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(perf_evsel__name(evsel), id_str[i])) {
 112                         ps->id = i;
 113                         break;
 114                 }
 115         }
 116 }
 117
 118 static void perf_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 perf_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         perf_evsel__reset_stat_priv(evsel);
 135         return 0;
 136 }
 137
 138 static void perf_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 perf_evsel__alloc_prev_raw_counts(struct evsel *evsel,
 148                                              int ncpus, int nthreads)
 149 {
 150         struct perf_counts *counts;
 151
 152         counts = perf_counts__new(ncpus, nthreads);
 153         if (counts)
 154                 evsel->prev_raw_counts = counts;
 155
 156         return counts ? 0 : -ENOMEM;
 157 }
 158
 159 static void perf_evsel__free_prev_raw_counts(struct evsel *evsel)
 160 {
 161         perf_counts__delete(evsel->prev_raw_counts);
 162         evsel->prev_raw_counts = NULL;
 163 }
 164
 165 static void perf_evsel__reset_prev_raw_counts(struct evsel *evsel)
 166 {
 167         if (evsel->prev_raw_counts) {
 168                 evsel->prev_raw_counts->aggr.val = 0;
 169                 evsel->prev_raw_counts->aggr.ena = 0;
 170                 evsel->prev_raw_counts->aggr.run = 0;
 171        }
 172 }
 173
 174 static int perf_evsel__alloc_stats(struct evsel *evsel, bool alloc_raw)
 175 {
 176         int ncpus = perf_evsel__nr_cpus(evsel);
 177         int nthreads = perf_thread_map__nr(evsel->core.threads);
 178
 179         if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
 180             perf_evsel__alloc_counts(evsel, ncpus, nthreads) < 0 ||
 181             (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel, ncpus, nthreads) < 0))
 182                 return -ENOMEM;
 183
 184         return 0;
 185 }
 186
 187 int perf_evlist__alloc_stats(struct evlist *evlist, bool alloc_raw)
 188 {
 189         struct evsel *evsel;
 190
 191         evlist__for_each_entry(evlist, evsel) {
 192                 if (perf_evsel__alloc_stats(evsel, alloc_raw))
 193                         goto out_free;
 194         }
 195
 196         return 0;
 197
 198 out_free:
 199         perf_evlist__free_stats(evlist);
 200         return -1;
 201 }
 202
 203 void perf_evlist__free_stats(struct evlist *evlist)
 204 {
 205         struct evsel *evsel;
 206
 207         evlist__for_each_entry(evlist, evsel) {
 208                 perf_evsel__free_stat_priv(evsel);
 209                 perf_evsel__free_counts(evsel);
 210                 perf_evsel__free_prev_raw_counts(evsel);
 211         }
 212 }
 213
 214 void perf_evlist__reset_stats(struct evlist *evlist)
 215 {
 216         struct evsel *evsel;
 217
 218         evlist__for_each_entry(evlist, evsel) {
 219                 perf_evsel__reset_stat_priv(evsel);
 220                 perf_evsel__reset_counts(evsel);
 221         }
 222 }
 223
 224 void perf_evlist__reset_prev_raw_counts(struct evlist *evlist)
 225 {
 226         struct evsel *evsel;
 227
 228         evlist__for_each_entry(evlist, evsel)
 229                 perf_evsel__reset_prev_raw_counts(evsel);
 230 }
 231
 232 static void zero_per_pkg(struct evsel *counter)
 233 {
 234         if (counter->per_pkg_mask)
 235                 memset(counter->per_pkg_mask, 0, cpu__max_cpu());
 236 }
 237
 238 static int check_per_pkg(struct evsel *counter,
 239                          struct perf_counts_values *vals, int cpu, bool *skip)
 240 {
 241         unsigned long *mask = counter->per_pkg_mask;
 242         struct perf_cpu_map *cpus = evsel__cpus(counter);
 243         int s;
 244
 245         *skip = false;
 246
 247         if (!counter->per_pkg)
 248                 return 0;
 249
 250         if (perf_cpu_map__empty(cpus))
 251                 return 0;
 252
 253         if (!mask) {
 254                 mask = zalloc(cpu__max_cpu());
 255                 if (!mask)
 256                         return -ENOMEM;
 257
 258                 counter->per_pkg_mask = mask;
 259         }
 260
 261         /*
 262          * we do not consider an event that has not run as a good
 263          * instance to mark a package as used (skip=1). Otherwise
 264          * we may run into a situation where the first CPU in a package
 265          * is not running anything, yet the second is, and this function
 266          * would mark the package as used after the first CPU and would
 267          * not read the values from the second CPU.
 268          */
 269         if (!(vals->run && vals->ena))
 270                 return 0;
 271
 272         s = cpu_map__get_socket(cpus, cpu, NULL);
 273         if (s < 0)
 274                 return -1;
 275
 276         *skip = test_and_set_bit(s, mask) == 1;
 277         return 0;
 278 }
 279
 280 static int
 281 process_counter_values(struct perf_stat_config *config, struct evsel *evsel,
 282                        int cpu, int thread,
 283                        struct perf_counts_values *count)
 284 {
 285         struct perf_counts_values *aggr = &evsel->counts->aggr;
 286         static struct perf_counts_values zero;
 287         bool skip = false;
 288
 289         if (check_per_pkg(evsel, count, cpu, &skip)) {
 290                 pr_err("failed to read per-pkg counter\n");
 291                 return -1;
 292         }
 293
 294         if (skip)
 295                 count = &zero;
 296
 297         switch (config->aggr_mode) {
 298         case AGGR_THREAD:
 299         case AGGR_CORE:
 300         case AGGR_DIE:
 301         case AGGR_SOCKET:
 302         case AGGR_NONE:
 303                 if (!evsel->snapshot)
 304                         perf_evsel__compute_deltas(evsel, cpu, thread, count);
 305                 perf_counts_values__scale(count, config->scale, NULL);
 306                 if ((config->aggr_mode == AGGR_NONE) && (!evsel->percore)) {
 307                         perf_stat__update_shadow_stats(evsel, count->val,
 308                                                        cpu, &rt_stat);
 309                 }
 310
 311                 if (config->aggr_mode == AGGR_THREAD) {
 312                         if (config->stats)
 313                                 perf_stat__update_shadow_stats(evsel,
 314                                         count->val, 0, &config->stats[thread]);
 315                         else
 316                                 perf_stat__update_shadow_stats(evsel,
 317                                         count->val, 0, &rt_stat);
 318                 }
 319                 break;
 320         case AGGR_GLOBAL:
 321                 aggr->val += count->val;
 322                 aggr->ena += count->ena;
 323                 aggr->run += count->run;
 324         case AGGR_UNSET:
 325         default:
 326                 break;
 327         }
 328
 329         return 0;
 330 }
 331
 332 static int process_counter_maps(struct perf_stat_config *config,
 333                                 struct evsel *counter)
 334 {
 335         int nthreads = perf_thread_map__nr(counter->core.threads);
 336         int ncpus = perf_evsel__nr_cpus(counter);
 337         int cpu, thread;
 338
 339         if (counter->core.system_wide)
 340                 nthreads = 1;
 341
 342         for (thread = 0; thread < nthreads; thread++) {
 343                 for (cpu = 0; cpu < ncpus; cpu++) {
 344                         if (process_counter_values(config, counter, cpu, thread,
 345                                                    perf_counts(counter->counts, cpu, thread)))
 346                                 return -1;
 347                 }
 348         }
 349
 350         return 0;
 351 }
 352
 353 int perf_stat_process_counter(struct perf_stat_config *config,
 354                               struct evsel *counter)
 355 {
 356         struct perf_counts_values *aggr = &counter->counts->aggr;
 357         struct perf_stat_evsel *ps = counter->stats;
 358         u64 *count = counter->counts->aggr.values;
 359         int i, ret;
 360
 361         aggr->val = aggr->ena = aggr->run = 0;
 362
 363         /*
 364          * We calculate counter's data every interval,
 365          * and the display code shows ps->res_stats
 366          * avg value. We need to zero the stats for
 367          * interval mode, otherwise overall avg running
 368          * averages will be shown for each interval.
 369          */
 370         if (config->interval)
 371                 init_stats(ps->res_stats);
 372
 373         if (counter->per_pkg)
 374                 zero_per_pkg(counter);
 375
 376         ret = process_counter_maps(config, counter);
 377         if (ret)
 378                 return ret;
 379
 380         if (config->aggr_mode != AGGR_GLOBAL)
 381                 return 0;
 382
 383         if (!counter->snapshot)
 384                 perf_evsel__compute_deltas(counter, -1, -1, aggr);
 385         perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled);
 386
 387         for (i = 0; i < 3; i++)
 388                 update_stats(&ps->res_stats[i], count[i]);
 389
 390         if (verbose > 0) {
 391                 fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 392                         perf_evsel__name(counter), count[0], count[1], count[2]);
 393         }
 394
 395         /*
 396          * Save the full runtime - to allow normalization during printout:
 397          */
 398         perf_stat__update_shadow_stats(counter, *count, 0, &rt_stat);
 399
 400         return 0;
 401 }
 402
 403 int perf_event__process_stat_event(struct perf_session *session,
 404                                    union perf_event *event)
 405 {
 406         struct perf_counts_values count;
 407         struct perf_record_stat *st = &event->stat;
 408         struct evsel *counter;
 409
 410         count.val = st->val;
 411         count.ena = st->ena;
 412         count.run = st->run;
 413
 414         counter = perf_evlist__id2evsel(session->evlist, st->id);
 415         if (!counter) {
 416                 pr_err("Failed to resolve counter for stat event.\n");
 417                 return -EINVAL;
 418         }
 419
 420         *perf_counts(counter->counts, st->cpu, st->thread) = count;
 421         counter->supported = true;
 422         return 0;
 423 }
 424
 425 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
 426 {
 427         struct perf_record_stat *st = (struct perf_record_stat *)event;
 428         size_t ret;
 429
 430         ret  = fprintf(fp, "\n... id %" PRI_lu64 ", cpu %d, thread %d\n",
 431                        st->id, st->cpu, st->thread);
 432         ret += fprintf(fp, "... value %" PRI_lu64 ", enabled %" PRI_lu64 ", running %" PRI_lu64 "\n",
 433                        st->val, st->ena, st->run);
 434
 435         return ret;
 436 }
 437
 438 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
 439 {
 440         struct perf_record_stat_round *rd = (struct perf_record_stat_round *)event;
 441         size_t ret;
 442
 443         ret = fprintf(fp, "\n... time %" PRI_lu64 ", type %s\n", rd->time,
 444                       rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
 445
 446         return ret;
 447 }
 448
 449 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
 450 {
 451         struct perf_stat_config sc;
 452         size_t ret;
 453
 454         perf_event__read_stat_config(&sc, &event->stat_config);
 455
 456         ret  = fprintf(fp, "\n");
 457         ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
 458         ret += fprintf(fp, "... scale     %d\n", sc.scale);
 459         ret += fprintf(fp, "... interval  %u\n", sc.interval);
 460
 461         return ret;
 462 }
 463
 464 int create_perf_stat_counter(struct evsel *evsel,
 465                              struct perf_stat_config *config,
 466                              struct target *target)
 467 {
 468         struct perf_event_attr *attr = &evsel->core.attr;
 469         struct evsel *leader = evsel->leader;
 470
 471         attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 472                             PERF_FORMAT_TOTAL_TIME_RUNNING;
 473
 474         /*
 475          * The event is part of non trivial group, let's enable
 476          * the group read (for leader) and ID retrieval for all
 477          * members.
 478          */
 479         if (leader->core.nr_members > 1)
 480                 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
 481
 482         attr->inherit = !config->no_inherit;
 483
 484         /*
 485          * Some events get initialized with sample_(period/type) set,
 486          * like tracepoints. Clear it up for counting.
 487          */
 488         attr->sample_period = 0;
 489
 490         if (config->identifier)
 491                 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
 492
 493         /*
 494          * Disabling all counters initially, they will be enabled
 495          * either manually by us or by kernel via enable_on_exec
 496          * set later.
 497          */
 498         if (perf_evsel__is_group_leader(evsel)) {
 499                 attr->disabled = 1;
 500
 501                 /*
 502                  * In case of initial_delay we enable tracee
 503                  * events manually.
 504                  */
 505                 if (target__none(target) && !config->initial_delay)
 506                         attr->enable_on_exec = 1;
 507         }
 508
 509         if (target__has_cpu(target) && !target__has_per_thread(target))
 510                 return perf_evsel__open_per_cpu(evsel, evsel__cpus(evsel));
 511
 512         return perf_evsel__open_per_thread(evsel, evsel->core.threads);
 513 }