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 #include <sys/mman.h>
  31
  32 static u_int32_t futex1 = 0, futex2 = 0;
  33
  34 static pthread_t *worker;
  35 static bool done = false;
  36 static pthread_mutex_t thread_lock;
  37 static pthread_cond_t thread_parent, thread_worker;
  38 static struct stats requeuetime_stats, requeued_stats;
  39 static unsigned int threads_starting;
  40 static int futex_flag = 0;
  41
  42 static struct bench_futex_parameters params = {
  43         /*
  44          * How many tasks to requeue at a time.
  45          * Default to 1 in order to make the kernel work more.
  46          */
  47         .nrequeue = 1,
  48 };
  49
  50 static const struct option options[] = {
  51         OPT_UINTEGER('t', "threads",  &params.nthreads, "Specify amount of threads"),
  52         OPT_UINTEGER('q', "nrequeue", &params.nrequeue, "Specify amount of threads to requeue at once"),
  53         OPT_BOOLEAN( 's', "silent",   &params.silent, "Silent mode: do not display data/details"),
  54         OPT_BOOLEAN( 'S', "shared",   &params.fshared, "Use shared futexes instead of private ones"),
  55         OPT_BOOLEAN( 'm', "mlockall", &params.mlockall, "Lock all current and future memory"),
  56         OPT_BOOLEAN( 'B', "broadcast", &params.broadcast, "Requeue all threads at once"),
  57         OPT_END()
  58 };
  59
  60 static const char * const bench_futex_requeue_usage[] = {
  61         "perf bench futex requeue <options>",
  62         NULL
  63 };
  64
  65 static void print_summary(void)
  66 {
  67         double requeuetime_avg = avg_stats(&requeuetime_stats);
  68         double requeuetime_stddev = stddev_stats(&requeuetime_stats);
  69         unsigned int requeued_avg = avg_stats(&requeued_stats);
  70
  71         printf("Requeued %d of %d threads in %.4f ms (+-%.2f%%)\n",
  72                requeued_avg,
  73                params.nthreads,
  74                requeuetime_avg / USEC_PER_MSEC,
  75                rel_stddev_stats(requeuetime_stddev, requeuetime_avg));
  76 }
  77
  78 static void *workerfn(void *arg __maybe_unused)
  79 {
  80         int ret;
  81
  82         pthread_mutex_lock(&thread_lock);
  83         threads_starting--;
  84         if (!threads_starting)
  85                 pthread_cond_signal(&thread_parent);
  86         pthread_cond_wait(&thread_worker, &thread_lock);
  87         pthread_mutex_unlock(&thread_lock);
  88
  89         while (1) {
  90                 ret = futex_wait(&futex1, 0, NULL, futex_flag);
  91                 if (!ret)
  92                         break;
  93
  94                 if (ret && errno != EAGAIN) {
  95                         if (!params.silent)
  96                                 warn("futex_wait");
  97                         break;
  98                 }
  99         }
 100
 101         return NULL;
 102 }
 103
 104 static void block_threads(pthread_t *w,
 105                           pthread_attr_t thread_attr, struct perf_cpu_map *cpu)
 106 {
 107         cpu_set_t cpuset;
 108         unsigned int i;
 109
 110         threads_starting = params.nthreads;
 111
 112         /* create and block all threads */
 113         for (i = 0; i < params.nthreads; i++) {
 114                 CPU_ZERO(&cpuset);
 115                 CPU_SET(cpu->map[i % cpu->nr], &cpuset);
 116
 117                 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset))
 118                         err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
 119
 120                 if (pthread_create(&w[i], &thread_attr, workerfn, NULL))
 121                         err(EXIT_FAILURE, "pthread_create");
 122         }
 123 }
 124
 125 static void toggle_done(int sig __maybe_unused,
 126                         siginfo_t *info __maybe_unused,
 127                         void *uc __maybe_unused)
 128 {
 129         done = true;
 130 }
 131
 132 int bench_futex_requeue(int argc, const char **argv)
 133 {
 134         int ret = 0;
 135         unsigned int i, j;
 136         struct sigaction act;
 137         pthread_attr_t thread_attr;
 138         struct perf_cpu_map *cpu;
 139
 140         argc = parse_options(argc, argv, options, bench_futex_requeue_usage, 0);
 141         if (argc)
 142                 goto err;
 143
 144         cpu = perf_cpu_map__new(NULL);
 145         if (!cpu)
 146                 err(EXIT_FAILURE, "cpu_map__new");
 147
 148         memset(&act, 0, sizeof(act));
 149         sigfillset(&act.sa_mask);
 150         act.sa_sigaction = toggle_done;
 151         sigaction(SIGINT, &act, NULL);
 152
 153         if (params.mlockall) {
 154                 if (mlockall(MCL_CURRENT | MCL_FUTURE))
 155                         err(EXIT_FAILURE, "mlockall");
 156         }
 157
 158         if (!params.nthreads)
 159                 params.nthreads = cpu->nr;
 160
 161         worker = calloc(params.nthreads, sizeof(*worker));
 162         if (!worker)
 163                 err(EXIT_FAILURE, "calloc");
 164
 165         if (!params.fshared)
 166                 futex_flag = FUTEX_PRIVATE_FLAG;
 167
 168         if (params.nrequeue > params.nthreads)
 169                 params.nrequeue = params.nthreads;
 170
 171         if (params.broadcast)
 172                 params.nrequeue = params.nthreads;
 173
 174         printf("Run summary [PID %d]: Requeuing %d threads (from [%s] %p to %p), "
 175                "%d at a time.\n\n",  getpid(), params.nthreads,
 176                params.fshared ? "shared":"private", &futex1, &futex2, params.nrequeue);
 177
 178         init_stats(&requeued_stats);
 179         init_stats(&requeuetime_stats);
 180         pthread_attr_init(&thread_attr);
 181         pthread_mutex_init(&thread_lock, NULL);
 182         pthread_cond_init(&thread_parent, NULL);
 183         pthread_cond_init(&thread_worker, NULL);
 184
 185         for (j = 0; j < bench_repeat && !done; j++) {
 186                 unsigned int nrequeued = 0;
 187                 struct timeval start, end, runtime;
 188
 189                 /* create, launch & block all threads */
 190                 block_threads(worker, thread_attr, cpu);
 191
 192                 /* make sure all threads are already blocked */
 193                 pthread_mutex_lock(&thread_lock);
 194                 while (threads_starting)
 195                         pthread_cond_wait(&thread_parent, &thread_lock);
 196                 pthread_cond_broadcast(&thread_worker);
 197                 pthread_mutex_unlock(&thread_lock);
 198
 199                 usleep(100000);
 200
 201                 /* Ok, all threads are patiently blocked, start requeueing */
 202                 gettimeofday(&start, NULL);
 203                 while (nrequeued < params.nthreads) {
 204                         /*
 205                          * Do not wakeup any tasks blocked on futex1, allowing
 206                          * us to really measure futex_wait functionality.
 207                          */
 208                         nrequeued += futex_cmp_requeue(&futex1, 0, &futex2, 0,
 209                                                        params.nrequeue,
 210                                                        futex_flag);
 211                 }
 212
 213                 gettimeofday(&end, NULL);
 214                 timersub(&end, &start, &runtime);
 215
 216                 update_stats(&requeued_stats, nrequeued);
 217                 update_stats(&requeuetime_stats, runtime.tv_usec);
 218
 219                 if (!params.silent) {
 220                         printf("[Run %d]: Requeued %d of %d threads in %.4f ms\n",
 221                                j + 1, nrequeued, params.nthreads,
 222                                runtime.tv_usec / (double)USEC_PER_MSEC);
 223                 }
 224
 225                 /* everybody should be blocked on futex2, wake'em up */
 226                 nrequeued = futex_wake(&futex2, nrequeued, futex_flag);
 227                 if (params.nthreads != nrequeued)
 228                         warnx("couldn't wakeup all tasks (%d/%d)",
 229                               nrequeued, params.nthreads);
 230
 231                 for (i = 0; i < params.nthreads; i++) {
 232                         ret = pthread_join(worker[i], NULL);
 233                         if (ret)
 234                                 err(EXIT_FAILURE, "pthread_join");
 235                 }
 236         }
 237
 238         /* cleanup & report results */
 239         pthread_cond_destroy(&thread_parent);
 240         pthread_cond_destroy(&thread_worker);
 241         pthread_mutex_destroy(&thread_lock);
 242         pthread_attr_destroy(&thread_attr);
 243
 244         print_summary();
 245
 246         free(worker);
 247         return ret;
 248 err:
 249         usage_with_options(bench_futex_requeue_usage, options);
 250         exit(EXIT_FAILURE);
 251 }