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