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_BOOLEAN( 'I', "immutable", ¶ms.buckets_immutable, "Make the hash buckets immutable"), 51 OPT_UINTEGER('t', "threads", ¶ms.nthreads, "Specify amount of threads"), 52 OPT_UINTEGER('r', "runtime", ¶ms.runtime, "Specify runtime (in seconds)"), 53 OPT_BOOLEAN( 'M', "multi", ¶ms.multi, "Use multiple futexes"), 54 OPT_BOOLEAN( 's', "silent", ¶ms.silent, "Silent mode: do not display data/details"), 55 OPT_BOOLEAN( 'S', "shared", ¶ms.fshared, "Use shared futexes instead of private ones"), 56 OPT_BOOLEAN( 'm', "mlockall", ¶ms.mlockall, "Lock all current and future memory"), 57 OPT_END() 58 }; 59 60 static const char * const bench_futex_lock_pi_usage[] = { 61 "perf bench futex lock-pi <options>", 62 NULL 63 }; 64 65 static void print_summary(void) 66 { 67 unsigned long avg = avg_stats(&throughput_stats); 68 double stddev = stddev_stats(&throughput_stats); 69 70 printf("%sAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n", 71 !params.silent ? "\n" : "", avg, rel_stddev_stats(stddev, avg), 72 (int)bench__runtime.tv_sec); 73 futex_print_nbuckets(¶ms); 74 } 75 76 static void toggle_done(int sig __maybe_unused, 77 siginfo_t *info __maybe_unused, 78 void *uc __maybe_unused) 79 { 80 /* inform all threads that we're done for the day */ 81 done = true; 82 gettimeofday(&bench__end, NULL); 83 timersub(&bench__end, &bench__start, &bench__runtime); 84 } 85 86 static void *workerfn(void *arg) 87 { 88 struct worker *w = (struct worker *) arg; 89 unsigned long ops = w->ops; 90 91 mutex_lock(&thread_lock); 92 threads_starting--; 93 if (!threads_starting) 94 cond_signal(&thread_parent); 95 cond_wait(&thread_worker, &thread_lock); 96 mutex_unlock(&thread_lock); 97 98 do { 99 int ret; 100 again: 101 ret = futex_lock_pi(w->futex, NULL, futex_flag); 102 103 if (ret) { /* handle lock acquisition */ 104 if (!params.silent) 105 warn("thread %d: Could not lock pi-lock for %p (%d)", 106 w->tid, w->futex, ret); 107 if (done) 108 break; 109 110 goto again; 111 } 112 113 usleep(1); 114 ret = futex_unlock_pi(w->futex, futex_flag); 115 if (ret && !params.silent) 116 warn("thread %d: Could not unlock pi-lock for %p (%d)", 117 w->tid, w->futex, ret); 118 ops++; /* account for thread's share of work */ 119 } while (!done); 120 121 w->ops = ops; 122 return NULL; 123 } 124 125 static void create_threads(struct worker *w, struct perf_cpu_map *cpu) 126 { 127 cpu_set_t *cpuset; 128 unsigned int i; 129 int nrcpus = cpu__max_cpu().cpu; 130 size_t size; 131 132 threads_starting = params.nthreads; 133 134 cpuset = CPU_ALLOC(nrcpus); 135 BUG_ON(!cpuset); 136 size = CPU_ALLOC_SIZE(nrcpus); 137 138 for (i = 0; i < params.nthreads; i++) { 139 pthread_attr_t thread_attr; 140 141 pthread_attr_init(&thread_attr); 142 worker[i].tid = i; 143 144 if (params.multi) { 145 worker[i].futex = calloc(1, sizeof(u_int32_t)); 146 if (!worker[i].futex) 147 err(EXIT_FAILURE, "calloc"); 148 } else 149 worker[i].futex = &global_futex; 150 151 CPU_ZERO_S(size, cpuset); 152 CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset); 153 154 if (pthread_attr_setaffinity_np(&thread_attr, size, cpuset)) { 155 CPU_FREE(cpuset); 156 err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); 157 } 158 159 if (pthread_create(&w[i].thread, &thread_attr, workerfn, &worker[i])) { 160 CPU_FREE(cpuset); 161 err(EXIT_FAILURE, "pthread_create"); 162 } 163 pthread_attr_destroy(&thread_attr); 164 } 165 CPU_FREE(cpuset); 166 } 167 168 int bench_futex_lock_pi(int argc, const char **argv) 169 { 170 int ret = 0; 171 unsigned int i; 172 struct sigaction act; 173 struct perf_cpu_map *cpu; 174 175 argc = parse_options(argc, argv, options, bench_futex_lock_pi_usage, 0); 176 if (argc) 177 goto err; 178 179 cpu = perf_cpu_map__new_online_cpus(); 180 if (!cpu) 181 err(EXIT_FAILURE, "calloc"); 182 183 memset(&act, 0, sizeof(act)); 184 sigfillset(&act.sa_mask); 185 act.sa_sigaction = toggle_done; 186 sigaction(SIGINT, &act, NULL); 187 188 if (params.mlockall) { 189 if (mlockall(MCL_CURRENT | MCL_FUTURE)) 190 err(EXIT_FAILURE, "mlockall"); 191 } 192 193 if (!params.nthreads) 194 params.nthreads = perf_cpu_map__nr(cpu); 195 196 worker = calloc(params.nthreads, sizeof(*worker)); 197 if (!worker) 198 err(EXIT_FAILURE, "calloc"); 199 200 if (!params.fshared) 201 futex_flag = FUTEX_PRIVATE_FLAG; 202 203 printf("Run summary [PID %d]: %d threads doing pi lock/unlock pairing for %d secs.\n\n", 204 getpid(), params.nthreads, params.runtime); 205 206 init_stats(&throughput_stats); 207 mutex_init(&thread_lock); 208 cond_init(&thread_parent); 209 cond_init(&thread_worker); 210 futex_set_nbuckets_param(¶ms); 211 212 threads_starting = params.nthreads; 213 gettimeofday(&bench__start, NULL); 214 215 create_threads(worker, cpu); 216 217 mutex_lock(&thread_lock); 218 while (threads_starting) 219 cond_wait(&thread_parent, &thread_lock); 220 cond_broadcast(&thread_worker); 221 mutex_unlock(&thread_lock); 222 223 sleep(params.runtime); 224 toggle_done(0, NULL, NULL); 225 226 for (i = 0; i < params.nthreads; i++) { 227 ret = pthread_join(worker[i].thread, NULL); 228 if (ret) 229 err(EXIT_FAILURE, "pthread_join"); 230 } 231 232 /* cleanup & report results */ 233 cond_destroy(&thread_parent); 234 cond_destroy(&thread_worker); 235 mutex_destroy(&thread_lock); 236 237 for (i = 0; i < params.nthreads; i++) { 238 unsigned long t = bench__runtime.tv_sec > 0 ? 239 worker[i].ops / bench__runtime.tv_sec : 0; 240 241 update_stats(&throughput_stats, t); 242 if (!params.silent) 243 printf("[thread %3d] futex: %p [ %ld ops/sec ]\n", 244 worker[i].tid, worker[i].futex, t); 245 246 if (params.multi) 247 zfree(&worker[i].futex); 248 } 249 250 print_summary(); 251 252 free(worker); 253 perf_cpu_map__put(cpu); 254 return ret; 255 err: 256 usage_with_options(bench_futex_lock_pi_usage, options); 257 exit(EXIT_FAILURE); 258 } 259