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