stat.c

  1// SPDX-License-Identifier: GPL-2.0
  2#include <errno.h>
  3#include <linux/err.h>
  4#include <inttypes.h>
  5#include <math.h>
  6#include <string.h>
  7#include "counts.h"
  8#include "cpumap.h"
  9#include "debug.h"
 10#include "header.h"
 11#include "stat.h"
 12#include "session.h"
 13#include "target.h"
 14#include "evlist.h"
 15#include "evsel.h"
 16#include "thread_map.h"
 17#include "util/hashmap.h"
 18#include <linux/zalloc.h>
 19
 20void update_stats(struct stats *stats, u64 val)
 21{
 22	double delta;
 23
 24	stats->n++;
 25	delta = val - stats->mean;
 26	stats->mean += delta / stats->n;
 27	stats->M2 += delta*(val - stats->mean);
 28
 29	if (val > stats->max)
 30		stats->max = val;
 31
 32	if (val < stats->min)
 33		stats->min = val;
 34}
 35
 36double avg_stats(struct stats *stats)
 37{
 38	return stats->mean;
 39}
 40
 41/*
 42 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
 43 *
 44 *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
 45 * s^2 = -------------------------------
 46 *                  n - 1
 47 *
 48 * http://en.wikipedia.org/wiki/Stddev
 49 *
 50 * The std dev of the mean is related to the std dev by:
 51 *
 52 *             s
 53 * s_mean = -------
 54 *          sqrt(n)
 55 *
 56 */
 57double stddev_stats(struct stats *stats)
 58{
 59	double variance, variance_mean;
 60
 61	if (stats->n < 2)
 62		return 0.0;
 63
 64	variance = stats->M2 / (stats->n - 1);
 65	variance_mean = variance / stats->n;
 66
 67	return sqrt(variance_mean);
 68}
 69
 70double rel_stddev_stats(double stddev, double avg)
 71{
 72	double pct = 0.0;
 73
 74	if (avg)
 75		pct = 100.0 * stddev/avg;
 76
 77	return pct;
 78}
 79
 80static void evsel__reset_aggr_stats(struct evsel *evsel)
 81{
 82	struct perf_stat_evsel *ps = evsel->stats;
 83	struct perf_stat_aggr *aggr = ps->aggr;
 84
 85	if (aggr)
 86		memset(aggr, 0, sizeof(*aggr) * ps->nr_aggr);
 87}
 88
 89static void evsel__reset_stat_priv(struct evsel *evsel)
 90{
 91	struct perf_stat_evsel *ps = evsel->stats;
 92
 93	init_stats(&ps->res_stats);
 94	evsel__reset_aggr_stats(evsel);
 95}
 96
 97static int evsel__alloc_aggr_stats(struct evsel *evsel, int nr_aggr)
 98{
 99	struct perf_stat_evsel *ps = evsel->stats;
100
101	if (ps == NULL)
102		return 0;
103
104	ps->nr_aggr = nr_aggr;
105	ps->aggr = calloc(nr_aggr, sizeof(*ps->aggr));
106	if (ps->aggr == NULL)
107		return -ENOMEM;
108
109	return 0;
110}
111
112int evlist__alloc_aggr_stats(struct evlist *evlist, int nr_aggr)
113{
114	struct evsel *evsel;
115
116	evlist__for_each_entry(evlist, evsel) {
117		if (evsel__alloc_aggr_stats(evsel, nr_aggr) < 0)
118			return -1;
119	}
120	return 0;
121}
122
123static int evsel__alloc_stat_priv(struct evsel *evsel, int nr_aggr)
124{
125	struct perf_stat_evsel *ps;
126
127	ps = zalloc(sizeof(*ps));
128	if (ps == NULL)
129		return -ENOMEM;
130
131	evsel->stats = ps;
132
133	if (nr_aggr && evsel__alloc_aggr_stats(evsel, nr_aggr) < 0) {
134		evsel->stats = NULL;
135		free(ps);
136		return -ENOMEM;
137	}
138
139	evsel__reset_stat_priv(evsel);
140	return 0;
141}
142
143static void evsel__free_stat_priv(struct evsel *evsel)
144{
145	struct perf_stat_evsel *ps = evsel->stats;
146
147	if (ps) {
148		zfree(&ps->aggr);
149		zfree(&ps->group_data);
150	}
151	zfree(&evsel->stats);
152}
153
154static int evsel__alloc_prev_raw_counts(struct evsel *evsel)
155{
156	int cpu_map_nr = evsel__nr_cpus(evsel);
157	int nthreads = perf_thread_map__nr(evsel->core.threads);
158	struct perf_counts *counts;
159
160	counts = perf_counts__new(cpu_map_nr, nthreads);
161	if (counts)
162		evsel->prev_raw_counts = counts;
163
164	return counts ? 0 : -ENOMEM;
165}
166
167static void evsel__free_prev_raw_counts(struct evsel *evsel)
168{
169	perf_counts__delete(evsel->prev_raw_counts);
170	evsel->prev_raw_counts = NULL;
171}
172
173static void evsel__reset_prev_raw_counts(struct evsel *evsel)
174{
175	if (evsel->prev_raw_counts)
176		perf_counts__reset(evsel->prev_raw_counts);
177}
178
179static int evsel__alloc_stats(struct evsel *evsel, int nr_aggr, bool alloc_raw)
180{
181	if (evsel__alloc_stat_priv(evsel, nr_aggr) < 0 ||
182	    evsel__alloc_counts(evsel) < 0 ||
183	    (alloc_raw && evsel__alloc_prev_raw_counts(evsel) < 0))
184		return -ENOMEM;
185
186	return 0;
187}
188
189int evlist__alloc_stats(struct perf_stat_config *config,
190			struct evlist *evlist, bool alloc_raw)
191{
192	struct evsel *evsel;
193	int nr_aggr = 0;
194
195	if (config && config->aggr_map)
196		nr_aggr = config->aggr_map->nr;
197
198	evlist__for_each_entry(evlist, evsel) {
199		if (evsel__alloc_stats(evsel, nr_aggr, alloc_raw))
200			goto out_free;
201	}
202
203	return 0;
204
205out_free:
206	evlist__free_stats(evlist);
207	return -1;
208}
209
210void evlist__free_stats(struct evlist *evlist)
211{
212	struct evsel *evsel;
213
214	evlist__for_each_entry(evlist, evsel) {
215		evsel__free_stat_priv(evsel);
216		evsel__free_counts(evsel);
217		evsel__free_prev_raw_counts(evsel);
218	}
219}
220
221void evlist__reset_stats(struct evlist *evlist)
222{
223	struct evsel *evsel;
224
225	evlist__for_each_entry(evlist, evsel) {
226		evsel__reset_stat_priv(evsel);
227		evsel__reset_counts(evsel);
228	}
229}
230
231void evlist__reset_aggr_stats(struct evlist *evlist)
232{
233	struct evsel *evsel;
234
235	evlist__for_each_entry(evlist, evsel)
236		evsel__reset_aggr_stats(evsel);
237}
238
239void evlist__reset_prev_raw_counts(struct evlist *evlist)
240{
241	struct evsel *evsel;
242
243	evlist__for_each_entry(evlist, evsel)
244		evsel__reset_prev_raw_counts(evsel);
245}
246
247static void evsel__copy_prev_raw_counts(struct evsel *evsel)
248{
249	int idx, nthreads = perf_thread_map__nr(evsel->core.threads);
250
251	for (int thread = 0; thread < nthreads; thread++) {
252		perf_cpu_map__for_each_idx(idx, evsel__cpus(evsel)) {
253			*perf_counts(evsel->counts, idx, thread) =
254				*perf_counts(evsel->prev_raw_counts, idx, thread);
255		}
256	}
257}
258
259void evlist__copy_prev_raw_counts(struct evlist *evlist)
260{
261	struct evsel *evsel;
262
263	evlist__for_each_entry(evlist, evsel)
264		evsel__copy_prev_raw_counts(evsel);
265}
266
267static void evsel__copy_res_stats(struct evsel *evsel)
268{
269	struct perf_stat_evsel *ps = evsel->stats;
270
271	/*
272	 * For GLOBAL aggregation mode, it updates the counts for each run
273	 * in the evsel->stats.res_stats.  See perf_stat_process_counter().
274	 */
275	*ps->aggr[0].counts.values = avg_stats(&ps->res_stats);
276}
277
278void evlist__copy_res_stats(struct perf_stat_config *config, struct evlist *evlist)
279{
280	struct evsel *evsel;
281
282	if (config->aggr_mode != AGGR_GLOBAL)
283		return;
284
285	evlist__for_each_entry(evlist, evsel)
286		evsel__copy_res_stats(evsel);
287}
288
289static size_t pkg_id_hash(long __key, void *ctx __maybe_unused)
290{
291	uint64_t *key = (uint64_t *) __key;
292
293	return *key & 0xffffffff;
294}
295
296static bool pkg_id_equal(long __key1, long __key2, void *ctx __maybe_unused)
297{
298	uint64_t *key1 = (uint64_t *) __key1;
299	uint64_t *key2 = (uint64_t *) __key2;
300
301	return *key1 == *key2;
302}
303
304static int check_per_pkg(struct evsel *counter, struct perf_counts_values *vals,
305			 int cpu_map_idx, bool *skip)
306{
307	struct hashmap *mask = counter->per_pkg_mask;
308	struct perf_cpu_map *cpus = evsel__cpus(counter);
309	struct perf_cpu cpu = perf_cpu_map__cpu(cpus, cpu_map_idx);
310	int s, d, ret = 0;
311	uint64_t *key;
312
313	*skip = false;
314
315	if (!counter->per_pkg)
316		return 0;
317
318	if (perf_cpu_map__is_any_cpu_or_is_empty(cpus))
319		return 0;
320
321	if (!mask) {
322		mask = hashmap__new(pkg_id_hash, pkg_id_equal, NULL);
323		if (IS_ERR(mask))
324			return -ENOMEM;
325
326		counter->per_pkg_mask = mask;
327	}
328
329	/*
330	 * we do not consider an event that has not run as a good
331	 * instance to mark a package as used (skip=1). Otherwise
332	 * we may run into a situation where the first CPU in a package
333	 * is not running anything, yet the second is, and this function
334	 * would mark the package as used after the first CPU and would
335	 * not read the values from the second CPU.
336	 */
337	if (!(vals->run && vals->ena))
338		return 0;
339
340	s = cpu__get_socket_id(cpu);
341	if (s < 0)
342		return -1;
343
344	/*
345	 * On multi-die system, die_id > 0. On no-die system, die_id = 0.
346	 * We use hashmap(socket, die) to check the used socket+die pair.
347	 */
348	d = cpu__get_die_id(cpu);
349	if (d < 0)
350		return -1;
351
352	key = malloc(sizeof(*key));
353	if (!key)
354		return -ENOMEM;
355
356	*key = (uint64_t)d << 32 | s;
357	if (hashmap__find(mask, key, NULL)) {
358		*skip = true;
359		free(key);
360	} else
361		ret = hashmap__add(mask, key, 1);
362
363	return ret;
364}
365
366static bool evsel__count_has_error(struct evsel *evsel,
367				   struct perf_counts_values *count,
368				   struct perf_stat_config *config)
369{
370	/* the evsel was failed already */
371	if (evsel->err || evsel->counts->scaled == -1)
372		return true;
373
374	/* this is meaningful for CPU aggregation modes only */
375	if (config->aggr_mode == AGGR_GLOBAL)
376		return false;
377
378	/* it's considered ok when it actually ran */
379	if (count->ena != 0 && count->run != 0)
380		return false;
381
382	return true;
383}
384
385static int
386process_counter_values(struct perf_stat_config *config, struct evsel *evsel,
387		       int cpu_map_idx, int thread,
388		       struct perf_counts_values *count)
389{
390	struct perf_stat_evsel *ps = evsel->stats;
391	static struct perf_counts_values zero;
392	bool skip = false;
393
394	if (check_per_pkg(evsel, count, cpu_map_idx, &skip)) {
395		pr_err("failed to read per-pkg counter\n");
396		return -1;
397	}
398
399	if (skip)
400		count = &zero;
401
402	if (!evsel->snapshot)
403		evsel__compute_deltas(evsel, cpu_map_idx, thread, count);
404	perf_counts_values__scale(count, config->scale, NULL);
405
406	if (config->aggr_mode == AGGR_THREAD) {
407		struct perf_counts_values *aggr_counts = &ps->aggr[thread].counts;
408
409		/*
410		 * Skip value 0 when enabling --per-thread globally,
411		 * otherwise too many 0 output.
412		 */
413		if (count->val == 0 && config->system_wide)
414			return 0;
415
416		ps->aggr[thread].nr++;
417
418		aggr_counts->val += count->val;
419		aggr_counts->ena += count->ena;
420		aggr_counts->run += count->run;
421		return 0;
422	}
423
424	if (ps->aggr) {
425		struct perf_cpu cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
426		struct aggr_cpu_id aggr_id = config->aggr_get_id(config, cpu);
427		struct perf_stat_aggr *ps_aggr;
428		int i;
429
430		for (i = 0; i < ps->nr_aggr; i++) {
431			if (!aggr_cpu_id__equal(&aggr_id, &config->aggr_map->map[i]))
432				continue;
433
434			ps_aggr = &ps->aggr[i];
435			ps_aggr->nr++;
436
437			/*
438			 * When any result is bad, make them all to give consistent output
439			 * in interval mode.  But per-task counters can have 0 enabled time
440			 * when some tasks are idle.
441			 */
442			if (evsel__count_has_error(evsel, count, config) && !ps_aggr->failed) {
443				ps_aggr->counts.val = 0;
444				ps_aggr->counts.ena = 0;
445				ps_aggr->counts.run = 0;
446				ps_aggr->failed = true;
447			}
448
449			if (!ps_aggr->failed) {
450				ps_aggr->counts.val += count->val;
451				ps_aggr->counts.ena += count->ena;
452				ps_aggr->counts.run += count->run;
453			}
454			break;
455		}
456	}
457
458	return 0;
459}
460
461static int process_counter_maps(struct perf_stat_config *config,
462				struct evsel *counter)
463{
464	int nthreads = perf_thread_map__nr(counter->core.threads);
465	int ncpus = evsel__nr_cpus(counter);
466	int idx, thread;
467
468	for (thread = 0; thread < nthreads; thread++) {
469		for (idx = 0; idx < ncpus; idx++) {
470			if (process_counter_values(config, counter, idx, thread,
471						   perf_counts(counter->counts, idx, thread)))
472				return -1;
473		}
474	}
475
476	return 0;
477}
478
479int perf_stat_process_counter(struct perf_stat_config *config,
480			      struct evsel *counter)
481{
482	struct perf_stat_evsel *ps = counter->stats;
483	u64 *count;
484	int ret;
485
486	if (counter->per_pkg)
487		evsel__zero_per_pkg(counter);
488
489	ret = process_counter_maps(config, counter);
490	if (ret)
491		return ret;
492
493	if (config->aggr_mode != AGGR_GLOBAL)
494		return 0;
495
496	/*
497	 * GLOBAL aggregation mode only has a single aggr counts,
498	 * so we can use ps->aggr[0] as the actual output.
499	 */
500	count = ps->aggr[0].counts.values;
501	update_stats(&ps->res_stats, *count);
502
503	if (verbose > 0) {
504		fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
505			evsel__name(counter), count[0], count[1], count[2]);
506	}
507
508	return 0;
509}
510
511static int evsel__merge_aggr_counters(struct evsel *evsel, struct evsel *alias)
512{
513	struct perf_stat_evsel *ps_a = evsel->stats;
514	struct perf_stat_evsel *ps_b = alias->stats;
515	int i;
516
517	if (ps_a->aggr == NULL && ps_b->aggr == NULL)
518		return 0;
519
520	if (ps_a->nr_aggr != ps_b->nr_aggr) {
521		pr_err("Unmatched aggregation mode between aliases\n");
522		return -1;
523	}
524
525	for (i = 0; i < ps_a->nr_aggr; i++) {
526		struct perf_counts_values *aggr_counts_a = &ps_a->aggr[i].counts;
527		struct perf_counts_values *aggr_counts_b = &ps_b->aggr[i].counts;
528
529		ps_a->aggr[i].nr += ps_b->aggr[i].nr;
530
531		aggr_counts_a->val += aggr_counts_b->val;
532		aggr_counts_a->ena += aggr_counts_b->ena;
533		aggr_counts_a->run += aggr_counts_b->run;
534	}
535
536	return 0;
537}
538
539static void evsel__merge_aliases(struct evsel *evsel)
540{
541	struct evlist *evlist = evsel->evlist;
542	struct evsel *alias;
543
544	alias = list_prepare_entry(evsel, &(evlist->core.entries), core.node);
545	list_for_each_entry_continue(alias, &evlist->core.entries, core.node) {
546		if (alias->first_wildcard_match == evsel) {
547			/* Merge the same events on different PMUs. */
548			evsel__merge_aggr_counters(evsel, alias);
549		}
550	}
551}
552
553static bool evsel__should_merge_hybrid(const struct evsel *evsel,
554				       const struct perf_stat_config *config)
555{
556	return config->hybrid_merge && evsel__is_hybrid(evsel);
557}
558
559static void evsel__merge_stats(struct evsel *evsel, struct perf_stat_config *config)
560{
561	if (!evsel->pmu || !evsel->pmu->is_core || evsel__should_merge_hybrid(evsel, config))
562		evsel__merge_aliases(evsel);
563}
564
565/* merge the same uncore and hybrid events if requested */
566void perf_stat_merge_counters(struct perf_stat_config *config, struct evlist *evlist)
567{
568	struct evsel *evsel;
569
570	if (config->aggr_mode == AGGR_NONE)
571		return;
572
573	evlist__for_each_entry(evlist, evsel)
574		evsel__merge_stats(evsel, config);
575}
576
577static void evsel__update_percore_stats(struct evsel *evsel, struct aggr_cpu_id *core_id)
578{
579	struct perf_stat_evsel *ps = evsel->stats;
580	struct perf_counts_values counts = { 0, };
581	struct aggr_cpu_id id;
582	struct perf_cpu cpu;
583	int idx;
584
585	/* collect per-core counts */
586	perf_cpu_map__for_each_cpu(cpu, idx, evsel->core.cpus) {
587		struct perf_stat_aggr *aggr = &ps->aggr[idx];
588
589		id = aggr_cpu_id__core(cpu, NULL);
590		if (!aggr_cpu_id__equal(core_id, &id))
591			continue;
592
593		counts.val += aggr->counts.val;
594		counts.ena += aggr->counts.ena;
595		counts.run += aggr->counts.run;
596	}
597
598	/* update aggregated per-core counts for each CPU */
599	perf_cpu_map__for_each_cpu(cpu, idx, evsel->core.cpus) {
600		struct perf_stat_aggr *aggr = &ps->aggr[idx];
601
602		id = aggr_cpu_id__core(cpu, NULL);
603		if (!aggr_cpu_id__equal(core_id, &id))
604			continue;
605
606		aggr->counts.val = counts.val;
607		aggr->counts.ena = counts.ena;
608		aggr->counts.run = counts.run;
609
610		aggr->used = true;
611	}
612}
613
614/* we have an aggr_map for cpu, but want to aggregate the counters per-core */
615static void evsel__process_percore(struct evsel *evsel)
616{
617	struct perf_stat_evsel *ps = evsel->stats;
618	struct aggr_cpu_id core_id;
619	struct perf_cpu cpu;
620	int idx;
621
622	if (!evsel->percore)
623		return;
624
625	perf_cpu_map__for_each_cpu(cpu, idx, evsel->core.cpus) {
626		struct perf_stat_aggr *aggr = &ps->aggr[idx];
627
628		if (aggr->used)
629			continue;
630
631		core_id = aggr_cpu_id__core(cpu, NULL);
632		evsel__update_percore_stats(evsel, &core_id);
633	}
634}
635
636/* process cpu stats on per-core events */
637void perf_stat_process_percore(struct perf_stat_config *config, struct evlist *evlist)
638{
639	struct evsel *evsel;
640
641	if (config->aggr_mode != AGGR_NONE)
642		return;
643
644	evlist__for_each_entry(evlist, evsel)
645		evsel__process_percore(evsel);
646}
647
648int perf_event__process_stat_event(const struct perf_tool *tool __maybe_unused,
649				   struct perf_session *session,
650				   union perf_event *event)
651{
652	struct perf_counts_values count, *ptr;
653	struct perf_record_stat *st = &event->stat;
654	struct evsel *counter;
655	int cpu_map_idx;
656
657	count.val = st->val;
658	count.ena = st->ena;
659	count.run = st->run;
660
661	counter = evlist__id2evsel(session->evlist, st->id);
662	if (!counter) {
663		pr_err("Failed to resolve counter for stat event.\n");
664		return -EINVAL;
665	}
666	cpu_map_idx = perf_cpu_map__idx(evsel__cpus(counter), (struct perf_cpu){.cpu = st->cpu});
667	if (cpu_map_idx == -1) {
668		pr_err("Invalid CPU %d for event %s.\n", st->cpu, evsel__name(counter));
669		return -EINVAL;
670	}
671	ptr = perf_counts(counter->counts, cpu_map_idx, st->thread);
672	if (ptr == NULL) {
673		pr_err("Failed to find perf count for CPU %d thread %d on event %s.\n",
674			st->cpu, st->thread, evsel__name(counter));
675		return -EINVAL;
676	}
677	*ptr = count;
678	counter->supported = true;
679	return 0;
680}
681
682size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
683{
684	struct perf_record_stat *st = (struct perf_record_stat *)event;
685	size_t ret;
686
687	ret  = fprintf(fp, "\n... id %" PRI_lu64 ", cpu %d, thread %d\n",
688		       st->id, st->cpu, st->thread);
689	ret += fprintf(fp, "... value %" PRI_lu64 ", enabled %" PRI_lu64 ", running %" PRI_lu64 "\n",
690		       st->val, st->ena, st->run);
691
692	return ret;
693}
694
695size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
696{
697	struct perf_record_stat_round *rd = (struct perf_record_stat_round *)event;
698	size_t ret;
699
700	ret = fprintf(fp, "\n... time %" PRI_lu64 ", type %s\n", rd->time,
701		      rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
702
703	return ret;
704}
705
706size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
707{
708	struct perf_stat_config sc = {};
709	size_t ret;
710
711	perf_event__read_stat_config(&sc, &event->stat_config);
712
713	ret  = fprintf(fp, "\n");
714	ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
715	ret += fprintf(fp, "... scale     %d\n", sc.scale);
716	ret += fprintf(fp, "... interval  %u\n", sc.interval);
717
718	return ret;
719}