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 <linux/zalloc.h>
  17
  18 void update_stats(struct stats *stats, u64 val)
  19 {
  20         double delta;
  21
  22         stats->n++;
  23         delta = val - stats->mean;
  24         stats->mean += delta / stats->n;
  25         stats->M2 += delta*(val - stats->mean);
  26
  27         if (val > stats->max)
  28                 stats->max = val;
  29
  30         if (val < stats->min)
  31                 stats->min = val;
  32 }
  33
  34 double avg_stats(struct stats *stats)
  35 {
  36         return stats->mean;
  37 }
  38
  39 /*
  40  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
  41  *
  42  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
  43  * s^2 = -------------------------------
  44  *                  n - 1
  45  *
  46  * http://en.wikipedia.org/wiki/Stddev
  47  *
  48  * The std dev of the mean is related to the std dev by:
  49  *
  50  *             s
  51  * s_mean = -------
  52  *          sqrt(n)
  53  *
  54  */
  55 double stddev_stats(struct stats *stats)
  56 {
  57         double variance, variance_mean;
  58
  59         if (stats->n < 2)
  60                 return 0.0;
  61
  62         variance = stats->M2 / (stats->n - 1);
  63         variance_mean = variance / stats->n;
  64
  65         return sqrt(variance_mean);
  66 }
  67
  68 double rel_stddev_stats(double stddev, double avg)
  69 {
  70         double pct = 0.0;
  71
  72         if (avg)
  73                 pct = 100.0 * stddev/avg;
  74
  75         return pct;
  76 }
  77
  78 bool __perf_evsel_stat__is(struct evsel *evsel,
  79                            enum perf_stat_evsel_id id)
  80 {
  81         struct perf_stat_evsel *ps = evsel->stats;
  82
  83         return ps->id == id;
  84 }
  85
  86 #define ID(id, name) [PERF_STAT_EVSEL_ID__##id] = #name
  87 static const char *id_str[PERF_STAT_EVSEL_ID__MAX] = {
  88         ID(NONE,                x),
  89         ID(CYCLES_IN_TX,        cpu/cycles-t/),
  90         ID(TRANSACTION_START,   cpu/tx-start/),
  91         ID(ELISION_START,       cpu/el-start/),
  92         ID(CYCLES_IN_TX_CP,     cpu/cycles-ct/),
  93         ID(TOPDOWN_TOTAL_SLOTS, topdown-total-slots),
  94         ID(TOPDOWN_SLOTS_ISSUED, topdown-slots-issued),
  95         ID(TOPDOWN_SLOTS_RETIRED, topdown-slots-retired),
  96         ID(TOPDOWN_FETCH_BUBBLES, topdown-fetch-bubbles),
  97         ID(TOPDOWN_RECOVERY_BUBBLES, topdown-recovery-bubbles),
  98         ID(SMI_NUM, msr/smi/),
  99         ID(APERF, msr/aperf/),
 100 };
 101 #undef ID
 102
 103 static void perf_stat_evsel_id_init(struct evsel *evsel)
 104 {
 105         struct perf_stat_evsel *ps = evsel->stats;
 106         int i;
 107
 108         /* ps->id is 0 hence PERF_STAT_EVSEL_ID__NONE by default */
 109
 110         for (i = 0; i < PERF_STAT_EVSEL_ID__MAX; i++) {
 111                 if (!strcmp(evsel__name(evsel), id_str[i])) {
 112                         ps->id = i;
 113                         break;
 114                 }
 115         }
 116 }
 117
 118 static void evsel__reset_stat_priv(struct evsel *evsel)
 119 {
 120         int i;
 121         struct perf_stat_evsel *ps = evsel->stats;
 122
 123         for (i = 0; i < 3; i++)
 124                 init_stats(&ps->res_stats[i]);
 125
 126         perf_stat_evsel_id_init(evsel);
 127 }
 128
 129 static int evsel__alloc_stat_priv(struct evsel *evsel)
 130 {
 131         evsel->stats = zalloc(sizeof(struct perf_stat_evsel));
 132         if (evsel->stats == NULL)
 133                 return -ENOMEM;
 134         evsel__reset_stat_priv(evsel);
 135         return 0;
 136 }
 137
 138 static void evsel__free_stat_priv(struct evsel *evsel)
 139 {
 140         struct perf_stat_evsel *ps = evsel->stats;
 141
 142         if (ps)
 143                 zfree(&ps->group_data);
 144         zfree(&evsel->stats);
 145 }
 146
 147 static int evsel__alloc_prev_raw_counts(struct evsel *evsel, int ncpus, int nthreads)
 148 {
 149         struct perf_counts *counts;
 150
 151         counts = perf_counts__new(ncpus, nthreads);
 152         if (counts)
 153                 evsel->prev_raw_counts = counts;
 154
 155         return counts ? 0 : -ENOMEM;
 156 }
 157
 158 static void evsel__free_prev_raw_counts(struct evsel *evsel)
 159 {
 160         perf_counts__delete(evsel->prev_raw_counts);
 161         evsel->prev_raw_counts = NULL;
 162 }
 163
 164 static void evsel__reset_prev_raw_counts(struct evsel *evsel)
 165 {
 166         if (evsel->prev_raw_counts)
 167                 perf_counts__reset(evsel->prev_raw_counts);
 168 }
 169
 170 static int evsel__alloc_stats(struct evsel *evsel, bool alloc_raw)
 171 {
 172         int ncpus = evsel__nr_cpus(evsel);
 173         int nthreads = perf_thread_map__nr(evsel->core.threads);
 174
 175         if (evsel__alloc_stat_priv(evsel) < 0 ||
 176             evsel__alloc_counts(evsel, ncpus, nthreads) < 0 ||
 177             (alloc_raw && evsel__alloc_prev_raw_counts(evsel, ncpus, nthreads) < 0))
 178                 return -ENOMEM;
 179
 180         return 0;
 181 }
 182
 183 int perf_evlist__alloc_stats(struct evlist *evlist, bool alloc_raw)
 184 {
 185         struct evsel *evsel;
 186
 187         evlist__for_each_entry(evlist, evsel) {
 188                 if (evsel__alloc_stats(evsel, alloc_raw))
 189                         goto out_free;
 190         }
 191
 192         return 0;
 193
 194 out_free:
 195         perf_evlist__free_stats(evlist);
 196         return -1;
 197 }
 198
 199 void perf_evlist__free_stats(struct evlist *evlist)
 200 {
 201         struct evsel *evsel;
 202
 203         evlist__for_each_entry(evlist, evsel) {
 204                 evsel__free_stat_priv(evsel);
 205                 evsel__free_counts(evsel);
 206                 evsel__free_prev_raw_counts(evsel);
 207         }
 208 }
 209
 210 void perf_evlist__reset_stats(struct evlist *evlist)
 211 {
 212         struct evsel *evsel;
 213
 214         evlist__for_each_entry(evlist, evsel) {
 215                 evsel__reset_stat_priv(evsel);
 216                 evsel__reset_counts(evsel);
 217         }
 218 }
 219
 220 void perf_evlist__reset_prev_raw_counts(struct evlist *evlist)
 221 {
 222         struct evsel *evsel;
 223
 224         evlist__for_each_entry(evlist, evsel)
 225                 evsel__reset_prev_raw_counts(evsel);
 226 }
 227
 228 static void perf_evsel__copy_prev_raw_counts(struct evsel *evsel)
 229 {
 230         int ncpus = evsel__nr_cpus(evsel);
 231         int nthreads = perf_thread_map__nr(evsel->core.threads);
 232
 233         for (int thread = 0; thread < nthreads; thread++) {
 234                 for (int cpu = 0; cpu < ncpus; cpu++) {
 235                         *perf_counts(evsel->counts, cpu, thread) =
 236                                 *perf_counts(evsel->prev_raw_counts, cpu,
 237                                              thread);
 238                 }
 239         }
 240
 241         evsel->counts->aggr = evsel->prev_raw_counts->aggr;
 242 }
 243
 244 void perf_evlist__copy_prev_raw_counts(struct evlist *evlist)
 245 {
 246         struct evsel *evsel;
 247
 248         evlist__for_each_entry(evlist, evsel)
 249                 perf_evsel__copy_prev_raw_counts(evsel);
 250 }
 251
 252 void perf_evlist__save_aggr_prev_raw_counts(struct evlist *evlist)
 253 {
 254         struct evsel *evsel;
 255
 256         /*
 257          * To collect the overall statistics for interval mode,
 258          * we copy the counts from evsel->prev_raw_counts to
 259          * evsel->counts. The perf_stat_process_counter creates
 260          * aggr values from per cpu values, but the per cpu values
 261          * are 0 for AGGR_GLOBAL. So we use a trick that saves the
 262          * previous aggr value to the first member of perf_counts,
 263          * then aggr calculation in process_counter_values can work
 264          * correctly.
 265          */
 266         evlist__for_each_entry(evlist, evsel) {
 267                 *perf_counts(evsel->prev_raw_counts, 0, 0) =
 268                         evsel->prev_raw_counts->aggr;
 269         }
 270 }
 271
 272 static void zero_per_pkg(struct evsel *counter)
 273 {
 274         if (counter->per_pkg_mask)
 275                 memset(counter->per_pkg_mask, 0, cpu__max_cpu());
 276 }
 277
 278 static int check_per_pkg(struct evsel *counter,
 279                          struct perf_counts_values *vals, int cpu, bool *skip)
 280 {
 281         unsigned long *mask = counter->per_pkg_mask;
 282         struct perf_cpu_map *cpus = evsel__cpus(counter);
 283         int s;
 284
 285         *skip = false;
 286
 287         if (!counter->per_pkg)
 288                 return 0;
 289
 290         if (perf_cpu_map__empty(cpus))
 291                 return 0;
 292
 293         if (!mask) {
 294                 mask = zalloc(cpu__max_cpu());
 295                 if (!mask)
 296                         return -ENOMEM;
 297
 298                 counter->per_pkg_mask = mask;
 299         }
 300
 301         /*
 302          * we do not consider an event that has not run as a good
 303          * instance to mark a package as used (skip=1). Otherwise
 304          * we may run into a situation where the first CPU in a package
 305          * is not running anything, yet the second is, and this function
 306          * would mark the package as used after the first CPU and would
 307          * not read the values from the second CPU.
 308          */
 309         if (!(vals->run && vals->ena))
 310                 return 0;
 311
 312         s = cpu_map__get_socket(cpus, cpu, NULL);
 313         if (s < 0)
 314                 return -1;
 315
 316         *skip = test_and_set_bit(s, mask) == 1;
 317         return 0;
 318 }
 319
 320 static int
 321 process_counter_values(struct perf_stat_config *config, struct evsel *evsel,
 322                        int cpu, int thread,
 323                        struct perf_counts_values *count)
 324 {
 325         struct perf_counts_values *aggr = &evsel->counts->aggr;
 326         static struct perf_counts_values zero;
 327         bool skip = false;
 328
 329         if (check_per_pkg(evsel, count, cpu, &skip)) {
 330                 pr_err("failed to read per-pkg counter\n");
 331                 return -1;
 332         }
 333
 334         if (skip)
 335                 count = &zero;
 336
 337         switch (config->aggr_mode) {
 338         case AGGR_THREAD:
 339         case AGGR_CORE:
 340         case AGGR_DIE:
 341         case AGGR_SOCKET:
 342         case AGGR_NODE:
 343         case AGGR_NONE:
 344                 if (!evsel->snapshot)
 345                         evsel__compute_deltas(evsel, cpu, thread, count);
 346                 perf_counts_values__scale(count, config->scale, NULL);
 347                 if ((config->aggr_mode == AGGR_NONE) && (!evsel->percore)) {
 348                         perf_stat__update_shadow_stats(evsel, count->val,
 349                                                        cpu, &rt_stat);
 350                 }
 351
 352                 if (config->aggr_mode == AGGR_THREAD) {
 353                         if (config->stats)
 354                                 perf_stat__update_shadow_stats(evsel,
 355                                         count->val, 0, &config->stats[thread]);
 356                         else
 357                                 perf_stat__update_shadow_stats(evsel,
 358                                         count->val, 0, &rt_stat);
 359                 }
 360                 break;
 361         case AGGR_GLOBAL:
 362                 aggr->val += count->val;
 363                 aggr->ena += count->ena;
 364                 aggr->run += count->run;
 365         case AGGR_UNSET:
 366         default:
 367                 break;
 368         }
 369
 370         return 0;
 371 }
 372
 373 static int process_counter_maps(struct perf_stat_config *config,
 374                                 struct evsel *counter)
 375 {
 376         int nthreads = perf_thread_map__nr(counter->core.threads);
 377         int ncpus = evsel__nr_cpus(counter);
 378         int cpu, thread;
 379
 380         if (counter->core.system_wide)
 381                 nthreads = 1;
 382
 383         for (thread = 0; thread < nthreads; thread++) {
 384                 for (cpu = 0; cpu < ncpus; cpu++) {
 385                         if (process_counter_values(config, counter, cpu, thread,
 386                                                    perf_counts(counter->counts, cpu, thread)))
 387                                 return -1;
 388                 }
 389         }
 390
 391         return 0;
 392 }
 393
 394 int perf_stat_process_counter(struct perf_stat_config *config,
 395                               struct evsel *counter)
 396 {
 397         struct perf_counts_values *aggr = &counter->counts->aggr;
 398         struct perf_stat_evsel *ps = counter->stats;
 399         u64 *count = counter->counts->aggr.values;
 400         int i, ret;
 401
 402         aggr->val = aggr->ena = aggr->run = 0;
 403
 404         /*
 405          * We calculate counter's data every interval,
 406          * and the display code shows ps->res_stats
 407          * avg value. We need to zero the stats for
 408          * interval mode, otherwise overall avg running
 409          * averages will be shown for each interval.
 410          */
 411         if (config->interval || config->summary) {
 412                 for (i = 0; i < 3; i++)
 413                         init_stats(&ps->res_stats[i]);
 414         }
 415
 416         if (counter->per_pkg)
 417                 zero_per_pkg(counter);
 418
 419         ret = process_counter_maps(config, counter);
 420         if (ret)
 421                 return ret;
 422
 423         if (config->aggr_mode != AGGR_GLOBAL)
 424                 return 0;
 425
 426         if (!counter->snapshot)
 427                 evsel__compute_deltas(counter, -1, -1, aggr);
 428         perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled);
 429
 430         for (i = 0; i < 3; i++)
 431                 update_stats(&ps->res_stats[i], count[i]);
 432
 433         if (verbose > 0) {
 434                 fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 435                         evsel__name(counter), count[0], count[1], count[2]);
 436         }
 437
 438         /*
 439          * Save the full runtime - to allow normalization during printout:
 440          */
 441         perf_stat__update_shadow_stats(counter, *count, 0, &rt_stat);
 442
 443         return 0;
 444 }
 445
 446 int perf_event__process_stat_event(struct perf_session *session,
 447                                    union perf_event *event)
 448 {
 449         struct perf_counts_values count;
 450         struct perf_record_stat *st = &event->stat;
 451         struct evsel *counter;
 452
 453         count.val = st->val;
 454         count.ena = st->ena;
 455         count.run = st->run;
 456
 457         counter = perf_evlist__id2evsel(session->evlist, st->id);
 458         if (!counter) {
 459                 pr_err("Failed to resolve counter for stat event.\n");
 460                 return -EINVAL;
 461         }
 462
 463         *perf_counts(counter->counts, st->cpu, st->thread) = count;
 464         counter->supported = true;
 465         return 0;
 466 }
 467
 468 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
 469 {
 470         struct perf_record_stat *st = (struct perf_record_stat *)event;
 471         size_t ret;
 472
 473         ret  = fprintf(fp, "\n... id %" PRI_lu64 ", cpu %d, thread %d\n",
 474                        st->id, st->cpu, st->thread);
 475         ret += fprintf(fp, "... value %" PRI_lu64 ", enabled %" PRI_lu64 ", running %" PRI_lu64 "\n",
 476                        st->val, st->ena, st->run);
 477
 478         return ret;
 479 }
 480
 481 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
 482 {
 483         struct perf_record_stat_round *rd = (struct perf_record_stat_round *)event;
 484         size_t ret;
 485
 486         ret = fprintf(fp, "\n... time %" PRI_lu64 ", type %s\n", rd->time,
 487                       rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
 488
 489         return ret;
 490 }
 491
 492 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
 493 {
 494         struct perf_stat_config sc;
 495         size_t ret;
 496
 497         perf_event__read_stat_config(&sc, &event->stat_config);
 498
 499         ret  = fprintf(fp, "\n");
 500         ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
 501         ret += fprintf(fp, "... scale     %d\n", sc.scale);
 502         ret += fprintf(fp, "... interval  %u\n", sc.interval);
 503
 504         return ret;
 505 }
 506
 507 int create_perf_stat_counter(struct evsel *evsel,
 508                              struct perf_stat_config *config,
 509                              struct target *target,
 510                              int cpu)
 511 {
 512         struct perf_event_attr *attr = &evsel->core.attr;
 513         struct evsel *leader = evsel->leader;
 514
 515         attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 516                             PERF_FORMAT_TOTAL_TIME_RUNNING;
 517
 518         /*
 519          * The event is part of non trivial group, let's enable
 520          * the group read (for leader) and ID retrieval for all
 521          * members.
 522          */
 523         if (leader->core.nr_members > 1)
 524                 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
 525
 526         attr->inherit = !config->no_inherit;
 527
 528         /*
 529          * Some events get initialized with sample_(period/type) set,
 530          * like tracepoints. Clear it up for counting.
 531          */
 532         attr->sample_period = 0;
 533
 534         if (config->identifier)
 535                 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
 536
 537         if (config->all_user) {
 538                 attr->exclude_kernel = 1;
 539                 attr->exclude_user   = 0;
 540         }
 541
 542         if (config->all_kernel) {
 543                 attr->exclude_kernel = 0;
 544                 attr->exclude_user   = 1;
 545         }
 546
 547         /*
 548          * Disabling all counters initially, they will be enabled
 549          * either manually by us or by kernel via enable_on_exec
 550          * set later.
 551          */
 552         if (evsel__is_group_leader(evsel)) {
 553                 attr->disabled = 1;
 554
 555                 /*
 556                  * In case of initial_delay we enable tracee
 557                  * events manually.
 558                  */
 559                 if (target__none(target) && !config->initial_delay)
 560                         attr->enable_on_exec = 1;
 561         }
 562
 563         if (target__has_cpu(target) && !target__has_per_thread(target))
 564                 return evsel__open_per_cpu(evsel, evsel__cpus(evsel), cpu);
 565
 566         return evsel__open_per_thread(evsel, evsel->core.threads);
 567 }