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(const void *__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(const void *__key1, const void *__key2,
 325                          void *ctx __maybe_unused)
 326 {
 327         uint64_t *key1 = (uint64_t *) __key1;
 328         uint64_t *key2 = (uint64_t *) __key2;
 329
 330         return *key1 == *key2;
 331 }
 332
 333 static int check_per_pkg(struct evsel *counter, struct perf_counts_values *vals,
 334                          int cpu_map_idx, bool *skip)
 335 {
 336         struct hashmap *mask = counter->per_pkg_mask;
 337         struct perf_cpu_map *cpus = evsel__cpus(counter);
 338         struct perf_cpu cpu = perf_cpu_map__cpu(cpus, cpu_map_idx);
 339         int s, d, ret = 0;
 340         uint64_t *key;
 341
 342         *skip = false;
 343
 344         if (!counter->per_pkg)
 345                 return 0;
 346
 347         if (perf_cpu_map__empty(cpus))
 348                 return 0;
 349
 350         if (!mask) {
 351                 mask = hashmap__new(pkg_id_hash, pkg_id_equal, NULL);
 352                 if (IS_ERR(mask))
 353                         return -ENOMEM;
 354
 355                 counter->per_pkg_mask = mask;
 356         }
 357
 358         /*
 359          * we do not consider an event that has not run as a good
 360          * instance to mark a package as used (skip=1). Otherwise
 361          * we may run into a situation where the first CPU in a package
 362          * is not running anything, yet the second is, and this function
 363          * would mark the package as used after the first CPU and would
 364          * not read the values from the second CPU.
 365          */
 366         if (!(vals->run && vals->ena))
 367                 return 0;
 368
 369         s = cpu__get_socket_id(cpu);
 370         if (s < 0)
 371                 return -1;
 372
 373         /*
 374          * On multi-die system, die_id > 0. On no-die system, die_id = 0.
 375          * We use hashmap(socket, die) to check the used socket+die pair.
 376          */
 377         d = cpu__get_die_id(cpu);
 378         if (d < 0)
 379                 return -1;
 380
 381         key = malloc(sizeof(*key));
 382         if (!key)
 383                 return -ENOMEM;
 384
 385         *key = (uint64_t)d << 32 | s;
 386         if (hashmap__find(mask, (void *)key, NULL)) {
 387                 *skip = true;
 388                 free(key);
 389         } else
 390                 ret = hashmap__add(mask, (void *)key, (void *)1);
 391
 392         return ret;
 393 }
 394
 395 static bool evsel__count_has_error(struct evsel *evsel,
 396                                    struct perf_counts_values *count,
 397                                    struct perf_stat_config *config)
 398 {
 399         /* the evsel was failed already */
 400         if (evsel->err || evsel->counts->scaled == -1)
 401                 return true;
 402
 403         /* this is meaningful for CPU aggregation modes only */
 404         if (config->aggr_mode == AGGR_GLOBAL)
 405                 return false;
 406
 407         /* it's considered ok when it actually ran */
 408         if (count->ena != 0 && count->run != 0)
 409                 return false;
 410
 411         return true;
 412 }
 413
 414 static int
 415 process_counter_values(struct perf_stat_config *config, struct evsel *evsel,
 416                        int cpu_map_idx, int thread,
 417                        struct perf_counts_values *count)
 418 {
 419         struct perf_stat_evsel *ps = evsel->stats;
 420         static struct perf_counts_values zero;
 421         bool skip = false;
 422
 423         if (check_per_pkg(evsel, count, cpu_map_idx, &skip)) {
 424                 pr_err("failed to read per-pkg counter\n");
 425                 return -1;
 426         }
 427
 428         if (skip)
 429                 count = &zero;
 430
 431         if (!evsel->snapshot)
 432                 evsel__compute_deltas(evsel, cpu_map_idx, thread, count);
 433         perf_counts_values__scale(count, config->scale, NULL);
 434
 435         if (config->aggr_mode == AGGR_THREAD) {
 436                 struct perf_counts_values *aggr_counts = &ps->aggr[thread].counts;
 437
 438                 /*
 439                  * Skip value 0 when enabling --per-thread globally,
 440                  * otherwise too many 0 output.
 441                  */
 442                 if (count->val == 0 && config->system_wide)
 443                         return 0;
 444
 445                 ps->aggr[thread].nr++;
 446
 447                 aggr_counts->val += count->val;
 448                 aggr_counts->ena += count->ena;
 449                 aggr_counts->run += count->run;
 450                 return 0;
 451         }
 452
 453         if (ps->aggr) {
 454                 struct perf_cpu cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
 455                 struct aggr_cpu_id aggr_id = config->aggr_get_id(config, cpu);
 456                 struct perf_stat_aggr *ps_aggr;
 457                 int i;
 458
 459                 for (i = 0; i < ps->nr_aggr; i++) {
 460                         if (!aggr_cpu_id__equal(&aggr_id, &config->aggr_map->map[i]))
 461                                 continue;
 462
 463                         ps_aggr = &ps->aggr[i];
 464                         ps_aggr->nr++;
 465
 466                         /*
 467                          * When any result is bad, make them all to give consistent output
 468                          * in interval mode.  But per-task counters can have 0 enabled time
 469                          * when some tasks are idle.
 470                          */
 471                         if (evsel__count_has_error(evsel, count, config) && !ps_aggr->failed) {
 472                                 ps_aggr->counts.val = 0;
 473                                 ps_aggr->counts.ena = 0;
 474                                 ps_aggr->counts.run = 0;
 475                                 ps_aggr->failed = true;
 476                         }
 477
 478                         if (!ps_aggr->failed) {
 479                                 ps_aggr->counts.val += count->val;
 480                                 ps_aggr->counts.ena += count->ena;
 481                                 ps_aggr->counts.run += count->run;
 482                         }
 483                         break;
 484                 }
 485         }
 486
 487         return 0;
 488 }
 489
 490 static int process_counter_maps(struct perf_stat_config *config,
 491                                 struct evsel *counter)
 492 {
 493         int nthreads = perf_thread_map__nr(counter->core.threads);
 494         int ncpus = evsel__nr_cpus(counter);
 495         int idx, thread;
 496
 497         for (thread = 0; thread < nthreads; thread++) {
 498                 for (idx = 0; idx < ncpus; idx++) {
 499                         if (process_counter_values(config, counter, idx, thread,
 500                                                    perf_counts(counter->counts, idx, thread)))
 501                                 return -1;
 502                 }
 503         }
 504
 505         return 0;
 506 }
 507
 508 int perf_stat_process_counter(struct perf_stat_config *config,
 509                               struct evsel *counter)
 510 {
 511         struct perf_stat_evsel *ps = counter->stats;
 512         u64 *count;
 513         int ret;
 514
 515         if (counter->per_pkg)
 516                 evsel__zero_per_pkg(counter);
 517
 518         ret = process_counter_maps(config, counter);
 519         if (ret)
 520                 return ret;
 521
 522         if (config->aggr_mode != AGGR_GLOBAL)
 523                 return 0;
 524
 525         /*
 526          * GLOBAL aggregation mode only has a single aggr counts,
 527          * so we can use ps->aggr[0] as the actual output.
 528          */
 529         count = ps->aggr[0].counts.values;
 530         update_stats(&ps->res_stats, *count);
 531
 532         if (verbose > 0) {
 533                 fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 534                         evsel__name(counter), count[0], count[1], count[2]);
 535         }
 536
 537         return 0;
 538 }
 539
 540 static int evsel__merge_aggr_counters(struct evsel *evsel, struct evsel *alias)
 541 {
 542         struct perf_stat_evsel *ps_a = evsel->stats;
 543         struct perf_stat_evsel *ps_b = alias->stats;
 544         int i;
 545
 546         if (ps_a->aggr == NULL && ps_b->aggr == NULL)
 547                 return 0;
 548
 549         if (ps_a->nr_aggr != ps_b->nr_aggr) {
 550                 pr_err("Unmatched aggregation mode between aliases\n");
 551                 return -1;
 552         }
 553
 554         for (i = 0; i < ps_a->nr_aggr; i++) {
 555                 struct perf_counts_values *aggr_counts_a = &ps_a->aggr[i].counts;
 556                 struct perf_counts_values *aggr_counts_b = &ps_b->aggr[i].counts;
 557
 558                 /* NB: don't increase aggr.nr for aliases */
 559
 560                 aggr_counts_a->val += aggr_counts_b->val;
 561                 aggr_counts_a->ena += aggr_counts_b->ena;
 562                 aggr_counts_a->run += aggr_counts_b->run;
 563         }
 564
 565         return 0;
 566 }
 567 /* events should have the same name, scale, unit, cgroup but on different PMUs */
 568 static bool evsel__is_alias(struct evsel *evsel_a, struct evsel *evsel_b)
 569 {
 570         if (strcmp(evsel__name(evsel_a), evsel__name(evsel_b)))
 571                 return false;
 572
 573         if (evsel_a->scale != evsel_b->scale)
 574                 return false;
 575
 576         if (evsel_a->cgrp != evsel_b->cgrp)
 577                 return false;
 578
 579         if (strcmp(evsel_a->unit, evsel_b->unit))
 580                 return false;
 581
 582         if (evsel__is_clock(evsel_a) != evsel__is_clock(evsel_b))
 583                 return false;
 584
 585         return !!strcmp(evsel_a->pmu_name, evsel_b->pmu_name);
 586 }
 587
 588 static void evsel__merge_aliases(struct evsel *evsel)
 589 {
 590         struct evlist *evlist = evsel->evlist;
 591         struct evsel *alias;
 592
 593         alias = list_prepare_entry(evsel, &(evlist->core.entries), core.node);
 594         list_for_each_entry_continue(alias, &evlist->core.entries, core.node) {
 595                 /* Merge the same events on different PMUs. */
 596                 if (evsel__is_alias(evsel, alias)) {
 597                         evsel__merge_aggr_counters(evsel, alias);
 598                         alias->merged_stat = true;
 599                 }
 600         }
 601 }
 602
 603 static bool evsel__should_merge_hybrid(const struct evsel *evsel,
 604                                        const struct perf_stat_config *config)
 605 {
 606         return config->hybrid_merge && evsel__is_hybrid(evsel);
 607 }
 608
 609 static void evsel__merge_stats(struct evsel *evsel, struct perf_stat_config *config)
 610 {
 611         /* this evsel is already merged */
 612         if (evsel->merged_stat)
 613                 return;
 614
 615         if (evsel->auto_merge_stats || evsel__should_merge_hybrid(evsel, config))
 616                 evsel__merge_aliases(evsel);
 617 }
 618
 619 /* merge the same uncore and hybrid events if requested */
 620 void perf_stat_merge_counters(struct perf_stat_config *config, struct evlist *evlist)
 621 {
 622         struct evsel *evsel;
 623
 624         if (config->no_merge)
 625                 return;
 626
 627         evlist__for_each_entry(evlist, evsel)
 628                 evsel__merge_stats(evsel, config);
 629 }
 630
 631 static void evsel__update_percore_stats(struct evsel *evsel, struct aggr_cpu_id *core_id)
 632 {
 633         struct perf_stat_evsel *ps = evsel->stats;
 634         struct perf_counts_values counts = { 0, };
 635         struct aggr_cpu_id id;
 636         struct perf_cpu cpu;
 637         int idx;
 638
 639         /* collect per-core counts */
 640         perf_cpu_map__for_each_cpu(cpu, idx, evsel->core.cpus) {
 641                 struct perf_stat_aggr *aggr = &ps->aggr[idx];
 642
 643                 id = aggr_cpu_id__core(cpu, NULL);
 644                 if (!aggr_cpu_id__equal(core_id, &id))
 645                         continue;
 646
 647                 counts.val += aggr->counts.val;
 648                 counts.ena += aggr->counts.ena;
 649                 counts.run += aggr->counts.run;
 650         }
 651
 652         /* update aggregated per-core counts for each CPU */
 653         perf_cpu_map__for_each_cpu(cpu, idx, evsel->core.cpus) {
 654                 struct perf_stat_aggr *aggr = &ps->aggr[idx];
 655
 656                 id = aggr_cpu_id__core(cpu, NULL);
 657                 if (!aggr_cpu_id__equal(core_id, &id))
 658                         continue;
 659
 660                 aggr->counts.val = counts.val;
 661                 aggr->counts.ena = counts.ena;
 662                 aggr->counts.run = counts.run;
 663
 664                 aggr->used = true;
 665         }
 666 }
 667
 668 /* we have an aggr_map for cpu, but want to aggregate the counters per-core */
 669 static void evsel__process_percore(struct evsel *evsel)
 670 {
 671         struct perf_stat_evsel *ps = evsel->stats;
 672         struct aggr_cpu_id core_id;
 673         struct perf_cpu cpu;
 674         int idx;
 675
 676         if (!evsel->percore)
 677                 return;
 678
 679         perf_cpu_map__for_each_cpu(cpu, idx, evsel->core.cpus) {
 680                 struct perf_stat_aggr *aggr = &ps->aggr[idx];
 681
 682                 if (aggr->used)
 683                         continue;
 684
 685                 core_id = aggr_cpu_id__core(cpu, NULL);
 686                 evsel__update_percore_stats(evsel, &core_id);
 687         }
 688 }
 689
 690 /* process cpu stats on per-core events */
 691 void perf_stat_process_percore(struct perf_stat_config *config, struct evlist *evlist)
 692 {
 693         struct evsel *evsel;
 694
 695         if (config->aggr_mode != AGGR_NONE)
 696                 return;
 697
 698         evlist__for_each_entry(evlist, evsel)
 699                 evsel__process_percore(evsel);
 700 }
 701
 702 static void evsel__update_shadow_stats(struct evsel *evsel)
 703 {
 704         struct perf_stat_evsel *ps = evsel->stats;
 705         int i;
 706
 707         if (ps->aggr == NULL)
 708                 return;
 709
 710         for (i = 0; i < ps->nr_aggr; i++) {
 711                 struct perf_counts_values *aggr_counts = &ps->aggr[i].counts;
 712
 713                 perf_stat__update_shadow_stats(evsel, aggr_counts->val, i, &rt_stat);
 714         }
 715 }
 716
 717 void perf_stat_process_shadow_stats(struct perf_stat_config *config __maybe_unused,
 718                                     struct evlist *evlist)
 719 {
 720         struct evsel *evsel;
 721
 722         evlist__for_each_entry(evlist, evsel)
 723                 evsel__update_shadow_stats(evsel);
 724 }
 725
 726 int perf_event__process_stat_event(struct perf_session *session,
 727                                    union perf_event *event)
 728 {
 729         struct perf_counts_values count, *ptr;
 730         struct perf_record_stat *st = &event->stat;
 731         struct evsel *counter;
 732         int cpu_map_idx;
 733
 734         count.val = st->val;
 735         count.ena = st->ena;
 736         count.run = st->run;
 737
 738         counter = evlist__id2evsel(session->evlist, st->id);
 739         if (!counter) {
 740                 pr_err("Failed to resolve counter for stat event.\n");
 741                 return -EINVAL;
 742         }
 743         cpu_map_idx = perf_cpu_map__idx(evsel__cpus(counter), (struct perf_cpu){.cpu = st->cpu});
 744         if (cpu_map_idx == -1) {
 745                 pr_err("Invalid CPU %d for event %s.\n", st->cpu, evsel__name(counter));
 746                 return -EINVAL;
 747         }
 748         ptr = perf_counts(counter->counts, cpu_map_idx, st->thread);
 749         if (ptr == NULL) {
 750                 pr_err("Failed to find perf count for CPU %d thread %d on event %s.\n",
 751                         st->cpu, st->thread, evsel__name(counter));
 752                 return -EINVAL;
 753         }
 754         *ptr = count;
 755         counter->supported = true;
 756         return 0;
 757 }
 758
 759 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
 760 {
 761         struct perf_record_stat *st = (struct perf_record_stat *)event;
 762         size_t ret;
 763
 764         ret  = fprintf(fp, "\n... id %" PRI_lu64 ", cpu %d, thread %d\n",
 765                        st->id, st->cpu, st->thread);
 766         ret += fprintf(fp, "... value %" PRI_lu64 ", enabled %" PRI_lu64 ", running %" PRI_lu64 "\n",
 767                        st->val, st->ena, st->run);
 768
 769         return ret;
 770 }
 771
 772 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
 773 {
 774         struct perf_record_stat_round *rd = (struct perf_record_stat_round *)event;
 775         size_t ret;
 776
 777         ret = fprintf(fp, "\n... time %" PRI_lu64 ", type %s\n", rd->time,
 778                       rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
 779
 780         return ret;
 781 }
 782
 783 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
 784 {
 785         struct perf_stat_config sc;
 786         size_t ret;
 787
 788         perf_event__read_stat_config(&sc, &event->stat_config);
 789
 790         ret  = fprintf(fp, "\n");
 791         ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
 792         ret += fprintf(fp, "... scale     %d\n", sc.scale);
 793         ret += fprintf(fp, "... interval  %u\n", sc.interval);
 794
 795         return ret;
 796 }
 797
 798 int create_perf_stat_counter(struct evsel *evsel,
 799                              struct perf_stat_config *config,
 800                              struct target *target,
 801                              int cpu_map_idx)
 802 {
 803         struct perf_event_attr *attr = &evsel->core.attr;
 804         struct evsel *leader = evsel__leader(evsel);
 805
 806         attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 807                             PERF_FORMAT_TOTAL_TIME_RUNNING;
 808
 809         /*
 810          * The event is part of non trivial group, let's enable
 811          * the group read (for leader) and ID retrieval for all
 812          * members.
 813          */
 814         if (leader->core.nr_members > 1)
 815                 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
 816
 817         attr->inherit = !config->no_inherit && list_empty(&evsel->bpf_counter_list);
 818
 819         /*
 820          * Some events get initialized with sample_(period/type) set,
 821          * like tracepoints. Clear it up for counting.
 822          */
 823         attr->sample_period = 0;
 824
 825         if (config->identifier)
 826                 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
 827
 828         if (config->all_user) {
 829                 attr->exclude_kernel = 1;
 830                 attr->exclude_user   = 0;
 831         }
 832
 833         if (config->all_kernel) {
 834                 attr->exclude_kernel = 0;
 835                 attr->exclude_user   = 1;
 836         }
 837
 838         /*
 839          * Disabling all counters initially, they will be enabled
 840          * either manually by us or by kernel via enable_on_exec
 841          * set later.
 842          */
 843         if (evsel__is_group_leader(evsel)) {
 844                 attr->disabled = 1;
 845
 846                 /*
 847                  * In case of initial_delay we enable tracee
 848                  * events manually.
 849                  */
 850                 if (target__none(target) && !config->initial_delay)
 851                         attr->enable_on_exec = 1;
 852         }
 853
 854         if (target__has_cpu(target) && !target__has_per_thread(target))
 855                 return evsel__open_per_cpu(evsel, evsel__cpus(evsel), cpu_map_idx);
 856
 857         return evsel__open_per_thread(evsel, evsel->core.threads);
 858 }