tools/perf/util/stat.c

   1 #include <math.h>
   2 #include "stat.h"
   3 #include "evlist.h"
   4 #include "evsel.h"
   5 #include "thread_map.h"
   6
   7 void update_stats(struct stats *stats, u64 val)
   8 {
   9         double delta;
  10
  11         stats->n++;
  12         delta = val - stats->mean;
  13         stats->mean += delta / stats->n;
  14         stats->M2 += delta*(val - stats->mean);
  15
  16         if (val > stats->max)
  17                 stats->max = val;
  18
  19         if (val < stats->min)
  20                 stats->min = val;
  21 }
  22
  23 double avg_stats(struct stats *stats)
  24 {
  25         return stats->mean;
  26 }
  27
  28 /*
  29  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
  30  *
  31  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
  32  * s^2 = -------------------------------
  33  *                  n - 1
  34  *
  35  * http://en.wikipedia.org/wiki/Stddev
  36  *
  37  * The std dev of the mean is related to the std dev by:
  38  *
  39  *             s
  40  * s_mean = -------
  41  *          sqrt(n)
  42  *
  43  */
  44 double stddev_stats(struct stats *stats)
  45 {
  46         double variance, variance_mean;
  47
  48         if (stats->n < 2)
  49                 return 0.0;
  50
  51         variance = stats->M2 / (stats->n - 1);
  52         variance_mean = variance / stats->n;
  53
  54         return sqrt(variance_mean);
  55 }
  56
  57 double rel_stddev_stats(double stddev, double avg)
  58 {
  59         double pct = 0.0;
  60
  61         if (avg)
  62                 pct = 100.0 * stddev/avg;
  63
  64         return pct;
  65 }
  66
  67 bool __perf_evsel_stat__is(struct perf_evsel *evsel,
  68                            enum perf_stat_evsel_id id)
  69 {
  70         struct perf_stat_evsel *ps = evsel->priv;
  71
  72         return ps->id == id;
  73 }
  74
  75 #define ID(id, name) [PERF_STAT_EVSEL_ID__##id] = #name
  76 static const char *id_str[PERF_STAT_EVSEL_ID__MAX] = {
  77         ID(NONE,                x),
  78         ID(CYCLES_IN_TX,        cpu/cycles-t/),
  79         ID(TRANSACTION_START,   cpu/tx-start/),
  80         ID(ELISION_START,       cpu/el-start/),
  81         ID(CYCLES_IN_TX_CP,     cpu/cycles-ct/),
  82         ID(TOPDOWN_TOTAL_SLOTS, topdown-total-slots),
  83         ID(TOPDOWN_SLOTS_ISSUED, topdown-slots-issued),
  84         ID(TOPDOWN_SLOTS_RETIRED, topdown-slots-retired),
  85         ID(TOPDOWN_FETCH_BUBBLES, topdown-fetch-bubbles),
  86         ID(TOPDOWN_RECOVERY_BUBBLES, topdown-recovery-bubbles),
  87 };
  88 #undef ID
  89
  90 void perf_stat_evsel_id_init(struct perf_evsel *evsel)
  91 {
  92         struct perf_stat_evsel *ps = evsel->priv;
  93         int i;
  94
  95         /* ps->id is 0 hence PERF_STAT_EVSEL_ID__NONE by default */
  96
  97         for (i = 0; i < PERF_STAT_EVSEL_ID__MAX; i++) {
  98                 if (!strcmp(perf_evsel__name(evsel), id_str[i])) {
  99                         ps->id = i;
 100                         break;
 101                 }
 102         }
 103 }
 104
 105 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
 106 {
 107         int i;
 108         struct perf_stat_evsel *ps = evsel->priv;
 109
 110         for (i = 0; i < 3; i++)
 111                 init_stats(&ps->res_stats[i]);
 112
 113         perf_stat_evsel_id_init(evsel);
 114 }
 115
 116 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
 117 {
 118         evsel->priv = zalloc(sizeof(struct perf_stat_evsel));
 119         if (evsel->priv == NULL)
 120                 return -ENOMEM;
 121         perf_evsel__reset_stat_priv(evsel);
 122         return 0;
 123 }
 124
 125 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
 126 {
 127         zfree(&evsel->priv);
 128 }
 129
 130 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel,
 131                                              int ncpus, int nthreads)
 132 {
 133         struct perf_counts *counts;
 134
 135         counts = perf_counts__new(ncpus, nthreads);
 136         if (counts)
 137                 evsel->prev_raw_counts = counts;
 138
 139         return counts ? 0 : -ENOMEM;
 140 }
 141
 142 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
 143 {
 144         perf_counts__delete(evsel->prev_raw_counts);
 145         evsel->prev_raw_counts = NULL;
 146 }
 147
 148 static int perf_evsel__alloc_stats(struct perf_evsel *evsel, bool alloc_raw)
 149 {
 150         int ncpus = perf_evsel__nr_cpus(evsel);
 151         int nthreads = thread_map__nr(evsel->threads);
 152
 153         if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
 154             perf_evsel__alloc_counts(evsel, ncpus, nthreads) < 0 ||
 155             (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel, ncpus, nthreads) < 0))
 156                 return -ENOMEM;
 157
 158         return 0;
 159 }
 160
 161 int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
 162 {
 163         struct perf_evsel *evsel;
 164
 165         evlist__for_each(evlist, evsel) {
 166                 if (perf_evsel__alloc_stats(evsel, alloc_raw))
 167                         goto out_free;
 168         }
 169
 170         return 0;
 171
 172 out_free:
 173         perf_evlist__free_stats(evlist);
 174         return -1;
 175 }
 176
 177 void perf_evlist__free_stats(struct perf_evlist *evlist)
 178 {
 179         struct perf_evsel *evsel;
 180
 181         evlist__for_each(evlist, evsel) {
 182                 perf_evsel__free_stat_priv(evsel);
 183                 perf_evsel__free_counts(evsel);
 184                 perf_evsel__free_prev_raw_counts(evsel);
 185         }
 186 }
 187
 188 void perf_evlist__reset_stats(struct perf_evlist *evlist)
 189 {
 190         struct perf_evsel *evsel;
 191
 192         evlist__for_each(evlist, evsel) {
 193                 perf_evsel__reset_stat_priv(evsel);
 194                 perf_evsel__reset_counts(evsel);
 195         }
 196 }
 197
 198 static void zero_per_pkg(struct perf_evsel *counter)
 199 {
 200         if (counter->per_pkg_mask)
 201                 memset(counter->per_pkg_mask, 0, MAX_NR_CPUS);
 202 }
 203
 204 static int check_per_pkg(struct perf_evsel *counter,
 205                          struct perf_counts_values *vals, int cpu, bool *skip)
 206 {
 207         unsigned long *mask = counter->per_pkg_mask;
 208         struct cpu_map *cpus = perf_evsel__cpus(counter);
 209         int s;
 210
 211         *skip = false;
 212
 213         if (!counter->per_pkg)
 214                 return 0;
 215
 216         if (cpu_map__empty(cpus))
 217                 return 0;
 218
 219         if (!mask) {
 220                 mask = zalloc(MAX_NR_CPUS);
 221                 if (!mask)
 222                         return -ENOMEM;
 223
 224                 counter->per_pkg_mask = mask;
 225         }
 226
 227         /*
 228          * we do not consider an event that has not run as a good
 229          * instance to mark a package as used (skip=1). Otherwise
 230          * we may run into a situation where the first CPU in a package
 231          * is not running anything, yet the second is, and this function
 232          * would mark the package as used after the first CPU and would
 233          * not read the values from the second CPU.
 234          */
 235         if (!(vals->run && vals->ena))
 236                 return 0;
 237
 238         s = cpu_map__get_socket(cpus, cpu, NULL);
 239         if (s < 0)
 240                 return -1;
 241
 242         *skip = test_and_set_bit(s, mask) == 1;
 243         return 0;
 244 }
 245
 246 static int
 247 process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel,
 248                        int cpu, int thread,
 249                        struct perf_counts_values *count)
 250 {
 251         struct perf_counts_values *aggr = &evsel->counts->aggr;
 252         static struct perf_counts_values zero;
 253         bool skip = false;
 254
 255         if (check_per_pkg(evsel, count, cpu, &skip)) {
 256                 pr_err("failed to read per-pkg counter\n");
 257                 return -1;
 258         }
 259
 260         if (skip)
 261                 count = &zero;
 262
 263         switch (config->aggr_mode) {
 264         case AGGR_THREAD:
 265         case AGGR_CORE:
 266         case AGGR_SOCKET:
 267         case AGGR_NONE:
 268                 if (!evsel->snapshot)
 269                         perf_evsel__compute_deltas(evsel, cpu, thread, count);
 270                 perf_counts_values__scale(count, config->scale, NULL);
 271                 if (config->aggr_mode == AGGR_NONE)
 272                         perf_stat__update_shadow_stats(evsel, count->values, cpu);
 273                 break;
 274         case AGGR_GLOBAL:
 275                 aggr->val += count->val;
 276                 if (config->scale) {
 277                         aggr->ena += count->ena;
 278                         aggr->run += count->run;
 279                 }
 280         case AGGR_UNSET:
 281         default:
 282                 break;
 283         }
 284
 285         return 0;
 286 }
 287
 288 static int process_counter_maps(struct perf_stat_config *config,
 289                                 struct perf_evsel *counter)
 290 {
 291         int nthreads = thread_map__nr(counter->threads);
 292         int ncpus = perf_evsel__nr_cpus(counter);
 293         int cpu, thread;
 294
 295         if (counter->system_wide)
 296                 nthreads = 1;
 297
 298         for (thread = 0; thread < nthreads; thread++) {
 299                 for (cpu = 0; cpu < ncpus; cpu++) {
 300                         if (process_counter_values(config, counter, cpu, thread,
 301                                                    perf_counts(counter->counts, cpu, thread)))
 302                                 return -1;
 303                 }
 304         }
 305
 306         return 0;
 307 }
 308
 309 int perf_stat_process_counter(struct perf_stat_config *config,
 310                               struct perf_evsel *counter)
 311 {
 312         struct perf_counts_values *aggr = &counter->counts->aggr;
 313         struct perf_stat_evsel *ps = counter->priv;
 314         u64 *count = counter->counts->aggr.values;
 315         u64 val;
 316         int i, ret;
 317
 318         aggr->val = aggr->ena = aggr->run = 0;
 319
 320         /*
 321          * We calculate counter's data every interval,
 322          * and the display code shows ps->res_stats
 323          * avg value. We need to zero the stats for
 324          * interval mode, otherwise overall avg running
 325          * averages will be shown for each interval.
 326          */
 327         if (config->interval)
 328                 init_stats(ps->res_stats);
 329
 330         if (counter->per_pkg)
 331                 zero_per_pkg(counter);
 332
 333         ret = process_counter_maps(config, counter);
 334         if (ret)
 335                 return ret;
 336
 337         if (config->aggr_mode != AGGR_GLOBAL)
 338                 return 0;
 339
 340         if (!counter->snapshot)
 341                 perf_evsel__compute_deltas(counter, -1, -1, aggr);
 342         perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled);
 343
 344         for (i = 0; i < 3; i++)
 345                 update_stats(&ps->res_stats[i], count[i]);
 346
 347         if (verbose) {
 348                 fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 349                         perf_evsel__name(counter), count[0], count[1], count[2]);
 350         }
 351
 352         /*
 353          * Save the full runtime - to allow normalization during printout:
 354          */
 355         val = counter->scale * *count;
 356         perf_stat__update_shadow_stats(counter, &val, 0);
 357
 358         return 0;
 359 }
 360
 361 int perf_event__process_stat_event(struct perf_tool *tool __maybe_unused,
 362                                    union perf_event *event,
 363                                    struct perf_session *session)
 364 {
 365         struct perf_counts_values count;
 366         struct stat_event *st = &event->stat;
 367         struct perf_evsel *counter;
 368
 369         count.val = st->val;
 370         count.ena = st->ena;
 371         count.run = st->run;
 372
 373         counter = perf_evlist__id2evsel(session->evlist, st->id);
 374         if (!counter) {
 375                 pr_err("Failed to resolve counter for stat event.\n");
 376                 return -EINVAL;
 377         }
 378
 379         *perf_counts(counter->counts, st->cpu, st->thread) = count;
 380         counter->supported = true;
 381         return 0;
 382 }
 383
 384 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
 385 {
 386         struct stat_event *st = (struct stat_event *) event;
 387         size_t ret;
 388
 389         ret  = fprintf(fp, "\n... id %" PRIu64 ", cpu %d, thread %d\n",
 390                        st->id, st->cpu, st->thread);
 391         ret += fprintf(fp, "... value %" PRIu64 ", enabled %" PRIu64 ", running %" PRIu64 "\n",
 392                        st->val, st->ena, st->run);
 393
 394         return ret;
 395 }
 396
 397 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
 398 {
 399         struct stat_round_event *rd = (struct stat_round_event *)event;
 400         size_t ret;
 401
 402         ret = fprintf(fp, "\n... time %" PRIu64 ", type %s\n", rd->time,
 403                       rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
 404
 405         return ret;
 406 }
 407
 408 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
 409 {
 410         struct perf_stat_config sc;
 411         size_t ret;
 412
 413         perf_event__read_stat_config(&sc, &event->stat_config);
 414
 415         ret  = fprintf(fp, "\n");
 416         ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
 417         ret += fprintf(fp, "... scale     %d\n", sc.scale);
 418         ret += fprintf(fp, "... interval  %u\n", sc.interval);
 419
 420         return ret;
 421 }