tools/perf/bench/futex-requeue.c

   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 -- thus mimicking a regular futex_wait.
  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
  31 static u_int32_t futex1 = 0, futex2 = 0;
  32
  33 /*
  34  * How many tasks to requeue at a time.
  35  * Default to 1 in order to make the kernel work more.
  36  */
  37 static unsigned int nrequeue = 1;
  38
  39 static 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 requeuetime_stats, requeued_stats;
  44 static unsigned int 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('q', "nrequeue", &nrequeue, "Specify amount of threads to requeue 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_requeue_usage[] = {
  56         "perf bench futex requeue <options>",
  57         NULL
  58 };
  59
  60 static void print_summary(void)
  61 {
  62         double requeuetime_avg = avg_stats(&requeuetime_stats);
  63         double requeuetime_stddev = stddev_stats(&requeuetime_stats);
  64         unsigned int requeued_avg = avg_stats(&requeued_stats);
  65
  66         printf("Requeued %d of %d threads in %.4f ms (+-%.2f%%)\n",
  67                requeued_avg,
  68                nthreads,
  69                requeuetime_avg / USEC_PER_MSEC,
  70                rel_stddev_stats(requeuetime_stddev, requeuetime_avg));
  71 }
  72
  73 static void *workerfn(void *arg __maybe_unused)
  74 {
  75         pthread_mutex_lock(&thread_lock);
  76         threads_starting--;
  77         if (!threads_starting)
  78                 pthread_cond_signal(&thread_parent);
  79         pthread_cond_wait(&thread_worker, &thread_lock);
  80         pthread_mutex_unlock(&thread_lock);
  81
  82         futex_wait(&futex1, 0, NULL, futex_flag);
  83         return NULL;
  84 }
  85
  86 static void block_threads(pthread_t *w,
  87                           pthread_attr_t thread_attr, struct perf_cpu_map *cpu)
  88 {
  89         cpu_set_t cpuset;
  90         unsigned int i;
  91
  92         threads_starting = nthreads;
  93
  94         /* create and block all threads */
  95         for (i = 0; i < nthreads; i++) {
  96                 CPU_ZERO(&cpuset);
  97                 CPU_SET(cpu->map[i % cpu->nr], &cpuset);
  98
  99                 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset))
 100                         err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
 101
 102                 if (pthread_create(&w[i], &thread_attr, workerfn, NULL))
 103                         err(EXIT_FAILURE, "pthread_create");
 104         }
 105 }
 106
 107 static void toggle_done(int sig __maybe_unused,
 108                         siginfo_t *info __maybe_unused,
 109                         void *uc __maybe_unused)
 110 {
 111         done = true;
 112 }
 113
 114 int bench_futex_requeue(int argc, const char **argv)
 115 {
 116         int ret = 0;
 117         unsigned int i, j;
 118         struct sigaction act;
 119         pthread_attr_t thread_attr;
 120         struct perf_cpu_map *cpu;
 121
 122         argc = parse_options(argc, argv, options, bench_futex_requeue_usage, 0);
 123         if (argc)
 124                 goto err;
 125
 126         cpu = perf_cpu_map__new(NULL);
 127         if (!cpu)
 128                 err(EXIT_FAILURE, "cpu_map__new");
 129
 130         memset(&act, 0, sizeof(act));
 131         sigfillset(&act.sa_mask);
 132         act.sa_sigaction = toggle_done;
 133         sigaction(SIGINT, &act, NULL);
 134
 135         if (!nthreads)
 136                 nthreads = cpu->nr;
 137
 138         worker = calloc(nthreads, sizeof(*worker));
 139         if (!worker)
 140                 err(EXIT_FAILURE, "calloc");
 141
 142         if (!fshared)
 143                 futex_flag = FUTEX_PRIVATE_FLAG;
 144
 145         if (nrequeue > nthreads)
 146                 nrequeue = nthreads;
 147
 148         printf("Run summary [PID %d]: Requeuing %d threads (from [%s] %p to %p), "
 149                "%d at a time.\n\n",  getpid(), nthreads,
 150                fshared ? "shared":"private", &futex1, &futex2, nrequeue);
 151
 152         init_stats(&requeued_stats);
 153         init_stats(&requeuetime_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 nrequeued = 0;
 161                 struct timeval start, end, runtime;
 162
 163                 /* create, launch & block all threads */
 164                 block_threads(worker, thread_attr, cpu);
 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 requeueing */
 176                 gettimeofday(&start, NULL);
 177                 while (nrequeued < nthreads) {
 178                         /*
 179                          * Do not wakeup any tasks blocked on futex1, allowing
 180                          * us to really measure futex_wait functionality.
 181                          */
 182                         nrequeued += futex_cmp_requeue(&futex1, 0, &futex2, 0,
 183                                                        nrequeue, futex_flag);
 184                 }
 185
 186                 gettimeofday(&end, NULL);
 187                 timersub(&end, &start, &runtime);
 188
 189                 update_stats(&requeued_stats, nrequeued);
 190                 update_stats(&requeuetime_stats, runtime.tv_usec);
 191
 192                 if (!silent) {
 193                         printf("[Run %d]: Requeued %d of %d threads in %.4f ms\n",
 194                                j + 1, nrequeued, nthreads, runtime.tv_usec / (double)USEC_PER_MSEC);
 195                 }
 196
 197                 /* everybody should be blocked on futex2, wake'em up */
 198                 nrequeued = futex_wake(&futex2, nrequeued, futex_flag);
 199                 if (nthreads != nrequeued)
 200                         warnx("couldn't wakeup all tasks (%d/%d)", nrequeued, nthreads);
 201
 202                 for (i = 0; i < nthreads; i++) {
 203                         ret = pthread_join(worker[i], NULL);
 204                         if (ret)
 205                                 err(EXIT_FAILURE, "pthread_join");
 206                 }
 207         }
 208
 209         /* cleanup & report results */
 210         pthread_cond_destroy(&thread_parent);
 211         pthread_cond_destroy(&thread_worker);
 212         pthread_mutex_destroy(&thread_lock);
 213         pthread_attr_destroy(&thread_attr);
 214
 215         print_summary();
 216
 217         free(worker);
 218         return ret;
 219 err:
 220         usage_with_options(bench_futex_requeue_usage, options);
 221         exit(EXIT_FAILURE);
 222 }