runtime.tv_sec > 0 ? 222 worker[i].ops / bench__runtime.tv_sec : 0; 223 224 update_stats(&throughput_stats, t); 225 if (!params.silent) 226 printf("[thread %3d] futex: %p [ %ld ops/sec ]\n", 227 worker[i].tid, worker[i].futex, t); 228 229 if (params.multi) 230 zfree(&worker[i].futex); 231 } 232 233 print_summary(); 234 235 free(worker); 236 perf_cpu_map__put(cpu); 237 return ret; 238 err: 239 usage_with_options(bench_futex_lock_pi_usage, options); 240 exit(EXIT_FAILURE); 241 } 242