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 /* For the CLR_() macros */
  12 #include <pthread.h>
  13
  14 #include <signal.h>
  15 #include "../util/stat.h"
  16 #include <subcmd/parse-options.h>
  17 #include <linux/compiler.h>
  18 #include <linux/kernel.h>
  19 #include <errno.h>
  20 #include "bench.h"
  21 #include "futex.h"
  22
  23 #include <err.h>
  24 #include <stdlib.h>
  25 #include <sys/time.h>
  26
  27 static u_int32_t futex1 = 0, futex2 = 0;
  28
  29 /*
  30  * How many tasks to requeue at a time.
  31  * Default to 1 in order to make the kernel work more.
  32  */
  33 static unsigned int nrequeue = 1;
  34
  35 static pthread_t *worker;
  36 static bool done = false, silent = false, fshared = false;
  37 static pthread_mutex_t thread_lock;
  38 static pthread_cond_t thread_parent, thread_worker;
  39 static struct stats requeuetime_stats, requeued_stats;
  40 static unsigned int ncpus, threads_starting, nthreads = 0;
  41 static int futex_flag = 0;
  42
  43 static const struct option options[] = {
  44         OPT_UINTEGER('t', "threads",  &nthreads, "Specify amount of threads"),
  45         OPT_UINTEGER('q', "nrequeue", &nrequeue, "Specify amount of threads to requeue at once"),
  46         OPT_BOOLEAN( 's', "silent",   &silent,   "Silent mode: do not display data/details"),
  47         OPT_BOOLEAN( 'S', "shared",   &fshared,  "Use shared futexes instead of private ones"),
  48         OPT_END()
  49 };
  50
  51 static const char * const bench_futex_requeue_usage[] = {
  52         "perf bench futex requeue <options>",
  53         NULL
  54 };
  55
  56 static void print_summary(void)
  57 {
  58         double requeuetime_avg = avg_stats(&requeuetime_stats);
  59         double requeuetime_stddev = stddev_stats(&requeuetime_stats);
  60         unsigned int requeued_avg = avg_stats(&requeued_stats);
  61
  62         printf("Requeued %d of %d threads in %.4f ms (+-%.2f%%)\n",
  63                requeued_avg,
  64                nthreads,
  65                requeuetime_avg/1e3,
  66                rel_stddev_stats(requeuetime_stddev, requeuetime_avg));
  67 }
  68
  69 static void *workerfn(void *arg __maybe_unused)
  70 {
  71         pthread_mutex_lock(&thread_lock);
  72         threads_starting--;
  73         if (!threads_starting)
  74                 pthread_cond_signal(&thread_parent);
  75         pthread_cond_wait(&thread_worker, &thread_lock);
  76         pthread_mutex_unlock(&thread_lock);
  77
  78         futex_wait(&futex1, 0, NULL, futex_flag);
  79         return NULL;
  80 }
  81
  82 static void block_threads(pthread_t *w,
  83                           pthread_attr_t thread_attr)
  84 {
  85         cpu_set_t cpu;
  86         unsigned int i;
  87
  88         threads_starting = nthreads;
  89
  90         /* create and block all threads */
  91         for (i = 0; i < nthreads; i++) {
  92                 CPU_ZERO(&cpu);
  93                 CPU_SET(i % ncpus, &cpu);
  94
  95                 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu))
  96                         err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
  97
  98                 if (pthread_create(&w[i], &thread_attr, workerfn, NULL))
  99                         err(EXIT_FAILURE, "pthread_create");
 100         }
 101 }
 102
 103 static void toggle_done(int sig __maybe_unused,
 104                         siginfo_t *info __maybe_unused,
 105                         void *uc __maybe_unused)
 106 {
 107         done = true;
 108 }
 109
 110 int bench_futex_requeue(int argc, const char **argv,
 111                         const char *prefix __maybe_unused)
 112 {
 113         int ret = 0;
 114         unsigned int i, j;
 115         struct sigaction act;
 116         pthread_attr_t thread_attr;
 117
 118         argc = parse_options(argc, argv, options, bench_futex_requeue_usage, 0);
 119         if (argc)
 120                 goto err;
 121
 122         ncpus = sysconf(_SC_NPROCESSORS_ONLN);
 123
 124         sigfillset(&act.sa_mask);
 125         act.sa_sigaction = toggle_done;
 126         sigaction(SIGINT, &act, NULL);
 127
 128         if (!nthreads)
 129                 nthreads = ncpus;
 130
 131         worker = calloc(nthreads, sizeof(*worker));
 132         if (!worker)
 133                 err(EXIT_FAILURE, "calloc");
 134
 135         if (!fshared)
 136                 futex_flag = FUTEX_PRIVATE_FLAG;
 137
 138         if (nrequeue > nthreads)
 139                 nrequeue = nthreads;
 140
 141         printf("Run summary [PID %d]: Requeuing %d threads (from [%s] %p to %p), "
 142                "%d at a time.\n\n",  getpid(), nthreads,
 143                fshared ? "shared":"private", &futex1, &futex2, nrequeue);
 144
 145         init_stats(&requeued_stats);
 146         init_stats(&requeuetime_stats);
 147         pthread_attr_init(&thread_attr);
 148         pthread_mutex_init(&thread_lock, NULL);
 149         pthread_cond_init(&thread_parent, NULL);
 150         pthread_cond_init(&thread_worker, NULL);
 151
 152         for (j = 0; j < bench_repeat && !done; j++) {
 153                 unsigned int nrequeued = 0;
 154                 struct timeval start, end, runtime;
 155
 156                 /* create, launch & block all threads */
 157                 block_threads(worker, thread_attr);
 158
 159                 /* make sure all threads are already blocked */
 160                 pthread_mutex_lock(&thread_lock);
 161                 while (threads_starting)
 162                         pthread_cond_wait(&thread_parent, &thread_lock);
 163                 pthread_cond_broadcast(&thread_worker);
 164                 pthread_mutex_unlock(&thread_lock);
 165
 166                 usleep(100000);
 167
 168                 /* Ok, all threads are patiently blocked, start requeueing */
 169                 gettimeofday(&start, NULL);
 170                 while (nrequeued < nthreads) {
 171                         /*
 172                          * Do not wakeup any tasks blocked on futex1, allowing
 173                          * us to really measure futex_wait functionality.
 174                          */
 175                         nrequeued += futex_cmp_requeue(&futex1, 0, &futex2, 0,
 176                                                        nrequeue, futex_flag);
 177                 }
 178
 179                 gettimeofday(&end, NULL);
 180                 timersub(&end, &start, &runtime);
 181
 182                 update_stats(&requeued_stats, nrequeued);
 183                 update_stats(&requeuetime_stats, runtime.tv_usec);
 184
 185                 if (!silent) {
 186                         printf("[Run %d]: Requeued %d of %d threads in %.4f ms\n",
 187                                j + 1, nrequeued, nthreads, runtime.tv_usec/1e3);
 188                 }
 189
 190                 /* everybody should be blocked on futex2, wake'em up */
 191                 nrequeued = futex_wake(&futex2, nrequeued, futex_flag);
 192                 if (nthreads != nrequeued)
 193                         warnx("couldn't wakeup all tasks (%d/%d)", nrequeued, nthreads);
 194
 195                 for (i = 0; i < nthreads; i++) {
 196                         ret = pthread_join(worker[i], NULL);
 197                         if (ret)
 198                                 err(EXIT_FAILURE, "pthread_join");
 199                 }
 200         }
 201
 202         /* cleanup & report results */
 203         pthread_cond_destroy(&thread_parent);
 204         pthread_cond_destroy(&thread_worker);
 205         pthread_mutex_destroy(&thread_lock);
 206         pthread_attr_destroy(&thread_attr);
 207
 208         print_summary();
 209
 210         free(worker);
 211         return ret;
 212 err:
 213         usage_with_options(bench_futex_requeue_usage, options);
 214         exit(EXIT_FAILURE);
 215 }