Quelle  bench.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0
/* Copyright (c) 2020 Facebook */
#define _GNU_SOURCE
#include <argp.h>
#include <linux/compiler.h>
#include <sys/time.h>
#include <sched.h>
#include <fcntl.h>
#include <pthread.h>
#include <sys/sysinfo.h>
#include <signal.h>
#include "bench.h"
#include "bpf_util.h"
#include "testing_helpers.h"

struct env env = {
 .warmup_sec = 1,
 .duration_sec = 5,
 .affinity = false,
 .quiet = false,
 .consumer_cnt = 0,
 .producer_cnt = 1,
};

static int libbpf_print_fn(enum libbpf_print_level level,
      const char *format, va_list args)
{
 if (level == LIBBPF_DEBUG && !env.verbose)
  return 0;
 return vfprintf(stderr, format, args);
}

void setup_libbpf(void)
{
 libbpf_set_strict_mode(LIBBPF_STRICT_ALL);
 libbpf_set_print(libbpf_print_fn);
}

void false_hits_report_progress(int iter, struct bench_res *res, long delta_ns)
{
 long total = res->false_hits  + res->hits + res->drops;

 printf("Iter %3d (%7.3lfus): ",
        iter, (delta_ns - 1000000000) / 1000.0);

 printf("%ld false hits of %ld total operations. Percentage = %2.2f %%\n",
        res->false_hits, total, ((float)res->false_hits / total) * 100);
}

void false_hits_report_final(struct bench_res res[], int res_cnt)
{
 long total_hits = 0, total_drops = 0, total_false_hits = 0, total_ops = 0;
 int i;

 for (i = 0; i < res_cnt; i++) {
  total_hits += res[i].hits;
  total_false_hits += res[i].false_hits;
  total_drops += res[i].drops;
 }
 total_ops = total_hits + total_false_hits + total_drops;

 printf("Summary: %ld false hits of %ld total operations. ",
        total_false_hits, total_ops);
 printf("Percentage = %2.2f %%\n",
        ((float)total_false_hits / total_ops) * 100);
}

void hits_drops_report_progress(int iter, struct bench_res *res, long delta_ns)
{
 double hits_per_sec, drops_per_sec;
 double hits_per_prod;

 hits_per_sec = res->hits / 1000000.0 / (delta_ns / 1000000000.0);
 hits_per_prod = hits_per_sec / env.producer_cnt;
 drops_per_sec = res->drops / 1000000.0 / (delta_ns / 1000000000.0);

 printf("Iter %3d (%7.3lfus): ",
        iter, (delta_ns - 1000000000) / 1000.0);

 printf("hits %8.3lfM/s (%7.3lfM/prod), drops %8.3lfM/s, total operations %8.3lfM/s\n",
        hits_per_sec, hits_per_prod, drops_per_sec, hits_per_sec + drops_per_sec);
}

void
grace_period_latency_basic_stats(struct bench_res res[], int res_cnt, struct basic_stats *gp_stat)
{
 int i;

 memset(gp_stat, 0, sizeof(struct basic_stats));

 for (i = 0; i < res_cnt; i++)
  gp_stat->mean += res[i].gp_ns / 1000.0 / (double)res[i].gp_ct / (0.0 + res_cnt);

#define IT_MEAN_DIFF (res[i].gp_ns / 1000.0 / (double)res[i].gp_ct - gp_stat->mean)
 if (res_cnt > 1) {
  for (i = 0; i < res_cnt; i++)
   gp_stat->stddev += (IT_MEAN_DIFF * IT_MEAN_DIFF) / (res_cnt - 1.0);
 }
 gp_stat->stddev = sqrt(gp_stat->stddev);
#undef IT_MEAN_DIFF
}

void
grace_period_ticks_basic_stats(struct bench_res res[], int res_cnt, struct basic_stats *gp_stat)
{
 int i;

 memset(gp_stat, 0, sizeof(struct basic_stats));
 for (i = 0; i < res_cnt; i++)
  gp_stat->mean += res[i].stime / (double)res[i].gp_ct / (0.0 + res_cnt);

#define IT_MEAN_DIFF (res[i].stime / (double)res[i].gp_ct - gp_stat->mean)
 if (res_cnt > 1) {
  for (i = 0; i < res_cnt; i++)
   gp_stat->stddev += (IT_MEAN_DIFF * IT_MEAN_DIFF) / (res_cnt - 1.0);
 }
 gp_stat->stddev = sqrt(gp_stat->stddev);
#undef IT_MEAN_DIFF
}

void hits_drops_report_final(struct bench_res res[], int res_cnt)
{
 int i;
 double hits_mean = 0.0, drops_mean = 0.0, total_ops_mean = 0.0;
 double hits_stddev = 0.0, drops_stddev = 0.0, total_ops_stddev = 0.0;
 double total_ops;

 for (i = 0; i < res_cnt; i++) {
  hits_mean += res[i].hits / 1000000.0 / (0.0 + res_cnt);
  drops_mean += res[i].drops / 1000000.0 / (0.0 + res_cnt);
 }
 total_ops_mean = hits_mean + drops_mean;

 if (res_cnt > 1)  {
  for (i = 0; i < res_cnt; i++) {
   hits_stddev += (hits_mean - res[i].hits / 1000000.0) *
           (hits_mean - res[i].hits / 1000000.0) /
           (res_cnt - 1.0);
   drops_stddev += (drops_mean - res[i].drops / 1000000.0) *
     (drops_mean - res[i].drops / 1000000.0) /
     (res_cnt - 1.0);
   total_ops = res[i].hits + res[i].drops;
   total_ops_stddev += (total_ops_mean - total_ops / 1000000.0) *
     (total_ops_mean - total_ops / 1000000.0) /
     (res_cnt - 1.0);
  }
  hits_stddev = sqrt(hits_stddev);
  drops_stddev = sqrt(drops_stddev);
  total_ops_stddev = sqrt(total_ops_stddev);
 }
 printf("Summary: hits %8.3lf \u00B1 %5.3lfM/s (%7.3lfM/prod), ",
        hits_mean, hits_stddev, hits_mean / env.producer_cnt);
 printf("drops %8.3lf \u00B1 %5.3lfM/s, ",
        drops_mean, drops_stddev);
 printf("total operations %8.3lf \u00B1 %5.3lfM/s\n",
        total_ops_mean, total_ops_stddev);
}

void ops_report_progress(int iter, struct bench_res *res, long delta_ns)
{
 double hits_per_sec, hits_per_prod;

 hits_per_sec = res->hits / 1000000.0 / (delta_ns / 1000000000.0);
 hits_per_prod = hits_per_sec / env.producer_cnt;

 printf("Iter %3d (%7.3lfus): ", iter, (delta_ns - 1000000000) / 1000.0);

 printf("hits %8.3lfM/s (%7.3lfM/prod)\n", hits_per_sec, hits_per_prod);
}

void ops_report_final(struct bench_res res[], int res_cnt)
{
 double hits_mean = 0.0, hits_stddev = 0.0;
 int i;

 for (i = 0; i < res_cnt; i++)
  hits_mean += res[i].hits / 1000000.0 / (0.0 + res_cnt);

 if (res_cnt > 1)  {
  for (i = 0; i < res_cnt; i++)
   hits_stddev += (hits_mean - res[i].hits / 1000000.0) *
           (hits_mean - res[i].hits / 1000000.0) /
           (res_cnt - 1.0);

  hits_stddev = sqrt(hits_stddev);
 }
 printf("Summary: throughput %8.3lf \u00B1 %5.3lf M ops/s (%7.3lfM ops/prod), ",
        hits_mean, hits_stddev, hits_mean / env.producer_cnt);
 printf("latency %8.3lf ns/op\n", 1000.0 / hits_mean * env.producer_cnt);
}

void local_storage_report_progress(int iter, struct bench_res *res,
       long delta_ns)
{
 double important_hits_per_sec, hits_per_sec;
 double delta_sec = delta_ns / 1000000000.0;

 hits_per_sec = res->hits / 1000000.0 / delta_sec;
 important_hits_per_sec = res->important_hits / 1000000.0 / delta_sec;

 printf("Iter %3d (%7.3lfus): ", iter, (delta_ns - 1000000000) / 1000.0);

 printf("hits %8.3lfM/s ", hits_per_sec);
 printf("important_hits %8.3lfM/s\n", important_hits_per_sec);
}

void local_storage_report_final(struct bench_res res[], int res_cnt)
{
 double important_hits_mean = 0.0, important_hits_stddev = 0.0;
 double hits_mean = 0.0, hits_stddev = 0.0;
 int i;

 for (i = 0; i < res_cnt; i++) {
  hits_mean += res[i].hits / 1000000.0 / (0.0 + res_cnt);
  important_hits_mean += res[i].important_hits / 1000000.0 / (0.0 + res_cnt);
 }

 if (res_cnt > 1)  {
  for (i = 0; i < res_cnt; i++) {
   hits_stddev += (hits_mean - res[i].hits / 1000000.0) *
           (hits_mean - res[i].hits / 1000000.0) /
           (res_cnt - 1.0);
   important_hits_stddev +=
           (important_hits_mean - res[i].important_hits / 1000000.0) *
           (important_hits_mean - res[i].important_hits / 1000000.0) /
           (res_cnt - 1.0);
  }

  hits_stddev = sqrt(hits_stddev);
  important_hits_stddev = sqrt(important_hits_stddev);
 }
 printf("Summary: hits throughput %8.3lf \u00B1 %5.3lf M ops/s, ",
        hits_mean, hits_stddev);
 printf("hits latency %8.3lf ns/op, ", 1000.0 / hits_mean);
 printf("important_hits throughput %8.3lf \u00B1 %5.3lf M ops/s\n",
        important_hits_mean, important_hits_stddev);
}

const char *argp_program_version = "benchmark";
const char *argp_program_bug_address = "";
const char argp_program_doc[] =
"benchmark Generic benchmarking framework.\n"
"\n"
"This tool runs benchmarks.\n"
"\n"
"USAGE: benchmark \n"
"\n"
"EXAMPLES:\n"
" # run 'count-local' benchmark with 1 producer and 1 consumer\n"
" benchmark count-local\n"
" # run 'count-local' with 16 producer and 8 consumer thread, pinned to CPUs\n"
" benchmark -p16 -c8 -a count-local\n";

enum {
 ARG_PROD_AFFINITY_SET = 1000,
 ARG_CONS_AFFINITY_SET = 1001,
};

static const struct argp_option opts[] = {
 { "list", 'l', NULL, 0, "List available benchmarks"},
 { "duration", 'd', "SEC", 0, "Duration of benchmark, seconds"},
 { "warmup", 'w', "SEC", 0, "Warm-up period, seconds"},
 { "producers", 'p', "NUM", 0, "Number of producer threads"},
 { "consumers", 'c', "NUM", 0, "Number of consumer threads"},
 { "verbose", 'v', NULL, 0, "Verbose debug output"},
 { "affinity", 'a', NULL, 0, "Set consumer/producer thread affinity"},
 { "quiet", 'q', NULL, 0, "Be more quiet"},
 { "prod-affinity", ARG_PROD_AFFINITY_SET, "CPUSET", 0,
   "Set of CPUs for producer threads; implies --affinity"},
 { "cons-affinity", ARG_CONS_AFFINITY_SET, "CPUSET", 0,
   "Set of CPUs for consumer threads; implies --affinity"},
 {},
};

extern struct argp bench_ringbufs_argp;
extern struct argp bench_bloom_map_argp;
extern struct argp bench_bpf_loop_argp;
extern struct argp bench_local_storage_argp;
extern struct argp bench_local_storage_rcu_tasks_trace_argp;
extern struct argp bench_strncmp_argp;
extern struct argp bench_hashmap_lookup_argp;
extern struct argp bench_local_storage_create_argp;
extern struct argp bench_htab_mem_argp;
extern struct argp bench_trigger_batch_argp;
extern struct argp bench_crypto_argp;
extern struct argp bench_sockmap_argp;

static const struct argp_child bench_parsers[] = {
 { &bench_ringbufs_argp, 0, "Ring buffers benchmark", 0 },
 { &bench_bloom_map_argp, 0, "Bloom filter map benchmark", 0 },
 { &bench_bpf_loop_argp, 0, "bpf_loop helper benchmark", 0 },
 { &bench_local_storage_argp, 0, "local_storage benchmark", 0 },
 { &bench_strncmp_argp, 0, "bpf_strncmp helper benchmark", 0 },
 { &bench_local_storage_rcu_tasks_trace_argp, 0,
  "local_storage RCU Tasks Trace slowdown benchmark", 0 },
 { &bench_hashmap_lookup_argp, 0, "Hashmap lookup benchmark", 0 },
 { &bench_local_storage_create_argp, 0, "local-storage-create benchmark", 0 },
 { &bench_htab_mem_argp, 0, "hash map memory benchmark", 0 },
 { &bench_trigger_batch_argp, 0, "BPF triggering benchmark", 0 },
 { &bench_crypto_argp, 0, "bpf crypto benchmark", 0 },
 { &bench_sockmap_argp, 0, "bpf sockmap benchmark", 0 },
 {},
};

/* Make pos_args global, so that we can run argp_parse twice, if necessary */
static int pos_args;

static error_t parse_arg(int key, char *arg, struct argp_state *state)
{
 switch (key) {
 case 'v':
  env.verbose = true;
  break;
 case 'l':
  env.list = true;
  break;
 case 'd':
  env.duration_sec = strtol(arg, NULL, 10);
  if (env.duration_sec <= 0) {
   fprintf(stderr, "Invalid duration: %s\n", arg);
   argp_usage(state);
  }
  break;
 case 'w':
  env.warmup_sec = strtol(arg, NULL, 10);
  if (env.warmup_sec <= 0) {
   fprintf(stderr, "Invalid warm-up duration: %s\n", arg);
   argp_usage(state);
  }
  break;
 case 'p':
  env.producer_cnt = strtol(arg, NULL, 10);
  if (env.producer_cnt < 0) {
   fprintf(stderr, "Invalid producer count: %s\n", arg);
   argp_usage(state);
  }
  break;
 case 'c':
  env.consumer_cnt = strtol(arg, NULL, 10);
  if (env.consumer_cnt < 0) {
   fprintf(stderr, "Invalid consumer count: %s\n", arg);
   argp_usage(state);
  }
  break;
 case 'a':
  env.affinity = true;
  break;
 case 'q':
  env.quiet = true;
  break;
 case ARG_PROD_AFFINITY_SET:
  env.affinity = true;
  if (parse_num_list(arg, &env.prod_cpus.cpus,
       &env.prod_cpus.cpus_len)) {
   fprintf(stderr, "Invalid format of CPU set for producers.");
   argp_usage(state);
  }
  break;
 case ARG_CONS_AFFINITY_SET:
  env.affinity = true;
  if (parse_num_list(arg, &env.cons_cpus.cpus,
       &env.cons_cpus.cpus_len)) {
   fprintf(stderr, "Invalid format of CPU set for consumers.");
   argp_usage(state);
  }
  break;
 case ARGP_KEY_ARG:
  if (pos_args++) {
   fprintf(stderr,
    "Unrecognized positional argument: %s\n", arg);
   argp_usage(state);
  }
  env.bench_name = strdup(arg);
  break;
 default:
  return ARGP_ERR_UNKNOWN;
 }
 return 0;
}

static void parse_cmdline_args_init(int argc, char **argv)
{
 static const struct argp argp = {
  .options = opts,
  .parser = parse_arg,
  .doc = argp_program_doc,
  .children = bench_parsers,
 };
 if (argp_parse(&argp, argc, argv, 0, NULL, NULL))
  exit(1);
}

static void parse_cmdline_args_final(int argc, char **argv)
{
 struct argp_child bench_parsers[2] = {};
 const struct argp argp = {
  .options = opts,
  .parser = parse_arg,
  .doc = argp_program_doc,
  .children = bench_parsers,
 };

 /* Parse arguments the second time with the correct set of parsers */
 if (bench->argp) {
  bench_parsers[0].argp = bench->argp;
  bench_parsers[0].header = bench->name;
  pos_args = 0;
  if (argp_parse(&argp, argc, argv, 0, NULL, NULL))
   exit(1);
 }
}

static void collect_measurements(long delta_ns);

static __u64 last_time_ns;
static void sigalarm_handler(int signo)
{
 long new_time_ns = get_time_ns();
 long delta_ns = new_time_ns - last_time_ns;

 collect_measurements(delta_ns);

 last_time_ns = new_time_ns;
}

/* set up periodic 1-second timer */
static void setup_timer()
{
 static struct sigaction sigalarm_action = {
  .sa_handler = sigalarm_handler,
 };
 struct itimerval timer_settings = {};
 int err;

 last_time_ns = get_time_ns();
 err = sigaction(SIGALRM, &sigalarm_action, NULL);
 if (err < 0) {
  fprintf(stderr, "failed to install SIGALRM handler: %d\n", -errno);
  exit(1);
 }
 timer_settings.it_interval.tv_sec = 1;
 timer_settings.it_value.tv_sec = 1;
 err = setitimer(ITIMER_REAL, &timer_settings, NULL);
 if (err < 0) {
  fprintf(stderr, "failed to arm interval timer: %d\n", -errno);
  exit(1);
 }
}

static void set_thread_affinity(pthread_t thread, int cpu)
{
 cpu_set_t cpuset;
 int err;

 CPU_ZERO(&cpuset);
 CPU_SET(cpu, &cpuset);
 err = pthread_setaffinity_np(thread, sizeof(cpuset), &cpuset);
 if (err) {
  fprintf(stderr, "setting affinity to CPU #%d failed: %d\n",
   cpu, -err);
  exit(1);
 }
}

static int next_cpu(struct cpu_set *cpu_set)
{
 if (cpu_set->cpus) {
  int i;

  /* find next available CPU */
  for (i = cpu_set->next_cpu; i < cpu_set->cpus_len; i++) {
   if (cpu_set->cpus[i]) {
    cpu_set->next_cpu = i + 1;
    return i;
   }
  }
  fprintf(stderr, "Not enough CPUs specified, need CPU #%d or higher.\n", i);
  exit(1);
 }

 return cpu_set->next_cpu++ % env.nr_cpus;
}

static struct bench_state {
 int res_cnt;
 struct bench_res *results;
 pthread_t *consumers;
 pthread_t *producers;
} state;

const struct bench *bench = NULL;

extern const struct bench bench_count_global;
extern const struct bench bench_count_local;
extern const struct bench bench_rename_base;
extern const struct bench bench_rename_kprobe;
extern const struct bench bench_rename_kretprobe;
extern const struct bench bench_rename_rawtp;
extern const struct bench bench_rename_fentry;
extern const struct bench bench_rename_fexit;

/* pure counting benchmarks to establish theoretical limits */
extern const struct bench bench_trig_usermode_count;
extern const struct bench bench_trig_syscall_count;
extern const struct bench bench_trig_kernel_count;

/* batched, staying mostly in-kernel benchmarks */
extern const struct bench bench_trig_kprobe;
extern const struct bench bench_trig_kretprobe;
extern const struct bench bench_trig_kprobe_multi;
extern const struct bench bench_trig_kretprobe_multi;
extern const struct bench bench_trig_fentry;
extern const struct bench bench_trig_fexit;
extern const struct bench bench_trig_fmodret;
extern const struct bench bench_trig_tp;
extern const struct bench bench_trig_rawtp;

/* uprobe/uretprobe benchmarks */
extern const struct bench bench_trig_uprobe_nop;
extern const struct bench bench_trig_uretprobe_nop;
extern const struct bench bench_trig_uprobe_push;
extern const struct bench bench_trig_uretprobe_push;
extern const struct bench bench_trig_uprobe_ret;
extern const struct bench bench_trig_uretprobe_ret;
extern const struct bench bench_trig_uprobe_multi_nop;
extern const struct bench bench_trig_uretprobe_multi_nop;
extern const struct bench bench_trig_uprobe_multi_push;
extern const struct bench bench_trig_uretprobe_multi_push;
extern const struct bench bench_trig_uprobe_multi_ret;
extern const struct bench bench_trig_uretprobe_multi_ret;
#ifdef __x86_64__
extern const struct bench bench_trig_uprobe_nop5;
extern const struct bench bench_trig_uretprobe_nop5;
extern const struct bench bench_trig_uprobe_multi_nop5;
extern const struct bench bench_trig_uretprobe_multi_nop5;
#endif

extern const struct bench bench_rb_libbpf;
extern const struct bench bench_rb_custom;
extern const struct bench bench_pb_libbpf;
extern const struct bench bench_pb_custom;
extern const struct bench bench_bloom_lookup;
extern const struct bench bench_bloom_update;
extern const struct bench bench_bloom_false_positive;
extern const struct bench bench_hashmap_without_bloom;
extern const struct bench bench_hashmap_with_bloom;
extern const struct bench bench_bpf_loop;
extern const struct bench bench_strncmp_no_helper;
extern const struct bench bench_strncmp_helper;
extern const struct bench bench_bpf_hashmap_full_update;
extern const struct bench bench_local_storage_cache_seq_get;
extern const struct bench bench_local_storage_cache_interleaved_get;
extern const struct bench bench_local_storage_cache_hashmap_control;
extern const struct bench bench_local_storage_tasks_trace;
extern const struct bench bench_bpf_hashmap_lookup;
extern const struct bench bench_local_storage_create;
extern const struct bench bench_htab_mem;
extern const struct bench bench_crypto_encrypt;
extern const struct bench bench_crypto_decrypt;
extern const struct bench bench_sockmap;

static const struct bench *benchs[] = {
 &bench_count_global,
 &bench_count_local,
 &bench_rename_base,
 &bench_rename_kprobe,
 &bench_rename_kretprobe,
 &bench_rename_rawtp,
 &bench_rename_fentry,
 &bench_rename_fexit,
 /* pure counting benchmarks for establishing theoretical limits */
 &bench_trig_usermode_count,
 &bench_trig_kernel_count,
 &bench_trig_syscall_count,
 /* batched, staying mostly in-kernel triggers */
 &bench_trig_kprobe,
 &bench_trig_kretprobe,
 &bench_trig_kprobe_multi,
 &bench_trig_kretprobe_multi,
 &bench_trig_fentry,
 &bench_trig_fexit,
 &bench_trig_fmodret,
 &bench_trig_tp,
 &bench_trig_rawtp,
 /* uprobes */
 &bench_trig_uprobe_nop,
 &bench_trig_uretprobe_nop,
 &bench_trig_uprobe_push,
 &bench_trig_uretprobe_push,
 &bench_trig_uprobe_ret,
 &bench_trig_uretprobe_ret,
 &bench_trig_uprobe_multi_nop,
 &bench_trig_uretprobe_multi_nop,
 &bench_trig_uprobe_multi_push,
 &bench_trig_uretprobe_multi_push,
 &bench_trig_uprobe_multi_ret,
 &bench_trig_uretprobe_multi_ret,
#ifdef __x86_64__
 &bench_trig_uprobe_nop5,
 &bench_trig_uretprobe_nop5,
 &bench_trig_uprobe_multi_nop5,
 &bench_trig_uretprobe_multi_nop5,
#endif
 /* ringbuf/perfbuf benchmarks */
 &bench_rb_libbpf,
 &bench_rb_custom,
 &bench_pb_libbpf,
 &bench_pb_custom,
 &bench_bloom_lookup,
 &bench_bloom_update,
 &bench_bloom_false_positive,
 &bench_hashmap_without_bloom,
 &bench_hashmap_with_bloom,
 &bench_bpf_loop,
 &bench_strncmp_no_helper,
 &bench_strncmp_helper,
 &bench_bpf_hashmap_full_update,
 &bench_local_storage_cache_seq_get,
 &bench_local_storage_cache_interleaved_get,
 &bench_local_storage_cache_hashmap_control,
 &bench_local_storage_tasks_trace,
 &bench_bpf_hashmap_lookup,
 &bench_local_storage_create,
 &bench_htab_mem,
 &bench_crypto_encrypt,
 &bench_crypto_decrypt,
 &bench_sockmap,
};

static void find_benchmark(void)
{
 int i;

 if (!env.bench_name) {
  fprintf(stderr, "benchmark name is not specified\n");
  exit(1);
 }
 for (i = 0; i < ARRAY_SIZE(benchs); i++) {
  if (strcmp(benchs[i]->name, env.bench_name) == 0) {
   bench = benchs[i];
   break;
  }
 }
 if (!bench) {
  fprintf(stderr, "benchmark '%s' not found\n", env.bench_name);
  exit(1);
 }
}

static void setup_benchmark(void)
{
 int i, err;

 if (!env.quiet)
  printf("Setting up benchmark '%s'...\n", bench->name);

 state.producers = calloc(env.producer_cnt, sizeof(*state.producers));
 state.consumers = calloc(env.consumer_cnt, sizeof(*state.consumers));
 state.results = calloc(env.duration_sec + env.warmup_sec + 2,
          sizeof(*state.results));
 if (!state.producers || !state.consumers || !state.results)
  exit(1);

 if (bench->validate)
  bench->validate();
 if (bench->setup)
  bench->setup();

 for (i = 0; i < env.consumer_cnt; i++) {
  if (!bench->consumer_thread) {
   fprintf(stderr, "benchmark doesn't support consumers!\n");
   exit(1);
  }
  err = pthread_create(&state.consumers[i], NULL,
         bench->consumer_thread, (void *)(long)i);
  if (err) {
   fprintf(stderr, "failed to create consumer thread #%d: %d\n",
    i, -err);
   exit(1);
  }
  if (env.affinity)
   set_thread_affinity(state.consumers[i],
         next_cpu(&env.cons_cpus));
 }

 /* unless explicit producer CPU list is specified, continue after
 * last consumer CPU
 */
 if (!env.prod_cpus.cpus)
  env.prod_cpus.next_cpu = env.cons_cpus.next_cpu;

 for (i = 0; i < env.producer_cnt; i++) {
  if (!bench->producer_thread) {
   fprintf(stderr, "benchmark doesn't support producers!\n");
   exit(1);
  }
  err = pthread_create(&state.producers[i], NULL,
         bench->producer_thread, (void *)(long)i);
  if (err) {
   fprintf(stderr, "failed to create producer thread #%d: %d\n",
    i, -err);
   exit(1);
  }
  if (env.affinity)
   set_thread_affinity(state.producers[i],
         next_cpu(&env.prod_cpus));
 }

 if (!env.quiet)
  printf("Benchmark '%s' started.\n", bench->name);
}

static pthread_mutex_t bench_done_mtx = PTHREAD_MUTEX_INITIALIZER;
static pthread_cond_t bench_done = PTHREAD_COND_INITIALIZER;

static void collect_measurements(long delta_ns) {
 int iter = state.res_cnt++;
 struct bench_res *res = &state.results[iter];

 bench->measure(res);

 if (bench->report_progress)
  bench->report_progress(iter, res, delta_ns);

 if (iter == env.duration_sec + env.warmup_sec) {
  pthread_mutex_lock(&bench_done_mtx);
  pthread_cond_signal(&bench_done);
  pthread_mutex_unlock(&bench_done_mtx);
 }
}

int main(int argc, char **argv)
{
 env.nr_cpus = get_nprocs();
 parse_cmdline_args_init(argc, argv);

 if (env.list) {
  int i;

  printf("Available benchmarks:\n");
  for (i = 0; i < ARRAY_SIZE(benchs); i++) {
   printf("- %s\n", benchs[i]->name);
  }
  return 0;
 }

 find_benchmark();
 parse_cmdline_args_final(argc, argv);

 setup_benchmark();

 setup_timer();

 pthread_mutex_lock(&bench_done_mtx);
 pthread_cond_wait(&bench_done, &bench_done_mtx);
 pthread_mutex_unlock(&bench_done_mtx);

 if (bench->report_final)
  /* skip first sample */
  bench->report_final(state.results + env.warmup_sec,
        state.res_cnt - env.warmup_sec);

 return 0;
}