tools/perf/util/stat-display.c

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