Merge remote-tracking branch 'torvalds/master' into perf/core
[sfrench/cifs-2.6.git] / tools / perf / util / 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
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                     (strstr(evsel__name(evsel), id_str[i]) && evsel->pmu_name
122                      && strstr(evsel__name(evsel), evsel->pmu_name))) {
123                         ps->id = i;
124                         break;
125                 }
126         }
127 }
128
129 static void evsel__reset_aggr_stats(struct evsel *evsel)
130 {
131         struct perf_stat_evsel *ps = evsel->stats;
132         struct perf_stat_aggr *aggr = ps->aggr;
133
134         if (aggr)
135                 memset(aggr, 0, sizeof(*aggr) * ps->nr_aggr);
136 }
137
138 static void evsel__reset_stat_priv(struct evsel *evsel)
139 {
140         struct perf_stat_evsel *ps = evsel->stats;
141
142         init_stats(&ps->res_stats);
143         evsel__reset_aggr_stats(evsel);
144 }
145
146 static int evsel__alloc_aggr_stats(struct evsel *evsel, int nr_aggr)
147 {
148         struct perf_stat_evsel *ps = evsel->stats;
149
150         if (ps == NULL)
151                 return 0;
152
153         ps->nr_aggr = nr_aggr;
154         ps->aggr = calloc(nr_aggr, sizeof(*ps->aggr));
155         if (ps->aggr == NULL)
156                 return -ENOMEM;
157
158         return 0;
159 }
160
161 int evlist__alloc_aggr_stats(struct evlist *evlist, int nr_aggr)
162 {
163         struct evsel *evsel;
164
165         evlist__for_each_entry(evlist, evsel) {
166                 if (evsel__alloc_aggr_stats(evsel, nr_aggr) < 0)
167                         return -1;
168         }
169         return 0;
170 }
171
172 static int evsel__alloc_stat_priv(struct evsel *evsel, int nr_aggr)
173 {
174         struct perf_stat_evsel *ps;
175
176         ps = zalloc(sizeof(*ps));
177         if (ps == NULL)
178                 return -ENOMEM;
179
180         evsel->stats = ps;
181
182         if (nr_aggr && evsel__alloc_aggr_stats(evsel, nr_aggr) < 0) {
183                 evsel->stats = NULL;
184                 free(ps);
185                 return -ENOMEM;
186         }
187
188         perf_stat_evsel_id_init(evsel);
189         evsel__reset_stat_priv(evsel);
190         return 0;
191 }
192
193 static void evsel__free_stat_priv(struct evsel *evsel)
194 {
195         struct perf_stat_evsel *ps = evsel->stats;
196
197         if (ps) {
198                 zfree(&ps->aggr);
199                 zfree(&ps->group_data);
200         }
201         zfree(&evsel->stats);
202 }
203
204 static int evsel__alloc_prev_raw_counts(struct evsel *evsel)
205 {
206         int cpu_map_nr = evsel__nr_cpus(evsel);
207         int nthreads = perf_thread_map__nr(evsel->core.threads);
208         struct perf_counts *counts;
209
210         counts = perf_counts__new(cpu_map_nr, nthreads);
211         if (counts)
212                 evsel->prev_raw_counts = counts;
213
214         return counts ? 0 : -ENOMEM;
215 }
216
217 static void evsel__free_prev_raw_counts(struct evsel *evsel)
218 {
219         perf_counts__delete(evsel->prev_raw_counts);
220         evsel->prev_raw_counts = NULL;
221 }
222
223 static void evsel__reset_prev_raw_counts(struct evsel *evsel)
224 {
225         if (evsel->prev_raw_counts)
226                 perf_counts__reset(evsel->prev_raw_counts);
227 }
228
229 static int evsel__alloc_stats(struct evsel *evsel, int nr_aggr, bool alloc_raw)
230 {
231         if (evsel__alloc_stat_priv(evsel, nr_aggr) < 0 ||
232             evsel__alloc_counts(evsel) < 0 ||
233             (alloc_raw && evsel__alloc_prev_raw_counts(evsel) < 0))
234                 return -ENOMEM;
235
236         return 0;
237 }
238
239 int evlist__alloc_stats(struct perf_stat_config *config,
240                         struct evlist *evlist, bool alloc_raw)
241 {
242         struct evsel *evsel;
243         int nr_aggr = 0;
244
245         if (config && config->aggr_map)
246                 nr_aggr = config->aggr_map->nr;
247
248         evlist__for_each_entry(evlist, evsel) {
249                 if (evsel__alloc_stats(evsel, nr_aggr, alloc_raw))
250                         goto out_free;
251         }
252
253         return 0;
254
255 out_free:
256         evlist__free_stats(evlist);
257         return -1;
258 }
259
260 void evlist__free_stats(struct evlist *evlist)
261 {
262         struct evsel *evsel;
263
264         evlist__for_each_entry(evlist, evsel) {
265                 evsel__free_stat_priv(evsel);
266                 evsel__free_counts(evsel);
267                 evsel__free_prev_raw_counts(evsel);
268         }
269 }
270
271 void evlist__reset_stats(struct evlist *evlist)
272 {
273         struct evsel *evsel;
274
275         evlist__for_each_entry(evlist, evsel) {
276                 evsel__reset_stat_priv(evsel);
277                 evsel__reset_counts(evsel);
278         }
279 }
280
281 void evlist__reset_aggr_stats(struct evlist *evlist)
282 {
283         struct evsel *evsel;
284
285         evlist__for_each_entry(evlist, evsel)
286                 evsel__reset_aggr_stats(evsel);
287 }
288
289 void evlist__reset_prev_raw_counts(struct evlist *evlist)
290 {
291         struct evsel *evsel;
292
293         evlist__for_each_entry(evlist, evsel)
294                 evsel__reset_prev_raw_counts(evsel);
295 }
296
297 static void evsel__copy_prev_raw_counts(struct evsel *evsel)
298 {
299         int idx, nthreads = perf_thread_map__nr(evsel->core.threads);
300
301         for (int thread = 0; thread < nthreads; thread++) {
302                 perf_cpu_map__for_each_idx(idx, evsel__cpus(evsel)) {
303                         *perf_counts(evsel->counts, idx, thread) =
304                                 *perf_counts(evsel->prev_raw_counts, idx, thread);
305                 }
306         }
307 }
308
309 void evlist__copy_prev_raw_counts(struct evlist *evlist)
310 {
311         struct evsel *evsel;
312
313         evlist__for_each_entry(evlist, evsel)
314                 evsel__copy_prev_raw_counts(evsel);
315 }
316
317 static size_t pkg_id_hash(long __key, void *ctx __maybe_unused)
318 {
319         uint64_t *key = (uint64_t *) __key;
320
321         return *key & 0xffffffff;
322 }
323
324 static bool pkg_id_equal(long __key1, long __key2, void *ctx __maybe_unused)
325 {
326         uint64_t *key1 = (uint64_t *) __key1;
327         uint64_t *key2 = (uint64_t *) __key2;
328
329         return *key1 == *key2;
330 }
331
332 static int check_per_pkg(struct evsel *counter, struct perf_counts_values *vals,
333                          int cpu_map_idx, bool *skip)
334 {
335         struct hashmap *mask = counter->per_pkg_mask;
336         struct perf_cpu_map *cpus = evsel__cpus(counter);
337         struct perf_cpu cpu = perf_cpu_map__cpu(cpus, cpu_map_idx);
338         int s, d, ret = 0;
339         uint64_t *key;
340
341         *skip = false;
342
343         if (!counter->per_pkg)
344                 return 0;
345
346         if (perf_cpu_map__empty(cpus))
347                 return 0;
348
349         if (!mask) {
350                 mask = hashmap__new(pkg_id_hash, pkg_id_equal, NULL);
351                 if (IS_ERR(mask))
352                         return -ENOMEM;
353
354                 counter->per_pkg_mask = mask;
355         }
356
357         /*
358          * we do not consider an event that has not run as a good
359          * instance to mark a package as used (skip=1). Otherwise
360          * we may run into a situation where the first CPU in a package
361          * is not running anything, yet the second is, and this function
362          * would mark the package as used after the first CPU and would
363          * not read the values from the second CPU.
364          */
365         if (!(vals->run && vals->ena))
366                 return 0;
367
368         s = cpu__get_socket_id(cpu);
369         if (s < 0)
370                 return -1;
371
372         /*
373          * On multi-die system, die_id > 0. On no-die system, die_id = 0.
374          * We use hashmap(socket, die) to check the used socket+die pair.
375          */
376         d = cpu__get_die_id(cpu);
377         if (d < 0)
378                 return -1;
379
380         key = malloc(sizeof(*key));
381         if (!key)
382                 return -ENOMEM;
383
384         *key = (uint64_t)d << 32 | s;
385         if (hashmap__find(mask, key, NULL)) {
386                 *skip = true;
387                 free(key);
388         } else
389                 ret = hashmap__add(mask, key, 1);
390
391         return ret;
392 }
393
394 static bool evsel__count_has_error(struct evsel *evsel,
395                                    struct perf_counts_values *count,
396                                    struct perf_stat_config *config)
397 {
398         /* the evsel was failed already */
399         if (evsel->err || evsel->counts->scaled == -1)
400                 return true;
401
402         /* this is meaningful for CPU aggregation modes only */
403         if (config->aggr_mode == AGGR_GLOBAL)
404                 return false;
405
406         /* it's considered ok when it actually ran */
407         if (count->ena != 0 && count->run != 0)
408                 return false;
409
410         return true;
411 }
412
413 static int
414 process_counter_values(struct perf_stat_config *config, struct evsel *evsel,
415                        int cpu_map_idx, int thread,
416                        struct perf_counts_values *count)
417 {
418         struct perf_stat_evsel *ps = evsel->stats;
419         static struct perf_counts_values zero;
420         bool skip = false;
421
422         if (check_per_pkg(evsel, count, cpu_map_idx, &skip)) {
423                 pr_err("failed to read per-pkg counter\n");
424                 return -1;
425         }
426
427         if (skip)
428                 count = &zero;
429
430         if (!evsel->snapshot)
431                 evsel__compute_deltas(evsel, cpu_map_idx, thread, count);
432         perf_counts_values__scale(count, config->scale, NULL);
433
434         if (config->aggr_mode == AGGR_THREAD) {
435                 struct perf_counts_values *aggr_counts = &ps->aggr[thread].counts;
436
437                 /*
438                  * Skip value 0 when enabling --per-thread globally,
439                  * otherwise too many 0 output.
440                  */
441                 if (count->val == 0 && config->system_wide)
442                         return 0;
443
444                 ps->aggr[thread].nr++;
445
446                 aggr_counts->val += count->val;
447                 aggr_counts->ena += count->ena;
448                 aggr_counts->run += count->run;
449                 return 0;
450         }
451
452         if (ps->aggr) {
453                 struct perf_cpu cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
454                 struct aggr_cpu_id aggr_id = config->aggr_get_id(config, cpu);
455                 struct perf_stat_aggr *ps_aggr;
456                 int i;
457
458                 for (i = 0; i < ps->nr_aggr; i++) {
459                         if (!aggr_cpu_id__equal(&aggr_id, &config->aggr_map->map[i]))
460                                 continue;
461
462                         ps_aggr = &ps->aggr[i];
463                         ps_aggr->nr++;
464
465                         /*
466                          * When any result is bad, make them all to give consistent output
467                          * in interval mode.  But per-task counters can have 0 enabled time
468                          * when some tasks are idle.
469                          */
470                         if (evsel__count_has_error(evsel, count, config) && !ps_aggr->failed) {
471                                 ps_aggr->counts.val = 0;
472                                 ps_aggr->counts.ena = 0;
473                                 ps_aggr->counts.run = 0;
474                                 ps_aggr->failed = true;
475                         }
476
477                         if (!ps_aggr->failed) {
478                                 ps_aggr->counts.val += count->val;
479                                 ps_aggr->counts.ena += count->ena;
480                                 ps_aggr->counts.run += count->run;
481                         }
482                         break;
483                 }
484         }
485
486         return 0;
487 }
488
489 static int process_counter_maps(struct perf_stat_config *config,
490                                 struct evsel *counter)
491 {
492         int nthreads = perf_thread_map__nr(counter->core.threads);
493         int ncpus = evsel__nr_cpus(counter);
494         int idx, thread;
495
496         for (thread = 0; thread < nthreads; thread++) {
497                 for (idx = 0; idx < ncpus; idx++) {
498                         if (process_counter_values(config, counter, idx, thread,
499                                                    perf_counts(counter->counts, idx, thread)))
500                                 return -1;
501                 }
502         }
503
504         return 0;
505 }
506
507 int perf_stat_process_counter(struct perf_stat_config *config,
508                               struct evsel *counter)
509 {
510         struct perf_stat_evsel *ps = counter->stats;
511         u64 *count;
512         int ret;
513
514         if (counter->per_pkg)
515                 evsel__zero_per_pkg(counter);
516
517         ret = process_counter_maps(config, counter);
518         if (ret)
519                 return ret;
520
521         if (config->aggr_mode != AGGR_GLOBAL)
522                 return 0;
523
524         /*
525          * GLOBAL aggregation mode only has a single aggr counts,
526          * so we can use ps->aggr[0] as the actual output.
527          */
528         count = ps->aggr[0].counts.values;
529         update_stats(&ps->res_stats, *count);
530
531         if (verbose > 0) {
532                 fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
533                         evsel__name(counter), count[0], count[1], count[2]);
534         }
535
536         return 0;
537 }
538
539 static int evsel__merge_aggr_counters(struct evsel *evsel, struct evsel *alias)
540 {
541         struct perf_stat_evsel *ps_a = evsel->stats;
542         struct perf_stat_evsel *ps_b = alias->stats;
543         int i;
544
545         if (ps_a->aggr == NULL && ps_b->aggr == NULL)
546                 return 0;
547
548         if (ps_a->nr_aggr != ps_b->nr_aggr) {
549                 pr_err("Unmatched aggregation mode between aliases\n");
550                 return -1;
551         }
552
553         for (i = 0; i < ps_a->nr_aggr; i++) {
554                 struct perf_counts_values *aggr_counts_a = &ps_a->aggr[i].counts;
555                 struct perf_counts_values *aggr_counts_b = &ps_b->aggr[i].counts;
556
557                 /* NB: don't increase aggr.nr for aliases */
558
559                 aggr_counts_a->val += aggr_counts_b->val;
560                 aggr_counts_a->ena += aggr_counts_b->ena;
561                 aggr_counts_a->run += aggr_counts_b->run;
562         }
563
564         return 0;
565 }
566 /* events should have the same name, scale, unit, cgroup but on different PMUs */
567 static bool evsel__is_alias(struct evsel *evsel_a, struct evsel *evsel_b)
568 {
569         if (strcmp(evsel__name(evsel_a), evsel__name(evsel_b)))
570                 return false;
571
572         if (evsel_a->scale != evsel_b->scale)
573                 return false;
574
575         if (evsel_a->cgrp != evsel_b->cgrp)
576                 return false;
577
578         if (strcmp(evsel_a->unit, evsel_b->unit))
579                 return false;
580
581         if (evsel__is_clock(evsel_a) != evsel__is_clock(evsel_b))
582                 return false;
583
584         return !!strcmp(evsel_a->pmu_name, evsel_b->pmu_name);
585 }
586
587 static void evsel__merge_aliases(struct evsel *evsel)
588 {
589         struct evlist *evlist = evsel->evlist;
590         struct evsel *alias;
591
592         alias = list_prepare_entry(evsel, &(evlist->core.entries), core.node);
593         list_for_each_entry_continue(alias, &evlist->core.entries, core.node) {
594                 /* Merge the same events on different PMUs. */
595                 if (evsel__is_alias(evsel, alias)) {
596                         evsel__merge_aggr_counters(evsel, alias);
597                         alias->merged_stat = true;
598                 }
599         }
600 }
601
602 static bool evsel__should_merge_hybrid(const struct evsel *evsel,
603                                        const struct perf_stat_config *config)
604 {
605         return config->hybrid_merge && evsel__is_hybrid(evsel);
606 }
607
608 static void evsel__merge_stats(struct evsel *evsel, struct perf_stat_config *config)
609 {
610         /* this evsel is already merged */
611         if (evsel->merged_stat)
612                 return;
613
614         if (evsel->auto_merge_stats || evsel__should_merge_hybrid(evsel, config))
615                 evsel__merge_aliases(evsel);
616 }
617
618 /* merge the same uncore and hybrid events if requested */
619 void perf_stat_merge_counters(struct perf_stat_config *config, struct evlist *evlist)
620 {
621         struct evsel *evsel;
622
623         if (config->no_merge)
624                 return;
625
626         evlist__for_each_entry(evlist, evsel)
627                 evsel__merge_stats(evsel, config);
628 }
629
630 static void evsel__update_percore_stats(struct evsel *evsel, struct aggr_cpu_id *core_id)
631 {
632         struct perf_stat_evsel *ps = evsel->stats;
633         struct perf_counts_values counts = { 0, };
634         struct aggr_cpu_id id;
635         struct perf_cpu cpu;
636         int idx;
637
638         /* collect per-core counts */
639         perf_cpu_map__for_each_cpu(cpu, idx, evsel->core.cpus) {
640                 struct perf_stat_aggr *aggr = &ps->aggr[idx];
641
642                 id = aggr_cpu_id__core(cpu, NULL);
643                 if (!aggr_cpu_id__equal(core_id, &id))
644                         continue;
645
646                 counts.val += aggr->counts.val;
647                 counts.ena += aggr->counts.ena;
648                 counts.run += aggr->counts.run;
649         }
650
651         /* update aggregated per-core counts for each CPU */
652         perf_cpu_map__for_each_cpu(cpu, idx, evsel->core.cpus) {
653                 struct perf_stat_aggr *aggr = &ps->aggr[idx];
654
655                 id = aggr_cpu_id__core(cpu, NULL);
656                 if (!aggr_cpu_id__equal(core_id, &id))
657                         continue;
658
659                 aggr->counts.val = counts.val;
660                 aggr->counts.ena = counts.ena;
661                 aggr->counts.run = counts.run;
662
663                 aggr->used = true;
664         }
665 }
666
667 /* we have an aggr_map for cpu, but want to aggregate the counters per-core */
668 static void evsel__process_percore(struct evsel *evsel)
669 {
670         struct perf_stat_evsel *ps = evsel->stats;
671         struct aggr_cpu_id core_id;
672         struct perf_cpu cpu;
673         int idx;
674
675         if (!evsel->percore)
676                 return;
677
678         perf_cpu_map__for_each_cpu(cpu, idx, evsel->core.cpus) {
679                 struct perf_stat_aggr *aggr = &ps->aggr[idx];
680
681                 if (aggr->used)
682                         continue;
683
684                 core_id = aggr_cpu_id__core(cpu, NULL);
685                 evsel__update_percore_stats(evsel, &core_id);
686         }
687 }
688
689 /* process cpu stats on per-core events */
690 void perf_stat_process_percore(struct perf_stat_config *config, struct evlist *evlist)
691 {
692         struct evsel *evsel;
693
694         if (config->aggr_mode != AGGR_NONE)
695                 return;
696
697         evlist__for_each_entry(evlist, evsel)
698                 evsel__process_percore(evsel);
699 }
700
701 static void evsel__update_shadow_stats(struct evsel *evsel)
702 {
703         struct perf_stat_evsel *ps = evsel->stats;
704         int i;
705
706         if (ps->aggr == NULL)
707                 return;
708
709         for (i = 0; i < ps->nr_aggr; i++) {
710                 struct perf_counts_values *aggr_counts = &ps->aggr[i].counts;
711
712                 perf_stat__update_shadow_stats(evsel, aggr_counts->val, i, &rt_stat);
713         }
714 }
715
716 void perf_stat_process_shadow_stats(struct perf_stat_config *config __maybe_unused,
717                                     struct evlist *evlist)
718 {
719         struct evsel *evsel;
720
721         evlist__for_each_entry(evlist, evsel)
722                 evsel__update_shadow_stats(evsel);
723 }
724
725 int perf_event__process_stat_event(struct perf_session *session,
726                                    union perf_event *event)
727 {
728         struct perf_counts_values count, *ptr;
729         struct perf_record_stat *st = &event->stat;
730         struct evsel *counter;
731         int cpu_map_idx;
732
733         count.val = st->val;
734         count.ena = st->ena;
735         count.run = st->run;
736
737         counter = evlist__id2evsel(session->evlist, st->id);
738         if (!counter) {
739                 pr_err("Failed to resolve counter for stat event.\n");
740                 return -EINVAL;
741         }
742         cpu_map_idx = perf_cpu_map__idx(evsel__cpus(counter), (struct perf_cpu){.cpu = st->cpu});
743         if (cpu_map_idx == -1) {
744                 pr_err("Invalid CPU %d for event %s.\n", st->cpu, evsel__name(counter));
745                 return -EINVAL;
746         }
747         ptr = perf_counts(counter->counts, cpu_map_idx, st->thread);
748         if (ptr == NULL) {
749                 pr_err("Failed to find perf count for CPU %d thread %d on event %s.\n",
750                         st->cpu, st->thread, evsel__name(counter));
751                 return -EINVAL;
752         }
753         *ptr = count;
754         counter->supported = true;
755         return 0;
756 }
757
758 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
759 {
760         struct perf_record_stat *st = (struct perf_record_stat *)event;
761         size_t ret;
762
763         ret  = fprintf(fp, "\n... id %" PRI_lu64 ", cpu %d, thread %d\n",
764                        st->id, st->cpu, st->thread);
765         ret += fprintf(fp, "... value %" PRI_lu64 ", enabled %" PRI_lu64 ", running %" PRI_lu64 "\n",
766                        st->val, st->ena, st->run);
767
768         return ret;
769 }
770
771 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
772 {
773         struct perf_record_stat_round *rd = (struct perf_record_stat_round *)event;
774         size_t ret;
775
776         ret = fprintf(fp, "\n... time %" PRI_lu64 ", type %s\n", rd->time,
777                       rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
778
779         return ret;
780 }
781
782 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
783 {
784         struct perf_stat_config sc;
785         size_t ret;
786
787         perf_event__read_stat_config(&sc, &event->stat_config);
788
789         ret  = fprintf(fp, "\n");
790         ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
791         ret += fprintf(fp, "... scale     %d\n", sc.scale);
792         ret += fprintf(fp, "... interval  %u\n", sc.interval);
793
794         return ret;
795 }
796
797 int create_perf_stat_counter(struct evsel *evsel,
798                              struct perf_stat_config *config,
799                              struct target *target,
800                              int cpu_map_idx)
801 {
802         struct perf_event_attr *attr = &evsel->core.attr;
803         struct evsel *leader = evsel__leader(evsel);
804
805         attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
806                             PERF_FORMAT_TOTAL_TIME_RUNNING;
807
808         /*
809          * The event is part of non trivial group, let's enable
810          * the group read (for leader) and ID retrieval for all
811          * members.
812          */
813         if (leader->core.nr_members > 1)
814                 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
815
816         attr->inherit = !config->no_inherit && list_empty(&evsel->bpf_counter_list);
817
818         /*
819          * Some events get initialized with sample_(period/type) set,
820          * like tracepoints. Clear it up for counting.
821          */
822         attr->sample_period = 0;
823
824         if (config->identifier)
825                 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
826
827         if (config->all_user) {
828                 attr->exclude_kernel = 1;
829                 attr->exclude_user   = 0;
830         }
831
832         if (config->all_kernel) {
833                 attr->exclude_kernel = 0;
834                 attr->exclude_user   = 1;
835         }
836
837         /*
838          * Disabling all counters initially, they will be enabled
839          * either manually by us or by kernel via enable_on_exec
840          * set later.
841          */
842         if (evsel__is_group_leader(evsel)) {
843                 attr->disabled = 1;
844
845                 /*
846                  * In case of initial_delay we enable tracee
847                  * events manually.
848                  */
849                 if (target__none(target) && !config->initial_delay)
850                         attr->enable_on_exec = 1;
851         }
852
853         if (target__has_cpu(target) && !target__has_per_thread(target))
854                 return evsel__open_per_cpu(evsel, evsel__cpus(evsel), cpu_map_idx);
855
856         return evsel__open_per_thread(evsel, evsel->core.threads);
857 }