1 // SPDX-License-Identifier: GPL-2.0
15 #include "thread_map.h"
17 #include <linux/zalloc.h>
19 void update_stats(struct stats *stats, u64 val)
24 delta = val - stats->mean;
25 stats->mean += delta / stats->n;
26 stats->M2 += delta*(val - stats->mean);
35 double avg_stats(struct stats *stats)
41 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
43 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
44 * s^2 = -------------------------------
47 * http://en.wikipedia.org/wiki/Stddev
49 * The std dev of the mean is related to the std dev by:
56 double stddev_stats(struct stats *stats)
58 double variance, variance_mean;
63 variance = stats->M2 / (stats->n - 1);
64 variance_mean = variance / stats->n;
66 return sqrt(variance_mean);
69 double rel_stddev_stats(double stddev, double avg)
74 pct = 100.0 * stddev/avg;
79 bool __perf_stat_evsel__is(struct evsel *evsel, 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(TOPDOWN_HEAVY_OPS, topdown-heavy-ops),
103 ID(TOPDOWN_BR_MISPREDICT, topdown-br-mispredict),
104 ID(TOPDOWN_FETCH_LAT, topdown-fetch-lat),
105 ID(TOPDOWN_MEM_BOUND, topdown-mem-bound),
106 ID(SMI_NUM, msr/smi/),
107 ID(APERF, msr/aperf/),
111 static void perf_stat_evsel_id_init(struct evsel *evsel)
113 struct perf_stat_evsel *ps = evsel->stats;
116 /* ps->id is 0 hence PERF_STAT_EVSEL_ID__NONE by default */
118 for (i = 0; i < PERF_STAT_EVSEL_ID__MAX; i++) {
119 if (!strcmp(evsel__name(evsel), id_str[i])) {
126 static void evsel__reset_stat_priv(struct evsel *evsel)
129 struct perf_stat_evsel *ps = evsel->stats;
131 for (i = 0; i < 3; i++)
132 init_stats(&ps->res_stats[i]);
134 perf_stat_evsel_id_init(evsel);
137 static int evsel__alloc_stat_priv(struct evsel *evsel)
139 evsel->stats = zalloc(sizeof(struct perf_stat_evsel));
140 if (evsel->stats == NULL)
142 evsel__reset_stat_priv(evsel);
146 static void evsel__free_stat_priv(struct evsel *evsel)
148 struct perf_stat_evsel *ps = evsel->stats;
151 zfree(&ps->group_data);
152 zfree(&evsel->stats);
155 static int evsel__alloc_prev_raw_counts(struct evsel *evsel)
157 int cpu_map_nr = evsel__nr_cpus(evsel);
158 int nthreads = perf_thread_map__nr(evsel->core.threads);
159 struct perf_counts *counts;
161 counts = perf_counts__new(cpu_map_nr, nthreads);
163 evsel->prev_raw_counts = counts;
165 return counts ? 0 : -ENOMEM;
168 static void evsel__free_prev_raw_counts(struct evsel *evsel)
170 perf_counts__delete(evsel->prev_raw_counts);
171 evsel->prev_raw_counts = NULL;
174 static void evsel__reset_prev_raw_counts(struct evsel *evsel)
176 if (evsel->prev_raw_counts)
177 perf_counts__reset(evsel->prev_raw_counts);
180 static int evsel__alloc_stats(struct evsel *evsel, bool alloc_raw)
182 if (evsel__alloc_stat_priv(evsel) < 0 ||
183 evsel__alloc_counts(evsel) < 0 ||
184 (alloc_raw && evsel__alloc_prev_raw_counts(evsel) < 0))
190 int evlist__alloc_stats(struct evlist *evlist, bool alloc_raw)
194 evlist__for_each_entry(evlist, evsel) {
195 if (evsel__alloc_stats(evsel, alloc_raw))
202 evlist__free_stats(evlist);
206 void evlist__free_stats(struct evlist *evlist)
210 evlist__for_each_entry(evlist, evsel) {
211 evsel__free_stat_priv(evsel);
212 evsel__free_counts(evsel);
213 evsel__free_prev_raw_counts(evsel);
217 void evlist__reset_stats(struct evlist *evlist)
221 evlist__for_each_entry(evlist, evsel) {
222 evsel__reset_stat_priv(evsel);
223 evsel__reset_counts(evsel);
227 void evlist__reset_prev_raw_counts(struct evlist *evlist)
231 evlist__for_each_entry(evlist, evsel)
232 evsel__reset_prev_raw_counts(evsel);
235 static void evsel__copy_prev_raw_counts(struct evsel *evsel)
237 int ncpus = evsel__nr_cpus(evsel);
238 int nthreads = perf_thread_map__nr(evsel->core.threads);
240 for (int thread = 0; thread < nthreads; thread++) {
241 for (int cpu = 0; cpu < ncpus; cpu++) {
242 *perf_counts(evsel->counts, cpu, thread) =
243 *perf_counts(evsel->prev_raw_counts, cpu,
248 evsel->counts->aggr = evsel->prev_raw_counts->aggr;
251 void evlist__copy_prev_raw_counts(struct evlist *evlist)
255 evlist__for_each_entry(evlist, evsel)
256 evsel__copy_prev_raw_counts(evsel);
259 void evlist__save_aggr_prev_raw_counts(struct evlist *evlist)
264 * To collect the overall statistics for interval mode,
265 * we copy the counts from evsel->prev_raw_counts to
266 * evsel->counts. The perf_stat_process_counter creates
267 * aggr values from per cpu values, but the per cpu values
268 * are 0 for AGGR_GLOBAL. So we use a trick that saves the
269 * previous aggr value to the first member of perf_counts,
270 * then aggr calculation in process_counter_values can work
273 evlist__for_each_entry(evlist, evsel) {
274 *perf_counts(evsel->prev_raw_counts, 0, 0) =
275 evsel->prev_raw_counts->aggr;
279 static size_t pkg_id_hash(const void *__key, void *ctx __maybe_unused)
281 uint64_t *key = (uint64_t *) __key;
283 return *key & 0xffffffff;
286 static bool pkg_id_equal(const void *__key1, const void *__key2,
287 void *ctx __maybe_unused)
289 uint64_t *key1 = (uint64_t *) __key1;
290 uint64_t *key2 = (uint64_t *) __key2;
292 return *key1 == *key2;
295 static int check_per_pkg(struct evsel *counter, struct perf_counts_values *vals,
296 int cpu_map_idx, bool *skip)
298 struct hashmap *mask = counter->per_pkg_mask;
299 struct perf_cpu_map *cpus = evsel__cpus(counter);
300 struct perf_cpu cpu = perf_cpu_map__cpu(cpus, cpu_map_idx);
306 if (!counter->per_pkg)
309 if (perf_cpu_map__empty(cpus))
313 mask = hashmap__new(pkg_id_hash, pkg_id_equal, NULL);
317 counter->per_pkg_mask = mask;
321 * we do not consider an event that has not run as a good
322 * instance to mark a package as used (skip=1). Otherwise
323 * we may run into a situation where the first CPU in a package
324 * is not running anything, yet the second is, and this function
325 * would mark the package as used after the first CPU and would
326 * not read the values from the second CPU.
328 if (!(vals->run && vals->ena))
331 s = cpu__get_socket_id(cpu);
336 * On multi-die system, die_id > 0. On no-die system, die_id = 0.
337 * We use hashmap(socket, die) to check the used socket+die pair.
339 d = cpu__get_die_id(cpu);
343 key = malloc(sizeof(*key));
347 *key = (uint64_t)d << 32 | s;
348 if (hashmap__find(mask, (void *)key, NULL)) {
352 ret = hashmap__add(mask, (void *)key, (void *)1);
358 process_counter_values(struct perf_stat_config *config, struct evsel *evsel,
359 int cpu_map_idx, int thread,
360 struct perf_counts_values *count)
362 struct perf_counts_values *aggr = &evsel->counts->aggr;
363 static struct perf_counts_values zero;
366 if (check_per_pkg(evsel, count, cpu_map_idx, &skip)) {
367 pr_err("failed to read per-pkg counter\n");
374 switch (config->aggr_mode) {
381 if (!evsel->snapshot)
382 evsel__compute_deltas(evsel, cpu_map_idx, thread, count);
383 perf_counts_values__scale(count, config->scale, NULL);
384 if ((config->aggr_mode == AGGR_NONE) && (!evsel->percore)) {
385 perf_stat__update_shadow_stats(evsel, count->val,
386 cpu_map_idx, &rt_stat);
389 if (config->aggr_mode == AGGR_THREAD) {
391 perf_stat__update_shadow_stats(evsel,
392 count->val, 0, &config->stats[thread]);
394 perf_stat__update_shadow_stats(evsel,
395 count->val, 0, &rt_stat);
399 aggr->val += count->val;
400 aggr->ena += count->ena;
401 aggr->run += count->run;
410 static int process_counter_maps(struct perf_stat_config *config,
411 struct evsel *counter)
413 int nthreads = perf_thread_map__nr(counter->core.threads);
414 int ncpus = evsel__nr_cpus(counter);
417 if (counter->core.system_wide)
420 for (thread = 0; thread < nthreads; thread++) {
421 for (idx = 0; idx < ncpus; idx++) {
422 if (process_counter_values(config, counter, idx, thread,
423 perf_counts(counter->counts, idx, thread)))
431 int perf_stat_process_counter(struct perf_stat_config *config,
432 struct evsel *counter)
434 struct perf_counts_values *aggr = &counter->counts->aggr;
435 struct perf_stat_evsel *ps = counter->stats;
436 u64 *count = counter->counts->aggr.values;
439 aggr->val = aggr->ena = aggr->run = 0;
441 if (counter->per_pkg)
442 evsel__zero_per_pkg(counter);
444 ret = process_counter_maps(config, counter);
448 if (config->aggr_mode != AGGR_GLOBAL)
451 if (!counter->snapshot)
452 evsel__compute_deltas(counter, -1, -1, aggr);
453 perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled);
455 for (i = 0; i < 3; i++)
456 update_stats(&ps->res_stats[i], count[i]);
459 fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
460 evsel__name(counter), count[0], count[1], count[2]);
464 * Save the full runtime - to allow normalization during printout:
466 perf_stat__update_shadow_stats(counter, *count, 0, &rt_stat);
471 int perf_event__process_stat_event(struct perf_session *session,
472 union perf_event *event)
474 struct perf_counts_values count;
475 struct perf_record_stat *st = &event->stat;
476 struct evsel *counter;
482 counter = evlist__id2evsel(session->evlist, st->id);
484 pr_err("Failed to resolve counter for stat event.\n");
488 *perf_counts(counter->counts, st->cpu, st->thread) = count;
489 counter->supported = true;
493 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
495 struct perf_record_stat *st = (struct perf_record_stat *)event;
498 ret = fprintf(fp, "\n... id %" PRI_lu64 ", cpu %d, thread %d\n",
499 st->id, st->cpu, st->thread);
500 ret += fprintf(fp, "... value %" PRI_lu64 ", enabled %" PRI_lu64 ", running %" PRI_lu64 "\n",
501 st->val, st->ena, st->run);
506 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
508 struct perf_record_stat_round *rd = (struct perf_record_stat_round *)event;
511 ret = fprintf(fp, "\n... time %" PRI_lu64 ", type %s\n", rd->time,
512 rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
517 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
519 struct perf_stat_config sc;
522 perf_event__read_stat_config(&sc, &event->stat_config);
524 ret = fprintf(fp, "\n");
525 ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
526 ret += fprintf(fp, "... scale %d\n", sc.scale);
527 ret += fprintf(fp, "... interval %u\n", sc.interval);
532 int create_perf_stat_counter(struct evsel *evsel,
533 struct perf_stat_config *config,
534 struct target *target,
537 struct perf_event_attr *attr = &evsel->core.attr;
538 struct evsel *leader = evsel__leader(evsel);
540 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
541 PERF_FORMAT_TOTAL_TIME_RUNNING;
544 * The event is part of non trivial group, let's enable
545 * the group read (for leader) and ID retrieval for all
548 if (leader->core.nr_members > 1)
549 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
551 attr->inherit = !config->no_inherit && list_empty(&evsel->bpf_counter_list);
554 * Some events get initialized with sample_(period/type) set,
555 * like tracepoints. Clear it up for counting.
557 attr->sample_period = 0;
559 if (config->identifier)
560 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
562 if (config->all_user) {
563 attr->exclude_kernel = 1;
564 attr->exclude_user = 0;
567 if (config->all_kernel) {
568 attr->exclude_kernel = 0;
569 attr->exclude_user = 1;
573 * Disabling all counters initially, they will be enabled
574 * either manually by us or by kernel via enable_on_exec
577 if (evsel__is_group_leader(evsel)) {
581 * In case of initial_delay we enable tracee
584 if (target__none(target) && !config->initial_delay)
585 attr->enable_on_exec = 1;
588 if (target__has_cpu(target) && !target__has_per_thread(target))
589 return evsel__open_per_cpu(evsel, evsel__cpus(evsel), cpu_map_idx);
591 return evsel__open_per_thread(evsel, evsel->core.threads);