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 <perf/cpumap.h> 24 25 #include "../util/stat.h" 26 #include <subcmd/parse-options.h> 27 #include "bench.h" 28 #include "futex.h" 29 30 #include <err.h> 31 32 static unsigned int nthreads = 0; 33 static unsigned int nsecs = 10; 34 /* amount of futexes per thread */ 35 static unsigned int nfutexes = 1024; 36 static bool fshared = false, done = false, silent = false; 37 static int futex_flag = 0; 38 39 struct timeval bench__start, bench__end, bench__runtime; 40 static pthread_mutex_t thread_lock; 41 static unsigned int threads_starting; 42 static struct stats throughput_stats; 43 static pthread_cond_t thread_parent, thread_worker; 44 45 struct worker { 46 int tid; 47 u_int32_t *futex; 48 pthread_t thread; 49 unsigned long ops; 50 }; 51 52 static const struct option options[] = { 53 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"), 54 OPT_UINTEGER('r', "runtime", &nsecs, "Specify runtime (in seconds)"), 55 OPT_UINTEGER('f', "futexes", &nfutexes, "Specify amount of futexes per threads"), 56 OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"), 57 OPT_BOOLEAN( 'S', "shared", &fshared, "Use shared futexes instead of private ones"), 58 OPT_END() 59 }; 60 61 static const char * const bench_futex_hash_usage[] = { 62 "perf bench futex hash <options>", 63 NULL 64 }; 65 66 static void *workerfn(void *arg) 67 { 68 int ret; 69 struct worker *w = (struct worker *) arg; 70 unsigned int i; 71 unsigned long ops = w->ops; /* avoid cacheline bouncing */ 72 73 pthread_mutex_lock(&thread_lock); 74 threads_starting--; 75 if (!threads_starting) 76 pthread_cond_signal(&thread_parent); 77 pthread_cond_wait(&thread_worker, &thread_lock); 78 pthread_mutex_unlock(&thread_lock); 79 80 do { 81 for (i = 0; i < nfutexes; i++, ops++) { 82 /* 83 * We want the futex calls to fail in order to stress 84 * the hashing of uaddr and not measure other steps, 85 * such as internal waitqueue handling, thus enlarging 86 * the critical region protected by hb->lock. 87 */ 88 ret = futex_wait(&w->futex[i], 1234, NULL, futex_flag); 89 if (!silent && 90 (!ret || errno != EAGAIN || errno != EWOULDBLOCK)) 91 warn("Non-expected futex return call"); 92 } 93 } while (!done); 94 95 w->ops = ops; 96 return NULL; 97 } 98 99 static void toggle_done(int sig __maybe_unused, 100 siginfo_t *info __maybe_unused, 101 void *uc __maybe_unused) 102 { 103 /* inform all threads that we're done for the day */ 104 done = true; 105 gettimeofday(&bench__end, NULL); 106 timersub(&bench__end, &bench__start, &bench__runtime); 107 } 108 109 static void print_summary(void) 110 { 111 unsigned long avg = avg_stats(&throughput_stats); 112 double stddev = stddev_stats(&throughput_stats); 113 114 printf("%sAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n", 115 !silent ? "\n" : "", avg, rel_stddev_stats(stddev, avg), 116 (int)bench__runtime.tv_sec); 117 } 118 119 int bench_futex_hash(int argc, const char **argv) 120 { 121 int ret = 0; 122 cpu_set_t cpuset; 123 struct sigaction act; 124 unsigned int i; 125 pthread_attr_t thread_attr; 126 struct worker *worker = NULL; 127 struct perf_cpu_map *cpu; 128 129 argc = parse_options(argc, argv, options, bench_futex_hash_usage, 0); 130 if (argc) { 131 usage_with_options(bench_futex_hash_usage, options); 132 exit(EXIT_FAILURE); 133 } 134 135 cpu = perf_cpu_map__new(NULL); 136 if (!cpu) 137 goto errmem; 138 139 memset(&act, 0, sizeof(act)); 140 sigfillset(&act.sa_mask); 141 act.sa_sigaction = toggle_done; 142 sigaction(SIGINT, &act, NULL); 143 144 if (!nthreads) /* default to the number of CPUs */ 145 nthreads = cpu->nr; 146 147 worker = calloc(nthreads, sizeof(*worker)); 148 if (!worker) 149 goto errmem; 150 151 if (!fshared) 152 futex_flag = FUTEX_PRIVATE_FLAG; 153 154 printf("Run summary [PID %d]: %d threads, each operating on %d [%s] futexes for %d secs.\n\n", 155 getpid(), nthreads, nfutexes, fshared ? "shared":"private", nsecs); 156 157 init_stats(&throughput_stats); 158 pthread_mutex_init(&thread_lock, NULL); 159 pthread_cond_init(&thread_parent, NULL); 160 pthread_cond_init(&thread_worker, NULL); 161 162 threads_starting = nthreads; 163 pthread_attr_init(&thread_attr); 164 gettimeofday(&bench__start, NULL); 165 for (i = 0; i < nthreads; i++) { 166 worker[i].tid = i; 167 worker[i].futex = calloc(nfutexes, sizeof(*worker[i].futex)); 168 if (!worker[i].futex) 169 goto errmem; 170 171 CPU_ZERO(&cpuset); 172 CPU_SET(cpu->map[i % cpu->nr], &cpuset); 173 174 ret = pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset); 175 if (ret) 176 err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); 177 178 ret = pthread_create(&worker[i].thread, &thread_attr, workerfn, 179 (void *)(struct worker *) &worker[i]); 180 if (ret) 181 err(EXIT_FAILURE, "pthread_create"); 182 183 } 184 pthread_attr_destroy(&thread_attr); 185 186 pthread_mutex_lock(&thread_lock); 187 while (threads_starting) 188 pthread_cond_wait(&thread_parent, &thread_lock); 189 pthread_cond_broadcast(&thread_worker); 190 pthread_mutex_unlock(&thread_lock); 191 192 sleep(nsecs); 193 toggle_done(0, NULL, NULL); 194 195 for (i = 0; i < nthreads; i++) { 196 ret = pthread_join(worker[i].thread, NULL); 197 if (ret) 198 err(EXIT_FAILURE, "pthread_join"); 199 } 200 201 /* cleanup & report results */ 202 pthread_cond_destroy(&thread_parent); 203 pthread_cond_destroy(&thread_worker); 204 pthread_mutex_destroy(&thread_lock); 205 206 for (i = 0; i < nthreads; i++) { 207 unsigned long t = bench__runtime.tv_sec > 0 ? 208 worker[i].ops / bench__runtime.tv_sec : 0; 209 update_stats(&throughput_stats, t); 210 if (!silent) { 211 if (nfutexes == 1) 212 printf("[thread %2d] futex: %p [ %ld ops/sec ]\n", 213 worker[i].tid, &worker[i].futex[0], t); 214 else 215 printf("[thread %2d] futexes: %p ... %p [ %ld ops/sec ]\n", 216 worker[i].tid, &worker[i].futex[0], 217 &worker[i].futex[nfutexes-1], t); 218 } 219 220 zfree(&worker[i].futex); 221 } 222 223 print_summary(); 224 225 free(worker); 226 free(cpu); 227 return ret; 228 errmem: 229 err(EXIT_FAILURE, "calloc"); 230 } 231