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 if (id > -1) { 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", 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 < 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, 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, evsel__name(counter)); 465 466 print_cgroup(config, counter); 467 468 if (!config->csv_output) 469 pm(config, &os, NULL, NULL, "", 0); 470 print_noise(config, counter, noise); 471 print_running(config, run, ena); 472 if (config->csv_output) 473 pm(config, &os, NULL, NULL, "", 0); 474 return; 475 } 476 477 if (!config->metric_only) 478 abs_printout(config, id, nr, counter, uval); 479 480 out.print_metric = pm; 481 out.new_line = nl; 482 out.ctx = &os; 483 out.force_header = false; 484 485 if (config->csv_output && !config->metric_only) { 486 print_noise(config, counter, noise); 487 print_running(config, run, ena); 488 } 489 490 perf_stat__print_shadow_stats(config, counter, uval, 491 first_shadow_cpu(config, counter, id), 492 &out, &config->metric_events, st); 493 if (!config->csv_output && !config->metric_only) { 494 print_noise(config, counter, noise); 495 print_running(config, run, ena); 496 } 497 } 498 499 static void aggr_update_shadow(struct perf_stat_config *config, 500 struct evlist *evlist) 501 { 502 int cpu, s2, id, s; 503 u64 val; 504 struct evsel *counter; 505 506 for (s = 0; s < config->aggr_map->nr; s++) { 507 id = config->aggr_map->map[s]; 508 evlist__for_each_entry(evlist, counter) { 509 val = 0; 510 for (cpu = 0; cpu < evsel__nr_cpus(counter); cpu++) { 511 s2 = config->aggr_get_id(config, evlist->core.cpus, cpu); 512 if (s2 != id) 513 continue; 514 val += perf_counts(counter->counts, cpu, 0)->val; 515 } 516 perf_stat__update_shadow_stats(counter, val, 517 first_shadow_cpu(config, counter, id), 518 &rt_stat); 519 } 520 } 521 } 522 523 static void uniquify_event_name(struct evsel *counter) 524 { 525 char *new_name; 526 char *config; 527 528 if (counter->uniquified_name || 529 !counter->pmu_name || !strncmp(counter->name, counter->pmu_name, 530 strlen(counter->pmu_name))) 531 return; 532 533 config = strchr(counter->name, '/'); 534 if (config) { 535 if (asprintf(&new_name, 536 "%s%s", counter->pmu_name, config) > 0) { 537 free(counter->name); 538 counter->name = new_name; 539 } 540 } else { 541 if (asprintf(&new_name, 542 "%s [%s]", counter->name, counter->pmu_name) > 0) { 543 free(counter->name); 544 counter->name = new_name; 545 } 546 } 547 548 counter->uniquified_name = true; 549 } 550 551 static void collect_all_aliases(struct perf_stat_config *config, struct evsel *counter, 552 void (*cb)(struct perf_stat_config *config, struct evsel *counter, void *data, 553 bool first), 554 void *data) 555 { 556 struct evlist *evlist = counter->evlist; 557 struct evsel *alias; 558 559 alias = list_prepare_entry(counter, &(evlist->core.entries), core.node); 560 list_for_each_entry_continue (alias, &evlist->core.entries, core.node) { 561 if (strcmp(evsel__name(alias), evsel__name(counter)) || 562 alias->scale != counter->scale || 563 alias->cgrp != counter->cgrp || 564 strcmp(alias->unit, counter->unit) || 565 evsel__is_clock(alias) != evsel__is_clock(counter) || 566 !strcmp(alias->pmu_name, counter->pmu_name)) 567 break; 568 alias->merged_stat = true; 569 cb(config, alias, data, false); 570 } 571 } 572 573 static bool collect_data(struct perf_stat_config *config, struct evsel *counter, 574 void (*cb)(struct perf_stat_config *config, struct evsel *counter, void *data, 575 bool first), 576 void *data) 577 { 578 if (counter->merged_stat) 579 return false; 580 cb(config, counter, data, true); 581 if (config->no_merge) 582 uniquify_event_name(counter); 583 else if (counter->auto_merge_stats) 584 collect_all_aliases(config, counter, cb, data); 585 return true; 586 } 587 588 struct aggr_data { 589 u64 ena, run, val; 590 int id; 591 int nr; 592 int cpu; 593 }; 594 595 static void aggr_cb(struct perf_stat_config *config, 596 struct evsel *counter, void *data, bool first) 597 { 598 struct aggr_data *ad = data; 599 int cpu, s2; 600 601 for (cpu = 0; cpu < evsel__nr_cpus(counter); cpu++) { 602 struct perf_counts_values *counts; 603 604 s2 = config->aggr_get_id(config, evsel__cpus(counter), cpu); 605 if (s2 != ad->id) 606 continue; 607 if (first) 608 ad->nr++; 609 counts = perf_counts(counter->counts, cpu, 0); 610 /* 611 * When any result is bad, make them all to give 612 * consistent output in interval mode. 613 */ 614 if (counts->ena == 0 || counts->run == 0 || 615 counter->counts->scaled == -1) { 616 ad->ena = 0; 617 ad->run = 0; 618 break; 619 } 620 ad->val += counts->val; 621 ad->ena += counts->ena; 622 ad->run += counts->run; 623 } 624 } 625 626 static void print_counter_aggrdata(struct perf_stat_config *config, 627 struct evsel *counter, int s, 628 char *prefix, bool metric_only, 629 bool *first, int cpu) 630 { 631 struct aggr_data ad; 632 FILE *output = config->output; 633 u64 ena, run, val; 634 int id, nr; 635 double uval; 636 637 ad.id = id = config->aggr_map->map[s]; 638 ad.val = ad.ena = ad.run = 0; 639 ad.nr = 0; 640 if (!collect_data(config, counter, aggr_cb, &ad)) 641 return; 642 643 nr = ad.nr; 644 ena = ad.ena; 645 run = ad.run; 646 val = ad.val; 647 if (*first && metric_only) { 648 *first = false; 649 aggr_printout(config, counter, id, nr); 650 } 651 if (prefix && !metric_only) 652 fprintf(output, "%s", prefix); 653 654 uval = val * counter->scale; 655 printout(config, cpu != -1 ? cpu : id, nr, counter, uval, prefix, 656 run, ena, 1.0, &rt_stat); 657 if (!metric_only) 658 fputc('\n', output); 659 } 660 661 static void print_aggr(struct perf_stat_config *config, 662 struct evlist *evlist, 663 char *prefix) 664 { 665 bool metric_only = config->metric_only; 666 FILE *output = config->output; 667 struct evsel *counter; 668 int s; 669 bool first; 670 671 if (!config->aggr_map || !config->aggr_get_id) 672 return; 673 674 aggr_update_shadow(config, evlist); 675 676 /* 677 * With metric_only everything is on a single line. 678 * Without each counter has its own line. 679 */ 680 for (s = 0; s < config->aggr_map->nr; s++) { 681 if (prefix && metric_only) 682 fprintf(output, "%s", prefix); 683 684 first = true; 685 evlist__for_each_entry(evlist, counter) { 686 print_counter_aggrdata(config, counter, s, 687 prefix, metric_only, 688 &first, -1); 689 } 690 if (metric_only) 691 fputc('\n', output); 692 } 693 } 694 695 static int cmp_val(const void *a, const void *b) 696 { 697 return ((struct perf_aggr_thread_value *)b)->val - 698 ((struct perf_aggr_thread_value *)a)->val; 699 } 700 701 static struct perf_aggr_thread_value *sort_aggr_thread( 702 struct evsel *counter, 703 int nthreads, int ncpus, 704 int *ret, 705 struct target *_target) 706 { 707 int cpu, thread, i = 0; 708 double uval; 709 struct perf_aggr_thread_value *buf; 710 711 buf = calloc(nthreads, sizeof(struct perf_aggr_thread_value)); 712 if (!buf) 713 return NULL; 714 715 for (thread = 0; thread < nthreads; thread++) { 716 u64 ena = 0, run = 0, val = 0; 717 718 for (cpu = 0; cpu < ncpus; cpu++) { 719 val += perf_counts(counter->counts, cpu, thread)->val; 720 ena += perf_counts(counter->counts, cpu, thread)->ena; 721 run += perf_counts(counter->counts, cpu, thread)->run; 722 } 723 724 uval = val * counter->scale; 725 726 /* 727 * Skip value 0 when enabling --per-thread globally, 728 * otherwise too many 0 output. 729 */ 730 if (uval == 0.0 && target__has_per_thread(_target)) 731 continue; 732 733 buf[i].counter = counter; 734 buf[i].id = thread; 735 buf[i].uval = uval; 736 buf[i].val = val; 737 buf[i].run = run; 738 buf[i].ena = ena; 739 i++; 740 } 741 742 qsort(buf, i, sizeof(struct perf_aggr_thread_value), cmp_val); 743 744 if (ret) 745 *ret = i; 746 747 return buf; 748 } 749 750 static void print_aggr_thread(struct perf_stat_config *config, 751 struct target *_target, 752 struct evsel *counter, char *prefix) 753 { 754 FILE *output = config->output; 755 int nthreads = perf_thread_map__nr(counter->core.threads); 756 int ncpus = perf_cpu_map__nr(counter->core.cpus); 757 int thread, sorted_threads, id; 758 struct perf_aggr_thread_value *buf; 759 760 buf = sort_aggr_thread(counter, nthreads, ncpus, &sorted_threads, _target); 761 if (!buf) { 762 perror("cannot sort aggr thread"); 763 return; 764 } 765 766 for (thread = 0; thread < sorted_threads; thread++) { 767 if (prefix) 768 fprintf(output, "%s", prefix); 769 770 id = buf[thread].id; 771 if (config->stats) 772 printout(config, id, 0, buf[thread].counter, buf[thread].uval, 773 prefix, buf[thread].run, buf[thread].ena, 1.0, 774 &config->stats[id]); 775 else 776 printout(config, id, 0, buf[thread].counter, buf[thread].uval, 777 prefix, buf[thread].run, buf[thread].ena, 1.0, 778 &rt_stat); 779 fputc('\n', output); 780 } 781 782 free(buf); 783 } 784 785 struct caggr_data { 786 double avg, avg_enabled, avg_running; 787 }; 788 789 static void counter_aggr_cb(struct perf_stat_config *config __maybe_unused, 790 struct evsel *counter, void *data, 791 bool first __maybe_unused) 792 { 793 struct caggr_data *cd = data; 794 struct perf_stat_evsel *ps = counter->stats; 795 796 cd->avg += avg_stats(&ps->res_stats[0]); 797 cd->avg_enabled += avg_stats(&ps->res_stats[1]); 798 cd->avg_running += avg_stats(&ps->res_stats[2]); 799 } 800 801 /* 802 * Print out the results of a single counter: 803 * aggregated counts in system-wide mode 804 */ 805 static void print_counter_aggr(struct perf_stat_config *config, 806 struct evsel *counter, char *prefix) 807 { 808 bool metric_only = config->metric_only; 809 FILE *output = config->output; 810 double uval; 811 struct caggr_data cd = { .avg = 0.0 }; 812 813 if (!collect_data(config, counter, counter_aggr_cb, &cd)) 814 return; 815 816 if (prefix && !metric_only) 817 fprintf(output, "%s", prefix); 818 819 uval = cd.avg * counter->scale; 820 printout(config, -1, 0, counter, uval, prefix, cd.avg_running, cd.avg_enabled, 821 cd.avg, &rt_stat); 822 if (!metric_only) 823 fprintf(output, "\n"); 824 } 825 826 static void counter_cb(struct perf_stat_config *config __maybe_unused, 827 struct evsel *counter, void *data, 828 bool first __maybe_unused) 829 { 830 struct aggr_data *ad = data; 831 832 ad->val += perf_counts(counter->counts, ad->cpu, 0)->val; 833 ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena; 834 ad->run += perf_counts(counter->counts, ad->cpu, 0)->run; 835 } 836 837 /* 838 * Print out the results of a single counter: 839 * does not use aggregated count in system-wide 840 */ 841 static void print_counter(struct perf_stat_config *config, 842 struct evsel *counter, char *prefix) 843 { 844 FILE *output = config->output; 845 u64 ena, run, val; 846 double uval; 847 int cpu; 848 849 for (cpu = 0; cpu < evsel__nr_cpus(counter); cpu++) { 850 struct aggr_data ad = { .cpu = cpu }; 851 852 if (!collect_data(config, counter, counter_cb, &ad)) 853 return; 854 val = ad.val; 855 ena = ad.ena; 856 run = ad.run; 857 858 if (prefix) 859 fprintf(output, "%s", prefix); 860 861 uval = val * counter->scale; 862 printout(config, cpu, 0, counter, uval, prefix, run, ena, 1.0, 863 &rt_stat); 864 865 fputc('\n', output); 866 } 867 } 868 869 static void print_no_aggr_metric(struct perf_stat_config *config, 870 struct evlist *evlist, 871 char *prefix) 872 { 873 int cpu; 874 int nrcpus = 0; 875 struct evsel *counter; 876 u64 ena, run, val; 877 double uval; 878 879 nrcpus = evlist->core.cpus->nr; 880 for (cpu = 0; cpu < nrcpus; cpu++) { 881 bool first = true; 882 883 if (prefix) 884 fputs(prefix, config->output); 885 evlist__for_each_entry(evlist, counter) { 886 if (first) { 887 aggr_printout(config, counter, cpu, 0); 888 first = false; 889 } 890 val = perf_counts(counter->counts, cpu, 0)->val; 891 ena = perf_counts(counter->counts, cpu, 0)->ena; 892 run = perf_counts(counter->counts, cpu, 0)->run; 893 894 uval = val * counter->scale; 895 printout(config, cpu, 0, counter, uval, prefix, run, ena, 1.0, 896 &rt_stat); 897 } 898 fputc('\n', config->output); 899 } 900 } 901 902 static int aggr_header_lens[] = { 903 [AGGR_CORE] = 24, 904 [AGGR_DIE] = 18, 905 [AGGR_SOCKET] = 12, 906 [AGGR_NONE] = 6, 907 [AGGR_THREAD] = 24, 908 [AGGR_GLOBAL] = 0, 909 }; 910 911 static const char *aggr_header_csv[] = { 912 [AGGR_CORE] = "core,cpus,", 913 [AGGR_DIE] = "die,cpus", 914 [AGGR_SOCKET] = "socket,cpus", 915 [AGGR_NONE] = "cpu,", 916 [AGGR_THREAD] = "comm-pid,", 917 [AGGR_GLOBAL] = "" 918 }; 919 920 static void print_metric_headers(struct perf_stat_config *config, 921 struct evlist *evlist, 922 const char *prefix, bool no_indent) 923 { 924 struct perf_stat_output_ctx out; 925 struct evsel *counter; 926 struct outstate os = { 927 .fh = config->output 928 }; 929 930 if (prefix) 931 fprintf(config->output, "%s", prefix); 932 933 if (!config->csv_output && !no_indent) 934 fprintf(config->output, "%*s", 935 aggr_header_lens[config->aggr_mode], ""); 936 if (config->csv_output) { 937 if (config->interval) 938 fputs("time,", config->output); 939 fputs(aggr_header_csv[config->aggr_mode], config->output); 940 } 941 942 /* Print metrics headers only */ 943 evlist__for_each_entry(evlist, counter) { 944 os.evsel = counter; 945 out.ctx = &os; 946 out.print_metric = print_metric_header; 947 out.new_line = new_line_metric; 948 out.force_header = true; 949 perf_stat__print_shadow_stats(config, counter, 0, 950 0, 951 &out, 952 &config->metric_events, 953 &rt_stat); 954 } 955 fputc('\n', config->output); 956 } 957 958 static void print_interval(struct perf_stat_config *config, 959 struct evlist *evlist, 960 char *prefix, struct timespec *ts) 961 { 962 bool metric_only = config->metric_only; 963 unsigned int unit_width = config->unit_width; 964 FILE *output = config->output; 965 static int num_print_interval; 966 967 if (config->interval_clear) 968 puts(CONSOLE_CLEAR); 969 970 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, config->csv_sep); 971 972 if ((num_print_interval == 0 && !config->csv_output) || config->interval_clear) { 973 switch (config->aggr_mode) { 974 case AGGR_NODE: 975 fprintf(output, "# time node cpus"); 976 if (!metric_only) 977 fprintf(output, " counts %*s events\n", unit_width, "unit"); 978 break; 979 case AGGR_SOCKET: 980 fprintf(output, "# time socket cpus"); 981 if (!metric_only) 982 fprintf(output, " counts %*s events\n", unit_width, "unit"); 983 break; 984 case AGGR_DIE: 985 fprintf(output, "# time die cpus"); 986 if (!metric_only) 987 fprintf(output, " counts %*s events\n", unit_width, "unit"); 988 break; 989 case AGGR_CORE: 990 fprintf(output, "# time core cpus"); 991 if (!metric_only) 992 fprintf(output, " counts %*s events\n", unit_width, "unit"); 993 break; 994 case AGGR_NONE: 995 fprintf(output, "# time CPU "); 996 if (!metric_only) 997 fprintf(output, " counts %*s events\n", unit_width, "unit"); 998 break; 999 case AGGR_THREAD: 1000 fprintf(output, "# time comm-pid"); 1001 if (!metric_only) 1002 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1003 break; 1004 case AGGR_GLOBAL: 1005 default: 1006 fprintf(output, "# time"); 1007 if (!metric_only) 1008 fprintf(output, " counts %*s events\n", unit_width, "unit"); 1009 case AGGR_UNSET: 1010 break; 1011 } 1012 } 1013 1014 if ((num_print_interval == 0 || config->interval_clear) && metric_only) 1015 print_metric_headers(config, evlist, " ", true); 1016 if (++num_print_interval == 25) 1017 num_print_interval = 0; 1018 } 1019 1020 static void print_header(struct perf_stat_config *config, 1021 struct target *_target, 1022 int argc, const char **argv) 1023 { 1024 FILE *output = config->output; 1025 int i; 1026 1027 fflush(stdout); 1028 1029 if (!config->csv_output) { 1030 fprintf(output, "\n"); 1031 fprintf(output, " Performance counter stats for "); 1032 if (_target->system_wide) 1033 fprintf(output, "\'system wide"); 1034 else if (_target->cpu_list) 1035 fprintf(output, "\'CPU(s) %s", _target->cpu_list); 1036 else if (!target__has_task(_target)) { 1037 fprintf(output, "\'%s", argv ? argv[0] : "pipe"); 1038 for (i = 1; argv && (i < argc); i++) 1039 fprintf(output, " %s", argv[i]); 1040 } else if (_target->pid) 1041 fprintf(output, "process id \'%s", _target->pid); 1042 else 1043 fprintf(output, "thread id \'%s", _target->tid); 1044 1045 fprintf(output, "\'"); 1046 if (config->run_count > 1) 1047 fprintf(output, " (%d runs)", config->run_count); 1048 fprintf(output, ":\n\n"); 1049 } 1050 } 1051 1052 static int get_precision(double num) 1053 { 1054 if (num > 1) 1055 return 0; 1056 1057 return lround(ceil(-log10(num))); 1058 } 1059 1060 static void print_table(struct perf_stat_config *config, 1061 FILE *output, int precision, double avg) 1062 { 1063 char tmp[64]; 1064 int idx, indent = 0; 1065 1066 scnprintf(tmp, 64, " %17.*f", precision, avg); 1067 while (tmp[indent] == ' ') 1068 indent++; 1069 1070 fprintf(output, "%*s# Table of individual measurements:\n", indent, ""); 1071 1072 for (idx = 0; idx < config->run_count; idx++) { 1073 double run = (double) config->walltime_run[idx] / NSEC_PER_SEC; 1074 int h, n = 1 + abs((int) (100.0 * (run - avg)/run) / 5); 1075 1076 fprintf(output, " %17.*f (%+.*f) ", 1077 precision, run, precision, run - avg); 1078 1079 for (h = 0; h < n; h++) 1080 fprintf(output, "#"); 1081 1082 fprintf(output, "\n"); 1083 } 1084 1085 fprintf(output, "\n%*s# Final result:\n", indent, ""); 1086 } 1087 1088 static double timeval2double(struct timeval *t) 1089 { 1090 return t->tv_sec + (double) t->tv_usec/USEC_PER_SEC; 1091 } 1092 1093 static void print_footer(struct perf_stat_config *config) 1094 { 1095 double avg = avg_stats(config->walltime_nsecs_stats) / NSEC_PER_SEC; 1096 FILE *output = config->output; 1097 1098 if (!config->null_run) 1099 fprintf(output, "\n"); 1100 1101 if (config->run_count == 1) { 1102 fprintf(output, " %17.9f seconds time elapsed", avg); 1103 1104 if (config->ru_display) { 1105 double ru_utime = timeval2double(&config->ru_data.ru_utime); 1106 double ru_stime = timeval2double(&config->ru_data.ru_stime); 1107 1108 fprintf(output, "\n\n"); 1109 fprintf(output, " %17.9f seconds user\n", ru_utime); 1110 fprintf(output, " %17.9f seconds sys\n", ru_stime); 1111 } 1112 } else { 1113 double sd = stddev_stats(config->walltime_nsecs_stats) / NSEC_PER_SEC; 1114 /* 1115 * Display at most 2 more significant 1116 * digits than the stddev inaccuracy. 1117 */ 1118 int precision = get_precision(sd) + 2; 1119 1120 if (config->walltime_run_table) 1121 print_table(config, output, precision, avg); 1122 1123 fprintf(output, " %17.*f +- %.*f seconds time elapsed", 1124 precision, avg, precision, sd); 1125 1126 print_noise_pct(config, sd, avg); 1127 } 1128 fprintf(output, "\n\n"); 1129 1130 if (config->print_free_counters_hint && sysctl__nmi_watchdog_enabled()) 1131 fprintf(output, 1132 "Some events weren't counted. Try disabling the NMI watchdog:\n" 1133 " echo 0 > /proc/sys/kernel/nmi_watchdog\n" 1134 " perf stat ...\n" 1135 " echo 1 > /proc/sys/kernel/nmi_watchdog\n"); 1136 1137 if (config->print_mixed_hw_group_error) 1138 fprintf(output, 1139 "The events in group usually have to be from " 1140 "the same PMU. Try reorganizing the group.\n"); 1141 } 1142 1143 static void print_percore_thread(struct perf_stat_config *config, 1144 struct evsel *counter, char *prefix) 1145 { 1146 int s, s2, id; 1147 bool first = true; 1148 1149 for (int i = 0; i < evsel__nr_cpus(counter); i++) { 1150 s2 = config->aggr_get_id(config, evsel__cpus(counter), i); 1151 for (s = 0; s < config->aggr_map->nr; s++) { 1152 id = config->aggr_map->map[s]; 1153 if (s2 == id) 1154 break; 1155 } 1156 1157 print_counter_aggrdata(config, counter, s, 1158 prefix, false, 1159 &first, i); 1160 } 1161 } 1162 1163 static void print_percore(struct perf_stat_config *config, 1164 struct evsel *counter, char *prefix) 1165 { 1166 bool metric_only = config->metric_only; 1167 FILE *output = config->output; 1168 int s; 1169 bool first = true; 1170 1171 if (!config->aggr_map || !config->aggr_get_id) 1172 return; 1173 1174 if (config->percore_show_thread) 1175 return print_percore_thread(config, counter, prefix); 1176 1177 for (s = 0; s < config->aggr_map->nr; s++) { 1178 if (prefix && metric_only) 1179 fprintf(output, "%s", prefix); 1180 1181 print_counter_aggrdata(config, counter, s, 1182 prefix, metric_only, 1183 &first, -1); 1184 } 1185 1186 if (metric_only) 1187 fputc('\n', output); 1188 } 1189 1190 void 1191 perf_evlist__print_counters(struct evlist *evlist, 1192 struct perf_stat_config *config, 1193 struct target *_target, 1194 struct timespec *ts, 1195 int argc, const char **argv) 1196 { 1197 bool metric_only = config->metric_only; 1198 int interval = config->interval; 1199 struct evsel *counter; 1200 char buf[64], *prefix = NULL; 1201 1202 if (interval) 1203 print_interval(config, evlist, prefix = buf, ts); 1204 else 1205 print_header(config, _target, argc, argv); 1206 1207 if (metric_only) { 1208 static int num_print_iv; 1209 1210 if (num_print_iv == 0 && !interval) 1211 print_metric_headers(config, evlist, prefix, false); 1212 if (num_print_iv++ == 25) 1213 num_print_iv = 0; 1214 if (config->aggr_mode == AGGR_GLOBAL && prefix) 1215 fprintf(config->output, "%s", prefix); 1216 } 1217 1218 switch (config->aggr_mode) { 1219 case AGGR_CORE: 1220 case AGGR_DIE: 1221 case AGGR_SOCKET: 1222 case AGGR_NODE: 1223 print_aggr(config, evlist, prefix); 1224 break; 1225 case AGGR_THREAD: 1226 evlist__for_each_entry(evlist, counter) { 1227 print_aggr_thread(config, _target, counter, prefix); 1228 } 1229 break; 1230 case AGGR_GLOBAL: 1231 evlist__for_each_entry(evlist, counter) { 1232 print_counter_aggr(config, counter, prefix); 1233 } 1234 if (metric_only) 1235 fputc('\n', config->output); 1236 break; 1237 case AGGR_NONE: 1238 if (metric_only) 1239 print_no_aggr_metric(config, evlist, prefix); 1240 else { 1241 evlist__for_each_entry(evlist, counter) { 1242 if (counter->percore) 1243 print_percore(config, counter, prefix); 1244 else 1245 print_counter(config, counter, prefix); 1246 } 1247 } 1248 break; 1249 case AGGR_UNSET: 1250 default: 1251 break; 1252 } 1253 1254 if (!interval && !config->csv_output) 1255 print_footer(config); 1256 1257 fflush(config->output); 1258 } 1259