1 // SPDX-License-Identifier: GPL-2.0
15 #include "thread_map.h"
16 #include <linux/zalloc.h>
18 void update_stats(struct stats *stats, u64 val)
23 delta = val - stats->mean;
24 stats->mean += delta / stats->n;
25 stats->M2 += delta*(val - stats->mean);
34 double avg_stats(struct stats *stats)
40 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
42 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
43 * s^2 = -------------------------------
46 * http://en.wikipedia.org/wiki/Stddev
48 * The std dev of the mean is related to the std dev by:
55 double stddev_stats(struct stats *stats)
57 double variance, variance_mean;
62 variance = stats->M2 / (stats->n - 1);
63 variance_mean = variance / stats->n;
65 return sqrt(variance_mean);
68 double rel_stddev_stats(double stddev, double avg)
73 pct = 100.0 * stddev/avg;
78 bool __perf_evsel_stat__is(struct evsel *evsel,
79 enum perf_stat_evsel_id id)
81 struct perf_stat_evsel *ps = evsel->stats;
86 #define ID(id, name) [PERF_STAT_EVSEL_ID__##id] = #name
87 static const char *id_str[PERF_STAT_EVSEL_ID__MAX] = {
89 ID(CYCLES_IN_TX, cpu/cycles-t/),
90 ID(TRANSACTION_START, cpu/tx-start/),
91 ID(ELISION_START, cpu/el-start/),
92 ID(CYCLES_IN_TX_CP, cpu/cycles-ct/),
93 ID(TOPDOWN_TOTAL_SLOTS, topdown-total-slots),
94 ID(TOPDOWN_SLOTS_ISSUED, topdown-slots-issued),
95 ID(TOPDOWN_SLOTS_RETIRED, topdown-slots-retired),
96 ID(TOPDOWN_FETCH_BUBBLES, topdown-fetch-bubbles),
97 ID(TOPDOWN_RECOVERY_BUBBLES, topdown-recovery-bubbles),
98 ID(SMI_NUM, msr/smi/),
99 ID(APERF, msr/aperf/),
103 static void perf_stat_evsel_id_init(struct evsel *evsel)
105 struct perf_stat_evsel *ps = evsel->stats;
108 /* ps->id is 0 hence PERF_STAT_EVSEL_ID__NONE by default */
110 for (i = 0; i < PERF_STAT_EVSEL_ID__MAX; i++) {
111 if (!strcmp(evsel__name(evsel), id_str[i])) {
118 static void evsel__reset_stat_priv(struct evsel *evsel)
121 struct perf_stat_evsel *ps = evsel->stats;
123 for (i = 0; i < 3; i++)
124 init_stats(&ps->res_stats[i]);
126 perf_stat_evsel_id_init(evsel);
129 static int evsel__alloc_stat_priv(struct evsel *evsel)
131 evsel->stats = zalloc(sizeof(struct perf_stat_evsel));
132 if (evsel->stats == NULL)
134 evsel__reset_stat_priv(evsel);
138 static void evsel__free_stat_priv(struct evsel *evsel)
140 struct perf_stat_evsel *ps = evsel->stats;
143 zfree(&ps->group_data);
144 zfree(&evsel->stats);
147 static int evsel__alloc_prev_raw_counts(struct evsel *evsel, int ncpus, int nthreads)
149 struct perf_counts *counts;
151 counts = perf_counts__new(ncpus, nthreads);
153 evsel->prev_raw_counts = counts;
155 return counts ? 0 : -ENOMEM;
158 static void evsel__free_prev_raw_counts(struct evsel *evsel)
160 perf_counts__delete(evsel->prev_raw_counts);
161 evsel->prev_raw_counts = NULL;
164 static void evsel__reset_prev_raw_counts(struct evsel *evsel)
166 if (evsel->prev_raw_counts) {
167 evsel->prev_raw_counts->aggr.val = 0;
168 evsel->prev_raw_counts->aggr.ena = 0;
169 evsel->prev_raw_counts->aggr.run = 0;
173 static int evsel__alloc_stats(struct evsel *evsel, bool alloc_raw)
175 int ncpus = evsel__nr_cpus(evsel);
176 int nthreads = perf_thread_map__nr(evsel->core.threads);
178 if (evsel__alloc_stat_priv(evsel) < 0 ||
179 evsel__alloc_counts(evsel, ncpus, nthreads) < 0 ||
180 (alloc_raw && evsel__alloc_prev_raw_counts(evsel, ncpus, nthreads) < 0))
186 int perf_evlist__alloc_stats(struct evlist *evlist, bool alloc_raw)
190 evlist__for_each_entry(evlist, evsel) {
191 if (evsel__alloc_stats(evsel, alloc_raw))
198 perf_evlist__free_stats(evlist);
202 void perf_evlist__free_stats(struct evlist *evlist)
206 evlist__for_each_entry(evlist, evsel) {
207 evsel__free_stat_priv(evsel);
208 evsel__free_counts(evsel);
209 evsel__free_prev_raw_counts(evsel);
213 void perf_evlist__reset_stats(struct evlist *evlist)
217 evlist__for_each_entry(evlist, evsel) {
218 evsel__reset_stat_priv(evsel);
219 evsel__reset_counts(evsel);
223 void perf_evlist__reset_prev_raw_counts(struct evlist *evlist)
227 evlist__for_each_entry(evlist, evsel)
228 evsel__reset_prev_raw_counts(evsel);
231 static void zero_per_pkg(struct evsel *counter)
233 if (counter->per_pkg_mask)
234 memset(counter->per_pkg_mask, 0, cpu__max_cpu());
237 static int check_per_pkg(struct evsel *counter,
238 struct perf_counts_values *vals, int cpu, bool *skip)
240 unsigned long *mask = counter->per_pkg_mask;
241 struct perf_cpu_map *cpus = evsel__cpus(counter);
246 if (!counter->per_pkg)
249 if (perf_cpu_map__empty(cpus))
253 mask = zalloc(cpu__max_cpu());
257 counter->per_pkg_mask = mask;
261 * we do not consider an event that has not run as a good
262 * instance to mark a package as used (skip=1). Otherwise
263 * we may run into a situation where the first CPU in a package
264 * is not running anything, yet the second is, and this function
265 * would mark the package as used after the first CPU and would
266 * not read the values from the second CPU.
268 if (!(vals->run && vals->ena))
271 s = cpu_map__get_socket(cpus, cpu, NULL);
275 *skip = test_and_set_bit(s, mask) == 1;
280 process_counter_values(struct perf_stat_config *config, struct evsel *evsel,
282 struct perf_counts_values *count)
284 struct perf_counts_values *aggr = &evsel->counts->aggr;
285 static struct perf_counts_values zero;
288 if (check_per_pkg(evsel, count, cpu, &skip)) {
289 pr_err("failed to read per-pkg counter\n");
296 switch (config->aggr_mode) {
303 if (!evsel->snapshot)
304 evsel__compute_deltas(evsel, cpu, thread, count);
305 perf_counts_values__scale(count, config->scale, NULL);
306 if ((config->aggr_mode == AGGR_NONE) && (!evsel->percore)) {
307 perf_stat__update_shadow_stats(evsel, count->val,
311 if (config->aggr_mode == AGGR_THREAD) {
313 perf_stat__update_shadow_stats(evsel,
314 count->val, 0, &config->stats[thread]);
316 perf_stat__update_shadow_stats(evsel,
317 count->val, 0, &rt_stat);
321 aggr->val += count->val;
322 aggr->ena += count->ena;
323 aggr->run += count->run;
332 static int process_counter_maps(struct perf_stat_config *config,
333 struct evsel *counter)
335 int nthreads = perf_thread_map__nr(counter->core.threads);
336 int ncpus = evsel__nr_cpus(counter);
339 if (counter->core.system_wide)
342 for (thread = 0; thread < nthreads; thread++) {
343 for (cpu = 0; cpu < ncpus; cpu++) {
344 if (process_counter_values(config, counter, cpu, thread,
345 perf_counts(counter->counts, cpu, thread)))
353 int perf_stat_process_counter(struct perf_stat_config *config,
354 struct evsel *counter)
356 struct perf_counts_values *aggr = &counter->counts->aggr;
357 struct perf_stat_evsel *ps = counter->stats;
358 u64 *count = counter->counts->aggr.values;
361 aggr->val = aggr->ena = aggr->run = 0;
364 * We calculate counter's data every interval,
365 * and the display code shows ps->res_stats
366 * avg value. We need to zero the stats for
367 * interval mode, otherwise overall avg running
368 * averages will be shown for each interval.
370 if (config->interval) {
371 for (i = 0; i < 3; i++)
372 init_stats(&ps->res_stats[i]);
375 if (counter->per_pkg)
376 zero_per_pkg(counter);
378 ret = process_counter_maps(config, counter);
382 if (config->aggr_mode != AGGR_GLOBAL)
385 if (!counter->snapshot)
386 evsel__compute_deltas(counter, -1, -1, aggr);
387 perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled);
389 for (i = 0; i < 3; i++)
390 update_stats(&ps->res_stats[i], count[i]);
393 fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
394 evsel__name(counter), count[0], count[1], count[2]);
398 * Save the full runtime - to allow normalization during printout:
400 perf_stat__update_shadow_stats(counter, *count, 0, &rt_stat);
405 int perf_event__process_stat_event(struct perf_session *session,
406 union perf_event *event)
408 struct perf_counts_values count;
409 struct perf_record_stat *st = &event->stat;
410 struct evsel *counter;
416 counter = perf_evlist__id2evsel(session->evlist, st->id);
418 pr_err("Failed to resolve counter for stat event.\n");
422 *perf_counts(counter->counts, st->cpu, st->thread) = count;
423 counter->supported = true;
427 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
429 struct perf_record_stat *st = (struct perf_record_stat *)event;
432 ret = fprintf(fp, "\n... id %" PRI_lu64 ", cpu %d, thread %d\n",
433 st->id, st->cpu, st->thread);
434 ret += fprintf(fp, "... value %" PRI_lu64 ", enabled %" PRI_lu64 ", running %" PRI_lu64 "\n",
435 st->val, st->ena, st->run);
440 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
442 struct perf_record_stat_round *rd = (struct perf_record_stat_round *)event;
445 ret = fprintf(fp, "\n... time %" PRI_lu64 ", type %s\n", rd->time,
446 rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
451 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
453 struct perf_stat_config sc;
456 perf_event__read_stat_config(&sc, &event->stat_config);
458 ret = fprintf(fp, "\n");
459 ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
460 ret += fprintf(fp, "... scale %d\n", sc.scale);
461 ret += fprintf(fp, "... interval %u\n", sc.interval);
466 int create_perf_stat_counter(struct evsel *evsel,
467 struct perf_stat_config *config,
468 struct target *target,
471 struct perf_event_attr *attr = &evsel->core.attr;
472 struct evsel *leader = evsel->leader;
474 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
475 PERF_FORMAT_TOTAL_TIME_RUNNING;
478 * The event is part of non trivial group, let's enable
479 * the group read (for leader) and ID retrieval for all
482 if (leader->core.nr_members > 1)
483 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
485 attr->inherit = !config->no_inherit;
488 * Some events get initialized with sample_(period/type) set,
489 * like tracepoints. Clear it up for counting.
491 attr->sample_period = 0;
493 if (config->identifier)
494 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
496 if (config->all_user) {
497 attr->exclude_kernel = 1;
498 attr->exclude_user = 0;
501 if (config->all_kernel) {
502 attr->exclude_kernel = 0;
503 attr->exclude_user = 1;
507 * Disabling all counters initially, they will be enabled
508 * either manually by us or by kernel via enable_on_exec
511 if (evsel__is_group_leader(evsel)) {
515 * In case of initial_delay we enable tracee
518 if (target__none(target) && !config->initial_delay)
519 attr->enable_on_exec = 1;
522 if (target__has_cpu(target) && !target__has_per_thread(target))
523 return evsel__open_per_cpu(evsel, evsel__cpus(evsel), cpu);
525 return evsel__open_per_thread(evsel, evsel->core.threads);