1 // SPDX-License-Identifier: GPL-2.0
8 #include "thread_map.h"
10 void update_stats(struct stats *stats, u64 val)
15 delta = val - stats->mean;
16 stats->mean += delta / stats->n;
17 stats->M2 += delta*(val - stats->mean);
26 double avg_stats(struct stats *stats)
32 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
34 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
35 * s^2 = -------------------------------
38 * http://en.wikipedia.org/wiki/Stddev
40 * The std dev of the mean is related to the std dev by:
47 double stddev_stats(struct stats *stats)
49 double variance, variance_mean;
54 variance = stats->M2 / (stats->n - 1);
55 variance_mean = variance / stats->n;
57 return sqrt(variance_mean);
60 double rel_stddev_stats(double stddev, double avg)
65 pct = 100.0 * stddev/avg;
70 bool __perf_evsel_stat__is(struct perf_evsel *evsel,
71 enum perf_stat_evsel_id id)
73 struct perf_stat_evsel *ps = evsel->stats;
78 #define ID(id, name) [PERF_STAT_EVSEL_ID__##id] = #name
79 static const char *id_str[PERF_STAT_EVSEL_ID__MAX] = {
81 ID(CYCLES_IN_TX, cpu/cycles-t/),
82 ID(TRANSACTION_START, cpu/tx-start/),
83 ID(ELISION_START, cpu/el-start/),
84 ID(CYCLES_IN_TX_CP, cpu/cycles-ct/),
85 ID(TOPDOWN_TOTAL_SLOTS, topdown-total-slots),
86 ID(TOPDOWN_SLOTS_ISSUED, topdown-slots-issued),
87 ID(TOPDOWN_SLOTS_RETIRED, topdown-slots-retired),
88 ID(TOPDOWN_FETCH_BUBBLES, topdown-fetch-bubbles),
89 ID(TOPDOWN_RECOVERY_BUBBLES, topdown-recovery-bubbles),
90 ID(SMI_NUM, msr/smi/),
91 ID(APERF, msr/aperf/),
95 static void perf_stat_evsel_id_init(struct perf_evsel *evsel)
97 struct perf_stat_evsel *ps = evsel->stats;
100 /* ps->id is 0 hence PERF_STAT_EVSEL_ID__NONE by default */
102 for (i = 0; i < PERF_STAT_EVSEL_ID__MAX; i++) {
103 if (!strcmp(perf_evsel__name(evsel), id_str[i])) {
110 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
113 struct perf_stat_evsel *ps = evsel->stats;
115 for (i = 0; i < 3; i++)
116 init_stats(&ps->res_stats[i]);
118 perf_stat_evsel_id_init(evsel);
121 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
123 evsel->stats = zalloc(sizeof(struct perf_stat_evsel));
124 if (evsel->stats == NULL)
126 perf_evsel__reset_stat_priv(evsel);
130 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
132 struct perf_stat_evsel *ps = evsel->stats;
135 free(ps->group_data);
136 zfree(&evsel->stats);
139 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel,
140 int ncpus, int nthreads)
142 struct perf_counts *counts;
144 counts = perf_counts__new(ncpus, nthreads);
146 evsel->prev_raw_counts = counts;
148 return counts ? 0 : -ENOMEM;
151 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
153 perf_counts__delete(evsel->prev_raw_counts);
154 evsel->prev_raw_counts = NULL;
157 static void perf_evsel__reset_prev_raw_counts(struct perf_evsel *evsel)
159 if (evsel->prev_raw_counts) {
160 evsel->prev_raw_counts->aggr.val = 0;
161 evsel->prev_raw_counts->aggr.ena = 0;
162 evsel->prev_raw_counts->aggr.run = 0;
166 static int perf_evsel__alloc_stats(struct perf_evsel *evsel, bool alloc_raw)
168 int ncpus = perf_evsel__nr_cpus(evsel);
169 int nthreads = thread_map__nr(evsel->threads);
171 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
172 perf_evsel__alloc_counts(evsel, ncpus, nthreads) < 0 ||
173 (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel, ncpus, nthreads) < 0))
179 int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
181 struct perf_evsel *evsel;
183 evlist__for_each_entry(evlist, evsel) {
184 if (perf_evsel__alloc_stats(evsel, alloc_raw))
191 perf_evlist__free_stats(evlist);
195 void perf_evlist__free_stats(struct perf_evlist *evlist)
197 struct perf_evsel *evsel;
199 evlist__for_each_entry(evlist, evsel) {
200 perf_evsel__free_stat_priv(evsel);
201 perf_evsel__free_counts(evsel);
202 perf_evsel__free_prev_raw_counts(evsel);
206 void perf_evlist__reset_stats(struct perf_evlist *evlist)
208 struct perf_evsel *evsel;
210 evlist__for_each_entry(evlist, evsel) {
211 perf_evsel__reset_stat_priv(evsel);
212 perf_evsel__reset_counts(evsel);
216 void perf_evlist__reset_prev_raw_counts(struct perf_evlist *evlist)
218 struct perf_evsel *evsel;
220 evlist__for_each_entry(evlist, evsel)
221 perf_evsel__reset_prev_raw_counts(evsel);
224 static void zero_per_pkg(struct perf_evsel *counter)
226 if (counter->per_pkg_mask)
227 memset(counter->per_pkg_mask, 0, MAX_NR_CPUS);
230 static int check_per_pkg(struct perf_evsel *counter,
231 struct perf_counts_values *vals, int cpu, bool *skip)
233 unsigned long *mask = counter->per_pkg_mask;
234 struct cpu_map *cpus = perf_evsel__cpus(counter);
239 if (!counter->per_pkg)
242 if (cpu_map__empty(cpus))
246 mask = zalloc(MAX_NR_CPUS);
250 counter->per_pkg_mask = mask;
254 * we do not consider an event that has not run as a good
255 * instance to mark a package as used (skip=1). Otherwise
256 * we may run into a situation where the first CPU in a package
257 * is not running anything, yet the second is, and this function
258 * would mark the package as used after the first CPU and would
259 * not read the values from the second CPU.
261 if (!(vals->run && vals->ena))
264 s = cpu_map__get_socket(cpus, cpu, NULL);
268 *skip = test_and_set_bit(s, mask) == 1;
273 process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel,
275 struct perf_counts_values *count)
277 struct perf_counts_values *aggr = &evsel->counts->aggr;
278 static struct perf_counts_values zero;
281 if (check_per_pkg(evsel, count, cpu, &skip)) {
282 pr_err("failed to read per-pkg counter\n");
289 switch (config->aggr_mode) {
294 if (!evsel->snapshot)
295 perf_evsel__compute_deltas(evsel, cpu, thread, count);
296 perf_counts_values__scale(count, config->scale, NULL);
297 if (config->aggr_mode == AGGR_NONE)
298 perf_stat__update_shadow_stats(evsel, count->val, cpu,
300 if (config->aggr_mode == AGGR_THREAD) {
302 perf_stat__update_shadow_stats(evsel,
303 count->val, 0, &config->stats[thread]);
305 perf_stat__update_shadow_stats(evsel,
306 count->val, 0, &rt_stat);
310 aggr->val += count->val;
312 aggr->ena += count->ena;
313 aggr->run += count->run;
323 static int process_counter_maps(struct perf_stat_config *config,
324 struct perf_evsel *counter)
326 int nthreads = thread_map__nr(counter->threads);
327 int ncpus = perf_evsel__nr_cpus(counter);
330 if (counter->system_wide)
333 for (thread = 0; thread < nthreads; thread++) {
334 for (cpu = 0; cpu < ncpus; cpu++) {
335 if (process_counter_values(config, counter, cpu, thread,
336 perf_counts(counter->counts, cpu, thread)))
344 int perf_stat_process_counter(struct perf_stat_config *config,
345 struct perf_evsel *counter)
347 struct perf_counts_values *aggr = &counter->counts->aggr;
348 struct perf_stat_evsel *ps = counter->stats;
349 u64 *count = counter->counts->aggr.values;
352 aggr->val = aggr->ena = aggr->run = 0;
355 * We calculate counter's data every interval,
356 * and the display code shows ps->res_stats
357 * avg value. We need to zero the stats for
358 * interval mode, otherwise overall avg running
359 * averages will be shown for each interval.
361 if (config->interval)
362 init_stats(ps->res_stats);
364 if (counter->per_pkg)
365 zero_per_pkg(counter);
367 ret = process_counter_maps(config, counter);
371 if (config->aggr_mode != AGGR_GLOBAL)
374 if (!counter->snapshot)
375 perf_evsel__compute_deltas(counter, -1, -1, aggr);
376 perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled);
378 for (i = 0; i < 3; i++)
379 update_stats(&ps->res_stats[i], count[i]);
382 fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
383 perf_evsel__name(counter), count[0], count[1], count[2]);
387 * Save the full runtime - to allow normalization during printout:
389 perf_stat__update_shadow_stats(counter, *count, 0, &rt_stat);
394 int perf_event__process_stat_event(struct perf_tool *tool __maybe_unused,
395 union perf_event *event,
396 struct perf_session *session)
398 struct perf_counts_values count;
399 struct stat_event *st = &event->stat;
400 struct perf_evsel *counter;
406 counter = perf_evlist__id2evsel(session->evlist, st->id);
408 pr_err("Failed to resolve counter for stat event.\n");
412 *perf_counts(counter->counts, st->cpu, st->thread) = count;
413 counter->supported = true;
417 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
419 struct stat_event *st = (struct stat_event *) event;
422 ret = fprintf(fp, "\n... id %" PRIu64 ", cpu %d, thread %d\n",
423 st->id, st->cpu, st->thread);
424 ret += fprintf(fp, "... value %" PRIu64 ", enabled %" PRIu64 ", running %" PRIu64 "\n",
425 st->val, st->ena, st->run);
430 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
432 struct stat_round_event *rd = (struct stat_round_event *)event;
435 ret = fprintf(fp, "\n... time %" PRIu64 ", type %s\n", rd->time,
436 rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
441 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
443 struct perf_stat_config sc;
446 perf_event__read_stat_config(&sc, &event->stat_config);
448 ret = fprintf(fp, "\n");
449 ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
450 ret += fprintf(fp, "... scale %d\n", sc.scale);
451 ret += fprintf(fp, "... interval %u\n", sc.interval);