tools/perf/bench/futex-wake.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2013  Davidlohr Bueso <davidlohr@hp.com>
   4  *
   5  * futex-wake: Block a bunch of threads on a futex and wake'em up, N at a time.
   6  *
   7  * This program is particularly useful to measure the latency of nthread wakeups
   8  * in non-error situations:  all waiters are queued and all wake calls wakeup
   9  * one or more tasks, and thus the waitqueue is never empty.
  10  */
  11
  12 /* For the CLR_() macros */
  13 #include <string.h>
  14 #include <pthread.h>
  15
  16 #include <signal.h>
  17 #include "../util/stat.h"
  18 #include <subcmd/parse-options.h>
  19 #include <linux/compiler.h>
  20 #include <linux/kernel.h>
  21 #include <linux/time64.h>
  22 #include <errno.h>
  23 #include <perf/cpumap.h>
  24 #include "bench.h"
  25 #include "futex.h"
  26
  27 #include <err.h>
  28 #include <stdlib.h>
  29 #include <sys/time.h>
  30 #include <sys/mman.h>
  31
  32 /* all threads will block on the same futex */
  33 static u_int32_t futex1 = 0;
  34
  35 static pthread_t *worker;
  36 static bool done = false;
  37 static pthread_mutex_t thread_lock;
  38 static pthread_cond_t thread_parent, thread_worker;
  39 static struct stats waketime_stats, wakeup_stats;
  40 static unsigned int threads_starting;
  41 static int futex_flag = 0;
  42
  43 static struct bench_futex_parameters params = {
  44         /*
  45          * How many wakeups to do at a time.
  46          * Default to 1 in order to make the kernel work more.
  47          */
  48         .nwakes  = 1,
  49 };
  50
  51 static const struct option options[] = {
  52         OPT_UINTEGER('t', "threads", &params.nthreads, "Specify amount of threads"),
  53         OPT_UINTEGER('w', "nwakes",  &params.nwakes, "Specify amount of threads to wake at once"),
  54         OPT_BOOLEAN( 's', "silent",  &params.silent, "Silent mode: do not display data/details"),
  55         OPT_BOOLEAN( 'S', "shared",  &params.fshared, "Use shared futexes instead of private ones"),
  56         OPT_BOOLEAN( 'm', "mlockall", &params.mlockall, "Lock all current and future memory"),
  57
  58         OPT_END()
  59 };
  60
  61 static const char * const bench_futex_wake_usage[] = {
  62         "perf bench futex wake <options>",
  63         NULL
  64 };
  65
  66 static void *workerfn(void *arg __maybe_unused)
  67 {
  68         pthread_mutex_lock(&thread_lock);
  69         threads_starting--;
  70         if (!threads_starting)
  71                 pthread_cond_signal(&thread_parent);
  72         pthread_cond_wait(&thread_worker, &thread_lock);
  73         pthread_mutex_unlock(&thread_lock);
  74
  75         while (1) {
  76                 if (futex_wait(&futex1, 0, NULL, futex_flag) != EINTR)
  77                         break;
  78         }
  79
  80         pthread_exit(NULL);
  81         return NULL;
  82 }
  83
  84 static void print_summary(void)
  85 {
  86         double waketime_avg = avg_stats(&waketime_stats);
  87         double waketime_stddev = stddev_stats(&waketime_stats);
  88         unsigned int wakeup_avg = avg_stats(&wakeup_stats);
  89
  90         printf("Wokeup %d of %d threads in %.4f ms (+-%.2f%%)\n",
  91                wakeup_avg,
  92                params.nthreads,
  93                waketime_avg / USEC_PER_MSEC,
  94                rel_stddev_stats(waketime_stddev, waketime_avg));
  95 }
  96
  97 static void block_threads(pthread_t *w,
  98                           pthread_attr_t thread_attr, struct perf_cpu_map *cpu)
  99 {
 100         cpu_set_t cpuset;
 101         unsigned int i;
 102
 103         threads_starting = params.nthreads;
 104
 105         /* create and block all threads */
 106         for (i = 0; i < params.nthreads; i++) {
 107                 CPU_ZERO(&cpuset);
 108                 CPU_SET(cpu->map[i % cpu->nr], &cpuset);
 109
 110                 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset))
 111                         err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
 112
 113                 if (pthread_create(&w[i], &thread_attr, workerfn, NULL))
 114                         err(EXIT_FAILURE, "pthread_create");
 115         }
 116 }
 117
 118 static void toggle_done(int sig __maybe_unused,
 119                         siginfo_t *info __maybe_unused,
 120                         void *uc __maybe_unused)
 121 {
 122         done = true;
 123 }
 124
 125 int bench_futex_wake(int argc, const char **argv)
 126 {
 127         int ret = 0;
 128         unsigned int i, j;
 129         struct sigaction act;
 130         pthread_attr_t thread_attr;
 131         struct perf_cpu_map *cpu;
 132
 133         argc = parse_options(argc, argv, options, bench_futex_wake_usage, 0);
 134         if (argc) {
 135                 usage_with_options(bench_futex_wake_usage, options);
 136                 exit(EXIT_FAILURE);
 137         }
 138
 139         cpu = perf_cpu_map__new(NULL);
 140         if (!cpu)
 141                 err(EXIT_FAILURE, "calloc");
 142
 143         memset(&act, 0, sizeof(act));
 144         sigfillset(&act.sa_mask);
 145         act.sa_sigaction = toggle_done;
 146         sigaction(SIGINT, &act, NULL);
 147
 148         if (params.mlockall) {
 149                 if (mlockall(MCL_CURRENT | MCL_FUTURE))
 150                         err(EXIT_FAILURE, "mlockall");
 151         }
 152
 153         if (!params.nthreads)
 154                 params.nthreads = cpu->nr;
 155
 156         worker = calloc(params.nthreads, sizeof(*worker));
 157         if (!worker)
 158                 err(EXIT_FAILURE, "calloc");
 159
 160         if (!params.fshared)
 161                 futex_flag = FUTEX_PRIVATE_FLAG;
 162
 163         printf("Run summary [PID %d]: blocking on %d threads (at [%s] futex %p), "
 164                "waking up %d at a time.\n\n",
 165                getpid(), params.nthreads, params.fshared ? "shared":"private",
 166                &futex1, params.nwakes);
 167
 168         init_stats(&wakeup_stats);
 169         init_stats(&waketime_stats);
 170         pthread_attr_init(&thread_attr);
 171         pthread_mutex_init(&thread_lock, NULL);
 172         pthread_cond_init(&thread_parent, NULL);
 173         pthread_cond_init(&thread_worker, NULL);
 174
 175         for (j = 0; j < bench_repeat && !done; j++) {
 176                 unsigned int nwoken = 0;
 177                 struct timeval start, end, runtime;
 178
 179                 /* create, launch & block all threads */
 180                 block_threads(worker, thread_attr, cpu);
 181
 182                 /* make sure all threads are already blocked */
 183                 pthread_mutex_lock(&thread_lock);
 184                 while (threads_starting)
 185                         pthread_cond_wait(&thread_parent, &thread_lock);
 186                 pthread_cond_broadcast(&thread_worker);
 187                 pthread_mutex_unlock(&thread_lock);
 188
 189                 usleep(100000);
 190
 191                 /* Ok, all threads are patiently blocked, start waking folks up */
 192                 gettimeofday(&start, NULL);
 193                 while (nwoken != params.nthreads)
 194                         nwoken += futex_wake(&futex1,
 195                                              params.nwakes, futex_flag);
 196                 gettimeofday(&end, NULL);
 197                 timersub(&end, &start, &runtime);
 198
 199                 update_stats(&wakeup_stats, nwoken);
 200                 update_stats(&waketime_stats, runtime.tv_usec);
 201
 202                 if (!params.silent) {
 203                         printf("[Run %d]: Wokeup %d of %d threads in %.4f ms\n",
 204                                j + 1, nwoken, params.nthreads,
 205                                runtime.tv_usec / (double)USEC_PER_MSEC);
 206                 }
 207
 208                 for (i = 0; i < params.nthreads; i++) {
 209                         ret = pthread_join(worker[i], NULL);
 210                         if (ret)
 211                                 err(EXIT_FAILURE, "pthread_join");
 212                 }
 213
 214         }
 215
 216         /* cleanup & report results */
 217         pthread_cond_destroy(&thread_parent);
 218         pthread_cond_destroy(&thread_worker);
 219         pthread_mutex_destroy(&thread_lock);
 220         pthread_attr_destroy(&thread_attr);
 221
 222         print_summary();
 223
 224         free(worker);
 225         perf_cpu_map__put(cpu);
 226         return ret;
 227 }