xref: /linux/tools/perf/util/stat.c (revision 704fd176204577459beadb37d46e164d376fabc3) 
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 "hashmap.h"
18 #include <linux/zalloc.h>
19 
20 void 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 
36 double 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  */
57 double 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 
70 double 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 
80 bool __perf_stat_evsel__is(struct evsel *evsel, enum perf_stat_evsel_id id)
81 {
82 	struct perf_stat_evsel *ps = evsel->stats;
83 
84 	return ps->id == id;
85 }
86 
87 #define ID(id, name) [PERF_STAT_EVSEL_ID__##id] = #name
88 static const char *id_str[PERF_STAT_EVSEL_ID__MAX] = {
89 	ID(NONE,		x),
90 	ID(CYCLES_IN_TX,	cpu/cycles-t/),
91 	ID(TRANSACTION_START,	cpu/tx-start/),
92 	ID(ELISION_START,	cpu/el-start/),
93 	ID(CYCLES_IN_TX_CP,	cpu/cycles-ct/),
94 	ID(TOPDOWN_TOTAL_SLOTS, topdown-total-slots),
95 	ID(TOPDOWN_SLOTS_ISSUED, topdown-slots-issued),
96 	ID(TOPDOWN_SLOTS_RETIRED, topdown-slots-retired),
97 	ID(TOPDOWN_FETCH_BUBBLES, topdown-fetch-bubbles),
98 	ID(TOPDOWN_RECOVERY_BUBBLES, topdown-recovery-bubbles),
99 	ID(TOPDOWN_RETIRING, topdown-retiring),
100 	ID(TOPDOWN_BAD_SPEC, topdown-bad-spec),
101 	ID(TOPDOWN_FE_BOUND, topdown-fe-bound),
102 	ID(TOPDOWN_BE_BOUND, topdown-be-bound),
103 	ID(TOPDOWN_HEAVY_OPS, topdown-heavy-ops),
104 	ID(TOPDOWN_BR_MISPREDICT, topdown-br-mispredict),
105 	ID(TOPDOWN_FETCH_LAT, topdown-fetch-lat),
106 	ID(TOPDOWN_MEM_BOUND, topdown-mem-bound),
107 	ID(SMI_NUM, msr/smi/),
108 	ID(APERF, msr/aperf/),
109 };
110 #undef ID
111 
112 static void perf_stat_evsel_id_init(struct evsel *evsel)
113 {
114 	struct perf_stat_evsel *ps = evsel->stats;
115 	int i;
116 
117 	/* ps->id is 0 hence PERF_STAT_EVSEL_ID__NONE by default */
118 
119 	for (i = 0; i < PERF_STAT_EVSEL_ID__MAX; i++) {
120 		if (!strcmp(evsel__name(evsel), id_str[i])) {
121 			ps->id = i;
122 			break;
123 		}
124 	}
125 }
126 
127 static void evsel__reset_stat_priv(struct evsel *evsel)
128 {
129 	int i;
130 	struct perf_stat_evsel *ps = evsel->stats;
131 
132 	for (i = 0; i < 3; i++)
133 		init_stats(&ps->res_stats[i]);
134 
135 	perf_stat_evsel_id_init(evsel);
136 }
137 
138 static int evsel__alloc_stat_priv(struct evsel *evsel)
139 {
140 	evsel->stats = zalloc(sizeof(struct perf_stat_evsel));
141 	if (evsel->stats == NULL)
142 		return -ENOMEM;
143 	evsel__reset_stat_priv(evsel);
144 	return 0;
145 }
146 
147 static void evsel__free_stat_priv(struct evsel *evsel)
148 {
149 	struct perf_stat_evsel *ps = evsel->stats;
150 
151 	if (ps)
152 		zfree(&ps->group_data);
153 	zfree(&evsel->stats);
154 }
155 
156 static int evsel__alloc_prev_raw_counts(struct evsel *evsel)
157 {
158 	int cpu_map_nr = evsel__nr_cpus(evsel);
159 	int nthreads = perf_thread_map__nr(evsel->core.threads);
160 	struct perf_counts *counts;
161 
162 	counts = perf_counts__new(cpu_map_nr, nthreads);
163 	if (counts)
164 		evsel->prev_raw_counts = counts;
165 
166 	return counts ? 0 : -ENOMEM;
167 }
168 
169 static void evsel__free_prev_raw_counts(struct evsel *evsel)
170 {
171 	perf_counts__delete(evsel->prev_raw_counts);
172 	evsel->prev_raw_counts = NULL;
173 }
174 
175 static void evsel__reset_prev_raw_counts(struct evsel *evsel)
176 {
177 	if (evsel->prev_raw_counts)
178 		perf_counts__reset(evsel->prev_raw_counts);
179 }
180 
181 static int evsel__alloc_stats(struct evsel *evsel, bool alloc_raw)
182 {
183 	if (evsel__alloc_stat_priv(evsel) < 0 ||
184 	    evsel__alloc_counts(evsel) < 0 ||
185 	    (alloc_raw && evsel__alloc_prev_raw_counts(evsel) < 0))
186 		return -ENOMEM;
187 
188 	return 0;
189 }
190 
191 int evlist__alloc_stats(struct evlist *evlist, bool alloc_raw)
192 {
193 	struct evsel *evsel;
194 
195 	evlist__for_each_entry(evlist, evsel) {
196 		if (evsel__alloc_stats(evsel, alloc_raw))
197 			goto out_free;
198 	}
199 
200 	return 0;
201 
202 out_free:
203 	evlist__free_stats(evlist);
204 	return -1;
205 }
206 
207 void evlist__free_stats(struct evlist *evlist)
208 {
209 	struct evsel *evsel;
210 
211 	evlist__for_each_entry(evlist, evsel) {
212 		evsel__free_stat_priv(evsel);
213 		evsel__free_counts(evsel);
214 		evsel__free_prev_raw_counts(evsel);
215 	}
216 }
217 
218 void evlist__reset_stats(struct evlist *evlist)
219 {
220 	struct evsel *evsel;
221 
222 	evlist__for_each_entry(evlist, evsel) {
223 		evsel__reset_stat_priv(evsel);
224 		evsel__reset_counts(evsel);
225 	}
226 }
227 
228 void evlist__reset_prev_raw_counts(struct evlist *evlist)
229 {
230 	struct evsel *evsel;
231 
232 	evlist__for_each_entry(evlist, evsel)
233 		evsel__reset_prev_raw_counts(evsel);
234 }
235 
236 static void evsel__copy_prev_raw_counts(struct evsel *evsel)
237 {
238 	int ncpus = evsel__nr_cpus(evsel);
239 	int nthreads = perf_thread_map__nr(evsel->core.threads);
240 
241 	for (int thread = 0; thread < nthreads; thread++) {
242 		for (int cpu = 0; cpu < ncpus; cpu++) {
243 			*perf_counts(evsel->counts, cpu, thread) =
244 				*perf_counts(evsel->prev_raw_counts, cpu,
245 					     thread);
246 		}
247 	}
248 
249 	evsel->counts->aggr = evsel->prev_raw_counts->aggr;
250 }
251 
252 void evlist__copy_prev_raw_counts(struct evlist *evlist)
253 {
254 	struct evsel *evsel;
255 
256 	evlist__for_each_entry(evlist, evsel)
257 		evsel__copy_prev_raw_counts(evsel);
258 }
259 
260 void evlist__save_aggr_prev_raw_counts(struct evlist *evlist)
261 {
262 	struct evsel *evsel;
263 
264 	/*
265 	 * To collect the overall statistics for interval mode,
266 	 * we copy the counts from evsel->prev_raw_counts to
267 	 * evsel->counts. The perf_stat_process_counter creates
268 	 * aggr values from per cpu values, but the per cpu values
269 	 * are 0 for AGGR_GLOBAL. So we use a trick that saves the
270 	 * previous aggr value to the first member of perf_counts,
271 	 * then aggr calculation in process_counter_values can work
272 	 * correctly.
273 	 */
274 	evlist__for_each_entry(evlist, evsel) {
275 		*perf_counts(evsel->prev_raw_counts, 0, 0) =
276 			evsel->prev_raw_counts->aggr;
277 	}
278 }
279 
280 static size_t pkg_id_hash(const void *__key, void *ctx __maybe_unused)
281 {
282 	uint64_t *key = (uint64_t *) __key;
283 
284 	return *key & 0xffffffff;
285 }
286 
287 static bool pkg_id_equal(const void *__key1, const void *__key2,
288 			 void *ctx __maybe_unused)
289 {
290 	uint64_t *key1 = (uint64_t *) __key1;
291 	uint64_t *key2 = (uint64_t *) __key2;
292 
293 	return *key1 == *key2;
294 }
295 
296 static int check_per_pkg(struct evsel *counter, struct perf_counts_values *vals,
297 			 int cpu_map_idx, bool *skip)
298 {
299 	struct hashmap *mask = counter->per_pkg_mask;
300 	struct perf_cpu_map *cpus = evsel__cpus(counter);
301 	struct perf_cpu cpu = perf_cpu_map__cpu(cpus, cpu_map_idx);
302 	int s, d, ret = 0;
303 	uint64_t *key;
304 
305 	*skip = false;
306 
307 	if (!counter->per_pkg)
308 		return 0;
309 
310 	if (perf_cpu_map__empty(cpus))
311 		return 0;
312 
313 	if (!mask) {
314 		mask = hashmap__new(pkg_id_hash, pkg_id_equal, NULL);
315 		if (IS_ERR(mask))
316 			return -ENOMEM;
317 
318 		counter->per_pkg_mask = mask;
319 	}
320 
321 	/*
322 	 * we do not consider an event that has not run as a good
323 	 * instance to mark a package as used (skip=1). Otherwise
324 	 * we may run into a situation where the first CPU in a package
325 	 * is not running anything, yet the second is, and this function
326 	 * would mark the package as used after the first CPU and would
327 	 * not read the values from the second CPU.
328 	 */
329 	if (!(vals->run && vals->ena))
330 		return 0;
331 
332 	s = cpu__get_socket_id(cpu);
333 	if (s < 0)
334 		return -1;
335 
336 	/*
337 	 * On multi-die system, die_id > 0. On no-die system, die_id = 0.
338 	 * We use hashmap(socket, die) to check the used socket+die pair.
339 	 */
340 	d = cpu__get_die_id(cpu);
341 	if (d < 0)
342 		return -1;
343 
344 	key = malloc(sizeof(*key));
345 	if (!key)
346 		return -ENOMEM;
347 
348 	*key = (uint64_t)d << 32 | s;
349 	if (hashmap__find(mask, (void *)key, NULL)) {
350 		*skip = true;
351 		free(key);
352 	} else
353 		ret = hashmap__add(mask, (void *)key, (void *)1);
354 
355 	return ret;
356 }
357 
358 static int
359 process_counter_values(struct perf_stat_config *config, struct evsel *evsel,
360 		       int cpu_map_idx, int thread,
361 		       struct perf_counts_values *count)
362 {
363 	struct perf_counts_values *aggr = &evsel->counts->aggr;
364 	static struct perf_counts_values zero;
365 	bool skip = false;
366 
367 	if (check_per_pkg(evsel, count, cpu_map_idx, &skip)) {
368 		pr_err("failed to read per-pkg counter\n");
369 		return -1;
370 	}
371 
372 	if (skip)
373 		count = &zero;
374 
375 	switch (config->aggr_mode) {
376 	case AGGR_THREAD:
377 	case AGGR_CORE:
378 	case AGGR_DIE:
379 	case AGGR_SOCKET:
380 	case AGGR_NODE:
381 	case AGGR_NONE:
382 		if (!evsel->snapshot)
383 			evsel__compute_deltas(evsel, cpu_map_idx, thread, count);
384 		perf_counts_values__scale(count, config->scale, NULL);
385 		if ((config->aggr_mode == AGGR_NONE) && (!evsel->percore)) {
386 			perf_stat__update_shadow_stats(evsel, count->val,
387 						       cpu_map_idx, &rt_stat);
388 		}
389 
390 		if (config->aggr_mode == AGGR_THREAD) {
391 			if (config->stats)
392 				perf_stat__update_shadow_stats(evsel,
393 					count->val, 0, &config->stats[thread]);
394 			else
395 				perf_stat__update_shadow_stats(evsel,
396 					count->val, 0, &rt_stat);
397 		}
398 		break;
399 	case AGGR_GLOBAL:
400 		aggr->val += count->val;
401 		aggr->ena += count->ena;
402 		aggr->run += count->run;
403 	case AGGR_UNSET:
404 	default:
405 		break;
406 	}
407 
408 	return 0;
409 }
410 
411 static int process_counter_maps(struct perf_stat_config *config,
412 				struct evsel *counter)
413 {
414 	int nthreads = perf_thread_map__nr(counter->core.threads);
415 	int ncpus = evsel__nr_cpus(counter);
416 	int idx, thread;
417 
418 	if (counter->core.system_wide)
419 		nthreads = 1;
420 
421 	for (thread = 0; thread < nthreads; thread++) {
422 		for (idx = 0; idx < ncpus; idx++) {
423 			if (process_counter_values(config, counter, idx, thread,
424 						   perf_counts(counter->counts, idx, thread)))
425 				return -1;
426 		}
427 	}
428 
429 	return 0;
430 }
431 
432 int perf_stat_process_counter(struct perf_stat_config *config,
433 			      struct evsel *counter)
434 {
435 	struct perf_counts_values *aggr = &counter->counts->aggr;
436 	struct perf_stat_evsel *ps = counter->stats;
437 	u64 *count = counter->counts->aggr.values;
438 	int i, ret;
439 
440 	aggr->val = aggr->ena = aggr->run = 0;
441 
442 	if (counter->per_pkg)
443 		evsel__zero_per_pkg(counter);
444 
445 	ret = process_counter_maps(config, counter);
446 	if (ret)
447 		return ret;
448 
449 	if (config->aggr_mode != AGGR_GLOBAL)
450 		return 0;
451 
452 	if (!counter->snapshot)
453 		evsel__compute_deltas(counter, -1, -1, aggr);
454 	perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled);
455 
456 	for (i = 0; i < 3; i++)
457 		update_stats(&ps->res_stats[i], count[i]);
458 
459 	if (verbose > 0) {
460 		fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
461 			evsel__name(counter), count[0], count[1], count[2]);
462 	}
463 
464 	/*
465 	 * Save the full runtime - to allow normalization during printout:
466 	 */
467 	perf_stat__update_shadow_stats(counter, *count, 0, &rt_stat);
468 
469 	return 0;
470 }
471 
472 int perf_event__process_stat_event(struct perf_session *session,
473 				   union perf_event *event)
474 {
475 	struct perf_counts_values count;
476 	struct perf_record_stat *st = &event->stat;
477 	struct evsel *counter;
478 
479 	count.val = st->val;
480 	count.ena = st->ena;
481 	count.run = st->run;
482 
483 	counter = evlist__id2evsel(session->evlist, st->id);
484 	if (!counter) {
485 		pr_err("Failed to resolve counter for stat event.\n");
486 		return -EINVAL;
487 	}
488 
489 	*perf_counts(counter->counts, st->cpu, st->thread) = count;
490 	counter->supported = true;
491 	return 0;
492 }
493 
494 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
495 {
496 	struct perf_record_stat *st = (struct perf_record_stat *)event;
497 	size_t ret;
498 
499 	ret  = fprintf(fp, "\n... id %" PRI_lu64 ", cpu %d, thread %d\n",
500 		       st->id, st->cpu, st->thread);
501 	ret += fprintf(fp, "... value %" PRI_lu64 ", enabled %" PRI_lu64 ", running %" PRI_lu64 "\n",
502 		       st->val, st->ena, st->run);
503 
504 	return ret;
505 }
506 
507 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
508 {
509 	struct perf_record_stat_round *rd = (struct perf_record_stat_round *)event;
510 	size_t ret;
511 
512 	ret = fprintf(fp, "\n... time %" PRI_lu64 ", type %s\n", rd->time,
513 		      rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
514 
515 	return ret;
516 }
517 
518 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
519 {
520 	struct perf_stat_config sc;
521 	size_t ret;
522 
523 	perf_event__read_stat_config(&sc, &event->stat_config);
524 
525 	ret  = fprintf(fp, "\n");
526 	ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
527 	ret += fprintf(fp, "... scale     %d\n", sc.scale);
528 	ret += fprintf(fp, "... interval  %u\n", sc.interval);
529 
530 	return ret;
531 }
532 
533 int create_perf_stat_counter(struct evsel *evsel,
534 			     struct perf_stat_config *config,
535 			     struct target *target,
536 			     int cpu_map_idx)
537 {
538 	struct perf_event_attr *attr = &evsel->core.attr;
539 	struct evsel *leader = evsel__leader(evsel);
540 
541 	attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
542 			    PERF_FORMAT_TOTAL_TIME_RUNNING;
543 
544 	/*
545 	 * The event is part of non trivial group, let's enable
546 	 * the group read (for leader) and ID retrieval for all
547 	 * members.
548 	 */
549 	if (leader->core.nr_members > 1)
550 		attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
551 
552 	attr->inherit = !config->no_inherit && list_empty(&evsel->bpf_counter_list);
553 
554 	/*
555 	 * Some events get initialized with sample_(period/type) set,
556 	 * like tracepoints. Clear it up for counting.
557 	 */
558 	attr->sample_period = 0;
559 
560 	if (config->identifier)
561 		attr->sample_type = PERF_SAMPLE_IDENTIFIER;
562 
563 	if (config->all_user) {
564 		attr->exclude_kernel = 1;
565 		attr->exclude_user   = 0;
566 	}
567 
568 	if (config->all_kernel) {
569 		attr->exclude_kernel = 0;
570 		attr->exclude_user   = 1;
571 	}
572 
573 	/*
574 	 * Disabling all counters initially, they will be enabled
575 	 * either manually by us or by kernel via enable_on_exec
576 	 * set later.
577 	 */
578 	if (evsel__is_group_leader(evsel)) {
579 		attr->disabled = 1;
580 
581 		/*
582 		 * In case of initial_delay we enable tracee
583 		 * events manually.
584 		 */
585 		if (target__none(target) && !config->initial_delay)
586 			attr->enable_on_exec = 1;
587 	}
588 
589 	if (target__has_cpu(target) && !target__has_per_thread(target))
590 		return evsel__open_per_cpu(evsel, evsel__cpus(evsel), cpu_map_idx);
591 
592 	return evsel__open_per_thread(evsel, evsel->core.threads);
593 }
594