1 #include <stdlib.h> 2 #include <stdio.h> 3 #include <inttypes.h> 4 #include <linux/string.h> 5 #include <linux/time64.h> 6 #include <math.h> 7 #include "color.h" 8 #include "counts.h" 9 #include "evlist.h" 10 #include "evsel.h" 11 #include "stat.h" 12 #include "top.h" 13 #include "thread_map.h" 14 #include "cpumap.h" 15 #include "string2.h" 16 #include <linux/ctype.h> 17 #include "cgroup.h" 18 #include <api/fs/fs.h> 19 #include "util.h" 20 21 #define CNTR_NOT_SUPPORTED "<not supported>" 22 #define CNTR_NOT_COUNTED "<not counted>" 23 24 static void print_running(struct perf_stat_config *config, 25 u64 run, u64 ena) 26 { 27 if (config->csv_output) { 28 fprintf(config->output, "%s%" PRIu64 "%s%.2f", 29 config->csv_sep, 30 run, 31 config->csv_sep, 32 ena ? 100.0 * run / ena : 100.0); 33 } else if (run != ena) { 34 fprintf(config->output, " (%.2f%%)", 100.0 * run / ena); 35 } 36 } 37 38 static void print_noise_pct(struct perf_stat_config *config, 39 double total, double avg) 40 { 41 double pct = rel_stddev_stats(total, avg); 42 43 if (config->csv_output) 44 fprintf(config->output, "%s%.2f%%", config->csv_sep, pct); 45 else if (pct) 46 fprintf(config->output, " ( +-%6.2f%% )", pct); 47 } 48 49 static void print_noise(struct perf_stat_config *config, 50 struct evsel *evsel, double avg) 51 { 52 struct perf_stat_evsel *ps; 53 54 if (config->run_count == 1) 55 return; 56 57 ps = evsel->stats; 58 print_noise_pct(config, stddev_stats(&ps->res_stats[0]), avg); 59 } 60 61 static void print_cgroup(struct perf_stat_config *config, struct evsel *evsel) 62 { 63 if (nr_cgroups) { 64 const char *cgrp_name = evsel->cgrp ? evsel->cgrp->name : ""; 65 fprintf(config->output, "%s%s", config->csv_sep, cgrp_name); 66 } 67 } 68 69 70 static void aggr_printout(struct perf_stat_config *config, 71 struct evsel *evsel, struct aggr_cpu_id id, int nr) 72 { 73 switch (config->aggr_mode) { 74 case AGGR_CORE: 75 fprintf(config->output, "S%d-D%d-C%*d%s%*d%s", 76 id.socket, 77 id.die, 78 config->csv_output ? 0 : -8, 79 id.core, 80 config->csv_sep, 81 config->csv_output ? 0 : 4, 82 nr, 83 config->csv_sep); 84 break; 85 case AGGR_DIE: 86 fprintf(config->output, "S%d-D%*d%s%*d%s", 87 id.socket, 88 config->csv_output ? 0 : -8, 89 id.die, 90 config->csv_sep, 91 config->csv_output ? 0 : 4, 92 nr, 93 config->csv_sep); 94 break; 95 case AGGR_SOCKET: 96 fprintf(config->output, "S%*d%s%*d%s", 97 config->csv_output ? 0 : -5, 98 id.socket, 99 config->csv_sep, 100 config->csv_output ? 0 : 4, 101 nr, 102 config->csv_sep); 103 break; 104 case AGGR_NODE: 105 fprintf(config->output, "N%*d%s%*d%s", 106 config->csv_output ? 0 : -5, 107 id.node, 108 config->csv_sep, 109 config->csv_output ? 0 : 4, 110 nr, 111 config->csv_sep); 112 break; 113 case AGGR_NONE: 114 if (evsel->percore && !config->percore_show_thread) { 115 fprintf(config->output, "S%d-D%d-C%*d%s", 116 id.socket, 117 id.die, 118 config->csv_output ? 0 : -3, 119 id.core, config->csv_sep); 120 } else if (id.core > -1) { 121 fprintf(config->output, "CPU%*d%s", 122 config->csv_output ? 0 : -7, 123 evsel__cpus(evsel)->map[id.core], 124 config->csv_sep); 125 } 126 break; 127 case AGGR_THREAD: 128 fprintf(config->output, "%*s-%*d%s", 129 config->csv_output ? 0 : 16, 130 perf_thread_map__comm(evsel->core.threads, id.thread), 131 config->csv_output ? 0 : -8, 132 perf_thread_map__pid(evsel->core.threads, id.thread), 133 config->csv_sep); 134 break; 135 case AGGR_GLOBAL: 136 case AGGR_UNSET: 137 default: 138 break; 139 } 140 } 141 142 struct outstate { 143 FILE *fh; 144 bool newline; 145 const char *prefix; 146 int nfields; 147 int nr; 148 struct aggr_cpu_id id; 149 struct evsel *evsel; 150 }; 151 152 #define METRIC_LEN 35 153 154 static void new_line_std(struct perf_stat_config *config __maybe_unused, 155 void *ctx) 156 { 157 struct outstate *os = ctx; 158 159 os->newline = true; 160 } 161 162 static void do_new_line_std(struct perf_stat_config *config, 163 struct outstate *os) 164 { 165 fputc('\n', os->fh); 166 fputs(os->prefix, os->fh); 167 aggr_printout(config, os->evsel, os->id, os->nr); 168 if (config->aggr_mode == AGGR_NONE) 169 fprintf(os->fh, " "); 170 fprintf(os->fh, " "); 171 } 172 173 static void print_metric_std(struct perf_stat_config *config, 174 void *ctx, const char *color, const char *fmt, 175 const char *unit, double val) 176 { 177 struct outstate *os = ctx; 178 FILE *out = os->fh; 179 int n; 180 bool newline = os->newline; 181 182 os->newline = false; 183 184 if (unit == NULL || fmt == NULL) { 185 fprintf(out, "%-*s", METRIC_LEN, ""); 186 return; 187 } 188 189 if (newline) 190 do_new_line_std(config, os); 191 192 n = fprintf(out, " # "); 193 if (color) 194 n += color_fprintf(out, color, fmt, val); 195 else 196 n += fprintf(out, fmt, val); 197 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit); 198 } 199 200 static void new_line_csv(struct perf_stat_config *config, void *ctx) 201 { 202 struct outstate *os = ctx; 203 int i; 204 205 fputc('\n', os->fh); 206 if (os->prefix) 207 fprintf(os->fh, "%s%s", os->prefix, config->csv_sep); 208 aggr_printout(config, os->evsel, os->id, os->nr); 209 for (i = 0; i < os->nfields; i++) 210 fputs(config->csv_sep, os->fh); 211 } 212 213 static void print_metric_csv(struct perf_stat_config *config __maybe_unused, 214 void *ctx, 215 const char *color __maybe_unused, 216 const char *fmt, const char *unit, double val) 217 { 218 struct outstate *os = ctx; 219 FILE *out = os->fh; 220 char buf[64], *vals, *ends; 221 222 if (unit == NULL || fmt == NULL) { 223 fprintf(out, "%s%s", config->csv_sep, config->csv_sep); 224 return; 225 } 226 snprintf(buf, sizeof(buf), fmt, val); 227 ends = vals = skip_spaces(buf); 228 while (isdigit(*ends) || *ends == '.') 229 ends++; 230 *ends = 0; 231 fprintf(out, "%s%s%s%s", config->csv_sep, vals, config->csv_sep, skip_spaces(unit)); 232 } 233 234 /* Filter out some columns that don't work well in metrics only mode */ 235 236 static bool valid_only_metric(const char *unit) 237 { 238 if (!unit) 239 return false; 240 if (strstr(unit, "/sec") || 241 strstr(unit, "CPUs utilized")) 242 return false; 243 return true; 244 } 245 246 static const char *fixunit(char *buf, struct evsel *evsel, 247 const char *unit) 248 { 249 if (!strncmp(unit, "of all", 6)) { 250 snprintf(buf, 1024, "%s %s", evsel__name(evsel), 251 unit); 252 return buf; 253 } 254 return unit; 255 } 256 257 static void print_metric_only(struct perf_stat_config *config, 258 void *ctx, const char *color, const char *fmt, 259 const char *unit, double val) 260 { 261 struct outstate *os = ctx; 262 FILE *out = os->fh; 263 char buf[1024], str[1024]; 264 unsigned mlen = config->metric_only_len; 265 266 if (!valid_only_metric(unit)) 267 return; 268 unit = fixunit(buf, os->evsel, unit); 269 if (mlen < strlen(unit)) 270 mlen = strlen(unit) + 1; 271 272 if (color) 273 mlen += strlen(color) + sizeof(PERF_COLOR_RESET) - 1; 274 275 color_snprintf(str, sizeof(str), color ?: "", fmt, val); 276 fprintf(out, "%*s ", mlen, str); 277 } 278 279 static void print_metric_only_csv(struct perf_stat_config *config __maybe_unused, 280 void *ctx, const char *color __maybe_unused, 281 const char *fmt, 282 const char *unit, double val) 283 { 284 struct outstate *os = ctx; 285 FILE *out = os->fh; 286 char buf[64], *vals, *ends; 287 char tbuf[1024]; 288 289 if (!valid_only_metric(unit)) 290 return; 291 unit = fixunit(tbuf, os->evsel, unit); 292 snprintf(buf, sizeof buf, fmt, val); 293 ends = vals = skip_spaces(buf); 294 while (isdigit(*ends) || *ends == '.') 295 ends++; 296 *ends = 0; 297 fprintf(out, "%s%s", vals, config->csv_sep); 298 } 299 300 static void new_line_metric(struct perf_stat_config *config __maybe_unused, 301 void *ctx __maybe_unused) 302 { 303 } 304 305 static void print_metric_header(struct perf_stat_config *config, 306 void *ctx, const char *color __maybe_unused, 307 const char *fmt __maybe_unused, 308 const char *unit, double val __maybe_unused) 309 { 310 struct outstate *os = ctx; 311 char tbuf[1024]; 312 313 if (!valid_only_metric(unit)) 314 return; 315 unit = fixunit(tbuf, os->evsel, unit); 316 if (config->csv_output) 317 fprintf(os->fh, "%s%s", unit, config->csv_sep); 318 else 319 fprintf(os->fh, "%*s ", config->metric_only_len, unit); 320 } 321 322 static int first_shadow_cpu(struct perf_stat_config *config, 323 struct evsel *evsel, struct aggr_cpu_id id) 324 { 325 struct evlist *evlist = evsel->evlist; 326 int i; 327 328 if (config->aggr_mode == AGGR_NONE) 329 return id.core; 330 331 if (!config->aggr_get_id) 332 return 0; 333 334 for (i = 0; i < evsel__nr_cpus(evsel); i++) { 335 int cpu2 = evsel__cpus(evsel)->map[i]; 336 337 if (cpu_map__compare_aggr_cpu_id( 338 config->aggr_get_id(config, evlist->core.cpus, cpu2), 339 id)) { 340 return cpu2; 341 } 342 } 343 return 0; 344 } 345 346 static void abs_printout(struct perf_stat_config *config, 347 struct aggr_cpu_id id, int nr, struct evsel *evsel, double avg) 348 { 349 FILE *output = config->output; 350 double sc = evsel->scale; 351 const char *fmt; 352 353 if (config->csv_output) { 354 fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s"; 355 } else { 356 if (config->big_num) 357 fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s"; 358 else 359 fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s"; 360 } 361 362 aggr_printout(config, evsel, id, nr); 363 364 fprintf(output, fmt, avg, config->csv_sep); 365 366 if (evsel->unit) 367 fprintf(output, "%-*s%s", 368 config->csv_output ? 0 : config->unit_width, 369 evsel->unit, config->csv_sep); 370 371 fprintf(output, "%-*s", config->csv_output ? 0 : 25, evsel__name(evsel)); 372 373 print_cgroup(config, evsel); 374 } 375 376 static bool is_mixed_hw_group(struct evsel *counter) 377 { 378 struct evlist *evlist = counter->evlist; 379 u32 pmu_type = counter->core.attr.type; 380 struct evsel *pos; 381 382 if (counter->core.nr_members < 2) 383 return false; 384 385 evlist__for_each_entry(evlist, pos) { 386 /* software events can be part of any hardware group */ 387 if (pos->core.attr.type == PERF_TYPE_SOFTWARE) 388 continue; 389 if (pmu_type == PERF_TYPE_SOFTWARE) { 390 pmu_type = pos->core.attr.type; 391 continue; 392 } 393 if (pmu_type != pos->core.attr.type) 394 return true; 395 } 396 397 return false; 398 } 399 400 static void printout(struct perf_stat_config *config, struct aggr_cpu_id id, int nr, 401 struct evsel *counter, double uval, 402 char *prefix, u64 run, u64 ena, double noise, 403 struct runtime_stat *st) 404 { 405 struct perf_stat_output_ctx out; 406 struct outstate os = { 407 .fh = config->output, 408 .prefix = prefix ? prefix : "", 409 .id = id, 410 .nr = nr, 411 .evsel = counter, 412 }; 413 print_metric_t pm = print_metric_std; 414 new_line_t nl; 415 416 if (config->metric_only) { 417 nl = new_line_metric; 418 if (config->csv_output) 419 pm = print_metric_only_csv; 420 else 421 pm = print_metric_only; 422 } else 423 nl = new_line_std; 424 425 if (config->csv_output && !config->metric_only) { 426 static int aggr_fields[] = { 427 [AGGR_GLOBAL] = 0, 428 [AGGR_THREAD] = 1, 429 [AGGR_NONE] = 1, 430 [AGGR_SOCKET] = 2, 431 [AGGR_DIE] = 2, 432 [AGGR_CORE] = 2, 433 }; 434 435 pm = print_metric_csv; 436 nl = new_line_csv; 437 os.nfields = 3; 438 os.nfields += aggr_fields[config->aggr_mode]; 439 if (counter->cgrp) 440 os.nfields++; 441 } 442 443 if (!config->no_csv_summary && config->csv_output && 444 config->summary && !config->interval) { 445 fprintf(config->output, "%16s%s", "summary", config->csv_sep); 446 } 447 448 if (run == 0 || ena == 0 || counter->counts->scaled == -1) { 449 if (config->metric_only) { 450 pm(config, &os, NULL, "", "", 0); 451 return; 452 } 453 aggr_printout(config, counter, id, nr); 454 455 fprintf(config->output, "%*s%s", 456 config->csv_output ? 0 : 18, 457 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED, 458 config->csv_sep); 459 460 if (counter->supported) { 461 config->print_free_counters_hint = 1; 462 if (is_mixed_hw_group(counter)) 463 config->print_mixed_hw_group_error = 1; 464 } 465 466 fprintf(config->output, "%-*s%s", 467 config->csv_output ? 0 : config->unit_width, 468 counter->unit, config->csv_sep); 469 470 fprintf(config->output, "%*s", 471 config->csv_output ? 0 : -25, evsel__name(counter)); 472 473 print_cgroup(config, counter); 474 475 if (!config->csv_output) 476 pm(config, &os, NULL, NULL, "", 0); 477 print_noise(config, counter, noise); 478 print_running(config, run, ena); 479 if (config->csv_output) 480 pm(config, &os, NULL, NULL, "", 0); 481 return; 482 } 483 484 if (!config->metric_only) 485 abs_printout(config, id, nr, counter, uval); 486 487 out.print_metric = pm; 488 out.new_line = nl; 489 out.ctx = &os; 490 out.force_header = false; 491 492 if (config->csv_output && !config->metric_only) { 493 print_noise(config, counter, noise); 494 print_running(config, run, ena); 495 } 496 497 perf_stat__print_shadow_stats(config, counter, uval, 498 first_shadow_cpu(config, counter, id), 499 &out, &config->metric_events, st); 500 if (!config->csv_output && !config->metric_only) { 501 print_noise(config, counter, noise); 502 print_running(config, run, ena); 503 } 504 } 505 506 static void aggr_update_shadow(struct perf_stat_config *config, 507 struct evlist *evlist) 508 { 509 int cpu, s; 510 struct aggr_cpu_id s2, id; 511 u64 val; 512 struct evsel *counter; 513 514 for (s = 0; s < config->aggr_map->nr; s++) { 515 id = config->aggr_map->map[s]; 516 evlist__for_each_entry(evlist, counter) { 517 val = 0; 518 for (cpu = 0; cpu < evsel__nr_cpus(counter); cpu++) { 519 s2 = config->aggr_get_id(config, evlist->core.cpus, cpu); 520 if (!cpu_map__compare_aggr_cpu_id(s2, id)) 521 continue; 522 val += perf_counts(counter->counts, cpu, 0)->val; 523 } 524 perf_stat__update_shadow_stats(counter, val, 525 first_shadow_cpu(config, counter, id), 526 &rt_stat); 527 } 528 } 529 } 530 531 static void uniquify_event_name(struct evsel *counter) 532 { 533 char *new_name; 534 char *config; 535 536 if (counter->uniquified_name || 537 !counter->pmu_name || !strncmp(counter->name, counter->pmu_name, 538 strlen(counter->pmu_name))) 539 return; 540 541 config = strchr(counter->name, '/'); 542 if (config) { 543 if (asprintf(&new_name, 544 "%s%s", counter->pmu_name, config) > 0) { 545 free(counter->name); 546 counter->name = new_name; 547 } 548 } else { 549 if (asprintf(&new_name, 550 "%s [%s]", counter->name, counter->pmu_name) > 0) { 551 free(counter->name); 552 counter->name = new_name; 553 } 554 } 555 556 counter->uniquified_name = true; 557 } 558 559 static void collect_all_aliases(struct perf_stat_config *config, struct evsel *counter, 560 void (*cb)(struct perf_stat_config *config, struct evsel *counter, void *data, 561 bool first), 562 void *data) 563 { 564 struct evlist *evlist = counter->evlist; 565 struct evsel *alias; 566 567 alias = list_prepare_entry(counter, &(evlist->core.entries), core.node); 568 list_for_each_entry_continue (alias, &evlist->core.entries, core.node) { 569 if (strcmp(evsel__name(alias), evsel__name(counter)) || 570 alias->scale != counter->scale || 571 alias->cgrp != counter->cgrp || 572 strcmp(alias->unit, counter->unit) || 573 evsel__is_clock(alias) != evsel__is_clock(counter) || 574 !strcmp(alias->pmu_name, counter->pmu_name)) 575 break; 576 alias->merged_stat = true; 577 cb(config, alias, data, false); 578 } 579 } 580 581 static bool collect_data(struct perf_stat_config *config, struct evsel *counter, 582 void (*cb)(struct perf_stat_config *config, struct evsel *counter, void *data, 583 bool first), 584 void *data) 585 { 586 if (counter->merged_stat) 587 return false; 588 cb(config, counter, data, true); 589 if (config->no_merge) 590 uniquify_event_name(counter); 591 else if (counter->auto_merge_stats) 592 collect_all_aliases(config, counter, cb, data); 593 return true; 594 } 595 596 struct aggr_data { 597 u64 ena, run, val; 598 struct aggr_cpu_id id; 599 int nr; 600 int cpu; 601 }; 602 603 static void aggr_cb(struct perf_stat_config *config, 604 struct evsel *counter, void *data, bool first) 605 { 606 struct aggr_data *ad = data; 607 int cpu; 608 struct aggr_cpu_id s2; 609 610 for (cpu = 0; cpu < evsel__nr_cpus(counter); cpu++) { 611 struct perf_counts_values *counts; 612 613 s2 = config->aggr_get_id(config, evsel__cpus(counter), cpu); 614 if (!cpu_map__compare_aggr_cpu_id(s2, ad->id)) 615 continue; 616 if (first) 617 ad->nr++; 618 counts = perf_counts(counter->counts, cpu, 0); 619 /* 620 * When any result is bad, make them all to give 621 * consistent output in interval mode. 622 */ 623 if (counts->ena == 0 || counts->run == 0 || 624 counter->counts->scaled == -1) { 625 ad->ena = 0; 626 ad->run = 0; 627 break; 628 } 629 ad->val += counts->val; 630 ad->ena += counts->ena; 631 ad->run += counts->run; 632 } 633 } 634 635 static void print_counter_aggrdata(struct perf_stat_config *config, 636 struct evsel *counter, int s, 637 char *prefix, bool metric_only, 638 bool *first, int cpu) 639 { 640 struct aggr_data ad; 641 FILE *output = config->output; 642 u64 ena, run, val; 643 int nr; 644 struct aggr_cpu_id id; 645 double uval; 646 647 ad.id = id = config->aggr_map->map[s]; 648 ad.val = ad.ena = ad.run = 0; 649 ad.nr = 0; 650 if (!collect_data(config, counter, aggr_cb, &ad)) 651 return; 652 653 nr = ad.nr; 654 ena = ad.ena; 655 run = ad.run; 656 val = ad.val; 657 if (*first && metric_only) { 658 *first = false; 659 aggr_printout(config, counter, id, nr); 660 } 661 if (prefix && !metric_only) 662 fprintf(output, "%s", prefix); 663 664 uval = val * counter->scale; 665 if (cpu != -1) { 666 id = cpu_map__empty_aggr_cpu_id(); 667 id.core = cpu; 668 } 669 printout(config, id, nr, counter, uval, 670 prefix, run, ena, 1.0, &rt_stat); 671 if (!metric_only) 672 fputc('\n', output); 673 } 674 675 static void print_aggr(struct perf_stat_config *config, 676 struct evlist *evlist, 677 char *prefix) 678 { 679 bool metric_only = config->metric_only; 680 FILE *output = config->output; 681 struct evsel *counter; 682 int s; 683 bool first; 684 685 if (!config->aggr_map || !config->aggr_get_id) 686 return; 687 688 aggr_update_shadow(config, evlist); 689 690 /* 691 * With metric_only everything is on a single line. 692 * Without each counter has its own line. 693 */ 694 for (s = 0; s < config->aggr_map->nr; s++) { 695 if (prefix && metric_only) 696 fprintf(output, "%s", prefix); 697 698 first = true; 699 evlist__for_each_entry(evlist, counter) { 700 print_counter_aggrdata(config, counter, s, 701 prefix, metric_only, 702 &first, -1); 703 } 704 if (metric_only) 705 fputc('\n', output); 706 } 707 } 708 709 static int cmp_val(const void *a, const void *b) 710 { 711 return ((struct perf_aggr_thread_value *)b)->val - 712 ((struct perf_aggr_thread_value *)a)->val; 713 } 714 715 static struct perf_aggr_thread_value *sort_aggr_thread( 716 struct evsel *counter, 717 int nthreads, int ncpus, 718 int *ret, 719 struct target *_target) 720 { 721 int cpu, thread, i = 0; 722 double uval; 723 struct perf_aggr_thread_value *buf; 724 725 buf = calloc(nthreads, sizeof(struct perf_aggr_thread_value)); 726 if (!buf) 727 return NULL; 728 729 for (thread = 0; thread < nthreads; thread++) { 730 u64 ena = 0, run = 0, val = 0; 731 732 for (cpu = 0; cpu < ncpus; cpu++) { 733 val += perf_counts(counter->counts, cpu, thread)->val; 734 ena += perf_counts(counter->counts, cpu, thread)->ena; 735 run += perf_counts(counter->counts, cpu, thread)->run; 736 } 737 738 uval = val * counter->scale; 739 740 /* 741 * Skip value 0 when enabling --per-thread globally, 742 * otherwise too many 0 output. 743 */ 744 if (uval == 0.0 && target__has_per_thread(_target)) 745 continue; 746 747 buf[i].counter = counter; 748 buf[i].id = cpu_map__empty_aggr_cpu_id(); 749 buf[i].id.thread = thread; 750 buf[i].uval = uval; 751 buf[i].val = val; 752 buf[i].run = run; 753 buf[i].ena = ena; 754 i++; 755 } 756 757 qsort(buf, i, sizeof(struct perf_aggr_thread_value), cmp_val); 758 759 if (ret) 760 *ret = i; 761 762 return buf; 763 } 764 765 static void print_aggr_thread(struct perf_stat_config *config, 766 struct target *_target, 767 struct evsel *counter, char *prefix) 768 { 769 FILE *output = config->output; 770 int nthreads = perf_thread_map__nr(counter->core.threads); 771 int ncpus = perf_cpu_map__nr(counter->core.cpus); 772 int thread, sorted_threads; 773 struct aggr_cpu_id id; 774 struct perf_aggr_thread_value *buf; 775 776 buf = sort_aggr_thread(counter, nthreads, ncpus, &sorted_threads, _target); 777 if (!buf) { 778 perror("cannot sort aggr thread"); 779 return; 780 } 781 782 for (thread = 0; thread < sorted_threads; thread++) { 783 if (prefix) 784 fprintf(output, "%s", prefix); 785 786 id = buf[thread].id; 787 if (config->stats) 788 printout(config, id, 0, buf[thread].counter, buf[thread].uval, 789 prefix, buf[thread].run, buf[thread].ena, 1.0, 790 &config->stats[id.thread]); 791 else 792 printout(config, id, 0, buf[thread].counter, buf[thread].uval, 793 prefix, buf[thread].run, buf[thread].ena, 1.0, 794 &rt_stat); 795 fputc('\n', output); 796 } 797 798 free(buf); 799 } 800 801 struct caggr_data { 802 double avg, avg_enabled, avg_running; 803 }; 804 805 static void counter_aggr_cb(struct perf_stat_config *config __maybe_unused, 806 struct evsel *counter, void *data, 807 bool first __maybe_unused) 808 { 809 struct caggr_data *cd = data; 810 struct perf_stat_evsel *ps = counter->stats; 811 812 cd->avg += avg_stats(&ps->res_stats[0]); 813 cd->avg_enabled += avg_stats(&ps->res_stats[1]); 814 cd->avg_running += avg_stats(&ps->res_stats[2]); 815 } 816 817 /* 818 * Print out the results of a single counter: 819 * aggregated counts in system-wide mode 820 */ 821 static void print_counter_aggr(struct perf_stat_config *config, 822 struct evsel *counter, char *prefix) 823 { 824 bool metric_only = config->metric_only; 825 FILE *output = config->output; 826 double uval; 827 struct caggr_data cd = { .avg = 0.0 }; 828 829 if (!collect_data(config, counter, counter_aggr_cb, &cd)) 830 return; 831 832 if (prefix && !metric_only) 833 fprintf(output, "%s", prefix); 834 835 uval = cd.avg * counter->scale; 836 printout(config, cpu_map__empty_aggr_cpu_id(), 0, counter, uval, prefix, cd.avg_running, 837 cd.avg_enabled, cd.avg, &rt_stat); 838 if (!metric_only) 839 fprintf(output, "\n"); 840 } 841 842 static void counter_cb(struct perf_stat_config *config __maybe_unused, 843 struct evsel *counter, void *data, 844 bool first __maybe_unused) 845 { 846 struct aggr_data *ad = data; 847 848 ad->val += perf_counts(counter->counts, ad->cpu, 0)->val; 849 ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena; 850 ad->run += perf_counts(counter->counts, ad->cpu, 0)->run; 851 } 852 853 /* 854 * Print out the results of a single counter: 855 * does not use aggregated count in system-wide 856 */ 857 static void print_counter(struct perf_stat_config *config, 858 struct evsel *counter, char *prefix) 859 { 860 FILE *output = config->output; 861 u64 ena, run, val; 862 double uval; 863 int cpu; 864 struct aggr_cpu_id id; 865 866 for (cpu = 0; cpu < evsel__nr_cpus(counter); cpu++) { 867 struct aggr_data ad = { .cpu = cpu }; 868 869 if (!collect_data(config, counter, counter_cb, &ad)) 870 return; 871 val = ad.val; 872 ena = ad.ena; 873 run = ad.run; 874 875 if (prefix) 876 fprintf(output, "%s", prefix); 877 878 uval = val * counter->scale; 879 id = cpu_map__empty_aggr_cpu_id(); 880 id.core = cpu; 881 printout(config, id, 0, counter, uval, prefix, 882 run, ena, 1.0, &rt_stat); 883 884 fputc('\n', output); 885 } 886 } 887 888 static void print_no_aggr_metric(struct perf_stat_config *config, 889 struct evlist *evlist, 890 char *prefix) 891 { 892 int cpu; 893 int nrcpus = 0; 894 struct evsel *counter; 895 u64 ena, run, val; 896 double uval; 897 struct aggr_cpu_id id; 898 899 nrcpus = evlist->core.cpus->nr; 900 for (cpu = 0; cpu < nrcpus; cpu++) { 901 bool first = true; 902 903 if (prefix) 904 fputs(prefix, config->output); 905 evlist__for_each_entry(evlist, counter) { 906 id = cpu_map__empty_aggr_cpu_id(); 907 id.core = cpu; 908 if (first) { 909 aggr_printout(config, counter, id, 0); 910 first = false; 911 } 912 val = perf_counts(counter->counts, cpu, 0)->val; 913 ena = perf_counts(counter->counts, cpu, 0)->ena; 914 run = perf_counts(counter->counts, cpu, 0)->run; 915 916 uval = val * counter->scale; 917 printout(config, id, 0, counter, uval, prefix, 918 run, ena, 1.0, &rt_stat); 919 } 920 fputc('\n', config->output); 921 } 922 } 923 924 static int aggr_header_lens[] = { 925 [AGGR_CORE] = 24, 926 [AGGR_DIE] = 18, 927 [AGGR_SOCKET] = 12, 928 [AGGR_NONE] = 6, 929 [AGGR_THREAD] = 24, 930 [AGGR_GLOBAL] = 0, 931 }; 932 933 static const char *aggr_header_csv[] = { 934 [AGGR_CORE] = "core,cpus,", 935 [AGGR_DIE] = "die,cpus", 936 [AGGR_SOCKET] = "socket,cpus", 937 [AGGR_NONE] = "cpu,", 938 [AGGR_THREAD] = "comm-pid,", 939 [AGGR_GLOBAL] = "" 940 }; 941 942 static void print_metric_headers(struct perf_stat_config *config, 943 struct evlist *evlist, 944 const char *prefix, bool no_indent) 945 { 946 struct perf_stat_output_ctx out; 947 struct evsel *counter; 948 struct outstate os = { 949 .fh = config->output 950 }; 951 952 if (prefix) 953 fprintf(config->output, "%s", prefix); 954 955 if (!config->csv_output && !no_indent) 956 fprintf(config->output, "%*s", 957 aggr_header_lens[config->aggr_mode], ""); 958 if (config->csv_output) { 959 if (config->interval) 960 fputs("time,", config->output); 961 fputs(aggr_header_csv[config->aggr_mode], config->output); 962 } 963 964 /* Print metrics headers only */ 965 evlist__for_each_entry(evlist, counter) { 966 os.evsel = counter; 967 out.ctx = &os; 968 out.print_metric = print_metric_header; 969 out.new_line = new_line_metric; 970 out.force_header = true; 971 perf_stat__print_shadow_stats(config, counter, 0, 972 0, 973 &out, 974 &config->metric_events, 975 &rt_stat); 976 } 977 fputc('\n', config->output); 978 } 979 980 static void print_interval(struct perf_stat_config *config, 981 struct evlist *evlist, 982 char *prefix, struct timespec *ts) 983 { 984 bool metric_only = config->metric_only; 985 unsigned int unit_width = config->unit_width; 986 FILE *output = config->output; 987 static int num_print_interval; 988 989 if (config->interval_clear) 990 puts(CONSOLE_CLEAR); 991 992 sprintf(prefix, "%6lu.%09lu%s", (unsigned long) ts->tv_sec, ts->tv_nsec, config->csv_sep); 993 994 if ((num_print_interval == 0 && !config->csv_output) || config->interval_clear) { 995 switch (config->aggr_mode) { 996 case AGGR_NODE: 997 fprintf(output, "# time node cpus"); 998 if (!metric_only) 999 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1000 break; 1001 case AGGR_SOCKET: 1002 fprintf(output, "# time socket cpus"); 1003 if (!metric_only) 1004 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1005 break; 1006 case AGGR_DIE: 1007 fprintf(output, "# time die cpus"); 1008 if (!metric_only) 1009 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1010 break; 1011 case AGGR_CORE: 1012 fprintf(output, "# time core cpus"); 1013 if (!metric_only) 1014 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1015 break; 1016 case AGGR_NONE: 1017 fprintf(output, "# time CPU "); 1018 if (!metric_only) 1019 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1020 break; 1021 case AGGR_THREAD: 1022 fprintf(output, "# time comm-pid"); 1023 if (!metric_only) 1024 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1025 break; 1026 case AGGR_GLOBAL: 1027 default: 1028 fprintf(output, "# time"); 1029 if (!metric_only) 1030 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1031 case AGGR_UNSET: 1032 break; 1033 } 1034 } 1035 1036 if ((num_print_interval == 0 || config->interval_clear) && metric_only) 1037 print_metric_headers(config, evlist, " ", true); 1038 if (++num_print_interval == 25) 1039 num_print_interval = 0; 1040 } 1041 1042 static void print_header(struct perf_stat_config *config, 1043 struct target *_target, 1044 int argc, const char **argv) 1045 { 1046 FILE *output = config->output; 1047 int i; 1048 1049 fflush(stdout); 1050 1051 if (!config->csv_output) { 1052 fprintf(output, "\n"); 1053 fprintf(output, " Performance counter stats for "); 1054 if (_target->bpf_str) 1055 fprintf(output, "\'BPF program(s) %s", _target->bpf_str); 1056 else if (_target->system_wide) 1057 fprintf(output, "\'system wide"); 1058 else if (_target->cpu_list) 1059 fprintf(output, "\'CPU(s) %s", _target->cpu_list); 1060 else if (!target__has_task(_target)) { 1061 fprintf(output, "\'%s", argv ? argv[0] : "pipe"); 1062 for (i = 1; argv && (i < argc); i++) 1063 fprintf(output, " %s", argv[i]); 1064 } else if (_target->pid) 1065 fprintf(output, "process id \'%s", _target->pid); 1066 else 1067 fprintf(output, "thread id \'%s", _target->tid); 1068 1069 fprintf(output, "\'"); 1070 if (config->run_count > 1) 1071 fprintf(output, " (%d runs)", config->run_count); 1072 fprintf(output, ":\n\n"); 1073 } 1074 } 1075 1076 static int get_precision(double num) 1077 { 1078 if (num > 1) 1079 return 0; 1080 1081 return lround(ceil(-log10(num))); 1082 } 1083 1084 static void print_table(struct perf_stat_config *config, 1085 FILE *output, int precision, double avg) 1086 { 1087 char tmp[64]; 1088 int idx, indent = 0; 1089 1090 scnprintf(tmp, 64, " %17.*f", precision, avg); 1091 while (tmp[indent] == ' ') 1092 indent++; 1093 1094 fprintf(output, "%*s# Table of individual measurements:\n", indent, ""); 1095 1096 for (idx = 0; idx < config->run_count; idx++) { 1097 double run = (double) config->walltime_run[idx] / NSEC_PER_SEC; 1098 int h, n = 1 + abs((int) (100.0 * (run - avg)/run) / 5); 1099 1100 fprintf(output, " %17.*f (%+.*f) ", 1101 precision, run, precision, run - avg); 1102 1103 for (h = 0; h < n; h++) 1104 fprintf(output, "#"); 1105 1106 fprintf(output, "\n"); 1107 } 1108 1109 fprintf(output, "\n%*s# Final result:\n", indent, ""); 1110 } 1111 1112 static double timeval2double(struct timeval *t) 1113 { 1114 return t->tv_sec + (double) t->tv_usec/USEC_PER_SEC; 1115 } 1116 1117 static void print_footer(struct perf_stat_config *config) 1118 { 1119 double avg = avg_stats(config->walltime_nsecs_stats) / NSEC_PER_SEC; 1120 FILE *output = config->output; 1121 1122 if (!config->null_run) 1123 fprintf(output, "\n"); 1124 1125 if (config->run_count == 1) { 1126 fprintf(output, " %17.9f seconds time elapsed", avg); 1127 1128 if (config->ru_display) { 1129 double ru_utime = timeval2double(&config->ru_data.ru_utime); 1130 double ru_stime = timeval2double(&config->ru_data.ru_stime); 1131 1132 fprintf(output, "\n\n"); 1133 fprintf(output, " %17.9f seconds user\n", ru_utime); 1134 fprintf(output, " %17.9f seconds sys\n", ru_stime); 1135 } 1136 } else { 1137 double sd = stddev_stats(config->walltime_nsecs_stats) / NSEC_PER_SEC; 1138 /* 1139 * Display at most 2 more significant 1140 * digits than the stddev inaccuracy. 1141 */ 1142 int precision = get_precision(sd) + 2; 1143 1144 if (config->walltime_run_table) 1145 print_table(config, output, precision, avg); 1146 1147 fprintf(output, " %17.*f +- %.*f seconds time elapsed", 1148 precision, avg, precision, sd); 1149 1150 print_noise_pct(config, sd, avg); 1151 } 1152 fprintf(output, "\n\n"); 1153 1154 if (config->print_free_counters_hint && sysctl__nmi_watchdog_enabled()) 1155 fprintf(output, 1156 "Some events weren't counted. Try disabling the NMI watchdog:\n" 1157 " echo 0 > /proc/sys/kernel/nmi_watchdog\n" 1158 " perf stat ...\n" 1159 " echo 1 > /proc/sys/kernel/nmi_watchdog\n"); 1160 1161 if (config->print_mixed_hw_group_error) 1162 fprintf(output, 1163 "The events in group usually have to be from " 1164 "the same PMU. Try reorganizing the group.\n"); 1165 } 1166 1167 static void print_percore_thread(struct perf_stat_config *config, 1168 struct evsel *counter, char *prefix) 1169 { 1170 int s; 1171 struct aggr_cpu_id s2, id; 1172 bool first = true; 1173 1174 for (int i = 0; i < evsel__nr_cpus(counter); i++) { 1175 s2 = config->aggr_get_id(config, evsel__cpus(counter), i); 1176 for (s = 0; s < config->aggr_map->nr; s++) { 1177 id = config->aggr_map->map[s]; 1178 if (cpu_map__compare_aggr_cpu_id(s2, id)) 1179 break; 1180 } 1181 1182 print_counter_aggrdata(config, counter, s, 1183 prefix, false, 1184 &first, i); 1185 } 1186 } 1187 1188 static void print_percore(struct perf_stat_config *config, 1189 struct evsel *counter, char *prefix) 1190 { 1191 bool metric_only = config->metric_only; 1192 FILE *output = config->output; 1193 int s; 1194 bool first = true; 1195 1196 if (!config->aggr_map || !config->aggr_get_id) 1197 return; 1198 1199 if (config->percore_show_thread) 1200 return print_percore_thread(config, counter, prefix); 1201 1202 for (s = 0; s < config->aggr_map->nr; s++) { 1203 if (prefix && metric_only) 1204 fprintf(output, "%s", prefix); 1205 1206 print_counter_aggrdata(config, counter, s, 1207 prefix, metric_only, 1208 &first, -1); 1209 } 1210 1211 if (metric_only) 1212 fputc('\n', output); 1213 } 1214 1215 void evlist__print_counters(struct evlist *evlist, struct perf_stat_config *config, 1216 struct target *_target, struct timespec *ts, int argc, const char **argv) 1217 { 1218 bool metric_only = config->metric_only; 1219 int interval = config->interval; 1220 struct evsel *counter; 1221 char buf[64], *prefix = NULL; 1222 1223 if (interval) 1224 print_interval(config, evlist, prefix = buf, ts); 1225 else 1226 print_header(config, _target, argc, argv); 1227 1228 if (metric_only) { 1229 static int num_print_iv; 1230 1231 if (num_print_iv == 0 && !interval) 1232 print_metric_headers(config, evlist, prefix, false); 1233 if (num_print_iv++ == 25) 1234 num_print_iv = 0; 1235 if (config->aggr_mode == AGGR_GLOBAL && prefix) 1236 fprintf(config->output, "%s", prefix); 1237 } 1238 1239 switch (config->aggr_mode) { 1240 case AGGR_CORE: 1241 case AGGR_DIE: 1242 case AGGR_SOCKET: 1243 case AGGR_NODE: 1244 print_aggr(config, evlist, prefix); 1245 break; 1246 case AGGR_THREAD: 1247 evlist__for_each_entry(evlist, counter) { 1248 print_aggr_thread(config, _target, counter, prefix); 1249 } 1250 break; 1251 case AGGR_GLOBAL: 1252 evlist__for_each_entry(evlist, counter) { 1253 print_counter_aggr(config, counter, prefix); 1254 } 1255 if (metric_only) 1256 fputc('\n', config->output); 1257 break; 1258 case AGGR_NONE: 1259 if (metric_only) 1260 print_no_aggr_metric(config, evlist, prefix); 1261 else { 1262 evlist__for_each_entry(evlist, counter) { 1263 if (counter->percore) 1264 print_percore(config, counter, prefix); 1265 else 1266 print_counter(config, counter, prefix); 1267 } 1268 } 1269 break; 1270 case AGGR_UNSET: 1271 default: 1272 break; 1273 } 1274 1275 if (!interval && !config->csv_output) 1276 print_footer(config); 1277 1278 fflush(config->output); 1279 } 1280