tools/perf/bench/futex-requeue.c

   1 /*
   2  * Copyright (C) 2013  Davidlohr Bueso <davidlohr@hp.com>
   3  *
   4  * futex-requeue: Block a bunch of threads on futex1 and requeue them
   5  *                on futex2, N at a time.
   6  *
   7  * This program is particularly useful to measure the latency of nthread
   8  * requeues without waking up any tasks -- thus mimicking a regular futex_wait.
   9  */
  10
  11 #include "../perf.h"
  12 #include "../util/util.h"
  13 #include "../util/stat.h"
  14 #include "../util/parse-options.h"
  15 #include "../util/header.h"
  16 #include "bench.h"
  17 #include "futex.h"
  18
  19 #include <err.h>
  20 #include <stdlib.h>
  21 #include <sys/time.h>
  22 #include <pthread.h>
  23
  24 static u_int32_t futex1 = 0, futex2 = 0;
  25
  26 /*
  27  * How many tasks to requeue at a time.
  28  * Default to 1 in order to make the kernel work more.
  29  */
  30 static unsigned int nrequeue = 1;
  31
  32 static pthread_t *worker;
  33 static bool done = false, silent = false, fshared = false;
  34 static pthread_mutex_t thread_lock;
  35 static pthread_cond_t thread_parent, thread_worker;
  36 static struct stats requeuetime_stats, requeued_stats;
  37 static unsigned int ncpus, threads_starting, nthreads = 0;
  38 static int futex_flag = 0;
  39
  40 static const struct option options[] = {
  41         OPT_UINTEGER('t', "threads",  &nthreads, "Specify amount of threads"),
  42         OPT_UINTEGER('q', "nrequeue", &nrequeue, "Specify amount of threads to requeue at once"),
  43         OPT_BOOLEAN( 's', "silent",   &silent,   "Silent mode: do not display data/details"),
  44         OPT_BOOLEAN( 'S', "shared",   &fshared,  "Use shared futexes instead of private ones"),
  45         OPT_END()
  46 };
  47
  48 static const char * const bench_futex_requeue_usage[] = {
  49         "perf bench futex requeue <options>",
  50         NULL
  51 };
  52
  53 static void print_summary(void)
  54 {
  55         double requeuetime_avg = avg_stats(&requeuetime_stats);
  56         double requeuetime_stddev = stddev_stats(&requeuetime_stats);
  57         unsigned int requeued_avg = avg_stats(&requeued_stats);
  58
  59         printf("Requeued %d of %d threads in %.4f ms (+-%.2f%%)\n",
  60                requeued_avg,
  61                nthreads,
  62                requeuetime_avg/1e3,
  63                rel_stddev_stats(requeuetime_stddev, requeuetime_avg));
  64 }
  65
  66 static void *workerfn(void *arg __maybe_unused)
  67 {
  68         pthread_mutex_lock(&thread_lock);
  69         threads_starting--;
  70         if (!threads_starting)
  71                 pthread_cond_signal(&thread_parent);
  72         pthread_cond_wait(&thread_worker, &thread_lock);
  73         pthread_mutex_unlock(&thread_lock);
  74
  75         futex_wait(&futex1, 0, NULL, futex_flag);
  76         return NULL;
  77 }
  78
  79 static void block_threads(pthread_t *w,
  80                           pthread_attr_t thread_attr)
  81 {
  82         cpu_set_t cpu;
  83         unsigned int i;
  84
  85         threads_starting = nthreads;
  86
  87         /* create and block all threads */
  88         for (i = 0; i < nthreads; i++) {
  89                 CPU_ZERO(&cpu);
  90                 CPU_SET(i % ncpus, &cpu);
  91
  92                 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu))
  93                         err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
  94
  95                 if (pthread_create(&w[i], &thread_attr, workerfn, NULL))
  96                         err(EXIT_FAILURE, "pthread_create");
  97         }
  98 }
  99
 100 static void toggle_done(int sig __maybe_unused,
 101                         siginfo_t *info __maybe_unused,
 102                         void *uc __maybe_unused)
 103 {
 104         done = true;
 105 }
 106
 107 int bench_futex_requeue(int argc, const char **argv,
 108                         const char *prefix __maybe_unused)
 109 {
 110         int ret = 0;
 111         unsigned int i, j;
 112         struct sigaction act;
 113         pthread_attr_t thread_attr;
 114
 115         argc = parse_options(argc, argv, options, bench_futex_requeue_usage, 0);
 116         if (argc)
 117                 goto err;
 118
 119         ncpus = sysconf(_SC_NPROCESSORS_ONLN);
 120
 121         sigfillset(&act.sa_mask);
 122         act.sa_sigaction = toggle_done;
 123         sigaction(SIGINT, &act, NULL);
 124
 125         if (!nthreads)
 126                 nthreads = ncpus;
 127
 128         worker = calloc(nthreads, sizeof(*worker));
 129         if (!worker)
 130                 err(EXIT_FAILURE, "calloc");
 131
 132         if (!fshared)
 133                 futex_flag = FUTEX_PRIVATE_FLAG;
 134
 135         if (nrequeue > nthreads)
 136                 nrequeue = nthreads;
 137
 138         printf("Run summary [PID %d]: Requeuing %d threads (from [%s] %p to %p), "
 139                "%d at a time.\n\n",  getpid(), nthreads,
 140                fshared ? "shared":"private", &futex1, &futex2, nrequeue);
 141
 142         init_stats(&requeued_stats);
 143         init_stats(&requeuetime_stats);
 144         pthread_attr_init(&thread_attr);
 145         pthread_mutex_init(&thread_lock, NULL);
 146         pthread_cond_init(&thread_parent, NULL);
 147         pthread_cond_init(&thread_worker, NULL);
 148
 149         for (j = 0; j < bench_repeat && !done; j++) {
 150                 unsigned int nrequeued = 0;
 151                 struct timeval start, end, runtime;
 152
 153                 /* create, launch & block all threads */
 154                 block_threads(worker, thread_attr);
 155
 156                 /* make sure all threads are already blocked */
 157                 pthread_mutex_lock(&thread_lock);
 158                 while (threads_starting)
 159                         pthread_cond_wait(&thread_parent, &thread_lock);
 160                 pthread_cond_broadcast(&thread_worker);
 161                 pthread_mutex_unlock(&thread_lock);
 162
 163                 usleep(100000);
 164
 165                 /* Ok, all threads are patiently blocked, start requeueing */
 166                 gettimeofday(&start, NULL);
 167                 while (nrequeued < nthreads) {
 168                         /*
 169                          * Do not wakeup any tasks blocked on futex1, allowing
 170                          * us to really measure futex_wait functionality.
 171                          */
 172                         nrequeued += futex_cmp_requeue(&futex1, 0, &futex2, 0,
 173                                                        nrequeue, futex_flag);
 174                 }
 175
 176                 gettimeofday(&end, NULL);
 177                 timersub(&end, &start, &runtime);
 178
 179                 update_stats(&requeued_stats, nrequeued);
 180                 update_stats(&requeuetime_stats, runtime.tv_usec);
 181
 182                 if (!silent) {
 183                         printf("[Run %d]: Requeued %d of %d threads in %.4f ms\n",
 184                                j + 1, nrequeued, nthreads, runtime.tv_usec/1e3);
 185                 }
 186
 187                 /* everybody should be blocked on futex2, wake'em up */
 188                 nrequeued = futex_wake(&futex2, nrequeued, futex_flag);
 189                 if (nthreads != nrequeued)
 190                         warnx("couldn't wakeup all tasks (%d/%d)", nrequeued, nthreads);
 191
 192                 for (i = 0; i < nthreads; i++) {
 193                         ret = pthread_join(worker[i], NULL);
 194                         if (ret)
 195                                 err(EXIT_FAILURE, "pthread_join");
 196                 }
 197         }
 198
 199         /* cleanup & report results */
 200         pthread_cond_destroy(&thread_parent);
 201         pthread_cond_destroy(&thread_worker);
 202         pthread_mutex_destroy(&thread_lock);
 203         pthread_attr_destroy(&thread_attr);
 204
 205         print_summary();
 206
 207         free(worker);
 208         return ret;
 209 err:
 210         usage_with_options(bench_futex_requeue_usage, options);
 211         exit(EXIT_FAILURE);
 212 }