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 int nrcpus = perf_cpu_map__nr(cpu); 126 size_t size; 127 128 threads_starting = params.nthreads; 129 130 cpuset = CPU_ALLOC(nrcpus); 131 BUG_ON(!cpuset); 132 size = CPU_ALLOC_SIZE(nrcpus); 133 134 for (i = 0; i < params.nthreads; i++) { 135 worker[i].tid = i; 136 137 if (params.multi) { 138 worker[i].futex = calloc(1, sizeof(u_int32_t)); 139 if (!worker[i].futex) 140 err(EXIT_FAILURE, "calloc"); 141 } else 142 worker[i].futex = &global_futex; 143 144 CPU_ZERO_S(size, cpuset); 145 CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset); 146 147 if (pthread_attr_setaffinity_np(&thread_attr, size, cpuset)) { 148 CPU_FREE(cpuset); 149 err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); 150 } 151 152 if (pthread_create(&w[i].thread, &thread_attr, workerfn, &worker[i])) { 153 CPU_FREE(cpuset); 154 err(EXIT_FAILURE, "pthread_create"); 155 } 156 } 157 CPU_FREE(cpuset); 158 } 159 160 int bench_futex_lock_pi(int argc, const char **argv) 161 { 162 int ret = 0; 163 unsigned int i; 164 struct sigaction act; 165 pthread_attr_t thread_attr; 166 struct perf_cpu_map *cpu; 167 168 argc = parse_options(argc, argv, options, bench_futex_lock_pi_usage, 0); 169 if (argc) 170 goto err; 171 172 cpu = perf_cpu_map__new(NULL); 173 if (!cpu) 174 err(EXIT_FAILURE, "calloc"); 175 176 memset(&act, 0, sizeof(act)); 177 sigfillset(&act.sa_mask); 178 act.sa_sigaction = toggle_done; 179 sigaction(SIGINT, &act, NULL); 180 181 if (params.mlockall) { 182 if (mlockall(MCL_CURRENT | MCL_FUTURE)) 183 err(EXIT_FAILURE, "mlockall"); 184 } 185 186 if (!params.nthreads) 187 params.nthreads = perf_cpu_map__nr(cpu); 188 189 worker = calloc(params.nthreads, sizeof(*worker)); 190 if (!worker) 191 err(EXIT_FAILURE, "calloc"); 192 193 if (!params.fshared) 194 futex_flag = FUTEX_PRIVATE_FLAG; 195 196 printf("Run summary [PID %d]: %d threads doing pi lock/unlock pairing for %d secs.\n\n", 197 getpid(), params.nthreads, params.runtime); 198 199 init_stats(&throughput_stats); 200 pthread_mutex_init(&thread_lock, NULL); 201 pthread_cond_init(&thread_parent, NULL); 202 pthread_cond_init(&thread_worker, NULL); 203 204 threads_starting = params.nthreads; 205 pthread_attr_init(&thread_attr); 206 gettimeofday(&bench__start, NULL); 207 208 create_threads(worker, thread_attr, cpu); 209 pthread_attr_destroy(&thread_attr); 210 211 pthread_mutex_lock(&thread_lock); 212 while (threads_starting) 213 pthread_cond_wait(&thread_parent, &thread_lock); 214 pthread_cond_broadcast(&thread_worker); 215 pthread_mutex_unlock(&thread_lock); 216 217 sleep(params.runtime); 218 toggle_done(0, NULL, NULL); 219 220 for (i = 0; i < params.nthreads; i++) { 221 ret = pthread_join(worker[i].thread, NULL); 222 if (ret) 223 err(EXIT_FAILURE, "pthread_join"); 224 } 225 226 /* cleanup & report results */ 227 pthread_cond_destroy(&thread_parent); 228 pthread_cond_destroy(&thread_worker); 229 pthread_mutex_destroy(&thread_lock); 230 231 for (i = 0; i < params.nthreads; i++) { 232 unsigned long t = bench__runtime.tv_sec > 0 ? 233 worker[i].ops / bench__runtime.tv_sec : 0; 234 235 update_stats(&throughput_stats, t); 236 if (!params.silent) 237 printf("[thread %3d] futex: %p [ %ld ops/sec ]\n", 238 worker[i].tid, worker[i].futex, t); 239 240 if (params.multi) 241 zfree(&worker[i].futex); 242 } 243 244 print_summary(); 245 246 free(worker); 247 perf_cpu_map__put(cpu); 248 return ret; 249 err: 250 usage_with_options(bench_futex_lock_pi_usage, options); 251 exit(EXIT_FAILURE); 252 } 253