tools/perf/bench/futex-hash.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2013  Davidlohr Bueso <davidlohr@hp.com>
   4  *
   5  * futex-hash: Stress the hell out of the Linux kernel futex uaddr hashing.
   6  *
   7  * This program is particularly useful for measuring the kernel's futex hash
   8  * table/function implementation. In order for it to make sense, use with as
   9  * many threads and futexes as possible.
  10  */
  11
  12 /* For the CLR_() macros */
  13 #include <string.h>
  14 #include <pthread.h>
  15
  16 #include <errno.h>
  17 #include <signal.h>
  18 #include <stdlib.h>
  19 #include <linux/compiler.h>
  20 #include <linux/kernel.h>
  21 #include <sys/time.h>
  22
  23 #include "../util/stat.h"
  24 #include <subcmd/parse-options.h>
  25 #include "bench.h"
  26 #include "futex.h"
  27
  28 #include <err.h>
  29 #include <sys/time.h>
  30
  31 static unsigned int nthreads = 0;
  32 static unsigned int nsecs    = 10;
  33 /* amount of futexes per thread */
  34 static unsigned int nfutexes = 1024;
  35 static bool fshared = false, done = false, silent = false;
  36 static int futex_flag = 0;
  37
  38 struct timeval start, end, runtime;
  39 static pthread_mutex_t thread_lock;
  40 static unsigned int threads_starting;
  41 static struct stats throughput_stats;
  42 static pthread_cond_t thread_parent, thread_worker;
  43
  44 struct worker {
  45         int tid;
  46         u_int32_t *futex;
  47         pthread_t thread;
  48         unsigned long ops;
  49 };
  50
  51 static const struct option options[] = {
  52         OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
  53         OPT_UINTEGER('r', "runtime", &nsecs,    "Specify runtime (in seconds)"),
  54         OPT_UINTEGER('f', "futexes", &nfutexes, "Specify amount of futexes per threads"),
  55         OPT_BOOLEAN( 's', "silent",  &silent,   "Silent mode: do not display data/details"),
  56         OPT_BOOLEAN( 'S', "shared",  &fshared,  "Use shared futexes instead of private ones"),
  57         OPT_END()
  58 };
  59
  60 static const char * const bench_futex_hash_usage[] = {
  61         "perf bench futex hash <options>",
  62         NULL
  63 };
  64
  65 static void *workerfn(void *arg)
  66 {
  67         int ret;
  68         struct worker *w = (struct worker *) arg;
  69         unsigned int i;
  70         unsigned long ops = w->ops; /* avoid cacheline bouncing */
  71
  72         pthread_mutex_lock(&thread_lock);
  73         threads_starting--;
  74         if (!threads_starting)
  75                 pthread_cond_signal(&thread_parent);
  76         pthread_cond_wait(&thread_worker, &thread_lock);
  77         pthread_mutex_unlock(&thread_lock);
  78
  79         do {
  80                 for (i = 0; i < nfutexes; i++, ops++) {
  81                         /*
  82                          * We want the futex calls to fail in order to stress
  83                          * the hashing of uaddr and not measure other steps,
  84                          * such as internal waitqueue handling, thus enlarging
  85                          * the critical region protected by hb->lock.
  86                          */
  87                         ret = futex_wait(&w->futex[i], 1234, NULL, futex_flag);
  88                         if (!silent &&
  89                             (!ret || errno != EAGAIN || errno != EWOULDBLOCK))
  90                                 warn("Non-expected futex return call");
  91                 }
  92         }  while (!done);
  93
  94         w->ops = ops;
  95         return NULL;
  96 }
  97
  98 static void toggle_done(int sig __maybe_unused,
  99                         siginfo_t *info __maybe_unused,
 100                         void *uc __maybe_unused)
 101 {
 102         /* inform all threads that we're done for the day */
 103         done = true;
 104         gettimeofday(&end, NULL);
 105         timersub(&end, &start, &runtime);
 106 }
 107
 108 static void print_summary(void)
 109 {
 110         unsigned long avg = avg_stats(&throughput_stats);
 111         double stddev = stddev_stats(&throughput_stats);
 112
 113         printf("%sAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n",
 114                !silent ? "\n" : "", avg, rel_stddev_stats(stddev, avg),
 115                (int) runtime.tv_sec);
 116 }
 117
 118 int bench_futex_hash(int argc, const char **argv)
 119 {
 120         int ret = 0;
 121         cpu_set_t cpu;
 122         struct sigaction act;
 123         unsigned int i, ncpus;
 124         pthread_attr_t thread_attr;
 125         struct worker *worker = NULL;
 126
 127         argc = parse_options(argc, argv, options, bench_futex_hash_usage, 0);
 128         if (argc) {
 129                 usage_with_options(bench_futex_hash_usage, options);
 130                 exit(EXIT_FAILURE);
 131         }
 132
 133         ncpus = sysconf(_SC_NPROCESSORS_ONLN);
 134
 135         sigfillset(&act.sa_mask);
 136         act.sa_sigaction = toggle_done;
 137         sigaction(SIGINT, &act, NULL);
 138
 139         if (!nthreads) /* default to the number of CPUs */
 140                 nthreads = ncpus;
 141
 142         worker = calloc(nthreads, sizeof(*worker));
 143         if (!worker)
 144                 goto errmem;
 145
 146         if (!fshared)
 147                 futex_flag = FUTEX_PRIVATE_FLAG;
 148
 149         printf("Run summary [PID %d]: %d threads, each operating on %d [%s] futexes for %d secs.\n\n",
 150                getpid(), nthreads, nfutexes, fshared ? "shared":"private", nsecs);
 151
 152         init_stats(&throughput_stats);
 153         pthread_mutex_init(&thread_lock, NULL);
 154         pthread_cond_init(&thread_parent, NULL);
 155         pthread_cond_init(&thread_worker, NULL);
 156
 157         threads_starting = nthreads;
 158         pthread_attr_init(&thread_attr);
 159         gettimeofday(&start, NULL);
 160         for (i = 0; i < nthreads; i++) {
 161                 worker[i].tid = i;
 162                 worker[i].futex = calloc(nfutexes, sizeof(*worker[i].futex));
 163                 if (!worker[i].futex)
 164                         goto errmem;
 165
 166                 CPU_ZERO(&cpu);
 167                 CPU_SET(i % ncpus, &cpu);
 168
 169                 ret = pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu);
 170                 if (ret)
 171                         err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
 172
 173                 ret = pthread_create(&worker[i].thread, &thread_attr, workerfn,
 174                                      (void *)(struct worker *) &worker[i]);
 175                 if (ret)
 176                         err(EXIT_FAILURE, "pthread_create");
 177
 178         }
 179         pthread_attr_destroy(&thread_attr);
 180
 181         pthread_mutex_lock(&thread_lock);
 182         while (threads_starting)
 183                 pthread_cond_wait(&thread_parent, &thread_lock);
 184         pthread_cond_broadcast(&thread_worker);
 185         pthread_mutex_unlock(&thread_lock);
 186
 187         sleep(nsecs);
 188         toggle_done(0, NULL, NULL);
 189
 190         for (i = 0; i < nthreads; i++) {
 191                 ret = pthread_join(worker[i].thread, NULL);
 192                 if (ret)
 193                         err(EXIT_FAILURE, "pthread_join");
 194         }
 195
 196         /* cleanup & report results */
 197         pthread_cond_destroy(&thread_parent);
 198         pthread_cond_destroy(&thread_worker);
 199         pthread_mutex_destroy(&thread_lock);
 200
 201         for (i = 0; i < nthreads; i++) {
 202                 unsigned long t = worker[i].ops/runtime.tv_sec;
 203                 update_stats(&throughput_stats, t);
 204                 if (!silent) {
 205                         if (nfutexes == 1)
 206                                 printf("[thread %2d] futex: %p [ %ld ops/sec ]\n",
 207                                        worker[i].tid, &worker[i].futex[0], t);
 208                         else
 209                                 printf("[thread %2d] futexes: %p ... %p [ %ld ops/sec ]\n",
 210                                        worker[i].tid, &worker[i].futex[0],
 211                                        &worker[i].futex[nfutexes-1], t);
 212                 }
 213
 214                 free(worker[i].futex);
 215         }
 216
 217         print_summary();
 218
 219         free(worker);
 220         return ret;
 221 errmem:
 222         err(EXIT_FAILURE, "calloc");
 223 }