1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (C) 2013 Davidlohr Bueso <davidlohr@hp.com> 4 * 5 * futex-hash: Stress the hell out of the Linux kernel futex uaddr hashing. 6 * 7 * This program is particularly useful for measuring the kernel's futex hash 8 * table/function implementation. In order for it to make sense, use with as 9 * many threads and futexes as possible. 10 */ 11 12 /* For the CLR_() macros */ 13 #include <string.h> 14 #include <pthread.h> 15 16 #include <errno.h> 17 #include <signal.h> 18 #include <stdlib.h> 19 #include <linux/compiler.h> 20 #include <linux/kernel.h> 21 #include <linux/zalloc.h> 22 #include <sys/time.h> 23 #include <sys/mman.h> 24 #include <sys/prctl.h> 25 #include <perf/cpumap.h> 26 27 #include "../util/mutex.h" 28 #include "../util/stat.h" 29 #include <subcmd/parse-options.h> 30 #include "bench.h" 31 #include "futex.h" 32 33 #include <err.h> 34 35 static bool done = false; 36 static int futex_flag = 0; 37 38 struct timeval bench__start, bench__end, bench__runtime; 39 static struct mutex thread_lock; 40 static unsigned int threads_starting; 41 static struct stats throughput_stats; 42 static struct cond thread_parent, thread_worker; 43 44 struct worker { 45 int tid; 46 u_int32_t *futex; 47 pthread_t thread; 48 unsigned long ops; 49 }; 50 51 static struct bench_futex_parameters params = { 52 .nfutexes = 1024, 53 .runtime = 10, 54 .nbuckets = -1, 55 }; 56 57 static const struct option options[] = { 58 OPT_INTEGER( 'b', "buckets", ¶ms.nbuckets, "Specify amount of hash buckets"), 59 OPT_BOOLEAN( 'I', "immutable", ¶ms.buckets_immutable, "Make the hash buckets immutable"), 60 OPT_UINTEGER('t', "threads", ¶ms.nthreads, "Specify amount of threads"), 61 OPT_UINTEGER('r', "runtime", ¶ms.runtime, "Specify runtime (in seconds)"), 62 OPT_UINTEGER('f', "futexes", ¶ms.nfutexes, "Specify amount of futexes per threads"), 63 OPT_BOOLEAN( 's', "silent", ¶ms.silent, "Silent mode: do not display data/details"), 64 OPT_BOOLEAN( 'S', "shared", ¶ms.fshared, "Use shared futexes instead of private ones"), 65 OPT_BOOLEAN( 'm', "mlockall", ¶ms.mlockall, "Lock all current and future memory"), 66 OPT_END() 67 }; 68 69 static const char * const bench_futex_hash_usage[] = { 70 "perf bench futex hash <options>", 71 NULL 72 }; 73 74 static void *workerfn(void *arg) 75 { 76 int ret; 77 struct worker *w = (struct worker *) arg; 78 unsigned int i; 79 unsigned long ops = w->ops; /* avoid cacheline bouncing */ 80 81 mutex_lock(&thread_lock); 82 threads_starting--; 83 if (!threads_starting) 84 cond_signal(&thread_parent); 85 cond_wait(&thread_worker, &thread_lock); 86 mutex_unlock(&thread_lock); 87 88 do { 89 for (i = 0; i < params.nfutexes; i++, ops++) { 90 /* 91 * We want the futex calls to fail in order to stress 92 * the hashing of uaddr and not measure other steps, 93 * such as internal waitqueue handling, thus enlarging 94 * the critical region protected by hb->lock. 95 */ 96 ret = futex_wait(&w->futex[i], 1234, NULL, futex_flag); 97 if (!params.silent && 98 (!ret || errno != EAGAIN || errno != EWOULDBLOCK)) 99 warn("Non-expected futex return call"); 100 } 101 } while (!done); 102 103 w->ops = ops; 104 return NULL; 105 } 106 107 static void toggle_done(int sig __maybe_unused, 108 siginfo_t *info __maybe_unused, 109 void *uc __maybe_unused) 110 { 111 /* inform all threads that we're done for the day */ 112 done = true; 113 gettimeofday(&bench__end, NULL); 114 timersub(&bench__end, &bench__start, &bench__runtime); 115 } 116 117 static void print_summary(void) 118 { 119 unsigned long avg = avg_stats(&throughput_stats); 120 double stddev = stddev_stats(&throughput_stats); 121 122 printf("%sAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n", 123 !params.silent ? "\n" : "", avg, rel_stddev_stats(stddev, avg), 124 (int)bench__runtime.tv_sec); 125 futex_print_nbuckets(¶ms); 126 } 127 128 int bench_futex_hash(int argc, const char **argv) 129 { 130 int ret = 0; 131 cpu_set_t *cpuset; 132 struct sigaction act; 133 unsigned int i; 134 pthread_attr_t thread_attr; 135 struct worker *worker = NULL; 136 struct perf_cpu_map *cpu; 137 int nrcpus; 138 size_t size; 139 140 argc = parse_options(argc, argv, options, bench_futex_hash_usage, 0); 141 if (argc) { 142 usage_with_options(bench_futex_hash_usage, options); 143 exit(EXIT_FAILURE); 144 } 145 146 cpu = perf_cpu_map__new_online_cpus(); 147 if (!cpu) 148 goto errmem; 149 150 memset(&act, 0, sizeof(act)); 151 sigfillset(&act.sa_mask); 152 act.sa_sigaction = toggle_done; 153 sigaction(SIGINT, &act, NULL); 154 155 if (params.mlockall) { 156 if (mlockall(MCL_CURRENT | MCL_FUTURE)) 157 err(EXIT_FAILURE, "mlockall"); 158 } 159 160 if (!params.nthreads) /* default to the number of CPUs */ 161 params.nthreads = perf_cpu_map__nr(cpu); 162 163 worker = calloc(params.nthreads, sizeof(*worker)); 164 if (!worker) 165 goto errmem; 166 167 if (!params.fshared) 168 futex_flag = FUTEX_PRIVATE_FLAG; 169 futex_set_nbuckets_param(¶ms); 170 171 printf("Run summary [PID %d]: %d threads, each operating on %d [%s] futexes for %d secs.\n\n", 172 getpid(), params.nthreads, params.nfutexes, params.fshared ? "shared":"private", params.runtime); 173 174 init_stats(&throughput_stats); 175 mutex_init(&thread_lock); 176 cond_init(&thread_parent); 177 cond_init(&thread_worker); 178 179 threads_starting = params.nthreads; 180 pthread_attr_init(&thread_attr); 181 gettimeofday(&bench__start, NULL); 182 183 nrcpus = cpu__max_cpu().cpu; 184 cpuset = CPU_ALLOC(nrcpus); 185 BUG_ON(!cpuset); 186 size = CPU_ALLOC_SIZE(nrcpus); 187 188 for (i = 0; i < params.nthreads; i++) { 189 worker[i].tid = i; 190 worker[i].futex = calloc(params.nfutexes, sizeof(*worker[i].futex)); 191 if (!worker[i].futex) 192 goto errmem; 193 194 CPU_ZERO_S(size, cpuset); 195 196 CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset); 197 ret = pthread_attr_setaffinity_np(&thread_attr, size, cpuset); 198 if (ret) { 199 CPU_FREE(cpuset); 200 err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); 201 } 202 ret = pthread_create(&worker[i].thread, &thread_attr, workerfn, 203 (void *)(struct worker *) &worker[i]); 204 if (ret) { 205 CPU_FREE(cpuset); 206 err(EXIT_FAILURE, "pthread_create"); 207 } 208 209 } 210 CPU_FREE(cpuset); 211 pthread_attr_destroy(&thread_attr); 212 213 mutex_lock(&thread_lock); 214 while (threads_starting) 215 cond_wait(&thread_parent, &thread_lock); 216 cond_broadcast(&thread_worker); 217 mutex_unlock(&thread_lock); 218 219 sleep(params.runtime); 220 toggle_done(0, NULL, NULL); 221 222 for (i = 0; i < params.nthreads; i++) { 223 ret = pthread_join(worker[i].thread, NULL); 224 if (ret) 225 err(EXIT_FAILURE, "pthread_join"); 226 } 227 228 /* cleanup & report results */ 229 cond_destroy(&thread_parent); 230 cond_destroy(&thread_worker); 231 mutex_destroy(&thread_lock); 232 233 for (i = 0; i < params.nthreads; i++) { 234 unsigned long t = bench__runtime.tv_sec > 0 ? 235 worker[i].ops / bench__runtime.tv_sec : 0; 236 update_stats(&throughput_stats, t); 237 if (!params.silent) { 238 if (params.nfutexes == 1) 239 printf("[thread %2d] futex: %p [ %ld ops/sec ]\n", 240 worker[i].tid, &worker[i].futex[0], t); 241 else 242 printf("[thread %2d] futexes: %p ... %p [ %ld ops/sec ]\n", 243 worker[i].tid, &worker[i].futex[0], 244 &worker[i].futex[params.nfutexes-1], t); 245 } 246 247 zfree(&worker[i].futex); 248 } 249 250 print_summary(); 251 252 free(worker); 253 free(cpu); 254 return ret; 255 errmem: 256 err(EXIT_FAILURE, "calloc"); 257 } 258