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         pthread_mutex_lock(&thread_lock);
  81         threads_starting--;
  82         if (!threads_starting)
  83                 pthread_cond_signal(&thread_parent);
  84         pthread_cond_wait(&thread_worker, &thread_lock);
  85         pthread_mutex_unlock(&thread_lock);
  86
  87         futex_wait(&futex1, 0, NULL, futex_flag);
  88         return NULL;
  89 }
  90
  91 static void block_threads(pthread_t *w,
  92                           pthread_attr_t thread_attr, struct perf_cpu_map *cpu)
  93 {
  94         cpu_set_t cpuset;
  95         unsigned int i;
  96
  97         threads_starting = params.nthreads;
  98
  99         /* create and block all threads */
 100         for (i = 0; i < params.nthreads; i++) {
 101                 CPU_ZERO(&cpuset);
 102                 CPU_SET(cpu->map[i % cpu->nr], &cpuset);
 103
 104                 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset))
 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_requeue(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         struct perf_cpu_map *cpu;
 126
 127         argc = parse_options(argc, argv, options, bench_futex_requeue_usage, 0);
 128         if (argc)
 129                 goto err;
 130
 131         cpu = perf_cpu_map__new(NULL);
 132         if (!cpu)
 133                 err(EXIT_FAILURE, "cpu_map__new");
 134
 135         memset(&act, 0, sizeof(act));
 136         sigfillset(&act.sa_mask);
 137         act.sa_sigaction = toggle_done;
 138         sigaction(SIGINT, &act, NULL);
 139
 140         if (params.mlockall) {
 141                 if (mlockall(MCL_CURRENT | MCL_FUTURE))
 142                         err(EXIT_FAILURE, "mlockall");
 143         }
 144
 145         if (!params.nthreads)
 146                 params.nthreads = cpu->nr;
 147
 148         worker = calloc(params.nthreads, sizeof(*worker));
 149         if (!worker)
 150                 err(EXIT_FAILURE, "calloc");
 151
 152         if (!params.fshared)
 153                 futex_flag = FUTEX_PRIVATE_FLAG;
 154
 155         if (params.nrequeue > params.nthreads)
 156                 params.nrequeue = params.nthreads;
 157
 158         if (params.broadcast)
 159                 params.nrequeue = params.nthreads;
 160
 161         printf("Run summary [PID %d]: Requeuing %d threads (from [%s] %p to %p), "
 162                "%d at a time.\n\n",  getpid(), params.nthreads,
 163                params.fshared ? "shared":"private", &futex1, &futex2, params.nrequeue);
 164
 165         init_stats(&requeued_stats);
 166         init_stats(&requeuetime_stats);
 167         pthread_attr_init(&thread_attr);
 168         pthread_mutex_init(&thread_lock, NULL);
 169         pthread_cond_init(&thread_parent, NULL);
 170         pthread_cond_init(&thread_worker, NULL);
 171
 172         for (j = 0; j < bench_repeat && !done; j++) {
 173                 unsigned int nrequeued = 0;
 174                 struct timeval start, end, runtime;
 175
 176                 /* create, launch & block all threads */
 177                 block_threads(worker, thread_attr, cpu);
 178
 179                 /* make sure all threads are already blocked */
 180                 pthread_mutex_lock(&thread_lock);
 181                 while (threads_starting)
 182                         pthread_cond_wait(&thread_parent, &thread_lock);
 183                 pthread_cond_broadcast(&thread_worker);
 184                 pthread_mutex_unlock(&thread_lock);
 185
 186                 usleep(100000);
 187
 188                 /* Ok, all threads are patiently blocked, start requeueing */
 189                 gettimeofday(&start, NULL);
 190                 while (nrequeued < params.nthreads) {
 191                         /*
 192                          * Do not wakeup any tasks blocked on futex1, allowing
 193                          * us to really measure futex_wait functionality.
 194                          */
 195                         nrequeued += futex_cmp_requeue(&futex1, 0, &futex2, 0,
 196                                                        params.nrequeue,
 197                                                        futex_flag);
 198                 }
 199
 200                 gettimeofday(&end, NULL);
 201                 timersub(&end, &start, &runtime);
 202
 203                 update_stats(&requeued_stats, nrequeued);
 204                 update_stats(&requeuetime_stats, runtime.tv_usec);
 205
 206                 if (!params.silent) {
 207                         printf("[Run %d]: Requeued %d of %d threads in %.4f ms\n",
 208                                j + 1, nrequeued, params.nthreads,
 209                                runtime.tv_usec / (double)USEC_PER_MSEC);
 210                 }
 211
 212                 /* everybody should be blocked on futex2, wake'em up */
 213                 nrequeued = futex_wake(&futex2, nrequeued, futex_flag);
 214                 if (params.nthreads != nrequeued)
 215                         warnx("couldn't wakeup all tasks (%d/%d)",
 216                               nrequeued, params.nthreads);
 217
 218                 for (i = 0; i < params.nthreads; i++) {
 219                         ret = pthread_join(worker[i], NULL);
 220                         if (ret)
 221                                 err(EXIT_FAILURE, "pthread_join");
 222                 }
 223         }
 224
 225         /* cleanup & report results */
 226         pthread_cond_destroy(&thread_parent);
 227         pthread_cond_destroy(&thread_worker);
 228         pthread_mutex_destroy(&thread_lock);
 229         pthread_attr_destroy(&thread_attr);
 230
 231         print_summary();
 232
 233         free(worker);
 234         return ret;
 235 err:
 236         usage_with_options(bench_futex_requeue_usage, options);
 237         exit(EXIT_FAILURE);
 238 }