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/mutex.h" 12 #include "../util/stat.h" 13 #include <subcmd/parse-options.h> 14 #include <linux/compiler.h> 15 #include <linux/kernel.h> 16 #include <linux/zalloc.h> 17 #include <errno.h> 18 #include <perf/cpumap.h> 19 #include "bench.h" 20 #include "futex.h" 21 22 #include <err.h> 23 #include <stdlib.h> 24 #include <sys/time.h> 25 #include <sys/mman.h> 26 27 struct worker { 28 int tid; 29 u_int32_t *futex; 30 pthread_t thread; 31 unsigned long ops; 32 }; 33 34 static u_int32_t global_futex = 0; 35 static struct worker *worker; 36 static bool done = false; 37 static int futex_flag = 0; 38 static struct mutex thread_lock; 39 static unsigned int threads_starting; 40 static struct stats throughput_stats; 41 static struct cond thread_parent, thread_worker; 42 43 static struct bench_futex_parameters params = { 44 .runtime = 10, 45 }; 46 47 static const struct option options[] = { 48 OPT_UINTEGER('t', "threads", ¶ms.nthreads, "Specify amount of threads"), 49 OPT_UINTEGER('r', "runtime", ¶ms.runtime, "Specify runtime (in seconds)"), 50 OPT_BOOLEAN( 'M', "multi", ¶ms.multi, "Use multiple futexes"), 51 OPT_BOOLEAN( 's', "silent", ¶ms.silent, "Silent mode: do not display data/details"), 52 OPT_BOOLEAN( 'S', "shared", ¶ms.fshared, "Use shared futexes instead of private ones"), 53 OPT_BOOLEAN( 'm', "mlockall", ¶ms.mlockall, "Lock all current and future memory"), 54 OPT_END() 55 }; 56 57 static const char * const bench_futex_lock_pi_usage[] = { 58 "perf bench futex lock-pi <options>", 59 NULL 60 }; 61 62 static void print_summary(void) 63 { 64 unsigned long avg = avg_stats(&throughput_stats); 65 double stddev = stddev_stats(&throughput_stats); 66 67 printf("%sAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n", 68 !params.silent ? "\n" : "", avg, rel_stddev_stats(stddev, avg), 69 (int)bench__runtime.tv_sec); 70 } 71 72 static void toggle_done(int sig __maybe_unused, 73 siginfo_t *info __maybe_unused, 74 void *uc __maybe_unused) 75 { 76 /* inform all threads that we're done for the day */ 77 done = true; 78 gettimeofday(&bench__end, NULL); 79 timersub(&bench__end, &bench__start, &bench__runtime); 80 } 81 82 static void *workerfn(void *arg) 83 { 84 struct worker *w = (struct worker *) arg; 85 unsigned long ops = w->ops; 86 87 mutex_lock(&thread_lock); 88 threads_starting--; 89 if (!threads_starting) 90 cond_signal(&thread_parent); 91 cond_wait(&thread_worker, &thread_lock); 92 mutex_unlock(&thread_lock); 93 94 do { 95 int ret; 96 again: 97 ret = futex_lock_pi(w->futex, NULL, futex_flag); 98 99 if (ret) { /* handle lock acquisition */ 100 if (!params.silent) 101 warn("thread %d: Could not lock pi-lock for %p (%d)", 102 w->tid, w->futex, ret); 103 if (done) 104 break; 105 106 goto again; 107 } 108 109 usleep(1); 110 ret = futex_unlock_pi(w->futex, futex_flag); 111 if (ret && !params.silent) 112 warn("thread %d: Could not unlock pi-lock for %p (%d)", 113 w->tid, w->futex, ret); 114 ops++; /* account for thread's share of work */ 115 } while (!done); 116 117 w->ops = ops; 118 return NULL; 119 } 120 121 static void create_threads(struct worker *w, pthread_attr_t thread_attr, 122 struct perf_cpu_map *cpu) 123 { 124 cpu_set_t *cpuset; 125 unsigned int i; 126 int nrcpus = perf_cpu_map__nr(cpu); 127 size_t size; 128 129 threads_starting = params.nthreads; 130 131 cpuset = CPU_ALLOC(nrcpus); 132 BUG_ON(!cpuset); 133 size = CPU_ALLOC_SIZE(nrcpus); 134 135 for (i = 0; i < params.nthreads; i++) { 136 worker[i].tid = i; 137 138 if (params.multi) { 139 worker[i].futex = calloc(1, sizeof(u_int32_t)); 140 if (!worker[i].futex) 141 err(EXIT_FAILURE, "calloc"); 142 } else 143 worker[i].futex = &global_futex; 144 145 CPU_ZERO_S(size, cpuset); 146 CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset); 147 148 if (pthread_attr_setaffinity_np(&thread_attr, size, cpuset)) { 149 CPU_FREE(cpuset); 150 err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); 151 } 152 153 if (pthread_create(&w[i].thread, &thread_attr, workerfn, &worker[i])) { 154 CPU_FREE(cpuset); 155 err(EXIT_FAILURE, "pthread_create"); 156 } 157 } 158 CPU_FREE(cpuset); 159 } 160 161 int bench_futex_lock_pi(int argc, const char **argv) 162 { 163 int ret = 0; 164 unsigned int i; 165 struct sigaction act; 166 pthread_attr_t thread_attr; 167 struct perf_cpu_map *cpu; 168 169 argc = parse_options(argc, argv, options, bench_futex_lock_pi_usage, 0); 170 if (argc) 171 goto err; 172 173 cpu = perf_cpu_map__new(NULL); 174 if (!cpu) 175 err(EXIT_FAILURE, "calloc"); 176 177 memset(&act, 0, sizeof(act)); 178 sigfillset(&act.sa_mask); 179 act.sa_sigaction = toggle_done; 180 sigaction(SIGINT, &act, NULL); 181 182 if (params.mlockall) { 183 if (mlockall(MCL_CURRENT | MCL_FUTURE)) 184 err(EXIT_FAILURE, "mlockall"); 185 } 186 187 if (!params.nthreads) 188 params.nthreads = perf_cpu_map__nr(cpu); 189 190 worker = calloc(params.nthreads, sizeof(*worker)); 191 if (!worker) 192 err(EXIT_FAILURE, "calloc"); 193 194 if (!params.fshared) 195 futex_flag = FUTEX_PRIVATE_FLAG; 196 197 printf("Run summary [PID %d]: %d threads doing pi lock/unlock pairing for %d secs.\n\n", 198 getpid(), params.nthreads, params.runtime); 199 200 init_stats(&throughput_stats); 201 mutex_init(&thread_lock); 202 cond_init(&thread_parent); 203 cond_init(&thread_worker); 204 205 threads_starting = params.nthreads; 206 pthread_attr_init(&thread_attr); 207 gettimeofday(&bench__start, NULL); 208 209 create_threads(worker, thread_attr, cpu); 210 pthread_attr_destroy(&thread_attr); 211 212 mutex_lock(&thread_lock); 213 while (threads_starting) 214 cond_wait(&thread_parent, &thread_lock); 215 cond_broadcast(&thread_worker); 216 mutex_unlock(&thread_lock); 217 218 sleep(params.runtime); 219 toggle_done(0, NULL, NULL); 220 221 for (i = 0; i < params.nthreads; i++) { 222 ret = pthread_join(worker[i].thread, NULL); 223 if (ret) 224 err(EXIT_FAILURE, "pthread_join"); 225 } 226 227 /* cleanup & report results */ 228 cond_destroy(&thread_parent); 229 cond_destroy(&thread_worker); 230 mutex_destroy(&thread_lock); 231 232 for (i = 0; i < params.nthreads; i++) { 233 unsigned long t = bench__runtime.tv_sec > 0 ? 234 worker[i].ops / bench__runtime.tv_sec : 0; 235 236 update_stats(&throughput_stats, t); 237 if (!params.silent) 238 printf("[thread %3d] futex: %p [ %ld ops/sec ]\n", 239 worker[i].tid, worker[i].futex, t); 240 241 if (params.multi) 242 zfree(&worker[i].futex); 243 } 244 245 print_summary(); 246 247 free(worker); 248 perf_cpu_map__put(cpu); 249 return ret; 250 err: 251 usage_with_options(bench_futex_lock_pi_usage, options); 252 exit(EXIT_FAILURE); 253 } 254