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 .nbuckets = -1, 45 .runtime = 10, 46 }; 47 48 static const struct option options[] = { 49 OPT_INTEGER( 'b', "buckets", ¶ms.nbuckets, "Specify amount of hash buckets"), 50 OPT_UINTEGER('t', "threads", ¶ms.nthreads, "Specify amount of threads"), 51 OPT_UINTEGER('r', "runtime", ¶ms.runtime, "Specify runtime (in seconds)"), 52 OPT_BOOLEAN( 'M', "multi", ¶ms.multi, "Use multiple futexes"), 53 OPT_BOOLEAN( 's', "silent", ¶ms.silent, "Silent mode: do not display data/details"), 54 OPT_BOOLEAN( 'S', "shared", ¶ms.fshared, "Use shared futexes instead of private ones"), 55 OPT_BOOLEAN( 'm', "mlockall", ¶ms.mlockall, "Lock all current and future memory"), 56 OPT_END() 57 }; 58 59 static const char * const bench_futex_lock_pi_usage[] = { 60 "perf bench futex lock-pi <options>", 61 NULL 62 }; 63 64 static void print_summary(void) 65 { 66 unsigned long avg = avg_stats(&throughput_stats); 67 double stddev = stddev_stats(&throughput_stats); 68 69 printf("%sAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n", 70 !params.silent ? "\n" : "", avg, rel_stddev_stats(stddev, avg), 71 (int)bench__runtime.tv_sec); 72 futex_print_nbuckets(¶ms); 73 } 74 75 static void toggle_done(int sig __maybe_unused, 76 siginfo_t *info __maybe_unused, 77 void *uc __maybe_unused) 78 { 79 /* inform all threads that we're done for the day */ 80 done = true; 81 gettimeofday(&bench__end, NULL); 82 timersub(&bench__end, &bench__start, &bench__runtime); 83 } 84 85 static void *workerfn(void *arg) 86 { 87 struct worker *w = (struct worker *) arg; 88 unsigned long ops = w->ops; 89 90 mutex_lock(&thread_lock); 91 threads_starting--; 92 if (!threads_starting) 93 cond_signal(&thread_parent); 94 cond_wait(&thread_worker, &thread_lock); 95 mutex_unlock(&thread_lock); 96 97 do { 98 int ret; 99 again: 100 ret = futex_lock_pi(w->futex, NULL, futex_flag); 101 102 if (ret) { /* handle lock acquisition */ 103 if (!params.silent) 104 warn("thread %d: Could not lock pi-lock for %p (%d)", 105 w->tid, w->futex, ret); 106 if (done) 107 break; 108 109 goto again; 110 } 111 112 usleep(1); 113 ret = futex_unlock_pi(w->futex, futex_flag); 114 if (ret && !params.silent) 115 warn("thread %d: Could not unlock pi-lock for %p (%d)", 116 w->tid, w->futex, ret); 117 ops++; /* account for thread's share of work */ 118 } while (!done); 119 120 w->ops = ops; 121 return NULL; 122 } 123 124 static void create_threads(struct worker *w, struct perf_cpu_map *cpu) 125 { 126 cpu_set_t *cpuset; 127 unsigned int i; 128 int nrcpus = cpu__max_cpu().cpu; 129 size_t size; 130 131 threads_starting = params.nthreads; 132 133 cpuset = CPU_ALLOC(nrcpus); 134 BUG_ON(!cpuset); 135 size = CPU_ALLOC_SIZE(nrcpus); 136 137 for (i = 0; i < params.nthreads; i++) { 138 pthread_attr_t thread_attr; 139 140 pthread_attr_init(&thread_attr); 141 worker[i].tid = i; 142 143 if (params.multi) { 144 worker[i].futex = calloc(1, sizeof(u_int32_t)); 145 if (!worker[i].futex) 146 err(EXIT_FAILURE, "calloc"); 147 } else 148 worker[i].futex = &global_futex; 149 150 CPU_ZERO_S(size, cpuset); 151 CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset); 152 153 if (pthread_attr_setaffinity_np(&thread_attr, size, cpuset)) { 154 CPU_FREE(cpuset); 155 err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); 156 } 157 158 if (pthread_create(&w[i].thread, &thread_attr, workerfn, &worker[i])) { 159 CPU_FREE(cpuset); 160 err(EXIT_FAILURE, "pthread_create"); 161 } 162 pthread_attr_destroy(&thread_attr); 163 } 164 CPU_FREE(cpuset); 165 } 166 167 int bench_futex_lock_pi(int argc, const char **argv) 168 { 169 int ret = 0; 170 unsigned int i; 171 struct sigaction act; 172 struct perf_cpu_map *cpu; 173 174 argc = parse_options(argc, argv, options, bench_futex_lock_pi_usage, 0); 175 if (argc) 176 goto err; 177 178 cpu = perf_cpu_map__new_online_cpus(); 179 if (!cpu) 180 err(EXIT_FAILURE, "calloc"); 181 182 memset(&act, 0, sizeof(act)); 183 sigfillset(&act.sa_mask); 184 act.sa_sigaction = toggle_done; 185 sigaction(SIGINT, &act, NULL); 186 187 if (params.mlockall) { 188 if (mlockall(MCL_CURRENT | MCL_FUTURE)) 189 err(EXIT_FAILURE, "mlockall"); 190 } 191 192 if (!params.nthreads) 193 params.nthreads = perf_cpu_map__nr(cpu); 194 195 worker = calloc(params.nthreads, sizeof(*worker)); 196 if (!worker) 197 err(EXIT_FAILURE, "calloc"); 198 199 if (!params.fshared) 200 futex_flag = FUTEX_PRIVATE_FLAG; 201 202 printf("Run summary [PID %d]: %d threads doing pi lock/unlock pairing for %d secs.\n\n", 203 getpid(), params.nthreads, params.runtime); 204 205 init_stats(&throughput_stats); 206 mutex_init(&thread_lock); 207 cond_init(&thread_parent); 208 cond_init(&thread_worker); 209 futex_set_nbuckets_param(¶ms); 210 211 threads_starting = params.nthreads; 212 gettimeofday(&bench__start, NULL); 213 214 create_threads(worker, cpu); 215 216 mutex_lock(&thread_lock); 217 while (threads_starting) 218 cond_wait(&thread_parent, &thread_lock); 219 cond_broadcast(&thread_worker); 220 mutex_unlock(&thread_lock); 221 222 sleep(params.runtime); 223 toggle_done(0, NULL, NULL); 224 225 for (i = 0; i < params.nthreads; i++) { 226 ret = pthread_join(worker[i].thread, NULL); 227 if (ret) 228 err(EXIT_FAILURE, "pthread_join"); 229 } 230 231 /* cleanup & report results */ 232 cond_destroy(&thread_parent); 233 cond_destroy(&thread_worker); 234 mutex_destroy(&thread_lock); 235 236 for (i = 0; i < params.nthreads; i++) { 237 unsigned long t = bench__runtime.tv_sec > 0 ? 238 worker[i].ops / bench__runtime.tv_sec : 0; 239 240 update_stats(&throughput_stats, t); 241 if (!params.silent) 242 printf("[thread %3d] futex: %p [ %ld ops/sec ]\n", 243 worker[i].tid, worker[i].futex, t); 244 245 if (params.multi) 246 zfree(&worker[i].futex); 247 } 248 249 print_summary(); 250 251 free(worker); 252 perf_cpu_map__put(cpu); 253 return ret; 254 err: 255 usage_with_options(bench_futex_lock_pi_usage, options); 256 exit(EXIT_FAILURE); 257 } 258