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