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