1 // SPDX-License-Identifier: GPL-2.0 2 #include <math.h> 3 #include <stdio.h> 4 #include "evsel.h" 5 #include "stat.h" 6 #include "color.h" 7 #include "debug.h" 8 #include "pmu.h" 9 #include "rblist.h" 10 #include "evlist.h" 11 #include "expr.h" 12 #include "metricgroup.h" 13 #include "cgroup.h" 14 #include "units.h" 15 #include <linux/zalloc.h> 16 #include "iostat.h" 17 18 /* 19 * AGGR_GLOBAL: Use CPU 0 20 * AGGR_SOCKET: Use first CPU of socket 21 * AGGR_DIE: Use first CPU of die 22 * AGGR_CORE: Use first CPU of core 23 * AGGR_NONE: Use matching CPU 24 * AGGR_THREAD: Not supported? 25 */ 26 27 struct runtime_stat rt_stat; 28 struct stats walltime_nsecs_stats; 29 30 struct saved_value { 31 struct rb_node rb_node; 32 struct evsel *evsel; 33 enum stat_type type; 34 int ctx; 35 int cpu; 36 struct cgroup *cgrp; 37 struct runtime_stat *stat; 38 struct stats stats; 39 u64 metric_total; 40 int metric_other; 41 }; 42 43 static int saved_value_cmp(struct rb_node *rb_node, const void *entry) 44 { 45 struct saved_value *a = container_of(rb_node, 46 struct saved_value, 47 rb_node); 48 const struct saved_value *b = entry; 49 50 if (a->cpu != b->cpu) 51 return a->cpu - b->cpu; 52 53 /* 54 * Previously the rbtree was used to link generic metrics. 55 * The keys were evsel/cpu. Now the rbtree is extended to support 56 * per-thread shadow stats. For shadow stats case, the keys 57 * are cpu/type/ctx/stat (evsel is NULL). For generic metrics 58 * case, the keys are still evsel/cpu (type/ctx/stat are 0 or NULL). 59 */ 60 if (a->type != b->type) 61 return a->type - b->type; 62 63 if (a->ctx != b->ctx) 64 return a->ctx - b->ctx; 65 66 if (a->cgrp != b->cgrp) 67 return (char *)a->cgrp < (char *)b->cgrp ? -1 : +1; 68 69 if (a->evsel == NULL && b->evsel == NULL) { 70 if (a->stat == b->stat) 71 return 0; 72 73 if ((char *)a->stat < (char *)b->stat) 74 return -1; 75 76 return 1; 77 } 78 79 if (a->evsel == b->evsel) 80 return 0; 81 if ((char *)a->evsel < (char *)b->evsel) 82 return -1; 83 return +1; 84 } 85 86 static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused, 87 const void *entry) 88 { 89 struct saved_value *nd = malloc(sizeof(struct saved_value)); 90 91 if (!nd) 92 return NULL; 93 memcpy(nd, entry, sizeof(struct saved_value)); 94 return &nd->rb_node; 95 } 96 97 static void saved_value_delete(struct rblist *rblist __maybe_unused, 98 struct rb_node *rb_node) 99 { 100 struct saved_value *v; 101 102 BUG_ON(!rb_node); 103 v = container_of(rb_node, struct saved_value, rb_node); 104 free(v); 105 } 106 107 static struct saved_value *saved_value_lookup(struct evsel *evsel, 108 int cpu, 109 bool create, 110 enum stat_type type, 111 int ctx, 112 struct runtime_stat *st, 113 struct cgroup *cgrp) 114 { 115 struct rblist *rblist; 116 struct rb_node *nd; 117 struct saved_value dm = { 118 .cpu = cpu, 119 .evsel = evsel, 120 .type = type, 121 .ctx = ctx, 122 .stat = st, 123 .cgrp = cgrp, 124 }; 125 126 rblist = &st->value_list; 127 128 /* don't use context info for clock events */ 129 if (type == STAT_NSECS) 130 dm.ctx = 0; 131 132 nd = rblist__find(rblist, &dm); 133 if (nd) 134 return container_of(nd, struct saved_value, rb_node); 135 if (create) { 136 rblist__add_node(rblist, &dm); 137 nd = rblist__find(rblist, &dm); 138 if (nd) 139 return container_of(nd, struct saved_value, rb_node); 140 } 141 return NULL; 142 } 143 144 void runtime_stat__init(struct runtime_stat *st) 145 { 146 struct rblist *rblist = &st->value_list; 147 148 rblist__init(rblist); 149 rblist->node_cmp = saved_value_cmp; 150 rblist->node_new = saved_value_new; 151 rblist->node_delete = saved_value_delete; 152 } 153 154 void runtime_stat__exit(struct runtime_stat *st) 155 { 156 rblist__exit(&st->value_list); 157 } 158 159 void perf_stat__init_shadow_stats(void) 160 { 161 runtime_stat__init(&rt_stat); 162 } 163 164 static int evsel_context(struct evsel *evsel) 165 { 166 int ctx = 0; 167 168 if (evsel->core.attr.exclude_kernel) 169 ctx |= CTX_BIT_KERNEL; 170 if (evsel->core.attr.exclude_user) 171 ctx |= CTX_BIT_USER; 172 if (evsel->core.attr.exclude_hv) 173 ctx |= CTX_BIT_HV; 174 if (evsel->core.attr.exclude_host) 175 ctx |= CTX_BIT_HOST; 176 if (evsel->core.attr.exclude_idle) 177 ctx |= CTX_BIT_IDLE; 178 179 return ctx; 180 } 181 182 static void reset_stat(struct runtime_stat *st) 183 { 184 struct rblist *rblist; 185 struct rb_node *pos, *next; 186 187 rblist = &st->value_list; 188 next = rb_first_cached(&rblist->entries); 189 while (next) { 190 pos = next; 191 next = rb_next(pos); 192 memset(&container_of(pos, struct saved_value, rb_node)->stats, 193 0, 194 sizeof(struct stats)); 195 } 196 } 197 198 void perf_stat__reset_shadow_stats(void) 199 { 200 reset_stat(&rt_stat); 201 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats)); 202 } 203 204 void perf_stat__reset_shadow_per_stat(struct runtime_stat *st) 205 { 206 reset_stat(st); 207 } 208 209 struct runtime_stat_data { 210 int ctx; 211 struct cgroup *cgrp; 212 }; 213 214 static void update_runtime_stat(struct runtime_stat *st, 215 enum stat_type type, 216 int cpu, u64 count, 217 struct runtime_stat_data *rsd) 218 { 219 struct saved_value *v = saved_value_lookup(NULL, cpu, true, type, 220 rsd->ctx, st, rsd->cgrp); 221 222 if (v) 223 update_stats(&v->stats, count); 224 } 225 226 /* 227 * Update various tracking values we maintain to print 228 * more semantic information such as miss/hit ratios, 229 * instruction rates, etc: 230 */ 231 void perf_stat__update_shadow_stats(struct evsel *counter, u64 count, 232 int cpu, struct runtime_stat *st) 233 { 234 u64 count_ns = count; 235 struct saved_value *v; 236 struct runtime_stat_data rsd = { 237 .ctx = evsel_context(counter), 238 .cgrp = counter->cgrp, 239 }; 240 241 count *= counter->scale; 242 243 if (evsel__is_clock(counter)) 244 update_runtime_stat(st, STAT_NSECS, cpu, count_ns, &rsd); 245 else if (evsel__match(counter, HARDWARE, HW_CPU_CYCLES)) 246 update_runtime_stat(st, STAT_CYCLES, cpu, count, &rsd); 247 else if (perf_stat_evsel__is(counter, CYCLES_IN_TX)) 248 update_runtime_stat(st, STAT_CYCLES_IN_TX, cpu, count, &rsd); 249 else if (perf_stat_evsel__is(counter, TRANSACTION_START)) 250 update_runtime_stat(st, STAT_TRANSACTION, cpu, count, &rsd); 251 else if (perf_stat_evsel__is(counter, ELISION_START)) 252 update_runtime_stat(st, STAT_ELISION, cpu, count, &rsd); 253 else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS)) 254 update_runtime_stat(st, STAT_TOPDOWN_TOTAL_SLOTS, 255 cpu, count, &rsd); 256 else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED)) 257 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_ISSUED, 258 cpu, count, &rsd); 259 else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED)) 260 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_RETIRED, 261 cpu, count, &rsd); 262 else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES)) 263 update_runtime_stat(st, STAT_TOPDOWN_FETCH_BUBBLES, 264 cpu, count, &rsd); 265 else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES)) 266 update_runtime_stat(st, STAT_TOPDOWN_RECOVERY_BUBBLES, 267 cpu, count, &rsd); 268 else if (perf_stat_evsel__is(counter, TOPDOWN_RETIRING)) 269 update_runtime_stat(st, STAT_TOPDOWN_RETIRING, 270 cpu, count, &rsd); 271 else if (perf_stat_evsel__is(counter, TOPDOWN_BAD_SPEC)) 272 update_runtime_stat(st, STAT_TOPDOWN_BAD_SPEC, 273 cpu, count, &rsd); 274 else if (perf_stat_evsel__is(counter, TOPDOWN_FE_BOUND)) 275 update_runtime_stat(st, STAT_TOPDOWN_FE_BOUND, 276 cpu, count, &rsd); 277 else if (perf_stat_evsel__is(counter, TOPDOWN_BE_BOUND)) 278 update_runtime_stat(st, STAT_TOPDOWN_BE_BOUND, 279 cpu, count, &rsd); 280 else if (perf_stat_evsel__is(counter, TOPDOWN_HEAVY_OPS)) 281 update_runtime_stat(st, STAT_TOPDOWN_HEAVY_OPS, 282 cpu, count, &rsd); 283 else if (perf_stat_evsel__is(counter, TOPDOWN_BR_MISPREDICT)) 284 update_runtime_stat(st, STAT_TOPDOWN_BR_MISPREDICT, 285 cpu, count, &rsd); 286 else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_LAT)) 287 update_runtime_stat(st, STAT_TOPDOWN_FETCH_LAT, 288 cpu, count, &rsd); 289 else if (perf_stat_evsel__is(counter, TOPDOWN_MEM_BOUND)) 290 update_runtime_stat(st, STAT_TOPDOWN_MEM_BOUND, 291 cpu, count, &rsd); 292 else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) 293 update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT, 294 cpu, count, &rsd); 295 else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND)) 296 update_runtime_stat(st, STAT_STALLED_CYCLES_BACK, 297 cpu, count, &rsd); 298 else if (evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS)) 299 update_runtime_stat(st, STAT_BRANCHES, cpu, count, &rsd); 300 else if (evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES)) 301 update_runtime_stat(st, STAT_CACHEREFS, cpu, count, &rsd); 302 else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1D)) 303 update_runtime_stat(st, STAT_L1_DCACHE, cpu, count, &rsd); 304 else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1I)) 305 update_runtime_stat(st, STAT_L1_ICACHE, cpu, count, &rsd); 306 else if (evsel__match(counter, HW_CACHE, HW_CACHE_LL)) 307 update_runtime_stat(st, STAT_LL_CACHE, cpu, count, &rsd); 308 else if (evsel__match(counter, HW_CACHE, HW_CACHE_DTLB)) 309 update_runtime_stat(st, STAT_DTLB_CACHE, cpu, count, &rsd); 310 else if (evsel__match(counter, HW_CACHE, HW_CACHE_ITLB)) 311 update_runtime_stat(st, STAT_ITLB_CACHE, cpu, count, &rsd); 312 else if (perf_stat_evsel__is(counter, SMI_NUM)) 313 update_runtime_stat(st, STAT_SMI_NUM, cpu, count, &rsd); 314 else if (perf_stat_evsel__is(counter, APERF)) 315 update_runtime_stat(st, STAT_APERF, cpu, count, &rsd); 316 317 if (counter->collect_stat) { 318 v = saved_value_lookup(counter, cpu, true, STAT_NONE, 0, st, 319 rsd.cgrp); 320 update_stats(&v->stats, count); 321 if (counter->metric_leader) 322 v->metric_total += count; 323 } else if (counter->metric_leader) { 324 v = saved_value_lookup(counter->metric_leader, 325 cpu, true, STAT_NONE, 0, st, rsd.cgrp); 326 v->metric_total += count; 327 v->metric_other++; 328 } 329 } 330 331 /* used for get_ratio_color() */ 332 enum grc_type { 333 GRC_STALLED_CYCLES_FE, 334 GRC_STALLED_CYCLES_BE, 335 GRC_CACHE_MISSES, 336 GRC_MAX_NR 337 }; 338 339 static const char *get_ratio_color(enum grc_type type, double ratio) 340 { 341 static const double grc_table[GRC_MAX_NR][3] = { 342 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 }, 343 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 }, 344 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 }, 345 }; 346 const char *color = PERF_COLOR_NORMAL; 347 348 if (ratio > grc_table[type][0]) 349 color = PERF_COLOR_RED; 350 else if (ratio > grc_table[type][1]) 351 color = PERF_COLOR_MAGENTA; 352 else if (ratio > grc_table[type][2]) 353 color = PERF_COLOR_YELLOW; 354 355 return color; 356 } 357 358 static struct evsel *perf_stat__find_event(struct evlist *evsel_list, 359 const char *name) 360 { 361 struct evsel *c2; 362 363 evlist__for_each_entry (evsel_list, c2) { 364 if (!strcasecmp(c2->name, name) && !c2->collect_stat) 365 return c2; 366 } 367 return NULL; 368 } 369 370 /* Mark MetricExpr target events and link events using them to them. */ 371 void perf_stat__collect_metric_expr(struct evlist *evsel_list) 372 { 373 struct evsel *counter, *leader, **metric_events, *oc; 374 bool found; 375 struct expr_parse_ctx *ctx; 376 struct hashmap_entry *cur; 377 size_t bkt; 378 int i; 379 380 ctx = expr__ctx_new(); 381 if (!ctx) { 382 pr_debug("expr__ctx_new failed"); 383 return; 384 } 385 evlist__for_each_entry(evsel_list, counter) { 386 bool invalid = false; 387 388 leader = evsel__leader(counter); 389 if (!counter->metric_expr) 390 continue; 391 392 expr__ctx_clear(ctx); 393 metric_events = counter->metric_events; 394 if (!metric_events) { 395 if (expr__find_ids(counter->metric_expr, 396 counter->name, 397 ctx, 1) < 0) 398 continue; 399 400 metric_events = calloc(sizeof(struct evsel *), 401 hashmap__size(ctx->ids) + 1); 402 if (!metric_events) { 403 expr__ctx_free(ctx); 404 return; 405 } 406 counter->metric_events = metric_events; 407 } 408 409 i = 0; 410 hashmap__for_each_entry(ctx->ids, cur, bkt) { 411 const char *metric_name = (const char *)cur->key; 412 413 found = false; 414 if (leader) { 415 /* Search in group */ 416 for_each_group_member (oc, leader) { 417 if (!strcasecmp(oc->name, 418 metric_name) && 419 !oc->collect_stat) { 420 found = true; 421 break; 422 } 423 } 424 } 425 if (!found) { 426 /* Search ignoring groups */ 427 oc = perf_stat__find_event(evsel_list, 428 metric_name); 429 } 430 if (!oc) { 431 /* Deduping one is good enough to handle duplicated PMUs. */ 432 static char *printed; 433 434 /* 435 * Adding events automatically would be difficult, because 436 * it would risk creating groups that are not schedulable. 437 * perf stat doesn't understand all the scheduling constraints 438 * of events. So we ask the user instead to add the missing 439 * events. 440 */ 441 if (!printed || 442 strcasecmp(printed, metric_name)) { 443 fprintf(stderr, 444 "Add %s event to groups to get metric expression for %s\n", 445 metric_name, 446 counter->name); 447 printed = strdup(metric_name); 448 } 449 invalid = true; 450 continue; 451 } 452 metric_events[i++] = oc; 453 oc->collect_stat = true; 454 } 455 metric_events[i] = NULL; 456 if (invalid) { 457 free(metric_events); 458 counter->metric_events = NULL; 459 counter->metric_expr = NULL; 460 } 461 } 462 expr__ctx_free(ctx); 463 } 464 465 static double runtime_stat_avg(struct runtime_stat *st, 466 enum stat_type type, int cpu, 467 struct runtime_stat_data *rsd) 468 { 469 struct saved_value *v; 470 471 v = saved_value_lookup(NULL, cpu, false, type, rsd->ctx, st, rsd->cgrp); 472 if (!v) 473 return 0.0; 474 475 return avg_stats(&v->stats); 476 } 477 478 static double runtime_stat_n(struct runtime_stat *st, 479 enum stat_type type, int cpu, 480 struct runtime_stat_data *rsd) 481 { 482 struct saved_value *v; 483 484 v = saved_value_lookup(NULL, cpu, false, type, rsd->ctx, st, rsd->cgrp); 485 if (!v) 486 return 0.0; 487 488 return v->stats.n; 489 } 490 491 static void print_stalled_cycles_frontend(struct perf_stat_config *config, 492 int cpu, double avg, 493 struct perf_stat_output_ctx *out, 494 struct runtime_stat *st, 495 struct runtime_stat_data *rsd) 496 { 497 double total, ratio = 0.0; 498 const char *color; 499 500 total = runtime_stat_avg(st, STAT_CYCLES, cpu, rsd); 501 502 if (total) 503 ratio = avg / total * 100.0; 504 505 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio); 506 507 if (ratio) 508 out->print_metric(config, out->ctx, color, "%7.2f%%", "frontend cycles idle", 509 ratio); 510 else 511 out->print_metric(config, out->ctx, NULL, NULL, "frontend cycles idle", 0); 512 } 513 514 static void print_stalled_cycles_backend(struct perf_stat_config *config, 515 int cpu, double avg, 516 struct perf_stat_output_ctx *out, 517 struct runtime_stat *st, 518 struct runtime_stat_data *rsd) 519 { 520 double total, ratio = 0.0; 521 const char *color; 522 523 total = runtime_stat_avg(st, STAT_CYCLES, cpu, rsd); 524 525 if (total) 526 ratio = avg / total * 100.0; 527 528 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio); 529 530 out->print_metric(config, out->ctx, color, "%7.2f%%", "backend cycles idle", ratio); 531 } 532 533 static void print_branch_misses(struct perf_stat_config *config, 534 int cpu, double avg, 535 struct perf_stat_output_ctx *out, 536 struct runtime_stat *st, 537 struct runtime_stat_data *rsd) 538 { 539 double total, ratio = 0.0; 540 const char *color; 541 542 total = runtime_stat_avg(st, STAT_BRANCHES, cpu, rsd); 543 544 if (total) 545 ratio = avg / total * 100.0; 546 547 color = get_ratio_color(GRC_CACHE_MISSES, ratio); 548 549 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all branches", ratio); 550 } 551 552 static void print_l1_dcache_misses(struct perf_stat_config *config, 553 int cpu, double avg, 554 struct perf_stat_output_ctx *out, 555 struct runtime_stat *st, 556 struct runtime_stat_data *rsd) 557 { 558 double total, ratio = 0.0; 559 const char *color; 560 561 total = runtime_stat_avg(st, STAT_L1_DCACHE, cpu, rsd); 562 563 if (total) 564 ratio = avg / total * 100.0; 565 566 color = get_ratio_color(GRC_CACHE_MISSES, ratio); 567 568 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-dcache accesses", ratio); 569 } 570 571 static void print_l1_icache_misses(struct perf_stat_config *config, 572 int cpu, double avg, 573 struct perf_stat_output_ctx *out, 574 struct runtime_stat *st, 575 struct runtime_stat_data *rsd) 576 { 577 double total, ratio = 0.0; 578 const char *color; 579 580 total = runtime_stat_avg(st, STAT_L1_ICACHE, cpu, rsd); 581 582 if (total) 583 ratio = avg / total * 100.0; 584 585 color = get_ratio_color(GRC_CACHE_MISSES, ratio); 586 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-icache accesses", ratio); 587 } 588 589 static void print_dtlb_cache_misses(struct perf_stat_config *config, 590 int cpu, double avg, 591 struct perf_stat_output_ctx *out, 592 struct runtime_stat *st, 593 struct runtime_stat_data *rsd) 594 { 595 double total, ratio = 0.0; 596 const char *color; 597 598 total = runtime_stat_avg(st, STAT_DTLB_CACHE, cpu, rsd); 599 600 if (total) 601 ratio = avg / total * 100.0; 602 603 color = get_ratio_color(GRC_CACHE_MISSES, ratio); 604 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all dTLB cache accesses", ratio); 605 } 606 607 static void print_itlb_cache_misses(struct perf_stat_config *config, 608 int cpu, double avg, 609 struct perf_stat_output_ctx *out, 610 struct runtime_stat *st, 611 struct runtime_stat_data *rsd) 612 { 613 double total, ratio = 0.0; 614 const char *color; 615 616 total = runtime_stat_avg(st, STAT_ITLB_CACHE, cpu, rsd); 617 618 if (total) 619 ratio = avg / total * 100.0; 620 621 color = get_ratio_color(GRC_CACHE_MISSES, ratio); 622 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all iTLB cache accesses", ratio); 623 } 624 625 static void print_ll_cache_misses(struct perf_stat_config *config, 626 int cpu, double avg, 627 struct perf_stat_output_ctx *out, 628 struct runtime_stat *st, 629 struct runtime_stat_data *rsd) 630 { 631 double total, ratio = 0.0; 632 const char *color; 633 634 total = runtime_stat_avg(st, STAT_LL_CACHE, cpu, rsd); 635 636 if (total) 637 ratio = avg / total * 100.0; 638 639 color = get_ratio_color(GRC_CACHE_MISSES, ratio); 640 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all LL-cache accesses", ratio); 641 } 642 643 /* 644 * High level "TopDown" CPU core pipe line bottleneck break down. 645 * 646 * Basic concept following 647 * Yasin, A Top Down Method for Performance analysis and Counter architecture 648 * ISPASS14 649 * 650 * The CPU pipeline is divided into 4 areas that can be bottlenecks: 651 * 652 * Frontend -> Backend -> Retiring 653 * BadSpeculation in addition means out of order execution that is thrown away 654 * (for example branch mispredictions) 655 * Frontend is instruction decoding. 656 * Backend is execution, like computation and accessing data in memory 657 * Retiring is good execution that is not directly bottlenecked 658 * 659 * The formulas are computed in slots. 660 * A slot is an entry in the pipeline each for the pipeline width 661 * (for example a 4-wide pipeline has 4 slots for each cycle) 662 * 663 * Formulas: 664 * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) / 665 * TotalSlots 666 * Retiring = SlotsRetired / TotalSlots 667 * FrontendBound = FetchBubbles / TotalSlots 668 * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound 669 * 670 * The kernel provides the mapping to the low level CPU events and any scaling 671 * needed for the CPU pipeline width, for example: 672 * 673 * TotalSlots = Cycles * 4 674 * 675 * The scaling factor is communicated in the sysfs unit. 676 * 677 * In some cases the CPU may not be able to measure all the formulas due to 678 * missing events. In this case multiple formulas are combined, as possible. 679 * 680 * Full TopDown supports more levels to sub-divide each area: for example 681 * BackendBound into computing bound and memory bound. For now we only 682 * support Level 1 TopDown. 683 */ 684 685 static double sanitize_val(double x) 686 { 687 if (x < 0 && x >= -0.02) 688 return 0.0; 689 return x; 690 } 691 692 static double td_total_slots(int cpu, struct runtime_stat *st, 693 struct runtime_stat_data *rsd) 694 { 695 return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, cpu, rsd); 696 } 697 698 static double td_bad_spec(int cpu, struct runtime_stat *st, 699 struct runtime_stat_data *rsd) 700 { 701 double bad_spec = 0; 702 double total_slots; 703 double total; 704 705 total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, cpu, rsd) - 706 runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, cpu, rsd) + 707 runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, cpu, rsd); 708 709 total_slots = td_total_slots(cpu, st, rsd); 710 if (total_slots) 711 bad_spec = total / total_slots; 712 return sanitize_val(bad_spec); 713 } 714 715 static double td_retiring(int cpu, struct runtime_stat *st, 716 struct runtime_stat_data *rsd) 717 { 718 double retiring = 0; 719 double total_slots = td_total_slots(cpu, st, rsd); 720 double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, 721 cpu, rsd); 722 723 if (total_slots) 724 retiring = ret_slots / total_slots; 725 return retiring; 726 } 727 728 static double td_fe_bound(int cpu, struct runtime_stat *st, 729 struct runtime_stat_data *rsd) 730 { 731 double fe_bound = 0; 732 double total_slots = td_total_slots(cpu, st, rsd); 733 double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES, 734 cpu, rsd); 735 736 if (total_slots) 737 fe_bound = fetch_bub / total_slots; 738 return fe_bound; 739 } 740 741 static double td_be_bound(int cpu, struct runtime_stat *st, 742 struct runtime_stat_data *rsd) 743 { 744 double sum = (td_fe_bound(cpu, st, rsd) + 745 td_bad_spec(cpu, st, rsd) + 746 td_retiring(cpu, st, rsd)); 747 if (sum == 0) 748 return 0; 749 return sanitize_val(1.0 - sum); 750 } 751 752 /* 753 * Kernel reports metrics multiplied with slots. To get back 754 * the ratios we need to recreate the sum. 755 */ 756 757 static double td_metric_ratio(int cpu, enum stat_type type, 758 struct runtime_stat *stat, 759 struct runtime_stat_data *rsd) 760 { 761 double sum = runtime_stat_avg(stat, STAT_TOPDOWN_RETIRING, cpu, rsd) + 762 runtime_stat_avg(stat, STAT_TOPDOWN_FE_BOUND, cpu, rsd) + 763 runtime_stat_avg(stat, STAT_TOPDOWN_BE_BOUND, cpu, rsd) + 764 runtime_stat_avg(stat, STAT_TOPDOWN_BAD_SPEC, cpu, rsd); 765 double d = runtime_stat_avg(stat, type, cpu, rsd); 766 767 if (sum) 768 return d / sum; 769 return 0; 770 } 771 772 /* 773 * ... but only if most of the values are actually available. 774 * We allow two missing. 775 */ 776 777 static bool full_td(int cpu, struct runtime_stat *stat, 778 struct runtime_stat_data *rsd) 779 { 780 int c = 0; 781 782 if (runtime_stat_avg(stat, STAT_TOPDOWN_RETIRING, cpu, rsd) > 0) 783 c++; 784 if (runtime_stat_avg(stat, STAT_TOPDOWN_BE_BOUND, cpu, rsd) > 0) 785 c++; 786 if (runtime_stat_avg(stat, STAT_TOPDOWN_FE_BOUND, cpu, rsd) > 0) 787 c++; 788 if (runtime_stat_avg(stat, STAT_TOPDOWN_BAD_SPEC, cpu, rsd) > 0) 789 c++; 790 return c >= 2; 791 } 792 793 static void print_smi_cost(struct perf_stat_config *config, int cpu, 794 struct perf_stat_output_ctx *out, 795 struct runtime_stat *st, 796 struct runtime_stat_data *rsd) 797 { 798 double smi_num, aperf, cycles, cost = 0.0; 799 const char *color = NULL; 800 801 smi_num = runtime_stat_avg(st, STAT_SMI_NUM, cpu, rsd); 802 aperf = runtime_stat_avg(st, STAT_APERF, cpu, rsd); 803 cycles = runtime_stat_avg(st, STAT_CYCLES, cpu, rsd); 804 805 if ((cycles == 0) || (aperf == 0)) 806 return; 807 808 if (smi_num) 809 cost = (aperf - cycles) / aperf * 100.00; 810 811 if (cost > 10) 812 color = PERF_COLOR_RED; 813 out->print_metric(config, out->ctx, color, "%8.1f%%", "SMI cycles%", cost); 814 out->print_metric(config, out->ctx, NULL, "%4.0f", "SMI#", smi_num); 815 } 816 817 static int prepare_metric(struct evsel **metric_events, 818 struct metric_ref *metric_refs, 819 struct expr_parse_ctx *pctx, 820 int cpu, 821 struct runtime_stat *st) 822 { 823 double scale; 824 char *n, *pn; 825 int i, j, ret; 826 827 for (i = 0; metric_events[i]; i++) { 828 struct saved_value *v; 829 struct stats *stats; 830 u64 metric_total = 0; 831 832 if (!strcmp(metric_events[i]->name, "duration_time")) { 833 stats = &walltime_nsecs_stats; 834 scale = 1e-9; 835 } else { 836 v = saved_value_lookup(metric_events[i], cpu, false, 837 STAT_NONE, 0, st, 838 metric_events[i]->cgrp); 839 if (!v) 840 break; 841 stats = &v->stats; 842 scale = 1.0; 843 844 if (v->metric_other) 845 metric_total = v->metric_total; 846 } 847 848 n = strdup(metric_events[i]->name); 849 if (!n) 850 return -ENOMEM; 851 /* 852 * This display code with --no-merge adds [cpu] postfixes. 853 * These are not supported by the parser. Remove everything 854 * after the space. 855 */ 856 pn = strchr(n, ' '); 857 if (pn) 858 *pn = 0; 859 860 if (metric_total) 861 expr__add_id_val(pctx, n, metric_total); 862 else 863 expr__add_id_val(pctx, n, avg_stats(stats)*scale); 864 } 865 866 for (j = 0; metric_refs && metric_refs[j].metric_name; j++) { 867 ret = expr__add_ref(pctx, &metric_refs[j]); 868 if (ret) 869 return ret; 870 } 871 872 return i; 873 } 874 875 static void generic_metric(struct perf_stat_config *config, 876 const char *metric_expr, 877 struct evsel **metric_events, 878 struct metric_ref *metric_refs, 879 char *name, 880 const char *metric_name, 881 const char *metric_unit, 882 int runtime, 883 int cpu, 884 struct perf_stat_output_ctx *out, 885 struct runtime_stat *st) 886 { 887 print_metric_t print_metric = out->print_metric; 888 struct expr_parse_ctx *pctx; 889 double ratio, scale; 890 int i; 891 void *ctxp = out->ctx; 892 893 pctx = expr__ctx_new(); 894 if (!pctx) 895 return; 896 897 i = prepare_metric(metric_events, metric_refs, pctx, cpu, st); 898 if (i < 0) { 899 expr__ctx_free(pctx); 900 return; 901 } 902 if (!metric_events[i]) { 903 if (expr__parse(&ratio, pctx, metric_expr, runtime) == 0) { 904 char *unit; 905 char metric_bf[64]; 906 907 if (metric_unit && metric_name) { 908 if (perf_pmu__convert_scale(metric_unit, 909 &unit, &scale) >= 0) { 910 ratio *= scale; 911 } 912 if (strstr(metric_expr, "?")) 913 scnprintf(metric_bf, sizeof(metric_bf), 914 "%s %s_%d", unit, metric_name, runtime); 915 else 916 scnprintf(metric_bf, sizeof(metric_bf), 917 "%s %s", unit, metric_name); 918 919 print_metric(config, ctxp, NULL, "%8.1f", 920 metric_bf, ratio); 921 } else { 922 print_metric(config, ctxp, NULL, "%8.2f", 923 metric_name ? 924 metric_name : 925 out->force_header ? name : "", 926 ratio); 927 } 928 } else { 929 print_metric(config, ctxp, NULL, NULL, 930 out->force_header ? 931 (metric_name ? metric_name : name) : "", 0); 932 } 933 } else { 934 print_metric(config, ctxp, NULL, NULL, 935 out->force_header ? 936 (metric_name ? metric_name : name) : "", 0); 937 } 938 939 expr__ctx_free(pctx); 940 } 941 942 double test_generic_metric(struct metric_expr *mexp, int cpu, struct runtime_stat *st) 943 { 944 struct expr_parse_ctx *pctx; 945 double ratio = 0.0; 946 947 pctx = expr__ctx_new(); 948 if (!pctx) 949 return NAN; 950 951 if (prepare_metric(mexp->metric_events, mexp->metric_refs, pctx, cpu, st) < 0) 952 goto out; 953 954 if (expr__parse(&ratio, pctx, mexp->metric_expr, 1)) 955 ratio = 0.0; 956 957 out: 958 expr__ctx_free(pctx); 959 return ratio; 960 } 961 962 void perf_stat__print_shadow_stats(struct perf_stat_config *config, 963 struct evsel *evsel, 964 double avg, int cpu, 965 struct perf_stat_output_ctx *out, 966 struct rblist *metric_events, 967 struct runtime_stat *st) 968 { 969 void *ctxp = out->ctx; 970 print_metric_t print_metric = out->print_metric; 971 double total, ratio = 0.0, total2; 972 const char *color = NULL; 973 struct runtime_stat_data rsd = { 974 .ctx = evsel_context(evsel), 975 .cgrp = evsel->cgrp, 976 }; 977 struct metric_event *me; 978 int num = 1; 979 980 if (config->iostat_run) { 981 iostat_print_metric(config, evsel, out); 982 } else if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) { 983 total = runtime_stat_avg(st, STAT_CYCLES, cpu, &rsd); 984 985 if (total) { 986 ratio = avg / total; 987 print_metric(config, ctxp, NULL, "%7.2f ", 988 "insn per cycle", ratio); 989 } else { 990 print_metric(config, ctxp, NULL, NULL, "insn per cycle", 0); 991 } 992 993 total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT, cpu, &rsd); 994 995 total = max(total, runtime_stat_avg(st, 996 STAT_STALLED_CYCLES_BACK, 997 cpu, &rsd)); 998 999 if (total && avg) { 1000 out->new_line(config, ctxp); 1001 ratio = total / avg; 1002 print_metric(config, ctxp, NULL, "%7.2f ", 1003 "stalled cycles per insn", 1004 ratio); 1005 } 1006 } else if (evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) { 1007 if (runtime_stat_n(st, STAT_BRANCHES, cpu, &rsd) != 0) 1008 print_branch_misses(config, cpu, avg, out, st, &rsd); 1009 else 1010 print_metric(config, ctxp, NULL, NULL, "of all branches", 0); 1011 } else if ( 1012 evsel->core.attr.type == PERF_TYPE_HW_CACHE && 1013 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_L1D | 1014 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 1015 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) { 1016 1017 if (runtime_stat_n(st, STAT_L1_DCACHE, cpu, &rsd) != 0) 1018 print_l1_dcache_misses(config, cpu, avg, out, st, &rsd); 1019 else 1020 print_metric(config, ctxp, NULL, NULL, "of all L1-dcache accesses", 0); 1021 } else if ( 1022 evsel->core.attr.type == PERF_TYPE_HW_CACHE && 1023 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_L1I | 1024 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 1025 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) { 1026 1027 if (runtime_stat_n(st, STAT_L1_ICACHE, cpu, &rsd) != 0) 1028 print_l1_icache_misses(config, cpu, avg, out, st, &rsd); 1029 else 1030 print_metric(config, ctxp, NULL, NULL, "of all L1-icache accesses", 0); 1031 } else if ( 1032 evsel->core.attr.type == PERF_TYPE_HW_CACHE && 1033 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_DTLB | 1034 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 1035 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) { 1036 1037 if (runtime_stat_n(st, STAT_DTLB_CACHE, cpu, &rsd) != 0) 1038 print_dtlb_cache_misses(config, cpu, avg, out, st, &rsd); 1039 else 1040 print_metric(config, ctxp, NULL, NULL, "of all dTLB cache accesses", 0); 1041 } else if ( 1042 evsel->core.attr.type == PERF_TYPE_HW_CACHE && 1043 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_ITLB | 1044 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 1045 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) { 1046 1047 if (runtime_stat_n(st, STAT_ITLB_CACHE, cpu, &rsd) != 0) 1048 print_itlb_cache_misses(config, cpu, avg, out, st, &rsd); 1049 else 1050 print_metric(config, ctxp, NULL, NULL, "of all iTLB cache accesses", 0); 1051 } else if ( 1052 evsel->core.attr.type == PERF_TYPE_HW_CACHE && 1053 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_LL | 1054 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | 1055 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) { 1056 1057 if (runtime_stat_n(st, STAT_LL_CACHE, cpu, &rsd) != 0) 1058 print_ll_cache_misses(config, cpu, avg, out, st, &rsd); 1059 else 1060 print_metric(config, ctxp, NULL, NULL, "of all LL-cache accesses", 0); 1061 } else if (evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) { 1062 total = runtime_stat_avg(st, STAT_CACHEREFS, cpu, &rsd); 1063 1064 if (total) 1065 ratio = avg * 100 / total; 1066 1067 if (runtime_stat_n(st, STAT_CACHEREFS, cpu, &rsd) != 0) 1068 print_metric(config, ctxp, NULL, "%8.3f %%", 1069 "of all cache refs", ratio); 1070 else 1071 print_metric(config, ctxp, NULL, NULL, "of all cache refs", 0); 1072 } else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) { 1073 print_stalled_cycles_frontend(config, cpu, avg, out, st, &rsd); 1074 } else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) { 1075 print_stalled_cycles_backend(config, cpu, avg, out, st, &rsd); 1076 } else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) { 1077 total = runtime_stat_avg(st, STAT_NSECS, cpu, &rsd); 1078 1079 if (total) { 1080 ratio = avg / total; 1081 print_metric(config, ctxp, NULL, "%8.3f", "GHz", ratio); 1082 } else { 1083 print_metric(config, ctxp, NULL, NULL, "Ghz", 0); 1084 } 1085 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) { 1086 total = runtime_stat_avg(st, STAT_CYCLES, cpu, &rsd); 1087 1088 if (total) 1089 print_metric(config, ctxp, NULL, 1090 "%7.2f%%", "transactional cycles", 1091 100.0 * (avg / total)); 1092 else 1093 print_metric(config, ctxp, NULL, NULL, "transactional cycles", 1094 0); 1095 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) { 1096 total = runtime_stat_avg(st, STAT_CYCLES, cpu, &rsd); 1097 total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, cpu, &rsd); 1098 1099 if (total2 < avg) 1100 total2 = avg; 1101 if (total) 1102 print_metric(config, ctxp, NULL, "%7.2f%%", "aborted cycles", 1103 100.0 * ((total2-avg) / total)); 1104 else 1105 print_metric(config, ctxp, NULL, NULL, "aborted cycles", 0); 1106 } else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) { 1107 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX, cpu, &rsd); 1108 1109 if (avg) 1110 ratio = total / avg; 1111 1112 if (runtime_stat_n(st, STAT_CYCLES_IN_TX, cpu, &rsd) != 0) 1113 print_metric(config, ctxp, NULL, "%8.0f", 1114 "cycles / transaction", ratio); 1115 else 1116 print_metric(config, ctxp, NULL, NULL, "cycles / transaction", 1117 0); 1118 } else if (perf_stat_evsel__is(evsel, ELISION_START)) { 1119 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX, cpu, &rsd); 1120 1121 if (avg) 1122 ratio = total / avg; 1123 1124 print_metric(config, ctxp, NULL, "%8.0f", "cycles / elision", ratio); 1125 } else if (evsel__is_clock(evsel)) { 1126 if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0) 1127 print_metric(config, ctxp, NULL, "%8.3f", "CPUs utilized", 1128 avg / (ratio * evsel->scale)); 1129 else 1130 print_metric(config, ctxp, NULL, NULL, "CPUs utilized", 0); 1131 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) { 1132 double fe_bound = td_fe_bound(cpu, st, &rsd); 1133 1134 if (fe_bound > 0.2) 1135 color = PERF_COLOR_RED; 1136 print_metric(config, ctxp, color, "%8.1f%%", "frontend bound", 1137 fe_bound * 100.); 1138 } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) { 1139 double retiring = td_retiring(cpu, st, &rsd); 1140 1141 if (retiring > 0.7) 1142 color = PERF_COLOR_GREEN; 1143 print_metric(config, ctxp, color, "%8.1f%%", "retiring", 1144 retiring * 100.); 1145 } else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) { 1146 double bad_spec = td_bad_spec(cpu, st, &rsd); 1147 1148 if (bad_spec > 0.1) 1149 color = PERF_COLOR_RED; 1150 print_metric(config, ctxp, color, "%8.1f%%", "bad speculation", 1151 bad_spec * 100.); 1152 } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) { 1153 double be_bound = td_be_bound(cpu, st, &rsd); 1154 const char *name = "backend bound"; 1155 static int have_recovery_bubbles = -1; 1156 1157 /* In case the CPU does not support topdown-recovery-bubbles */ 1158 if (have_recovery_bubbles < 0) 1159 have_recovery_bubbles = pmu_have_event("cpu", 1160 "topdown-recovery-bubbles"); 1161 if (!have_recovery_bubbles) 1162 name = "backend bound/bad spec"; 1163 1164 if (be_bound > 0.2) 1165 color = PERF_COLOR_RED; 1166 if (td_total_slots(cpu, st, &rsd) > 0) 1167 print_metric(config, ctxp, color, "%8.1f%%", name, 1168 be_bound * 100.); 1169 else 1170 print_metric(config, ctxp, NULL, NULL, name, 0); 1171 } else if (perf_stat_evsel__is(evsel, TOPDOWN_RETIRING) && 1172 full_td(cpu, st, &rsd)) { 1173 double retiring = td_metric_ratio(cpu, 1174 STAT_TOPDOWN_RETIRING, st, 1175 &rsd); 1176 if (retiring > 0.7) 1177 color = PERF_COLOR_GREEN; 1178 print_metric(config, ctxp, color, "%8.1f%%", "retiring", 1179 retiring * 100.); 1180 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FE_BOUND) && 1181 full_td(cpu, st, &rsd)) { 1182 double fe_bound = td_metric_ratio(cpu, 1183 STAT_TOPDOWN_FE_BOUND, st, 1184 &rsd); 1185 if (fe_bound > 0.2) 1186 color = PERF_COLOR_RED; 1187 print_metric(config, ctxp, color, "%8.1f%%", "frontend bound", 1188 fe_bound * 100.); 1189 } else if (perf_stat_evsel__is(evsel, TOPDOWN_BE_BOUND) && 1190 full_td(cpu, st, &rsd)) { 1191 double be_bound = td_metric_ratio(cpu, 1192 STAT_TOPDOWN_BE_BOUND, st, 1193 &rsd); 1194 if (be_bound > 0.2) 1195 color = PERF_COLOR_RED; 1196 print_metric(config, ctxp, color, "%8.1f%%", "backend bound", 1197 be_bound * 100.); 1198 } else if (perf_stat_evsel__is(evsel, TOPDOWN_BAD_SPEC) && 1199 full_td(cpu, st, &rsd)) { 1200 double bad_spec = td_metric_ratio(cpu, 1201 STAT_TOPDOWN_BAD_SPEC, st, 1202 &rsd); 1203 if (bad_spec > 0.1) 1204 color = PERF_COLOR_RED; 1205 print_metric(config, ctxp, color, "%8.1f%%", "bad speculation", 1206 bad_spec * 100.); 1207 } else if (perf_stat_evsel__is(evsel, TOPDOWN_HEAVY_OPS) && 1208 full_td(cpu, st, &rsd) && (config->topdown_level > 1)) { 1209 double retiring = td_metric_ratio(cpu, 1210 STAT_TOPDOWN_RETIRING, st, 1211 &rsd); 1212 double heavy_ops = td_metric_ratio(cpu, 1213 STAT_TOPDOWN_HEAVY_OPS, st, 1214 &rsd); 1215 double light_ops = retiring - heavy_ops; 1216 1217 if (retiring > 0.7 && heavy_ops > 0.1) 1218 color = PERF_COLOR_GREEN; 1219 print_metric(config, ctxp, color, "%8.1f%%", "heavy operations", 1220 heavy_ops * 100.); 1221 if (retiring > 0.7 && light_ops > 0.6) 1222 color = PERF_COLOR_GREEN; 1223 else 1224 color = NULL; 1225 print_metric(config, ctxp, color, "%8.1f%%", "light operations", 1226 light_ops * 100.); 1227 } else if (perf_stat_evsel__is(evsel, TOPDOWN_BR_MISPREDICT) && 1228 full_td(cpu, st, &rsd) && (config->topdown_level > 1)) { 1229 double bad_spec = td_metric_ratio(cpu, 1230 STAT_TOPDOWN_BAD_SPEC, st, 1231 &rsd); 1232 double br_mis = td_metric_ratio(cpu, 1233 STAT_TOPDOWN_BR_MISPREDICT, st, 1234 &rsd); 1235 double m_clears = bad_spec - br_mis; 1236 1237 if (bad_spec > 0.1 && br_mis > 0.05) 1238 color = PERF_COLOR_RED; 1239 print_metric(config, ctxp, color, "%8.1f%%", "branch mispredict", 1240 br_mis * 100.); 1241 if (bad_spec > 0.1 && m_clears > 0.05) 1242 color = PERF_COLOR_RED; 1243 else 1244 color = NULL; 1245 print_metric(config, ctxp, color, "%8.1f%%", "machine clears", 1246 m_clears * 100.); 1247 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_LAT) && 1248 full_td(cpu, st, &rsd) && (config->topdown_level > 1)) { 1249 double fe_bound = td_metric_ratio(cpu, 1250 STAT_TOPDOWN_FE_BOUND, st, 1251 &rsd); 1252 double fetch_lat = td_metric_ratio(cpu, 1253 STAT_TOPDOWN_FETCH_LAT, st, 1254 &rsd); 1255 double fetch_bw = fe_bound - fetch_lat; 1256 1257 if (fe_bound > 0.2 && fetch_lat > 0.15) 1258 color = PERF_COLOR_RED; 1259 print_metric(config, ctxp, color, "%8.1f%%", "fetch latency", 1260 fetch_lat * 100.); 1261 if (fe_bound > 0.2 && fetch_bw > 0.1) 1262 color = PERF_COLOR_RED; 1263 else 1264 color = NULL; 1265 print_metric(config, ctxp, color, "%8.1f%%", "fetch bandwidth", 1266 fetch_bw * 100.); 1267 } else if (perf_stat_evsel__is(evsel, TOPDOWN_MEM_BOUND) && 1268 full_td(cpu, st, &rsd) && (config->topdown_level > 1)) { 1269 double be_bound = td_metric_ratio(cpu, 1270 STAT_TOPDOWN_BE_BOUND, st, 1271 &rsd); 1272 double mem_bound = td_metric_ratio(cpu, 1273 STAT_TOPDOWN_MEM_BOUND, st, 1274 &rsd); 1275 double core_bound = be_bound - mem_bound; 1276 1277 if (be_bound > 0.2 && mem_bound > 0.2) 1278 color = PERF_COLOR_RED; 1279 print_metric(config, ctxp, color, "%8.1f%%", "memory bound", 1280 mem_bound * 100.); 1281 if (be_bound > 0.2 && core_bound > 0.1) 1282 color = PERF_COLOR_RED; 1283 else 1284 color = NULL; 1285 print_metric(config, ctxp, color, "%8.1f%%", "Core bound", 1286 core_bound * 100.); 1287 } else if (evsel->metric_expr) { 1288 generic_metric(config, evsel->metric_expr, evsel->metric_events, NULL, 1289 evsel->name, evsel->metric_name, NULL, 1, cpu, out, st); 1290 } else if (runtime_stat_n(st, STAT_NSECS, cpu, &rsd) != 0) { 1291 char unit = ' '; 1292 char unit_buf[10] = "/sec"; 1293 1294 total = runtime_stat_avg(st, STAT_NSECS, cpu, &rsd); 1295 if (total) 1296 ratio = convert_unit_double(1000000000.0 * avg / total, &unit); 1297 1298 if (unit != ' ') 1299 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit); 1300 print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio); 1301 } else if (perf_stat_evsel__is(evsel, SMI_NUM)) { 1302 print_smi_cost(config, cpu, out, st, &rsd); 1303 } else { 1304 num = 0; 1305 } 1306 1307 if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) { 1308 struct metric_expr *mexp; 1309 1310 list_for_each_entry (mexp, &me->head, nd) { 1311 if (num++ > 0) 1312 out->new_line(config, ctxp); 1313 generic_metric(config, mexp->metric_expr, mexp->metric_events, 1314 mexp->metric_refs, evsel->name, mexp->metric_name, 1315 mexp->metric_unit, mexp->runtime, cpu, out, st); 1316 } 1317 } 1318 if (num == 0) 1319 print_metric(config, ctxp, NULL, NULL, NULL, 0); 1320 } 1321