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