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 <linux/zalloc.h>
  22 #include <sys/time.h>
  23 #include <sys/mman.h>
  24 #include <perf/cpumap.h>
  25
  26 #include "../util/stat.h"
  27 #include <subcmd/parse-options.h>
  28 #include "bench.h"
  29 #include "futex.h"
  30
  31 #include <err.h>
  32
  33 static bool done = false;
  34 static int futex_flag = 0;
  35
  36 struct timeval bench__start, bench__end, bench__runtime;
  37 static pthread_mutex_t thread_lock;
  38 static unsigned int threads_starting;
  39 static struct stats throughput_stats;
  40 static pthread_cond_t thread_parent, thread_worker;
  41
  42 struct worker {
  43         int tid;
  44         u_int32_t *futex;
  45         pthread_t thread;
  46         unsigned long ops;
  47 };
  48
  49 static struct bench_futex_parameters params = {
  50         .nfutexes = 1024,
  51         .runtime  = 10,
  52 };
  53
  54 static const struct option options[] = {
  55         OPT_UINTEGER('t', "threads", &params.nthreads, "Specify amount of threads"),
  56         OPT_UINTEGER('r', "runtime", &params.runtime, "Specify runtime (in seconds)"),
  57         OPT_UINTEGER('f', "futexes", &params.nfutexes, "Specify amount of futexes per threads"),
  58         OPT_BOOLEAN( 's', "silent",  &params.silent, "Silent mode: do not display data/details"),
  59         OPT_BOOLEAN( 'S', "shared",  &params.fshared, "Use shared futexes instead of private ones"),
  60         OPT_BOOLEAN( 'm', "mlockall", &params.mlockall, "Lock all current and future memory"),
  61         OPT_END()
  62 };
  63
  64 static const char * const bench_futex_hash_usage[] = {
  65         "perf bench futex hash <options>",
  66         NULL
  67 };
  68
  69 static void *workerfn(void *arg)
  70 {
  71         int ret;
  72         struct worker *w = (struct worker *) arg;
  73         unsigned int i;
  74         unsigned long ops = w->ops; /* avoid cacheline bouncing */
  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         do {
  84                 for (i = 0; i < params.nfutexes; i++, ops++) {
  85                         /*
  86                          * We want the futex calls to fail in order to stress
  87                          * the hashing of uaddr and not measure other steps,
  88                          * such as internal waitqueue handling, thus enlarging
  89                          * the critical region protected by hb->lock.
  90                          */
  91                         ret = futex_wait(&w->futex[i], 1234, NULL, futex_flag);
  92                         if (!params.silent &&
  93                             (!ret || errno != EAGAIN || errno != EWOULDBLOCK))
  94                                 warn("Non-expected futex return call");
  95                 }
  96         }  while (!done);
  97
  98         w->ops = ops;
  99         return NULL;
 100 }
 101
 102 static void toggle_done(int sig __maybe_unused,
 103                         siginfo_t *info __maybe_unused,
 104                         void *uc __maybe_unused)
 105 {
 106         /* inform all threads that we're done for the day */
 107         done = true;
 108         gettimeofday(&bench__end, NULL);
 109         timersub(&bench__end, &bench__start, &bench__runtime);
 110 }
 111
 112 static void print_summary(void)
 113 {
 114         unsigned long avg = avg_stats(&throughput_stats);
 115         double stddev = stddev_stats(&throughput_stats);
 116
 117         printf("%sAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n",
 118                !params.silent ? "\n" : "", avg, rel_stddev_stats(stddev, avg),
 119                (int)bench__runtime.tv_sec);
 120 }
 121
 122 int bench_futex_hash(int argc, const char **argv)
 123 {
 124         int ret = 0;
 125         cpu_set_t cpuset;
 126         struct sigaction act;
 127         unsigned int i;
 128         pthread_attr_t thread_attr;
 129         struct worker *worker = NULL;
 130         struct perf_cpu_map *cpu;
 131
 132         argc = parse_options(argc, argv, options, bench_futex_hash_usage, 0);
 133         if (argc) {
 134                 usage_with_options(bench_futex_hash_usage, options);
 135                 exit(EXIT_FAILURE);
 136         }
 137
 138         cpu = perf_cpu_map__new(NULL);
 139         if (!cpu)
 140                 goto errmem;
 141
 142         memset(&act, 0, sizeof(act));
 143         sigfillset(&act.sa_mask);
 144         act.sa_sigaction = toggle_done;
 145         sigaction(SIGINT, &act, NULL);
 146
 147         if (params.mlockall) {
 148                 if (mlockall(MCL_CURRENT | MCL_FUTURE))
 149                         err(EXIT_FAILURE, "mlockall");
 150         }
 151
 152         if (!params.nthreads) /* default to the number of CPUs */
 153                 params.nthreads = perf_cpu_map__nr(cpu);
 154
 155         worker = calloc(params.nthreads, sizeof(*worker));
 156         if (!worker)
 157                 goto errmem;
 158
 159         if (!params.fshared)
 160                 futex_flag = FUTEX_PRIVATE_FLAG;
 161
 162         printf("Run summary [PID %d]: %d threads, each operating on %d [%s] futexes for %d secs.\n\n",
 163                getpid(), params.nthreads, params.nfutexes, params.fshared ? "shared":"private", params.runtime);
 164
 165         init_stats(&throughput_stats);
 166         pthread_mutex_init(&thread_lock, NULL);
 167         pthread_cond_init(&thread_parent, NULL);
 168         pthread_cond_init(&thread_worker, NULL);
 169
 170         threads_starting = params.nthreads;
 171         pthread_attr_init(&thread_attr);
 172         gettimeofday(&bench__start, NULL);
 173         for (i = 0; i < params.nthreads; i++) {
 174                 worker[i].tid = i;
 175                 worker[i].futex = calloc(params.nfutexes, sizeof(*worker[i].futex));
 176                 if (!worker[i].futex)
 177                         goto errmem;
 178
 179                 CPU_ZERO(&cpuset);
 180                 CPU_SET(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, &cpuset);
 181
 182                 ret = pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset);
 183                 if (ret)
 184                         err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
 185
 186                 ret = pthread_create(&worker[i].thread, &thread_attr, workerfn,
 187                                      (void *)(struct worker *) &worker[i]);
 188                 if (ret)
 189                         err(EXIT_FAILURE, "pthread_create");
 190
 191         }
 192         pthread_attr_destroy(&thread_attr);
 193
 194         pthread_mutex_lock(&thread_lock);
 195         while (threads_starting)
 196                 pthread_cond_wait(&thread_parent, &thread_lock);
 197         pthread_cond_broadcast(&thread_worker);
 198         pthread_mutex_unlock(&thread_lock);
 199
 200         sleep(params.runtime);
 201         toggle_done(0, NULL, NULL);
 202
 203         for (i = 0; i < params.nthreads; i++) {
 204                 ret = pthread_join(worker[i].thread, NULL);
 205                 if (ret)
 206                         err(EXIT_FAILURE, "pthread_join");
 207         }
 208
 209         /* cleanup & report results */
 210         pthread_cond_destroy(&thread_parent);
 211         pthread_cond_destroy(&thread_worker);
 212         pthread_mutex_destroy(&thread_lock);
 213
 214         for (i = 0; i < params.nthreads; i++) {
 215                 unsigned long t = bench__runtime.tv_sec > 0 ?
 216                         worker[i].ops / bench__runtime.tv_sec : 0;
 217                 update_stats(&throughput_stats, t);
 218                 if (!params.silent) {
 219                         if (params.nfutexes == 1)
 220                                 printf("[thread %2d] futex: %p [ %ld ops/sec ]\n",
 221                                        worker[i].tid, &worker[i].futex[0], t);
 222                         else
 223                                 printf("[thread %2d] futexes: %p ... %p [ %ld ops/sec ]\n",
 224                                        worker[i].tid, &worker[i].futex[0],
 225                                        &worker[i].futex[params.nfutexes-1], t);
 226                 }
 227
 228                 zfree(&worker[i].futex);
 229         }
 230
 231         print_summary();
 232
 233         free(worker);
 234         free(cpu);
 235         return ret;
 236 errmem:
 237         err(EXIT_FAILURE, "calloc");
 238 }