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 <perf/cpumap.h> 8 #include "color.h" 9 #include "counts.h" 10 #include "evlist.h" 11 #include "evsel.h" 12 #include "stat.h" 13 #include "top.h" 14 #include "thread_map.h" 15 #include "cpumap.h" 16 #include "string2.h" 17 #include <linux/ctype.h> 18 #include "cgroup.h" 19 #include <api/fs/fs.h> 20 #include "util.h" 21 #include "iostat.h" 22 #include "pmu-hybrid.h" 23 #include "evlist-hybrid.h" 24 25 #define CNTR_NOT_SUPPORTED "<not supported>" 26 #define CNTR_NOT_COUNTED "<not counted>" 27 28 static void print_running(struct perf_stat_config *config, 29 u64 run, u64 ena) 30 { 31 if (config->csv_output) { 32 fprintf(config->output, "%s%" PRIu64 "%s%.2f", 33 config->csv_sep, 34 run, 35 config->csv_sep, 36 ena ? 100.0 * run / ena : 100.0); 37 } else if (run != ena) { 38 fprintf(config->output, " (%.2f%%)", 100.0 * run / ena); 39 } 40 } 41 42 static void print_noise_pct(struct perf_stat_config *config, 43 double total, double avg) 44 { 45 double pct = rel_stddev_stats(total, avg); 46 47 if (config->csv_output) 48 fprintf(config->output, "%s%.2f%%", config->csv_sep, pct); 49 else if (pct) 50 fprintf(config->output, " ( +-%6.2f%% )", pct); 51 } 52 53 static void print_noise(struct perf_stat_config *config, 54 struct evsel *evsel, double avg) 55 { 56 struct perf_stat_evsel *ps; 57 58 if (config->run_count == 1) 59 return; 60 61 ps = evsel->stats; 62 print_noise_pct(config, stddev_stats(&ps->res_stats[0]), avg); 63 } 64 65 static void print_cgroup(struct perf_stat_config *config, struct evsel *evsel) 66 { 67 if (nr_cgroups) { 68 const char *cgrp_name = evsel->cgrp ? evsel->cgrp->name : ""; 69 fprintf(config->output, "%s%s", config->csv_sep, cgrp_name); 70 } 71 } 72 73 74 static void aggr_printout(struct perf_stat_config *config, 75 struct evsel *evsel, struct aggr_cpu_id id, int nr) 76 { 77 switch (config->aggr_mode) { 78 case AGGR_CORE: 79 fprintf(config->output, "S%d-D%d-C%*d%s%*d%s", 80 id.socket, 81 id.die, 82 config->csv_output ? 0 : -8, 83 id.core, 84 config->csv_sep, 85 config->csv_output ? 0 : 4, 86 nr, 87 config->csv_sep); 88 break; 89 case AGGR_DIE: 90 fprintf(config->output, "S%d-D%*d%s%*d%s", 91 id.socket, 92 config->csv_output ? 0 : -8, 93 id.die, 94 config->csv_sep, 95 config->csv_output ? 0 : 4, 96 nr, 97 config->csv_sep); 98 break; 99 case AGGR_SOCKET: 100 fprintf(config->output, "S%*d%s%*d%s", 101 config->csv_output ? 0 : -5, 102 id.socket, 103 config->csv_sep, 104 config->csv_output ? 0 : 4, 105 nr, 106 config->csv_sep); 107 break; 108 case AGGR_NODE: 109 fprintf(config->output, "N%*d%s%*d%s", 110 config->csv_output ? 0 : -5, 111 id.node, 112 config->csv_sep, 113 config->csv_output ? 0 : 4, 114 nr, 115 config->csv_sep); 116 break; 117 case AGGR_NONE: 118 if (evsel->percore && !config->percore_show_thread) { 119 fprintf(config->output, "S%d-D%d-C%*d%s", 120 id.socket, 121 id.die, 122 config->csv_output ? 0 : -3, 123 id.core, config->csv_sep); 124 } else if (id.cpu.cpu > -1) { 125 fprintf(config->output, "CPU%*d%s", 126 config->csv_output ? 0 : -7, 127 id.cpu.cpu, config->csv_sep); 128 } 129 break; 130 case AGGR_THREAD: 131 fprintf(config->output, "%*s-%*d%s", 132 config->csv_output ? 0 : 16, 133 perf_thread_map__comm(evsel->core.threads, id.thread), 134 config->csv_output ? 0 : -8, 135 perf_thread_map__pid(evsel->core.threads, id.thread), 136 config->csv_sep); 137 break; 138 case AGGR_GLOBAL: 139 case AGGR_UNSET: 140 default: 141 break; 142 } 143 } 144 145 struct outstate { 146 FILE *fh; 147 bool newline; 148 const char *prefix; 149 int nfields; 150 int nr; 151 struct aggr_cpu_id id; 152 struct evsel *evsel; 153 }; 154 155 #define METRIC_LEN 35 156 157 static void new_line_std(struct perf_stat_config *config __maybe_unused, 158 void *ctx) 159 { 160 struct outstate *os = ctx; 161 162 os->newline = true; 163 } 164 165 static void do_new_line_std(struct perf_stat_config *config, 166 struct outstate *os) 167 { 168 fputc('\n', os->fh); 169 fputs(os->prefix, os->fh); 170 aggr_printout(config, os->evsel, os->id, os->nr); 171 if (config->aggr_mode == AGGR_NONE) 172 fprintf(os->fh, " "); 173 fprintf(os->fh, " "); 174 } 175 176 static void print_metric_std(struct perf_stat_config *config, 177 void *ctx, const char *color, const char *fmt, 178 const char *unit, double val) 179 { 180 struct outstate *os = ctx; 181 FILE *out = os->fh; 182 int n; 183 bool newline = os->newline; 184 185 os->newline = false; 186 187 if (unit == NULL || fmt == NULL) { 188 fprintf(out, "%-*s", METRIC_LEN, ""); 189 return; 190 } 191 192 if (newline) 193 do_new_line_std(config, os); 194 195 n = fprintf(out, " # "); 196 if (color) 197 n += color_fprintf(out, color, fmt, val); 198 else 199 n += fprintf(out, fmt, val); 200 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit); 201 } 202 203 static void new_line_csv(struct perf_stat_config *config, void *ctx) 204 { 205 struct outstate *os = ctx; 206 int i; 207 208 fputc('\n', os->fh); 209 if (os->prefix) 210 fprintf(os->fh, "%s%s", os->prefix, config->csv_sep); 211 aggr_printout(config, os->evsel, os->id, os->nr); 212 for (i = 0; i < os->nfields; i++) 213 fputs(config->csv_sep, os->fh); 214 } 215 216 static void print_metric_csv(struct perf_stat_config *config __maybe_unused, 217 void *ctx, 218 const char *color __maybe_unused, 219 const char *fmt, const char *unit, double val) 220 { 221 struct outstate *os = ctx; 222 FILE *out = os->fh; 223 char buf[64], *vals, *ends; 224 225 if (unit == NULL || fmt == NULL) { 226 fprintf(out, "%s%s", config->csv_sep, config->csv_sep); 227 return; 228 } 229 snprintf(buf, sizeof(buf), fmt, val); 230 ends = vals = skip_spaces(buf); 231 while (isdigit(*ends) || *ends == '.') 232 ends++; 233 *ends = 0; 234 fprintf(out, "%s%s%s%s", config->csv_sep, vals, config->csv_sep, skip_spaces(unit)); 235 } 236 237 /* Filter out some columns that don't work well in metrics only mode */ 238 239 static bool valid_only_metric(const char *unit) 240 { 241 if (!unit) 242 return false; 243 if (strstr(unit, "/sec") || 244 strstr(unit, "CPUs utilized")) 245 return false; 246 return true; 247 } 248 249 static const char *fixunit(char *buf, struct evsel *evsel, 250 const char *unit) 251 { 252 if (!strncmp(unit, "of all", 6)) { 253 snprintf(buf, 1024, "%s %s", evsel__name(evsel), 254 unit); 255 return buf; 256 } 257 return unit; 258 } 259 260 static void print_metric_only(struct perf_stat_config *config, 261 void *ctx, const char *color, const char *fmt, 262 const char *unit, double val) 263 { 264 struct outstate *os = ctx; 265 FILE *out = os->fh; 266 char buf[1024], str[1024]; 267 unsigned mlen = config->metric_only_len; 268 269 if (!valid_only_metric(unit)) 270 return; 271 unit = fixunit(buf, os->evsel, unit); 272 if (mlen < strlen(unit)) 273 mlen = strlen(unit) + 1; 274 275 if (color) 276 mlen += strlen(color) + sizeof(PERF_COLOR_RESET) - 1; 277 278 color_snprintf(str, sizeof(str), color ?: "", fmt, val); 279 fprintf(out, "%*s ", mlen, str); 280 } 281 282 static void print_metric_only_csv(struct perf_stat_config *config __maybe_unused, 283 void *ctx, const char *color __maybe_unused, 284 const char *fmt, 285 const char *unit, double val) 286 { 287 struct outstate *os = ctx; 288 FILE *out = os->fh; 289 char buf[64], *vals, *ends; 290 char tbuf[1024]; 291 292 if (!valid_only_metric(unit)) 293 return; 294 unit = fixunit(tbuf, os->evsel, unit); 295 snprintf(buf, sizeof buf, fmt, val); 296 ends = vals = skip_spaces(buf); 297 while (isdigit(*ends) || *ends == '.') 298 ends++; 299 *ends = 0; 300 fprintf(out, "%s%s", vals, config->csv_sep); 301 } 302 303 static void new_line_metric(struct perf_stat_config *config __maybe_unused, 304 void *ctx __maybe_unused) 305 { 306 } 307 308 static void print_metric_header(struct perf_stat_config *config, 309 void *ctx, const char *color __maybe_unused, 310 const char *fmt __maybe_unused, 311 const char *unit, double val __maybe_unused) 312 { 313 struct outstate *os = ctx; 314 char tbuf[1024]; 315 316 /* In case of iostat, print metric header for first root port only */ 317 if (config->iostat_run && 318 os->evsel->priv != os->evsel->evlist->selected->priv) 319 return; 320 321 if (!valid_only_metric(unit)) 322 return; 323 unit = fixunit(tbuf, os->evsel, unit); 324 if (config->csv_output) 325 fprintf(os->fh, "%s%s", unit, config->csv_sep); 326 else 327 fprintf(os->fh, "%*s ", config->metric_only_len, unit); 328 } 329 330 static int first_shadow_cpu_map_idx(struct perf_stat_config *config, 331 struct evsel *evsel, const struct aggr_cpu_id *id) 332 { 333 struct perf_cpu_map *cpus = evsel__cpus(evsel); 334 struct perf_cpu cpu; 335 int idx; 336 337 if (config->aggr_mode == AGGR_NONE) 338 return perf_cpu_map__idx(cpus, id->cpu); 339 340 if (!config->aggr_get_id) 341 return 0; 342 343 perf_cpu_map__for_each_cpu(cpu, idx, cpus) { 344 struct aggr_cpu_id cpu_id = config->aggr_get_id(config, cpu); 345 346 if (aggr_cpu_id__equal(&cpu_id, id)) 347 return idx; 348 } 349 return 0; 350 } 351 352 static void abs_printout(struct perf_stat_config *config, 353 struct aggr_cpu_id id, int nr, struct evsel *evsel, double avg) 354 { 355 FILE *output = config->output; 356 double sc = evsel->scale; 357 const char *fmt; 358 359 if (config->csv_output) { 360 fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s"; 361 } else { 362 if (config->big_num) 363 fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s"; 364 else 365 fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s"; 366 } 367 368 aggr_printout(config, evsel, id, nr); 369 370 fprintf(output, fmt, avg, config->csv_sep); 371 372 if (evsel->unit) 373 fprintf(output, "%-*s%s", 374 config->csv_output ? 0 : config->unit_width, 375 evsel->unit, config->csv_sep); 376 377 fprintf(output, "%-*s", config->csv_output ? 0 : 32, evsel__name(evsel)); 378 379 print_cgroup(config, evsel); 380 } 381 382 static bool is_mixed_hw_group(struct evsel *counter) 383 { 384 struct evlist *evlist = counter->evlist; 385 u32 pmu_type = counter->core.attr.type; 386 struct evsel *pos; 387 388 if (counter->core.nr_members < 2) 389 return false; 390 391 evlist__for_each_entry(evlist, pos) { 392 /* software events can be part of any hardware group */ 393 if (pos->core.attr.type == PERF_TYPE_SOFTWARE) 394 continue; 395 if (pmu_type == PERF_TYPE_SOFTWARE) { 396 pmu_type = pos->core.attr.type; 397 continue; 398 } 399 if (pmu_type != pos->core.attr.type) 400 return true; 401 } 402 403 return false; 404 } 405 406 static void printout(struct perf_stat_config *config, struct aggr_cpu_id id, int nr, 407 struct evsel *counter, double uval, 408 char *prefix, u64 run, u64 ena, double noise, 409 struct runtime_stat *st) 410 { 411 struct perf_stat_output_ctx out; 412 struct outstate os = { 413 .fh = config->output, 414 .prefix = prefix ? prefix : "", 415 .id = id, 416 .nr = nr, 417 .evsel = counter, 418 }; 419 print_metric_t pm = print_metric_std; 420 new_line_t nl; 421 422 if (config->metric_only) { 423 nl = new_line_metric; 424 if (config->csv_output) 425 pm = print_metric_only_csv; 426 else 427 pm = print_metric_only; 428 } else 429 nl = new_line_std; 430 431 if (config->csv_output && !config->metric_only) { 432 static int aggr_fields[] = { 433 [AGGR_GLOBAL] = 0, 434 [AGGR_THREAD] = 1, 435 [AGGR_NONE] = 1, 436 [AGGR_SOCKET] = 2, 437 [AGGR_DIE] = 2, 438 [AGGR_CORE] = 2, 439 }; 440 441 pm = print_metric_csv; 442 nl = new_line_csv; 443 os.nfields = 3; 444 os.nfields += aggr_fields[config->aggr_mode]; 445 if (counter->cgrp) 446 os.nfields++; 447 } 448 449 if (!config->no_csv_summary && config->csv_output && 450 config->summary && !config->interval) { 451 fprintf(config->output, "%16s%s", "summary", config->csv_sep); 452 } 453 454 if (run == 0 || ena == 0 || counter->counts->scaled == -1) { 455 if (config->metric_only) { 456 pm(config, &os, NULL, "", "", 0); 457 return; 458 } 459 aggr_printout(config, counter, id, nr); 460 461 fprintf(config->output, "%*s%s", 462 config->csv_output ? 0 : 18, 463 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, 464 config->csv_sep); 465 466 if (counter->supported) { 467 if (!evlist__has_hybrid(counter->evlist)) { 468 config->print_free_counters_hint = 1; 469 if (is_mixed_hw_group(counter)) 470 config->print_mixed_hw_group_error = 1; 471 } 472 } 473 474 fprintf(config->output, "%-*s%s", 475 config->csv_output ? 0 : config->unit_width, 476 counter->unit, config->csv_sep); 477 478 fprintf(config->output, "%*s", 479 config->csv_output ? 0 : -25, evsel__name(counter)); 480 481 print_cgroup(config, counter); 482 483 if (!config->csv_output) 484 pm(config, &os, NULL, NULL, "", 0); 485 print_noise(config, counter, noise); 486 print_running(config, run, ena); 487 if (config->csv_output) 488 pm(config, &os, NULL, NULL, "", 0); 489 return; 490 } 491 492 if (!config->metric_only) 493 abs_printout(config, id, nr, counter, uval); 494 495 out.print_metric = pm; 496 out.new_line = nl; 497 out.ctx = &os; 498 out.force_header = false; 499 500 if (config->csv_output && !config->metric_only) { 501 print_noise(config, counter, noise); 502 print_running(config, run, ena); 503 } 504 505 perf_stat__print_shadow_stats(config, counter, uval, 506 first_shadow_cpu_map_idx(config, counter, &id), 507 &out, &config->metric_events, st); 508 if (!config->csv_output && !config->metric_only) { 509 print_noise(config, counter, noise); 510 print_running(config, run, ena); 511 } 512 } 513 514 static void aggr_update_shadow(struct perf_stat_config *config, 515 struct evlist *evlist) 516 { 517 int idx, s; 518 struct perf_cpu cpu; 519 struct aggr_cpu_id s2, id; 520 u64 val; 521 struct evsel *counter; 522 struct perf_cpu_map *cpus; 523 524 for (s = 0; s < config->aggr_map->nr; s++) { 525 id = config->aggr_map->map[s]; 526 evlist__for_each_entry(evlist, counter) { 527 cpus = evsel__cpus(counter); 528 val = 0; 529 perf_cpu_map__for_each_cpu(cpu, idx, cpus) { 530 s2 = config->aggr_get_id(config, cpu); 531 if (!aggr_cpu_id__equal(&s2, &id)) 532 continue; 533 val += perf_counts(counter->counts, idx, 0)->val; 534 } 535 perf_stat__update_shadow_stats(counter, val, 536 first_shadow_cpu_map_idx(config, counter, &id), 537 &rt_stat); 538 } 539 } 540 } 541 542 static void uniquify_event_name(struct evsel *counter) 543 { 544 char *new_name; 545 char *config; 546 int ret = 0; 547 548 if (counter->uniquified_name || counter->use_config_name || 549 !counter->pmu_name || !strncmp(counter->name, counter->pmu_name, 550 strlen(counter->pmu_name))) 551 return; 552 553 config = strchr(counter->name, '/'); 554 if (config) { 555 if (asprintf(&new_name, 556 "%s%s", counter->pmu_name, config) > 0) { 557 free(counter->name); 558 counter->name = new_name; 559 } 560 } else { 561 if (perf_pmu__has_hybrid()) { 562 ret = asprintf(&new_name, "%s/%s/", 563 counter->pmu_name, counter->name); 564 } else { 565 ret = asprintf(&new_name, "%s [%s]", 566 counter->name, counter->pmu_name); 567 } 568 569 if (ret) { 570 free(counter->name); 571 counter->name = new_name; 572 } 573 } 574 575 counter->uniquified_name = true; 576 } 577 578 static void collect_all_aliases(struct perf_stat_config *config, struct evsel *counter, 579 void (*cb)(struct perf_stat_config *config, struct evsel *counter, void *data, 580 bool first), 581 void *data) 582 { 583 struct evlist *evlist = counter->evlist; 584 struct evsel *alias; 585 586 alias = list_prepare_entry(counter, &(evlist->core.entries), core.node); 587 list_for_each_entry_continue (alias, &evlist->core.entries, core.node) { 588 /* Merge events with the same name, etc. but on different PMUs. */ 589 if (!strcmp(evsel__name(alias), evsel__name(counter)) && 590 alias->scale == counter->scale && 591 alias->cgrp == counter->cgrp && 592 !strcmp(alias->unit, counter->unit) && 593 evsel__is_clock(alias) == evsel__is_clock(counter) && 594 strcmp(alias->pmu_name, counter->pmu_name)) { 595 alias->merged_stat = true; 596 cb(config, alias, data, false); 597 } 598 } 599 } 600 601 static bool is_uncore(struct evsel *evsel) 602 { 603 struct perf_pmu *pmu = evsel__find_pmu(evsel); 604 605 return pmu && pmu->is_uncore; 606 } 607 608 static bool hybrid_uniquify(struct evsel *evsel) 609 { 610 return perf_pmu__has_hybrid() && !is_uncore(evsel); 611 } 612 613 static bool hybrid_merge(struct evsel *counter, struct perf_stat_config *config, 614 bool check) 615 { 616 if (hybrid_uniquify(counter)) { 617 if (check) 618 return config && config->hybrid_merge; 619 else 620 return config && !config->hybrid_merge; 621 } 622 623 return false; 624 } 625 626 static bool collect_data(struct perf_stat_config *config, struct evsel *counter, 627 void (*cb)(struct perf_stat_config *config, struct evsel *counter, void *data, 628 bool first), 629 void *data) 630 { 631 if (counter->merged_stat) 632 return false; 633 cb(config, counter, data, true); 634 if (config->no_merge || hybrid_merge(counter, config, false)) 635 uniquify_event_name(counter); 636 else if (counter->auto_merge_stats || hybrid_merge(counter, config, true)) 637 collect_all_aliases(config, counter, cb, data); 638 return true; 639 } 640 641 struct aggr_data { 642 u64 ena, run, val; 643 struct aggr_cpu_id id; 644 int nr; 645 int cpu_map_idx; 646 }; 647 648 static void aggr_cb(struct perf_stat_config *config, 649 struct evsel *counter, void *data, bool first) 650 { 651 struct aggr_data *ad = data; 652 int idx; 653 struct perf_cpu cpu; 654 struct perf_cpu_map *cpus; 655 struct aggr_cpu_id s2; 656 657 cpus = evsel__cpus(counter); 658 perf_cpu_map__for_each_cpu(cpu, idx, cpus) { 659 struct perf_counts_values *counts; 660 661 s2 = config->aggr_get_id(config, cpu); 662 if (!aggr_cpu_id__equal(&s2, &ad->id)) 663 continue; 664 if (first) 665 ad->nr++; 666 counts = perf_counts(counter->counts, idx, 0); 667 /* 668 * When any result is bad, make them all to give 669 * consistent output in interval mode. 670 */ 671 if (counts->ena == 0 || counts->run == 0 || 672 counter->counts->scaled == -1) { 673 ad->ena = 0; 674 ad->run = 0; 675 break; 676 } 677 ad->val += counts->val; 678 ad->ena += counts->ena; 679 ad->run += counts->run; 680 } 681 } 682 683 static void print_counter_aggrdata(struct perf_stat_config *config, 684 struct evsel *counter, int s, 685 char *prefix, bool metric_only, 686 bool *first, struct perf_cpu cpu) 687 { 688 struct aggr_data ad; 689 FILE *output = config->output; 690 u64 ena, run, val; 691 int nr; 692 struct aggr_cpu_id id; 693 double uval; 694 695 ad.id = id = config->aggr_map->map[s]; 696 ad.val = ad.ena = ad.run = 0; 697 ad.nr = 0; 698 if (!collect_data(config, counter, aggr_cb, &ad)) 699 return; 700 701 if (perf_pmu__has_hybrid() && ad.ena == 0) 702 return; 703 704 nr = ad.nr; 705 ena = ad.ena; 706 run = ad.run; 707 val = ad.val; 708 if (*first && metric_only) { 709 *first = false; 710 aggr_printout(config, counter, id, nr); 711 } 712 if (prefix && !metric_only) 713 fprintf(output, "%s", prefix); 714 715 uval = val * counter->scale; 716 if (cpu.cpu != -1) 717 id = aggr_cpu_id__cpu(cpu, /*data=*/NULL); 718 719 printout(config, id, nr, counter, uval, 720 prefix, run, ena, 1.0, &rt_stat); 721 if (!metric_only) 722 fputc('\n', output); 723 } 724 725 static void print_aggr(struct perf_stat_config *config, 726 struct evlist *evlist, 727 char *prefix) 728 { 729 bool metric_only = config->metric_only; 730 FILE *output = config->output; 731 struct evsel *counter; 732 int s; 733 bool first; 734 735 if (!config->aggr_map || !config->aggr_get_id) 736 return; 737 738 aggr_update_shadow(config, evlist); 739 740 /* 741 * With metric_only everything is on a single line. 742 * Without each counter has its own line. 743 */ 744 for (s = 0; s < config->aggr_map->nr; s++) { 745 if (prefix && metric_only) 746 fprintf(output, "%s", prefix); 747 748 first = true; 749 evlist__for_each_entry(evlist, counter) { 750 print_counter_aggrdata(config, counter, s, 751 prefix, metric_only, 752 &first, (struct perf_cpu){ .cpu = -1 }); 753 } 754 if (metric_only) 755 fputc('\n', output); 756 } 757 } 758 759 static int cmp_val(const void *a, const void *b) 760 { 761 return ((struct perf_aggr_thread_value *)b)->val - 762 ((struct perf_aggr_thread_value *)a)->val; 763 } 764 765 static struct perf_aggr_thread_value *sort_aggr_thread( 766 struct evsel *counter, 767 int *ret, 768 struct target *_target) 769 { 770 int nthreads = perf_thread_map__nr(counter->core.threads); 771 int i = 0; 772 double uval; 773 struct perf_aggr_thread_value *buf; 774 775 buf = calloc(nthreads, sizeof(struct perf_aggr_thread_value)); 776 if (!buf) 777 return NULL; 778 779 for (int thread = 0; thread < nthreads; thread++) { 780 int idx; 781 u64 ena = 0, run = 0, val = 0; 782 783 perf_cpu_map__for_each_idx(idx, evsel__cpus(counter)) { 784 struct perf_counts_values *counts = 785 perf_counts(counter->counts, idx, thread); 786 787 val += counts->val; 788 ena += counts->ena; 789 run += counts->run; 790 } 791 792 uval = val * counter->scale; 793 794 /* 795 * Skip value 0 when enabling --per-thread globally, 796 * otherwise too many 0 output. 797 */ 798 if (uval == 0.0 && target__has_per_thread(_target)) 799 continue; 800 801 buf[i].counter = counter; 802 buf[i].id = aggr_cpu_id__empty(); 803 buf[i].id.thread = thread; 804 buf[i].uval = uval; 805 buf[i].val = val; 806 buf[i].run = run; 807 buf[i].ena = ena; 808 i++; 809 } 810 811 qsort(buf, i, sizeof(struct perf_aggr_thread_value), cmp_val); 812 813 if (ret) 814 *ret = i; 815 816 return buf; 817 } 818 819 static void print_aggr_thread(struct perf_stat_config *config, 820 struct target *_target, 821 struct evsel *counter, char *prefix) 822 { 823 FILE *output = config->output; 824 int thread, sorted_threads; 825 struct aggr_cpu_id id; 826 struct perf_aggr_thread_value *buf; 827 828 buf = sort_aggr_thread(counter, &sorted_threads, _target); 829 if (!buf) { 830 perror("cannot sort aggr thread"); 831 return; 832 } 833 834 for (thread = 0; thread < sorted_threads; thread++) { 835 if (prefix) 836 fprintf(output, "%s", prefix); 837 838 id = buf[thread].id; 839 if (config->stats) 840 printout(config, id, 0, buf[thread].counter, buf[thread].uval, 841 prefix, buf[thread].run, buf[thread].ena, 1.0, 842 &config->stats[id.thread]); 843 else 844 printout(config, id, 0, buf[thread].counter, buf[thread].uval, 845 prefix, buf[thread].run, buf[thread].ena, 1.0, 846 &rt_stat); 847 fputc('\n', output); 848 } 849 850 free(buf); 851 } 852 853 struct caggr_data { 854 double avg, avg_enabled, avg_running; 855 }; 856 857 static void counter_aggr_cb(struct perf_stat_config *config __maybe_unused, 858 struct evsel *counter, void *data, 859 bool first __maybe_unused) 860 { 861 struct caggr_data *cd = data; 862 struct perf_counts_values *aggr = &counter->counts->aggr; 863 864 cd->avg += aggr->val; 865 cd->avg_enabled += aggr->ena; 866 cd->avg_running += aggr->run; 867 } 868 869 /* 870 * Print out the results of a single counter: 871 * aggregated counts in system-wide mode 872 */ 873 static void print_counter_aggr(struct perf_stat_config *config, 874 struct evsel *counter, char *prefix) 875 { 876 bool metric_only = config->metric_only; 877 FILE *output = config->output; 878 double uval; 879 struct caggr_data cd = { .avg = 0.0 }; 880 881 if (!collect_data(config, counter, counter_aggr_cb, &cd)) 882 return; 883 884 if (prefix && !metric_only) 885 fprintf(output, "%s", prefix); 886 887 uval = cd.avg * counter->scale; 888 printout(config, aggr_cpu_id__empty(), 0, counter, uval, prefix, cd.avg_running, 889 cd.avg_enabled, cd.avg, &rt_stat); 890 if (!metric_only) 891 fprintf(output, "\n"); 892 } 893 894 static void counter_cb(struct perf_stat_config *config __maybe_unused, 895 struct evsel *counter, void *data, 896 bool first __maybe_unused) 897 { 898 struct aggr_data *ad = data; 899 900 ad->val += perf_counts(counter->counts, ad->cpu_map_idx, 0)->val; 901 ad->ena += perf_counts(counter->counts, ad->cpu_map_idx, 0)->ena; 902 ad->run += perf_counts(counter->counts, ad->cpu_map_idx, 0)->run; 903 } 904 905 /* 906 * Print out the results of a single counter: 907 * does not use aggregated count in system-wide 908 */ 909 static void print_counter(struct perf_stat_config *config, 910 struct evsel *counter, char *prefix) 911 { 912 FILE *output = config->output; 913 u64 ena, run, val; 914 double uval; 915 int idx; 916 struct perf_cpu cpu; 917 struct aggr_cpu_id id; 918 919 perf_cpu_map__for_each_cpu(cpu, idx, evsel__cpus(counter)) { 920 struct aggr_data ad = { .cpu_map_idx = idx }; 921 922 if (!collect_data(config, counter, counter_cb, &ad)) 923 return; 924 val = ad.val; 925 ena = ad.ena; 926 run = ad.run; 927 928 if (prefix) 929 fprintf(output, "%s", prefix); 930 931 uval = val * counter->scale; 932 id = aggr_cpu_id__cpu(cpu, /*data=*/NULL); 933 printout(config, id, 0, counter, uval, prefix, 934 run, ena, 1.0, &rt_stat); 935 936 fputc('\n', output); 937 } 938 } 939 940 static void print_no_aggr_metric(struct perf_stat_config *config, 941 struct evlist *evlist, 942 char *prefix) 943 { 944 int all_idx; 945 struct perf_cpu cpu; 946 947 perf_cpu_map__for_each_cpu(cpu, all_idx, evlist->core.user_requested_cpus) { 948 struct evsel *counter; 949 bool first = true; 950 951 evlist__for_each_entry(evlist, counter) { 952 u64 ena, run, val; 953 double uval; 954 struct aggr_cpu_id id; 955 int counter_idx = perf_cpu_map__idx(evsel__cpus(counter), cpu); 956 957 if (counter_idx < 0) 958 continue; 959 960 id = aggr_cpu_id__cpu(cpu, /*data=*/NULL); 961 if (first) { 962 if (prefix) 963 fputs(prefix, config->output); 964 aggr_printout(config, counter, id, 0); 965 first = false; 966 } 967 val = perf_counts(counter->counts, counter_idx, 0)->val; 968 ena = perf_counts(counter->counts, counter_idx, 0)->ena; 969 run = perf_counts(counter->counts, counter_idx, 0)->run; 970 971 uval = val * counter->scale; 972 printout(config, id, 0, counter, uval, prefix, 973 run, ena, 1.0, &rt_stat); 974 } 975 if (!first) 976 fputc('\n', config->output); 977 } 978 } 979 980 static int aggr_header_lens[] = { 981 [AGGR_CORE] = 24, 982 [AGGR_DIE] = 18, 983 [AGGR_SOCKET] = 12, 984 [AGGR_NONE] = 6, 985 [AGGR_THREAD] = 24, 986 [AGGR_GLOBAL] = 0, 987 }; 988 989 static const char *aggr_header_csv[] = { 990 [AGGR_CORE] = "core,cpus,", 991 [AGGR_DIE] = "die,cpus", 992 [AGGR_SOCKET] = "socket,cpus", 993 [AGGR_NONE] = "cpu,", 994 [AGGR_THREAD] = "comm-pid,", 995 [AGGR_GLOBAL] = "" 996 }; 997 998 static void print_metric_headers(struct perf_stat_config *config, 999 struct evlist *evlist, 1000 const char *prefix, bool no_indent) 1001 { 1002 struct perf_stat_output_ctx out; 1003 struct evsel *counter; 1004 struct outstate os = { 1005 .fh = config->output 1006 }; 1007 1008 if (prefix) 1009 fprintf(config->output, "%s", prefix); 1010 1011 if (!config->csv_output && !no_indent) 1012 fprintf(config->output, "%*s", 1013 aggr_header_lens[config->aggr_mode], ""); 1014 if (config->csv_output) { 1015 if (config->interval) 1016 fputs("time,", config->output); 1017 if (!config->iostat_run) 1018 fputs(aggr_header_csv[config->aggr_mode], config->output); 1019 } 1020 if (config->iostat_run) 1021 iostat_print_header_prefix(config); 1022 1023 /* Print metrics headers only */ 1024 evlist__for_each_entry(evlist, counter) { 1025 os.evsel = counter; 1026 out.ctx = &os; 1027 out.print_metric = print_metric_header; 1028 out.new_line = new_line_metric; 1029 out.force_header = true; 1030 perf_stat__print_shadow_stats(config, counter, 0, 1031 0, 1032 &out, 1033 &config->metric_events, 1034 &rt_stat); 1035 } 1036 fputc('\n', config->output); 1037 } 1038 1039 static void print_interval(struct perf_stat_config *config, 1040 struct evlist *evlist, 1041 char *prefix, struct timespec *ts) 1042 { 1043 bool metric_only = config->metric_only; 1044 unsigned int unit_width = config->unit_width; 1045 FILE *output = config->output; 1046 static int num_print_interval; 1047 1048 if (config->interval_clear) 1049 puts(CONSOLE_CLEAR); 1050 1051 if (!config->iostat_run) 1052 sprintf(prefix, "%6lu.%09lu%s", (unsigned long) ts->tv_sec, ts->tv_nsec, config->csv_sep); 1053 1054 if ((num_print_interval == 0 && !config->csv_output) || config->interval_clear) { 1055 switch (config->aggr_mode) { 1056 case AGGR_NODE: 1057 fprintf(output, "# time node cpus"); 1058 if (!metric_only) 1059 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1060 break; 1061 case AGGR_SOCKET: 1062 fprintf(output, "# time socket cpus"); 1063 if (!metric_only) 1064 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1065 break; 1066 case AGGR_DIE: 1067 fprintf(output, "# time die cpus"); 1068 if (!metric_only) 1069 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1070 break; 1071 case AGGR_CORE: 1072 fprintf(output, "# time core cpus"); 1073 if (!metric_only) 1074 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1075 break; 1076 case AGGR_NONE: 1077 fprintf(output, "# time CPU "); 1078 if (!metric_only) 1079 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1080 break; 1081 case AGGR_THREAD: 1082 fprintf(output, "# time comm-pid"); 1083 if (!metric_only) 1084 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1085 break; 1086 case AGGR_GLOBAL: 1087 default: 1088 if (!config->iostat_run) { 1089 fprintf(output, "# time"); 1090 if (!metric_only) 1091 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1092 } 1093 case AGGR_UNSET: 1094 break; 1095 } 1096 } 1097 1098 if ((num_print_interval == 0 || config->interval_clear) && metric_only) 1099 print_metric_headers(config, evlist, " ", true); 1100 if (++num_print_interval == 25) 1101 num_print_interval = 0; 1102 } 1103 1104 static void print_header(struct perf_stat_config *config, 1105 struct target *_target, 1106 int argc, const char **argv) 1107 { 1108 FILE *output = config->output; 1109 int i; 1110 1111 fflush(stdout); 1112 1113 if (!config->csv_output) { 1114 fprintf(output, "\n"); 1115 fprintf(output, " Performance counter stats for "); 1116 if (_target->bpf_str) 1117 fprintf(output, "\'BPF program(s) %s", _target->bpf_str); 1118 else if (_target->system_wide) 1119 fprintf(output, "\'system wide"); 1120 else if (_target->cpu_list) 1121 fprintf(output, "\'CPU(s) %s", _target->cpu_list); 1122 else if (!target__has_task(_target)) { 1123 fprintf(output, "\'%s", argv ? argv[0] : "pipe"); 1124 for (i = 1; argv && (i < argc); i++) 1125 fprintf(output, " %s", argv[i]); 1126 } else if (_target->pid) 1127 fprintf(output, "process id \'%s", _target->pid); 1128 else 1129 fprintf(output, "thread id \'%s", _target->tid); 1130 1131 fprintf(output, "\'"); 1132 if (config->run_count > 1) 1133 fprintf(output, " (%d runs)", config->run_count); 1134 fprintf(output, ":\n\n"); 1135 } 1136 } 1137 1138 static int get_precision(double num) 1139 { 1140 if (num > 1) 1141 return 0; 1142 1143 return lround(ceil(-log10(num))); 1144 } 1145 1146 static void print_table(struct perf_stat_config *config, 1147 FILE *output, int precision, double avg) 1148 { 1149 char tmp[64]; 1150 int idx, indent = 0; 1151 1152 scnprintf(tmp, 64, " %17.*f", precision, avg); 1153 while (tmp[indent] == ' ') 1154 indent++; 1155 1156 fprintf(output, "%*s# Table of individual measurements:\n", indent, ""); 1157 1158 for (idx = 0; idx < config->run_count; idx++) { 1159 double run = (double) config->walltime_run[idx] / NSEC_PER_SEC; 1160 int h, n = 1 + abs((int) (100.0 * (run - avg)/run) / 5); 1161 1162 fprintf(output, " %17.*f (%+.*f) ", 1163 precision, run, precision, run - avg); 1164 1165 for (h = 0; h < n; h++) 1166 fprintf(output, "#"); 1167 1168 fprintf(output, "\n"); 1169 } 1170 1171 fprintf(output, "\n%*s# Final result:\n", indent, ""); 1172 } 1173 1174 static double timeval2double(struct timeval *t) 1175 { 1176 return t->tv_sec + (double) t->tv_usec/USEC_PER_SEC; 1177 } 1178 1179 static void print_footer(struct perf_stat_config *config) 1180 { 1181 double avg = avg_stats(config->walltime_nsecs_stats) / NSEC_PER_SEC; 1182 FILE *output = config->output; 1183 1184 if (!config->null_run) 1185 fprintf(output, "\n"); 1186 1187 if (config->run_count == 1) { 1188 fprintf(output, " %17.9f seconds time elapsed", avg); 1189 1190 if (config->ru_display) { 1191 double ru_utime = timeval2double(&config->ru_data.ru_utime); 1192 double ru_stime = timeval2double(&config->ru_data.ru_stime); 1193 1194 fprintf(output, "\n\n"); 1195 fprintf(output, " %17.9f seconds user\n", ru_utime); 1196 fprintf(output, " %17.9f seconds sys\n", ru_stime); 1197 } 1198 } else { 1199 double sd = stddev_stats(config->walltime_nsecs_stats) / NSEC_PER_SEC; 1200 /* 1201 * Display at most 2 more significant 1202 * digits than the stddev inaccuracy. 1203 */ 1204 int precision = get_precision(sd) + 2; 1205 1206 if (config->walltime_run_table) 1207 print_table(config, output, precision, avg); 1208 1209 fprintf(output, " %17.*f +- %.*f seconds time elapsed", 1210 precision, avg, precision, sd); 1211 1212 print_noise_pct(config, sd, avg); 1213 } 1214 fprintf(output, "\n\n"); 1215 1216 if (config->print_free_counters_hint && sysctl__nmi_watchdog_enabled()) 1217 fprintf(output, 1218 "Some events weren't counted. Try disabling the NMI watchdog:\n" 1219 " echo 0 > /proc/sys/kernel/nmi_watchdog\n" 1220 " perf stat ...\n" 1221 " echo 1 > /proc/sys/kernel/nmi_watchdog\n"); 1222 1223 if (config->print_mixed_hw_group_error) 1224 fprintf(output, 1225 "The events in group usually have to be from " 1226 "the same PMU. Try reorganizing the group.\n"); 1227 } 1228 1229 static void print_percore_thread(struct perf_stat_config *config, 1230 struct evsel *counter, char *prefix) 1231 { 1232 int s; 1233 struct aggr_cpu_id s2, id; 1234 struct perf_cpu_map *cpus; 1235 bool first = true; 1236 int idx; 1237 struct perf_cpu cpu; 1238 1239 cpus = evsel__cpus(counter); 1240 perf_cpu_map__for_each_cpu(cpu, idx, cpus) { 1241 s2 = config->aggr_get_id(config, cpu); 1242 for (s = 0; s < config->aggr_map->nr; s++) { 1243 id = config->aggr_map->map[s]; 1244 if (aggr_cpu_id__equal(&s2, &id)) 1245 break; 1246 } 1247 1248 print_counter_aggrdata(config, counter, s, 1249 prefix, false, 1250 &first, cpu); 1251 } 1252 } 1253 1254 static void print_percore(struct perf_stat_config *config, 1255 struct evsel *counter, char *prefix) 1256 { 1257 bool metric_only = config->metric_only; 1258 FILE *output = config->output; 1259 int s; 1260 bool first = true; 1261 1262 if (!config->aggr_map || !config->aggr_get_id) 1263 return; 1264 1265 if (config->percore_show_thread) 1266 return print_percore_thread(config, counter, prefix); 1267 1268 for (s = 0; s < config->aggr_map->nr; s++) { 1269 if (prefix && metric_only) 1270 fprintf(output, "%s", prefix); 1271 1272 print_counter_aggrdata(config, counter, s, 1273 prefix, metric_only, 1274 &first, (struct perf_cpu){ .cpu = -1 }); 1275 } 1276 1277 if (metric_only) 1278 fputc('\n', output); 1279 } 1280 1281 void evlist__print_counters(struct evlist *evlist, struct perf_stat_config *config, 1282 struct target *_target, struct timespec *ts, int argc, const char **argv) 1283 { 1284 bool metric_only = config->metric_only; 1285 int interval = config->interval; 1286 struct evsel *counter; 1287 char buf[64], *prefix = NULL; 1288 1289 if (config->iostat_run) 1290 evlist->selected = evlist__first(evlist); 1291 1292 if (interval) 1293 print_interval(config, evlist, prefix = buf, ts); 1294 else 1295 print_header(config, _target, argc, argv); 1296 1297 if (metric_only) { 1298 static int num_print_iv; 1299 1300 if (num_print_iv == 0 && !interval) 1301 print_metric_headers(config, evlist, prefix, false); 1302 if (num_print_iv++ == 25) 1303 num_print_iv = 0; 1304 if (config->aggr_mode == AGGR_GLOBAL && prefix && !config->iostat_run) 1305 fprintf(config->output, "%s", prefix); 1306 } 1307 1308 switch (config->aggr_mode) { 1309 case AGGR_CORE: 1310 case AGGR_DIE: 1311 case AGGR_SOCKET: 1312 case AGGR_NODE: 1313 print_aggr(config, evlist, prefix); 1314 break; 1315 case AGGR_THREAD: 1316 evlist__for_each_entry(evlist, counter) { 1317 print_aggr_thread(config, _target, counter, prefix); 1318 } 1319 break; 1320 case AGGR_GLOBAL: 1321 if (config->iostat_run) 1322 iostat_print_counters(evlist, config, ts, prefix = buf, 1323 print_counter_aggr); 1324 else { 1325 evlist__for_each_entry(evlist, counter) { 1326 print_counter_aggr(config, counter, prefix); 1327 } 1328 if (metric_only) 1329 fputc('\n', config->output); 1330 } 1331 break; 1332 case AGGR_NONE: 1333 if (metric_only) 1334 print_no_aggr_metric(config, evlist, prefix); 1335 else { 1336 evlist__for_each_entry(evlist, counter) { 1337 if (counter->percore) 1338 print_percore(config, counter, prefix); 1339 else 1340 print_counter(config, counter, prefix); 1341 } 1342 } 1343 break; 1344 case AGGR_UNSET: 1345 default: 1346 break; 1347 } 1348 1349 if (!interval && !config->csv_output) 1350 print_footer(config); 1351 1352 fflush(config->output); 1353 } 1354