tools/perf/bench/futex-hash.c

   1 /*
   2  * Copyright (C) 2013  Davidlohr Bueso <davidlohr@hp.com>
   3  *
   4  * futex-hash: Stress the hell out of the Linux kernel futex uaddr hashing.
   5  *
   6  * This program is particularly useful for measuring the kernel's futex hash
   7  * table/function implementation. In order for it to make sense, use with as
   8  * many threads and futexes as possible.
   9  */
  10
  11 /* For the CLR_() macros */
  12 #include <string.h>
  13 #include <pthread.h>
  14
  15 #include <errno.h>
  16 #include <signal.h>
  17 #include <stdlib.h>
  18 #include <linux/compiler.h>
  19 #include <linux/kernel.h>
  20 #include <sys/time.h>
  21
  22 #include "../util/stat.h"
  23 #include <subcmd/parse-options.h>
  24 #include "bench.h"
  25 #include "futex.h"
  26
  27 #include <err.h>
  28 #include <sys/time.h>
  29
  30 static unsigned int nthreads = 0;
  31 static unsigned int nsecs    = 10;
  32 /* amount of futexes per thread */
  33 static unsigned int nfutexes = 1024;
  34 static bool fshared = false, done = false, silent = false;
  35 static int futex_flag = 0;
  36
  37 struct timeval start, end, runtime;
  38 static pthread_mutex_t thread_lock;
  39 static unsigned int threads_starting;
  40 static struct stats throughput_stats;
  41 static pthread_cond_t thread_parent, thread_worker;
  42
  43 struct worker {
  44         int tid;
  45         u_int32_t *futex;
  46         pthread_t thread;
  47         unsigned long ops;
  48 };
  49
  50 static const struct option options[] = {
  51         OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
  52         OPT_UINTEGER('r', "runtime", &nsecs,    "Specify runtime (in seconds)"),
  53         OPT_UINTEGER('f', "futexes", &nfutexes, "Specify amount of futexes per threads"),
  54         OPT_BOOLEAN( 's', "silent",  &silent,   "Silent mode: do not display data/details"),
  55         OPT_BOOLEAN( 'S', "shared",  &fshared,  "Use shared futexes instead of private ones"),
  56         OPT_END()
  57 };
  58
  59 static const char * const bench_futex_hash_usage[] = {
  60         "perf bench futex hash <options>",
  61         NULL
  62 };
  63
  64 static void *workerfn(void *arg)
  65 {
  66         int ret;
  67         struct worker *w = (struct worker *) arg;
  68         unsigned int i;
  69         unsigned long ops = w->ops; /* avoid cacheline bouncing */
  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         do {
  79                 for (i = 0; i < nfutexes; i++, ops++) {
  80                         /*
  81                          * We want the futex calls to fail in order to stress
  82                          * the hashing of uaddr and not measure other steps,
  83                          * such as internal waitqueue handling, thus enlarging
  84                          * the critical region protected by hb->lock.
  85                          */
  86                         ret = futex_wait(&w->futex[i], 1234, NULL, futex_flag);
  87                         if (!silent &&
  88                             (!ret || errno != EAGAIN || errno != EWOULDBLOCK))
  89                                 warn("Non-expected futex return call");
  90                 }
  91         }  while (!done);
  92
  93         w->ops = ops;
  94         return NULL;
  95 }
  96
  97 static void toggle_done(int sig __maybe_unused,
  98                         siginfo_t *info __maybe_unused,
  99                         void *uc __maybe_unused)
 100 {
 101         /* inform all threads that we're done for the day */
 102         done = true;
 103         gettimeofday(&end, NULL);
 104         timersub(&end, &start, &runtime);
 105 }
 106
 107 static void print_summary(void)
 108 {
 109         unsigned long avg = avg_stats(&throughput_stats);
 110         double stddev = stddev_stats(&throughput_stats);
 111
 112         printf("%sAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n",
 113                !silent ? "\n" : "", avg, rel_stddev_stats(stddev, avg),
 114                (int) runtime.tv_sec);
 115 }
 116
 117 int bench_futex_hash(int argc, const char **argv)
 118 {
 119         int ret = 0;
 120         cpu_set_t cpu;
 121         struct sigaction act;
 122         unsigned int i, ncpus;
 123         pthread_attr_t thread_attr;
 124         struct worker *worker = NULL;
 125
 126         argc = parse_options(argc, argv, options, bench_futex_hash_usage, 0);
 127         if (argc) {
 128                 usage_with_options(bench_futex_hash_usage, options);
 129                 exit(EXIT_FAILURE);
 130         }
 131
 132         ncpus = sysconf(_SC_NPROCESSORS_ONLN);
 133
 134         sigfillset(&act.sa_mask);
 135         act.sa_sigaction = toggle_done;
 136         sigaction(SIGINT, &act, NULL);
 137
 138         if (!nthreads) /* default to the number of CPUs */
 139                 nthreads = ncpus;
 140
 141         worker = calloc(nthreads, sizeof(*worker));
 142         if (!worker)
 143                 goto errmem;
 144
 145         if (!fshared)
 146                 futex_flag = FUTEX_PRIVATE_FLAG;
 147
 148         printf("Run summary [PID %d]: %d threads, each operating on %d [%s] futexes for %d secs.\n\n",
 149                getpid(), nthreads, nfutexes, fshared ? "shared":"private", nsecs);
 150
 151         init_stats(&throughput_stats);
 152         pthread_mutex_init(&thread_lock, NULL);
 153         pthread_cond_init(&thread_parent, NULL);
 154         pthread_cond_init(&thread_worker, NULL);
 155
 156         threads_starting = nthreads;
 157         pthread_attr_init(&thread_attr);
 158         gettimeofday(&start, NULL);
 159         for (i = 0; i < nthreads; i++) {
 160                 worker[i].tid = i;
 161                 worker[i].futex = calloc(nfutexes, sizeof(*worker[i].futex));
 162                 if (!worker[i].futex)
 163                         goto errmem;
 164
 165                 CPU_ZERO(&cpu);
 166                 CPU_SET(i % ncpus, &cpu);
 167
 168                 ret = pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu);
 169                 if (ret)
 170                         err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
 171
 172                 ret = pthread_create(&worker[i].thread, &thread_attr, workerfn,
 173                                      (void *)(struct worker *) &worker[i]);
 174                 if (ret)
 175                         err(EXIT_FAILURE, "pthread_create");
 176
 177         }
 178         pthread_attr_destroy(&thread_attr);
 179
 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         sleep(nsecs);
 187         toggle_done(0, NULL, NULL);
 188
 189         for (i = 0; i < nthreads; i++) {
 190                 ret = pthread_join(worker[i].thread, NULL);
 191                 if (ret)
 192                         err(EXIT_FAILURE, "pthread_join");
 193         }
 194
 195         /* cleanup & report results */
 196         pthread_cond_destroy(&thread_parent);
 197         pthread_cond_destroy(&thread_worker);
 198         pthread_mutex_destroy(&thread_lock);
 199
 200         for (i = 0; i < nthreads; i++) {
 201                 unsigned long t = worker[i].ops/runtime.tv_sec;
 202                 update_stats(&throughput_stats, t);
 203                 if (!silent) {
 204                         if (nfutexes == 1)
 205                                 printf("[thread %2d] futex: %p [ %ld ops/sec ]\n",
 206                                        worker[i].tid, &worker[i].futex[0], t);
 207                         else
 208                                 printf("[thread %2d] futexes: %p ... %p [ %ld ops/sec ]\n",
 209                                        worker[i].tid, &worker[i].futex[0],
 210                                        &worker[i].futex[nfutexes-1], t);
 211                 }
 212
 213                 free(worker[i].futex);
 214         }
 215
 216         print_summary();
 217
 218         free(worker);
 219         return ret;
 220 errmem:
 221         err(EXIT_FAILURE, "calloc");
 222 }