tools/perf/bench/futex-wake-parallel.c

   1 /*
   2  * Copyright (C) 2015 Davidlohr Bueso.
   3  *
   4  * Block a bunch of threads and let parallel waker threads wakeup an
   5  * equal amount of them. The program output reflects the avg latency
   6  * for each individual thread to service its share of work. Ultimately
   7  * it can be used to measure futex_wake() changes.
   8  */
   9
  10 /* For the CLR_() macros */
  11 #include <string.h>
  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 <linux/time64.h>
  20 #include <errno.h>
  21 #include "bench.h"
  22 #include "futex.h"
  23
  24 #include <err.h>
  25 #include <stdlib.h>
  26 #include <sys/time.h>
  27
  28 struct thread_data {
  29         pthread_t worker;
  30         unsigned int nwoken;
  31         struct timeval runtime;
  32 };
  33
  34 static unsigned int nwakes = 1;
  35
  36 /* all threads will block on the same futex -- hash bucket chaos ;) */
  37 static u_int32_t futex = 0;
  38
  39 static pthread_t *blocked_worker;
  40 static bool done = false, silent = false, fshared = false;
  41 static unsigned int nblocked_threads = 0, nwaking_threads = 0;
  42 static pthread_mutex_t thread_lock;
  43 static pthread_cond_t thread_parent, thread_worker;
  44 static struct stats waketime_stats, wakeup_stats;
  45 static unsigned int ncpus, threads_starting;
  46 static int futex_flag = 0;
  47
  48 static const struct option options[] = {
  49         OPT_UINTEGER('t', "threads", &nblocked_threads, "Specify amount of threads"),
  50         OPT_UINTEGER('w', "nwakers", &nwaking_threads, "Specify amount of waking threads"),
  51         OPT_BOOLEAN( 's', "silent",  &silent,   "Silent mode: do not display data/details"),
  52         OPT_BOOLEAN( 'S', "shared",  &fshared,  "Use shared futexes instead of private ones"),
  53         OPT_END()
  54 };
  55
  56 static const char * const bench_futex_wake_parallel_usage[] = {
  57         "perf bench futex wake-parallel <options>",
  58         NULL
  59 };
  60
  61 static void *waking_workerfn(void *arg)
  62 {
  63         struct thread_data *waker = (struct thread_data *) arg;
  64         struct timeval start, end;
  65
  66         gettimeofday(&start, NULL);
  67
  68         waker->nwoken = futex_wake(&futex, nwakes, futex_flag);
  69         if (waker->nwoken != nwakes)
  70                 warnx("couldn't wakeup all tasks (%d/%d)",
  71                       waker->nwoken, nwakes);
  72
  73         gettimeofday(&end, NULL);
  74         timersub(&end, &start, &waker->runtime);
  75
  76         pthread_exit(NULL);
  77         return NULL;
  78 }
  79
  80 static void wakeup_threads(struct thread_data *td, pthread_attr_t thread_attr)
  81 {
  82         unsigned int i;
  83
  84         pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
  85
  86         /* create and block all threads */
  87         for (i = 0; i < nwaking_threads; i++) {
  88                 /*
  89                  * Thread creation order will impact per-thread latency
  90                  * as it will affect the order to acquire the hb spinlock.
  91                  * For now let the scheduler decide.
  92                  */
  93                 if (pthread_create(&td[i].worker, &thread_attr,
  94                                    waking_workerfn, (void *)&td[i]))
  95                         err(EXIT_FAILURE, "pthread_create");
  96         }
  97
  98         for (i = 0; i < nwaking_threads; i++)
  99                 if (pthread_join(td[i].worker, NULL))
 100                         err(EXIT_FAILURE, "pthread_join");
 101 }
 102
 103 static void *blocked_workerfn(void *arg __maybe_unused)
 104 {
 105         pthread_mutex_lock(&thread_lock);
 106         threads_starting--;
 107         if (!threads_starting)
 108                 pthread_cond_signal(&thread_parent);
 109         pthread_cond_wait(&thread_worker, &thread_lock);
 110         pthread_mutex_unlock(&thread_lock);
 111
 112         while (1) { /* handle spurious wakeups */
 113                 if (futex_wait(&futex, 0, NULL, futex_flag) != EINTR)
 114                         break;
 115         }
 116
 117         pthread_exit(NULL);
 118         return NULL;
 119 }
 120
 121 static void block_threads(pthread_t *w, pthread_attr_t thread_attr)
 122 {
 123         cpu_set_t cpu;
 124         unsigned int i;
 125
 126         threads_starting = nblocked_threads;
 127
 128         /* create and block all threads */
 129         for (i = 0; i < nblocked_threads; i++) {
 130                 CPU_ZERO(&cpu);
 131                 CPU_SET(i % ncpus, &cpu);
 132
 133                 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu))
 134                         err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
 135
 136                 if (pthread_create(&w[i], &thread_attr, blocked_workerfn, NULL))
 137                         err(EXIT_FAILURE, "pthread_create");
 138         }
 139 }
 140
 141 static void print_run(struct thread_data *waking_worker, unsigned int run_num)
 142 {
 143         unsigned int i, wakeup_avg;
 144         double waketime_avg, waketime_stddev;
 145         struct stats __waketime_stats, __wakeup_stats;
 146
 147         init_stats(&__wakeup_stats);
 148         init_stats(&__waketime_stats);
 149
 150         for (i = 0; i < nwaking_threads; i++) {
 151                 update_stats(&__waketime_stats, waking_worker[i].runtime.tv_usec);
 152                 update_stats(&__wakeup_stats, waking_worker[i].nwoken);
 153         }
 154
 155         waketime_avg = avg_stats(&__waketime_stats);
 156         waketime_stddev = stddev_stats(&__waketime_stats);
 157         wakeup_avg = avg_stats(&__wakeup_stats);
 158
 159         printf("[Run %d]: Avg per-thread latency (waking %d/%d threads) "
 160                "in %.4f ms (+-%.2f%%)\n", run_num + 1, wakeup_avg,
 161                nblocked_threads, waketime_avg / USEC_PER_MSEC,
 162                rel_stddev_stats(waketime_stddev, waketime_avg));
 163 }
 164
 165 static void print_summary(void)
 166 {
 167         unsigned int wakeup_avg;
 168         double waketime_avg, waketime_stddev;
 169
 170         waketime_avg = avg_stats(&waketime_stats);
 171         waketime_stddev = stddev_stats(&waketime_stats);
 172         wakeup_avg = avg_stats(&wakeup_stats);
 173
 174         printf("Avg per-thread latency (waking %d/%d threads) in %.4f ms (+-%.2f%%)\n",
 175                wakeup_avg,
 176                nblocked_threads,
 177                waketime_avg / USEC_PER_MSEC,
 178                rel_stddev_stats(waketime_stddev, waketime_avg));
 179 }
 180
 181
 182 static void do_run_stats(struct thread_data *waking_worker)
 183 {
 184         unsigned int i;
 185
 186         for (i = 0; i < nwaking_threads; i++) {
 187                 update_stats(&waketime_stats, waking_worker[i].runtime.tv_usec);
 188                 update_stats(&wakeup_stats, waking_worker[i].nwoken);
 189         }
 190
 191 }
 192
 193 static void toggle_done(int sig __maybe_unused,
 194                         siginfo_t *info __maybe_unused,
 195                         void *uc __maybe_unused)
 196 {
 197         done = true;
 198 }
 199
 200 int bench_futex_wake_parallel(int argc, const char **argv,
 201                               const char *prefix __maybe_unused)
 202 {
 203         int ret = 0;
 204         unsigned int i, j;
 205         struct sigaction act;
 206         pthread_attr_t thread_attr;
 207         struct thread_data *waking_worker;
 208
 209         argc = parse_options(argc, argv, options,
 210                              bench_futex_wake_parallel_usage, 0);
 211         if (argc) {
 212                 usage_with_options(bench_futex_wake_parallel_usage, options);
 213                 exit(EXIT_FAILURE);
 214         }
 215
 216         sigfillset(&act.sa_mask);
 217         act.sa_sigaction = toggle_done;
 218         sigaction(SIGINT, &act, NULL);
 219
 220         ncpus = sysconf(_SC_NPROCESSORS_ONLN);
 221         if (!nblocked_threads)
 222                 nblocked_threads = ncpus;
 223
 224         /* some sanity checks */
 225         if (nwaking_threads > nblocked_threads || !nwaking_threads)
 226                 nwaking_threads = nblocked_threads;
 227
 228         if (nblocked_threads % nwaking_threads)
 229                 errx(EXIT_FAILURE, "Must be perfectly divisible");
 230         /*
 231          * Each thread will wakeup nwakes tasks in
 232          * a single futex_wait call.
 233          */
 234         nwakes = nblocked_threads/nwaking_threads;
 235
 236         blocked_worker = calloc(nblocked_threads, sizeof(*blocked_worker));
 237         if (!blocked_worker)
 238                 err(EXIT_FAILURE, "calloc");
 239
 240         if (!fshared)
 241                 futex_flag = FUTEX_PRIVATE_FLAG;
 242
 243         printf("Run summary [PID %d]: blocking on %d threads (at [%s] "
 244                "futex %p), %d threads waking up %d at a time.\n\n",
 245                getpid(), nblocked_threads, fshared ? "shared":"private",
 246                &futex, nwaking_threads, nwakes);
 247
 248         init_stats(&wakeup_stats);
 249         init_stats(&waketime_stats);
 250
 251         pthread_attr_init(&thread_attr);
 252         pthread_mutex_init(&thread_lock, NULL);
 253         pthread_cond_init(&thread_parent, NULL);
 254         pthread_cond_init(&thread_worker, NULL);
 255
 256         for (j = 0; j < bench_repeat && !done; j++) {
 257                 waking_worker = calloc(nwaking_threads, sizeof(*waking_worker));
 258                 if (!waking_worker)
 259                         err(EXIT_FAILURE, "calloc");
 260
 261                 /* create, launch & block all threads */
 262                 block_threads(blocked_worker, thread_attr);
 263
 264                 /* make sure all threads are already blocked */
 265                 pthread_mutex_lock(&thread_lock);
 266                 while (threads_starting)
 267                         pthread_cond_wait(&thread_parent, &thread_lock);
 268                 pthread_cond_broadcast(&thread_worker);
 269                 pthread_mutex_unlock(&thread_lock);
 270
 271                 usleep(100000);
 272
 273                 /* Ok, all threads are patiently blocked, start waking folks up */
 274                 wakeup_threads(waking_worker, thread_attr);
 275
 276                 for (i = 0; i < nblocked_threads; i++) {
 277                         ret = pthread_join(blocked_worker[i], NULL);
 278                         if (ret)
 279                                 err(EXIT_FAILURE, "pthread_join");
 280                 }
 281
 282                 do_run_stats(waking_worker);
 283                 if (!silent)
 284                         print_run(waking_worker, j);
 285
 286                 free(waking_worker);
 287         }
 288
 289         /* cleanup & report results */
 290         pthread_cond_destroy(&thread_parent);
 291         pthread_cond_destroy(&thread_worker);
 292         pthread_mutex_destroy(&thread_lock);
 293         pthread_attr_destroy(&thread_attr);
 294
 295         print_summary();
 296
 297         free(blocked_worker);
 298         return ret;
 299 }