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(TOPDOWN_RETIRING, topdown-retiring),
99 ID(TOPDOWN_BAD_SPEC, topdown-bad-spec),
100 ID(TOPDOWN_FE_BOUND, topdown-fe-bound),
101 ID(TOPDOWN_BE_BOUND, topdown-be-bound),
102 ID(SMI_NUM, msr/smi/),
103 ID(APERF, msr/aperf/),
107 static void perf_stat_evsel_id_init(struct evsel *evsel)
109 struct perf_stat_evsel *ps = evsel->stats;
112 /* ps->id is 0 hence PERF_STAT_EVSEL_ID__NONE by default */
114 for (i = 0; i < PERF_STAT_EVSEL_ID__MAX; i++) {
115 if (!strcmp(evsel__name(evsel), id_str[i])) {
122 static void evsel__reset_stat_priv(struct evsel *evsel)
125 struct perf_stat_evsel *ps = evsel->stats;
127 for (i = 0; i < 3; i++)
128 init_stats(&ps->res_stats[i]);
130 perf_stat_evsel_id_init(evsel);
133 static int evsel__alloc_stat_priv(struct evsel *evsel)
135 evsel->stats = zalloc(sizeof(struct perf_stat_evsel));
136 if (evsel->stats == NULL)
138 evsel__reset_stat_priv(evsel);
142 static void evsel__free_stat_priv(struct evsel *evsel)
144 struct perf_stat_evsel *ps = evsel->stats;
147 zfree(&ps->group_data);
148 zfree(&evsel->stats);
151 static int evsel__alloc_prev_raw_counts(struct evsel *evsel, int ncpus, int nthreads)
153 struct perf_counts *counts;
155 counts = perf_counts__new(ncpus, nthreads);
157 evsel->prev_raw_counts = counts;
159 return counts ? 0 : -ENOMEM;
162 static void evsel__free_prev_raw_counts(struct evsel *evsel)
164 perf_counts__delete(evsel->prev_raw_counts);
165 evsel->prev_raw_counts = NULL;
168 static void evsel__reset_prev_raw_counts(struct evsel *evsel)
170 if (evsel->prev_raw_counts)
171 perf_counts__reset(evsel->prev_raw_counts);
174 static int evsel__alloc_stats(struct evsel *evsel, bool alloc_raw)
176 int ncpus = evsel__nr_cpus(evsel);
177 int nthreads = perf_thread_map__nr(evsel->core.threads);
179 if (evsel__alloc_stat_priv(evsel) < 0 ||
180 evsel__alloc_counts(evsel, ncpus, nthreads) < 0 ||
181 (alloc_raw && evsel__alloc_prev_raw_counts(evsel, ncpus, nthreads) < 0))
187 int perf_evlist__alloc_stats(struct evlist *evlist, bool alloc_raw)
191 evlist__for_each_entry(evlist, evsel) {
192 if (evsel__alloc_stats(evsel, alloc_raw))
199 perf_evlist__free_stats(evlist);
203 void perf_evlist__free_stats(struct evlist *evlist)
207 evlist__for_each_entry(evlist, evsel) {
208 evsel__free_stat_priv(evsel);
209 evsel__free_counts(evsel);
210 evsel__free_prev_raw_counts(evsel);
214 void perf_evlist__reset_stats(struct evlist *evlist)
218 evlist__for_each_entry(evlist, evsel) {
219 evsel__reset_stat_priv(evsel);
220 evsel__reset_counts(evsel);
224 void perf_evlist__reset_prev_raw_counts(struct evlist *evlist)
228 evlist__for_each_entry(evlist, evsel)
229 evsel__reset_prev_raw_counts(evsel);
232 static void perf_evsel__copy_prev_raw_counts(struct evsel *evsel)
234 int ncpus = evsel__nr_cpus(evsel);
235 int nthreads = perf_thread_map__nr(evsel->core.threads);
237 for (int thread = 0; thread < nthreads; thread++) {
238 for (int cpu = 0; cpu < ncpus; cpu++) {
239 *perf_counts(evsel->counts, cpu, thread) =
240 *perf_counts(evsel->prev_raw_counts, cpu,
245 evsel->counts->aggr = evsel->prev_raw_counts->aggr;
248 void perf_evlist__copy_prev_raw_counts(struct evlist *evlist)
252 evlist__for_each_entry(evlist, evsel)
253 perf_evsel__copy_prev_raw_counts(evsel);
256 void perf_evlist__save_aggr_prev_raw_counts(struct evlist *evlist)
261 * To collect the overall statistics for interval mode,
262 * we copy the counts from evsel->prev_raw_counts to
263 * evsel->counts. The perf_stat_process_counter creates
264 * aggr values from per cpu values, but the per cpu values
265 * are 0 for AGGR_GLOBAL. So we use a trick that saves the
266 * previous aggr value to the first member of perf_counts,
267 * then aggr calculation in process_counter_values can work
270 evlist__for_each_entry(evlist, evsel) {
271 *perf_counts(evsel->prev_raw_counts, 0, 0) =
272 evsel->prev_raw_counts->aggr;
276 static void zero_per_pkg(struct evsel *counter)
278 if (counter->per_pkg_mask)
279 memset(counter->per_pkg_mask, 0, cpu__max_cpu());
282 static int check_per_pkg(struct evsel *counter,
283 struct perf_counts_values *vals, int cpu, bool *skip)
285 unsigned long *mask = counter->per_pkg_mask;
286 struct perf_cpu_map *cpus = evsel__cpus(counter);
291 if (!counter->per_pkg)
294 if (perf_cpu_map__empty(cpus))
298 mask = zalloc(cpu__max_cpu());
302 counter->per_pkg_mask = mask;
306 * we do not consider an event that has not run as a good
307 * instance to mark a package as used (skip=1). Otherwise
308 * we may run into a situation where the first CPU in a package
309 * is not running anything, yet the second is, and this function
310 * would mark the package as used after the first CPU and would
311 * not read the values from the second CPU.
313 if (!(vals->run && vals->ena))
316 s = cpu_map__get_socket(cpus, cpu, NULL);
320 *skip = test_and_set_bit(s, mask) == 1;
325 process_counter_values(struct perf_stat_config *config, struct evsel *evsel,
327 struct perf_counts_values *count)
329 struct perf_counts_values *aggr = &evsel->counts->aggr;
330 static struct perf_counts_values zero;
333 if (check_per_pkg(evsel, count, cpu, &skip)) {
334 pr_err("failed to read per-pkg counter\n");
341 switch (config->aggr_mode) {
348 if (!evsel->snapshot)
349 evsel__compute_deltas(evsel, cpu, thread, count);
350 perf_counts_values__scale(count, config->scale, NULL);
351 if ((config->aggr_mode == AGGR_NONE) && (!evsel->percore)) {
352 perf_stat__update_shadow_stats(evsel, count->val,
356 if (config->aggr_mode == AGGR_THREAD) {
358 perf_stat__update_shadow_stats(evsel,
359 count->val, 0, &config->stats[thread]);
361 perf_stat__update_shadow_stats(evsel,
362 count->val, 0, &rt_stat);
366 aggr->val += count->val;
367 aggr->ena += count->ena;
368 aggr->run += count->run;
377 static int process_counter_maps(struct perf_stat_config *config,
378 struct evsel *counter)
380 int nthreads = perf_thread_map__nr(counter->core.threads);
381 int ncpus = evsel__nr_cpus(counter);
384 if (counter->core.system_wide)
387 for (thread = 0; thread < nthreads; thread++) {
388 for (cpu = 0; cpu < ncpus; cpu++) {
389 if (process_counter_values(config, counter, cpu, thread,
390 perf_counts(counter->counts, cpu, thread)))
398 int perf_stat_process_counter(struct perf_stat_config *config,
399 struct evsel *counter)
401 struct perf_counts_values *aggr = &counter->counts->aggr;
402 struct perf_stat_evsel *ps = counter->stats;
403 u64 *count = counter->counts->aggr.values;
406 aggr->val = aggr->ena = aggr->run = 0;
409 * We calculate counter's data every interval,
410 * and the display code shows ps->res_stats
411 * avg value. We need to zero the stats for
412 * interval mode, otherwise overall avg running
413 * averages will be shown for each interval.
415 if (config->interval || config->summary) {
416 for (i = 0; i < 3; i++)
417 init_stats(&ps->res_stats[i]);
420 if (counter->per_pkg)
421 zero_per_pkg(counter);
423 ret = process_counter_maps(config, counter);
427 if (config->aggr_mode != AGGR_GLOBAL)
430 if (!counter->snapshot)
431 evsel__compute_deltas(counter, -1, -1, aggr);
432 perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled);
434 for (i = 0; i < 3; i++)
435 update_stats(&ps->res_stats[i], count[i]);
438 fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
439 evsel__name(counter), count[0], count[1], count[2]);
443 * Save the full runtime - to allow normalization during printout:
445 perf_stat__update_shadow_stats(counter, *count, 0, &rt_stat);
450 int perf_event__process_stat_event(struct perf_session *session,
451 union perf_event *event)
453 struct perf_counts_values count;
454 struct perf_record_stat *st = &event->stat;
455 struct evsel *counter;
461 counter = perf_evlist__id2evsel(session->evlist, st->id);
463 pr_err("Failed to resolve counter for stat event.\n");
467 *perf_counts(counter->counts, st->cpu, st->thread) = count;
468 counter->supported = true;
472 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
474 struct perf_record_stat *st = (struct perf_record_stat *)event;
477 ret = fprintf(fp, "\n... id %" PRI_lu64 ", cpu %d, thread %d\n",
478 st->id, st->cpu, st->thread);
479 ret += fprintf(fp, "... value %" PRI_lu64 ", enabled %" PRI_lu64 ", running %" PRI_lu64 "\n",
480 st->val, st->ena, st->run);
485 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
487 struct perf_record_stat_round *rd = (struct perf_record_stat_round *)event;
490 ret = fprintf(fp, "\n... time %" PRI_lu64 ", type %s\n", rd->time,
491 rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
496 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
498 struct perf_stat_config sc;
501 perf_event__read_stat_config(&sc, &event->stat_config);
503 ret = fprintf(fp, "\n");
504 ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
505 ret += fprintf(fp, "... scale %d\n", sc.scale);
506 ret += fprintf(fp, "... interval %u\n", sc.interval);
511 int create_perf_stat_counter(struct evsel *evsel,
512 struct perf_stat_config *config,
513 struct target *target,
516 struct perf_event_attr *attr = &evsel->core.attr;
517 struct evsel *leader = evsel->leader;
519 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
520 PERF_FORMAT_TOTAL_TIME_RUNNING;
523 * The event is part of non trivial group, let's enable
524 * the group read (for leader) and ID retrieval for all
527 if (leader->core.nr_members > 1)
528 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
530 attr->inherit = !config->no_inherit;
533 * Some events get initialized with sample_(period/type) set,
534 * like tracepoints. Clear it up for counting.
536 attr->sample_period = 0;
538 if (config->identifier)
539 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
541 if (config->all_user) {
542 attr->exclude_kernel = 1;
543 attr->exclude_user = 0;
546 if (config->all_kernel) {
547 attr->exclude_kernel = 0;
548 attr->exclude_user = 1;
552 * Disabling all counters initially, they will be enabled
553 * either manually by us or by kernel via enable_on_exec
556 if (evsel__is_group_leader(evsel)) {
560 * In case of initial_delay we enable tracee
563 if (target__none(target) && !config->initial_delay)
564 attr->enable_on_exec = 1;
567 if (target__has_cpu(target) && !target__has_per_thread(target))
568 return evsel__open_per_cpu(evsel, evsel__cpus(evsel), cpu);
570 return evsel__open_per_thread(evsel, evsel->core.threads);