tools/perf/util/stat-display.c

   1 #include <stdlib.h>
   2 #include <stdio.h>
   3 #include <inttypes.h>
   4 #include <linux/string.h>
   5 #include <linux/time64.h>
   6 #include <math.h>
   7 #include "color.h"
   8 #include "counts.h"
   9 #include "evlist.h"
  10 #include "evsel.h"
  11 #include "stat.h"
  12 #include "top.h"
  13 #include "thread_map.h"
  14 #include "cpumap.h"
  15 #include "string2.h"
  16 #include <linux/ctype.h>
  17 #include "cgroup.h"
  18 #include <api/fs/fs.h>
  19 #include "util.h"
  20 #include "iostat.h"
  21 #include "pmu-hybrid.h"
  22 #include "evlist-hybrid.h"
  23
  24 #define CNTR_NOT_SUPPORTED      "<not supported>"
  25 #define CNTR_NOT_COUNTED        "<not counted>"
  26
  27 static void print_running(struct perf_stat_config *config,
  28                           u64 run, u64 ena)
  29 {
  30         if (config->csv_output) {
  31                 fprintf(config->output, "%s%" PRIu64 "%s%.2f",
  32                                         config->csv_sep,
  33                                         run,
  34                                         config->csv_sep,
  35                                         ena ? 100.0 * run / ena : 100.0);
  36         } else if (run != ena) {
  37                 fprintf(config->output, "  (%.2f%%)", 100.0 * run / ena);
  38         }
  39 }
  40
  41 static void print_noise_pct(struct perf_stat_config *config,
  42                             double total, double avg)
  43 {
  44         double pct = rel_stddev_stats(total, avg);
  45
  46         if (config->csv_output)
  47                 fprintf(config->output, "%s%.2f%%", config->csv_sep, pct);
  48         else if (pct)
  49                 fprintf(config->output, "  ( +-%6.2f%% )", pct);
  50 }
  51
  52 static void print_noise(struct perf_stat_config *config,
  53                         struct evsel *evsel, double avg)
  54 {
  55         struct perf_stat_evsel *ps;
  56
  57         if (config->run_count == 1)
  58                 return;
  59
  60         ps = evsel->stats;
  61         print_noise_pct(config, stddev_stats(&ps->res_stats[0]), avg);
  62 }
  63
  64 static void print_cgroup(struct perf_stat_config *config, struct evsel *evsel)
  65 {
  66         if (nr_cgroups) {
  67                 const char *cgrp_name = evsel->cgrp ? evsel->cgrp->name  : "";
  68                 fprintf(config->output, "%s%s", config->csv_sep, cgrp_name);
  69         }
  70 }
  71
  72
  73 static void aggr_printout(struct perf_stat_config *config,
  74                           struct evsel *evsel, struct aggr_cpu_id id, int nr)
  75 {
  76         switch (config->aggr_mode) {
  77         case AGGR_CORE:
  78                 fprintf(config->output, "S%d-D%d-C%*d%s%*d%s",
  79                         id.socket,
  80                         id.die,
  81                         config->csv_output ? 0 : -8,
  82                         id.core,
  83                         config->csv_sep,
  84                         config->csv_output ? 0 : 4,
  85                         nr,
  86                         config->csv_sep);
  87                 break;
  88         case AGGR_DIE:
  89                 fprintf(config->output, "S%d-D%*d%s%*d%s",
  90                         id.socket,
  91                         config->csv_output ? 0 : -8,
  92                         id.die,
  93                         config->csv_sep,
  94                         config->csv_output ? 0 : 4,
  95                         nr,
  96                         config->csv_sep);
  97                 break;
  98         case AGGR_SOCKET:
  99                 fprintf(config->output, "S%*d%s%*d%s",
 100                         config->csv_output ? 0 : -5,
 101                         id.socket,
 102                         config->csv_sep,
 103                         config->csv_output ? 0 : 4,
 104                         nr,
 105                         config->csv_sep);
 106                         break;
 107         case AGGR_NODE:
 108                 fprintf(config->output, "N%*d%s%*d%s",
 109                         config->csv_output ? 0 : -5,
 110                         id.node,
 111                         config->csv_sep,
 112                         config->csv_output ? 0 : 4,
 113                         nr,
 114                         config->csv_sep);
 115                         break;
 116         case AGGR_NONE:
 117                 if (evsel->percore && !config->percore_show_thread) {
 118                         fprintf(config->output, "S%d-D%d-C%*d%s",
 119                                 id.socket,
 120                                 id.die,
 121                                 config->csv_output ? 0 : -3,
 122                                 id.core, config->csv_sep);
 123                 } else if (id.core > -1) {
 124                         fprintf(config->output, "CPU%*d%s",
 125                                 config->csv_output ? 0 : -7,
 126                                 evsel__cpus(evsel)->map[id.core],
 127                                 config->csv_sep);
 128                 }
 129                 break;
 130         case AGGR_THREAD:
 131                 fprintf(config->output, "%*s-%*d%s",
 132                         config->csv_output ? 0 : 16,
 133                         perf_thread_map__comm(evsel->core.threads, id.thread),
 134                         config->csv_output ? 0 : -8,
 135                         perf_thread_map__pid(evsel->core.threads, id.thread),
 136                         config->csv_sep);
 137                 break;
 138         case AGGR_GLOBAL:
 139         case AGGR_UNSET:
 140         default:
 141                 break;
 142         }
 143 }
 144
 145 struct outstate {
 146         FILE *fh;
 147         bool newline;
 148         const char *prefix;
 149         int  nfields;
 150         int  nr;
 151         struct aggr_cpu_id id;
 152         struct evsel *evsel;
 153 };
 154
 155 #define METRIC_LEN  35
 156
 157 static void new_line_std(struct perf_stat_config *config __maybe_unused,
 158                          void *ctx)
 159 {
 160         struct outstate *os = ctx;
 161
 162         os->newline = true;
 163 }
 164
 165 static void do_new_line_std(struct perf_stat_config *config,
 166                             struct outstate *os)
 167 {
 168         fputc('\n', os->fh);
 169         fputs(os->prefix, os->fh);
 170         aggr_printout(config, os->evsel, os->id, os->nr);
 171         if (config->aggr_mode == AGGR_NONE)
 172                 fprintf(os->fh, "        ");
 173         fprintf(os->fh, "                                                 ");
 174 }
 175
 176 static void print_metric_std(struct perf_stat_config *config,
 177                              void *ctx, const char *color, const char *fmt,
 178                              const char *unit, double val)
 179 {
 180         struct outstate *os = ctx;
 181         FILE *out = os->fh;
 182         int n;
 183         bool newline = os->newline;
 184
 185         os->newline = false;
 186
 187         if (unit == NULL || fmt == NULL) {
 188                 fprintf(out, "%-*s", METRIC_LEN, "");
 189                 return;
 190         }
 191
 192         if (newline)
 193                 do_new_line_std(config, os);
 194
 195         n = fprintf(out, " # ");
 196         if (color)
 197                 n += color_fprintf(out, color, fmt, val);
 198         else
 199                 n += fprintf(out, fmt, val);
 200         fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
 201 }
 202
 203 static void new_line_csv(struct perf_stat_config *config, void *ctx)
 204 {
 205         struct outstate *os = ctx;
 206         int i;
 207
 208         fputc('\n', os->fh);
 209         if (os->prefix)
 210                 fprintf(os->fh, "%s", os->prefix);
 211         aggr_printout(config, os->evsel, os->id, os->nr);
 212         for (i = 0; i < os->nfields; i++)
 213                 fputs(config->csv_sep, os->fh);
 214 }
 215
 216 static void print_metric_csv(struct perf_stat_config *config __maybe_unused,
 217                              void *ctx,
 218                              const char *color __maybe_unused,
 219                              const char *fmt, const char *unit, double val)
 220 {
 221         struct outstate *os = ctx;
 222         FILE *out = os->fh;
 223         char buf[64], *vals, *ends;
 224
 225         if (unit == NULL || fmt == NULL) {
 226                 fprintf(out, "%s%s", config->csv_sep, config->csv_sep);
 227                 return;
 228         }
 229         snprintf(buf, sizeof(buf), fmt, val);
 230         ends = vals = skip_spaces(buf);
 231         while (isdigit(*ends) || *ends == '.')
 232                 ends++;
 233         *ends = 0;
 234         fprintf(out, "%s%s%s%s", config->csv_sep, vals, config->csv_sep, skip_spaces(unit));
 235 }
 236
 237 /* Filter out some columns that don't work well in metrics only mode */
 238
 239 static bool valid_only_metric(const char *unit)
 240 {
 241         if (!unit)
 242                 return false;
 243         if (strstr(unit, "/sec") ||
 244             strstr(unit, "CPUs utilized"))
 245                 return false;
 246         return true;
 247 }
 248
 249 static const char *fixunit(char *buf, struct evsel *evsel,
 250                            const char *unit)
 251 {
 252         if (!strncmp(unit, "of all", 6)) {
 253                 snprintf(buf, 1024, "%s %s", evsel__name(evsel),
 254                          unit);
 255                 return buf;
 256         }
 257         return unit;
 258 }
 259
 260 static void print_metric_only(struct perf_stat_config *config,
 261                               void *ctx, const char *color, const char *fmt,
 262                               const char *unit, double val)
 263 {
 264         struct outstate *os = ctx;
 265         FILE *out = os->fh;
 266         char buf[1024], str[1024];
 267         unsigned mlen = config->metric_only_len;
 268
 269         if (!valid_only_metric(unit))
 270                 return;
 271         unit = fixunit(buf, os->evsel, unit);
 272         if (mlen < strlen(unit))
 273                 mlen = strlen(unit) + 1;
 274
 275         if (color)
 276                 mlen += strlen(color) + sizeof(PERF_COLOR_RESET) - 1;
 277
 278         color_snprintf(str, sizeof(str), color ?: "", fmt, val);
 279         fprintf(out, "%*s ", mlen, str);
 280 }
 281
 282 static void print_metric_only_csv(struct perf_stat_config *config __maybe_unused,
 283                                   void *ctx, const char *color __maybe_unused,
 284                                   const char *fmt,
 285                                   const char *unit, double val)
 286 {
 287         struct outstate *os = ctx;
 288         FILE *out = os->fh;
 289         char buf[64], *vals, *ends;
 290         char tbuf[1024];
 291
 292         if (!valid_only_metric(unit))
 293                 return;
 294         unit = fixunit(tbuf, os->evsel, unit);
 295         snprintf(buf, sizeof buf, fmt, val);
 296         ends = vals = skip_spaces(buf);
 297         while (isdigit(*ends) || *ends == '.')
 298                 ends++;
 299         *ends = 0;
 300         fprintf(out, "%s%s", vals, config->csv_sep);
 301 }
 302
 303 static void new_line_metric(struct perf_stat_config *config __maybe_unused,
 304                             void *ctx __maybe_unused)
 305 {
 306 }
 307
 308 static void print_metric_header(struct perf_stat_config *config,
 309                                 void *ctx, const char *color __maybe_unused,
 310                                 const char *fmt __maybe_unused,
 311                                 const char *unit, double val __maybe_unused)
 312 {
 313         struct outstate *os = ctx;
 314         char tbuf[1024];
 315
 316         /* In case of iostat, print metric header for first root port only */
 317         if (config->iostat_run &&
 318             os->evsel->priv != os->evsel->evlist->selected->priv)
 319                 return;
 320
 321         if (!valid_only_metric(unit))
 322                 return;
 323         unit = fixunit(tbuf, os->evsel, unit);
 324         if (config->csv_output)
 325                 fprintf(os->fh, "%s%s", unit, config->csv_sep);
 326         else
 327                 fprintf(os->fh, "%*s ", config->metric_only_len, unit);
 328 }
 329
 330 static int first_shadow_cpu(struct perf_stat_config *config,
 331                             struct evsel *evsel, struct aggr_cpu_id id)
 332 {
 333         struct evlist *evlist = evsel->evlist;
 334         int i;
 335
 336         if (config->aggr_mode == AGGR_NONE)
 337                 return id.core;
 338
 339         if (!config->aggr_get_id)
 340                 return 0;
 341
 342         for (i = 0; i < evsel__nr_cpus(evsel); i++) {
 343                 int cpu2 = evsel__cpus(evsel)->map[i];
 344
 345                 if (cpu_map__compare_aggr_cpu_id(
 346                                         config->aggr_get_id(config, evlist->core.cpus, cpu2),
 347                                         id)) {
 348                         return cpu2;
 349                 }
 350         }
 351         return 0;
 352 }
 353
 354 static void abs_printout(struct perf_stat_config *config,
 355                          struct aggr_cpu_id id, int nr, struct evsel *evsel, double avg)
 356 {
 357         FILE *output = config->output;
 358         double sc =  evsel->scale;
 359         const char *fmt;
 360
 361         if (config->csv_output) {
 362                 fmt = floor(sc) != sc ?  "%.2f%s" : "%.0f%s";
 363         } else {
 364                 if (config->big_num)
 365                         fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
 366                 else
 367                         fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
 368         }
 369
 370         aggr_printout(config, evsel, id, nr);
 371
 372         fprintf(output, fmt, avg, config->csv_sep);
 373
 374         if (evsel->unit)
 375                 fprintf(output, "%-*s%s",
 376                         config->csv_output ? 0 : config->unit_width,
 377                         evsel->unit, config->csv_sep);
 378
 379         fprintf(output, "%-*s", config->csv_output ? 0 : 25, evsel__name(evsel));
 380
 381         print_cgroup(config, evsel);
 382 }
 383
 384 static bool is_mixed_hw_group(struct evsel *counter)
 385 {
 386         struct evlist *evlist = counter->evlist;
 387         u32 pmu_type = counter->core.attr.type;
 388         struct evsel *pos;
 389
 390         if (counter->core.nr_members < 2)
 391                 return false;
 392
 393         evlist__for_each_entry(evlist, pos) {
 394                 /* software events can be part of any hardware group */
 395                 if (pos->core.attr.type == PERF_TYPE_SOFTWARE)
 396                         continue;
 397                 if (pmu_type == PERF_TYPE_SOFTWARE) {
 398                         pmu_type = pos->core.attr.type;
 399                         continue;
 400                 }
 401                 if (pmu_type != pos->core.attr.type)
 402                         return true;
 403         }
 404
 405         return false;
 406 }
 407
 408 static void printout(struct perf_stat_config *config, struct aggr_cpu_id id, int nr,
 409                      struct evsel *counter, double uval,
 410                      char *prefix, u64 run, u64 ena, double noise,
 411                      struct runtime_stat *st)
 412 {
 413         struct perf_stat_output_ctx out;
 414         struct outstate os = {
 415                 .fh = config->output,
 416                 .prefix = prefix ? prefix : "",
 417                 .id = id,
 418                 .nr = nr,
 419                 .evsel = counter,
 420         };
 421         print_metric_t pm = print_metric_std;
 422         new_line_t nl;
 423
 424         if (config->metric_only) {
 425                 nl = new_line_metric;
 426                 if (config->csv_output)
 427                         pm = print_metric_only_csv;
 428                 else
 429                         pm = print_metric_only;
 430         } else
 431                 nl = new_line_std;
 432
 433         if (config->csv_output && !config->metric_only) {
 434                 static int aggr_fields[] = {
 435                         [AGGR_GLOBAL] = 0,
 436                         [AGGR_THREAD] = 1,
 437                         [AGGR_NONE] = 1,
 438                         [AGGR_SOCKET] = 2,
 439                         [AGGR_DIE] = 2,
 440                         [AGGR_CORE] = 2,
 441                 };
 442
 443                 pm = print_metric_csv;
 444                 nl = new_line_csv;
 445                 os.nfields = 3;
 446                 os.nfields += aggr_fields[config->aggr_mode];
 447                 if (counter->cgrp)
 448                         os.nfields++;
 449         }
 450
 451         if (!config->no_csv_summary && config->csv_output &&
 452             config->summary && !config->interval) {
 453                 fprintf(config->output, "%16s%s", "summary", config->csv_sep);
 454         }
 455
 456         if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
 457                 if (config->metric_only) {
 458                         pm(config, &os, NULL, "", "", 0);
 459                         return;
 460                 }
 461                 aggr_printout(config, counter, id, nr);
 462
 463                 fprintf(config->output, "%*s%s",
 464                         config->csv_output ? 0 : 18,
 465                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
 466                         config->csv_sep);
 467
 468                 if (counter->supported) {
 469                         if (!evlist__has_hybrid(counter->evlist)) {
 470                                 config->print_free_counters_hint = 1;
 471                                 if (is_mixed_hw_group(counter))
 472                                         config->print_mixed_hw_group_error = 1;
 473                         }
 474                 }
 475
 476                 fprintf(config->output, "%-*s%s",
 477                         config->csv_output ? 0 : config->unit_width,
 478                         counter->unit, config->csv_sep);
 479
 480                 fprintf(config->output, "%*s",
 481                         config->csv_output ? 0 : -25, evsel__name(counter));
 482
 483                 print_cgroup(config, counter);
 484
 485                 if (!config->csv_output)
 486                         pm(config, &os, NULL, NULL, "", 0);
 487                 print_noise(config, counter, noise);
 488                 print_running(config, run, ena);
 489                 if (config->csv_output)
 490                         pm(config, &os, NULL, NULL, "", 0);
 491                 return;
 492         }
 493
 494         if (!config->metric_only)
 495                 abs_printout(config, id, nr, counter, uval);
 496
 497         out.print_metric = pm;
 498         out.new_line = nl;
 499         out.ctx = &os;
 500         out.force_header = false;
 501
 502         if (config->csv_output && !config->metric_only) {
 503                 print_noise(config, counter, noise);
 504                 print_running(config, run, ena);
 505         }
 506
 507         perf_stat__print_shadow_stats(config, counter, uval,
 508                                 first_shadow_cpu(config, counter, id),
 509                                 &out, &config->metric_events, st);
 510         if (!config->csv_output && !config->metric_only) {
 511                 print_noise(config, counter, noise);
 512                 print_running(config, run, ena);
 513         }
 514 }
 515
 516 static void aggr_update_shadow(struct perf_stat_config *config,
 517                                struct evlist *evlist)
 518 {
 519         int cpu, s;
 520         struct aggr_cpu_id s2, id;
 521         u64 val;
 522         struct evsel *counter;
 523
 524         for (s = 0; s < config->aggr_map->nr; s++) {
 525                 id = config->aggr_map->map[s];
 526                 evlist__for_each_entry(evlist, counter) {
 527                         val = 0;
 528                         for (cpu = 0; cpu < evsel__nr_cpus(counter); cpu++) {
 529                                 s2 = config->aggr_get_id(config, evlist->core.cpus, cpu);
 530                                 if (!cpu_map__compare_aggr_cpu_id(s2, id))
 531                                         continue;
 532                                 val += perf_counts(counter->counts, cpu, 0)->val;
 533                         }
 534                         perf_stat__update_shadow_stats(counter, val,
 535                                         first_shadow_cpu(config, counter, id),
 536                                         &rt_stat);
 537                 }
 538         }
 539 }
 540
 541 static void uniquify_event_name(struct evsel *counter)
 542 {
 543         char *new_name;
 544         char *config;
 545         int ret = 0;
 546
 547         if (counter->uniquified_name || counter->use_config_name ||
 548             !counter->pmu_name || !strncmp(counter->name, counter->pmu_name,
 549                                            strlen(counter->pmu_name)))
 550                 return;
 551
 552         config = strchr(counter->name, '/');
 553         if (config) {
 554                 if (asprintf(&new_name,
 555                              "%s%s", counter->pmu_name, config) > 0) {
 556                         free(counter->name);
 557                         counter->name = new_name;
 558                 }
 559         } else {
 560                 if (perf_pmu__has_hybrid()) {
 561                         ret = asprintf(&new_name, "%s/%s/",
 562                                        counter->pmu_name, counter->name);
 563                 } else {
 564                         ret = asprintf(&new_name, "%s [%s]",
 565                                        counter->name, counter->pmu_name);
 566                 }
 567
 568                 if (ret) {
 569                         free(counter->name);
 570                         counter->name = new_name;
 571                 }
 572         }
 573
 574         counter->uniquified_name = true;
 575 }
 576
 577 static void collect_all_aliases(struct perf_stat_config *config, struct evsel *counter,
 578                             void (*cb)(struct perf_stat_config *config, struct evsel *counter, void *data,
 579                                        bool first),
 580                             void *data)
 581 {
 582         struct evlist *evlist = counter->evlist;
 583         struct evsel *alias;
 584
 585         alias = list_prepare_entry(counter, &(evlist->core.entries), core.node);
 586         list_for_each_entry_continue (alias, &evlist->core.entries, core.node) {
 587                 /* Merge events with the same name, etc. but on different PMUs. */
 588                 if (!strcmp(evsel__name(alias), evsel__name(counter)) &&
 589                         alias->scale == counter->scale &&
 590                         alias->cgrp == counter->cgrp &&
 591                         !strcmp(alias->unit, counter->unit) &&
 592                         evsel__is_clock(alias) == evsel__is_clock(counter) &&
 593                         strcmp(alias->pmu_name, counter->pmu_name)) {
 594                         alias->merged_stat = true;
 595                         cb(config, alias, data, false);
 596                 }
 597         }
 598 }
 599
 600 static bool is_uncore(struct evsel *evsel)
 601 {
 602         struct perf_pmu *pmu = evsel__find_pmu(evsel);
 603
 604         return pmu && pmu->is_uncore;
 605 }
 606
 607 static bool hybrid_uniquify(struct evsel *evsel)
 608 {
 609         return perf_pmu__has_hybrid() && !is_uncore(evsel);
 610 }
 611
 612 static bool collect_data(struct perf_stat_config *config, struct evsel *counter,
 613                             void (*cb)(struct perf_stat_config *config, struct evsel *counter, void *data,
 614                                        bool first),
 615                             void *data)
 616 {
 617         if (counter->merged_stat)
 618                 return false;
 619         cb(config, counter, data, true);
 620         if (config->no_merge || hybrid_uniquify(counter))
 621                 uniquify_event_name(counter);
 622         else if (counter->auto_merge_stats)
 623                 collect_all_aliases(config, counter, cb, data);
 624         return true;
 625 }
 626
 627 struct aggr_data {
 628         u64 ena, run, val;
 629         struct aggr_cpu_id id;
 630         int nr;
 631         int cpu;
 632 };
 633
 634 static void aggr_cb(struct perf_stat_config *config,
 635                     struct evsel *counter, void *data, bool first)
 636 {
 637         struct aggr_data *ad = data;
 638         int cpu;
 639         struct aggr_cpu_id s2;
 640
 641         for (cpu = 0; cpu < evsel__nr_cpus(counter); cpu++) {
 642                 struct perf_counts_values *counts;
 643
 644                 s2 = config->aggr_get_id(config, evsel__cpus(counter), cpu);
 645                 if (!cpu_map__compare_aggr_cpu_id(s2, ad->id))
 646                         continue;
 647                 if (first)
 648                         ad->nr++;
 649                 counts = perf_counts(counter->counts, cpu, 0);
 650                 /*
 651                  * When any result is bad, make them all to give
 652                  * consistent output in interval mode.
 653                  */
 654                 if (counts->ena == 0 || counts->run == 0 ||
 655                     counter->counts->scaled == -1) {
 656                         ad->ena = 0;
 657                         ad->run = 0;
 658                         break;
 659                 }
 660                 ad->val += counts->val;
 661                 ad->ena += counts->ena;
 662                 ad->run += counts->run;
 663         }
 664 }
 665
 666 static void print_counter_aggrdata(struct perf_stat_config *config,
 667                                    struct evsel *counter, int s,
 668                                    char *prefix, bool metric_only,
 669                                    bool *first, int cpu)
 670 {
 671         struct aggr_data ad;
 672         FILE *output = config->output;
 673         u64 ena, run, val;
 674         int nr;
 675         struct aggr_cpu_id id;
 676         double uval;
 677
 678         ad.id = id = config->aggr_map->map[s];
 679         ad.val = ad.ena = ad.run = 0;
 680         ad.nr = 0;
 681         if (!collect_data(config, counter, aggr_cb, &ad))
 682                 return;
 683
 684         if (perf_pmu__has_hybrid() && ad.ena == 0)
 685                 return;
 686
 687         nr = ad.nr;
 688         ena = ad.ena;
 689         run = ad.run;
 690         val = ad.val;
 691         if (*first && metric_only) {
 692                 *first = false;
 693                 aggr_printout(config, counter, id, nr);
 694         }
 695         if (prefix && !metric_only)
 696                 fprintf(output, "%s", prefix);
 697
 698         uval = val * counter->scale;
 699         if (cpu != -1) {
 700                 id = cpu_map__empty_aggr_cpu_id();
 701                 id.core = cpu;
 702         }
 703         printout(config, id, nr, counter, uval,
 704                  prefix, run, ena, 1.0, &rt_stat);
 705         if (!metric_only)
 706                 fputc('\n', output);
 707 }
 708
 709 static void print_aggr(struct perf_stat_config *config,
 710                        struct evlist *evlist,
 711                        char *prefix)
 712 {
 713         bool metric_only = config->metric_only;
 714         FILE *output = config->output;
 715         struct evsel *counter;
 716         int s;
 717         bool first;
 718
 719         if (!config->aggr_map || !config->aggr_get_id)
 720                 return;
 721
 722         aggr_update_shadow(config, evlist);
 723
 724         /*
 725          * With metric_only everything is on a single line.
 726          * Without each counter has its own line.
 727          */
 728         for (s = 0; s < config->aggr_map->nr; s++) {
 729                 if (prefix && metric_only)
 730                         fprintf(output, "%s", prefix);
 731
 732                 first = true;
 733                 evlist__for_each_entry(evlist, counter) {
 734                         print_counter_aggrdata(config, counter, s,
 735                                                prefix, metric_only,
 736                                                &first, -1);
 737                 }
 738                 if (metric_only)
 739                         fputc('\n', output);
 740         }
 741 }
 742
 743 static int cmp_val(const void *a, const void *b)
 744 {
 745         return ((struct perf_aggr_thread_value *)b)->val -
 746                 ((struct perf_aggr_thread_value *)a)->val;
 747 }
 748
 749 static struct perf_aggr_thread_value *sort_aggr_thread(
 750                                         struct evsel *counter,
 751                                         int nthreads, int ncpus,
 752                                         int *ret,
 753                                         struct target *_target)
 754 {
 755         int cpu, thread, i = 0;
 756         double uval;
 757         struct perf_aggr_thread_value *buf;
 758
 759         buf = calloc(nthreads, sizeof(struct perf_aggr_thread_value));
 760         if (!buf)
 761                 return NULL;
 762
 763         for (thread = 0; thread < nthreads; thread++) {
 764                 u64 ena = 0, run = 0, val = 0;
 765
 766                 for (cpu = 0; cpu < ncpus; cpu++) {
 767                         val += perf_counts(counter->counts, cpu, thread)->val;
 768                         ena += perf_counts(counter->counts, cpu, thread)->ena;
 769                         run += perf_counts(counter->counts, cpu, thread)->run;
 770                 }
 771
 772                 uval = val * counter->scale;
 773
 774                 /*
 775                  * Skip value 0 when enabling --per-thread globally,
 776                  * otherwise too many 0 output.
 777                  */
 778                 if (uval == 0.0 && target__has_per_thread(_target))
 779                         continue;
 780
 781                 buf[i].counter = counter;
 782                 buf[i].id = cpu_map__empty_aggr_cpu_id();
 783                 buf[i].id.thread = thread;
 784                 buf[i].uval = uval;
 785                 buf[i].val = val;
 786                 buf[i].run = run;
 787                 buf[i].ena = ena;
 788                 i++;
 789         }
 790
 791         qsort(buf, i, sizeof(struct perf_aggr_thread_value), cmp_val);
 792
 793         if (ret)
 794                 *ret = i;
 795
 796         return buf;
 797 }
 798
 799 static void print_aggr_thread(struct perf_stat_config *config,
 800                               struct target *_target,
 801                               struct evsel *counter, char *prefix)
 802 {
 803         FILE *output = config->output;
 804         int nthreads = perf_thread_map__nr(counter->core.threads);
 805         int ncpus = perf_cpu_map__nr(counter->core.cpus);
 806         int thread, sorted_threads;
 807         struct aggr_cpu_id id;
 808         struct perf_aggr_thread_value *buf;
 809
 810         buf = sort_aggr_thread(counter, nthreads, ncpus, &sorted_threads, _target);
 811         if (!buf) {
 812                 perror("cannot sort aggr thread");
 813                 return;
 814         }
 815
 816         for (thread = 0; thread < sorted_threads; thread++) {
 817                 if (prefix)
 818                         fprintf(output, "%s", prefix);
 819
 820                 id = buf[thread].id;
 821                 if (config->stats)
 822                         printout(config, id, 0, buf[thread].counter, buf[thread].uval,
 823                                  prefix, buf[thread].run, buf[thread].ena, 1.0,
 824                                  &config->stats[id.thread]);
 825                 else
 826                         printout(config, id, 0, buf[thread].counter, buf[thread].uval,
 827                                  prefix, buf[thread].run, buf[thread].ena, 1.0,
 828                                  &rt_stat);
 829                 fputc('\n', output);
 830         }
 831
 832         free(buf);
 833 }
 834
 835 struct caggr_data {
 836         double avg, avg_enabled, avg_running;
 837 };
 838
 839 static void counter_aggr_cb(struct perf_stat_config *config __maybe_unused,
 840                             struct evsel *counter, void *data,
 841                             bool first __maybe_unused)
 842 {
 843         struct caggr_data *cd = data;
 844         struct perf_counts_values *aggr = &counter->counts->aggr;
 845
 846         cd->avg += aggr->val;
 847         cd->avg_enabled += aggr->ena;
 848         cd->avg_running += aggr->run;
 849 }
 850
 851 /*
 852  * Print out the results of a single counter:
 853  * aggregated counts in system-wide mode
 854  */
 855 static void print_counter_aggr(struct perf_stat_config *config,
 856                                struct evsel *counter, char *prefix)
 857 {
 858         bool metric_only = config->metric_only;
 859         FILE *output = config->output;
 860         double uval;
 861         struct caggr_data cd = { .avg = 0.0 };
 862
 863         if (!collect_data(config, counter, counter_aggr_cb, &cd))
 864                 return;
 865
 866         if (prefix && !metric_only)
 867                 fprintf(output, "%s", prefix);
 868
 869         uval = cd.avg * counter->scale;
 870         printout(config, cpu_map__empty_aggr_cpu_id(), 0, counter, uval, prefix, cd.avg_running,
 871                  cd.avg_enabled, cd.avg, &rt_stat);
 872         if (!metric_only)
 873                 fprintf(output, "\n");
 874 }
 875
 876 static void counter_cb(struct perf_stat_config *config __maybe_unused,
 877                        struct evsel *counter, void *data,
 878                        bool first __maybe_unused)
 879 {
 880         struct aggr_data *ad = data;
 881
 882         ad->val += perf_counts(counter->counts, ad->cpu, 0)->val;
 883         ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena;
 884         ad->run += perf_counts(counter->counts, ad->cpu, 0)->run;
 885 }
 886
 887 /*
 888  * Print out the results of a single counter:
 889  * does not use aggregated count in system-wide
 890  */
 891 static void print_counter(struct perf_stat_config *config,
 892                           struct evsel *counter, char *prefix)
 893 {
 894         FILE *output = config->output;
 895         u64 ena, run, val;
 896         double uval;
 897         int cpu;
 898         struct aggr_cpu_id id;
 899
 900         for (cpu = 0; cpu < evsel__nr_cpus(counter); cpu++) {
 901                 struct aggr_data ad = { .cpu = cpu };
 902
 903                 if (!collect_data(config, counter, counter_cb, &ad))
 904                         return;
 905                 val = ad.val;
 906                 ena = ad.ena;
 907                 run = ad.run;
 908
 909                 if (prefix)
 910                         fprintf(output, "%s", prefix);
 911
 912                 uval = val * counter->scale;
 913                 id = cpu_map__empty_aggr_cpu_id();
 914                 id.core = cpu;
 915                 printout(config, id, 0, counter, uval, prefix,
 916                          run, ena, 1.0, &rt_stat);
 917
 918                 fputc('\n', output);
 919         }
 920 }
 921
 922 static void print_no_aggr_metric(struct perf_stat_config *config,
 923                                  struct evlist *evlist,
 924                                  char *prefix)
 925 {
 926         int cpu;
 927         int nrcpus = 0;
 928         struct evsel *counter;
 929         u64 ena, run, val;
 930         double uval;
 931         struct aggr_cpu_id id;
 932
 933         nrcpus = evlist->core.cpus->nr;
 934         for (cpu = 0; cpu < nrcpus; cpu++) {
 935                 bool first = true;
 936
 937                 if (prefix)
 938                         fputs(prefix, config->output);
 939                 evlist__for_each_entry(evlist, counter) {
 940                         id = cpu_map__empty_aggr_cpu_id();
 941                         id.core = cpu;
 942                         if (first) {
 943                                 aggr_printout(config, counter, id, 0);
 944                                 first = false;
 945                         }
 946                         val = perf_counts(counter->counts, cpu, 0)->val;
 947                         ena = perf_counts(counter->counts, cpu, 0)->ena;
 948                         run = perf_counts(counter->counts, cpu, 0)->run;
 949
 950                         uval = val * counter->scale;
 951                         printout(config, id, 0, counter, uval, prefix,
 952                                  run, ena, 1.0, &rt_stat);
 953                 }
 954                 fputc('\n', config->output);
 955         }
 956 }
 957
 958 static int aggr_header_lens[] = {
 959         [AGGR_CORE] = 24,
 960         [AGGR_DIE] = 18,
 961         [AGGR_SOCKET] = 12,
 962         [AGGR_NONE] = 6,
 963         [AGGR_THREAD] = 24,
 964         [AGGR_GLOBAL] = 0,
 965 };
 966
 967 static const char *aggr_header_csv[] = {
 968         [AGGR_CORE]     =       "core,cpus,",
 969         [AGGR_DIE]      =       "die,cpus",
 970         [AGGR_SOCKET]   =       "socket,cpus",
 971         [AGGR_NONE]     =       "cpu,",
 972         [AGGR_THREAD]   =       "comm-pid,",
 973         [AGGR_GLOBAL]   =       ""
 974 };
 975
 976 static void print_metric_headers(struct perf_stat_config *config,
 977                                  struct evlist *evlist,
 978                                  const char *prefix, bool no_indent)
 979 {
 980         struct perf_stat_output_ctx out;
 981         struct evsel *counter;
 982         struct outstate os = {
 983                 .fh = config->output
 984         };
 985
 986         if (prefix)
 987                 fprintf(config->output, "%s", prefix);
 988
 989         if (!config->csv_output && !no_indent)
 990                 fprintf(config->output, "%*s",
 991                         aggr_header_lens[config->aggr_mode], "");
 992         if (config->csv_output) {
 993                 if (config->interval)
 994                         fputs("time,", config->output);
 995                 if (!config->iostat_run)
 996                         fputs(aggr_header_csv[config->aggr_mode], config->output);
 997         }
 998         if (config->iostat_run)
 999                 iostat_print_header_prefix(config);
1000
1001         /* Print metrics headers only */
1002         evlist__for_each_entry(evlist, counter) {
1003                 os.evsel = counter;
1004                 out.ctx = &os;
1005                 out.print_metric = print_metric_header;
1006                 out.new_line = new_line_metric;
1007                 out.force_header = true;
1008                 perf_stat__print_shadow_stats(config, counter, 0,
1009                                               0,
1010                                               &out,
1011                                               &config->metric_events,
1012                                               &rt_stat);
1013         }
1014         fputc('\n', config->output);
1015 }
1016
1017 static void print_interval(struct perf_stat_config *config,
1018                            struct evlist *evlist,
1019                            char *prefix, struct timespec *ts)
1020 {
1021         bool metric_only = config->metric_only;
1022         unsigned int unit_width = config->unit_width;
1023         FILE *output = config->output;
1024         static int num_print_interval;
1025
1026         if (config->interval_clear)
1027                 puts(CONSOLE_CLEAR);
1028
1029         if (!config->iostat_run)
1030                 sprintf(prefix, "%6lu.%09lu%s", (unsigned long) ts->tv_sec, ts->tv_nsec, config->csv_sep);
1031
1032         if ((num_print_interval == 0 && !config->csv_output) || config->interval_clear) {
1033                 switch (config->aggr_mode) {
1034                 case AGGR_NODE:
1035                         fprintf(output, "#           time node   cpus");
1036                         if (!metric_only)
1037                                 fprintf(output, "             counts %*s events\n", unit_width, "unit");
1038                         break;
1039                 case AGGR_SOCKET:
1040                         fprintf(output, "#           time socket cpus");
1041                         if (!metric_only)
1042                                 fprintf(output, "             counts %*s events\n", unit_width, "unit");
1043                         break;
1044                 case AGGR_DIE:
1045                         fprintf(output, "#           time die          cpus");
1046                         if (!metric_only)
1047                                 fprintf(output, "             counts %*s events\n", unit_width, "unit");
1048                         break;
1049                 case AGGR_CORE:
1050                         fprintf(output, "#           time core            cpus");
1051                         if (!metric_only)
1052                                 fprintf(output, "             counts %*s events\n", unit_width, "unit");
1053                         break;
1054                 case AGGR_NONE:
1055                         fprintf(output, "#           time CPU    ");
1056                         if (!metric_only)
1057                                 fprintf(output, "                counts %*s events\n", unit_width, "unit");
1058                         break;
1059                 case AGGR_THREAD:
1060                         fprintf(output, "#           time             comm-pid");
1061                         if (!metric_only)
1062                                 fprintf(output, "                  counts %*s events\n", unit_width, "unit");
1063                         break;
1064                 case AGGR_GLOBAL:
1065                 default:
1066                         if (!config->iostat_run) {
1067                                 fprintf(output, "#           time");
1068                                 if (!metric_only)
1069                                         fprintf(output, "             counts %*s events\n", unit_width, "unit");
1070                         }
1071                 case AGGR_UNSET:
1072                         break;
1073                 }
1074         }
1075
1076         if ((num_print_interval == 0 || config->interval_clear) && metric_only)
1077                 print_metric_headers(config, evlist, " ", true);
1078         if (++num_print_interval == 25)
1079                 num_print_interval = 0;
1080 }
1081
1082 static void print_header(struct perf_stat_config *config,
1083                          struct target *_target,
1084                          int argc, const char **argv)
1085 {
1086         FILE *output = config->output;
1087         int i;
1088
1089         fflush(stdout);
1090
1091         if (!config->csv_output) {
1092                 fprintf(output, "\n");
1093                 fprintf(output, " Performance counter stats for ");
1094                 if (_target->bpf_str)
1095                         fprintf(output, "\'BPF program(s) %s", _target->bpf_str);
1096                 else if (_target->system_wide)
1097                         fprintf(output, "\'system wide");
1098                 else if (_target->cpu_list)
1099                         fprintf(output, "\'CPU(s) %s", _target->cpu_list);
1100                 else if (!target__has_task(_target)) {
1101                         fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1102                         for (i = 1; argv && (i < argc); i++)
1103                                 fprintf(output, " %s", argv[i]);
1104                 } else if (_target->pid)
1105                         fprintf(output, "process id \'%s", _target->pid);
1106                 else
1107                         fprintf(output, "thread id \'%s", _target->tid);
1108
1109                 fprintf(output, "\'");
1110                 if (config->run_count > 1)
1111                         fprintf(output, " (%d runs)", config->run_count);
1112                 fprintf(output, ":\n\n");
1113         }
1114 }
1115
1116 static int get_precision(double num)
1117 {
1118         if (num > 1)
1119                 return 0;
1120
1121         return lround(ceil(-log10(num)));
1122 }
1123
1124 static void print_table(struct perf_stat_config *config,
1125                         FILE *output, int precision, double avg)
1126 {
1127         char tmp[64];
1128         int idx, indent = 0;
1129
1130         scnprintf(tmp, 64, " %17.*f", precision, avg);
1131         while (tmp[indent] == ' ')
1132                 indent++;
1133
1134         fprintf(output, "%*s# Table of individual measurements:\n", indent, "");
1135
1136         for (idx = 0; idx < config->run_count; idx++) {
1137                 double run = (double) config->walltime_run[idx] / NSEC_PER_SEC;
1138                 int h, n = 1 + abs((int) (100.0 * (run - avg)/run) / 5);
1139
1140                 fprintf(output, " %17.*f (%+.*f) ",
1141                         precision, run, precision, run - avg);
1142
1143                 for (h = 0; h < n; h++)
1144                         fprintf(output, "#");
1145
1146                 fprintf(output, "\n");
1147         }
1148
1149         fprintf(output, "\n%*s# Final result:\n", indent, "");
1150 }
1151
1152 static double timeval2double(struct timeval *t)
1153 {
1154         return t->tv_sec + (double) t->tv_usec/USEC_PER_SEC;
1155 }
1156
1157 static void print_footer(struct perf_stat_config *config)
1158 {
1159         double avg = avg_stats(config->walltime_nsecs_stats) / NSEC_PER_SEC;
1160         FILE *output = config->output;
1161
1162         if (!config->null_run)
1163                 fprintf(output, "\n");
1164
1165         if (config->run_count == 1) {
1166                 fprintf(output, " %17.9f seconds time elapsed", avg);
1167
1168                 if (config->ru_display) {
1169                         double ru_utime = timeval2double(&config->ru_data.ru_utime);
1170                         double ru_stime = timeval2double(&config->ru_data.ru_stime);
1171
1172                         fprintf(output, "\n\n");
1173                         fprintf(output, " %17.9f seconds user\n", ru_utime);
1174                         fprintf(output, " %17.9f seconds sys\n", ru_stime);
1175                 }
1176         } else {
1177                 double sd = stddev_stats(config->walltime_nsecs_stats) / NSEC_PER_SEC;
1178                 /*
1179                  * Display at most 2 more significant
1180                  * digits than the stddev inaccuracy.
1181                  */
1182                 int precision = get_precision(sd) + 2;
1183
1184                 if (config->walltime_run_table)
1185                         print_table(config, output, precision, avg);
1186
1187                 fprintf(output, " %17.*f +- %.*f seconds time elapsed",
1188                         precision, avg, precision, sd);
1189
1190                 print_noise_pct(config, sd, avg);
1191         }
1192         fprintf(output, "\n\n");
1193
1194         if (config->print_free_counters_hint && sysctl__nmi_watchdog_enabled())
1195                 fprintf(output,
1196 "Some events weren't counted. Try disabling the NMI watchdog:\n"
1197 "       echo 0 > /proc/sys/kernel/nmi_watchdog\n"
1198 "       perf stat ...\n"
1199 "       echo 1 > /proc/sys/kernel/nmi_watchdog\n");
1200
1201         if (config->print_mixed_hw_group_error)
1202                 fprintf(output,
1203                         "The events in group usually have to be from "
1204                         "the same PMU. Try reorganizing the group.\n");
1205 }
1206
1207 static void print_percore_thread(struct perf_stat_config *config,
1208                                  struct evsel *counter, char *prefix)
1209 {
1210         int s;
1211         struct aggr_cpu_id s2, id;
1212         bool first = true;
1213
1214         for (int i = 0; i < evsel__nr_cpus(counter); i++) {
1215                 s2 = config->aggr_get_id(config, evsel__cpus(counter), i);
1216                 for (s = 0; s < config->aggr_map->nr; s++) {
1217                         id = config->aggr_map->map[s];
1218                         if (cpu_map__compare_aggr_cpu_id(s2, id))
1219                                 break;
1220                 }
1221
1222                 print_counter_aggrdata(config, counter, s,
1223                                        prefix, false,
1224                                        &first, i);
1225         }
1226 }
1227
1228 static void print_percore(struct perf_stat_config *config,
1229                           struct evsel *counter, char *prefix)
1230 {
1231         bool metric_only = config->metric_only;
1232         FILE *output = config->output;
1233         int s;
1234         bool first = true;
1235
1236         if (!config->aggr_map || !config->aggr_get_id)
1237                 return;
1238
1239         if (config->percore_show_thread)
1240                 return print_percore_thread(config, counter, prefix);
1241
1242         for (s = 0; s < config->aggr_map->nr; s++) {
1243                 if (prefix && metric_only)
1244                         fprintf(output, "%s", prefix);
1245
1246                 print_counter_aggrdata(config, counter, s,
1247                                        prefix, metric_only,
1248                                        &first, -1);
1249         }
1250
1251         if (metric_only)
1252                 fputc('\n', output);
1253 }
1254
1255 void evlist__print_counters(struct evlist *evlist, struct perf_stat_config *config,
1256                             struct target *_target, struct timespec *ts, int argc, const char **argv)
1257 {
1258         bool metric_only = config->metric_only;
1259         int interval = config->interval;
1260         struct evsel *counter;
1261         char buf[64], *prefix = NULL;
1262
1263         if (config->iostat_run)
1264                 evlist->selected = evlist__first(evlist);
1265
1266         if (interval)
1267                 print_interval(config, evlist, prefix = buf, ts);
1268         else
1269                 print_header(config, _target, argc, argv);
1270
1271         if (metric_only) {
1272                 static int num_print_iv;
1273
1274                 if (num_print_iv == 0 && !interval)
1275                         print_metric_headers(config, evlist, prefix, false);
1276                 if (num_print_iv++ == 25)
1277                         num_print_iv = 0;
1278                 if (config->aggr_mode == AGGR_GLOBAL && prefix && !config->iostat_run)
1279                         fprintf(config->output, "%s", prefix);
1280         }
1281
1282         switch (config->aggr_mode) {
1283         case AGGR_CORE:
1284         case AGGR_DIE:
1285         case AGGR_SOCKET:
1286         case AGGR_NODE:
1287                 print_aggr(config, evlist, prefix);
1288                 break;
1289         case AGGR_THREAD:
1290                 evlist__for_each_entry(evlist, counter) {
1291                         print_aggr_thread(config, _target, counter, prefix);
1292                 }
1293                 break;
1294         case AGGR_GLOBAL:
1295                 if (config->iostat_run)
1296                         iostat_print_counters(evlist, config, ts, prefix = buf,
1297                                               print_counter_aggr);
1298                 else {
1299                         evlist__for_each_entry(evlist, counter) {
1300                                 print_counter_aggr(config, counter, prefix);
1301                         }
1302                         if (metric_only)
1303                                 fputc('\n', config->output);
1304                 }
1305                 break;
1306         case AGGR_NONE:
1307                 if (metric_only)
1308                         print_no_aggr_metric(config, evlist, prefix);
1309                 else {
1310                         evlist__for_each_entry(evlist, counter) {
1311                                 if (counter->percore)
1312                                         print_percore(config, counter, prefix);
1313                                 else
1314                                         print_counter(config, counter, prefix);
1315                         }
1316                 }
1317                 break;
1318         case AGGR_UNSET:
1319         default:
1320                 break;
1321         }
1322
1323         if (!interval && !config->csv_output)
1324                 print_footer(config);
1325
1326         fflush(config->output);
1327 }