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 <internal/cpumap.h>
  24 #include <perf/cpumap.h>
  25 #include "bench.h"
  26 #include "futex.h"
  27
  28 #include <err.h>
  29 #include <stdlib.h>
  30 #include <sys/time.h>
  31
  32 static u_int32_t futex1 = 0, futex2 = 0;
  33
  34 /*
  35  * How many tasks to requeue at a time.
  36  * Default to 1 in order to make the kernel work more.
  37  */
  38 static unsigned int nrequeue = 1;
  39
  40 static pthread_t *worker;
  41 static bool done = false, silent = false, fshared = false;
  42 static pthread_mutex_t thread_lock;
  43 static pthread_cond_t thread_parent, thread_worker;
  44 static struct stats requeuetime_stats, requeued_stats;
  45 static unsigned int threads_starting, nthreads = 0;
  46 static int futex_flag = 0;
  47
  48 static const struct option options[] = {
  49         OPT_UINTEGER('t', "threads",  &nthreads, "Specify amount of threads"),
  50         OPT_UINTEGER('q', "nrequeue", &nrequeue, "Specify amount of threads to requeue at once"),
  51         OPT_BOOLEAN( 's', "silent",   &silent,   "Silent mode: do not display data/details"),
  52         OPT_BOOLEAN( 'S', "shared",   &fshared,  "Use shared futexes instead of private ones"),
  53         OPT_END()
  54 };
  55
  56 static const char * const bench_futex_requeue_usage[] = {
  57         "perf bench futex requeue <options>",
  58         NULL
  59 };
  60
  61 static void print_summary(void)
  62 {
  63         double requeuetime_avg = avg_stats(&requeuetime_stats);
  64         double requeuetime_stddev = stddev_stats(&requeuetime_stats);
  65         unsigned int requeued_avg = avg_stats(&requeued_stats);
  66
  67         printf("Requeued %d of %d threads in %.4f ms (+-%.2f%%)\n",
  68                requeued_avg,
  69                nthreads,
  70                requeuetime_avg / USEC_PER_MSEC,
  71                rel_stddev_stats(requeuetime_stddev, requeuetime_avg));
  72 }
  73
  74 static void *workerfn(void *arg __maybe_unused)
  75 {
  76         pthread_mutex_lock(&thread_lock);
  77         threads_starting--;
  78         if (!threads_starting)
  79                 pthread_cond_signal(&thread_parent);
  80         pthread_cond_wait(&thread_worker, &thread_lock);
  81         pthread_mutex_unlock(&thread_lock);
  82
  83         futex_wait(&futex1, 0, NULL, futex_flag);
  84         return NULL;
  85 }
  86
  87 static void block_threads(pthread_t *w,
  88                           pthread_attr_t thread_attr, struct perf_cpu_map *cpu)
  89 {
  90         cpu_set_t cpuset;
  91         unsigned int i;
  92
  93         threads_starting = nthreads;
  94
  95         /* create and block all threads */
  96         for (i = 0; i < nthreads; i++) {
  97                 CPU_ZERO(&cpuset);
  98                 CPU_SET(cpu->map[i % cpu->nr], &cpuset);
  99
 100                 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset))
 101                         err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
 102
 103                 if (pthread_create(&w[i], &thread_attr, workerfn, NULL))
 104                         err(EXIT_FAILURE, "pthread_create");
 105         }
 106 }
 107
 108 static void toggle_done(int sig __maybe_unused,
 109                         siginfo_t *info __maybe_unused,
 110                         void *uc __maybe_unused)
 111 {
 112         done = true;
 113 }
 114
 115 int bench_futex_requeue(int argc, const char **argv)
 116 {
 117         int ret = 0;
 118         unsigned int i, j;
 119         struct sigaction act;
 120         pthread_attr_t thread_attr;
 121         struct perf_cpu_map *cpu;
 122
 123         argc = parse_options(argc, argv, options, bench_futex_requeue_usage, 0);
 124         if (argc)
 125                 goto err;
 126
 127         cpu = perf_cpu_map__new(NULL);
 128         if (!cpu)
 129                 err(EXIT_FAILURE, "cpu_map__new");
 130
 131         memset(&act, 0, sizeof(act));
 132         sigfillset(&act.sa_mask);
 133         act.sa_sigaction = toggle_done;
 134         sigaction(SIGINT, &act, NULL);
 135
 136         if (!nthreads)
 137                 nthreads = cpu->nr;
 138
 139         worker = calloc(nthreads, sizeof(*worker));
 140         if (!worker)
 141                 err(EXIT_FAILURE, "calloc");
 142
 143         if (!fshared)
 144                 futex_flag = FUTEX_PRIVATE_FLAG;
 145
 146         if (nrequeue > nthreads)
 147                 nrequeue = nthreads;
 148
 149         printf("Run summary [PID %d]: Requeuing %d threads (from [%s] %p to %p), "
 150                "%d at a time.\n\n",  getpid(), nthreads,
 151                fshared ? "shared":"private", &futex1, &futex2, nrequeue);
 152
 153         init_stats(&requeued_stats);
 154         init_stats(&requeuetime_stats);
 155         pthread_attr_init(&thread_attr);
 156         pthread_mutex_init(&thread_lock, NULL);
 157         pthread_cond_init(&thread_parent, NULL);
 158         pthread_cond_init(&thread_worker, NULL);
 159
 160         for (j = 0; j < bench_repeat && !done; j++) {
 161                 unsigned int nrequeued = 0;
 162                 struct timeval start, end, runtime;
 163
 164                 /* create, launch & block all threads */
 165                 block_threads(worker, thread_attr, cpu);
 166
 167                 /* make sure all threads are already blocked */
 168                 pthread_mutex_lock(&thread_lock);
 169                 while (threads_starting)
 170                         pthread_cond_wait(&thread_parent, &thread_lock);
 171                 pthread_cond_broadcast(&thread_worker);
 172                 pthread_mutex_unlock(&thread_lock);
 173
 174                 usleep(100000);
 175
 176                 /* Ok, all threads are patiently blocked, start requeueing */
 177                 gettimeofday(&start, NULL);
 178                 while (nrequeued < nthreads) {
 179                         /*
 180                          * Do not wakeup any tasks blocked on futex1, allowing
 181                          * us to really measure futex_wait functionality.
 182                          */
 183                         nrequeued += futex_cmp_requeue(&futex1, 0, &futex2, 0,
 184                                                        nrequeue, futex_flag);
 185                 }
 186
 187                 gettimeofday(&end, NULL);
 188                 timersub(&end, &start, &runtime);
 189
 190                 update_stats(&requeued_stats, nrequeued);
 191                 update_stats(&requeuetime_stats, runtime.tv_usec);
 192
 193                 if (!silent) {
 194                         printf("[Run %d]: Requeued %d of %d threads in %.4f ms\n",
 195                                j + 1, nrequeued, nthreads, runtime.tv_usec / (double)USEC_PER_MSEC);
 196                 }
 197
 198                 /* everybody should be blocked on futex2, wake'em up */
 199                 nrequeued = futex_wake(&futex2, nrequeued, futex_flag);
 200                 if (nthreads != nrequeued)
 201                         warnx("couldn't wakeup all tasks (%d/%d)", nrequeued, nthreads);
 202
 203                 for (i = 0; i < nthreads; i++) {
 204                         ret = pthread_join(worker[i], NULL);
 205                         if (ret)
 206                                 err(EXIT_FAILURE, "pthread_join");
 207                 }
 208         }
 209
 210         /* cleanup & report results */
 211         pthread_cond_destroy(&thread_parent);
 212         pthread_cond_destroy(&thread_worker);
 213         pthread_mutex_destroy(&thread_lock);
 214         pthread_attr_destroy(&thread_attr);
 215
 216         print_summary();
 217
 218         free(worker);
 219         return ret;
 220 err:
 221         usage_with_options(bench_futex_requeue_usage, options);
 222         exit(EXIT_FAILURE);
 223 }