Quelle  synthesize.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0
/*
 * Benchmark synthesis of perf events such as at the start of a 'perf
 * record'. Synthesis is done on the current process and the 'dummy' event
 * handlers are invoked that support dump_trace but otherwise do nothing.
 *
 * Copyright 2019 Google LLC.
 */
#include <stdio.h>
#include "bench.h"
#include "../util/debug.h"
#include "../util/session.h"
#include "../util/stat.h"
#include "../util/synthetic-events.h"
#include "../util/target.h"
#include "../util/thread_map.h"
#include "../util/tool.h"
#include "../util/util.h"
#include <linux/atomic.h>
#include <linux/err.h>
#include <linux/time64.h>
#include <subcmd/parse-options.h>

static unsigned int min_threads = 1;
static unsigned int max_threads = UINT_MAX;
static unsigned int single_iterations = 10000;
static unsigned int multi_iterations = 10;
static bool run_st;
static bool run_mt;

static const struct option options[] = {
 OPT_BOOLEAN('s', "st", &run_st, "Run single threaded benchmark"),
 OPT_BOOLEAN('t', "mt", &run_mt, "Run multi-threaded benchmark"),
 OPT_UINTEGER('m', "min-threads", &min_threads,
  "Minimum number of threads in multithreaded bench"),
 OPT_UINTEGER('M', "max-threads", &max_threads,
  "Maximum number of threads in multithreaded bench"),
 OPT_UINTEGER('i', "single-iterations", &single_iterations,
  "Number of iterations used to compute single-threaded average"),
 OPT_UINTEGER('I', "multi-iterations", &multi_iterations,
  "Number of iterations used to compute multi-threaded average"),
 OPT_END()
};

static const char *const bench_usage[] = {
 "perf bench internals synthesize ",
 NULL
};

static atomic_t event_count;

static int process_synthesized_event(const struct perf_tool *tool __maybe_unused,
         union perf_event *event __maybe_unused,
         struct perf_sample *sample __maybe_unused,
         struct machine *machine __maybe_unused)
{
 atomic_inc(&event_count);
 return 0;
}

static int do_run_single_threaded(struct perf_session *session,
    struct perf_thread_map *threads,
    struct target *target, bool data_mmap)
{
 const unsigned int nr_threads_synthesize = 1;
 struct timeval start, end, diff;
 u64 runtime_us;
 unsigned int i;
 double time_average, time_stddev, event_average, event_stddev;
 int err;
 struct stats time_stats, event_stats;

 init_stats(&time_stats);
 init_stats(&event_stats);

 for (i = 0; i < single_iterations; i++) {
  atomic_set(&event_count, 0);
  gettimeofday(&start, NULL);
  err = __machine__synthesize_threads(&session->machines.host,
      NULL,
      target, threads,
      process_synthesized_event,
      true, data_mmap,
      nr_threads_synthesize);
  if (err)
   return err;

  gettimeofday(&end, NULL);
  timersub(&end, &start, &diff);
  runtime_us = diff.tv_sec * USEC_PER_SEC + diff.tv_usec;
  update_stats(&time_stats, runtime_us);
  update_stats(&event_stats, atomic_read(&event_count));
 }

 time_average = avg_stats(&time_stats);
 time_stddev = stddev_stats(&time_stats);
 printf(" Average %ssynthesis took: %.3f usec (+- %.3f usec)\n",
  data_mmap ? "data " : "", time_average, time_stddev);

 event_average = avg_stats(&event_stats);
 event_stddev = stddev_stats(&event_stats);
 printf(" Average num. events: %.3f (+- %.3f)\n",
  event_average, event_stddev);

 printf(" Average time per event %.3f usec\n",
  time_average / event_average);
 return 0;
}

static int run_single_threaded(void)
{
 struct perf_session *session;
 struct target target = {
  .pid = "self",
 };
 struct perf_thread_map *threads;
 struct perf_env host_env;
 int err;

 perf_set_singlethreaded();
 perf_env__init(&host_env);
 session = __perf_session__new(/*data=*/NULL, /*tool=*/NULL,
          /*trace_event_repipe=*/false, &host_env);
 if (IS_ERR(session)) {
  pr_err("Session creation failed.\n");
  perf_env__exit(&host_env);
  return PTR_ERR(session);
 }
 threads = thread_map__new_by_pid(getpid());
 if (!threads) {
  pr_err("Thread map creation failed.\n");
  err = -ENOMEM;
  goto err_out;
 }

 puts(
"Computing performance of single threaded perf event synthesis by\n"
"synthesizing events on the perf process itself:");

 err = do_run_single_threaded(session, threads, &target, false);
 if (err)
  goto err_out;

 err = do_run_single_threaded(session, threads, &target, true);

err_out:
 if (threads)
  perf_thread_map__put(threads);

 perf_session__delete(session);
 perf_env__exit(&host_env);
 return err;
}

static int do_run_multi_threaded(struct target *target,
    unsigned int nr_threads_synthesize)
{
 struct timeval start, end, diff;
 u64 runtime_us;
 unsigned int i;
 double time_average, time_stddev, event_average, event_stddev;
 int err = 0;
 struct stats time_stats, event_stats;
 struct perf_session *session;
 struct perf_env host_env;

 perf_env__init(&host_env);
 init_stats(&time_stats);
 init_stats(&event_stats);
 for (i = 0; i < multi_iterations; i++) {
  session = __perf_session__new(/*data=*/NULL, /*tool=*/NULL,
           /*trace_event_repipe=*/false, &host_env);
  if (IS_ERR(session)) {
   err = PTR_ERR(session);
   goto err_out;
  }
  atomic_set(&event_count, 0);
  gettimeofday(&start, NULL);
  err = __machine__synthesize_threads(&session->machines.host,
      NULL,
      target, NULL,
      process_synthesized_event,
      true, false,
      nr_threads_synthesize);
  if (err) {
   perf_session__delete(session);
   goto err_out;
  }

  gettimeofday(&end, NULL);
  timersub(&end, &start, &diff);
  runtime_us = diff.tv_sec * USEC_PER_SEC + diff.tv_usec;
  update_stats(&time_stats, runtime_us);
  update_stats(&event_stats, atomic_read(&event_count));
  perf_session__delete(session);
 }

 time_average = avg_stats(&time_stats);
 time_stddev = stddev_stats(&time_stats);
 printf(" Average synthesis took: %.3f usec (+- %.3f usec)\n",
  time_average, time_stddev);

 event_average = avg_stats(&event_stats);
 event_stddev = stddev_stats(&event_stats);
 printf(" Average num. events: %.3f (+- %.3f)\n",
  event_average, event_stddev);

 printf(" Average time per event %.3f usec\n",
  time_average / event_average);
err_out:
 perf_env__exit(&host_env);
 return err;
}

static int run_multi_threaded(void)
{
 struct target target = {
  .cpu_list = "0"
 };
 unsigned int nr_threads_synthesize;
 int err;

 if (max_threads == UINT_MAX)
  max_threads = sysconf(_SC_NPROCESSORS_ONLN);

 puts(
"Computing performance of multi threaded perf event synthesis by\n"
"synthesizing events on CPU 0:");

 for (nr_threads_synthesize = min_threads;
      nr_threads_synthesize <= max_threads;
      nr_threads_synthesize++) {
  if (nr_threads_synthesize == 1)
   perf_set_singlethreaded();
  else
   perf_set_multithreaded();

  printf(" Number of synthesis threads: %u\n",
   nr_threads_synthesize);

  err = do_run_multi_threaded(&target, nr_threads_synthesize);
  if (err)
   return err;
 }
 perf_set_singlethreaded();
 return 0;
}

int bench_synthesize(int argc, const char **argv)
{
 int err = 0;

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

 /*
 * If neither single threaded or multi-threaded are specified, default
 * to running just single threaded.
 */
 if (!run_st && !run_mt)
  run_st = true;

 if (run_st)
  err = run_single_threaded();

 if (!err && run_mt)
  err = run_multi_threaded();

 return err;
}