7 #include "thread_map.h"
9 void update_stats(struct stats *stats, u64 val)
14 delta = val - stats->mean;
15 stats->mean += delta / stats->n;
16 stats->M2 += delta*(val - stats->mean);
25 double avg_stats(struct stats *stats)
31 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
33 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
34 * s^2 = -------------------------------
37 * http://en.wikipedia.org/wiki/Stddev
39 * The std dev of the mean is related to the std dev by:
46 double stddev_stats(struct stats *stats)
48 double variance, variance_mean;
53 variance = stats->M2 / (stats->n - 1);
54 variance_mean = variance / stats->n;
56 return sqrt(variance_mean);
59 double rel_stddev_stats(double stddev, double avg)
64 pct = 100.0 * stddev/avg;
69 bool __perf_evsel_stat__is(struct perf_evsel *evsel,
70 enum perf_stat_evsel_id id)
72 struct perf_stat_evsel *ps = evsel->priv;
77 #define ID(id, name) [PERF_STAT_EVSEL_ID__##id] = #name
78 static const char *id_str[PERF_STAT_EVSEL_ID__MAX] = {
80 ID(CYCLES_IN_TX, cpu/cycles-t/),
81 ID(TRANSACTION_START, cpu/tx-start/),
82 ID(ELISION_START, cpu/el-start/),
83 ID(CYCLES_IN_TX_CP, cpu/cycles-ct/),
84 ID(TOPDOWN_TOTAL_SLOTS, topdown-total-slots),
85 ID(TOPDOWN_SLOTS_ISSUED, topdown-slots-issued),
86 ID(TOPDOWN_SLOTS_RETIRED, topdown-slots-retired),
87 ID(TOPDOWN_FETCH_BUBBLES, topdown-fetch-bubbles),
88 ID(TOPDOWN_RECOVERY_BUBBLES, topdown-recovery-bubbles),
92 void perf_stat_evsel_id_init(struct perf_evsel *evsel)
94 struct perf_stat_evsel *ps = evsel->priv;
97 /* ps->id is 0 hence PERF_STAT_EVSEL_ID__NONE by default */
99 for (i = 0; i < PERF_STAT_EVSEL_ID__MAX; i++) {
100 if (!strcmp(perf_evsel__name(evsel), id_str[i])) {
107 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
110 struct perf_stat_evsel *ps = evsel->priv;
112 for (i = 0; i < 3; i++)
113 init_stats(&ps->res_stats[i]);
115 perf_stat_evsel_id_init(evsel);
118 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
120 evsel->priv = zalloc(sizeof(struct perf_stat_evsel));
121 if (evsel->priv == NULL)
123 perf_evsel__reset_stat_priv(evsel);
127 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
132 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel,
133 int ncpus, int nthreads)
135 struct perf_counts *counts;
137 counts = perf_counts__new(ncpus, nthreads);
139 evsel->prev_raw_counts = counts;
141 return counts ? 0 : -ENOMEM;
144 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
146 perf_counts__delete(evsel->prev_raw_counts);
147 evsel->prev_raw_counts = NULL;
150 static int perf_evsel__alloc_stats(struct perf_evsel *evsel, bool alloc_raw)
152 int ncpus = perf_evsel__nr_cpus(evsel);
153 int nthreads = thread_map__nr(evsel->threads);
155 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
156 perf_evsel__alloc_counts(evsel, ncpus, nthreads) < 0 ||
157 (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel, ncpus, nthreads) < 0))
163 int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
165 struct perf_evsel *evsel;
167 evlist__for_each_entry(evlist, evsel) {
168 if (perf_evsel__alloc_stats(evsel, alloc_raw))
175 perf_evlist__free_stats(evlist);
179 void perf_evlist__free_stats(struct perf_evlist *evlist)
181 struct perf_evsel *evsel;
183 evlist__for_each_entry(evlist, evsel) {
184 perf_evsel__free_stat_priv(evsel);
185 perf_evsel__free_counts(evsel);
186 perf_evsel__free_prev_raw_counts(evsel);
190 void perf_evlist__reset_stats(struct perf_evlist *evlist)
192 struct perf_evsel *evsel;
194 evlist__for_each_entry(evlist, evsel) {
195 perf_evsel__reset_stat_priv(evsel);
196 perf_evsel__reset_counts(evsel);
200 static void zero_per_pkg(struct perf_evsel *counter)
202 if (counter->per_pkg_mask)
203 memset(counter->per_pkg_mask, 0, MAX_NR_CPUS);
206 static int check_per_pkg(struct perf_evsel *counter,
207 struct perf_counts_values *vals, int cpu, bool *skip)
209 unsigned long *mask = counter->per_pkg_mask;
210 struct cpu_map *cpus = perf_evsel__cpus(counter);
215 if (!counter->per_pkg)
218 if (cpu_map__empty(cpus))
222 mask = zalloc(MAX_NR_CPUS);
226 counter->per_pkg_mask = mask;
230 * we do not consider an event that has not run as a good
231 * instance to mark a package as used (skip=1). Otherwise
232 * we may run into a situation where the first CPU in a package
233 * is not running anything, yet the second is, and this function
234 * would mark the package as used after the first CPU and would
235 * not read the values from the second CPU.
237 if (!(vals->run && vals->ena))
240 s = cpu_map__get_socket(cpus, cpu, NULL);
244 *skip = test_and_set_bit(s, mask) == 1;
249 process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel,
251 struct perf_counts_values *count)
253 struct perf_counts_values *aggr = &evsel->counts->aggr;
254 static struct perf_counts_values zero;
257 if (check_per_pkg(evsel, count, cpu, &skip)) {
258 pr_err("failed to read per-pkg counter\n");
265 switch (config->aggr_mode) {
270 if (!evsel->snapshot)
271 perf_evsel__compute_deltas(evsel, cpu, thread, count);
272 perf_counts_values__scale(count, config->scale, NULL);
273 if (config->aggr_mode == AGGR_NONE)
274 perf_stat__update_shadow_stats(evsel, count->values, cpu);
277 aggr->val += count->val;
279 aggr->ena += count->ena;
280 aggr->run += count->run;
290 static int process_counter_maps(struct perf_stat_config *config,
291 struct perf_evsel *counter)
293 int nthreads = thread_map__nr(counter->threads);
294 int ncpus = perf_evsel__nr_cpus(counter);
297 if (counter->system_wide)
300 for (thread = 0; thread < nthreads; thread++) {
301 for (cpu = 0; cpu < ncpus; cpu++) {
302 if (process_counter_values(config, counter, cpu, thread,
303 perf_counts(counter->counts, cpu, thread)))
311 int perf_stat_process_counter(struct perf_stat_config *config,
312 struct perf_evsel *counter)
314 struct perf_counts_values *aggr = &counter->counts->aggr;
315 struct perf_stat_evsel *ps = counter->priv;
316 u64 *count = counter->counts->aggr.values;
320 aggr->val = aggr->ena = aggr->run = 0;
323 * We calculate counter's data every interval,
324 * and the display code shows ps->res_stats
325 * avg value. We need to zero the stats for
326 * interval mode, otherwise overall avg running
327 * averages will be shown for each interval.
329 if (config->interval)
330 init_stats(ps->res_stats);
332 if (counter->per_pkg)
333 zero_per_pkg(counter);
335 ret = process_counter_maps(config, counter);
339 if (config->aggr_mode != AGGR_GLOBAL)
342 if (!counter->snapshot)
343 perf_evsel__compute_deltas(counter, -1, -1, aggr);
344 perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled);
346 for (i = 0; i < 3; i++)
347 update_stats(&ps->res_stats[i], count[i]);
350 fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
351 perf_evsel__name(counter), count[0], count[1], count[2]);
355 * Save the full runtime - to allow normalization during printout:
357 val = counter->scale * *count;
358 perf_stat__update_shadow_stats(counter, &val, 0);
363 int perf_event__process_stat_event(struct perf_tool *tool __maybe_unused,
364 union perf_event *event,
365 struct perf_session *session)
367 struct perf_counts_values count;
368 struct stat_event *st = &event->stat;
369 struct perf_evsel *counter;
375 counter = perf_evlist__id2evsel(session->evlist, st->id);
377 pr_err("Failed to resolve counter for stat event.\n");
381 *perf_counts(counter->counts, st->cpu, st->thread) = count;
382 counter->supported = true;
386 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
388 struct stat_event *st = (struct stat_event *) event;
391 ret = fprintf(fp, "\n... id %" PRIu64 ", cpu %d, thread %d\n",
392 st->id, st->cpu, st->thread);
393 ret += fprintf(fp, "... value %" PRIu64 ", enabled %" PRIu64 ", running %" PRIu64 "\n",
394 st->val, st->ena, st->run);
399 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
401 struct stat_round_event *rd = (struct stat_round_event *)event;
404 ret = fprintf(fp, "\n... time %" PRIu64 ", type %s\n", rd->time,
405 rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
410 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
412 struct perf_stat_config sc;
415 perf_event__read_stat_config(&sc, &event->stat_config);
417 ret = fprintf(fp, "\n");
418 ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
419 ret += fprintf(fp, "... scale %d\n", sc.scale);
420 ret += fprintf(fp, "... interval %u\n", sc.interval);