tools/perf/util/stat.c

   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 "hashmap.h"
  17 #include <linux/zalloc.h>
  18
  19 void update_stats(struct stats *stats, u64 val)
  20 {
  21         double delta;
  22
  23         stats->n++;
  24         delta = val - stats->mean;
  25         stats->mean += delta / stats->n;
  26         stats->M2 += delta*(val - stats->mean);
  27
  28         if (val > stats->max)
  29                 stats->max = val;
  30
  31         if (val < stats->min)
  32                 stats->min = val;
  33 }
  34
  35 double avg_stats(struct stats *stats)
  36 {
  37         return stats->mean;
  38 }
  39
  40 /*
  41  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
  42  *
  43  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
  44  * s^2 = -------------------------------
  45  *                  n - 1
  46  *
  47  * http://en.wikipedia.org/wiki/Stddev
  48  *
  49  * The std dev of the mean is related to the std dev by:
  50  *
  51  *             s
  52  * s_mean = -------
  53  *          sqrt(n)
  54  *
  55  */
  56 double stddev_stats(struct stats *stats)
  57 {
  58         double variance, variance_mean;
  59
  60         if (stats->n < 2)
  61                 return 0.0;
  62
  63         variance = stats->M2 / (stats->n - 1);
  64         variance_mean = variance / stats->n;
  65
  66         return sqrt(variance_mean);
  67 }
  68
  69 double rel_stddev_stats(double stddev, double avg)
  70 {
  71         double pct = 0.0;
  72
  73         if (avg)
  74                 pct = 100.0 * stddev/avg;
  75
  76         return pct;
  77 }
  78
  79 bool __perf_stat_evsel__is(struct evsel *evsel, 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(TOPDOWN_RETIRING, topdown-retiring),
  99         ID(TOPDOWN_BAD_SPEC, topdown-bad-spec),
 100         ID(TOPDOWN_FE_BOUND, topdown-fe-bound),
 101         ID(TOPDOWN_BE_BOUND, topdown-be-bound),
 102         ID(TOPDOWN_HEAVY_OPS, topdown-heavy-ops),
 103         ID(TOPDOWN_BR_MISPREDICT, topdown-br-mispredict),
 104         ID(TOPDOWN_FETCH_LAT, topdown-fetch-lat),
 105         ID(TOPDOWN_MEM_BOUND, topdown-mem-bound),
 106         ID(SMI_NUM, msr/smi/),
 107         ID(APERF, msr/aperf/),
 108 };
 109 #undef ID
 110
 111 static void perf_stat_evsel_id_init(struct evsel *evsel)
 112 {
 113         struct perf_stat_evsel *ps = evsel->stats;
 114         int i;
 115
 116         /* ps->id is 0 hence PERF_STAT_EVSEL_ID__NONE by default */
 117
 118         for (i = 0; i < PERF_STAT_EVSEL_ID__MAX; i++) {
 119                 if (!strcmp(evsel__name(evsel), id_str[i])) {
 120                         ps->id = i;
 121                         break;
 122                 }
 123         }
 124 }
 125
 126 static void evsel__reset_stat_priv(struct evsel *evsel)
 127 {
 128         int i;
 129         struct perf_stat_evsel *ps = evsel->stats;
 130
 131         for (i = 0; i < 3; i++)
 132                 init_stats(&ps->res_stats[i]);
 133
 134         perf_stat_evsel_id_init(evsel);
 135 }
 136
 137 static int evsel__alloc_stat_priv(struct evsel *evsel)
 138 {
 139         evsel->stats = zalloc(sizeof(struct perf_stat_evsel));
 140         if (evsel->stats == NULL)
 141                 return -ENOMEM;
 142         evsel__reset_stat_priv(evsel);
 143         return 0;
 144 }
 145
 146 static void evsel__free_stat_priv(struct evsel *evsel)
 147 {
 148         struct perf_stat_evsel *ps = evsel->stats;
 149
 150         if (ps)
 151                 zfree(&ps->group_data);
 152         zfree(&evsel->stats);
 153 }
 154
 155 static int evsel__alloc_prev_raw_counts(struct evsel *evsel)
 156 {
 157         int cpu_map_nr = evsel__nr_cpus(evsel);
 158         int nthreads = perf_thread_map__nr(evsel->core.threads);
 159         struct perf_counts *counts;
 160
 161         counts = perf_counts__new(cpu_map_nr, nthreads);
 162         if (counts)
 163                 evsel->prev_raw_counts = counts;
 164
 165         return counts ? 0 : -ENOMEM;
 166 }
 167
 168 static void evsel__free_prev_raw_counts(struct evsel *evsel)
 169 {
 170         perf_counts__delete(evsel->prev_raw_counts);
 171         evsel->prev_raw_counts = NULL;
 172 }
 173
 174 static void evsel__reset_prev_raw_counts(struct evsel *evsel)
 175 {
 176         if (evsel->prev_raw_counts)
 177                 perf_counts__reset(evsel->prev_raw_counts);
 178 }
 179
 180 static int evsel__alloc_stats(struct evsel *evsel, bool alloc_raw)
 181 {
 182         if (evsel__alloc_stat_priv(evsel) < 0 ||
 183             evsel__alloc_counts(evsel) < 0 ||
 184             (alloc_raw && evsel__alloc_prev_raw_counts(evsel) < 0))
 185                 return -ENOMEM;
 186
 187         return 0;
 188 }
 189
 190 int evlist__alloc_stats(struct evlist *evlist, bool alloc_raw)
 191 {
 192         struct evsel *evsel;
 193
 194         evlist__for_each_entry(evlist, evsel) {
 195                 if (evsel__alloc_stats(evsel, alloc_raw))
 196                         goto out_free;
 197         }
 198
 199         return 0;
 200
 201 out_free:
 202         evlist__free_stats(evlist);
 203         return -1;
 204 }
 205
 206 void evlist__free_stats(struct evlist *evlist)
 207 {
 208         struct evsel *evsel;
 209
 210         evlist__for_each_entry(evlist, evsel) {
 211                 evsel__free_stat_priv(evsel);
 212                 evsel__free_counts(evsel);
 213                 evsel__free_prev_raw_counts(evsel);
 214         }
 215 }
 216
 217 void evlist__reset_stats(struct evlist *evlist)
 218 {
 219         struct evsel *evsel;
 220
 221         evlist__for_each_entry(evlist, evsel) {
 222                 evsel__reset_stat_priv(evsel);
 223                 evsel__reset_counts(evsel);
 224         }
 225 }
 226
 227 void evlist__reset_prev_raw_counts(struct evlist *evlist)
 228 {
 229         struct evsel *evsel;
 230
 231         evlist__for_each_entry(evlist, evsel)
 232                 evsel__reset_prev_raw_counts(evsel);
 233 }
 234
 235 static void evsel__copy_prev_raw_counts(struct evsel *evsel)
 236 {
 237         int ncpus = evsel__nr_cpus(evsel);
 238         int nthreads = perf_thread_map__nr(evsel->core.threads);
 239
 240         for (int thread = 0; thread < nthreads; thread++) {
 241                 for (int cpu = 0; cpu < ncpus; cpu++) {
 242                         *perf_counts(evsel->counts, cpu, thread) =
 243                                 *perf_counts(evsel->prev_raw_counts, cpu,
 244                                              thread);
 245                 }
 246         }
 247
 248         evsel->counts->aggr = evsel->prev_raw_counts->aggr;
 249 }
 250
 251 void evlist__copy_prev_raw_counts(struct evlist *evlist)
 252 {
 253         struct evsel *evsel;
 254
 255         evlist__for_each_entry(evlist, evsel)
 256                 evsel__copy_prev_raw_counts(evsel);
 257 }
 258
 259 void evlist__save_aggr_prev_raw_counts(struct evlist *evlist)
 260 {
 261         struct evsel *evsel;
 262
 263         /*
 264          * To collect the overall statistics for interval mode,
 265          * we copy the counts from evsel->prev_raw_counts to
 266          * evsel->counts. The perf_stat_process_counter creates
 267          * aggr values from per cpu values, but the per cpu values
 268          * are 0 for AGGR_GLOBAL. So we use a trick that saves the
 269          * previous aggr value to the first member of perf_counts,
 270          * then aggr calculation in process_counter_values can work
 271          * correctly.
 272          */
 273         evlist__for_each_entry(evlist, evsel) {
 274                 *perf_counts(evsel->prev_raw_counts, 0, 0) =
 275                         evsel->prev_raw_counts->aggr;
 276         }
 277 }
 278
 279 static size_t pkg_id_hash(const void *__key, void *ctx __maybe_unused)
 280 {
 281         uint64_t *key = (uint64_t *) __key;
 282
 283         return *key & 0xffffffff;
 284 }
 285
 286 static bool pkg_id_equal(const void *__key1, const void *__key2,
 287                          void *ctx __maybe_unused)
 288 {
 289         uint64_t *key1 = (uint64_t *) __key1;
 290         uint64_t *key2 = (uint64_t *) __key2;
 291
 292         return *key1 == *key2;
 293 }
 294
 295 static int check_per_pkg(struct evsel *counter, struct perf_counts_values *vals,
 296                          int cpu_map_idx, bool *skip)
 297 {
 298         struct hashmap *mask = counter->per_pkg_mask;
 299         struct perf_cpu_map *cpus = evsel__cpus(counter);
 300         struct perf_cpu cpu = perf_cpu_map__cpu(cpus, cpu_map_idx);
 301         int s, d, ret = 0;
 302         uint64_t *key;
 303
 304         *skip = false;
 305
 306         if (!counter->per_pkg)
 307                 return 0;
 308
 309         if (perf_cpu_map__empty(cpus))
 310                 return 0;
 311
 312         if (!mask) {
 313                 mask = hashmap__new(pkg_id_hash, pkg_id_equal, NULL);
 314                 if (!mask)
 315                         return -ENOMEM;
 316
 317                 counter->per_pkg_mask = mask;
 318         }
 319
 320         /*
 321          * we do not consider an event that has not run as a good
 322          * instance to mark a package as used (skip=1). Otherwise
 323          * we may run into a situation where the first CPU in a package
 324          * is not running anything, yet the second is, and this function
 325          * would mark the package as used after the first CPU and would
 326          * not read the values from the second CPU.
 327          */
 328         if (!(vals->run && vals->ena))
 329                 return 0;
 330
 331         s = cpu__get_socket_id(cpu);
 332         if (s < 0)
 333                 return -1;
 334
 335         /*
 336          * On multi-die system, die_id > 0. On no-die system, die_id = 0.
 337          * We use hashmap(socket, die) to check the used socket+die pair.
 338          */
 339         d = cpu__get_die_id(cpu);
 340         if (d < 0)
 341                 return -1;
 342
 343         key = malloc(sizeof(*key));
 344         if (!key)
 345                 return -ENOMEM;
 346
 347         *key = (uint64_t)d << 32 | s;
 348         if (hashmap__find(mask, (void *)key, NULL)) {
 349                 *skip = true;
 350                 free(key);
 351         } else
 352                 ret = hashmap__add(mask, (void *)key, (void *)1);
 353
 354         return ret;
 355 }
 356
 357 static int
 358 process_counter_values(struct perf_stat_config *config, struct evsel *evsel,
 359                        int cpu_map_idx, int thread,
 360                        struct perf_counts_values *count)
 361 {
 362         struct perf_counts_values *aggr = &evsel->counts->aggr;
 363         static struct perf_counts_values zero;
 364         bool skip = false;
 365
 366         if (check_per_pkg(evsel, count, cpu_map_idx, &skip)) {
 367                 pr_err("failed to read per-pkg counter\n");
 368                 return -1;
 369         }
 370
 371         if (skip)
 372                 count = &zero;
 373
 374         switch (config->aggr_mode) {
 375         case AGGR_THREAD:
 376         case AGGR_CORE:
 377         case AGGR_DIE:
 378         case AGGR_SOCKET:
 379         case AGGR_NODE:
 380         case AGGR_NONE:
 381                 if (!evsel->snapshot)
 382                         evsel__compute_deltas(evsel, cpu_map_idx, thread, count);
 383                 perf_counts_values__scale(count, config->scale, NULL);
 384                 if ((config->aggr_mode == AGGR_NONE) && (!evsel->percore)) {
 385                         perf_stat__update_shadow_stats(evsel, count->val,
 386                                                        cpu_map_idx, &rt_stat);
 387                 }
 388
 389                 if (config->aggr_mode == AGGR_THREAD) {
 390                         if (config->stats)
 391                                 perf_stat__update_shadow_stats(evsel,
 392                                         count->val, 0, &config->stats[thread]);
 393                         else
 394                                 perf_stat__update_shadow_stats(evsel,
 395                                         count->val, 0, &rt_stat);
 396                 }
 397                 break;
 398         case AGGR_GLOBAL:
 399                 aggr->val += count->val;
 400                 aggr->ena += count->ena;
 401                 aggr->run += count->run;
 402         case AGGR_UNSET:
 403         default:
 404                 break;
 405         }
 406
 407         return 0;
 408 }
 409
 410 static int process_counter_maps(struct perf_stat_config *config,
 411                                 struct evsel *counter)
 412 {
 413         int nthreads = perf_thread_map__nr(counter->core.threads);
 414         int ncpus = evsel__nr_cpus(counter);
 415         int idx, thread;
 416
 417         if (counter->core.system_wide)
 418                 nthreads = 1;
 419
 420         for (thread = 0; thread < nthreads; thread++) {
 421                 for (idx = 0; idx < ncpus; idx++) {
 422                         if (process_counter_values(config, counter, idx, thread,
 423                                                    perf_counts(counter->counts, idx, thread)))
 424                                 return -1;
 425                 }
 426         }
 427
 428         return 0;
 429 }
 430
 431 int perf_stat_process_counter(struct perf_stat_config *config,
 432                               struct evsel *counter)
 433 {
 434         struct perf_counts_values *aggr = &counter->counts->aggr;
 435         struct perf_stat_evsel *ps = counter->stats;
 436         u64 *count = counter->counts->aggr.values;
 437         int i, ret;
 438
 439         aggr->val = aggr->ena = aggr->run = 0;
 440
 441         if (counter->per_pkg)
 442                 evsel__zero_per_pkg(counter);
 443
 444         ret = process_counter_maps(config, counter);
 445         if (ret)
 446                 return ret;
 447
 448         if (config->aggr_mode != AGGR_GLOBAL)
 449                 return 0;
 450
 451         if (!counter->snapshot)
 452                 evsel__compute_deltas(counter, -1, -1, aggr);
 453         perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled);
 454
 455         for (i = 0; i < 3; i++)
 456                 update_stats(&ps->res_stats[i], count[i]);
 457
 458         if (verbose > 0) {
 459                 fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 460                         evsel__name(counter), count[0], count[1], count[2]);
 461         }
 462
 463         /*
 464          * Save the full runtime - to allow normalization during printout:
 465          */
 466         perf_stat__update_shadow_stats(counter, *count, 0, &rt_stat);
 467
 468         return 0;
 469 }
 470
 471 int perf_event__process_stat_event(struct perf_session *session,
 472                                    union perf_event *event)
 473 {
 474         struct perf_counts_values count;
 475         struct perf_record_stat *st = &event->stat;
 476         struct evsel *counter;
 477
 478         count.val = st->val;
 479         count.ena = st->ena;
 480         count.run = st->run;
 481
 482         counter = evlist__id2evsel(session->evlist, st->id);
 483         if (!counter) {
 484                 pr_err("Failed to resolve counter for stat event.\n");
 485                 return -EINVAL;
 486         }
 487
 488         *perf_counts(counter->counts, st->cpu, st->thread) = count;
 489         counter->supported = true;
 490         return 0;
 491 }
 492
 493 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
 494 {
 495         struct perf_record_stat *st = (struct perf_record_stat *)event;
 496         size_t ret;
 497
 498         ret  = fprintf(fp, "\n... id %" PRI_lu64 ", cpu %d, thread %d\n",
 499                        st->id, st->cpu, st->thread);
 500         ret += fprintf(fp, "... value %" PRI_lu64 ", enabled %" PRI_lu64 ", running %" PRI_lu64 "\n",
 501                        st->val, st->ena, st->run);
 502
 503         return ret;
 504 }
 505
 506 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
 507 {
 508         struct perf_record_stat_round *rd = (struct perf_record_stat_round *)event;
 509         size_t ret;
 510
 511         ret = fprintf(fp, "\n... time %" PRI_lu64 ", type %s\n", rd->time,
 512                       rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
 513
 514         return ret;
 515 }
 516
 517 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
 518 {
 519         struct perf_stat_config sc;
 520         size_t ret;
 521
 522         perf_event__read_stat_config(&sc, &event->stat_config);
 523
 524         ret  = fprintf(fp, "\n");
 525         ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
 526         ret += fprintf(fp, "... scale     %d\n", sc.scale);
 527         ret += fprintf(fp, "... interval  %u\n", sc.interval);
 528
 529         return ret;
 530 }
 531
 532 int create_perf_stat_counter(struct evsel *evsel,
 533                              struct perf_stat_config *config,
 534                              struct target *target,
 535                              int cpu_map_idx)
 536 {
 537         struct perf_event_attr *attr = &evsel->core.attr;
 538         struct evsel *leader = evsel__leader(evsel);
 539
 540         attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 541                             PERF_FORMAT_TOTAL_TIME_RUNNING;
 542
 543         /*
 544          * The event is part of non trivial group, let's enable
 545          * the group read (for leader) and ID retrieval for all
 546          * members.
 547          */
 548         if (leader->core.nr_members > 1)
 549                 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
 550
 551         attr->inherit = !config->no_inherit && list_empty(&evsel->bpf_counter_list);
 552
 553         /*
 554          * Some events get initialized with sample_(period/type) set,
 555          * like tracepoints. Clear it up for counting.
 556          */
 557         attr->sample_period = 0;
 558
 559         if (config->identifier)
 560                 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
 561
 562         if (config->all_user) {
 563                 attr->exclude_kernel = 1;
 564                 attr->exclude_user   = 0;
 565         }
 566
 567         if (config->all_kernel) {
 568                 attr->exclude_kernel = 0;
 569                 attr->exclude_user   = 1;
 570         }
 571
 572         /*
 573          * Disabling all counters initially, they will be enabled
 574          * either manually by us or by kernel via enable_on_exec
 575          * set later.
 576          */
 577         if (evsel__is_group_leader(evsel)) {
 578                 attr->disabled = 1;
 579
 580                 /*
 581                  * In case of initial_delay we enable tracee
 582                  * events manually.
 583                  */
 584                 if (target__none(target) && !config->initial_delay)
 585                         attr->enable_on_exec = 1;
 586         }
 587
 588         if (target__has_cpu(target) && !target__has_per_thread(target))
 589                 return evsel__open_per_cpu(evsel, evsel__cpus(evsel), cpu_map_idx);
 590
 591         return evsel__open_per_thread(evsel, evsel->core.threads);
 592 }