5 #include "thread_map.h"
7 void update_stats(struct stats *stats, u64 val)
12 delta = val - stats->mean;
13 stats->mean += delta / stats->n;
14 stats->M2 += delta*(val - stats->mean);
23 double avg_stats(struct stats *stats)
29 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
31 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
32 * s^2 = -------------------------------
35 * http://en.wikipedia.org/wiki/Stddev
37 * The std dev of the mean is related to the std dev by:
44 double stddev_stats(struct stats *stats)
46 double variance, variance_mean;
51 variance = stats->M2 / (stats->n - 1);
52 variance_mean = variance / stats->n;
54 return sqrt(variance_mean);
57 double rel_stddev_stats(double stddev, double avg)
62 pct = 100.0 * stddev/avg;
67 bool __perf_evsel_stat__is(struct perf_evsel *evsel,
68 enum perf_stat_evsel_id id)
70 struct perf_stat_evsel *ps = evsel->priv;
75 #define ID(id, name) [PERF_STAT_EVSEL_ID__##id] = #name
76 static const char *id_str[PERF_STAT_EVSEL_ID__MAX] = {
78 ID(CYCLES_IN_TX, cpu/cycles-t/),
79 ID(TRANSACTION_START, cpu/tx-start/),
80 ID(ELISION_START, cpu/el-start/),
81 ID(CYCLES_IN_TX_CP, cpu/cycles-ct/),
82 ID(TOPDOWN_TOTAL_SLOTS, topdown-total-slots),
83 ID(TOPDOWN_SLOTS_ISSUED, topdown-slots-issued),
84 ID(TOPDOWN_SLOTS_RETIRED, topdown-slots-retired),
85 ID(TOPDOWN_FETCH_BUBBLES, topdown-fetch-bubbles),
86 ID(TOPDOWN_RECOVERY_BUBBLES, topdown-recovery-bubbles),
90 void perf_stat_evsel_id_init(struct perf_evsel *evsel)
92 struct perf_stat_evsel *ps = evsel->priv;
95 /* ps->id is 0 hence PERF_STAT_EVSEL_ID__NONE by default */
97 for (i = 0; i < PERF_STAT_EVSEL_ID__MAX; i++) {
98 if (!strcmp(perf_evsel__name(evsel), id_str[i])) {
105 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
108 struct perf_stat_evsel *ps = evsel->priv;
110 for (i = 0; i < 3; i++)
111 init_stats(&ps->res_stats[i]);
113 perf_stat_evsel_id_init(evsel);
116 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
118 evsel->priv = zalloc(sizeof(struct perf_stat_evsel));
119 if (evsel->priv == NULL)
121 perf_evsel__reset_stat_priv(evsel);
125 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
130 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel,
131 int ncpus, int nthreads)
133 struct perf_counts *counts;
135 counts = perf_counts__new(ncpus, nthreads);
137 evsel->prev_raw_counts = counts;
139 return counts ? 0 : -ENOMEM;
142 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
144 perf_counts__delete(evsel->prev_raw_counts);
145 evsel->prev_raw_counts = NULL;
148 static int perf_evsel__alloc_stats(struct perf_evsel *evsel, bool alloc_raw)
150 int ncpus = perf_evsel__nr_cpus(evsel);
151 int nthreads = thread_map__nr(evsel->threads);
153 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
154 perf_evsel__alloc_counts(evsel, ncpus, nthreads) < 0 ||
155 (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel, ncpus, nthreads) < 0))
161 int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
163 struct perf_evsel *evsel;
165 evlist__for_each_entry(evlist, evsel) {
166 if (perf_evsel__alloc_stats(evsel, alloc_raw))
173 perf_evlist__free_stats(evlist);
177 void perf_evlist__free_stats(struct perf_evlist *evlist)
179 struct perf_evsel *evsel;
181 evlist__for_each_entry(evlist, evsel) {
182 perf_evsel__free_stat_priv(evsel);
183 perf_evsel__free_counts(evsel);
184 perf_evsel__free_prev_raw_counts(evsel);
188 void perf_evlist__reset_stats(struct perf_evlist *evlist)
190 struct perf_evsel *evsel;
192 evlist__for_each_entry(evlist, evsel) {
193 perf_evsel__reset_stat_priv(evsel);
194 perf_evsel__reset_counts(evsel);
198 static void zero_per_pkg(struct perf_evsel *counter)
200 if (counter->per_pkg_mask)
201 memset(counter->per_pkg_mask, 0, MAX_NR_CPUS);
204 static int check_per_pkg(struct perf_evsel *counter,
205 struct perf_counts_values *vals, int cpu, bool *skip)
207 unsigned long *mask = counter->per_pkg_mask;
208 struct cpu_map *cpus = perf_evsel__cpus(counter);
213 if (!counter->per_pkg)
216 if (cpu_map__empty(cpus))
220 mask = zalloc(MAX_NR_CPUS);
224 counter->per_pkg_mask = mask;
228 * we do not consider an event that has not run as a good
229 * instance to mark a package as used (skip=1). Otherwise
230 * we may run into a situation where the first CPU in a package
231 * is not running anything, yet the second is, and this function
232 * would mark the package as used after the first CPU and would
233 * not read the values from the second CPU.
235 if (!(vals->run && vals->ena))
238 s = cpu_map__get_socket(cpus, cpu, NULL);
242 *skip = test_and_set_bit(s, mask) == 1;
247 process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel,
249 struct perf_counts_values *count)
251 struct perf_counts_values *aggr = &evsel->counts->aggr;
252 static struct perf_counts_values zero;
255 if (check_per_pkg(evsel, count, cpu, &skip)) {
256 pr_err("failed to read per-pkg counter\n");
263 switch (config->aggr_mode) {
268 if (!evsel->snapshot)
269 perf_evsel__compute_deltas(evsel, cpu, thread, count);
270 perf_counts_values__scale(count, config->scale, NULL);
271 if (config->aggr_mode == AGGR_NONE)
272 perf_stat__update_shadow_stats(evsel, count->values, cpu);
275 aggr->val += count->val;
277 aggr->ena += count->ena;
278 aggr->run += count->run;
288 static int process_counter_maps(struct perf_stat_config *config,
289 struct perf_evsel *counter)
291 int nthreads = thread_map__nr(counter->threads);
292 int ncpus = perf_evsel__nr_cpus(counter);
295 if (counter->system_wide)
298 for (thread = 0; thread < nthreads; thread++) {
299 for (cpu = 0; cpu < ncpus; cpu++) {
300 if (process_counter_values(config, counter, cpu, thread,
301 perf_counts(counter->counts, cpu, thread)))
309 int perf_stat_process_counter(struct perf_stat_config *config,
310 struct perf_evsel *counter)
312 struct perf_counts_values *aggr = &counter->counts->aggr;
313 struct perf_stat_evsel *ps = counter->priv;
314 u64 *count = counter->counts->aggr.values;
318 aggr->val = aggr->ena = aggr->run = 0;
321 * We calculate counter's data every interval,
322 * and the display code shows ps->res_stats
323 * avg value. We need to zero the stats for
324 * interval mode, otherwise overall avg running
325 * averages will be shown for each interval.
327 if (config->interval)
328 init_stats(ps->res_stats);
330 if (counter->per_pkg)
331 zero_per_pkg(counter);
333 ret = process_counter_maps(config, counter);
337 if (config->aggr_mode != AGGR_GLOBAL)
340 if (!counter->snapshot)
341 perf_evsel__compute_deltas(counter, -1, -1, aggr);
342 perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled);
344 for (i = 0; i < 3; i++)
345 update_stats(&ps->res_stats[i], count[i]);
348 fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
349 perf_evsel__name(counter), count[0], count[1], count[2]);
353 * Save the full runtime - to allow normalization during printout:
355 val = counter->scale * *count;
356 perf_stat__update_shadow_stats(counter, &val, 0);
361 int perf_event__process_stat_event(struct perf_tool *tool __maybe_unused,
362 union perf_event *event,
363 struct perf_session *session)
365 struct perf_counts_values count;
366 struct stat_event *st = &event->stat;
367 struct perf_evsel *counter;
373 counter = perf_evlist__id2evsel(session->evlist, st->id);
375 pr_err("Failed to resolve counter for stat event.\n");
379 *perf_counts(counter->counts, st->cpu, st->thread) = count;
380 counter->supported = true;
384 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
386 struct stat_event *st = (struct stat_event *) event;
389 ret = fprintf(fp, "\n... id %" PRIu64 ", cpu %d, thread %d\n",
390 st->id, st->cpu, st->thread);
391 ret += fprintf(fp, "... value %" PRIu64 ", enabled %" PRIu64 ", running %" PRIu64 "\n",
392 st->val, st->ena, st->run);
397 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
399 struct stat_round_event *rd = (struct stat_round_event *)event;
402 ret = fprintf(fp, "\n... time %" PRIu64 ", type %s\n", rd->time,
403 rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
408 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
410 struct perf_stat_config sc;
413 perf_event__read_stat_config(&sc, &event->stat_config);
415 ret = fprintf(fp, "\n");
416 ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
417 ret += fprintf(fp, "... scale %d\n", sc.scale);
418 ret += fprintf(fp, "... interval %u\n", sc.interval);