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 void perf_evsel__reset_prev_raw_counts(struct perf_evsel *evsel)
 158 {
 159         if (evsel->prev_raw_counts) {
 160                 evsel->prev_raw_counts->aggr.val = 0;
 161                 evsel->prev_raw_counts->aggr.ena = 0;
 162                 evsel->prev_raw_counts->aggr.run = 0;
 163        }
 164 }
 165
 166 static int perf_evsel__alloc_stats(struct perf_evsel *evsel, bool alloc_raw)
 167 {
 168         int ncpus = perf_evsel__nr_cpus(evsel);
 169         int nthreads = thread_map__nr(evsel->threads);
 170
 171         if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
 172             perf_evsel__alloc_counts(evsel, ncpus, nthreads) < 0 ||
 173             (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel, ncpus, nthreads) < 0))
 174                 return -ENOMEM;
 175
 176         return 0;
 177 }
 178
 179 int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
 180 {
 181         struct perf_evsel *evsel;
 182
 183         evlist__for_each_entry(evlist, evsel) {
 184                 if (perf_evsel__alloc_stats(evsel, alloc_raw))
 185                         goto out_free;
 186         }
 187
 188         return 0;
 189
 190 out_free:
 191         perf_evlist__free_stats(evlist);
 192         return -1;
 193 }
 194
 195 void perf_evlist__free_stats(struct perf_evlist *evlist)
 196 {
 197         struct perf_evsel *evsel;
 198
 199         evlist__for_each_entry(evlist, evsel) {
 200                 perf_evsel__free_stat_priv(evsel);
 201                 perf_evsel__free_counts(evsel);
 202                 perf_evsel__free_prev_raw_counts(evsel);
 203         }
 204 }
 205
 206 void perf_evlist__reset_stats(struct perf_evlist *evlist)
 207 {
 208         struct perf_evsel *evsel;
 209
 210         evlist__for_each_entry(evlist, evsel) {
 211                 perf_evsel__reset_stat_priv(evsel);
 212                 perf_evsel__reset_counts(evsel);
 213         }
 214 }
 215
 216 void perf_evlist__reset_prev_raw_counts(struct perf_evlist *evlist)
 217 {
 218         struct perf_evsel *evsel;
 219
 220         evlist__for_each_entry(evlist, evsel)
 221                 perf_evsel__reset_prev_raw_counts(evsel);
 222 }
 223
 224 static void zero_per_pkg(struct perf_evsel *counter)
 225 {
 226         if (counter->per_pkg_mask)
 227                 memset(counter->per_pkg_mask, 0, MAX_NR_CPUS);
 228 }
 229
 230 static int check_per_pkg(struct perf_evsel *counter,
 231                          struct perf_counts_values *vals, int cpu, bool *skip)
 232 {
 233         unsigned long *mask = counter->per_pkg_mask;
 234         struct cpu_map *cpus = perf_evsel__cpus(counter);
 235         int s;
 236
 237         *skip = false;
 238
 239         if (!counter->per_pkg)
 240                 return 0;
 241
 242         if (cpu_map__empty(cpus))
 243                 return 0;
 244
 245         if (!mask) {
 246                 mask = zalloc(MAX_NR_CPUS);
 247                 if (!mask)
 248                         return -ENOMEM;
 249
 250                 counter->per_pkg_mask = mask;
 251         }
 252
 253         /*
 254          * we do not consider an event that has not run as a good
 255          * instance to mark a package as used (skip=1). Otherwise
 256          * we may run into a situation where the first CPU in a package
 257          * is not running anything, yet the second is, and this function
 258          * would mark the package as used after the first CPU and would
 259          * not read the values from the second CPU.
 260          */
 261         if (!(vals->run && vals->ena))
 262                 return 0;
 263
 264         s = cpu_map__get_socket(cpus, cpu, NULL);
 265         if (s < 0)
 266                 return -1;
 267
 268         *skip = test_and_set_bit(s, mask) == 1;
 269         return 0;
 270 }
 271
 272 static int
 273 process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel,
 274                        int cpu, int thread,
 275                        struct perf_counts_values *count)
 276 {
 277         struct perf_counts_values *aggr = &evsel->counts->aggr;
 278         static struct perf_counts_values zero;
 279         bool skip = false;
 280
 281         if (check_per_pkg(evsel, count, cpu, &skip)) {
 282                 pr_err("failed to read per-pkg counter\n");
 283                 return -1;
 284         }
 285
 286         if (skip)
 287                 count = &zero;
 288
 289         switch (config->aggr_mode) {
 290         case AGGR_THREAD:
 291         case AGGR_CORE:
 292         case AGGR_SOCKET:
 293         case AGGR_NONE:
 294                 if (!evsel->snapshot)
 295                         perf_evsel__compute_deltas(evsel, cpu, thread, count);
 296                 perf_counts_values__scale(count, config->scale, NULL);
 297                 if (config->aggr_mode == AGGR_NONE)
 298                         perf_stat__update_shadow_stats(evsel, count->val, cpu,
 299                                                        &rt_stat);
 300                 if (config->aggr_mode == AGGR_THREAD) {
 301                         if (config->stats)
 302                                 perf_stat__update_shadow_stats(evsel,
 303                                         count->val, 0, &config->stats[thread]);
 304                         else
 305                                 perf_stat__update_shadow_stats(evsel,
 306                                         count->val, 0, &rt_stat);
 307                 }
 308                 break;
 309         case AGGR_GLOBAL:
 310                 aggr->val += count->val;
 311                 if (config->scale) {
 312                         aggr->ena += count->ena;
 313                         aggr->run += count->run;
 314                 }
 315         case AGGR_UNSET:
 316         default:
 317                 break;
 318         }
 319
 320         return 0;
 321 }
 322
 323 static int process_counter_maps(struct perf_stat_config *config,
 324                                 struct perf_evsel *counter)
 325 {
 326         int nthreads = thread_map__nr(counter->threads);
 327         int ncpus = perf_evsel__nr_cpus(counter);
 328         int cpu, thread;
 329
 330         if (counter->system_wide)
 331                 nthreads = 1;
 332
 333         for (thread = 0; thread < nthreads; thread++) {
 334                 for (cpu = 0; cpu < ncpus; cpu++) {
 335                         if (process_counter_values(config, counter, cpu, thread,
 336                                                    perf_counts(counter->counts, cpu, thread)))
 337                                 return -1;
 338                 }
 339         }
 340
 341         return 0;
 342 }
 343
 344 int perf_stat_process_counter(struct perf_stat_config *config,
 345                               struct perf_evsel *counter)
 346 {
 347         struct perf_counts_values *aggr = &counter->counts->aggr;
 348         struct perf_stat_evsel *ps = counter->stats;
 349         u64 *count = counter->counts->aggr.values;
 350         int i, ret;
 351
 352         aggr->val = aggr->ena = aggr->run = 0;
 353
 354         /*
 355          * We calculate counter's data every interval,
 356          * and the display code shows ps->res_stats
 357          * avg value. We need to zero the stats for
 358          * interval mode, otherwise overall avg running
 359          * averages will be shown for each interval.
 360          */
 361         if (config->interval)
 362                 init_stats(ps->res_stats);
 363
 364         if (counter->per_pkg)
 365                 zero_per_pkg(counter);
 366
 367         ret = process_counter_maps(config, counter);
 368         if (ret)
 369                 return ret;
 370
 371         if (config->aggr_mode != AGGR_GLOBAL)
 372                 return 0;
 373
 374         if (!counter->snapshot)
 375                 perf_evsel__compute_deltas(counter, -1, -1, aggr);
 376         perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled);
 377
 378         for (i = 0; i < 3; i++)
 379                 update_stats(&ps->res_stats[i], count[i]);
 380
 381         if (verbose > 0) {
 382                 fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 383                         perf_evsel__name(counter), count[0], count[1], count[2]);
 384         }
 385
 386         /*
 387          * Save the full runtime - to allow normalization during printout:
 388          */
 389         perf_stat__update_shadow_stats(counter, *count, 0, &rt_stat);
 390
 391         return 0;
 392 }
 393
 394 int perf_event__process_stat_event(struct perf_tool *tool __maybe_unused,
 395                                    union perf_event *event,
 396                                    struct perf_session *session)
 397 {
 398         struct perf_counts_values count;
 399         struct stat_event *st = &event->stat;
 400         struct perf_evsel *counter;
 401
 402         count.val = st->val;
 403         count.ena = st->ena;
 404         count.run = st->run;
 405
 406         counter = perf_evlist__id2evsel(session->evlist, st->id);
 407         if (!counter) {
 408                 pr_err("Failed to resolve counter for stat event.\n");
 409                 return -EINVAL;
 410         }
 411
 412         *perf_counts(counter->counts, st->cpu, st->thread) = count;
 413         counter->supported = true;
 414         return 0;
 415 }
 416
 417 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
 418 {
 419         struct stat_event *st = (struct stat_event *) event;
 420         size_t ret;
 421
 422         ret  = fprintf(fp, "\n... id %" PRIu64 ", cpu %d, thread %d\n",
 423                        st->id, st->cpu, st->thread);
 424         ret += fprintf(fp, "... value %" PRIu64 ", enabled %" PRIu64 ", running %" PRIu64 "\n",
 425                        st->val, st->ena, st->run);
 426
 427         return ret;
 428 }
 429
 430 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
 431 {
 432         struct stat_round_event *rd = (struct stat_round_event *)event;
 433         size_t ret;
 434
 435         ret = fprintf(fp, "\n... time %" PRIu64 ", type %s\n", rd->time,
 436                       rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
 437
 438         return ret;
 439 }
 440
 441 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
 442 {
 443         struct perf_stat_config sc;
 444         size_t ret;
 445
 446         perf_event__read_stat_config(&sc, &event->stat_config);
 447
 448         ret  = fprintf(fp, "\n");
 449         ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
 450         ret += fprintf(fp, "... scale     %d\n", sc.scale);
 451         ret += fprintf(fp, "... interval  %u\n", sc.interval);
 452
 453         return ret;
 454 }