xref: /openbmc/linux/tools/perf/util/stat.c (revision f125e2d4)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <errno.h>
3 #include <inttypes.h>
4 #include <math.h>
5 #include <string.h>
6 #include "counts.h"
7 #include "cpumap.h"
8 #include "debug.h"
9 #include "header.h"
10 #include "stat.h"
11 #include "session.h"
12 #include "target.h"
13 #include "evlist.h"
14 #include "evsel.h"
15 #include "thread_map.h"
16 #include <linux/zalloc.h>
17 
18 void update_stats(struct stats *stats, u64 val)
19 {
20 	double delta;
21 
22 	stats->n++;
23 	delta = val - stats->mean;
24 	stats->mean += delta / stats->n;
25 	stats->M2 += delta*(val - stats->mean);
26 
27 	if (val > stats->max)
28 		stats->max = val;
29 
30 	if (val < stats->min)
31 		stats->min = val;
32 }
33 
34 double avg_stats(struct stats *stats)
35 {
36 	return stats->mean;
37 }
38 
39 /*
40  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
41  *
42  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
43  * s^2 = -------------------------------
44  *                  n - 1
45  *
46  * http://en.wikipedia.org/wiki/Stddev
47  *
48  * The std dev of the mean is related to the std dev by:
49  *
50  *             s
51  * s_mean = -------
52  *          sqrt(n)
53  *
54  */
55 double stddev_stats(struct stats *stats)
56 {
57 	double variance, variance_mean;
58 
59 	if (stats->n < 2)
60 		return 0.0;
61 
62 	variance = stats->M2 / (stats->n - 1);
63 	variance_mean = variance / stats->n;
64 
65 	return sqrt(variance_mean);
66 }
67 
68 double rel_stddev_stats(double stddev, double avg)
69 {
70 	double pct = 0.0;
71 
72 	if (avg)
73 		pct = 100.0 * stddev/avg;
74 
75 	return pct;
76 }
77 
78 bool __perf_evsel_stat__is(struct evsel *evsel,
79 			   enum perf_stat_evsel_id id)
80 {
81 	struct perf_stat_evsel *ps = evsel->stats;
82 
83 	return ps->id == id;
84 }
85 
86 #define ID(id, name) [PERF_STAT_EVSEL_ID__##id] = #name
87 static const char *id_str[PERF_STAT_EVSEL_ID__MAX] = {
88 	ID(NONE,		x),
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(SMI_NUM, msr/smi/),
99 	ID(APERF, msr/aperf/),
100 };
101 #undef ID
102 
103 static void perf_stat_evsel_id_init(struct evsel *evsel)
104 {
105 	struct perf_stat_evsel *ps = evsel->stats;
106 	int i;
107 
108 	/* ps->id is 0 hence PERF_STAT_EVSEL_ID__NONE by default */
109 
110 	for (i = 0; i < PERF_STAT_EVSEL_ID__MAX; i++) {
111 		if (!strcmp(perf_evsel__name(evsel), id_str[i])) {
112 			ps->id = i;
113 			break;
114 		}
115 	}
116 }
117 
118 static void perf_evsel__reset_stat_priv(struct evsel *evsel)
119 {
120 	int i;
121 	struct perf_stat_evsel *ps = evsel->stats;
122 
123 	for (i = 0; i < 3; i++)
124 		init_stats(&ps->res_stats[i]);
125 
126 	perf_stat_evsel_id_init(evsel);
127 }
128 
129 static int perf_evsel__alloc_stat_priv(struct evsel *evsel)
130 {
131 	evsel->stats = zalloc(sizeof(struct perf_stat_evsel));
132 	if (evsel->stats == NULL)
133 		return -ENOMEM;
134 	perf_evsel__reset_stat_priv(evsel);
135 	return 0;
136 }
137 
138 static void perf_evsel__free_stat_priv(struct evsel *evsel)
139 {
140 	struct perf_stat_evsel *ps = evsel->stats;
141 
142 	if (ps)
143 		zfree(&ps->group_data);
144 	zfree(&evsel->stats);
145 }
146 
147 static int perf_evsel__alloc_prev_raw_counts(struct evsel *evsel,
148 					     int ncpus, int nthreads)
149 {
150 	struct perf_counts *counts;
151 
152 	counts = perf_counts__new(ncpus, nthreads);
153 	if (counts)
154 		evsel->prev_raw_counts = counts;
155 
156 	return counts ? 0 : -ENOMEM;
157 }
158 
159 static void perf_evsel__free_prev_raw_counts(struct evsel *evsel)
160 {
161 	perf_counts__delete(evsel->prev_raw_counts);
162 	evsel->prev_raw_counts = NULL;
163 }
164 
165 static void perf_evsel__reset_prev_raw_counts(struct evsel *evsel)
166 {
167 	if (evsel->prev_raw_counts) {
168 		evsel->prev_raw_counts->aggr.val = 0;
169 		evsel->prev_raw_counts->aggr.ena = 0;
170 		evsel->prev_raw_counts->aggr.run = 0;
171        }
172 }
173 
174 static int perf_evsel__alloc_stats(struct evsel *evsel, bool alloc_raw)
175 {
176 	int ncpus = perf_evsel__nr_cpus(evsel);
177 	int nthreads = perf_thread_map__nr(evsel->core.threads);
178 
179 	if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
180 	    perf_evsel__alloc_counts(evsel, ncpus, nthreads) < 0 ||
181 	    (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel, ncpus, nthreads) < 0))
182 		return -ENOMEM;
183 
184 	return 0;
185 }
186 
187 int perf_evlist__alloc_stats(struct evlist *evlist, bool alloc_raw)
188 {
189 	struct evsel *evsel;
190 
191 	evlist__for_each_entry(evlist, evsel) {
192 		if (perf_evsel__alloc_stats(evsel, alloc_raw))
193 			goto out_free;
194 	}
195 
196 	return 0;
197 
198 out_free:
199 	perf_evlist__free_stats(evlist);
200 	return -1;
201 }
202 
203 void perf_evlist__free_stats(struct evlist *evlist)
204 {
205 	struct evsel *evsel;
206 
207 	evlist__for_each_entry(evlist, evsel) {
208 		perf_evsel__free_stat_priv(evsel);
209 		perf_evsel__free_counts(evsel);
210 		perf_evsel__free_prev_raw_counts(evsel);
211 	}
212 }
213 
214 void perf_evlist__reset_stats(struct evlist *evlist)
215 {
216 	struct evsel *evsel;
217 
218 	evlist__for_each_entry(evlist, evsel) {
219 		perf_evsel__reset_stat_priv(evsel);
220 		perf_evsel__reset_counts(evsel);
221 	}
222 }
223 
224 void perf_evlist__reset_prev_raw_counts(struct evlist *evlist)
225 {
226 	struct evsel *evsel;
227 
228 	evlist__for_each_entry(evlist, evsel)
229 		perf_evsel__reset_prev_raw_counts(evsel);
230 }
231 
232 static void zero_per_pkg(struct evsel *counter)
233 {
234 	if (counter->per_pkg_mask)
235 		memset(counter->per_pkg_mask, 0, cpu__max_cpu());
236 }
237 
238 static int check_per_pkg(struct evsel *counter,
239 			 struct perf_counts_values *vals, int cpu, bool *skip)
240 {
241 	unsigned long *mask = counter->per_pkg_mask;
242 	struct perf_cpu_map *cpus = evsel__cpus(counter);
243 	int s;
244 
245 	*skip = false;
246 
247 	if (!counter->per_pkg)
248 		return 0;
249 
250 	if (perf_cpu_map__empty(cpus))
251 		return 0;
252 
253 	if (!mask) {
254 		mask = zalloc(cpu__max_cpu());
255 		if (!mask)
256 			return -ENOMEM;
257 
258 		counter->per_pkg_mask = mask;
259 	}
260 
261 	/*
262 	 * we do not consider an event that has not run as a good
263 	 * instance to mark a package as used (skip=1). Otherwise
264 	 * we may run into a situation where the first CPU in a package
265 	 * is not running anything, yet the second is, and this function
266 	 * would mark the package as used after the first CPU and would
267 	 * not read the values from the second CPU.
268 	 */
269 	if (!(vals->run && vals->ena))
270 		return 0;
271 
272 	s = cpu_map__get_socket(cpus, cpu, NULL);
273 	if (s < 0)
274 		return -1;
275 
276 	*skip = test_and_set_bit(s, mask) == 1;
277 	return 0;
278 }
279 
280 static int
281 process_counter_values(struct perf_stat_config *config, struct evsel *evsel,
282 		       int cpu, int thread,
283 		       struct perf_counts_values *count)
284 {
285 	struct perf_counts_values *aggr = &evsel->counts->aggr;
286 	static struct perf_counts_values zero;
287 	bool skip = false;
288 
289 	if (check_per_pkg(evsel, count, cpu, &skip)) {
290 		pr_err("failed to read per-pkg counter\n");
291 		return -1;
292 	}
293 
294 	if (skip)
295 		count = &zero;
296 
297 	switch (config->aggr_mode) {
298 	case AGGR_THREAD:
299 	case AGGR_CORE:
300 	case AGGR_DIE:
301 	case AGGR_SOCKET:
302 	case AGGR_NODE:
303 	case AGGR_NONE:
304 		if (!evsel->snapshot)
305 			perf_evsel__compute_deltas(evsel, cpu, thread, count);
306 		perf_counts_values__scale(count, config->scale, NULL);
307 		if ((config->aggr_mode == AGGR_NONE) && (!evsel->percore)) {
308 			perf_stat__update_shadow_stats(evsel, count->val,
309 						       cpu, &rt_stat);
310 		}
311 
312 		if (config->aggr_mode == AGGR_THREAD) {
313 			if (config->stats)
314 				perf_stat__update_shadow_stats(evsel,
315 					count->val, 0, &config->stats[thread]);
316 			else
317 				perf_stat__update_shadow_stats(evsel,
318 					count->val, 0, &rt_stat);
319 		}
320 		break;
321 	case AGGR_GLOBAL:
322 		aggr->val += count->val;
323 		aggr->ena += count->ena;
324 		aggr->run += count->run;
325 	case AGGR_UNSET:
326 	default:
327 		break;
328 	}
329 
330 	return 0;
331 }
332 
333 static int process_counter_maps(struct perf_stat_config *config,
334 				struct evsel *counter)
335 {
336 	int nthreads = perf_thread_map__nr(counter->core.threads);
337 	int ncpus = perf_evsel__nr_cpus(counter);
338 	int cpu, thread;
339 
340 	if (counter->core.system_wide)
341 		nthreads = 1;
342 
343 	for (thread = 0; thread < nthreads; thread++) {
344 		for (cpu = 0; cpu < ncpus; cpu++) {
345 			if (process_counter_values(config, counter, cpu, thread,
346 						   perf_counts(counter->counts, cpu, thread)))
347 				return -1;
348 		}
349 	}
350 
351 	return 0;
352 }
353 
354 int perf_stat_process_counter(struct perf_stat_config *config,
355 			      struct evsel *counter)
356 {
357 	struct perf_counts_values *aggr = &counter->counts->aggr;
358 	struct perf_stat_evsel *ps = counter->stats;
359 	u64 *count = counter->counts->aggr.values;
360 	int i, ret;
361 
362 	aggr->val = aggr->ena = aggr->run = 0;
363 
364 	/*
365 	 * We calculate counter's data every interval,
366 	 * and the display code shows ps->res_stats
367 	 * avg value. We need to zero the stats for
368 	 * interval mode, otherwise overall avg running
369 	 * averages will be shown for each interval.
370 	 */
371 	if (config->interval)
372 		init_stats(ps->res_stats);
373 
374 	if (counter->per_pkg)
375 		zero_per_pkg(counter);
376 
377 	ret = process_counter_maps(config, counter);
378 	if (ret)
379 		return ret;
380 
381 	if (config->aggr_mode != AGGR_GLOBAL)
382 		return 0;
383 
384 	if (!counter->snapshot)
385 		perf_evsel__compute_deltas(counter, -1, -1, aggr);
386 	perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled);
387 
388 	for (i = 0; i < 3; i++)
389 		update_stats(&ps->res_stats[i], count[i]);
390 
391 	if (verbose > 0) {
392 		fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
393 			perf_evsel__name(counter), count[0], count[1], count[2]);
394 	}
395 
396 	/*
397 	 * Save the full runtime - to allow normalization during printout:
398 	 */
399 	perf_stat__update_shadow_stats(counter, *count, 0, &rt_stat);
400 
401 	return 0;
402 }
403 
404 int perf_event__process_stat_event(struct perf_session *session,
405 				   union perf_event *event)
406 {
407 	struct perf_counts_values count;
408 	struct perf_record_stat *st = &event->stat;
409 	struct evsel *counter;
410 
411 	count.val = st->val;
412 	count.ena = st->ena;
413 	count.run = st->run;
414 
415 	counter = perf_evlist__id2evsel(session->evlist, st->id);
416 	if (!counter) {
417 		pr_err("Failed to resolve counter for stat event.\n");
418 		return -EINVAL;
419 	}
420 
421 	*perf_counts(counter->counts, st->cpu, st->thread) = count;
422 	counter->supported = true;
423 	return 0;
424 }
425 
426 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
427 {
428 	struct perf_record_stat *st = (struct perf_record_stat *)event;
429 	size_t ret;
430 
431 	ret  = fprintf(fp, "\n... id %" PRI_lu64 ", cpu %d, thread %d\n",
432 		       st->id, st->cpu, st->thread);
433 	ret += fprintf(fp, "... value %" PRI_lu64 ", enabled %" PRI_lu64 ", running %" PRI_lu64 "\n",
434 		       st->val, st->ena, st->run);
435 
436 	return ret;
437 }
438 
439 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
440 {
441 	struct perf_record_stat_round *rd = (struct perf_record_stat_round *)event;
442 	size_t ret;
443 
444 	ret = fprintf(fp, "\n... time %" PRI_lu64 ", type %s\n", rd->time,
445 		      rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
446 
447 	return ret;
448 }
449 
450 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
451 {
452 	struct perf_stat_config sc;
453 	size_t ret;
454 
455 	perf_event__read_stat_config(&sc, &event->stat_config);
456 
457 	ret  = fprintf(fp, "\n");
458 	ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
459 	ret += fprintf(fp, "... scale     %d\n", sc.scale);
460 	ret += fprintf(fp, "... interval  %u\n", sc.interval);
461 
462 	return ret;
463 }
464 
465 int create_perf_stat_counter(struct evsel *evsel,
466 			     struct perf_stat_config *config,
467 			     struct target *target,
468 			     int cpu)
469 {
470 	struct perf_event_attr *attr = &evsel->core.attr;
471 	struct evsel *leader = evsel->leader;
472 
473 	attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
474 			    PERF_FORMAT_TOTAL_TIME_RUNNING;
475 
476 	/*
477 	 * The event is part of non trivial group, let's enable
478 	 * the group read (for leader) and ID retrieval for all
479 	 * members.
480 	 */
481 	if (leader->core.nr_members > 1)
482 		attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
483 
484 	attr->inherit = !config->no_inherit;
485 
486 	/*
487 	 * Some events get initialized with sample_(period/type) set,
488 	 * like tracepoints. Clear it up for counting.
489 	 */
490 	attr->sample_period = 0;
491 
492 	if (config->identifier)
493 		attr->sample_type = PERF_SAMPLE_IDENTIFIER;
494 
495 	if (config->all_user) {
496 		attr->exclude_kernel = 1;
497 		attr->exclude_user   = 0;
498 	}
499 
500 	if (config->all_kernel) {
501 		attr->exclude_kernel = 0;
502 		attr->exclude_user   = 1;
503 	}
504 
505 	/*
506 	 * Disabling all counters initially, they will be enabled
507 	 * either manually by us or by kernel via enable_on_exec
508 	 * set later.
509 	 */
510 	if (perf_evsel__is_group_leader(evsel)) {
511 		attr->disabled = 1;
512 
513 		/*
514 		 * In case of initial_delay we enable tracee
515 		 * events manually.
516 		 */
517 		if (target__none(target) && !config->initial_delay)
518 			attr->enable_on_exec = 1;
519 	}
520 
521 	if (target__has_cpu(target) && !target__has_per_thread(target))
522 		return perf_evsel__open_per_cpu(evsel, evsel__cpus(evsel), cpu);
523 
524 	return perf_evsel__open_per_thread(evsel, evsel->core.threads);
525 }
526