1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (C) 2013 Davidlohr Bueso <davidlohr@hp.com> 4 * 5 * futex-requeue: Block a bunch of threads on futex1 and requeue them 6 * on futex2, N at a time. 7 * 8 * This program is particularly useful to measure the latency of nthread 9 * requeues without waking up any tasks (in the non-pi case) -- thus 10 * mimicking a regular futex_wait. 11 */ 12 13 /* For the CLR_() macros */ 14 #include <string.h> 15 #include <pthread.h> 16 17 #include <signal.h> 18 #include "../util/mutex.h" 19 #include "../util/stat.h" 20 #include <subcmd/parse-options.h> 21 #include <linux/compiler.h> 22 #include <linux/kernel.h> 23 #include <linux/time64.h> 24 #include <errno.h> 25 #include <perf/cpumap.h> 26 #include "bench.h" 27 #include "futex.h" 28 29 #include <err.h> 30 #include <stdlib.h> 31 #include <sys/time.h> 32 #include <sys/mman.h> 33 34 static u_int32_t futex1 = 0, futex2 = 0; 35 36 static pthread_t *worker; 37 static bool done = false; 38 static struct mutex thread_lock; 39 static struct cond thread_parent, thread_worker; 40 static struct stats requeuetime_stats, requeued_stats; 41 static unsigned int threads_starting; 42 static int futex_flag = 0; 43 44 static struct bench_futex_parameters params = { 45 .nbuckets = -1, 46 /* 47 * How many tasks to requeue at a time. 48 * Default to 1 in order to make the kernel work more. 49 */ 50 .nrequeue = 1, 51 }; 52 53 static const struct option options[] = { 54 OPT_INTEGER( 'b', "buckets", ¶ms.nbuckets, "Specify amount of hash buckets"), 55 OPT_BOOLEAN( 'I', "immutable", ¶ms.buckets_immutable, "Make the hash buckets immutable"), 56 OPT_UINTEGER('t', "threads", ¶ms.nthreads, "Specify amount of threads"), 57 OPT_UINTEGER('q', "nrequeue", ¶ms.nrequeue, "Specify amount of threads to requeue at once"), 58 OPT_BOOLEAN( 's', "silent", ¶ms.silent, "Silent mode: do not display data/details"), 59 OPT_BOOLEAN( 'S', "shared", ¶ms.fshared, "Use shared futexes instead of private ones"), 60 OPT_BOOLEAN( 'm', "mlockall", ¶ms.mlockall, "Lock all current and future memory"), 61 OPT_BOOLEAN( 'B', "broadcast", ¶ms.broadcast, "Requeue all threads at once"), 62 OPT_BOOLEAN( 'p', "pi", ¶ms.pi, "Use PI-aware variants of FUTEX_CMP_REQUEUE"), 63 64 OPT_END() 65 }; 66 67 static const char * const bench_futex_requeue_usage[] = { 68 "perf bench futex requeue <options>", 69 NULL 70 }; 71 72 static void print_summary(void) 73 { 74 double requeuetime_avg = avg_stats(&requeuetime_stats); 75 double requeuetime_stddev = stddev_stats(&requeuetime_stats); 76 unsigned int requeued_avg = avg_stats(&requeued_stats); 77 78 printf("Requeued %d of %d threads in %.4f ms (+-%.2f%%)\n", 79 requeued_avg, 80 params.nthreads, 81 requeuetime_avg / USEC_PER_MSEC, 82 rel_stddev_stats(requeuetime_stddev, requeuetime_avg)); 83 futex_print_nbuckets(¶ms); 84 } 85 86 static void *workerfn(void *arg __maybe_unused) 87 { 88 int ret; 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 while (1) { 98 if (!params.pi) { 99 ret = futex_wait(&futex1, 0, NULL, futex_flag); 100 if (!ret) 101 break; 102 103 if (ret && errno != EAGAIN) { 104 if (!params.silent) 105 warnx("futex_wait"); 106 break; 107 } 108 } else { 109 ret = futex_wait_requeue_pi(&futex1, 0, &futex2, 110 NULL, futex_flag); 111 if (!ret) { 112 /* got the lock at futex2 */ 113 futex_unlock_pi(&futex2, futex_flag); 114 break; 115 } 116 117 if (ret && errno != EAGAIN) { 118 if (!params.silent) 119 warnx("futex_wait_requeue_pi"); 120 break; 121 } 122 } 123 } 124 125 return NULL; 126 } 127 128 static void block_threads(pthread_t *w, struct perf_cpu_map *cpu) 129 { 130 cpu_set_t *cpuset; 131 unsigned int i; 132 int nrcpus = cpu__max_cpu().cpu; 133 size_t size; 134 135 threads_starting = params.nthreads; 136 137 cpuset = CPU_ALLOC(nrcpus); 138 BUG_ON(!cpuset); 139 size = CPU_ALLOC_SIZE(nrcpus); 140 141 /* create and block all threads */ 142 for (i = 0; i < params.nthreads; i++) { 143 pthread_attr_t thread_attr; 144 145 pthread_attr_init(&thread_attr); 146 CPU_ZERO_S(size, cpuset); 147 CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset); 148 149 if (pthread_attr_setaffinity_np(&thread_attr, size, cpuset)) { 150 CPU_FREE(cpuset); 151 err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); 152 } 153 154 if (pthread_create(&w[i], &thread_attr, workerfn, NULL)) { 155 CPU_FREE(cpuset); 156 err(EXIT_FAILURE, "pthread_create"); 157 } 158 pthread_attr_destroy(&thread_attr); 159 } 160 CPU_FREE(cpuset); 161 } 162 163 static void toggle_done(int sig __maybe_unused, 164 siginfo_t *info __maybe_unused, 165 void *uc __maybe_unused) 166 { 167 done = true; 168 } 169 170 int bench_futex_requeue(int argc, const char **argv) 171 { 172 int ret = 0; 173 unsigned int i, j; 174 struct sigaction act; 175 struct perf_cpu_map *cpu; 176 177 argc = parse_options(argc, argv, options, bench_futex_requeue_usage, 0); 178 if (argc) 179 goto err; 180 181 cpu = perf_cpu_map__new_online_cpus(); 182 if (!cpu) 183 err(EXIT_FAILURE, "cpu_map__new"); 184 185 memset(&act, 0, sizeof(act)); 186 sigfillset(&act.sa_mask); 187 act.sa_sigaction = toggle_done; 188 sigaction(SIGINT, &act, NULL); 189 190 if (params.mlockall) { 191 if (mlockall(MCL_CURRENT | MCL_FUTURE)) 192 err(EXIT_FAILURE, "mlockall"); 193 } 194 195 if (!params.nthreads) 196 params.nthreads = perf_cpu_map__nr(cpu); 197 198 worker = calloc(params.nthreads, sizeof(*worker)); 199 if (!worker) 200 err(EXIT_FAILURE, "calloc"); 201 202 if (!params.fshared) 203 futex_flag = FUTEX_PRIVATE_FLAG; 204 205 if (params.nrequeue > params.nthreads) 206 params.nrequeue = params.nthreads; 207 208 if (params.broadcast) 209 params.nrequeue = params.nthreads; 210 211 futex_set_nbuckets_param(¶ms); 212 213 printf("Run summary [PID %d]: Requeuing %d threads (from [%s] %p to %s%p), " 214 "%d at a time.\n\n", getpid(), params.nthreads, 215 params.fshared ? "shared":"private", &futex1, 216 params.pi ? "PI ": "", &futex2, params.nrequeue); 217 218 init_stats(&requeued_stats); 219 init_stats(&requeuetime_stats); 220 mutex_init(&thread_lock); 221 cond_init(&thread_parent); 222 cond_init(&thread_worker); 223 224 for (j = 0; j < bench_repeat && !done; j++) { 225 unsigned int nrequeued = 0, wakeups = 0; 226 struct timeval start, end, runtime; 227 228 /* create, launch & block all threads */ 229 block_threads(worker, cpu); 230 231 /* make sure all threads are already blocked */ 232 mutex_lock(&thread_lock); 233 while (threads_starting) 234 cond_wait(&thread_parent, &thread_lock); 235 cond_broadcast(&thread_worker); 236 mutex_unlock(&thread_lock); 237 238 usleep(100000); 239 240 /* Ok, all threads are patiently blocked, start requeueing */ 241 gettimeofday(&start, NULL); 242 while (nrequeued < params.nthreads) { 243 int r; 244 245 /* 246 * For the regular non-pi case, do not wakeup any tasks 247 * blocked on futex1, allowing us to really measure 248 * futex_wait functionality. For the PI case the first 249 * waiter is always awoken. 250 */ 251 if (!params.pi) { 252 r = futex_cmp_requeue(&futex1, 0, &futex2, 0, 253 params.nrequeue, 254 futex_flag); 255 } else { 256 r = futex_cmp_requeue_pi(&futex1, 0, &futex2, 257 params.nrequeue, 258 futex_flag); 259 wakeups++; /* assume no error */ 260 } 261 262 if (r < 0) 263 err(EXIT_FAILURE, "couldn't requeue from %p to %p", 264 &futex1, &futex2); 265 266 nrequeued += r; 267 } 268 269 gettimeofday(&end, NULL); 270 timersub(&end, &start, &runtime); 271 272 update_stats(&requeued_stats, nrequeued); 273 update_stats(&requeuetime_stats, runtime.tv_usec); 274 275 if (!params.silent) { 276 if (!params.pi) 277 printf("[Run %d]: Requeued %d of %d threads in " 278 "%.4f ms\n", j + 1, nrequeued, 279 params.nthreads, 280 runtime.tv_usec / (double)USEC_PER_MSEC); 281 else { 282 nrequeued -= wakeups; 283 printf("[Run %d]: Awoke and Requeued (%d+%d) of " 284 "%d threads in %.4f ms\n", 285 j + 1, wakeups, nrequeued, 286 params.nthreads, 287 runtime.tv_usec / (double)USEC_PER_MSEC); 288 } 289 290 } 291 292 if (!params.pi) { 293 /* everybody should be blocked on futex2, wake'em up */ 294 nrequeued = futex_wake(&futex2, nrequeued, futex_flag); 295 if (params.nthreads != nrequeued) 296 warnx("couldn't wakeup all tasks (%d/%d)", 297 nrequeued, params.nthreads); 298 } 299 300 for (i = 0; i < params.nthreads; i++) { 301 ret = pthread_join(worker[i], NULL); 302 if (ret) 303 err(EXIT_FAILURE, "pthread_join"); 304 } 305 } 306 307 /* cleanup & report results */ 308 cond_destroy(&thread_parent); 309 cond_destroy(&thread_worker); 310 mutex_destroy(&thread_lock); 311 312 print_summary(); 313 314 free(worker); 315 perf_cpu_map__put(cpu); 316 return ret; 317 err: 318 usage_with_options(bench_futex_requeue_usage, options); 319 exit(EXIT_FAILURE); 320 } 321