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 "bench.h"
  24 #include "futex.h"
  25
  26 #include <err.h>
  27 #include <stdlib.h>
  28 #include <sys/time.h>
  29
  30 /* all threads will block on the same futex */
  31 static u_int32_t futex1 = 0;
  32
  33 /*
  34  * How many wakeups to do at a time.
  35  * Default to 1 in order to make the kernel work more.
  36  */
  37 static unsigned int nwakes = 1;
  38
  39 pthread_t *worker;
  40 static bool done = false, silent = false, fshared = false;
  41 static pthread_mutex_t thread_lock;
  42 static pthread_cond_t thread_parent, thread_worker;
  43 static struct stats waketime_stats, wakeup_stats;
  44 static unsigned int ncpus, threads_starting, nthreads = 0;
  45 static int futex_flag = 0;
  46
  47 static const struct option options[] = {
  48         OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
  49         OPT_UINTEGER('w', "nwakes",  &nwakes,   "Specify amount of threads to wake at once"),
  50         OPT_BOOLEAN( 's', "silent",  &silent,   "Silent mode: do not display data/details"),
  51         OPT_BOOLEAN( 'S', "shared",  &fshared,  "Use shared futexes instead of private ones"),
  52         OPT_END()
  53 };
  54
  55 static const char * const bench_futex_wake_usage[] = {
  56         "perf bench futex wake <options>",
  57         NULL
  58 };
  59
  60 static void *workerfn(void *arg __maybe_unused)
  61 {
  62         pthread_mutex_lock(&thread_lock);
  63         threads_starting--;
  64         if (!threads_starting)
  65                 pthread_cond_signal(&thread_parent);
  66         pthread_cond_wait(&thread_worker, &thread_lock);
  67         pthread_mutex_unlock(&thread_lock);
  68
  69         while (1) {
  70                 if (futex_wait(&futex1, 0, NULL, futex_flag) != EINTR)
  71                         break;
  72         }
  73
  74         pthread_exit(NULL);
  75         return NULL;
  76 }
  77
  78 static void print_summary(void)
  79 {
  80         double waketime_avg = avg_stats(&waketime_stats);
  81         double waketime_stddev = stddev_stats(&waketime_stats);
  82         unsigned int wakeup_avg = avg_stats(&wakeup_stats);
  83
  84         printf("Wokeup %d of %d threads in %.4f ms (+-%.2f%%)\n",
  85                wakeup_avg,
  86                nthreads,
  87                waketime_avg / USEC_PER_MSEC,
  88                rel_stddev_stats(waketime_stddev, waketime_avg));
  89 }
  90
  91 static void block_threads(pthread_t *w,
  92                           pthread_attr_t thread_attr)
  93 {
  94         cpu_set_t cpu;
  95         unsigned int i;
  96
  97         threads_starting = nthreads;
  98
  99         /* create and block all threads */
 100         for (i = 0; i < nthreads; i++) {
 101                 CPU_ZERO(&cpu);
 102                 CPU_SET(i % ncpus, &cpu);
 103
 104                 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu))
 105                         err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
 106
 107                 if (pthread_create(&w[i], &thread_attr, workerfn, NULL))
 108                         err(EXIT_FAILURE, "pthread_create");
 109         }
 110 }
 111
 112 static void toggle_done(int sig __maybe_unused,
 113                         siginfo_t *info __maybe_unused,
 114                         void *uc __maybe_unused)
 115 {
 116         done = true;
 117 }
 118
 119 int bench_futex_wake(int argc, const char **argv)
 120 {
 121         int ret = 0;
 122         unsigned int i, j;
 123         struct sigaction act;
 124         pthread_attr_t thread_attr;
 125
 126         argc = parse_options(argc, argv, options, bench_futex_wake_usage, 0);
 127         if (argc) {
 128                 usage_with_options(bench_futex_wake_usage, options);
 129                 exit(EXIT_FAILURE);
 130         }
 131
 132         ncpus = sysconf(_SC_NPROCESSORS_ONLN);
 133
 134         sigfillset(&act.sa_mask);
 135         act.sa_sigaction = toggle_done;
 136         sigaction(SIGINT, &act, NULL);
 137
 138         if (!nthreads)
 139                 nthreads = ncpus;
 140
 141         worker = calloc(nthreads, sizeof(*worker));
 142         if (!worker)
 143                 err(EXIT_FAILURE, "calloc");
 144
 145         if (!fshared)
 146                 futex_flag = FUTEX_PRIVATE_FLAG;
 147
 148         printf("Run summary [PID %d]: blocking on %d threads (at [%s] futex %p), "
 149                "waking up %d at a time.\n\n",
 150                getpid(), nthreads, fshared ? "shared":"private",  &futex1, nwakes);
 151
 152         init_stats(&wakeup_stats);
 153         init_stats(&waketime_stats);
 154         pthread_attr_init(&thread_attr);
 155         pthread_mutex_init(&thread_lock, NULL);
 156         pthread_cond_init(&thread_parent, NULL);
 157         pthread_cond_init(&thread_worker, NULL);
 158
 159         for (j = 0; j < bench_repeat && !done; j++) {
 160                 unsigned int nwoken = 0;
 161                 struct timeval start, end, runtime;
 162
 163                 /* create, launch & block all threads */
 164                 block_threads(worker, thread_attr);
 165
 166                 /* make sure all threads are already blocked */
 167                 pthread_mutex_lock(&thread_lock);
 168                 while (threads_starting)
 169                         pthread_cond_wait(&thread_parent, &thread_lock);
 170                 pthread_cond_broadcast(&thread_worker);
 171                 pthread_mutex_unlock(&thread_lock);
 172
 173                 usleep(100000);
 174
 175                 /* Ok, all threads are patiently blocked, start waking folks up */
 176                 gettimeofday(&start, NULL);
 177                 while (nwoken != nthreads)
 178                         nwoken += futex_wake(&futex1, nwakes, futex_flag);
 179                 gettimeofday(&end, NULL);
 180                 timersub(&end, &start, &runtime);
 181
 182                 update_stats(&wakeup_stats, nwoken);
 183                 update_stats(&waketime_stats, runtime.tv_usec);
 184
 185                 if (!silent) {
 186                         printf("[Run %d]: Wokeup %d of %d threads in %.4f ms\n",
 187                                j + 1, nwoken, nthreads, runtime.tv_usec / (double)USEC_PER_MSEC);
 188                 }
 189
 190                 for (i = 0; i < nthreads; i++) {
 191                         ret = pthread_join(worker[i], NULL);
 192                         if (ret)
 193                                 err(EXIT_FAILURE, "pthread_join");
 194                 }
 195
 196         }
 197
 198         /* cleanup & report results */
 199         pthread_cond_destroy(&thread_parent);
 200         pthread_cond_destroy(&thread_worker);
 201         pthread_mutex_destroy(&thread_lock);
 202         pthread_attr_destroy(&thread_attr);
 203
 204         print_summary();
 205
 206         free(worker);
 207         return ret;
 208 }