void perf_stat_dump(
FILE *yaml,
const stress_t stressors[],
const proc_info_t procs[STRESS_MAX],
const int32_t max_procs,
const double duration)
{
int32_t i;
bool no_perf_stats = true;
setlocale(LC_ALL, "");
pr_yaml(yaml, "perfstats:\n");
for (i = 0; i < STRESS_MAX; i++) {
int p;
uint64_t counter_totals[STRESS_PERF_MAX];
uint64_t total_cpu_cycles = 0;
uint64_t total_cache_refs = 0;
uint64_t total_branches = 0;
int ids[STRESS_PERF_MAX];
bool got_data = false;
char *munged;
memset(counter_totals, 0, sizeof(counter_totals));
/* Sum totals across all instances of the stressor */
for (p = 0; p < STRESS_PERF_MAX; p++) {
int32_t j, n = (i * max_procs);
stress_perf_t *sp = &shared->stats[n].sp;
if (!perf_stat_succeeded(sp))
continue;
ids[p] = ~0;
for (j = 0; j < procs[i].started_procs; j++, n++) {
uint64_t counter;
if (perf_get_counter_by_index(sp, p,
&counter, &ids[p]) < 0)
break;
if (counter == STRESS_PERF_INVALID) {
counter_totals[p] = STRESS_PERF_INVALID;
break;
}
counter_totals[p] += counter;
got_data |= (counter > 0);
}
if (ids[p] == STRESS_PERF_HW_CPU_CYCLES)
total_cpu_cycles = counter_totals[p];
if (ids[p] == STRESS_PERF_HW_CACHE_REFERENCES)
total_cache_refs = counter_totals[p];
if (ids[p] == STRESS_PERF_HW_BRANCH_INSTRUCTIONS)
total_branches = counter_totals[p];
}
if (!got_data)
continue;
munged = munge_underscore((char *)stressors[i].name);
pr_inf(stdout, "%s:\n", munged);
pr_yaml(yaml, " - stressor: %s\n", munged);
pr_yaml(yaml, " duration: %f\n", duration);
for (p = 0; p < STRESS_PERF_MAX; p++) {
const char *l = perf_get_label_by_index(p);
uint64_t ct = counter_totals[p];
if (l && (ct != STRESS_PERF_INVALID)) {
char extra[32];
char yaml_label[128];
*extra = '\0';
no_perf_stats = false;
if ((ids[p] == STRESS_PERF_HW_INSTRUCTIONS) &&
(total_cpu_cycles > 0))
snprintf(extra, sizeof(extra),
" (%.3f instr. per cycle)",
(double)ct / (double)total_cpu_cycles);
if ((ids[p] == STRESS_PERF_HW_CACHE_MISSES) &&
(total_cache_refs > 0))
snprintf(extra, sizeof(extra),
" (%5.2f%%)",
100.0 * (double)ct / (double)total_cache_refs);
if ((ids[p] == STRESS_PERF_HW_BRANCH_MISSES) &&
(total_branches > 0))
snprintf(extra, sizeof(extra),
" (%5.2f%%)",
100.0 * (double)ct / (double)total_branches);
pr_inf(stdout, "%'26" PRIu64 " %-23s %s%s\n",
ct, l, perf_stat_scale(ct, duration),
extra);
perf_yaml_label(yaml_label, l, sizeof(yaml_label));
pr_yaml(yaml, " %s_total: %" PRIu64
"\n", yaml_label, ct);
pr_yaml(yaml, " %s_per_second: %f\n",
yaml_label, (double)ct / duration);
}
}
pr_yaml(yaml, "\n");
}
if (no_perf_stats)
pr_inf(stdout, "perf counters are not available "
"on this device\n");
}
void perf_stat_dump(FILE *yaml, proc_info_t *procs_head, const double duration)
{
bool no_perf_stats = true;
proc_info_t *pi;
#if defined(HAVE_LOCALE_H)
(void)setlocale(LC_ALL, "");
#endif
pr_yaml(yaml, "perfstats:\n");
for (pi = procs_head; pi; pi = pi->next) {
int p;
uint64_t counter_totals[STRESS_PERF_MAX];
uint64_t total_cpu_cycles = 0;
uint64_t total_cache_refs = 0;
uint64_t total_branches = 0;
bool got_data = false;
char *munged;
(void)memset(counter_totals, 0, sizeof(counter_totals));
/* Sum totals across all instances of the stressor */
for (p = 0; p < STRESS_PERF_MAX && perf_info[p].label; p++) {
int32_t j;
stress_perf_t *sp = &pi->stats[0]->sp;
if (!perf_stat_succeeded(sp))
continue;
for (j = 0; j < pi->started_procs; j++) {
const uint64_t counter = sp->perf_stat[p].counter;
if (counter == STRESS_PERF_INVALID) {
counter_totals[p] = STRESS_PERF_INVALID;
break;
}
counter_totals[p] += counter;
got_data |= (counter > 0);
}
if (perf_info[p].type == PERF_TYPE_HARDWARE) {
unsigned long config = perf_info[p].config;
if (config == PERF_COUNT_HW_CPU_CYCLES)
total_cpu_cycles = counter_totals[p];
else if (config == PERF_COUNT_HW_CACHE_REFERENCES)
total_cache_refs = counter_totals[p];
else if (config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS)
total_branches = counter_totals[p];
}
}
if (!got_data)
continue;
munged = stress_munge_underscore(pi->stressor->name);
pr_inf("%s:\n", munged);
pr_yaml(yaml, " - stressor: %s\n", munged);
pr_yaml(yaml, " duration: %f\n", duration);
for (p = 0; p < STRESS_PERF_MAX && perf_info[p].label; p++) {
const char *l = perf_info[p].label;
uint64_t ct = counter_totals[p];
if (l && (ct != STRESS_PERF_INVALID)) {
char extra[32];
char yaml_label[128];
*extra = '\0';
no_perf_stats = false;
if (perf_info[p].type == PERF_TYPE_HARDWARE) {
unsigned long config = perf_info[p].config;
if ((config == PERF_COUNT_HW_INSTRUCTIONS) &&
(total_cpu_cycles > 0))
(void)snprintf(extra, sizeof(extra),
" (%.3f instr. per cycle)",
(double)ct / (double)total_cpu_cycles);
else if ((config == PERF_COUNT_HW_CACHE_MISSES) &&
(total_cache_refs > 0))
(void)snprintf(extra, sizeof(extra),
" (%5.2f%%)",
100.0 * (double)ct / (double)total_cache_refs);
else if ((config == PERF_COUNT_HW_BRANCH_MISSES) &&
(total_branches > 0))
(void)snprintf(extra, sizeof(extra),
" (%5.2f%%)",
100.0 * (double)ct / (double)total_branches);
}
pr_inf("%'26" PRIu64 " %-24s %s%s\n",
ct, l, perf_stat_scale(ct, duration),
extra);
perf_yaml_label(yaml_label, l, sizeof(yaml_label));
pr_yaml(yaml, " %s_total: %" PRIu64
"\n", yaml_label, ct);
pr_yaml(yaml, " %s_per_second: %f\n",
yaml_label, (double)ct / duration);
}
}
pr_yaml(yaml, "\n");
}
if (no_perf_stats) {
if (geteuid() != 0) {
char buffer[64];
int ret;
bool paranoid = false;
int level = 0;
static char *path = "/proc/sys/kernel/perf_event_paranoid";
ret = system_read(path, buffer, sizeof(buffer) - 1);
if (ret > 0) {
if (sscanf(buffer, "%5d", &level) == 1)
paranoid = true;
}
if (paranoid & (level > 1)) {
pr_inf("Cannot read perf counters, "
"do not have CAP_SYS_ADMIN capability "
"or %s is set too high (%d)\n",
path, level);
}
} else {
pr_inf("perf counters are not available "
"on this device\n");
}
}
}