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 "debug.h"
   8 #include "header.h"
   9 #include "stat.h"
  10 #include "session.h"
  11 #include "target.h"
  12 #include "evlist.h"
  13 #include "evsel.h"
  14 #include "thread_map.h"
  15 #include <linux/zalloc.h>
  16
  17 void update_stats(struct stats *stats, u64 val)
  18 {
  19         double delta;
  20
  21         stats->n++;
  22         delta = val - stats->mean;
  23         stats->mean += delta / stats->n;
  24         stats->M2 += delta*(val - stats->mean);
  25
  26         if (val > stats->max)
  27                 stats->max = val;
  28
  29         if (val < stats->min)
  30                 stats->min = val;
  31 }
  32
  33 double avg_stats(struct stats *stats)
  34 {
  35         return stats->mean;
  36 }
  37
  38 /*
  39  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
  40  *
  41  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
  42  * s^2 = -------------------------------
  43  *                  n - 1
  44  *
  45  * http://en.wikipedia.org/wiki/Stddev
  46  *
  47  * The std dev of the mean is related to the std dev by:
  48  *
  49  *             s
  50  * s_mean = -------
  51  *          sqrt(n)
  52  *
  53  */
  54 double stddev_stats(struct stats *stats)
  55 {
  56         double variance, variance_mean;
  57
  58         if (stats->n < 2)
  59                 return 0.0;
  60
  61         variance = stats->M2 / (stats->n - 1);
  62         variance_mean = variance / stats->n;
  63
  64         return sqrt(variance_mean);
  65 }
  66
  67 double rel_stddev_stats(double stddev, double avg)
  68 {
  69         double pct = 0.0;
  70
  71         if (avg)
  72                 pct = 100.0 * stddev/avg;
  73
  74         return pct;
  75 }
  76
  77 bool __perf_evsel_stat__is(struct evsel *evsel,
  78                            enum perf_stat_evsel_id id)
  79 {
  80         struct perf_stat_evsel *ps = evsel->stats;
  81
  82         return ps->id == id;
  83 }
  84
  85 #define ID(id, name) [PERF_STAT_EVSEL_ID__##id] = #name
  86 static const char *id_str[PERF_STAT_EVSEL_ID__MAX] = {
  87         ID(NONE,                x),
  88         ID(CYCLES_IN_TX,        cpu/cycles-t/),
  89         ID(TRANSACTION_START,   cpu/tx-start/),
  90         ID(ELISION_START,       cpu/el-start/),
  91         ID(CYCLES_IN_TX_CP,     cpu/cycles-ct/),
  92         ID(TOPDOWN_TOTAL_SLOTS, topdown-total-slots),
  93         ID(TOPDOWN_SLOTS_ISSUED, topdown-slots-issued),
  94         ID(TOPDOWN_SLOTS_RETIRED, topdown-slots-retired),
  95         ID(TOPDOWN_FETCH_BUBBLES, topdown-fetch-bubbles),
  96         ID(TOPDOWN_RECOVERY_BUBBLES, topdown-recovery-bubbles),
  97         ID(SMI_NUM, msr/smi/),
  98         ID(APERF, msr/aperf/),
  99 };
 100 #undef ID
 101
 102 static void perf_stat_evsel_id_init(struct evsel *evsel)
 103 {
 104         struct perf_stat_evsel *ps = evsel->stats;
 105         int i;
 106
 107         /* ps->id is 0 hence PERF_STAT_EVSEL_ID__NONE by default */
 108
 109         for (i = 0; i < PERF_STAT_EVSEL_ID__MAX; i++) {
 110                 if (!strcmp(perf_evsel__name(evsel), id_str[i])) {
 111                         ps->id = i;
 112                         break;
 113                 }
 114         }
 115 }
 116
 117 static void perf_evsel__reset_stat_priv(struct evsel *evsel)
 118 {
 119         int i;
 120         struct perf_stat_evsel *ps = evsel->stats;
 121
 122         for (i = 0; i < 3; i++)
 123                 init_stats(&ps->res_stats[i]);
 124
 125         perf_stat_evsel_id_init(evsel);
 126 }
 127
 128 static int perf_evsel__alloc_stat_priv(struct evsel *evsel)
 129 {
 130         evsel->stats = zalloc(sizeof(struct perf_stat_evsel));
 131         if (evsel->stats == NULL)
 132                 return -ENOMEM;
 133         perf_evsel__reset_stat_priv(evsel);
 134         return 0;
 135 }
 136
 137 static void perf_evsel__free_stat_priv(struct evsel *evsel)
 138 {
 139         struct perf_stat_evsel *ps = evsel->stats;
 140
 141         if (ps)
 142                 zfree(&ps->group_data);
 143         zfree(&evsel->stats);
 144 }
 145
 146 static int perf_evsel__alloc_prev_raw_counts(struct evsel *evsel,
 147                                              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 perf_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 int perf_evsel__alloc_stats(struct evsel *evsel, bool alloc_raw)
 165 {
 166         int ncpus = perf_evsel__nr_cpus(evsel);
 167         int nthreads = perf_thread_map__nr(evsel->core.threads);
 168
 169         if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
 170             perf_evsel__alloc_counts(evsel, ncpus, nthreads) < 0 ||
 171             (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel, ncpus, nthreads) < 0))
 172                 return -ENOMEM;
 173
 174         return 0;
 175 }
 176
 177 int perf_evlist__alloc_stats(struct evlist *evlist, bool alloc_raw)
 178 {
 179         struct evsel *evsel;
 180
 181         evlist__for_each_entry(evlist, evsel) {
 182                 if (perf_evsel__alloc_stats(evsel, alloc_raw))
 183                         goto out_free;
 184         }
 185
 186         return 0;
 187
 188 out_free:
 189         perf_evlist__free_stats(evlist);
 190         return -1;
 191 }
 192
 193 void perf_evlist__free_stats(struct evlist *evlist)
 194 {
 195         struct evsel *evsel;
 196
 197         evlist__for_each_entry(evlist, evsel) {
 198                 perf_evsel__free_stat_priv(evsel);
 199                 perf_evsel__free_counts(evsel);
 200                 perf_evsel__free_prev_raw_counts(evsel);
 201         }
 202 }
 203
 204 void perf_evlist__reset_stats(struct evlist *evlist)
 205 {
 206         struct evsel *evsel;
 207
 208         evlist__for_each_entry(evlist, evsel) {
 209                 perf_evsel__reset_stat_priv(evsel);
 210                 perf_evsel__reset_counts(evsel);
 211         }
 212 }
 213
 214 static void zero_per_pkg(struct evsel *counter)
 215 {
 216         if (counter->per_pkg_mask)
 217                 memset(counter->per_pkg_mask, 0, cpu__max_cpu());
 218 }
 219
 220 static int check_per_pkg(struct evsel *counter,
 221                          struct perf_counts_values *vals, int cpu, bool *skip)
 222 {
 223         unsigned long *mask = counter->per_pkg_mask;
 224         struct perf_cpu_map *cpus = evsel__cpus(counter);
 225         int s;
 226
 227         *skip = false;
 228
 229         if (!counter->per_pkg)
 230                 return 0;
 231
 232         if (perf_cpu_map__empty(cpus))
 233                 return 0;
 234
 235         if (!mask) {
 236                 mask = zalloc(cpu__max_cpu());
 237                 if (!mask)
 238                         return -ENOMEM;
 239
 240                 counter->per_pkg_mask = mask;
 241         }
 242
 243         /*
 244          * we do not consider an event that has not run as a good
 245          * instance to mark a package as used (skip=1). Otherwise
 246          * we may run into a situation where the first CPU in a package
 247          * is not running anything, yet the second is, and this function
 248          * would mark the package as used after the first CPU and would
 249          * not read the values from the second CPU.
 250          */
 251         if (!(vals->run && vals->ena))
 252                 return 0;
 253
 254         s = cpu_map__get_socket(cpus, cpu, NULL);
 255         if (s < 0)
 256                 return -1;
 257
 258         *skip = test_and_set_bit(s, mask) == 1;
 259         return 0;
 260 }
 261
 262 static int
 263 process_counter_values(struct perf_stat_config *config, struct evsel *evsel,
 264                        int cpu, int thread,
 265                        struct perf_counts_values *count)
 266 {
 267         struct perf_counts_values *aggr = &evsel->counts->aggr;
 268         static struct perf_counts_values zero;
 269         bool skip = false;
 270
 271         if (check_per_pkg(evsel, count, cpu, &skip)) {
 272                 pr_err("failed to read per-pkg counter\n");
 273                 return -1;
 274         }
 275
 276         if (skip)
 277                 count = &zero;
 278
 279         switch (config->aggr_mode) {
 280         case AGGR_THREAD:
 281         case AGGR_CORE:
 282         case AGGR_DIE:
 283         case AGGR_SOCKET:
 284         case AGGR_NONE:
 285                 if (!evsel->snapshot)
 286                         perf_evsel__compute_deltas(evsel, cpu, thread, count);
 287                 perf_counts_values__scale(count, config->scale, NULL);
 288                 if ((config->aggr_mode == AGGR_NONE) && (!evsel->percore)) {
 289                         perf_stat__update_shadow_stats(evsel, count->val,
 290                                                        cpu, &rt_stat);
 291                 }
 292
 293                 if (config->aggr_mode == AGGR_THREAD) {
 294                         if (config->stats)
 295                                 perf_stat__update_shadow_stats(evsel,
 296                                         count->val, 0, &config->stats[thread]);
 297                         else
 298                                 perf_stat__update_shadow_stats(evsel,
 299                                         count->val, 0, &rt_stat);
 300                 }
 301                 break;
 302         case AGGR_GLOBAL:
 303                 aggr->val += count->val;
 304                 aggr->ena += count->ena;
 305                 aggr->run += count->run;
 306         case AGGR_UNSET:
 307         default:
 308                 break;
 309         }
 310
 311         return 0;
 312 }
 313
 314 static int process_counter_maps(struct perf_stat_config *config,
 315                                 struct evsel *counter)
 316 {
 317         int nthreads = perf_thread_map__nr(counter->core.threads);
 318         int ncpus = perf_evsel__nr_cpus(counter);
 319         int cpu, thread;
 320
 321         if (counter->system_wide)
 322                 nthreads = 1;
 323
 324         for (thread = 0; thread < nthreads; thread++) {
 325                 for (cpu = 0; cpu < ncpus; cpu++) {
 326                         if (process_counter_values(config, counter, cpu, thread,
 327                                                    perf_counts(counter->counts, cpu, thread)))
 328                                 return -1;
 329                 }
 330         }
 331
 332         return 0;
 333 }
 334
 335 int perf_stat_process_counter(struct perf_stat_config *config,
 336                               struct evsel *counter)
 337 {
 338         struct perf_counts_values *aggr = &counter->counts->aggr;
 339         struct perf_stat_evsel *ps = counter->stats;
 340         u64 *count = counter->counts->aggr.values;
 341         int i, ret;
 342
 343         aggr->val = aggr->ena = aggr->run = 0;
 344
 345         /*
 346          * We calculate counter's data every interval,
 347          * and the display code shows ps->res_stats
 348          * avg value. We need to zero the stats for
 349          * interval mode, otherwise overall avg running
 350          * averages will be shown for each interval.
 351          */
 352         if (config->interval)
 353                 init_stats(ps->res_stats);
 354
 355         if (counter->per_pkg)
 356                 zero_per_pkg(counter);
 357
 358         ret = process_counter_maps(config, counter);
 359         if (ret)
 360                 return ret;
 361
 362         if (config->aggr_mode != AGGR_GLOBAL)
 363                 return 0;
 364
 365         if (!counter->snapshot)
 366                 perf_evsel__compute_deltas(counter, -1, -1, aggr);
 367         perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled);
 368
 369         for (i = 0; i < 3; i++)
 370                 update_stats(&ps->res_stats[i], count[i]);
 371
 372         if (verbose > 0) {
 373                 fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 374                         perf_evsel__name(counter), count[0], count[1], count[2]);
 375         }
 376
 377         /*
 378          * Save the full runtime - to allow normalization during printout:
 379          */
 380         perf_stat__update_shadow_stats(counter, *count, 0, &rt_stat);
 381
 382         return 0;
 383 }
 384
 385 int perf_event__process_stat_event(struct perf_session *session,
 386                                    union perf_event *event)
 387 {
 388         struct perf_counts_values count;
 389         struct perf_record_stat *st = &event->stat;
 390         struct evsel *counter;
 391
 392         count.val = st->val;
 393         count.ena = st->ena;
 394         count.run = st->run;
 395
 396         counter = perf_evlist__id2evsel(session->evlist, st->id);
 397         if (!counter) {
 398                 pr_err("Failed to resolve counter for stat event.\n");
 399                 return -EINVAL;
 400         }
 401
 402         *perf_counts(counter->counts, st->cpu, st->thread) = count;
 403         counter->supported = true;
 404         return 0;
 405 }
 406
 407 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
 408 {
 409         struct perf_record_stat *st = (struct perf_record_stat *)event;
 410         size_t ret;
 411
 412         ret  = fprintf(fp, "\n... id %" PRI_lu64 ", cpu %d, thread %d\n",
 413                        st->id, st->cpu, st->thread);
 414         ret += fprintf(fp, "... value %" PRI_lu64 ", enabled %" PRI_lu64 ", running %" PRI_lu64 "\n",
 415                        st->val, st->ena, st->run);
 416
 417         return ret;
 418 }
 419
 420 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
 421 {
 422         struct perf_record_stat_round *rd = (struct perf_record_stat_round *)event;
 423         size_t ret;
 424
 425         ret = fprintf(fp, "\n... time %" PRI_lu64 ", type %s\n", rd->time,
 426                       rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
 427
 428         return ret;
 429 }
 430
 431 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
 432 {
 433         struct perf_stat_config sc;
 434         size_t ret;
 435
 436         perf_event__read_stat_config(&sc, &event->stat_config);
 437
 438         ret  = fprintf(fp, "\n");
 439         ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
 440         ret += fprintf(fp, "... scale     %d\n", sc.scale);
 441         ret += fprintf(fp, "... interval  %u\n", sc.interval);
 442
 443         return ret;
 444 }
 445
 446 int create_perf_stat_counter(struct evsel *evsel,
 447                              struct perf_stat_config *config,
 448                              struct target *target)
 449 {
 450         struct perf_event_attr *attr = &evsel->core.attr;
 451         struct evsel *leader = evsel->leader;
 452
 453         attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 454                             PERF_FORMAT_TOTAL_TIME_RUNNING;
 455
 456         /*
 457          * The event is part of non trivial group, let's enable
 458          * the group read (for leader) and ID retrieval for all
 459          * members.
 460          */
 461         if (leader->core.nr_members > 1)
 462                 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
 463
 464         attr->inherit = !config->no_inherit;
 465
 466         /*
 467          * Some events get initialized with sample_(period/type) set,
 468          * like tracepoints. Clear it up for counting.
 469          */
 470         attr->sample_period = 0;
 471
 472         if (config->identifier)
 473                 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
 474
 475         /*
 476          * Disabling all counters initially, they will be enabled
 477          * either manually by us or by kernel via enable_on_exec
 478          * set later.
 479          */
 480         if (perf_evsel__is_group_leader(evsel)) {
 481                 attr->disabled = 1;
 482
 483                 /*
 484                  * In case of initial_delay we enable tracee
 485                  * events manually.
 486                  */
 487                 if (target__none(target) && !config->initial_delay)
 488                         attr->enable_on_exec = 1;
 489         }
 490
 491         if (target__has_cpu(target) && !target__has_per_thread(target))
 492                 return perf_evsel__open_per_cpu(evsel, evsel__cpus(evsel));
 493
 494         return perf_evsel__open_per_thread(evsel, evsel->core.threads);
 495 }
 496
 497 int perf_stat_synthesize_config(struct perf_stat_config *config,
 498                                 struct perf_tool *tool,
 499                                 struct evlist *evlist,
 500                                 perf_event__handler_t process,
 501                                 bool attrs)
 502 {
 503         int err;
 504
 505         if (attrs) {
 506                 err = perf_event__synthesize_attrs(tool, evlist, process);
 507                 if (err < 0) {
 508                         pr_err("Couldn't synthesize attrs.\n");
 509                         return err;
 510                 }
 511         }
 512
 513         err = perf_event__synthesize_extra_attr(tool, evlist, process,
 514                                                 attrs);
 515
 516         err = perf_event__synthesize_thread_map2(tool, evlist->core.threads,
 517                                                  process, NULL);
 518         if (err < 0) {
 519                 pr_err("Couldn't synthesize thread map.\n");
 520                 return err;
 521         }
 522
 523         err = perf_event__synthesize_cpu_map(tool, evlist->core.cpus,
 524                                              process, NULL);
 525         if (err < 0) {
 526                 pr_err("Couldn't synthesize thread map.\n");
 527                 return err;
 528         }
 529
 530         err = perf_event__synthesize_stat_config(tool, config, process, NULL);
 531         if (err < 0) {
 532                 pr_err("Couldn't synthesize config.\n");
 533                 return err;
 534         }
 535
 536         return 0;
 537 }