tools/perf/util/stat.c

   1 #include <errno.h>
   2 #include <inttypes.h>
   3 #include <math.h>
   4 #include "stat.h"
   5 #include "evlist.h"
   6 #include "evsel.h"
   7 #include "thread_map.h"
   8
   9 void update_stats(struct stats *stats, u64 val)
  10 {
  11         double delta;
  12
  13         stats->n++;
  14         delta = val - stats->mean;
  15         stats->mean += delta / stats->n;
  16         stats->M2 += delta*(val - stats->mean);
  17
  18         if (val > stats->max)
  19                 stats->max = val;
  20
  21         if (val < stats->min)
  22                 stats->min = val;
  23 }
  24
  25 double avg_stats(struct stats *stats)
  26 {
  27         return stats->mean;
  28 }
  29
  30 /*
  31  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
  32  *
  33  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
  34  * s^2 = -------------------------------
  35  *                  n - 1
  36  *
  37  * http://en.wikipedia.org/wiki/Stddev
  38  *
  39  * The std dev of the mean is related to the std dev by:
  40  *
  41  *             s
  42  * s_mean = -------
  43  *          sqrt(n)
  44  *
  45  */
  46 double stddev_stats(struct stats *stats)
  47 {
  48         double variance, variance_mean;
  49
  50         if (stats->n < 2)
  51                 return 0.0;
  52
  53         variance = stats->M2 / (stats->n - 1);
  54         variance_mean = variance / stats->n;
  55
  56         return sqrt(variance_mean);
  57 }
  58
  59 double rel_stddev_stats(double stddev, double avg)
  60 {
  61         double pct = 0.0;
  62
  63         if (avg)
  64                 pct = 100.0 * stddev/avg;
  65
  66         return pct;
  67 }
  68
  69 bool __perf_evsel_stat__is(struct perf_evsel *evsel,
  70                            enum perf_stat_evsel_id id)
  71 {
  72         struct perf_stat_evsel *ps = evsel->priv;
  73
  74         return ps->id == id;
  75 }
  76
  77 #define ID(id, name) [PERF_STAT_EVSEL_ID__##id] = #name
  78 static const char *id_str[PERF_STAT_EVSEL_ID__MAX] = {
  79         ID(NONE,                x),
  80         ID(CYCLES_IN_TX,        cpu/cycles-t/),
  81         ID(TRANSACTION_START,   cpu/tx-start/),
  82         ID(ELISION_START,       cpu/el-start/),
  83         ID(CYCLES_IN_TX_CP,     cpu/cycles-ct/),
  84         ID(TOPDOWN_TOTAL_SLOTS, topdown-total-slots),
  85         ID(TOPDOWN_SLOTS_ISSUED, topdown-slots-issued),
  86         ID(TOPDOWN_SLOTS_RETIRED, topdown-slots-retired),
  87         ID(TOPDOWN_FETCH_BUBBLES, topdown-fetch-bubbles),
  88         ID(TOPDOWN_RECOVERY_BUBBLES, topdown-recovery-bubbles),
  89         ID(SMI_NUM, msr/smi/),
  90         ID(APERF, msr/aperf/),
  91 };
  92 #undef ID
  93
  94 void perf_stat_evsel_id_init(struct perf_evsel *evsel)
  95 {
  96         struct perf_stat_evsel *ps = evsel->priv;
  97         int i;
  98
  99         /* ps->id is 0 hence PERF_STAT_EVSEL_ID__NONE by default */
 100
 101         for (i = 0; i < PERF_STAT_EVSEL_ID__MAX; i++) {
 102                 if (!strcmp(perf_evsel__name(evsel), id_str[i])) {
 103                         ps->id = i;
 104                         break;
 105                 }
 106         }
 107 }
 108
 109 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
 110 {
 111         int i;
 112         struct perf_stat_evsel *ps = evsel->priv;
 113
 114         for (i = 0; i < 3; i++)
 115                 init_stats(&ps->res_stats[i]);
 116
 117         perf_stat_evsel_id_init(evsel);
 118 }
 119
 120 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
 121 {
 122         evsel->priv = zalloc(sizeof(struct perf_stat_evsel));
 123         if (evsel->priv == NULL)
 124                 return -ENOMEM;
 125         perf_evsel__reset_stat_priv(evsel);
 126         return 0;
 127 }
 128
 129 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
 130 {
 131         zfree(&evsel->priv);
 132 }
 133
 134 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel,
 135                                              int ncpus, int nthreads)
 136 {
 137         struct perf_counts *counts;
 138
 139         counts = perf_counts__new(ncpus, nthreads);
 140         if (counts)
 141                 evsel->prev_raw_counts = counts;
 142
 143         return counts ? 0 : -ENOMEM;
 144 }
 145
 146 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
 147 {
 148         perf_counts__delete(evsel->prev_raw_counts);
 149         evsel->prev_raw_counts = NULL;
 150 }
 151
 152 static int perf_evsel__alloc_stats(struct perf_evsel *evsel, bool alloc_raw)
 153 {
 154         int ncpus = perf_evsel__nr_cpus(evsel);
 155         int nthreads = thread_map__nr(evsel->threads);
 156
 157         if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
 158             perf_evsel__alloc_counts(evsel, ncpus, nthreads) < 0 ||
 159             (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel, ncpus, nthreads) < 0))
 160                 return -ENOMEM;
 161
 162         return 0;
 163 }
 164
 165 int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
 166 {
 167         struct perf_evsel *evsel;
 168
 169         evlist__for_each_entry(evlist, evsel) {
 170                 if (perf_evsel__alloc_stats(evsel, alloc_raw))
 171                         goto out_free;
 172         }
 173
 174         return 0;
 175
 176 out_free:
 177         perf_evlist__free_stats(evlist);
 178         return -1;
 179 }
 180
 181 void perf_evlist__free_stats(struct perf_evlist *evlist)
 182 {
 183         struct perf_evsel *evsel;
 184
 185         evlist__for_each_entry(evlist, evsel) {
 186                 perf_evsel__free_stat_priv(evsel);
 187                 perf_evsel__free_counts(evsel);
 188                 perf_evsel__free_prev_raw_counts(evsel);
 189         }
 190 }
 191
 192 void perf_evlist__reset_stats(struct perf_evlist *evlist)
 193 {
 194         struct perf_evsel *evsel;
 195
 196         evlist__for_each_entry(evlist, evsel) {
 197                 perf_evsel__reset_stat_priv(evsel);
 198                 perf_evsel__reset_counts(evsel);
 199         }
 200 }
 201
 202 static void zero_per_pkg(struct perf_evsel *counter)
 203 {
 204         if (counter->per_pkg_mask)
 205                 memset(counter->per_pkg_mask, 0, MAX_NR_CPUS);
 206 }
 207
 208 static int check_per_pkg(struct perf_evsel *counter,
 209                          struct perf_counts_values *vals, int cpu, bool *skip)
 210 {
 211         unsigned long *mask = counter->per_pkg_mask;
 212         struct cpu_map *cpus = perf_evsel__cpus(counter);
 213         int s;
 214
 215         *skip = false;
 216
 217         if (!counter->per_pkg)
 218                 return 0;
 219
 220         if (cpu_map__empty(cpus))
 221                 return 0;
 222
 223         if (!mask) {
 224                 mask = zalloc(MAX_NR_CPUS);
 225                 if (!mask)
 226                         return -ENOMEM;
 227
 228                 counter->per_pkg_mask = mask;
 229         }
 230
 231         /*
 232          * we do not consider an event that has not run as a good
 233          * instance to mark a package as used (skip=1). Otherwise
 234          * we may run into a situation where the first CPU in a package
 235          * is not running anything, yet the second is, and this function
 236          * would mark the package as used after the first CPU and would
 237          * not read the values from the second CPU.
 238          */
 239         if (!(vals->run && vals->ena))
 240                 return 0;
 241
 242         s = cpu_map__get_socket(cpus, cpu, NULL);
 243         if (s < 0)
 244                 return -1;
 245
 246         *skip = test_and_set_bit(s, mask) == 1;
 247         return 0;
 248 }
 249
 250 static int
 251 process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel,
 252                        int cpu, int thread,
 253                        struct perf_counts_values *count)
 254 {
 255         struct perf_counts_values *aggr = &evsel->counts->aggr;
 256         static struct perf_counts_values zero;
 257         bool skip = false;
 258
 259         if (check_per_pkg(evsel, count, cpu, &skip)) {
 260                 pr_err("failed to read per-pkg counter\n");
 261                 return -1;
 262         }
 263
 264         if (skip)
 265                 count = &zero;
 266
 267         switch (config->aggr_mode) {
 268         case AGGR_THREAD:
 269         case AGGR_CORE:
 270         case AGGR_SOCKET:
 271         case AGGR_NONE:
 272                 if (!evsel->snapshot)
 273                         perf_evsel__compute_deltas(evsel, cpu, thread, count);
 274                 perf_counts_values__scale(count, config->scale, NULL);
 275                 if (config->aggr_mode == AGGR_NONE)
 276                         perf_stat__update_shadow_stats(evsel, count->values, cpu);
 277                 break;
 278         case AGGR_GLOBAL:
 279                 aggr->val += count->val;
 280                 if (config->scale) {
 281                         aggr->ena += count->ena;
 282                         aggr->run += count->run;
 283                 }
 284         case AGGR_UNSET:
 285         default:
 286                 break;
 287         }
 288
 289         return 0;
 290 }
 291
 292 static int process_counter_maps(struct perf_stat_config *config,
 293                                 struct perf_evsel *counter)
 294 {
 295         int nthreads = thread_map__nr(counter->threads);
 296         int ncpus = perf_evsel__nr_cpus(counter);
 297         int cpu, thread;
 298
 299         if (counter->system_wide)
 300                 nthreads = 1;
 301
 302         for (thread = 0; thread < nthreads; thread++) {
 303                 for (cpu = 0; cpu < ncpus; cpu++) {
 304                         if (process_counter_values(config, counter, cpu, thread,
 305                                                    perf_counts(counter->counts, cpu, thread)))
 306                                 return -1;
 307                 }
 308         }
 309
 310         return 0;
 311 }
 312
 313 int perf_stat_process_counter(struct perf_stat_config *config,
 314                               struct perf_evsel *counter)
 315 {
 316         struct perf_counts_values *aggr = &counter->counts->aggr;
 317         struct perf_stat_evsel *ps = counter->priv;
 318         u64 *count = counter->counts->aggr.values;
 319         u64 val;
 320         int i, ret;
 321
 322         aggr->val = aggr->ena = aggr->run = 0;
 323
 324         /*
 325          * We calculate counter's data every interval,
 326          * and the display code shows ps->res_stats
 327          * avg value. We need to zero the stats for
 328          * interval mode, otherwise overall avg running
 329          * averages will be shown for each interval.
 330          */
 331         if (config->interval)
 332                 init_stats(ps->res_stats);
 333
 334         if (counter->per_pkg)
 335                 zero_per_pkg(counter);
 336
 337         ret = process_counter_maps(config, counter);
 338         if (ret)
 339                 return ret;
 340
 341         if (config->aggr_mode != AGGR_GLOBAL)
 342                 return 0;
 343
 344         if (!counter->snapshot)
 345                 perf_evsel__compute_deltas(counter, -1, -1, aggr);
 346         perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled);
 347
 348         for (i = 0; i < 3; i++)
 349                 update_stats(&ps->res_stats[i], count[i]);
 350
 351         if (verbose > 0) {
 352                 fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 353                         perf_evsel__name(counter), count[0], count[1], count[2]);
 354         }
 355
 356         /*
 357          * Save the full runtime - to allow normalization during printout:
 358          */
 359         val = counter->scale * *count;
 360         perf_stat__update_shadow_stats(counter, &val, 0);
 361
 362         return 0;
 363 }
 364
 365 int perf_event__process_stat_event(struct perf_tool *tool __maybe_unused,
 366                                    union perf_event *event,
 367                                    struct perf_session *session)
 368 {
 369         struct perf_counts_values count;
 370         struct stat_event *st = &event->stat;
 371         struct perf_evsel *counter;
 372
 373         count.val = st->val;
 374         count.ena = st->ena;
 375         count.run = st->run;
 376
 377         counter = perf_evlist__id2evsel(session->evlist, st->id);
 378         if (!counter) {
 379                 pr_err("Failed to resolve counter for stat event.\n");
 380                 return -EINVAL;
 381         }
 382
 383         *perf_counts(counter->counts, st->cpu, st->thread) = count;
 384         counter->supported = true;
 385         return 0;
 386 }
 387
 388 size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp)
 389 {
 390         struct stat_event *st = (struct stat_event *) event;
 391         size_t ret;
 392
 393         ret  = fprintf(fp, "\n... id %" PRIu64 ", cpu %d, thread %d\n",
 394                        st->id, st->cpu, st->thread);
 395         ret += fprintf(fp, "... value %" PRIu64 ", enabled %" PRIu64 ", running %" PRIu64 "\n",
 396                        st->val, st->ena, st->run);
 397
 398         return ret;
 399 }
 400
 401 size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp)
 402 {
 403         struct stat_round_event *rd = (struct stat_round_event *)event;
 404         size_t ret;
 405
 406         ret = fprintf(fp, "\n... time %" PRIu64 ", type %s\n", rd->time,
 407                       rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL");
 408
 409         return ret;
 410 }
 411
 412 size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp)
 413 {
 414         struct perf_stat_config sc;
 415         size_t ret;
 416
 417         perf_event__read_stat_config(&sc, &event->stat_config);
 418
 419         ret  = fprintf(fp, "\n");
 420         ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode);
 421         ret += fprintf(fp, "... scale     %d\n", sc.scale);
 422         ret += fprintf(fp, "... interval  %u\n", sc.interval);
 423
 424         return ret;
 425 }