Quelle  futex-wake-parallel.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0
/*
 * Copyright (C) 2015 Davidlohr Bueso.
 *
 * Block a bunch of threads and let parallel waker threads wakeup an
 * equal amount of them. The program output reflects the avg latency
 * for each individual thread to service its share of work. Ultimately
 * it can be used to measure futex_wake() changes.
 */
#include "bench.h"
#include <linux/compiler.h>
#include "../util/debug.h"
#include "../util/mutex.h"

#ifndef HAVE_PTHREAD_BARRIER
int bench_futex_wake_parallel(int argc __maybe_unused, const char **argv __maybe_unused)
{
 pr_err("%s: pthread_barrier_t unavailable, disabling this test...\n", __func__);
 return 0;
}
#else /* HAVE_PTHREAD_BARRIER */
/* For the CLR_() macros */
#include <string.h>
#include <pthread.h>

#include <signal.h>
#include "../util/stat.h"
#include <subcmd/parse-options.h>
#include <linux/kernel.h>
#include <linux/time64.h>
#include <errno.h>
#include "futex.h"
#include <perf/cpumap.h>

#include <err.h>
#include <stdlib.h>
#include <sys/time.h>
#include <sys/mman.h>

struct thread_data {
 pthread_t worker;
 unsigned int nwoken;
 struct timeval runtime;
};

static unsigned int nwakes = 1;

/* all threads will block on the same futex -- hash bucket chaos ;) */
static u_int32_t futex = 0;

static pthread_t *blocked_worker;
static bool done = false;
static struct mutex thread_lock;
static struct cond thread_parent, thread_worker;
static pthread_barrier_t barrier;
static struct stats waketime_stats, wakeup_stats;
static unsigned int threads_starting;
static int futex_flag = 0;

static struct bench_futex_parameters params = {
 .nbuckets = -1,
};

static const struct option options[] = {
 OPT_INTEGER( 'b', "buckets", ¶ms.nbuckets, "Specify amount of hash buckets"),
 OPT_UINTEGER('t', "threads", ¶ms.nthreads, "Specify amount of threads"),
 OPT_UINTEGER('w', "nwakers", ¶ms.nwakes, "Specify amount of waking threads"),
 OPT_BOOLEAN( 's', "silent",  ¶ms.silent, "Silent mode: do not display data/details"),
 OPT_BOOLEAN( 'S', "shared",  ¶ms.fshared, "Use shared futexes instead of private ones"),
 OPT_BOOLEAN( 'm', "mlockall", ¶ms.mlockall, "Lock all current and future memory"),

 OPT_END()
};

static const char * const bench_futex_wake_parallel_usage[] = {
 "perf bench futex wake-parallel ",
 NULL
};

static void *waking_workerfn(void *arg)
{
 struct thread_data *waker = (struct thread_data *) arg;
 struct timeval start, end;

 pthread_barrier_wait(&barrier);

 gettimeofday(&start, NULL);

 waker->nwoken = futex_wake(&futex, nwakes, futex_flag);
 if (waker->nwoken != nwakes)
  warnx("couldn't wakeup all tasks (%d/%d)",
        waker->nwoken, nwakes);

 gettimeofday(&end, NULL);
 timersub(&end, &start, &waker->runtime);

 pthread_exit(NULL);
 return NULL;
}

static void wakeup_threads(struct thread_data *td)
{
 unsigned int i;
 pthread_attr_t thread_attr;

 pthread_attr_init(&thread_attr);
 pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);

 pthread_barrier_init(&barrier, NULL, params.nwakes + 1);

 /* create and block all threads */
 for (i = 0; i < params.nwakes; i++) {
  /*
 * Thread creation order will impact per-thread latency
 * as it will affect the order to acquire the hb spinlock.
 * For now let the scheduler decide.
 */
  if (pthread_create(&td[i].worker, &thread_attr,
       waking_workerfn, (void *)&td[i]))
   err(EXIT_FAILURE, "pthread_create");
 }

 pthread_barrier_wait(&barrier);

 for (i = 0; i < params.nwakes; i++)
  if (pthread_join(td[i].worker, NULL))
   err(EXIT_FAILURE, "pthread_join");

 pthread_barrier_destroy(&barrier);
 pthread_attr_destroy(&thread_attr);
}

static void *blocked_workerfn(void *arg __maybe_unused)
{
 mutex_lock(&thread_lock);
 threads_starting--;
 if (!threads_starting)
  cond_signal(&thread_parent);
 cond_wait(&thread_worker, &thread_lock);
 mutex_unlock(&thread_lock);

 while (1) { /* handle spurious wakeups */
  if (futex_wait(&futex, 0, NULL, futex_flag) != EINTR)
   break;
 }

 pthread_exit(NULL);
 return NULL;
}

static void block_threads(pthread_t *w, struct perf_cpu_map *cpu)
{
 cpu_set_t *cpuset;
 unsigned int i;
 int nrcpus = cpu__max_cpu().cpu;
 size_t size;

 threads_starting = params.nthreads;

 cpuset = CPU_ALLOC(nrcpus);
 BUG_ON(!cpuset);
 size = CPU_ALLOC_SIZE(nrcpus);

 /* create and block all threads */
 for (i = 0; i < params.nthreads; i++) {
  pthread_attr_t thread_attr;

  pthread_attr_init(&thread_attr);
  CPU_ZERO_S(size, cpuset);
  CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset);

  if (pthread_attr_setaffinity_np(&thread_attr, size, cpuset)) {
   CPU_FREE(cpuset);
   err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
  }

  if (pthread_create(&w[i], &thread_attr, blocked_workerfn, NULL)) {
   CPU_FREE(cpuset);
   err(EXIT_FAILURE, "pthread_create");
  }
  pthread_attr_destroy(&thread_attr);
 }
 CPU_FREE(cpuset);
}

static void print_run(struct thread_data *waking_worker, unsigned int run_num)
{
 unsigned int i, wakeup_avg;
 double waketime_avg, waketime_stddev;
 struct stats __waketime_stats, __wakeup_stats;

 init_stats(&__wakeup_stats);
 init_stats(&__waketime_stats);

 for (i = 0; i < params.nwakes; i++) {
  update_stats(&__waketime_stats, waking_worker[i].runtime.tv_usec);
  update_stats(&__wakeup_stats, waking_worker[i].nwoken);
 }

 waketime_avg = avg_stats(&__waketime_stats);
 waketime_stddev = stddev_stats(&__waketime_stats);
 wakeup_avg = avg_stats(&__wakeup_stats);

 printf("[Run %d]: Avg per-thread latency (waking %d/%d threads) "
        "in %.4f ms (+-%.2f%%)\n", run_num + 1, wakeup_avg,
        params.nthreads, waketime_avg / USEC_PER_MSEC,
        rel_stddev_stats(waketime_stddev, waketime_avg));
}

static void print_summary(void)
{
 unsigned int wakeup_avg;
 double waketime_avg, waketime_stddev;

 waketime_avg = avg_stats(&waketime_stats);
 waketime_stddev = stddev_stats(&waketime_stats);
 wakeup_avg = avg_stats(&wakeup_stats);

 printf("Avg per-thread latency (waking %d/%d threads) in %.4f ms (+-%.2f%%)\n",
        wakeup_avg,
        params.nthreads,
        waketime_avg / USEC_PER_MSEC,
        rel_stddev_stats(waketime_stddev, waketime_avg));
 futex_print_nbuckets(¶ms);
}


static void do_run_stats(struct thread_data *waking_worker)
{
 unsigned int i;

 for (i = 0; i < params.nwakes; i++) {
  update_stats(&waketime_stats, waking_worker[i].runtime.tv_usec);
  update_stats(&wakeup_stats, waking_worker[i].nwoken);
 }

}

static void toggle_done(int sig __maybe_unused,
   siginfo_t *info __maybe_unused,
   void *uc __maybe_unused)
{
 done = true;
}

int bench_futex_wake_parallel(int argc, const char **argv)
{
 int ret = 0;
 unsigned int i, j;
 struct sigaction act;
 struct thread_data *waking_worker;
 struct perf_cpu_map *cpu;

 argc = parse_options(argc, argv, options,
        bench_futex_wake_parallel_usage, 0);
 if (argc) {
  usage_with_options(bench_futex_wake_parallel_usage, options);
  exit(EXIT_FAILURE);
 }

 memset(&act, 0, sizeof(act));
 sigfillset(&act.sa_mask);
 act.sa_sigaction = toggle_done;
 sigaction(SIGINT, &act, NULL);

 if (params.mlockall) {
  if (mlockall(MCL_CURRENT | MCL_FUTURE))
   err(EXIT_FAILURE, "mlockall");
 }

 cpu = perf_cpu_map__new_online_cpus();
 if (!cpu)
  err(EXIT_FAILURE, "calloc");

 if (!params.nthreads)
  params.nthreads = perf_cpu_map__nr(cpu);

 /* some sanity checks */
 if (params.nwakes > params.nthreads ||
     !params.nwakes)
  params.nwakes = params.nthreads;

 if (params.nthreads % params.nwakes)
  errx(EXIT_FAILURE, "Must be perfectly divisible");
 /*
 * Each thread will wakeup nwakes tasks in
 * a single futex_wait call.
 */
 nwakes = params.nthreads/params.nwakes;

 blocked_worker = calloc(params.nthreads, sizeof(*blocked_worker));
 if (!blocked_worker)
  err(EXIT_FAILURE, "calloc");

 if (!params.fshared)
  futex_flag = FUTEX_PRIVATE_FLAG;

 futex_set_nbuckets_param(¶ms);

 printf("Run summary [PID %d]: blocking on %d threads (at [%s] "
        "futex %p), %d threads waking up %d at a time.\n\n",
        getpid(), params.nthreads, params.fshared ? "shared":"private",
        &futex, params.nwakes, nwakes);

 init_stats(&wakeup_stats);
 init_stats(&waketime_stats);

 mutex_init(&thread_lock);
 cond_init(&thread_parent);
 cond_init(&thread_worker);

 for (j = 0; j < bench_repeat && !done; j++) {
  waking_worker = calloc(params.nwakes, sizeof(*waking_worker));
  if (!waking_worker)
   err(EXIT_FAILURE, "calloc");

  /* create, launch & block all threads */
  block_threads(blocked_worker, cpu);

  /* make sure all threads are already blocked */
  mutex_lock(&thread_lock);
  while (threads_starting)
   cond_wait(&thread_parent, &thread_lock);
  cond_broadcast(&thread_worker);
  mutex_unlock(&thread_lock);

  usleep(200000);

  /* Ok, all threads are patiently blocked, start waking folks up */
  wakeup_threads(waking_worker);

  for (i = 0; i < params.nthreads; i++) {
   ret = pthread_join(blocked_worker[i], NULL);
   if (ret)
    err(EXIT_FAILURE, "pthread_join");
  }

  do_run_stats(waking_worker);
  if (!params.silent)
   print_run(waking_worker, j);

  free(waking_worker);
 }

 /* cleanup & report results */
 cond_destroy(&thread_parent);
 cond_destroy(&thread_worker);
 mutex_destroy(&thread_lock);

 print_summary();

 free(blocked_worker);
 perf_cpu_map__put(cpu);
 return ret;
}
#endif /* HAVE_PTHREAD_BARRIER */