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 : 25, 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 collect_data(struct perf_stat_config *config, struct evsel *counter, 614 void (*cb)(struct perf_stat_config *config, struct evsel *counter, void *data, 615 bool first), 616 void *data) 617 { 618 if (counter->merged_stat) 619 return false; 620 cb(config, counter, data, true); 621 if (config->no_merge || hybrid_uniquify(counter)) 622 uniquify_event_name(counter); 623 else if (counter->auto_merge_stats) 624 collect_all_aliases(config, counter, cb, data); 625 return true; 626 } 627 628 struct aggr_data { 629 u64 ena, run, val; 630 struct aggr_cpu_id id; 631 int nr; 632 int cpu_map_idx; 633 }; 634 635 static void aggr_cb(struct perf_stat_config *config, 636 struct evsel *counter, void *data, bool first) 637 { 638 struct aggr_data *ad = data; 639 int idx; 640 struct perf_cpu cpu; 641 struct perf_cpu_map *cpus; 642 struct aggr_cpu_id s2; 643 644 cpus = evsel__cpus(counter); 645 perf_cpu_map__for_each_cpu(cpu, idx, cpus) { 646 struct perf_counts_values *counts; 647 648 s2 = config->aggr_get_id(config, cpu); 649 if (!aggr_cpu_id__equal(&s2, &ad->id)) 650 continue; 651 if (first) 652 ad->nr++; 653 counts = perf_counts(counter->counts, idx, 0); 654 /* 655 * When any result is bad, make them all to give 656 * consistent output in interval mode. 657 */ 658 if (counts->ena == 0 || counts->run == 0 || 659 counter->counts->scaled == -1) { 660 ad->ena = 0; 661 ad->run = 0; 662 break; 663 } 664 ad->val += counts->val; 665 ad->ena += counts->ena; 666 ad->run += counts->run; 667 } 668 } 669 670 static void print_counter_aggrdata(struct perf_stat_config *config, 671 struct evsel *counter, int s, 672 char *prefix, bool metric_only, 673 bool *first, struct perf_cpu cpu) 674 { 675 struct aggr_data ad; 676 FILE *output = config->output; 677 u64 ena, run, val; 678 int nr; 679 struct aggr_cpu_id id; 680 double uval; 681 682 ad.id = id = config->aggr_map->map[s]; 683 ad.val = ad.ena = ad.run = 0; 684 ad.nr = 0; 685 if (!collect_data(config, counter, aggr_cb, &ad)) 686 return; 687 688 if (perf_pmu__has_hybrid() && ad.ena == 0) 689 return; 690 691 nr = ad.nr; 692 ena = ad.ena; 693 run = ad.run; 694 val = ad.val; 695 if (*first && metric_only) { 696 *first = false; 697 aggr_printout(config, counter, id, nr); 698 } 699 if (prefix && !metric_only) 700 fprintf(output, "%s", prefix); 701 702 uval = val * counter->scale; 703 if (cpu.cpu != -1) 704 id = aggr_cpu_id__cpu(cpu, /*data=*/NULL); 705 706 printout(config, id, nr, counter, uval, 707 prefix, run, ena, 1.0, &rt_stat); 708 if (!metric_only) 709 fputc('\n', output); 710 } 711 712 static void print_aggr(struct perf_stat_config *config, 713 struct evlist *evlist, 714 char *prefix) 715 { 716 bool metric_only = config->metric_only; 717 FILE *output = config->output; 718 struct evsel *counter; 719 int s; 720 bool first; 721 722 if (!config->aggr_map || !config->aggr_get_id) 723 return; 724 725 aggr_update_shadow(config, evlist); 726 727 /* 728 * With metric_only everything is on a single line. 729 * Without each counter has its own line. 730 */ 731 for (s = 0; s < config->aggr_map->nr; s++) { 732 if (prefix && metric_only) 733 fprintf(output, "%s", prefix); 734 735 first = true; 736 evlist__for_each_entry(evlist, counter) { 737 print_counter_aggrdata(config, counter, s, 738 prefix, metric_only, 739 &first, (struct perf_cpu){ .cpu = -1 }); 740 } 741 if (metric_only) 742 fputc('\n', output); 743 } 744 } 745 746 static int cmp_val(const void *a, const void *b) 747 { 748 return ((struct perf_aggr_thread_value *)b)->val - 749 ((struct perf_aggr_thread_value *)a)->val; 750 } 751 752 static struct perf_aggr_thread_value *sort_aggr_thread( 753 struct evsel *counter, 754 int nthreads, int ncpus, 755 int *ret, 756 struct target *_target) 757 { 758 int cpu, thread, i = 0; 759 double uval; 760 struct perf_aggr_thread_value *buf; 761 762 buf = calloc(nthreads, sizeof(struct perf_aggr_thread_value)); 763 if (!buf) 764 return NULL; 765 766 for (thread = 0; thread < nthreads; thread++) { 767 u64 ena = 0, run = 0, val = 0; 768 769 for (cpu = 0; cpu < ncpus; cpu++) { 770 val += perf_counts(counter->counts, cpu, thread)->val; 771 ena += perf_counts(counter->counts, cpu, thread)->ena; 772 run += perf_counts(counter->counts, cpu, thread)->run; 773 } 774 775 uval = val * counter->scale; 776 777 /* 778 * Skip value 0 when enabling --per-thread globally, 779 * otherwise too many 0 output. 780 */ 781 if (uval == 0.0 && target__has_per_thread(_target)) 782 continue; 783 784 buf[i].counter = counter; 785 buf[i].id = aggr_cpu_id__empty(); 786 buf[i].id.thread = thread; 787 buf[i].uval = uval; 788 buf[i].val = val; 789 buf[i].run = run; 790 buf[i].ena = ena; 791 i++; 792 } 793 794 qsort(buf, i, sizeof(struct perf_aggr_thread_value), cmp_val); 795 796 if (ret) 797 *ret = i; 798 799 return buf; 800 } 801 802 static void print_aggr_thread(struct perf_stat_config *config, 803 struct target *_target, 804 struct evsel *counter, char *prefix) 805 { 806 FILE *output = config->output; 807 int nthreads = perf_thread_map__nr(counter->core.threads); 808 int ncpus = perf_cpu_map__nr(counter->core.cpus); 809 int thread, sorted_threads; 810 struct aggr_cpu_id id; 811 struct perf_aggr_thread_value *buf; 812 813 buf = sort_aggr_thread(counter, nthreads, ncpus, &sorted_threads, _target); 814 if (!buf) { 815 perror("cannot sort aggr thread"); 816 return; 817 } 818 819 for (thread = 0; thread < sorted_threads; thread++) { 820 if (prefix) 821 fprintf(output, "%s", prefix); 822 823 id = buf[thread].id; 824 if (config->stats) 825 printout(config, id, 0, buf[thread].counter, buf[thread].uval, 826 prefix, buf[thread].run, buf[thread].ena, 1.0, 827 &config->stats[id.thread]); 828 else 829 printout(config, id, 0, buf[thread].counter, buf[thread].uval, 830 prefix, buf[thread].run, buf[thread].ena, 1.0, 831 &rt_stat); 832 fputc('\n', output); 833 } 834 835 free(buf); 836 } 837 838 struct caggr_data { 839 double avg, avg_enabled, avg_running; 840 }; 841 842 static void counter_aggr_cb(struct perf_stat_config *config __maybe_unused, 843 struct evsel *counter, void *data, 844 bool first __maybe_unused) 845 { 846 struct caggr_data *cd = data; 847 struct perf_counts_values *aggr = &counter->counts->aggr; 848 849 cd->avg += aggr->val; 850 cd->avg_enabled += aggr->ena; 851 cd->avg_running += aggr->run; 852 } 853 854 /* 855 * Print out the results of a single counter: 856 * aggregated counts in system-wide mode 857 */ 858 static void print_counter_aggr(struct perf_stat_config *config, 859 struct evsel *counter, char *prefix) 860 { 861 bool metric_only = config->metric_only; 862 FILE *output = config->output; 863 double uval; 864 struct caggr_data cd = { .avg = 0.0 }; 865 866 if (!collect_data(config, counter, counter_aggr_cb, &cd)) 867 return; 868 869 if (prefix && !metric_only) 870 fprintf(output, "%s", prefix); 871 872 uval = cd.avg * counter->scale; 873 printout(config, aggr_cpu_id__empty(), 0, counter, uval, prefix, cd.avg_running, 874 cd.avg_enabled, cd.avg, &rt_stat); 875 if (!metric_only) 876 fprintf(output, "\n"); 877 } 878 879 static void counter_cb(struct perf_stat_config *config __maybe_unused, 880 struct evsel *counter, void *data, 881 bool first __maybe_unused) 882 { 883 struct aggr_data *ad = data; 884 885 ad->val += perf_counts(counter->counts, ad->cpu_map_idx, 0)->val; 886 ad->ena += perf_counts(counter->counts, ad->cpu_map_idx, 0)->ena; 887 ad->run += perf_counts(counter->counts, ad->cpu_map_idx, 0)->run; 888 } 889 890 /* 891 * Print out the results of a single counter: 892 * does not use aggregated count in system-wide 893 */ 894 static void print_counter(struct perf_stat_config *config, 895 struct evsel *counter, char *prefix) 896 { 897 FILE *output = config->output; 898 u64 ena, run, val; 899 double uval; 900 int idx; 901 struct perf_cpu cpu; 902 struct aggr_cpu_id id; 903 904 perf_cpu_map__for_each_cpu(cpu, idx, evsel__cpus(counter)) { 905 struct aggr_data ad = { .cpu_map_idx = idx }; 906 907 if (!collect_data(config, counter, counter_cb, &ad)) 908 return; 909 val = ad.val; 910 ena = ad.ena; 911 run = ad.run; 912 913 if (prefix) 914 fprintf(output, "%s", prefix); 915 916 uval = val * counter->scale; 917 id = aggr_cpu_id__cpu(cpu, /*data=*/NULL); 918 printout(config, id, 0, counter, uval, prefix, 919 run, ena, 1.0, &rt_stat); 920 921 fputc('\n', output); 922 } 923 } 924 925 static void print_no_aggr_metric(struct perf_stat_config *config, 926 struct evlist *evlist, 927 char *prefix) 928 { 929 int all_idx; 930 struct perf_cpu cpu; 931 932 perf_cpu_map__for_each_cpu(cpu, all_idx, evlist->core.user_requested_cpus) { 933 struct evsel *counter; 934 bool first = true; 935 936 if (prefix) 937 fputs(prefix, config->output); 938 evlist__for_each_entry(evlist, counter) { 939 u64 ena, run, val; 940 double uval; 941 struct aggr_cpu_id id; 942 int counter_idx = perf_cpu_map__idx(evsel__cpus(counter), cpu); 943 944 if (counter_idx < 0) 945 continue; 946 947 id = aggr_cpu_id__cpu(cpu, /*data=*/NULL); 948 if (first) { 949 aggr_printout(config, counter, id, 0); 950 first = false; 951 } 952 val = perf_counts(counter->counts, counter_idx, 0)->val; 953 ena = perf_counts(counter->counts, counter_idx, 0)->ena; 954 run = perf_counts(counter->counts, counter_idx, 0)->run; 955 956 uval = val * counter->scale; 957 printout(config, id, 0, counter, uval, prefix, 958 run, ena, 1.0, &rt_stat); 959 } 960 fputc('\n', config->output); 961 } 962 } 963 964 static int aggr_header_lens[] = { 965 [AGGR_CORE] = 24, 966 [AGGR_DIE] = 18, 967 [AGGR_SOCKET] = 12, 968 [AGGR_NONE] = 6, 969 [AGGR_THREAD] = 24, 970 [AGGR_GLOBAL] = 0, 971 }; 972 973 static const char *aggr_header_csv[] = { 974 [AGGR_CORE] = "core,cpus,", 975 [AGGR_DIE] = "die,cpus", 976 [AGGR_SOCKET] = "socket,cpus", 977 [AGGR_NONE] = "cpu,", 978 [AGGR_THREAD] = "comm-pid,", 979 [AGGR_GLOBAL] = "" 980 }; 981 982 static void print_metric_headers(struct perf_stat_config *config, 983 struct evlist *evlist, 984 const char *prefix, bool no_indent) 985 { 986 struct perf_stat_output_ctx out; 987 struct evsel *counter; 988 struct outstate os = { 989 .fh = config->output 990 }; 991 992 if (prefix) 993 fprintf(config->output, "%s", prefix); 994 995 if (!config->csv_output && !no_indent) 996 fprintf(config->output, "%*s", 997 aggr_header_lens[config->aggr_mode], ""); 998 if (config->csv_output) { 999 if (config->interval) 1000 fputs("time,", config->output); 1001 if (!config->iostat_run) 1002 fputs(aggr_header_csv[config->aggr_mode], config->output); 1003 } 1004 if (config->iostat_run) 1005 iostat_print_header_prefix(config); 1006 1007 /* Print metrics headers only */ 1008 evlist__for_each_entry(evlist, counter) { 1009 os.evsel = counter; 1010 out.ctx = &os; 1011 out.print_metric = print_metric_header; 1012 out.new_line = new_line_metric; 1013 out.force_header = true; 1014 perf_stat__print_shadow_stats(config, counter, 0, 1015 0, 1016 &out, 1017 &config->metric_events, 1018 &rt_stat); 1019 } 1020 fputc('\n', config->output); 1021 } 1022 1023 static void print_interval(struct perf_stat_config *config, 1024 struct evlist *evlist, 1025 char *prefix, struct timespec *ts) 1026 { 1027 bool metric_only = config->metric_only; 1028 unsigned int unit_width = config->unit_width; 1029 FILE *output = config->output; 1030 static int num_print_interval; 1031 1032 if (config->interval_clear) 1033 puts(CONSOLE_CLEAR); 1034 1035 if (!config->iostat_run) 1036 sprintf(prefix, "%6lu.%09lu%s", (unsigned long) ts->tv_sec, ts->tv_nsec, config->csv_sep); 1037 1038 if ((num_print_interval == 0 && !config->csv_output) || config->interval_clear) { 1039 switch (config->aggr_mode) { 1040 case AGGR_NODE: 1041 fprintf(output, "# time node cpus"); 1042 if (!metric_only) 1043 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1044 break; 1045 case AGGR_SOCKET: 1046 fprintf(output, "# time socket cpus"); 1047 if (!metric_only) 1048 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1049 break; 1050 case AGGR_DIE: 1051 fprintf(output, "# time die cpus"); 1052 if (!metric_only) 1053 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1054 break; 1055 case AGGR_CORE: 1056 fprintf(output, "# time core cpus"); 1057 if (!metric_only) 1058 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1059 break; 1060 case AGGR_NONE: 1061 fprintf(output, "# time CPU "); 1062 if (!metric_only) 1063 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1064 break; 1065 case AGGR_THREAD: 1066 fprintf(output, "# time comm-pid"); 1067 if (!metric_only) 1068 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1069 break; 1070 case AGGR_GLOBAL: 1071 default: 1072 if (!config->iostat_run) { 1073 fprintf(output, "# time"); 1074 if (!metric_only) 1075 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1076 } 1077 case AGGR_UNSET: 1078 break; 1079 } 1080 } 1081 1082 if ((num_print_interval == 0 || config->interval_clear) && metric_only) 1083 print_metric_headers(config, evlist, " ", true); 1084 if (++num_print_interval == 25) 1085 num_print_interval = 0; 1086 } 1087 1088 static void print_header(struct perf_stat_config *config, 1089 struct target *_target, 1090 int argc, const char **argv) 1091 { 1092 FILE *output = config->output; 1093 int i; 1094 1095 fflush(stdout); 1096 1097 if (!config->csv_output) { 1098 fprintf(output, "\n"); 1099 fprintf(output, " Performance counter stats for "); 1100 if (_target->bpf_str) 1101 fprintf(output, "\'BPF program(s) %s", _target->bpf_str); 1102 else if (_target->system_wide) 1103 fprintf(output, "\'system wide"); 1104 else if (_target->cpu_list) 1105 fprintf(output, "\'CPU(s) %s", _target->cpu_list); 1106 else if (!target__has_task(_target)) { 1107 fprintf(output, "\'%s", argv ? argv[0] : "pipe"); 1108 for (i = 1; argv && (i < argc); i++) 1109 fprintf(output, " %s", argv[i]); 1110 } else if (_target->pid) 1111 fprintf(output, "process id \'%s", _target->pid); 1112 else 1113 fprintf(output, "thread id \'%s", _target->tid); 1114 1115 fprintf(output, "\'"); 1116 if (config->run_count > 1) 1117 fprintf(output, " (%d runs)", config->run_count); 1118 fprintf(output, ":\n\n"); 1119 } 1120 } 1121 1122 static int get_precision(double num) 1123 { 1124 if (num > 1) 1125 return 0; 1126 1127 return lround(ceil(-log10(num))); 1128 } 1129 1130 static void print_table(struct perf_stat_config *config, 1131 FILE *output, int precision, double avg) 1132 { 1133 char tmp[64]; 1134 int idx, indent = 0; 1135 1136 scnprintf(tmp, 64, " %17.*f", precision, avg); 1137 while (tmp[indent] == ' ') 1138 indent++; 1139 1140 fprintf(output, "%*s# Table of individual measurements:\n", indent, ""); 1141 1142 for (idx = 0; idx < config->run_count; idx++) { 1143 double run = (double) config->walltime_run[idx] / NSEC_PER_SEC; 1144 int h, n = 1 + abs((int) (100.0 * (run - avg)/run) / 5); 1145 1146 fprintf(output, " %17.*f (%+.*f) ", 1147 precision, run, precision, run - avg); 1148 1149 for (h = 0; h < n; h++) 1150 fprintf(output, "#"); 1151 1152 fprintf(output, "\n"); 1153 } 1154 1155 fprintf(output, "\n%*s# Final result:\n", indent, ""); 1156 } 1157 1158 static double timeval2double(struct timeval *t) 1159 { 1160 return t->tv_sec + (double) t->tv_usec/USEC_PER_SEC; 1161 } 1162 1163 static void print_footer(struct perf_stat_config *config) 1164 { 1165 double avg = avg_stats(config->walltime_nsecs_stats) / NSEC_PER_SEC; 1166 FILE *output = config->output; 1167 1168 if (!config->null_run) 1169 fprintf(output, "\n"); 1170 1171 if (config->run_count == 1) { 1172 fprintf(output, " %17.9f seconds time elapsed", avg); 1173 1174 if (config->ru_display) { 1175 double ru_utime = timeval2double(&config->ru_data.ru_utime); 1176 double ru_stime = timeval2double(&config->ru_data.ru_stime); 1177 1178 fprintf(output, "\n\n"); 1179 fprintf(output, " %17.9f seconds user\n", ru_utime); 1180 fprintf(output, " %17.9f seconds sys\n", ru_stime); 1181 } 1182 } else { 1183 double sd = stddev_stats(config->walltime_nsecs_stats) / NSEC_PER_SEC; 1184 /* 1185 * Display at most 2 more significant 1186 * digits than the stddev inaccuracy. 1187 */ 1188 int precision = get_precision(sd) + 2; 1189 1190 if (config->walltime_run_table) 1191 print_table(config, output, precision, avg); 1192 1193 fprintf(output, " %17.*f +- %.*f seconds time elapsed", 1194 precision, avg, precision, sd); 1195 1196 print_noise_pct(config, sd, avg); 1197 } 1198 fprintf(output, "\n\n"); 1199 1200 if (config->print_free_counters_hint && sysctl__nmi_watchdog_enabled()) 1201 fprintf(output, 1202 "Some events weren't counted. Try disabling the NMI watchdog:\n" 1203 " echo 0 > /proc/sys/kernel/nmi_watchdog\n" 1204 " perf stat ...\n" 1205 " echo 1 > /proc/sys/kernel/nmi_watchdog\n"); 1206 1207 if (config->print_mixed_hw_group_error) 1208 fprintf(output, 1209 "The events in group usually have to be from " 1210 "the same PMU. Try reorganizing the group.\n"); 1211 } 1212 1213 static void print_percore_thread(struct perf_stat_config *config, 1214 struct evsel *counter, char *prefix) 1215 { 1216 int s; 1217 struct aggr_cpu_id s2, id; 1218 struct perf_cpu_map *cpus; 1219 bool first = true; 1220 int idx; 1221 struct perf_cpu cpu; 1222 1223 cpus = evsel__cpus(counter); 1224 perf_cpu_map__for_each_cpu(cpu, idx, cpus) { 1225 s2 = config->aggr_get_id(config, cpu); 1226 for (s = 0; s < config->aggr_map->nr; s++) { 1227 id = config->aggr_map->map[s]; 1228 if (aggr_cpu_id__equal(&s2, &id)) 1229 break; 1230 } 1231 1232 print_counter_aggrdata(config, counter, s, 1233 prefix, false, 1234 &first, cpu); 1235 } 1236 } 1237 1238 static void print_percore(struct perf_stat_config *config, 1239 struct evsel *counter, char *prefix) 1240 { 1241 bool metric_only = config->metric_only; 1242 FILE *output = config->output; 1243 int s; 1244 bool first = true; 1245 1246 if (!config->aggr_map || !config->aggr_get_id) 1247 return; 1248 1249 if (config->percore_show_thread) 1250 return print_percore_thread(config, counter, prefix); 1251 1252 for (s = 0; s < config->aggr_map->nr; s++) { 1253 if (prefix && metric_only) 1254 fprintf(output, "%s", prefix); 1255 1256 print_counter_aggrdata(config, counter, s, 1257 prefix, metric_only, 1258 &first, (struct perf_cpu){ .cpu = -1 }); 1259 } 1260 1261 if (metric_only) 1262 fputc('\n', output); 1263 } 1264 1265 void evlist__print_counters(struct evlist *evlist, struct perf_stat_config *config, 1266 struct target *_target, struct timespec *ts, int argc, const char **argv) 1267 { 1268 bool metric_only = config->metric_only; 1269 int interval = config->interval; 1270 struct evsel *counter; 1271 char buf[64], *prefix = NULL; 1272 1273 if (config->iostat_run) 1274 evlist->selected = evlist__first(evlist); 1275 1276 if (interval) 1277 print_interval(config, evlist, prefix = buf, ts); 1278 else 1279 print_header(config, _target, argc, argv); 1280 1281 if (metric_only) { 1282 static int num_print_iv; 1283 1284 if (num_print_iv == 0 && !interval) 1285 print_metric_headers(config, evlist, prefix, false); 1286 if (num_print_iv++ == 25) 1287 num_print_iv = 0; 1288 if (config->aggr_mode == AGGR_GLOBAL && prefix && !config->iostat_run) 1289 fprintf(config->output, "%s", prefix); 1290 } 1291 1292 switch (config->aggr_mode) { 1293 case AGGR_CORE: 1294 case AGGR_DIE: 1295 case AGGR_SOCKET: 1296 case AGGR_NODE: 1297 print_aggr(config, evlist, prefix); 1298 break; 1299 case AGGR_THREAD: 1300 evlist__for_each_entry(evlist, counter) { 1301 print_aggr_thread(config, _target, counter, prefix); 1302 } 1303 break; 1304 case AGGR_GLOBAL: 1305 if (config->iostat_run) 1306 iostat_print_counters(evlist, config, ts, prefix = buf, 1307 print_counter_aggr); 1308 else { 1309 evlist__for_each_entry(evlist, counter) { 1310 print_counter_aggr(config, counter, prefix); 1311 } 1312 if (metric_only) 1313 fputc('\n', config->output); 1314 } 1315 break; 1316 case AGGR_NONE: 1317 if (metric_only) 1318 print_no_aggr_metric(config, evlist, prefix); 1319 else { 1320 evlist__for_each_entry(evlist, counter) { 1321 if (counter->percore) 1322 print_percore(config, counter, prefix); 1323 else 1324 print_counter(config, counter, prefix); 1325 } 1326 } 1327 break; 1328 case AGGR_UNSET: 1329 default: 1330 break; 1331 } 1332 1333 if (!interval && !config->csv_output) 1334 print_footer(config); 1335 1336 fflush(config->output); 1337 } 1338