xref: /openbmc/linux/tools/perf/util/stat.c (revision cd238eff)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <errno.h>
3 #include <inttypes.h>
4 #include <math.h>
5 #include "stat.h"
6 #include "evlist.h"
7 #include "evsel.h"
8 #include "thread_map.h"
9 
10 void update_stats(struct stats *stats, u64 val)
11 {
12 	double delta;
13 
14 	stats->n++;
15 	delta = val - stats->mean;
16 	stats->mean += delta / stats->n;
17 	stats->M2 += delta*(val - stats->mean);
18 
19 	if (val > stats->max)
20 		stats->max = val;
21 
22 	if (val < stats->min)
23 		stats->min = val;
24 }
25 
26 double avg_stats(struct stats *stats)
27 {
28 	return stats->mean;
29 }
30 
31 /*
32  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
33  *
34  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
35  * s^2 = -------------------------------
36  *                  n - 1
37  *
38  * http://en.wikipedia.org/wiki/Stddev
39  *
40  * The std dev of the mean is related to the std dev by:
41  *
42  *             s
43  * s_mean = -------
44  *          sqrt(n)
45  *
46  */
47 double stddev_stats(struct stats *stats)
48 {
49 	double variance, variance_mean;
50 
51 	if (stats->n < 2)
52 		return 0.0;
53 
54 	variance = stats->M2 / (stats->n - 1);
55 	variance_mean = variance / stats->n;
56 
57 	return sqrt(variance_mean);
58 }
59 
60 double rel_stddev_stats(double stddev, double avg)
61 {
62 	double pct = 0.0;
63 
64 	if (avg)
65 		pct = 100.0 * stddev/avg;
66 
67 	return pct;
68 }
69 
70 bool __perf_evsel_stat__is(struct perf_evsel *evsel,
71 			   enum perf_stat_evsel_id id)
72 {
73 	struct perf_stat_evsel *ps = evsel->stats;
74 
75 	return ps->id == id;
76 }
77 
78 #define ID(id, name) [PERF_STAT_EVSEL_ID__##id] = #name
79 static const char *id_str[PERF_STAT_EVSEL_ID__MAX] = {
80 	ID(NONE,		x),
81 	ID(CYCLES_IN_TX,	cpu/cycles-t/),
82 	ID(TRANSACTION_START,	cpu/tx-start/),
83 	ID(ELISION_START,	cpu/el-start/),
84 	ID(CYCLES_IN_TX_CP,	cpu/cycles-ct/),
85 	ID(TOPDOWN_TOTAL_SLOTS, topdown-total-slots),
86 	ID(TOPDOWN_SLOTS_ISSUED, topdown-slots-issued),
87 	ID(TOPDOWN_SLOTS_RETIRED, topdown-slots-retired),
88 	ID(TOPDOWN_FETCH_BUBBLES, topdown-fetch-bubbles),
89 	ID(TOPDOWN_RECOVERY_BUBBLES, topdown-recovery-bubbles),
90 	ID(SMI_NUM, msr/smi/),
91 	ID(APERF, msr/aperf/),
92 };
93 #undef ID
94 
95 static void perf_stat_evsel_id_init(struct perf_evsel *evsel)
96 {
97 	struct perf_stat_evsel *ps = evsel->stats;
98 	int i;
99 
100 	/* ps->id is 0 hence PERF_STAT_EVSEL_ID__NONE by default */
101 
102 	for (i = 0; i < PERF_STAT_EVSEL_ID__MAX; i++) {
103 		if (!strcmp(perf_evsel__name(evsel), id_str[i])) {
104 			ps->id = i;
105 			break;
106 		}
107 	}
108 }
109 
110 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
111 {
112 	int i;
113 	struct perf_stat_evsel *ps = evsel->stats;
114 
115 	for (i = 0; i < 3; i++)
116 		init_stats(&ps->res_stats[i]);
117 
118 	perf_stat_evsel_id_init(evsel);
119 }
120 
121 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
122 {
123 	evsel->stats = zalloc(sizeof(struct perf_stat_evsel));
124 	if (evsel->stats == NULL)
125 		return -ENOMEM;
126 	perf_evsel__reset_stat_priv(evsel);
127 	return 0;
128 }
129 
130 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
131 {
132 	struct perf_stat_evsel *ps = evsel->stats;
133 
134 	if (ps)
135 		free(ps->group_data);
136 	zfree(&evsel->stats);
137 }
138 
139 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel,
140 					     int ncpus, int nthreads)
141 {
142 	struct perf_counts *counts;
143 
144 	counts = perf_counts__new(ncpus, nthreads);
145 	if (counts)
146 		evsel->prev_raw_counts = counts;
147 
148 	return counts ? 0 : -ENOMEM;
149 }
150 
151 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
152 {
153 	perf_counts__delete(evsel->prev_raw_counts);
154 	evsel->prev_raw_counts = NULL;
155 }
156 
157 static int perf_evsel__alloc_stats(struct perf_evsel *evsel, bool alloc_raw)
158 {
159 	int ncpus = perf_evsel__nr_cpus(evsel);
160 	int nthreads = thread_map__nr(evsel->threads);
161 
162 	if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
163 	    perf_evsel__alloc_counts(evsel, ncpus, nthreads) < 0 ||
164 	    (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel, ncpus, nthreads) < 0))
165 		return -ENOMEM;
166 
167 	return 0;
168 }
169 
170 int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
171 {
172 	struct perf_evsel *evsel;
173 
174 	evlist__for_each_entry(evlist, evsel) {
175 		if (perf_evsel__alloc_stats(evsel, alloc_raw))
176 			goto out_free;
177 	}
178 
179 	return 0;
180 
181 out_free:
182 	perf_evlist__free_stats(evlist);
183 	return -1;
184 }
185 
186 void perf_evlist__free_stats(struct perf_evlist *evlist)
187 {
188 	struct perf_evsel *evsel;
189 
190 	evlist__for_each_entry(evlist, evsel) {
191 		perf_evsel__free_stat_priv(evsel);
192 		perf_evsel__free_counts(evsel);
193 		perf_evsel__free_prev_raw_counts(evsel);
194 	}
195 }
196 
197 void perf_evlist__reset_stats(struct perf_evlist *evlist)
198 {
199 	struct perf_evsel *evsel;
200 
201 	evlist__for_each_entry(evlist, evsel) {
202 		perf_evsel__reset_stat_priv(evsel);
203 		perf_evsel__reset_counts(evsel);
204 	}
205 }
206 
207 static void zero_per_pkg(struct perf_evsel *counter)
208 {
209 	if (counter->per_pkg_mask)
210 		memset(counter->per_pkg_mask, 0, MAX_NR_CPUS);
211 }
212 
213 static int check_per_pkg(struct perf_evsel *counter,
214 			 struct perf_counts_values *vals, int cpu, bool *skip)
215 {
216 	unsigned long *mask = counter->per_pkg_mask;
217 	struct cpu_map *cpus = perf_evsel__cpus(counter);
218 	int s;
219 
220 	*skip = false;
221 
222 	if (!counter->per_pkg)
223 		return 0;
224 
225 	if (cpu_map__empty(cpus))
226 		return 0;
227 
228 	if (!mask) {
229 		mask = zalloc(MAX_NR_CPUS);
230 		if (!mask)
231 			return -ENOMEM;
232 
233 		counter->per_pkg_mask = mask;
234 	}
235 
236 	/*
237 	 * we do not consider an event that has not run as a good
238 	 * instance to mark a package as used (skip=1). Otherwise
239 	 * we may run into a situation where the first CPU in a package
240 	 * is not running anything, yet the second is, and this function
241 	 * would mark the package as used after the first CPU and would
242 	 * not read the values from the second CPU.
243 	 */
244 	if (!(vals->run && vals->ena))
245 		return 0;
246 
247 	s = cpu_map__get_socket(cpus, cpu, NULL);
248 	if (s < 0)
249 		return -1;
250 
251 	*skip = test_and_set_bit(s, mask) == 1;
252 	return 0;
253 }
254 
255 static int
256 process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel,
257 		       int cpu, int thread,
258 		       struct perf_counts_values *count)
259 {
260 	struct perf_counts_values *aggr = &evsel->counts->aggr;
261 	static struct perf_counts_values zero;
262 	bool skip = false;
263 
264 	if (check_per_pkg(evsel, count, cpu, &skip)) {
265 		pr_err("failed to read per-pkg counter\n");
266 		return -1;
267 	}
268 
269 	if (skip)
270 		count = &zero;
271 
272 	switch (config->aggr_mode) {
273 	case AGGR_THREAD:
274 	case AGGR_CORE:
275 	case AGGR_SOCKET:
276 	case AGGR_NONE:
277 		if (!evsel->snapshot)
278 			perf_evsel__compute_deltas(evsel, cpu, thread, count);
279 		perf_counts_values__scale(count, config->scale, NULL);
280 		if ((config->aggr_mode == AGGR_NONE) && (!evsel->percore)) {
281 			perf_stat__update_shadow_stats(evsel, count->val,
282 						       cpu, &rt_stat);
283 		}
284 
285 		if (config->aggr_mode == AGGR_THREAD) {
286 			if (config->stats)
287 				perf_stat__update_shadow_stats(evsel,
288 					count->val, 0, &config->stats[thread]);
289 			else
290 				perf_stat__update_shadow_stats(evsel,
291 					count->val, 0, &rt_stat);
292 		}
293 		break;
294 	case AGGR_GLOBAL:
295 		aggr->val += count->val;
296 		aggr->ena += count->ena;
297 		aggr->run += count->run;
298 	case AGGR_UNSET:
299 	default:
300 		break;
301 	}
302 
303 	return 0;
304 }
305 
306 static int process_counter_maps(struct perf_stat_config *config,
307 				struct perf_evsel *counter)
308 {
309 	int nthreads = thread_map__nr(counter->threads);
310 	int ncpus = perf_evsel__nr_cpus(counter);
311 	int cpu, thread;
312 
313 	if (counter->system_wide)
314 		nthreads = 1;
315 
316 	for (thread = 0; thread < nthreads; thread++) {
317 		for (cpu = 0; cpu < ncpus; cpu++) {
318 			if (process_counter_values(config, counter, cpu, thread,
319 						   perf_counts(counter->counts, cpu, thread)))
320 				return -1;
321 		}
322 	}
323 
324 	return 0;
325 }
326 
327 int perf_stat_process_counter(struct perf_stat_config *config,
328 			      struct perf_evsel *counter)
329 {
330 	struct perf_counts_values *aggr = &counter->counts->aggr;
331 	struct perf_stat_evsel *ps = counter->stats;
332 	u64 *count = counter->counts->aggr.values;
333 	int i, ret;
334 
335 	aggr->val = aggr->ena = aggr->run = 0;
336 
337 	/*
338 	 * We calculate counter's data every interval,
339 	 * and the display code shows ps->res_stats
340 	 * avg value. We need to zero the stats for
341 	 * interval mode, otherwise overall avg running
342 	 * averages will be shown for each interval.
343 	 */
344 	if (config->interval)
345 		init_stats(ps->res_stats);
346 
347 	if (counter->per_pkg)
348 		zero_per_pkg(counter);
349 
350 	ret = process_counter_maps(config, counter);
351 	if (ret)
352 		return ret;
353 
354 	if (config->aggr_mode != AGGR_GLOBAL)
355 		return 0;
356 
357 	if (!counter->snapshot)
358 		perf_evsel__compute_deltas(counter, -1, -1, aggr);
359 	perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled);
360 
361 	for (i = 0; i < 3; i++)
362 		update_stats(&ps->res_stats[i], count[i]);
363 
364 	if (verbose > 0) {
365 		fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
366 			perf_evsel__name(counter), count[0], count[1], count[2]);
367 	}
368 
369 	/*
370 	 * Save the full runtime - to allow normalization during printout:
371 	 */
372 	perf_stat__update_shadow_stats(counter, *count, 0, &rt_stat);
373 
374 	return 0;
375 }
376 
377 int perf_event__process_stat_event(struct perf_session *session,
378 				   union perf_event *event)
379 {
380 	struct perf_counts_values count;
381 	struct stat_event *st = &event->stat;
382 	struct perf_evsel *counter;
383 
384 	count.val = st->val;
385 	count.ena = st->ena;
386 	count.run = st->run;
387 
388 	counter = perf_evlist__id2evsel(session->evlist, st->id);
389 	if (!counter) {
390 		pr_err("Failed to resolve counter for stat event.\n");
391 		return -EINVAL;
392 	}
393 
394 	*perf_counts(counter->counts, st->cpu, st->thread) = count;
395 	counter->supported = true;
396 	return 0;
397 }
398 
399 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
400 {
401 	struct stat_event *st = (struct stat_event *) event;
402 	size_t ret;
403 
404 	ret  = fprintf(fp, "\n... id %" PRIu64 ", cpu %d, thread %d\n",
405 		       st->id, st->cpu, st->thread);
406 	ret += fprintf(fp, "... value %" PRIu64 ", enabled %" PRIu64 ", running %" PRIu64 "\n",
407 		       st->val, st->ena, st->run);
408 
409 	return ret;
410 }
411 
412 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
413 {
414 	struct stat_round_event *rd = (struct stat_round_event *)event;
415 	size_t ret;
416 
417 	ret = fprintf(fp, "\n... time %" PRIu64 ", type %s\n", rd->time,
418 		      rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
419 
420 	return ret;
421 }
422 
423 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
424 {
425 	struct perf_stat_config sc;
426 	size_t ret;
427 
428 	perf_event__read_stat_config(&sc, &event->stat_config);
429 
430 	ret  = fprintf(fp, "\n");
431 	ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
432 	ret += fprintf(fp, "... scale     %d\n", sc.scale);
433 	ret += fprintf(fp, "... interval  %u\n", sc.interval);
434 
435 	return ret;
436 }
437 
438 int create_perf_stat_counter(struct perf_evsel *evsel,
439 			     struct perf_stat_config *config,
440 			     struct target *target)
441 {
442 	struct perf_event_attr *attr = &evsel->attr;
443 	struct perf_evsel *leader = evsel->leader;
444 
445 	attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
446 			    PERF_FORMAT_TOTAL_TIME_RUNNING;
447 
448 	/*
449 	 * The event is part of non trivial group, let's enable
450 	 * the group read (for leader) and ID retrieval for all
451 	 * members.
452 	 */
453 	if (leader->nr_members > 1)
454 		attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
455 
456 	attr->inherit = !config->no_inherit;
457 
458 	/*
459 	 * Some events get initialized with sample_(period/type) set,
460 	 * like tracepoints. Clear it up for counting.
461 	 */
462 	attr->sample_period = 0;
463 
464 	if (config->identifier)
465 		attr->sample_type = PERF_SAMPLE_IDENTIFIER;
466 
467 	/*
468 	 * Disabling all counters initially, they will be enabled
469 	 * either manually by us or by kernel via enable_on_exec
470 	 * set later.
471 	 */
472 	if (perf_evsel__is_group_leader(evsel)) {
473 		attr->disabled = 1;
474 
475 		/*
476 		 * In case of initial_delay we enable tracee
477 		 * events manually.
478 		 */
479 		if (target__none(target) && !config->initial_delay)
480 			attr->enable_on_exec = 1;
481 	}
482 
483 	if (target__has_cpu(target) && !target__has_per_thread(target))
484 		return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
485 
486 	return perf_evsel__open_per_thread(evsel, evsel->threads);
487 }
488 
489 int perf_stat_synthesize_config(struct perf_stat_config *config,
490 				struct perf_tool *tool,
491 				struct perf_evlist *evlist,
492 				perf_event__handler_t process,
493 				bool attrs)
494 {
495 	int err;
496 
497 	if (attrs) {
498 		err = perf_event__synthesize_attrs(tool, evlist, process);
499 		if (err < 0) {
500 			pr_err("Couldn't synthesize attrs.\n");
501 			return err;
502 		}
503 	}
504 
505 	err = perf_event__synthesize_extra_attr(tool, evlist, process,
506 						attrs);
507 
508 	err = perf_event__synthesize_thread_map2(tool, evlist->threads,
509 						 process, NULL);
510 	if (err < 0) {
511 		pr_err("Couldn't synthesize thread map.\n");
512 		return err;
513 	}
514 
515 	err = perf_event__synthesize_cpu_map(tool, evlist->cpus,
516 					     process, NULL);
517 	if (err < 0) {
518 		pr_err("Couldn't synthesize thread map.\n");
519 		return err;
520 	}
521 
522 	err = perf_event__synthesize_stat_config(tool, config, process, NULL);
523 	if (err < 0) {
524 		pr_err("Couldn't synthesize config.\n");
525 		return err;
526 	}
527 
528 	return 0;
529 }
530