1// SPDX-License-Identifier: GPL-2.0 2/* 3 * This is rewrite of original c2c tool introduced in here: 4 * http://lwn.net/Articles/588866/ 5 * 6 * The original tool was changed to fit in current perf state. 7 * 8 * Original authors: 9 * Don Zickus <dzickus@redhat.com> 10 * Dick Fowles <fowles@inreach.com> 11 * Joe Mario <jmario@redhat.com> 12 */ 13#include <errno.h> 14#include <inttypes.h> 15#include <linux/compiler.h> 16#include <linux/err.h> 17#include <linux/kernel.h> 18#include <linux/stringify.h> 19#include <linux/zalloc.h> 20#include <asm/bug.h> 21#include <sys/param.h> 22#include "debug.h" 23#include "builtin.h" 24#include <perf/cpumap.h> 25#include <subcmd/pager.h> 26#include <subcmd/parse-options.h> 27#include "map_symbol.h" 28#include "mem-events.h" 29#include "session.h" 30#include "hist.h" 31#include "sort.h" 32#include "tool.h" 33#include "cacheline.h" 34#include "data.h" 35#include "event.h" 36#include "evlist.h" 37#include "evsel.h" 38#include "ui/browsers/hists.h" 39#include "thread.h" 40#include "mem2node.h" 41#include "mem-info.h" 42#include "symbol.h" 43#include "ui/ui.h" 44#include "ui/progress.h" 45#include "pmus.h" 46#include "string2.h" 47#include "util/util.h" 48 49struct c2c_hists { 50 struct hists hists; 51 struct perf_hpp_list list; 52 struct c2c_stats stats; 53}; 54 55struct compute_stats { 56 struct stats lcl_hitm; 57 struct stats rmt_hitm; 58 struct stats lcl_peer; 59 struct stats rmt_peer; 60 struct stats load; 61}; 62 63struct c2c_hist_entry { 64 struct c2c_hists *hists; 65 struct c2c_stats stats; 66 unsigned long *cpuset; 67 unsigned long *nodeset; 68 struct c2c_stats *node_stats; 69 unsigned int cacheline_idx; 70 71 struct compute_stats cstats; 72 73 unsigned long paddr; 74 unsigned long paddr_cnt; 75 bool paddr_zero; 76 char *nodestr; 77 78 /* 79 * must be at the end, 80 * because of its callchain dynamic entry 81 */ 82 struct hist_entry he; 83}; 84 85static char const *coalesce_default = "iaddr"; 86 87struct perf_c2c { 88 struct perf_tool tool; 89 struct c2c_hists hists; 90 struct mem2node mem2node; 91 92 unsigned long **nodes; 93 int nodes_cnt; 94 int cpus_cnt; 95 int *cpu2node; 96 int node_info; 97 98 bool show_src; 99 bool show_all; 100 bool use_stdio; 101 bool stats_only; 102 bool symbol_full; 103 bool stitch_lbr; 104 105 /* Shared cache line stats */ 106 struct c2c_stats shared_clines_stats; 107 int shared_clines; 108 109 int display; 110 111 const char *coalesce; 112 char *cl_sort; 113 char *cl_resort; 114 char *cl_output; 115}; 116 117enum { 118 DISPLAY_LCL_HITM, 119 DISPLAY_RMT_HITM, 120 DISPLAY_TOT_HITM, 121 DISPLAY_SNP_PEER, 122 DISPLAY_MAX, 123}; 124 125static const char *display_str[DISPLAY_MAX] = { 126 [DISPLAY_LCL_HITM] = "Local HITMs", 127 [DISPLAY_RMT_HITM] = "Remote HITMs", 128 [DISPLAY_TOT_HITM] = "Total HITMs", 129 [DISPLAY_SNP_PEER] = "Peer Snoop", 130}; 131 132static const struct option c2c_options[] = { 133 OPT_INCR('v', "verbose", &verbose, "be more verbose (show counter open errors, etc)"), 134 OPT_END() 135}; 136 137static struct perf_c2c c2c; 138 139static void *c2c_he_zalloc(size_t size) 140{ 141 struct c2c_hist_entry *c2c_he; 142 143 c2c_he = zalloc(size + sizeof(*c2c_he)); 144 if (!c2c_he) 145 return NULL; 146 147 c2c_he->cpuset = bitmap_zalloc(c2c.cpus_cnt); 148 if (!c2c_he->cpuset) 149 goto out_free; 150 151 c2c_he->nodeset = bitmap_zalloc(c2c.nodes_cnt); 152 if (!c2c_he->nodeset) 153 goto out_free; 154 155 c2c_he->node_stats = zalloc(c2c.nodes_cnt * sizeof(*c2c_he->node_stats)); 156 if (!c2c_he->node_stats) 157 goto out_free; 158 159 init_stats(&c2c_he->cstats.lcl_hitm); 160 init_stats(&c2c_he->cstats.rmt_hitm); 161 init_stats(&c2c_he->cstats.lcl_peer); 162 init_stats(&c2c_he->cstats.rmt_peer); 163 init_stats(&c2c_he->cstats.load); 164 165 return &c2c_he->he; 166 167out_free: 168 zfree(&c2c_he->nodeset); 169 zfree(&c2c_he->cpuset); 170 free(c2c_he); 171 return NULL; 172} 173 174static void c2c_he_free(void *he) 175{ 176 struct c2c_hist_entry *c2c_he; 177 178 c2c_he = container_of(he, struct c2c_hist_entry, he); 179 if (c2c_he->hists) { 180 hists__delete_entries(&c2c_he->hists->hists); 181 zfree(&c2c_he->hists); 182 } 183 184 zfree(&c2c_he->cpuset); 185 zfree(&c2c_he->nodeset); 186 zfree(&c2c_he->nodestr); 187 zfree(&c2c_he->node_stats); 188 free(c2c_he); 189} 190 191static struct hist_entry_ops c2c_entry_ops = { 192 .new = c2c_he_zalloc, 193 .free = c2c_he_free, 194}; 195 196static int c2c_hists__init(struct c2c_hists *hists, 197 const char *sort, 198 int nr_header_lines); 199 200static struct c2c_hists* 201he__get_c2c_hists(struct hist_entry *he, 202 const char *sort, 203 int nr_header_lines) 204{ 205 struct c2c_hist_entry *c2c_he; 206 struct c2c_hists *hists; 207 int ret; 208 209 c2c_he = container_of(he, struct c2c_hist_entry, he); 210 if (c2c_he->hists) 211 return c2c_he->hists; 212 213 hists = c2c_he->hists = zalloc(sizeof(*hists)); 214 if (!hists) 215 return NULL; 216 217 ret = c2c_hists__init(hists, sort, nr_header_lines); 218 if (ret) { 219 free(hists); 220 return NULL; 221 } 222 223 return hists; 224} 225 226static void c2c_he__set_cpu(struct c2c_hist_entry *c2c_he, 227 struct perf_sample *sample) 228{ 229 if (WARN_ONCE(sample->cpu == (unsigned int) -1, 230 "WARNING: no sample cpu value")) 231 return; 232 233 __set_bit(sample->cpu, c2c_he->cpuset); 234} 235 236static void c2c_he__set_node(struct c2c_hist_entry *c2c_he, 237 struct perf_sample *sample) 238{ 239 int node; 240 241 if (!sample->phys_addr) { 242 c2c_he->paddr_zero = true; 243 return; 244 } 245 246 node = mem2node__node(&c2c.mem2node, sample->phys_addr); 247 if (WARN_ONCE(node < 0, "WARNING: failed to find node\n")) 248 return; 249 250 __set_bit(node, c2c_he->nodeset); 251 252 if (c2c_he->paddr != sample->phys_addr) { 253 c2c_he->paddr_cnt++; 254 c2c_he->paddr = sample->phys_addr; 255 } 256} 257 258static void compute_stats(struct c2c_hist_entry *c2c_he, 259 struct c2c_stats *stats, 260 u64 weight) 261{ 262 struct compute_stats *cstats = &c2c_he->cstats; 263 264 if (stats->rmt_hitm) 265 update_stats(&cstats->rmt_hitm, weight); 266 else if (stats->lcl_hitm) 267 update_stats(&cstats->lcl_hitm, weight); 268 else if (stats->rmt_peer) 269 update_stats(&cstats->rmt_peer, weight); 270 else if (stats->lcl_peer) 271 update_stats(&cstats->lcl_peer, weight); 272 else if (stats->load) 273 update_stats(&cstats->load, weight); 274} 275 276static int process_sample_event(struct perf_tool *tool __maybe_unused, 277 union perf_event *event, 278 struct perf_sample *sample, 279 struct evsel *evsel, 280 struct machine *machine) 281{ 282 struct c2c_hists *c2c_hists = &c2c.hists; 283 struct c2c_hist_entry *c2c_he; 284 struct c2c_stats stats = { .nr_entries = 0, }; 285 struct hist_entry *he; 286 struct addr_location al; 287 struct mem_info *mi, *mi_dup; 288 struct callchain_cursor *cursor; 289 int ret; 290 291 addr_location__init(&al); 292 if (machine__resolve(machine, &al, sample) < 0) { 293 pr_debug("problem processing %d event, skipping it.\n", 294 event->header.type); 295 ret = -1; 296 goto out; 297 } 298 299 if (c2c.stitch_lbr) 300 thread__set_lbr_stitch_enable(al.thread, true); 301 302 cursor = get_tls_callchain_cursor(); 303 ret = sample__resolve_callchain(sample, cursor, NULL, 304 evsel, &al, sysctl_perf_event_max_stack); 305 if (ret) 306 goto out; 307 308 mi = sample__resolve_mem(sample, &al); 309 if (mi == NULL) { 310 ret = -ENOMEM; 311 goto out; 312 } 313 314 /* 315 * The mi object is released in hists__add_entry_ops, 316 * if it gets sorted out into existing data, so we need 317 * to take the copy now. 318 */ 319 mi_dup = mem_info__get(mi); 320 321 c2c_decode_stats(&stats, mi); 322 323 he = hists__add_entry_ops(&c2c_hists->hists, &c2c_entry_ops, 324 &al, NULL, NULL, mi, NULL, 325 sample, true); 326 if (he == NULL) 327 goto free_mi; 328 329 c2c_he = container_of(he, struct c2c_hist_entry, he); 330 c2c_add_stats(&c2c_he->stats, &stats); 331 c2c_add_stats(&c2c_hists->stats, &stats); 332 333 c2c_he__set_cpu(c2c_he, sample); 334 c2c_he__set_node(c2c_he, sample); 335 336 hists__inc_nr_samples(&c2c_hists->hists, he->filtered); 337 ret = hist_entry__append_callchain(he, sample); 338 339 if (!ret) { 340 /* 341 * There's already been warning about missing 342 * sample's cpu value. Let's account all to 343 * node 0 in this case, without any further 344 * warning. 345 * 346 * Doing node stats only for single callchain data. 347 */ 348 int cpu = sample->cpu == (unsigned int) -1 ? 0 : sample->cpu; 349 int node = c2c.cpu2node[cpu]; 350 351 mi = mi_dup; 352 353 c2c_hists = he__get_c2c_hists(he, c2c.cl_sort, 2); 354 if (!c2c_hists) 355 goto free_mi; 356 357 he = hists__add_entry_ops(&c2c_hists->hists, &c2c_entry_ops, 358 &al, NULL, NULL, mi, NULL, 359 sample, true); 360 if (he == NULL) 361 goto free_mi; 362 363 c2c_he = container_of(he, struct c2c_hist_entry, he); 364 c2c_add_stats(&c2c_he->stats, &stats); 365 c2c_add_stats(&c2c_hists->stats, &stats); 366 c2c_add_stats(&c2c_he->node_stats[node], &stats); 367 368 compute_stats(c2c_he, &stats, sample->weight); 369 370 c2c_he__set_cpu(c2c_he, sample); 371 c2c_he__set_node(c2c_he, sample); 372 373 hists__inc_nr_samples(&c2c_hists->hists, he->filtered); 374 ret = hist_entry__append_callchain(he, sample); 375 } 376 377out: 378 addr_location__exit(&al); 379 return ret; 380 381free_mi: 382 mem_info__put(mi_dup); 383 mem_info__put(mi); 384 ret = -ENOMEM; 385 goto out; 386} 387 388static struct perf_c2c c2c = { 389 .tool = { 390 .sample = process_sample_event, 391 .mmap = perf_event__process_mmap, 392 .mmap2 = perf_event__process_mmap2, 393 .comm = perf_event__process_comm, 394 .exit = perf_event__process_exit, 395 .fork = perf_event__process_fork, 396 .lost = perf_event__process_lost, 397 .attr = perf_event__process_attr, 398 .auxtrace_info = perf_event__process_auxtrace_info, 399 .auxtrace = perf_event__process_auxtrace, 400 .auxtrace_error = perf_event__process_auxtrace_error, 401 .ordered_events = true, 402 .ordering_requires_timestamps = true, 403 }, 404}; 405 406static const char * const c2c_usage[] = { 407 "perf c2c {record|report}", 408 NULL 409}; 410 411static const char * const __usage_report[] = { 412 "perf c2c report", 413 NULL 414}; 415 416static const char * const *report_c2c_usage = __usage_report; 417 418#define C2C_HEADER_MAX 2 419 420struct c2c_header { 421 struct { 422 const char *text; 423 int span; 424 } line[C2C_HEADER_MAX]; 425}; 426 427struct c2c_dimension { 428 struct c2c_header header; 429 const char *name; 430 int width; 431 struct sort_entry *se; 432 433 int64_t (*cmp)(struct perf_hpp_fmt *fmt, 434 struct hist_entry *, struct hist_entry *); 435 int (*entry)(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 436 struct hist_entry *he); 437 int (*color)(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 438 struct hist_entry *he); 439}; 440 441struct c2c_fmt { 442 struct perf_hpp_fmt fmt; 443 struct c2c_dimension *dim; 444}; 445 446#define SYMBOL_WIDTH 30 447 448static struct c2c_dimension dim_symbol; 449static struct c2c_dimension dim_srcline; 450 451static int symbol_width(struct hists *hists, struct sort_entry *se) 452{ 453 int width = hists__col_len(hists, se->se_width_idx); 454 455 if (!c2c.symbol_full) 456 width = MIN(width, SYMBOL_WIDTH); 457 458 return width; 459} 460 461static int c2c_width(struct perf_hpp_fmt *fmt, 462 struct perf_hpp *hpp __maybe_unused, 463 struct hists *hists) 464{ 465 struct c2c_fmt *c2c_fmt; 466 struct c2c_dimension *dim; 467 468 c2c_fmt = container_of(fmt, struct c2c_fmt, fmt); 469 dim = c2c_fmt->dim; 470 471 if (dim == &dim_symbol || dim == &dim_srcline) 472 return symbol_width(hists, dim->se); 473 474 return dim->se ? hists__col_len(hists, dim->se->se_width_idx) : 475 c2c_fmt->dim->width; 476} 477 478static int c2c_header(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 479 struct hists *hists, int line, int *span) 480{ 481 struct perf_hpp_list *hpp_list = hists->hpp_list; 482 struct c2c_fmt *c2c_fmt; 483 struct c2c_dimension *dim; 484 const char *text = NULL; 485 int width = c2c_width(fmt, hpp, hists); 486 487 c2c_fmt = container_of(fmt, struct c2c_fmt, fmt); 488 dim = c2c_fmt->dim; 489 490 if (dim->se) { 491 text = dim->header.line[line].text; 492 /* Use the last line from sort_entry if not defined. */ 493 if (!text && (line == hpp_list->nr_header_lines - 1)) 494 text = dim->se->se_header; 495 } else { 496 text = dim->header.line[line].text; 497 498 if (*span) { 499 (*span)--; 500 return 0; 501 } else { 502 *span = dim->header.line[line].span; 503 } 504 } 505 506 if (text == NULL) 507 text = ""; 508 509 return scnprintf(hpp->buf, hpp->size, "%*s", width, text); 510} 511 512#define HEX_STR(__s, __v) \ 513({ \ 514 scnprintf(__s, sizeof(__s), "0x%" PRIx64, __v); \ 515 __s; \ 516}) 517 518static int64_t 519dcacheline_cmp(struct perf_hpp_fmt *fmt __maybe_unused, 520 struct hist_entry *left, struct hist_entry *right) 521{ 522 return sort__dcacheline_cmp(left, right); 523} 524 525static int dcacheline_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 526 struct hist_entry *he) 527{ 528 uint64_t addr = 0; 529 int width = c2c_width(fmt, hpp, he->hists); 530 char buf[20]; 531 532 if (he->mem_info) 533 addr = cl_address(mem_info__daddr(he->mem_info)->addr, chk_double_cl); 534 535 return scnprintf(hpp->buf, hpp->size, "%*s", width, HEX_STR(buf, addr)); 536} 537 538static int 539dcacheline_node_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 540 struct hist_entry *he) 541{ 542 struct c2c_hist_entry *c2c_he; 543 int width = c2c_width(fmt, hpp, he->hists); 544 545 c2c_he = container_of(he, struct c2c_hist_entry, he); 546 if (WARN_ON_ONCE(!c2c_he->nodestr)) 547 return 0; 548 549 return scnprintf(hpp->buf, hpp->size, "%*s", width, c2c_he->nodestr); 550} 551 552static int 553dcacheline_node_count(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 554 struct hist_entry *he) 555{ 556 struct c2c_hist_entry *c2c_he; 557 int width = c2c_width(fmt, hpp, he->hists); 558 559 c2c_he = container_of(he, struct c2c_hist_entry, he); 560 return scnprintf(hpp->buf, hpp->size, "%*lu", width, c2c_he->paddr_cnt); 561} 562 563static int offset_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 564 struct hist_entry *he) 565{ 566 uint64_t addr = 0; 567 int width = c2c_width(fmt, hpp, he->hists); 568 char buf[20]; 569 570 if (he->mem_info) 571 addr = cl_offset(mem_info__daddr(he->mem_info)->al_addr, chk_double_cl); 572 573 return scnprintf(hpp->buf, hpp->size, "%*s", width, HEX_STR(buf, addr)); 574} 575 576static int64_t 577offset_cmp(struct perf_hpp_fmt *fmt __maybe_unused, 578 struct hist_entry *left, struct hist_entry *right) 579{ 580 uint64_t l = 0, r = 0; 581 582 if (left->mem_info) 583 l = cl_offset(mem_info__daddr(left->mem_info)->addr, chk_double_cl); 584 585 if (right->mem_info) 586 r = cl_offset(mem_info__daddr(right->mem_info)->addr, chk_double_cl); 587 588 return (int64_t)(r - l); 589} 590 591static int 592iaddr_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 593 struct hist_entry *he) 594{ 595 uint64_t addr = 0; 596 int width = c2c_width(fmt, hpp, he->hists); 597 char buf[20]; 598 599 if (he->mem_info) 600 addr = mem_info__iaddr(he->mem_info)->addr; 601 602 return scnprintf(hpp->buf, hpp->size, "%*s", width, HEX_STR(buf, addr)); 603} 604 605static int64_t 606iaddr_cmp(struct perf_hpp_fmt *fmt __maybe_unused, 607 struct hist_entry *left, struct hist_entry *right) 608{ 609 return sort__iaddr_cmp(left, right); 610} 611 612static int 613tot_hitm_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 614 struct hist_entry *he) 615{ 616 struct c2c_hist_entry *c2c_he; 617 int width = c2c_width(fmt, hpp, he->hists); 618 unsigned int tot_hitm; 619 620 c2c_he = container_of(he, struct c2c_hist_entry, he); 621 tot_hitm = c2c_he->stats.lcl_hitm + c2c_he->stats.rmt_hitm; 622 623 return scnprintf(hpp->buf, hpp->size, "%*u", width, tot_hitm); 624} 625 626static int64_t 627tot_hitm_cmp(struct perf_hpp_fmt *fmt __maybe_unused, 628 struct hist_entry *left, struct hist_entry *right) 629{ 630 struct c2c_hist_entry *c2c_left; 631 struct c2c_hist_entry *c2c_right; 632 uint64_t tot_hitm_left; 633 uint64_t tot_hitm_right; 634 635 c2c_left = container_of(left, struct c2c_hist_entry, he); 636 c2c_right = container_of(right, struct c2c_hist_entry, he); 637 638 tot_hitm_left = c2c_left->stats.lcl_hitm + c2c_left->stats.rmt_hitm; 639 tot_hitm_right = c2c_right->stats.lcl_hitm + c2c_right->stats.rmt_hitm; 640 641 return tot_hitm_left - tot_hitm_right; 642} 643 644#define STAT_FN_ENTRY(__f) \ 645static int \ 646__f ## _entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, \ 647 struct hist_entry *he) \ 648{ \ 649 struct c2c_hist_entry *c2c_he; \ 650 int width = c2c_width(fmt, hpp, he->hists); \ 651 \ 652 c2c_he = container_of(he, struct c2c_hist_entry, he); \ 653 return scnprintf(hpp->buf, hpp->size, "%*u", width, \ 654 c2c_he->stats.__f); \ 655} 656 657#define STAT_FN_CMP(__f) \ 658static int64_t \ 659__f ## _cmp(struct perf_hpp_fmt *fmt __maybe_unused, \ 660 struct hist_entry *left, struct hist_entry *right) \ 661{ \ 662 struct c2c_hist_entry *c2c_left, *c2c_right; \ 663 \ 664 c2c_left = container_of(left, struct c2c_hist_entry, he); \ 665 c2c_right = container_of(right, struct c2c_hist_entry, he); \ 666 return (uint64_t) c2c_left->stats.__f - \ 667 (uint64_t) c2c_right->stats.__f; \ 668} 669 670#define STAT_FN(__f) \ 671 STAT_FN_ENTRY(__f) \ 672 STAT_FN_CMP(__f) 673 674STAT_FN(rmt_hitm) 675STAT_FN(lcl_hitm) 676STAT_FN(rmt_peer) 677STAT_FN(lcl_peer) 678STAT_FN(tot_peer) 679STAT_FN(store) 680STAT_FN(st_l1hit) 681STAT_FN(st_l1miss) 682STAT_FN(st_na) 683STAT_FN(ld_fbhit) 684STAT_FN(ld_l1hit) 685STAT_FN(ld_l2hit) 686STAT_FN(ld_llchit) 687STAT_FN(rmt_hit) 688 689static uint64_t get_load_llc_misses(struct c2c_stats *stats) 690{ 691 return stats->lcl_dram + 692 stats->rmt_dram + 693 stats->rmt_hitm + 694 stats->rmt_hit; 695} 696 697static uint64_t get_load_cache_hits(struct c2c_stats *stats) 698{ 699 return stats->ld_fbhit + 700 stats->ld_l1hit + 701 stats->ld_l2hit + 702 stats->ld_llchit + 703 stats->lcl_hitm; 704} 705 706static uint64_t get_stores(struct c2c_stats *stats) 707{ 708 return stats->st_l1hit + 709 stats->st_l1miss + 710 stats->st_na; 711} 712 713static uint64_t total_records(struct c2c_stats *stats) 714{ 715 return get_load_llc_misses(stats) + 716 get_load_cache_hits(stats) + 717 get_stores(stats); 718} 719 720static int 721tot_recs_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 722 struct hist_entry *he) 723{ 724 struct c2c_hist_entry *c2c_he; 725 int width = c2c_width(fmt, hpp, he->hists); 726 uint64_t tot_recs; 727 728 c2c_he = container_of(he, struct c2c_hist_entry, he); 729 tot_recs = total_records(&c2c_he->stats); 730 731 return scnprintf(hpp->buf, hpp->size, "%*" PRIu64, width, tot_recs); 732} 733 734static int64_t 735tot_recs_cmp(struct perf_hpp_fmt *fmt __maybe_unused, 736 struct hist_entry *left, struct hist_entry *right) 737{ 738 struct c2c_hist_entry *c2c_left; 739 struct c2c_hist_entry *c2c_right; 740 uint64_t tot_recs_left; 741 uint64_t tot_recs_right; 742 743 c2c_left = container_of(left, struct c2c_hist_entry, he); 744 c2c_right = container_of(right, struct c2c_hist_entry, he); 745 746 tot_recs_left = total_records(&c2c_left->stats); 747 tot_recs_right = total_records(&c2c_right->stats); 748 749 return tot_recs_left - tot_recs_right; 750} 751 752static uint64_t total_loads(struct c2c_stats *stats) 753{ 754 return get_load_llc_misses(stats) + 755 get_load_cache_hits(stats); 756} 757 758static int 759tot_loads_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 760 struct hist_entry *he) 761{ 762 struct c2c_hist_entry *c2c_he; 763 int width = c2c_width(fmt, hpp, he->hists); 764 uint64_t tot_recs; 765 766 c2c_he = container_of(he, struct c2c_hist_entry, he); 767 tot_recs = total_loads(&c2c_he->stats); 768 769 return scnprintf(hpp->buf, hpp->size, "%*" PRIu64, width, tot_recs); 770} 771 772static int64_t 773tot_loads_cmp(struct perf_hpp_fmt *fmt __maybe_unused, 774 struct hist_entry *left, struct hist_entry *right) 775{ 776 struct c2c_hist_entry *c2c_left; 777 struct c2c_hist_entry *c2c_right; 778 uint64_t tot_recs_left; 779 uint64_t tot_recs_right; 780 781 c2c_left = container_of(left, struct c2c_hist_entry, he); 782 c2c_right = container_of(right, struct c2c_hist_entry, he); 783 784 tot_recs_left = total_loads(&c2c_left->stats); 785 tot_recs_right = total_loads(&c2c_right->stats); 786 787 return tot_recs_left - tot_recs_right; 788} 789 790typedef double (get_percent_cb)(struct c2c_hist_entry *); 791 792static int 793percent_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 794 struct hist_entry *he, get_percent_cb get_percent) 795{ 796 struct c2c_hist_entry *c2c_he; 797 int width = c2c_width(fmt, hpp, he->hists); 798 double per; 799 800 c2c_he = container_of(he, struct c2c_hist_entry, he); 801 per = get_percent(c2c_he); 802 803#ifdef HAVE_SLANG_SUPPORT 804 if (use_browser) 805 return __hpp__slsmg_color_printf(hpp, "%*.2f%%", width - 1, per); 806#endif 807 return hpp_color_scnprintf(hpp, "%*.2f%%", width - 1, per); 808} 809 810static double percent_costly_snoop(struct c2c_hist_entry *c2c_he) 811{ 812 struct c2c_hists *hists; 813 struct c2c_stats *stats; 814 struct c2c_stats *total; 815 int tot = 0, st = 0; 816 double p; 817 818 hists = container_of(c2c_he->he.hists, struct c2c_hists, hists); 819 stats = &c2c_he->stats; 820 total = &hists->stats; 821 822 switch (c2c.display) { 823 case DISPLAY_RMT_HITM: 824 st = stats->rmt_hitm; 825 tot = total->rmt_hitm; 826 break; 827 case DISPLAY_LCL_HITM: 828 st = stats->lcl_hitm; 829 tot = total->lcl_hitm; 830 break; 831 case DISPLAY_TOT_HITM: 832 st = stats->tot_hitm; 833 tot = total->tot_hitm; 834 break; 835 case DISPLAY_SNP_PEER: 836 st = stats->tot_peer; 837 tot = total->tot_peer; 838 break; 839 default: 840 break; 841 } 842 843 p = tot ? (double) st / tot : 0; 844 845 return 100 * p; 846} 847 848#define PERC_STR(__s, __v) \ 849({ \ 850 scnprintf(__s, sizeof(__s), "%.2F%%", __v); \ 851 __s; \ 852}) 853 854static int 855percent_costly_snoop_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 856 struct hist_entry *he) 857{ 858 struct c2c_hist_entry *c2c_he; 859 int width = c2c_width(fmt, hpp, he->hists); 860 char buf[10]; 861 double per; 862 863 c2c_he = container_of(he, struct c2c_hist_entry, he); 864 per = percent_costly_snoop(c2c_he); 865 return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per)); 866} 867 868static int 869percent_costly_snoop_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 870 struct hist_entry *he) 871{ 872 return percent_color(fmt, hpp, he, percent_costly_snoop); 873} 874 875static int64_t 876percent_costly_snoop_cmp(struct perf_hpp_fmt *fmt __maybe_unused, 877 struct hist_entry *left, struct hist_entry *right) 878{ 879 struct c2c_hist_entry *c2c_left; 880 struct c2c_hist_entry *c2c_right; 881 double per_left; 882 double per_right; 883 884 c2c_left = container_of(left, struct c2c_hist_entry, he); 885 c2c_right = container_of(right, struct c2c_hist_entry, he); 886 887 per_left = percent_costly_snoop(c2c_left); 888 per_right = percent_costly_snoop(c2c_right); 889 890 return per_left - per_right; 891} 892 893static struct c2c_stats *he_stats(struct hist_entry *he) 894{ 895 struct c2c_hist_entry *c2c_he; 896 897 c2c_he = container_of(he, struct c2c_hist_entry, he); 898 return &c2c_he->stats; 899} 900 901static struct c2c_stats *total_stats(struct hist_entry *he) 902{ 903 struct c2c_hists *hists; 904 905 hists = container_of(he->hists, struct c2c_hists, hists); 906 return &hists->stats; 907} 908 909static double percent(u32 st, u32 tot) 910{ 911 return tot ? 100. * (double) st / (double) tot : 0; 912} 913 914#define PERCENT(__h, __f) percent(he_stats(__h)->__f, total_stats(__h)->__f) 915 916#define PERCENT_FN(__f) \ 917static double percent_ ## __f(struct c2c_hist_entry *c2c_he) \ 918{ \ 919 struct c2c_hists *hists; \ 920 \ 921 hists = container_of(c2c_he->he.hists, struct c2c_hists, hists); \ 922 return percent(c2c_he->stats.__f, hists->stats.__f); \ 923} 924 925PERCENT_FN(rmt_hitm) 926PERCENT_FN(lcl_hitm) 927PERCENT_FN(rmt_peer) 928PERCENT_FN(lcl_peer) 929PERCENT_FN(st_l1hit) 930PERCENT_FN(st_l1miss) 931PERCENT_FN(st_na) 932 933static int 934percent_rmt_hitm_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 935 struct hist_entry *he) 936{ 937 int width = c2c_width(fmt, hpp, he->hists); 938 double per = PERCENT(he, rmt_hitm); 939 char buf[10]; 940 941 return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per)); 942} 943 944static int 945percent_rmt_hitm_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 946 struct hist_entry *he) 947{ 948 return percent_color(fmt, hpp, he, percent_rmt_hitm); 949} 950 951static int64_t 952percent_rmt_hitm_cmp(struct perf_hpp_fmt *fmt __maybe_unused, 953 struct hist_entry *left, struct hist_entry *right) 954{ 955 double per_left; 956 double per_right; 957 958 per_left = PERCENT(left, rmt_hitm); 959 per_right = PERCENT(right, rmt_hitm); 960 961 return per_left - per_right; 962} 963 964static int 965percent_lcl_hitm_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 966 struct hist_entry *he) 967{ 968 int width = c2c_width(fmt, hpp, he->hists); 969 double per = PERCENT(he, lcl_hitm); 970 char buf[10]; 971 972 return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per)); 973} 974 975static int 976percent_lcl_hitm_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 977 struct hist_entry *he) 978{ 979 return percent_color(fmt, hpp, he, percent_lcl_hitm); 980} 981 982static int64_t 983percent_lcl_hitm_cmp(struct perf_hpp_fmt *fmt __maybe_unused, 984 struct hist_entry *left, struct hist_entry *right) 985{ 986 double per_left; 987 double per_right; 988 989 per_left = PERCENT(left, lcl_hitm); 990 per_right = PERCENT(right, lcl_hitm); 991 992 return per_left - per_right; 993} 994 995static int 996percent_lcl_peer_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 997 struct hist_entry *he) 998{ 999 int width = c2c_width(fmt, hpp, he->hists); 1000 double per = PERCENT(he, lcl_peer); 1001 char buf[10]; 1002 1003 return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per)); 1004} 1005 1006static int 1007percent_lcl_peer_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 1008 struct hist_entry *he) 1009{ 1010 return percent_color(fmt, hpp, he, percent_lcl_peer); 1011} 1012 1013static int64_t 1014percent_lcl_peer_cmp(struct perf_hpp_fmt *fmt __maybe_unused, 1015 struct hist_entry *left, struct hist_entry *right) 1016{ 1017 double per_left; 1018 double per_right; 1019 1020 per_left = PERCENT(left, lcl_peer); 1021 per_right = PERCENT(right, lcl_peer); 1022 1023 return per_left - per_right; 1024} 1025 1026static int 1027percent_rmt_peer_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 1028 struct hist_entry *he) 1029{ 1030 int width = c2c_width(fmt, hpp, he->hists); 1031 double per = PERCENT(he, rmt_peer); 1032 char buf[10]; 1033 1034 return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per)); 1035} 1036 1037static int 1038percent_rmt_peer_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 1039 struct hist_entry *he) 1040{ 1041 return percent_color(fmt, hpp, he, percent_rmt_peer); 1042} 1043 1044static int64_t 1045percent_rmt_peer_cmp(struct perf_hpp_fmt *fmt __maybe_unused, 1046 struct hist_entry *left, struct hist_entry *right) 1047{ 1048 double per_left; 1049 double per_right; 1050 1051 per_left = PERCENT(left, rmt_peer); 1052 per_right = PERCENT(right, rmt_peer); 1053 1054 return per_left - per_right; 1055} 1056 1057static int 1058percent_stores_l1hit_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 1059 struct hist_entry *he) 1060{ 1061 int width = c2c_width(fmt, hpp, he->hists); 1062 double per = PERCENT(he, st_l1hit); 1063 char buf[10]; 1064 1065 return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per)); 1066} 1067 1068static int 1069percent_stores_l1hit_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 1070 struct hist_entry *he) 1071{ 1072 return percent_color(fmt, hpp, he, percent_st_l1hit); 1073} 1074 1075static int64_t 1076percent_stores_l1hit_cmp(struct perf_hpp_fmt *fmt __maybe_unused, 1077 struct hist_entry *left, struct hist_entry *right) 1078{ 1079 double per_left; 1080 double per_right; 1081 1082 per_left = PERCENT(left, st_l1hit); 1083 per_right = PERCENT(right, st_l1hit); 1084 1085 return per_left - per_right; 1086} 1087 1088static int 1089percent_stores_l1miss_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 1090 struct hist_entry *he) 1091{ 1092 int width = c2c_width(fmt, hpp, he->hists); 1093 double per = PERCENT(he, st_l1miss); 1094 char buf[10]; 1095 1096 return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per)); 1097} 1098 1099static int 1100percent_stores_l1miss_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 1101 struct hist_entry *he) 1102{ 1103 return percent_color(fmt, hpp, he, percent_st_l1miss); 1104} 1105 1106static int64_t 1107percent_stores_l1miss_cmp(struct perf_hpp_fmt *fmt __maybe_unused, 1108 struct hist_entry *left, struct hist_entry *right) 1109{ 1110 double per_left; 1111 double per_right; 1112 1113 per_left = PERCENT(left, st_l1miss); 1114 per_right = PERCENT(right, st_l1miss); 1115 1116 return per_left - per_right; 1117} 1118 1119static int 1120percent_stores_na_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 1121 struct hist_entry *he) 1122{ 1123 int width = c2c_width(fmt, hpp, he->hists); 1124 double per = PERCENT(he, st_na); 1125 char buf[10]; 1126 1127 return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per)); 1128} 1129 1130static int 1131percent_stores_na_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 1132 struct hist_entry *he) 1133{ 1134 return percent_color(fmt, hpp, he, percent_st_na); 1135} 1136 1137static int64_t 1138percent_stores_na_cmp(struct perf_hpp_fmt *fmt __maybe_unused, 1139 struct hist_entry *left, struct hist_entry *right) 1140{ 1141 double per_left; 1142 double per_right; 1143 1144 per_left = PERCENT(left, st_na); 1145 per_right = PERCENT(right, st_na); 1146 1147 return per_left - per_right; 1148} 1149 1150STAT_FN(lcl_dram) 1151STAT_FN(rmt_dram) 1152 1153static int 1154pid_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 1155 struct hist_entry *he) 1156{ 1157 int width = c2c_width(fmt, hpp, he->hists); 1158 1159 return scnprintf(hpp->buf, hpp->size, "%*d", width, thread__pid(he->thread)); 1160} 1161 1162static int64_t 1163pid_cmp(struct perf_hpp_fmt *fmt __maybe_unused, 1164 struct hist_entry *left, struct hist_entry *right) 1165{ 1166 return thread__pid(left->thread) - thread__pid(right->thread); 1167} 1168 1169static int64_t 1170empty_cmp(struct perf_hpp_fmt *fmt __maybe_unused, 1171 struct hist_entry *left __maybe_unused, 1172 struct hist_entry *right __maybe_unused) 1173{ 1174 return 0; 1175} 1176 1177static int display_metrics(struct perf_hpp *hpp, u32 val, u32 sum) 1178{ 1179 int ret; 1180 1181 if (sum != 0) 1182 ret = scnprintf(hpp->buf, hpp->size, "%5.1f%% ", 1183 percent(val, sum)); 1184 else 1185 ret = scnprintf(hpp->buf, hpp->size, "%6s ", "n/a"); 1186 1187 return ret; 1188} 1189 1190static int 1191node_entry(struct perf_hpp_fmt *fmt __maybe_unused, struct perf_hpp *hpp, 1192 struct hist_entry *he) 1193{ 1194 struct c2c_hist_entry *c2c_he; 1195 bool first = true; 1196 int node; 1197 int ret = 0; 1198 1199 c2c_he = container_of(he, struct c2c_hist_entry, he); 1200 1201 for (node = 0; node < c2c.nodes_cnt; node++) { 1202 DECLARE_BITMAP(set, c2c.cpus_cnt); 1203 1204 bitmap_zero(set, c2c.cpus_cnt); 1205 bitmap_and(set, c2c_he->cpuset, c2c.nodes[node], c2c.cpus_cnt); 1206 1207 if (bitmap_empty(set, c2c.cpus_cnt)) { 1208 if (c2c.node_info == 1) { 1209 ret = scnprintf(hpp->buf, hpp->size, "%21s", " "); 1210 advance_hpp(hpp, ret); 1211 } 1212 continue; 1213 } 1214 1215 if (!first) { 1216 ret = scnprintf(hpp->buf, hpp->size, " "); 1217 advance_hpp(hpp, ret); 1218 } 1219 1220 switch (c2c.node_info) { 1221 case 0: 1222 ret = scnprintf(hpp->buf, hpp->size, "%2d", node); 1223 advance_hpp(hpp, ret); 1224 break; 1225 case 1: 1226 { 1227 int num = bitmap_weight(set, c2c.cpus_cnt); 1228 struct c2c_stats *stats = &c2c_he->node_stats[node]; 1229 1230 ret = scnprintf(hpp->buf, hpp->size, "%2d{%2d ", node, num); 1231 advance_hpp(hpp, ret); 1232 1233 switch (c2c.display) { 1234 case DISPLAY_RMT_HITM: 1235 ret = display_metrics(hpp, stats->rmt_hitm, 1236 c2c_he->stats.rmt_hitm); 1237 break; 1238 case DISPLAY_LCL_HITM: 1239 ret = display_metrics(hpp, stats->lcl_hitm, 1240 c2c_he->stats.lcl_hitm); 1241 break; 1242 case DISPLAY_TOT_HITM: 1243 ret = display_metrics(hpp, stats->tot_hitm, 1244 c2c_he->stats.tot_hitm); 1245 break; 1246 case DISPLAY_SNP_PEER: 1247 ret = display_metrics(hpp, stats->tot_peer, 1248 c2c_he->stats.tot_peer); 1249 break; 1250 default: 1251 break; 1252 } 1253 1254 advance_hpp(hpp, ret); 1255 1256 if (c2c_he->stats.store > 0) { 1257 ret = scnprintf(hpp->buf, hpp->size, "%5.1f%%}", 1258 percent(stats->store, c2c_he->stats.store)); 1259 } else { 1260 ret = scnprintf(hpp->buf, hpp->size, "%6s}", "n/a"); 1261 } 1262 1263 advance_hpp(hpp, ret); 1264 break; 1265 } 1266 case 2: 1267 ret = scnprintf(hpp->buf, hpp->size, "%2d{", node); 1268 advance_hpp(hpp, ret); 1269 1270 ret = bitmap_scnprintf(set, c2c.cpus_cnt, hpp->buf, hpp->size); 1271 advance_hpp(hpp, ret); 1272 1273 ret = scnprintf(hpp->buf, hpp->size, "}"); 1274 advance_hpp(hpp, ret); 1275 break; 1276 default: 1277 break; 1278 } 1279 1280 first = false; 1281 } 1282 1283 return 0; 1284} 1285 1286static int 1287mean_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 1288 struct hist_entry *he, double mean) 1289{ 1290 int width = c2c_width(fmt, hpp, he->hists); 1291 char buf[10]; 1292 1293 scnprintf(buf, 10, "%6.0f", mean); 1294 return scnprintf(hpp->buf, hpp->size, "%*s", width, buf); 1295} 1296 1297#define MEAN_ENTRY(__func, __val) \ 1298static int \ 1299__func(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, struct hist_entry *he) \ 1300{ \ 1301 struct c2c_hist_entry *c2c_he; \ 1302 c2c_he = container_of(he, struct c2c_hist_entry, he); \ 1303 return mean_entry(fmt, hpp, he, avg_stats(&c2c_he->cstats.__val)); \ 1304} 1305 1306MEAN_ENTRY(mean_rmt_entry, rmt_hitm); 1307MEAN_ENTRY(mean_lcl_entry, lcl_hitm); 1308MEAN_ENTRY(mean_load_entry, load); 1309MEAN_ENTRY(mean_rmt_peer_entry, rmt_peer); 1310MEAN_ENTRY(mean_lcl_peer_entry, lcl_peer); 1311 1312static int 1313cpucnt_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 1314 struct hist_entry *he) 1315{ 1316 struct c2c_hist_entry *c2c_he; 1317 int width = c2c_width(fmt, hpp, he->hists); 1318 char buf[10]; 1319 1320 c2c_he = container_of(he, struct c2c_hist_entry, he); 1321 1322 scnprintf(buf, 10, "%d", bitmap_weight(c2c_he->cpuset, c2c.cpus_cnt)); 1323 return scnprintf(hpp->buf, hpp->size, "%*s", width, buf); 1324} 1325 1326static int 1327cl_idx_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 1328 struct hist_entry *he) 1329{ 1330 struct c2c_hist_entry *c2c_he; 1331 int width = c2c_width(fmt, hpp, he->hists); 1332 char buf[10]; 1333 1334 c2c_he = container_of(he, struct c2c_hist_entry, he); 1335 1336 scnprintf(buf, 10, "%u", c2c_he->cacheline_idx); 1337 return scnprintf(hpp->buf, hpp->size, "%*s", width, buf); 1338} 1339 1340static int 1341cl_idx_empty_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 1342 struct hist_entry *he) 1343{ 1344 int width = c2c_width(fmt, hpp, he->hists); 1345 1346 return scnprintf(hpp->buf, hpp->size, "%*s", width, ""); 1347} 1348 1349#define HEADER_LOW(__h) \ 1350 { \ 1351 .line[1] = { \ 1352 .text = __h, \ 1353 }, \ 1354 } 1355 1356#define HEADER_BOTH(__h0, __h1) \ 1357 { \ 1358 .line[0] = { \ 1359 .text = __h0, \ 1360 }, \ 1361 .line[1] = { \ 1362 .text = __h1, \ 1363 }, \ 1364 } 1365 1366#define HEADER_SPAN(__h0, __h1, __s) \ 1367 { \ 1368 .line[0] = { \ 1369 .text = __h0, \ 1370 .span = __s, \ 1371 }, \ 1372 .line[1] = { \ 1373 .text = __h1, \ 1374 }, \ 1375 } 1376 1377#define HEADER_SPAN_LOW(__h) \ 1378 { \ 1379 .line[1] = { \ 1380 .text = __h, \ 1381 }, \ 1382 } 1383 1384static struct c2c_dimension dim_dcacheline = { 1385 .header = HEADER_SPAN("--- Cacheline ----", "Address", 2), 1386 .name = "dcacheline", 1387 .cmp = dcacheline_cmp, 1388 .entry = dcacheline_entry, 1389 .width = 18, 1390}; 1391 1392static struct c2c_dimension dim_dcacheline_node = { 1393 .header = HEADER_LOW("Node"), 1394 .name = "dcacheline_node", 1395 .cmp = empty_cmp, 1396 .entry = dcacheline_node_entry, 1397 .width = 4, 1398}; 1399 1400static struct c2c_dimension dim_dcacheline_count = { 1401 .header = HEADER_LOW("PA cnt"), 1402 .name = "dcacheline_count", 1403 .cmp = empty_cmp, 1404 .entry = dcacheline_node_count, 1405 .width = 6, 1406}; 1407 1408static struct c2c_header header_offset_tui = HEADER_SPAN("-----", "Off", 2); 1409 1410static struct c2c_dimension dim_offset = { 1411 .header = HEADER_SPAN("--- Data address -", "Offset", 2), 1412 .name = "offset", 1413 .cmp = offset_cmp, 1414 .entry = offset_entry, 1415 .width = 18, 1416}; 1417 1418static struct c2c_dimension dim_offset_node = { 1419 .header = HEADER_LOW("Node"), 1420 .name = "offset_node", 1421 .cmp = empty_cmp, 1422 .entry = dcacheline_node_entry, 1423 .width = 4, 1424}; 1425 1426static struct c2c_dimension dim_iaddr = { 1427 .header = HEADER_LOW("Code address"), 1428 .name = "iaddr", 1429 .cmp = iaddr_cmp, 1430 .entry = iaddr_entry, 1431 .width = 18, 1432}; 1433 1434static struct c2c_dimension dim_tot_hitm = { 1435 .header = HEADER_SPAN("------- Load Hitm -------", "Total", 2), 1436 .name = "tot_hitm", 1437 .cmp = tot_hitm_cmp, 1438 .entry = tot_hitm_entry, 1439 .width = 7, 1440}; 1441 1442static struct c2c_dimension dim_lcl_hitm = { 1443 .header = HEADER_SPAN_LOW("LclHitm"), 1444 .name = "lcl_hitm", 1445 .cmp = lcl_hitm_cmp, 1446 .entry = lcl_hitm_entry, 1447 .width = 7, 1448}; 1449 1450static struct c2c_dimension dim_rmt_hitm = { 1451 .header = HEADER_SPAN_LOW("RmtHitm"), 1452 .name = "rmt_hitm", 1453 .cmp = rmt_hitm_cmp, 1454 .entry = rmt_hitm_entry, 1455 .width = 7, 1456}; 1457 1458static struct c2c_dimension dim_tot_peer = { 1459 .header = HEADER_SPAN("------- Load Peer -------", "Total", 2), 1460 .name = "tot_peer", 1461 .cmp = tot_peer_cmp, 1462 .entry = tot_peer_entry, 1463 .width = 7, 1464}; 1465 1466static struct c2c_dimension dim_lcl_peer = { 1467 .header = HEADER_SPAN_LOW("Local"), 1468 .name = "lcl_peer", 1469 .cmp = lcl_peer_cmp, 1470 .entry = lcl_peer_entry, 1471 .width = 7, 1472}; 1473 1474static struct c2c_dimension dim_rmt_peer = { 1475 .header = HEADER_SPAN_LOW("Remote"), 1476 .name = "rmt_peer", 1477 .cmp = rmt_peer_cmp, 1478 .entry = rmt_peer_entry, 1479 .width = 7, 1480}; 1481 1482static struct c2c_dimension dim_cl_rmt_hitm = { 1483 .header = HEADER_SPAN("----- HITM -----", "Rmt", 1), 1484 .name = "cl_rmt_hitm", 1485 .cmp = rmt_hitm_cmp, 1486 .entry = rmt_hitm_entry, 1487 .width = 7, 1488}; 1489 1490static struct c2c_dimension dim_cl_lcl_hitm = { 1491 .header = HEADER_SPAN_LOW("Lcl"), 1492 .name = "cl_lcl_hitm", 1493 .cmp = lcl_hitm_cmp, 1494 .entry = lcl_hitm_entry, 1495 .width = 7, 1496}; 1497 1498static struct c2c_dimension dim_cl_rmt_peer = { 1499 .header = HEADER_SPAN("----- Peer -----", "Rmt", 1), 1500 .name = "cl_rmt_peer", 1501 .cmp = rmt_peer_cmp, 1502 .entry = rmt_peer_entry, 1503 .width = 7, 1504}; 1505 1506static struct c2c_dimension dim_cl_lcl_peer = { 1507 .header = HEADER_SPAN_LOW("Lcl"), 1508 .name = "cl_lcl_peer", 1509 .cmp = lcl_peer_cmp, 1510 .entry = lcl_peer_entry, 1511 .width = 7, 1512}; 1513 1514static struct c2c_dimension dim_tot_stores = { 1515 .header = HEADER_BOTH("Total", "Stores"), 1516 .name = "tot_stores", 1517 .cmp = store_cmp, 1518 .entry = store_entry, 1519 .width = 7, 1520}; 1521 1522static struct c2c_dimension dim_stores_l1hit = { 1523 .header = HEADER_SPAN("--------- Stores --------", "L1Hit", 2), 1524 .name = "stores_l1hit", 1525 .cmp = st_l1hit_cmp, 1526 .entry = st_l1hit_entry, 1527 .width = 7, 1528}; 1529 1530static struct c2c_dimension dim_stores_l1miss = { 1531 .header = HEADER_SPAN_LOW("L1Miss"), 1532 .name = "stores_l1miss", 1533 .cmp = st_l1miss_cmp, 1534 .entry = st_l1miss_entry, 1535 .width = 7, 1536}; 1537 1538static struct c2c_dimension dim_stores_na = { 1539 .header = HEADER_SPAN_LOW("N/A"), 1540 .name = "stores_na", 1541 .cmp = st_na_cmp, 1542 .entry = st_na_entry, 1543 .width = 7, 1544}; 1545 1546static struct c2c_dimension dim_cl_stores_l1hit = { 1547 .header = HEADER_SPAN("------- Store Refs ------", "L1 Hit", 2), 1548 .name = "cl_stores_l1hit", 1549 .cmp = st_l1hit_cmp, 1550 .entry = st_l1hit_entry, 1551 .width = 7, 1552}; 1553 1554static struct c2c_dimension dim_cl_stores_l1miss = { 1555 .header = HEADER_SPAN_LOW("L1 Miss"), 1556 .name = "cl_stores_l1miss", 1557 .cmp = st_l1miss_cmp, 1558 .entry = st_l1miss_entry, 1559 .width = 7, 1560}; 1561 1562static struct c2c_dimension dim_cl_stores_na = { 1563 .header = HEADER_SPAN_LOW("N/A"), 1564 .name = "cl_stores_na", 1565 .cmp = st_na_cmp, 1566 .entry = st_na_entry, 1567 .width = 7, 1568}; 1569 1570static struct c2c_dimension dim_ld_fbhit = { 1571 .header = HEADER_SPAN("----- Core Load Hit -----", "FB", 2), 1572 .name = "ld_fbhit", 1573 .cmp = ld_fbhit_cmp, 1574 .entry = ld_fbhit_entry, 1575 .width = 7, 1576}; 1577 1578static struct c2c_dimension dim_ld_l1hit = { 1579 .header = HEADER_SPAN_LOW("L1"), 1580 .name = "ld_l1hit", 1581 .cmp = ld_l1hit_cmp, 1582 .entry = ld_l1hit_entry, 1583 .width = 7, 1584}; 1585 1586static struct c2c_dimension dim_ld_l2hit = { 1587 .header = HEADER_SPAN_LOW("L2"), 1588 .name = "ld_l2hit", 1589 .cmp = ld_l2hit_cmp, 1590 .entry = ld_l2hit_entry, 1591 .width = 7, 1592}; 1593 1594static struct c2c_dimension dim_ld_llchit = { 1595 .header = HEADER_SPAN("- LLC Load Hit --", "LclHit", 1), 1596 .name = "ld_lclhit", 1597 .cmp = ld_llchit_cmp, 1598 .entry = ld_llchit_entry, 1599 .width = 8, 1600}; 1601 1602static struct c2c_dimension dim_ld_rmthit = { 1603 .header = HEADER_SPAN("- RMT Load Hit --", "RmtHit", 1), 1604 .name = "ld_rmthit", 1605 .cmp = rmt_hit_cmp, 1606 .entry = rmt_hit_entry, 1607 .width = 8, 1608}; 1609 1610static struct c2c_dimension dim_tot_recs = { 1611 .header = HEADER_BOTH("Total", "records"), 1612 .name = "tot_recs", 1613 .cmp = tot_recs_cmp, 1614 .entry = tot_recs_entry, 1615 .width = 7, 1616}; 1617 1618static struct c2c_dimension dim_tot_loads = { 1619 .header = HEADER_BOTH("Total", "Loads"), 1620 .name = "tot_loads", 1621 .cmp = tot_loads_cmp, 1622 .entry = tot_loads_entry, 1623 .width = 7, 1624}; 1625 1626static struct c2c_header percent_costly_snoop_header[] = { 1627 [DISPLAY_LCL_HITM] = HEADER_BOTH("Lcl", "Hitm"), 1628 [DISPLAY_RMT_HITM] = HEADER_BOTH("Rmt", "Hitm"), 1629 [DISPLAY_TOT_HITM] = HEADER_BOTH("Tot", "Hitm"), 1630 [DISPLAY_SNP_PEER] = HEADER_BOTH("Peer", "Snoop"), 1631}; 1632 1633static struct c2c_dimension dim_percent_costly_snoop = { 1634 .name = "percent_costly_snoop", 1635 .cmp = percent_costly_snoop_cmp, 1636 .entry = percent_costly_snoop_entry, 1637 .color = percent_costly_snoop_color, 1638 .width = 7, 1639}; 1640 1641static struct c2c_dimension dim_percent_rmt_hitm = { 1642 .header = HEADER_SPAN("----- HITM -----", "RmtHitm", 1), 1643 .name = "percent_rmt_hitm", 1644 .cmp = percent_rmt_hitm_cmp, 1645 .entry = percent_rmt_hitm_entry, 1646 .color = percent_rmt_hitm_color, 1647 .width = 7, 1648}; 1649 1650static struct c2c_dimension dim_percent_lcl_hitm = { 1651 .header = HEADER_SPAN_LOW("LclHitm"), 1652 .name = "percent_lcl_hitm", 1653 .cmp = percent_lcl_hitm_cmp, 1654 .entry = percent_lcl_hitm_entry, 1655 .color = percent_lcl_hitm_color, 1656 .width = 7, 1657}; 1658 1659static struct c2c_dimension dim_percent_rmt_peer = { 1660 .header = HEADER_SPAN("-- Peer Snoop --", "Rmt", 1), 1661 .name = "percent_rmt_peer", 1662 .cmp = percent_rmt_peer_cmp, 1663 .entry = percent_rmt_peer_entry, 1664 .color = percent_rmt_peer_color, 1665 .width = 7, 1666}; 1667 1668static struct c2c_dimension dim_percent_lcl_peer = { 1669 .header = HEADER_SPAN_LOW("Lcl"), 1670 .name = "percent_lcl_peer", 1671 .cmp = percent_lcl_peer_cmp, 1672 .entry = percent_lcl_peer_entry, 1673 .color = percent_lcl_peer_color, 1674 .width = 7, 1675}; 1676 1677static struct c2c_dimension dim_percent_stores_l1hit = { 1678 .header = HEADER_SPAN("------- Store Refs ------", "L1 Hit", 2), 1679 .name = "percent_stores_l1hit", 1680 .cmp = percent_stores_l1hit_cmp, 1681 .entry = percent_stores_l1hit_entry, 1682 .color = percent_stores_l1hit_color, 1683 .width = 7, 1684}; 1685 1686static struct c2c_dimension dim_percent_stores_l1miss = { 1687 .header = HEADER_SPAN_LOW("L1 Miss"), 1688 .name = "percent_stores_l1miss", 1689 .cmp = percent_stores_l1miss_cmp, 1690 .entry = percent_stores_l1miss_entry, 1691 .color = percent_stores_l1miss_color, 1692 .width = 7, 1693}; 1694 1695static struct c2c_dimension dim_percent_stores_na = { 1696 .header = HEADER_SPAN_LOW("N/A"), 1697 .name = "percent_stores_na", 1698 .cmp = percent_stores_na_cmp, 1699 .entry = percent_stores_na_entry, 1700 .color = percent_stores_na_color, 1701 .width = 7, 1702}; 1703 1704static struct c2c_dimension dim_dram_lcl = { 1705 .header = HEADER_SPAN("--- Load Dram ----", "Lcl", 1), 1706 .name = "dram_lcl", 1707 .cmp = lcl_dram_cmp, 1708 .entry = lcl_dram_entry, 1709 .width = 8, 1710}; 1711 1712static struct c2c_dimension dim_dram_rmt = { 1713 .header = HEADER_SPAN_LOW("Rmt"), 1714 .name = "dram_rmt", 1715 .cmp = rmt_dram_cmp, 1716 .entry = rmt_dram_entry, 1717 .width = 8, 1718}; 1719 1720static struct c2c_dimension dim_pid = { 1721 .header = HEADER_LOW("Pid"), 1722 .name = "pid", 1723 .cmp = pid_cmp, 1724 .entry = pid_entry, 1725 .width = 7, 1726}; 1727 1728static struct c2c_dimension dim_tid = { 1729 .header = HEADER_LOW("Tid"), 1730 .name = "tid", 1731 .se = &sort_thread, 1732}; 1733 1734static struct c2c_dimension dim_symbol = { 1735 .name = "symbol", 1736 .se = &sort_sym, 1737}; 1738 1739static struct c2c_dimension dim_dso = { 1740 .header = HEADER_BOTH("Shared", "Object"), 1741 .name = "dso", 1742 .se = &sort_dso, 1743}; 1744 1745static struct c2c_dimension dim_node = { 1746 .name = "node", 1747 .cmp = empty_cmp, 1748 .entry = node_entry, 1749 .width = 4, 1750}; 1751 1752static struct c2c_dimension dim_mean_rmt = { 1753 .header = HEADER_SPAN("---------- cycles ----------", "rmt hitm", 2), 1754 .name = "mean_rmt", 1755 .cmp = empty_cmp, 1756 .entry = mean_rmt_entry, 1757 .width = 8, 1758}; 1759 1760static struct c2c_dimension dim_mean_lcl = { 1761 .header = HEADER_SPAN_LOW("lcl hitm"), 1762 .name = "mean_lcl", 1763 .cmp = empty_cmp, 1764 .entry = mean_lcl_entry, 1765 .width = 8, 1766}; 1767 1768static struct c2c_dimension dim_mean_load = { 1769 .header = HEADER_SPAN_LOW("load"), 1770 .name = "mean_load", 1771 .cmp = empty_cmp, 1772 .entry = mean_load_entry, 1773 .width = 8, 1774}; 1775 1776static struct c2c_dimension dim_mean_rmt_peer = { 1777 .header = HEADER_SPAN("---------- cycles ----------", "rmt peer", 2), 1778 .name = "mean_rmt_peer", 1779 .cmp = empty_cmp, 1780 .entry = mean_rmt_peer_entry, 1781 .width = 8, 1782}; 1783 1784static struct c2c_dimension dim_mean_lcl_peer = { 1785 .header = HEADER_SPAN_LOW("lcl peer"), 1786 .name = "mean_lcl_peer", 1787 .cmp = empty_cmp, 1788 .entry = mean_lcl_peer_entry, 1789 .width = 8, 1790}; 1791 1792static struct c2c_dimension dim_cpucnt = { 1793 .header = HEADER_BOTH("cpu", "cnt"), 1794 .name = "cpucnt", 1795 .cmp = empty_cmp, 1796 .entry = cpucnt_entry, 1797 .width = 8, 1798}; 1799 1800static struct c2c_dimension dim_srcline = { 1801 .name = "cl_srcline", 1802 .se = &sort_srcline, 1803}; 1804 1805static struct c2c_dimension dim_dcacheline_idx = { 1806 .header = HEADER_LOW("Index"), 1807 .name = "cl_idx", 1808 .cmp = empty_cmp, 1809 .entry = cl_idx_entry, 1810 .width = 5, 1811}; 1812 1813static struct c2c_dimension dim_dcacheline_num = { 1814 .header = HEADER_LOW("Num"), 1815 .name = "cl_num", 1816 .cmp = empty_cmp, 1817 .entry = cl_idx_entry, 1818 .width = 5, 1819}; 1820 1821static struct c2c_dimension dim_dcacheline_num_empty = { 1822 .header = HEADER_LOW("Num"), 1823 .name = "cl_num_empty", 1824 .cmp = empty_cmp, 1825 .entry = cl_idx_empty_entry, 1826 .width = 5, 1827}; 1828 1829static struct c2c_dimension *dimensions[] = { 1830 &dim_dcacheline, 1831 &dim_dcacheline_node, 1832 &dim_dcacheline_count, 1833 &dim_offset, 1834 &dim_offset_node, 1835 &dim_iaddr, 1836 &dim_tot_hitm, 1837 &dim_lcl_hitm, 1838 &dim_rmt_hitm, 1839 &dim_tot_peer, 1840 &dim_lcl_peer, 1841 &dim_rmt_peer, 1842 &dim_cl_lcl_hitm, 1843 &dim_cl_rmt_hitm, 1844 &dim_cl_lcl_peer, 1845 &dim_cl_rmt_peer, 1846 &dim_tot_stores, 1847 &dim_stores_l1hit, 1848 &dim_stores_l1miss, 1849 &dim_stores_na, 1850 &dim_cl_stores_l1hit, 1851 &dim_cl_stores_l1miss, 1852 &dim_cl_stores_na, 1853 &dim_ld_fbhit, 1854 &dim_ld_l1hit, 1855 &dim_ld_l2hit, 1856 &dim_ld_llchit, 1857 &dim_ld_rmthit, 1858 &dim_tot_recs, 1859 &dim_tot_loads, 1860 &dim_percent_costly_snoop, 1861 &dim_percent_rmt_hitm, 1862 &dim_percent_lcl_hitm, 1863 &dim_percent_rmt_peer, 1864 &dim_percent_lcl_peer, 1865 &dim_percent_stores_l1hit, 1866 &dim_percent_stores_l1miss, 1867 &dim_percent_stores_na, 1868 &dim_dram_lcl, 1869 &dim_dram_rmt, 1870 &dim_pid, 1871 &dim_tid, 1872 &dim_symbol, 1873 &dim_dso, 1874 &dim_node, 1875 &dim_mean_rmt, 1876 &dim_mean_lcl, 1877 &dim_mean_rmt_peer, 1878 &dim_mean_lcl_peer, 1879 &dim_mean_load, 1880 &dim_cpucnt, 1881 &dim_srcline, 1882 &dim_dcacheline_idx, 1883 &dim_dcacheline_num, 1884 &dim_dcacheline_num_empty, 1885 NULL, 1886}; 1887 1888static void fmt_free(struct perf_hpp_fmt *fmt) 1889{ 1890 struct c2c_fmt *c2c_fmt; 1891 1892 c2c_fmt = container_of(fmt, struct c2c_fmt, fmt); 1893 free(c2c_fmt); 1894} 1895 1896static bool fmt_equal(struct perf_hpp_fmt *a, struct perf_hpp_fmt *b) 1897{ 1898 struct c2c_fmt *c2c_a = container_of(a, struct c2c_fmt, fmt); 1899 struct c2c_fmt *c2c_b = container_of(b, struct c2c_fmt, fmt); 1900 1901 return c2c_a->dim == c2c_b->dim; 1902} 1903 1904static struct c2c_dimension *get_dimension(const char *name) 1905{ 1906 unsigned int i; 1907 1908 for (i = 0; dimensions[i]; i++) { 1909 struct c2c_dimension *dim = dimensions[i]; 1910 1911 if (!strcmp(dim->name, name)) 1912 return dim; 1913 } 1914 1915 return NULL; 1916} 1917 1918static int c2c_se_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, 1919 struct hist_entry *he) 1920{ 1921 struct c2c_fmt *c2c_fmt = container_of(fmt, struct c2c_fmt, fmt); 1922 struct c2c_dimension *dim = c2c_fmt->dim; 1923 size_t len = fmt->user_len; 1924 1925 if (!len) { 1926 len = hists__col_len(he->hists, dim->se->se_width_idx); 1927 1928 if (dim == &dim_symbol || dim == &dim_srcline) 1929 len = symbol_width(he->hists, dim->se); 1930 } 1931 1932 return dim->se->se_snprintf(he, hpp->buf, hpp->size, len); 1933} 1934 1935static int64_t c2c_se_cmp(struct perf_hpp_fmt *fmt, 1936 struct hist_entry *a, struct hist_entry *b) 1937{ 1938 struct c2c_fmt *c2c_fmt = container_of(fmt, struct c2c_fmt, fmt); 1939 struct c2c_dimension *dim = c2c_fmt->dim; 1940 1941 return dim->se->se_cmp(a, b); 1942} 1943 1944static int64_t c2c_se_collapse(struct perf_hpp_fmt *fmt, 1945 struct hist_entry *a, struct hist_entry *b) 1946{ 1947 struct c2c_fmt *c2c_fmt = container_of(fmt, struct c2c_fmt, fmt); 1948 struct c2c_dimension *dim = c2c_fmt->dim; 1949 int64_t (*collapse_fn)(struct hist_entry *, struct hist_entry *); 1950 1951 collapse_fn = dim->se->se_collapse ?: dim->se->se_cmp; 1952 return collapse_fn(a, b); 1953} 1954 1955static struct c2c_fmt *get_format(const char *name) 1956{ 1957 struct c2c_dimension *dim = get_dimension(name); 1958 struct c2c_fmt *c2c_fmt; 1959 struct perf_hpp_fmt *fmt; 1960 1961 if (!dim) 1962 return NULL; 1963 1964 c2c_fmt = zalloc(sizeof(*c2c_fmt)); 1965 if (!c2c_fmt) 1966 return NULL; 1967 1968 c2c_fmt->dim = dim; 1969 1970 fmt = &c2c_fmt->fmt; 1971 INIT_LIST_HEAD(&fmt->list); 1972 INIT_LIST_HEAD(&fmt->sort_list); 1973 1974 fmt->cmp = dim->se ? c2c_se_cmp : dim->cmp; 1975 fmt->sort = dim->se ? c2c_se_cmp : dim->cmp; 1976 fmt->color = dim->se ? NULL : dim->color; 1977 fmt->entry = dim->se ? c2c_se_entry : dim->entry; 1978 fmt->header = c2c_header; 1979 fmt->width = c2c_width; 1980 fmt->collapse = dim->se ? c2c_se_collapse : dim->cmp; 1981 fmt->equal = fmt_equal; 1982 fmt->free = fmt_free; 1983 1984 return c2c_fmt; 1985} 1986 1987static int c2c_hists__init_output(struct perf_hpp_list *hpp_list, char *name) 1988{ 1989 struct c2c_fmt *c2c_fmt = get_format(name); 1990 1991 if (!c2c_fmt) { 1992 reset_dimensions(); 1993 return output_field_add(hpp_list, name); 1994 } 1995 1996 perf_hpp_list__column_register(hpp_list, &c2c_fmt->fmt); 1997 return 0; 1998} 1999 2000static int c2c_hists__init_sort(struct perf_hpp_list *hpp_list, char *name) 2001{ 2002 struct c2c_fmt *c2c_fmt = get_format(name); 2003 struct c2c_dimension *dim; 2004 2005 if (!c2c_fmt) { 2006 reset_dimensions(); 2007 return sort_dimension__add(hpp_list, name, NULL, 0); 2008 } 2009 2010 dim = c2c_fmt->dim; 2011 if (dim == &dim_dso) 2012 hpp_list->dso = 1; 2013 2014 perf_hpp_list__register_sort_field(hpp_list, &c2c_fmt->fmt); 2015 return 0; 2016} 2017 2018#define PARSE_LIST(_list, _fn) \ 2019 do { \ 2020 char *tmp, *tok; \ 2021 ret = 0; \ 2022 \ 2023 if (!_list) \ 2024 break; \ 2025 \ 2026 for (tok = strtok_r((char *)_list, ", ", &tmp); \ 2027 tok; tok = strtok_r(NULL, ", ", &tmp)) { \ 2028 ret = _fn(hpp_list, tok); \ 2029 if (ret == -EINVAL) { \ 2030 pr_err("Invalid --fields key: `%s'", tok); \ 2031 break; \ 2032 } else if (ret == -ESRCH) { \ 2033 pr_err("Unknown --fields key: `%s'", tok); \ 2034 break; \ 2035 } \ 2036 } \ 2037 } while (0) 2038 2039static int hpp_list__parse(struct perf_hpp_list *hpp_list, 2040 const char *output_, 2041 const char *sort_) 2042{ 2043 char *output = output_ ? strdup(output_) : NULL; 2044 char *sort = sort_ ? strdup(sort_) : NULL; 2045 int ret; 2046 2047 PARSE_LIST(output, c2c_hists__init_output); 2048 PARSE_LIST(sort, c2c_hists__init_sort); 2049 2050 /* copy sort keys to output fields */ 2051 perf_hpp__setup_output_field(hpp_list); 2052 2053 /* 2054 * We don't need other sorting keys other than those 2055 * we already specified. It also really slows down 2056 * the processing a lot with big number of output 2057 * fields, so switching this off for c2c. 2058 */ 2059 2060#if 0 2061 /* and then copy output fields to sort keys */ 2062 perf_hpp__append_sort_keys(&hists->list); 2063#endif 2064 2065 free(output); 2066 free(sort); 2067 return ret; 2068} 2069 2070static int c2c_hists__init(struct c2c_hists *hists, 2071 const char *sort, 2072 int nr_header_lines) 2073{ 2074 __hists__init(&hists->hists, &hists->list); 2075 2076 /* 2077 * Initialize only with sort fields, we need to resort 2078 * later anyway, and that's where we add output fields 2079 * as well. 2080 */ 2081 perf_hpp_list__init(&hists->list); 2082 2083 /* Overload number of header lines.*/ 2084 hists->list.nr_header_lines = nr_header_lines; 2085 2086 return hpp_list__parse(&hists->list, NULL, sort); 2087} 2088 2089static int c2c_hists__reinit(struct c2c_hists *c2c_hists, 2090 const char *output, 2091 const char *sort) 2092{ 2093 perf_hpp__reset_output_field(&c2c_hists->list); 2094 return hpp_list__parse(&c2c_hists->list, output, sort); 2095} 2096 2097#define DISPLAY_LINE_LIMIT 0.001 2098 2099static u8 filter_display(u32 val, u32 sum) 2100{ 2101 if (sum == 0 || ((double)val / sum) < DISPLAY_LINE_LIMIT) 2102 return HIST_FILTER__C2C; 2103 2104 return 0; 2105} 2106 2107static bool he__display(struct hist_entry *he, struct c2c_stats *stats) 2108{ 2109 struct c2c_hist_entry *c2c_he; 2110 2111 if (c2c.show_all) 2112 return true; 2113 2114 c2c_he = container_of(he, struct c2c_hist_entry, he); 2115 2116 switch (c2c.display) { 2117 case DISPLAY_LCL_HITM: 2118 he->filtered = filter_display(c2c_he->stats.lcl_hitm, 2119 stats->lcl_hitm); 2120 break; 2121 case DISPLAY_RMT_HITM: 2122 he->filtered = filter_display(c2c_he->stats.rmt_hitm, 2123 stats->rmt_hitm); 2124 break; 2125 case DISPLAY_TOT_HITM: 2126 he->filtered = filter_display(c2c_he->stats.tot_hitm, 2127 stats->tot_hitm); 2128 break; 2129 case DISPLAY_SNP_PEER: 2130 he->filtered = filter_display(c2c_he->stats.tot_peer, 2131 stats->tot_peer); 2132 break; 2133 default: 2134 break; 2135 } 2136 2137 return he->filtered == 0; 2138} 2139 2140static inline bool is_valid_hist_entry(struct hist_entry *he) 2141{ 2142 struct c2c_hist_entry *c2c_he; 2143 bool has_record = false; 2144 2145 c2c_he = container_of(he, struct c2c_hist_entry, he); 2146 2147 /* It's a valid entry if contains stores */ 2148 if (c2c_he->stats.store) 2149 return true; 2150 2151 switch (c2c.display) { 2152 case DISPLAY_LCL_HITM: 2153 has_record = !!c2c_he->stats.lcl_hitm; 2154 break; 2155 case DISPLAY_RMT_HITM: 2156 has_record = !!c2c_he->stats.rmt_hitm; 2157 break; 2158 case DISPLAY_TOT_HITM: 2159 has_record = !!c2c_he->stats.tot_hitm; 2160 break; 2161 case DISPLAY_SNP_PEER: 2162 has_record = !!c2c_he->stats.tot_peer; 2163 default: 2164 break; 2165 } 2166 2167 return has_record; 2168} 2169 2170static void set_node_width(struct c2c_hist_entry *c2c_he, int len) 2171{ 2172 struct c2c_dimension *dim; 2173 2174 dim = &c2c.hists == c2c_he->hists ? 2175 &dim_dcacheline_node : &dim_offset_node; 2176 2177 if (len > dim->width) 2178 dim->width = len; 2179} 2180 2181static int set_nodestr(struct c2c_hist_entry *c2c_he) 2182{ 2183 char buf[30]; 2184 int len; 2185 2186 if (c2c_he->nodestr) 2187 return 0; 2188 2189 if (!bitmap_empty(c2c_he->nodeset, c2c.nodes_cnt)) { 2190 len = bitmap_scnprintf(c2c_he->nodeset, c2c.nodes_cnt, 2191 buf, sizeof(buf)); 2192 } else { 2193 len = scnprintf(buf, sizeof(buf), "N/A"); 2194 } 2195 2196 set_node_width(c2c_he, len); 2197 c2c_he->nodestr = strdup(buf); 2198 return c2c_he->nodestr ? 0 : -ENOMEM; 2199} 2200 2201static void calc_width(struct c2c_hist_entry *c2c_he) 2202{ 2203 struct c2c_hists *c2c_hists; 2204 2205 c2c_hists = container_of(c2c_he->he.hists, struct c2c_hists, hists); 2206 hists__calc_col_len(&c2c_hists->hists, &c2c_he->he); 2207 set_nodestr(c2c_he); 2208} 2209 2210static int filter_cb(struct hist_entry *he, void *arg __maybe_unused) 2211{ 2212 struct c2c_hist_entry *c2c_he; 2213 2214 c2c_he = container_of(he, struct c2c_hist_entry, he); 2215 2216 if (c2c.show_src && !he->srcline) 2217 he->srcline = hist_entry__srcline(he); 2218 2219 calc_width(c2c_he); 2220 2221 if (!is_valid_hist_entry(he)) 2222 he->filtered = HIST_FILTER__C2C; 2223 2224 return 0; 2225} 2226 2227static int resort_cl_cb(struct hist_entry *he, void *arg __maybe_unused) 2228{ 2229 struct c2c_hist_entry *c2c_he; 2230 struct c2c_hists *c2c_hists; 2231 bool display = he__display(he, &c2c.shared_clines_stats); 2232 2233 c2c_he = container_of(he, struct c2c_hist_entry, he); 2234 c2c_hists = c2c_he->hists; 2235 2236 if (display && c2c_hists) { 2237 static unsigned int idx; 2238 2239 c2c_he->cacheline_idx = idx++; 2240 calc_width(c2c_he); 2241 2242 c2c_hists__reinit(c2c_hists, c2c.cl_output, c2c.cl_resort); 2243 2244 hists__collapse_resort(&c2c_hists->hists, NULL); 2245 hists__output_resort_cb(&c2c_hists->hists, NULL, filter_cb); 2246 } 2247 2248 return 0; 2249} 2250 2251static struct c2c_header header_node_0 = HEADER_LOW("Node"); 2252static struct c2c_header header_node_1_hitms_stores = 2253 HEADER_LOW("Node{cpus %hitms %stores}"); 2254static struct c2c_header header_node_1_peers_stores = 2255 HEADER_LOW("Node{cpus %peers %stores}"); 2256static struct c2c_header header_node_2 = HEADER_LOW("Node{cpu list}"); 2257 2258static void setup_nodes_header(void) 2259{ 2260 switch (c2c.node_info) { 2261 case 0: 2262 dim_node.header = header_node_0; 2263 break; 2264 case 1: 2265 if (c2c.display == DISPLAY_SNP_PEER) 2266 dim_node.header = header_node_1_peers_stores; 2267 else 2268 dim_node.header = header_node_1_hitms_stores; 2269 break; 2270 case 2: 2271 dim_node.header = header_node_2; 2272 break; 2273 default: 2274 break; 2275 } 2276 2277 return; 2278} 2279 2280static int setup_nodes(struct perf_session *session) 2281{ 2282 struct numa_node *n; 2283 unsigned long **nodes; 2284 int node, idx; 2285 struct perf_cpu cpu; 2286 int *cpu2node; 2287 2288 if (c2c.node_info > 2) 2289 c2c.node_info = 2; 2290 2291 c2c.nodes_cnt = session->header.env.nr_numa_nodes; 2292 c2c.cpus_cnt = session->header.env.nr_cpus_avail; 2293 2294 n = session->header.env.numa_nodes; 2295 if (!n) 2296 return -EINVAL; 2297 2298 nodes = zalloc(sizeof(unsigned long *) * c2c.nodes_cnt); 2299 if (!nodes) 2300 return -ENOMEM; 2301 2302 c2c.nodes = nodes; 2303 2304 cpu2node = zalloc(sizeof(int) * c2c.cpus_cnt); 2305 if (!cpu2node) 2306 return -ENOMEM; 2307 2308 for (idx = 0; idx < c2c.cpus_cnt; idx++) 2309 cpu2node[idx] = -1; 2310 2311 c2c.cpu2node = cpu2node; 2312 2313 for (node = 0; node < c2c.nodes_cnt; node++) { 2314 struct perf_cpu_map *map = n[node].map; 2315 unsigned long *set; 2316 2317 set = bitmap_zalloc(c2c.cpus_cnt); 2318 if (!set) 2319 return -ENOMEM; 2320 2321 nodes[node] = set; 2322 2323 perf_cpu_map__for_each_cpu_skip_any(cpu, idx, map) { 2324 __set_bit(cpu.cpu, set); 2325 2326 if (WARN_ONCE(cpu2node[cpu.cpu] != -1, "node/cpu topology bug")) 2327 return -EINVAL; 2328 2329 cpu2node[cpu.cpu] = node; 2330 } 2331 } 2332 2333 setup_nodes_header(); 2334 return 0; 2335} 2336 2337#define HAS_HITMS(__h) ((__h)->stats.lcl_hitm || (__h)->stats.rmt_hitm) 2338#define HAS_PEER(__h) ((__h)->stats.lcl_peer || (__h)->stats.rmt_peer) 2339 2340static int resort_shared_cl_cb(struct hist_entry *he, void *arg __maybe_unused) 2341{ 2342 struct c2c_hist_entry *c2c_he; 2343 c2c_he = container_of(he, struct c2c_hist_entry, he); 2344 2345 if (HAS_HITMS(c2c_he) || HAS_PEER(c2c_he)) { 2346 c2c.shared_clines++; 2347 c2c_add_stats(&c2c.shared_clines_stats, &c2c_he->stats); 2348 } 2349 2350 return 0; 2351} 2352 2353static int hists__iterate_cb(struct hists *hists, hists__resort_cb_t cb) 2354{ 2355 struct rb_node *next = rb_first_cached(&hists->entries); 2356 int ret = 0; 2357 2358 while (next) { 2359 struct hist_entry *he; 2360 2361 he = rb_entry(next, struct hist_entry, rb_node); 2362 ret = cb(he, NULL); 2363 if (ret) 2364 break; 2365 next = rb_next(&he->rb_node); 2366 } 2367 2368 return ret; 2369} 2370 2371static void print_c2c__display_stats(FILE *out) 2372{ 2373 int llc_misses; 2374 struct c2c_stats *stats = &c2c.hists.stats; 2375 2376 llc_misses = get_load_llc_misses(stats); 2377 2378 fprintf(out, "=================================================\n"); 2379 fprintf(out, " Trace Event Information \n"); 2380 fprintf(out, "=================================================\n"); 2381 fprintf(out, " Total records : %10d\n", stats->nr_entries); 2382 fprintf(out, " Locked Load/Store Operations : %10d\n", stats->locks); 2383 fprintf(out, " Load Operations : %10d\n", stats->load); 2384 fprintf(out, " Loads - uncacheable : %10d\n", stats->ld_uncache); 2385 fprintf(out, " Loads - IO : %10d\n", stats->ld_io); 2386 fprintf(out, " Loads - Miss : %10d\n", stats->ld_miss); 2387 fprintf(out, " Loads - no mapping : %10d\n", stats->ld_noadrs); 2388 fprintf(out, " Load Fill Buffer Hit : %10d\n", stats->ld_fbhit); 2389 fprintf(out, " Load L1D hit : %10d\n", stats->ld_l1hit); 2390 fprintf(out, " Load L2D hit : %10d\n", stats->ld_l2hit); 2391 fprintf(out, " Load LLC hit : %10d\n", stats->ld_llchit + stats->lcl_hitm); 2392 fprintf(out, " Load Local HITM : %10d\n", stats->lcl_hitm); 2393 fprintf(out, " Load Remote HITM : %10d\n", stats->rmt_hitm); 2394 fprintf(out, " Load Remote HIT : %10d\n", stats->rmt_hit); 2395 fprintf(out, " Load Local DRAM : %10d\n", stats->lcl_dram); 2396 fprintf(out, " Load Remote DRAM : %10d\n", stats->rmt_dram); 2397 fprintf(out, " Load MESI State Exclusive : %10d\n", stats->ld_excl); 2398 fprintf(out, " Load MESI State Shared : %10d\n", stats->ld_shared); 2399 fprintf(out, " Load LLC Misses : %10d\n", llc_misses); 2400 fprintf(out, " Load access blocked by data : %10d\n", stats->blk_data); 2401 fprintf(out, " Load access blocked by address : %10d\n", stats->blk_addr); 2402 fprintf(out, " Load HIT Local Peer : %10d\n", stats->lcl_peer); 2403 fprintf(out, " Load HIT Remote Peer : %10d\n", stats->rmt_peer); 2404 fprintf(out, " LLC Misses to Local DRAM : %10.1f%%\n", ((double)stats->lcl_dram/(double)llc_misses) * 100.); 2405 fprintf(out, " LLC Misses to Remote DRAM : %10.1f%%\n", ((double)stats->rmt_dram/(double)llc_misses) * 100.); 2406 fprintf(out, " LLC Misses to Remote cache (HIT) : %10.1f%%\n", ((double)stats->rmt_hit /(double)llc_misses) * 100.); 2407 fprintf(out, " LLC Misses to Remote cache (HITM) : %10.1f%%\n", ((double)stats->rmt_hitm/(double)llc_misses) * 100.); 2408 fprintf(out, " Store Operations : %10d\n", stats->store); 2409 fprintf(out, " Store - uncacheable : %10d\n", stats->st_uncache); 2410 fprintf(out, " Store - no mapping : %10d\n", stats->st_noadrs); 2411 fprintf(out, " Store L1D Hit : %10d\n", stats->st_l1hit); 2412 fprintf(out, " Store L1D Miss : %10d\n", stats->st_l1miss); 2413 fprintf(out, " Store No available memory level : %10d\n", stats->st_na); 2414 fprintf(out, " No Page Map Rejects : %10d\n", stats->nomap); 2415 fprintf(out, " Unable to parse data source : %10d\n", stats->noparse); 2416} 2417 2418static void print_shared_cacheline_info(FILE *out) 2419{ 2420 struct c2c_stats *stats = &c2c.shared_clines_stats; 2421 int hitm_cnt = stats->lcl_hitm + stats->rmt_hitm; 2422 2423 fprintf(out, "=================================================\n"); 2424 fprintf(out, " Global Shared Cache Line Event Information \n"); 2425 fprintf(out, "=================================================\n"); 2426 fprintf(out, " Total Shared Cache Lines : %10d\n", c2c.shared_clines); 2427 fprintf(out, " Load HITs on shared lines : %10d\n", stats->load); 2428 fprintf(out, " Fill Buffer Hits on shared lines : %10d\n", stats->ld_fbhit); 2429 fprintf(out, " L1D hits on shared lines : %10d\n", stats->ld_l1hit); 2430 fprintf(out, " L2D hits on shared lines : %10d\n", stats->ld_l2hit); 2431 fprintf(out, " LLC hits on shared lines : %10d\n", stats->ld_llchit + stats->lcl_hitm); 2432 fprintf(out, " Load hits on peer cache or nodes : %10d\n", stats->lcl_peer + stats->rmt_peer); 2433 fprintf(out, " Locked Access on shared lines : %10d\n", stats->locks); 2434 fprintf(out, " Blocked Access on shared lines : %10d\n", stats->blk_data + stats->blk_addr); 2435 fprintf(out, " Store HITs on shared lines : %10d\n", stats->store); 2436 fprintf(out, " Store L1D hits on shared lines : %10d\n", stats->st_l1hit); 2437 fprintf(out, " Store No available memory level : %10d\n", stats->st_na); 2438 fprintf(out, " Total Merged records : %10d\n", hitm_cnt + stats->store); 2439} 2440 2441static void print_cacheline(struct c2c_hists *c2c_hists, 2442 struct hist_entry *he_cl, 2443 struct perf_hpp_list *hpp_list, 2444 FILE *out) 2445{ 2446 char bf[1000]; 2447 struct perf_hpp hpp = { 2448 .buf = bf, 2449 .size = 1000, 2450 }; 2451 static bool once; 2452 2453 if (!once) { 2454 hists__fprintf_headers(&c2c_hists->hists, out); 2455 once = true; 2456 } else { 2457 fprintf(out, "\n"); 2458 } 2459 2460 fprintf(out, " ----------------------------------------------------------------------\n"); 2461 __hist_entry__snprintf(he_cl, &hpp, hpp_list); 2462 fprintf(out, "%s\n", bf); 2463 fprintf(out, " ----------------------------------------------------------------------\n"); 2464 2465 hists__fprintf(&c2c_hists->hists, false, 0, 0, 0, out, false); 2466} 2467 2468static void print_pareto(FILE *out) 2469{ 2470 struct perf_hpp_list hpp_list; 2471 struct rb_node *nd; 2472 int ret; 2473 const char *cl_output; 2474 2475 if (c2c.display != DISPLAY_SNP_PEER) 2476 cl_output = "cl_num," 2477 "cl_rmt_hitm," 2478 "cl_lcl_hitm," 2479 "cl_stores_l1hit," 2480 "cl_stores_l1miss," 2481 "cl_stores_na," 2482 "dcacheline"; 2483 else 2484 cl_output = "cl_num," 2485 "cl_rmt_peer," 2486 "cl_lcl_peer," 2487 "cl_stores_l1hit," 2488 "cl_stores_l1miss," 2489 "cl_stores_na," 2490 "dcacheline"; 2491 2492 perf_hpp_list__init(&hpp_list); 2493 ret = hpp_list__parse(&hpp_list, cl_output, NULL); 2494 2495 if (WARN_ONCE(ret, "failed to setup sort entries\n")) 2496 return; 2497 2498 nd = rb_first_cached(&c2c.hists.hists.entries); 2499 2500 for (; nd; nd = rb_next(nd)) { 2501 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node); 2502 struct c2c_hist_entry *c2c_he; 2503 2504 if (he->filtered) 2505 continue; 2506 2507 c2c_he = container_of(he, struct c2c_hist_entry, he); 2508 print_cacheline(c2c_he->hists, he, &hpp_list, out); 2509 } 2510} 2511 2512static void print_c2c_info(FILE *out, struct perf_session *session) 2513{ 2514 struct evlist *evlist = session->evlist; 2515 struct evsel *evsel; 2516 bool first = true; 2517 2518 fprintf(out, "=================================================\n"); 2519 fprintf(out, " c2c details \n"); 2520 fprintf(out, "=================================================\n"); 2521 2522 evlist__for_each_entry(evlist, evsel) { 2523 fprintf(out, "%-36s: %s\n", first ? " Events" : "", evsel__name(evsel)); 2524 first = false; 2525 } 2526 fprintf(out, " Cachelines sort on : %s\n", 2527 display_str[c2c.display]); 2528 fprintf(out, " Cacheline data grouping : %s\n", c2c.cl_sort); 2529} 2530 2531static void perf_c2c__hists_fprintf(FILE *out, struct perf_session *session) 2532{ 2533 setup_pager(); 2534 2535 print_c2c__display_stats(out); 2536 fprintf(out, "\n"); 2537 print_shared_cacheline_info(out); 2538 fprintf(out, "\n"); 2539 print_c2c_info(out, session); 2540 2541 if (c2c.stats_only) 2542 return; 2543 2544 fprintf(out, "\n"); 2545 fprintf(out, "=================================================\n"); 2546 fprintf(out, " Shared Data Cache Line Table \n"); 2547 fprintf(out, "=================================================\n"); 2548 fprintf(out, "#\n"); 2549 2550 hists__fprintf(&c2c.hists.hists, true, 0, 0, 0, stdout, true); 2551 2552 fprintf(out, "\n"); 2553 fprintf(out, "=================================================\n"); 2554 fprintf(out, " Shared Cache Line Distribution Pareto \n"); 2555 fprintf(out, "=================================================\n"); 2556 fprintf(out, "#\n"); 2557 2558 print_pareto(out); 2559} 2560 2561#ifdef HAVE_SLANG_SUPPORT 2562static void c2c_browser__update_nr_entries(struct hist_browser *hb) 2563{ 2564 u64 nr_entries = 0; 2565 struct rb_node *nd = rb_first_cached(&hb->hists->entries); 2566 2567 while (nd) { 2568 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node); 2569 2570 if (!he->filtered) 2571 nr_entries++; 2572 2573 nd = rb_next(nd); 2574 } 2575 2576 hb->nr_non_filtered_entries = nr_entries; 2577} 2578 2579struct c2c_cacheline_browser { 2580 struct hist_browser hb; 2581 struct hist_entry *he; 2582}; 2583 2584static int 2585perf_c2c_cacheline_browser__title(struct hist_browser *browser, 2586 char *bf, size_t size) 2587{ 2588 struct c2c_cacheline_browser *cl_browser; 2589 struct hist_entry *he; 2590 uint64_t addr = 0; 2591 2592 cl_browser = container_of(browser, struct c2c_cacheline_browser, hb); 2593 he = cl_browser->he; 2594 2595 if (he->mem_info) 2596 addr = cl_address(mem_info__daddr(he->mem_info)->addr, chk_double_cl); 2597 2598 scnprintf(bf, size, "Cacheline 0x%lx", addr); 2599 return 0; 2600} 2601 2602static struct c2c_cacheline_browser* 2603c2c_cacheline_browser__new(struct hists *hists, struct hist_entry *he) 2604{ 2605 struct c2c_cacheline_browser *browser; 2606 2607 browser = zalloc(sizeof(*browser)); 2608 if (browser) { 2609 hist_browser__init(&browser->hb, hists); 2610 browser->hb.c2c_filter = true; 2611 browser->hb.title = perf_c2c_cacheline_browser__title; 2612 browser->he = he; 2613 } 2614 2615 return browser; 2616} 2617 2618static int perf_c2c__browse_cacheline(struct hist_entry *he) 2619{ 2620 struct c2c_hist_entry *c2c_he; 2621 struct c2c_hists *c2c_hists; 2622 struct c2c_cacheline_browser *cl_browser; 2623 struct hist_browser *browser; 2624 int key = -1; 2625 static const char help[] = 2626 " ENTER Toggle callchains (if present) \n" 2627 " n Toggle Node details info \n" 2628 " s Toggle full length of symbol and source line columns \n" 2629 " q Return back to cacheline list \n"; 2630 2631 if (!he) 2632 return 0; 2633 2634 /* Display compact version first. */ 2635 c2c.symbol_full = false; 2636 2637 c2c_he = container_of(he, struct c2c_hist_entry, he); 2638 c2c_hists = c2c_he->hists; 2639 2640 cl_browser = c2c_cacheline_browser__new(&c2c_hists->hists, he); 2641 if (cl_browser == NULL) 2642 return -1; 2643 2644 browser = &cl_browser->hb; 2645 2646 /* reset abort key so that it can get Ctrl-C as a key */ 2647 SLang_reset_tty(); 2648 SLang_init_tty(0, 0, 0); 2649 2650 c2c_browser__update_nr_entries(browser); 2651 2652 while (1) { 2653 key = hist_browser__run(browser, "? - help", true, 0); 2654 2655 switch (key) { 2656 case 's': 2657 c2c.symbol_full = !c2c.symbol_full; 2658 break; 2659 case 'n': 2660 c2c.node_info = (c2c.node_info + 1) % 3; 2661 setup_nodes_header(); 2662 break; 2663 case 'q': 2664 goto out; 2665 case '?': 2666 ui_browser__help_window(&browser->b, help); 2667 break; 2668 default: 2669 break; 2670 } 2671 } 2672 2673out: 2674 free(cl_browser); 2675 return 0; 2676} 2677 2678static int perf_c2c_browser__title(struct hist_browser *browser, 2679 char *bf, size_t size) 2680{ 2681 scnprintf(bf, size, 2682 "Shared Data Cache Line Table " 2683 "(%lu entries, sorted on %s)", 2684 browser->nr_non_filtered_entries, 2685 display_str[c2c.display]); 2686 return 0; 2687} 2688 2689static struct hist_browser* 2690perf_c2c_browser__new(struct hists *hists) 2691{ 2692 struct hist_browser *browser = hist_browser__new(hists); 2693 2694 if (browser) { 2695 browser->title = perf_c2c_browser__title; 2696 browser->c2c_filter = true; 2697 } 2698 2699 return browser; 2700} 2701 2702static int perf_c2c__hists_browse(struct hists *hists) 2703{ 2704 struct hist_browser *browser; 2705 int key = -1; 2706 static const char help[] = 2707 " d Display cacheline details \n" 2708 " ENTER Toggle callchains (if present) \n" 2709 " q Quit \n"; 2710 2711 browser = perf_c2c_browser__new(hists); 2712 if (browser == NULL) 2713 return -1; 2714 2715 /* reset abort key so that it can get Ctrl-C as a key */ 2716 SLang_reset_tty(); 2717 SLang_init_tty(0, 0, 0); 2718 2719 c2c_browser__update_nr_entries(browser); 2720 2721 while (1) { 2722 key = hist_browser__run(browser, "? - help", true, 0); 2723 2724 switch (key) { 2725 case 'q': 2726 goto out; 2727 case 'd': 2728 perf_c2c__browse_cacheline(browser->he_selection); 2729 break; 2730 case '?': 2731 ui_browser__help_window(&browser->b, help); 2732 break; 2733 default: 2734 break; 2735 } 2736 } 2737 2738out: 2739 hist_browser__delete(browser); 2740 return 0; 2741} 2742 2743static void perf_c2c_display(struct perf_session *session) 2744{ 2745 if (use_browser == 0) 2746 perf_c2c__hists_fprintf(stdout, session); 2747 else 2748 perf_c2c__hists_browse(&c2c.hists.hists); 2749} 2750#else 2751static void perf_c2c_display(struct perf_session *session) 2752{ 2753 use_browser = 0; 2754 perf_c2c__hists_fprintf(stdout, session); 2755} 2756#endif /* HAVE_SLANG_SUPPORT */ 2757 2758static char *fill_line(const char *orig, int len) 2759{ 2760 int i, j, olen = strlen(orig); 2761 char *buf; 2762 2763 buf = zalloc(len + 1); 2764 if (!buf) 2765 return NULL; 2766 2767 j = len / 2 - olen / 2; 2768 2769 for (i = 0; i < j - 1; i++) 2770 buf[i] = '-'; 2771 2772 buf[i++] = ' '; 2773 2774 strcpy(buf + i, orig); 2775 2776 i += olen; 2777 2778 buf[i++] = ' '; 2779 2780 for (; i < len; i++) 2781 buf[i] = '-'; 2782 2783 return buf; 2784} 2785 2786static int ui_quirks(void) 2787{ 2788 const char *nodestr = "Data address"; 2789 char *buf; 2790 2791 if (!c2c.use_stdio) { 2792 dim_offset.width = 5; 2793 dim_offset.header = header_offset_tui; 2794 nodestr = chk_double_cl ? "Double-CL" : "CL"; 2795 } 2796 2797 dim_percent_costly_snoop.header = percent_costly_snoop_header[c2c.display]; 2798 2799 /* Fix the zero line for dcacheline column. */ 2800 buf = fill_line(chk_double_cl ? "Double-Cacheline" : "Cacheline", 2801 dim_dcacheline.width + 2802 dim_dcacheline_node.width + 2803 dim_dcacheline_count.width + 4); 2804 if (!buf) 2805 return -ENOMEM; 2806 2807 dim_dcacheline.header.line[0].text = buf; 2808 2809 /* Fix the zero line for offset column. */ 2810 buf = fill_line(nodestr, dim_offset.width + 2811 dim_offset_node.width + 2812 dim_dcacheline_count.width + 4); 2813 if (!buf) 2814 return -ENOMEM; 2815 2816 dim_offset.header.line[0].text = buf; 2817 2818 return 0; 2819} 2820 2821#define CALLCHAIN_DEFAULT_OPT "graph,0.5,caller,function,percent" 2822 2823const char callchain_help[] = "Display call graph (stack chain/backtrace):\n\n" 2824 CALLCHAIN_REPORT_HELP 2825 "\n\t\t\t\tDefault: " CALLCHAIN_DEFAULT_OPT; 2826 2827static int 2828parse_callchain_opt(const struct option *opt, const char *arg, int unset) 2829{ 2830 struct callchain_param *callchain = opt->value; 2831 2832 callchain->enabled = !unset; 2833 /* 2834 * --no-call-graph 2835 */ 2836 if (unset) { 2837 symbol_conf.use_callchain = false; 2838 callchain->mode = CHAIN_NONE; 2839 return 0; 2840 } 2841 2842 return parse_callchain_report_opt(arg); 2843} 2844 2845static int setup_callchain(struct evlist *evlist) 2846{ 2847 u64 sample_type = evlist__combined_sample_type(evlist); 2848 enum perf_call_graph_mode mode = CALLCHAIN_NONE; 2849 2850 if ((sample_type & PERF_SAMPLE_REGS_USER) && 2851 (sample_type & PERF_SAMPLE_STACK_USER)) { 2852 mode = CALLCHAIN_DWARF; 2853 dwarf_callchain_users = true; 2854 } else if (sample_type & PERF_SAMPLE_BRANCH_STACK) 2855 mode = CALLCHAIN_LBR; 2856 else if (sample_type & PERF_SAMPLE_CALLCHAIN) 2857 mode = CALLCHAIN_FP; 2858 2859 if (!callchain_param.enabled && 2860 callchain_param.mode != CHAIN_NONE && 2861 mode != CALLCHAIN_NONE) { 2862 symbol_conf.use_callchain = true; 2863 if (callchain_register_param(&callchain_param) < 0) { 2864 ui__error("Can't register callchain params.\n"); 2865 return -EINVAL; 2866 } 2867 } 2868 2869 if (c2c.stitch_lbr && (mode != CALLCHAIN_LBR)) { 2870 ui__warning("Can't find LBR callchain. Switch off --stitch-lbr.\n" 2871 "Please apply --call-graph lbr when recording.\n"); 2872 c2c.stitch_lbr = false; 2873 } 2874 2875 callchain_param.record_mode = mode; 2876 callchain_param.min_percent = 0; 2877 return 0; 2878} 2879 2880static int setup_display(const char *str) 2881{ 2882 const char *display = str; 2883 2884 if (!strcmp(display, "tot")) 2885 c2c.display = DISPLAY_TOT_HITM; 2886 else if (!strcmp(display, "rmt")) 2887 c2c.display = DISPLAY_RMT_HITM; 2888 else if (!strcmp(display, "lcl")) 2889 c2c.display = DISPLAY_LCL_HITM; 2890 else if (!strcmp(display, "peer")) 2891 c2c.display = DISPLAY_SNP_PEER; 2892 else { 2893 pr_err("failed: unknown display type: %s\n", str); 2894 return -1; 2895 } 2896 2897 return 0; 2898} 2899 2900#define for_each_token(__tok, __buf, __sep, __tmp) \ 2901 for (__tok = strtok_r(__buf, __sep, &__tmp); __tok; \ 2902 __tok = strtok_r(NULL, __sep, &__tmp)) 2903 2904static int build_cl_output(char *cl_sort, bool no_source) 2905{ 2906 char *tok, *tmp, *buf = strdup(cl_sort); 2907 bool add_pid = false; 2908 bool add_tid = false; 2909 bool add_iaddr = false; 2910 bool add_sym = false; 2911 bool add_dso = false; 2912 bool add_src = false; 2913 int ret = 0; 2914 2915 if (!buf) 2916 return -ENOMEM; 2917 2918 for_each_token(tok, buf, ",", tmp) { 2919 if (!strcmp(tok, "tid")) { 2920 add_tid = true; 2921 } else if (!strcmp(tok, "pid")) { 2922 add_pid = true; 2923 } else if (!strcmp(tok, "iaddr")) { 2924 add_iaddr = true; 2925 add_sym = true; 2926 add_dso = true; 2927 add_src = no_source ? false : true; 2928 } else if (!strcmp(tok, "dso")) { 2929 add_dso = true; 2930 } else if (strcmp(tok, "offset")) { 2931 pr_err("unrecognized sort token: %s\n", tok); 2932 ret = -EINVAL; 2933 goto err; 2934 } 2935 } 2936 2937 if (asprintf(&c2c.cl_output, 2938 "%s%s%s%s%s%s%s%s%s%s%s%s", 2939 c2c.use_stdio ? "cl_num_empty," : "", 2940 c2c.display == DISPLAY_SNP_PEER ? "percent_rmt_peer," 2941 "percent_lcl_peer," : 2942 "percent_rmt_hitm," 2943 "percent_lcl_hitm,", 2944 "percent_stores_l1hit," 2945 "percent_stores_l1miss," 2946 "percent_stores_na," 2947 "offset,offset_node,dcacheline_count,", 2948 add_pid ? "pid," : "", 2949 add_tid ? "tid," : "", 2950 add_iaddr ? "iaddr," : "", 2951 c2c.display == DISPLAY_SNP_PEER ? "mean_rmt_peer," 2952 "mean_lcl_peer," : 2953 "mean_rmt," 2954 "mean_lcl,", 2955 "mean_load," 2956 "tot_recs," 2957 "cpucnt,", 2958 add_sym ? "symbol," : "", 2959 add_dso ? "dso," : "", 2960 add_src ? "cl_srcline," : "", 2961 "node") < 0) { 2962 ret = -ENOMEM; 2963 goto err; 2964 } 2965 2966 c2c.show_src = add_src; 2967err: 2968 free(buf); 2969 return ret; 2970} 2971 2972static int setup_coalesce(const char *coalesce, bool no_source) 2973{ 2974 const char *c = coalesce ?: coalesce_default; 2975 const char *sort_str = NULL; 2976 2977 if (asprintf(&c2c.cl_sort, "offset,%s", c) < 0) 2978 return -ENOMEM; 2979 2980 if (build_cl_output(c2c.cl_sort, no_source)) 2981 return -1; 2982 2983 if (c2c.display == DISPLAY_TOT_HITM) 2984 sort_str = "tot_hitm"; 2985 else if (c2c.display == DISPLAY_RMT_HITM) 2986 sort_str = "rmt_hitm,lcl_hitm"; 2987 else if (c2c.display == DISPLAY_LCL_HITM) 2988 sort_str = "lcl_hitm,rmt_hitm"; 2989 else if (c2c.display == DISPLAY_SNP_PEER) 2990 sort_str = "tot_peer"; 2991 2992 if (asprintf(&c2c.cl_resort, "offset,%s", sort_str) < 0) 2993 return -ENOMEM; 2994 2995 pr_debug("coalesce sort fields: %s\n", c2c.cl_sort); 2996 pr_debug("coalesce resort fields: %s\n", c2c.cl_resort); 2997 pr_debug("coalesce output fields: %s\n", c2c.cl_output); 2998 return 0; 2999} 3000 3001static int perf_c2c__report(int argc, const char **argv) 3002{ 3003 struct itrace_synth_opts itrace_synth_opts = { 3004 .set = true, 3005 .mem = true, /* Only enable memory event */ 3006 .default_no_sample = true, 3007 }; 3008 3009 struct perf_session *session; 3010 struct ui_progress prog; 3011 struct perf_data data = { 3012 .mode = PERF_DATA_MODE_READ, 3013 }; 3014 char callchain_default_opt[] = CALLCHAIN_DEFAULT_OPT; 3015 const char *display = NULL; 3016 const char *coalesce = NULL; 3017 bool no_source = false; 3018 const struct option options[] = { 3019 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name, 3020 "file", "vmlinux pathname"), 3021 OPT_STRING('i', "input", &input_name, "file", 3022 "the input file to process"), 3023 OPT_INCR('N', "node-info", &c2c.node_info, 3024 "show extra node info in report (repeat for more info)"), 3025 OPT_BOOLEAN(0, "stdio", &c2c.use_stdio, "Use the stdio interface"), 3026 OPT_BOOLEAN(0, "stats", &c2c.stats_only, 3027 "Display only statistic tables (implies --stdio)"), 3028 OPT_BOOLEAN(0, "full-symbols", &c2c.symbol_full, 3029 "Display full length of symbols"), 3030 OPT_BOOLEAN(0, "no-source", &no_source, 3031 "Do not display Source Line column"), 3032 OPT_BOOLEAN(0, "show-all", &c2c.show_all, 3033 "Show all captured HITM lines."), 3034 OPT_CALLBACK_DEFAULT('g', "call-graph", &callchain_param, 3035 "print_type,threshold[,print_limit],order,sort_key[,branch],value", 3036 callchain_help, &parse_callchain_opt, 3037 callchain_default_opt), 3038 OPT_STRING('d', "display", &display, "Switch HITM output type", "tot,lcl,rmt,peer"), 3039 OPT_STRING('c', "coalesce", &coalesce, "coalesce fields", 3040 "coalesce fields: pid,tid,iaddr,dso"), 3041 OPT_BOOLEAN('f', "force", &symbol_conf.force, "don't complain, do it"), 3042 OPT_BOOLEAN(0, "stitch-lbr", &c2c.stitch_lbr, 3043 "Enable LBR callgraph stitching approach"), 3044 OPT_BOOLEAN(0, "double-cl", &chk_double_cl, "Detect adjacent cacheline false sharing"), 3045 OPT_PARENT(c2c_options), 3046 OPT_END() 3047 }; 3048 int err = 0; 3049 const char *output_str, *sort_str = NULL; 3050 3051 argc = parse_options(argc, argv, options, report_c2c_usage, 3052 PARSE_OPT_STOP_AT_NON_OPTION); 3053 if (argc) 3054 usage_with_options(report_c2c_usage, options); 3055 3056#ifndef HAVE_SLANG_SUPPORT 3057 c2c.use_stdio = true; 3058#endif 3059 3060 if (c2c.stats_only) 3061 c2c.use_stdio = true; 3062 3063 err = symbol__validate_sym_arguments(); 3064 if (err) 3065 goto out; 3066 3067 if (!input_name || !strlen(input_name)) 3068 input_name = "perf.data"; 3069 3070 data.path = input_name; 3071 data.force = symbol_conf.force; 3072 3073 session = perf_session__new(&data, &c2c.tool); 3074 if (IS_ERR(session)) { 3075 err = PTR_ERR(session); 3076 pr_debug("Error creating perf session\n"); 3077 goto out; 3078 } 3079 3080 /* 3081 * Use the 'tot' as default display type if user doesn't specify it; 3082 * since Arm64 platform doesn't support HITMs flag, use 'peer' as the 3083 * default display type. 3084 */ 3085 if (!display) { 3086 if (!strcmp(perf_env__arch(&session->header.env), "arm64")) 3087 display = "peer"; 3088 else 3089 display = "tot"; 3090 } 3091 3092 err = setup_display(display); 3093 if (err) 3094 goto out_session; 3095 3096 err = setup_coalesce(coalesce, no_source); 3097 if (err) { 3098 pr_debug("Failed to initialize hists\n"); 3099 goto out_session; 3100 } 3101 3102 err = c2c_hists__init(&c2c.hists, "dcacheline", 2); 3103 if (err) { 3104 pr_debug("Failed to initialize hists\n"); 3105 goto out_session; 3106 } 3107 3108 session->itrace_synth_opts = &itrace_synth_opts; 3109 3110 err = setup_nodes(session); 3111 if (err) { 3112 pr_err("Failed setup nodes\n"); 3113 goto out_session; 3114 } 3115 3116 err = mem2node__init(&c2c.mem2node, &session->header.env); 3117 if (err) 3118 goto out_session; 3119 3120 err = setup_callchain(session->evlist); 3121 if (err) 3122 goto out_mem2node; 3123 3124 if (symbol__init(&session->header.env) < 0) 3125 goto out_mem2node; 3126 3127 /* No pipe support at the moment. */ 3128 if (perf_data__is_pipe(session->data)) { 3129 pr_debug("No pipe support at the moment.\n"); 3130 goto out_mem2node; 3131 } 3132 3133 if (c2c.use_stdio) 3134 use_browser = 0; 3135 else 3136 use_browser = 1; 3137 3138 setup_browser(false); 3139 3140 err = perf_session__process_events(session); 3141 if (err) { 3142 pr_err("failed to process sample\n"); 3143 goto out_mem2node; 3144 } 3145 3146 if (c2c.display != DISPLAY_SNP_PEER) 3147 output_str = "cl_idx," 3148 "dcacheline," 3149 "dcacheline_node," 3150 "dcacheline_count," 3151 "percent_costly_snoop," 3152 "tot_hitm,lcl_hitm,rmt_hitm," 3153 "tot_recs," 3154 "tot_loads," 3155 "tot_stores," 3156 "stores_l1hit,stores_l1miss,stores_na," 3157 "ld_fbhit,ld_l1hit,ld_l2hit," 3158 "ld_lclhit,lcl_hitm," 3159 "ld_rmthit,rmt_hitm," 3160 "dram_lcl,dram_rmt"; 3161 else 3162 output_str = "cl_idx," 3163 "dcacheline," 3164 "dcacheline_node," 3165 "dcacheline_count," 3166 "percent_costly_snoop," 3167 "tot_peer,lcl_peer,rmt_peer," 3168 "tot_recs," 3169 "tot_loads," 3170 "tot_stores," 3171 "stores_l1hit,stores_l1miss,stores_na," 3172 "ld_fbhit,ld_l1hit,ld_l2hit," 3173 "ld_lclhit,lcl_hitm," 3174 "ld_rmthit,rmt_hitm," 3175 "dram_lcl,dram_rmt"; 3176 3177 if (c2c.display == DISPLAY_TOT_HITM) 3178 sort_str = "tot_hitm"; 3179 else if (c2c.display == DISPLAY_RMT_HITM) 3180 sort_str = "rmt_hitm"; 3181 else if (c2c.display == DISPLAY_LCL_HITM) 3182 sort_str = "lcl_hitm"; 3183 else if (c2c.display == DISPLAY_SNP_PEER) 3184 sort_str = "tot_peer"; 3185 3186 c2c_hists__reinit(&c2c.hists, output_str, sort_str); 3187 3188 ui_progress__init(&prog, c2c.hists.hists.nr_entries, "Sorting..."); 3189 3190 hists__collapse_resort(&c2c.hists.hists, NULL); 3191 hists__output_resort_cb(&c2c.hists.hists, &prog, resort_shared_cl_cb); 3192 hists__iterate_cb(&c2c.hists.hists, resort_cl_cb); 3193 3194 ui_progress__finish(); 3195 3196 if (ui_quirks()) { 3197 pr_err("failed to setup UI\n"); 3198 goto out_mem2node; 3199 } 3200 3201 perf_c2c_display(session); 3202 3203out_mem2node: 3204 mem2node__exit(&c2c.mem2node); 3205out_session: 3206 perf_session__delete(session); 3207out: 3208 return err; 3209} 3210 3211static int parse_record_events(const struct option *opt, 3212 const char *str, int unset __maybe_unused) 3213{ 3214 bool *event_set = (bool *) opt->value; 3215 struct perf_pmu *pmu; 3216 3217 pmu = perf_mem_events_find_pmu(); 3218 if (!pmu) { 3219 pr_err("failed: there is no PMU that supports perf c2c\n"); 3220 exit(-1); 3221 } 3222 3223 if (!strcmp(str, "list")) { 3224 perf_pmu__mem_events_list(pmu); 3225 exit(0); 3226 } 3227 if (perf_pmu__mem_events_parse(pmu, str)) 3228 exit(-1); 3229 3230 *event_set = true; 3231 return 0; 3232} 3233 3234 3235static const char * const __usage_record[] = { 3236 "perf c2c record [<options>] [<command>]", 3237 "perf c2c record [<options>] -- <command> [<options>]", 3238 NULL 3239}; 3240 3241static const char * const *record_mem_usage = __usage_record; 3242 3243static int perf_c2c__record(int argc, const char **argv) 3244{ 3245 int rec_argc, i = 0, j; 3246 const char **rec_argv; 3247 int ret; 3248 bool all_user = false, all_kernel = false; 3249 bool event_set = false; 3250 struct perf_mem_event *e; 3251 struct perf_pmu *pmu; 3252 struct option options[] = { 3253 OPT_CALLBACK('e', "event", &event_set, "event", 3254 "event selector. Use 'perf c2c record -e list' to list available events", 3255 parse_record_events), 3256 OPT_BOOLEAN('u', "all-user", &all_user, "collect only user level data"), 3257 OPT_BOOLEAN('k', "all-kernel", &all_kernel, "collect only kernel level data"), 3258 OPT_UINTEGER('l', "ldlat", &perf_mem_events__loads_ldlat, "setup mem-loads latency"), 3259 OPT_PARENT(c2c_options), 3260 OPT_END() 3261 }; 3262 3263 pmu = perf_mem_events_find_pmu(); 3264 if (!pmu) { 3265 pr_err("failed: no PMU supports the memory events\n"); 3266 return -1; 3267 } 3268 3269 if (perf_pmu__mem_events_init(pmu)) { 3270 pr_err("failed: memory events not supported\n"); 3271 return -1; 3272 } 3273 3274 argc = parse_options(argc, argv, options, record_mem_usage, 3275 PARSE_OPT_KEEP_UNKNOWN); 3276 3277 /* Max number of arguments multiplied by number of PMUs that can support them. */ 3278 rec_argc = argc + 11 * (perf_pmu__mem_events_num_mem_pmus(pmu) + 1); 3279 3280 rec_argv = calloc(rec_argc + 1, sizeof(char *)); 3281 if (!rec_argv) 3282 return -1; 3283 3284 rec_argv[i++] = "record"; 3285 3286 if (!event_set) { 3287 e = perf_pmu__mem_events_ptr(pmu, PERF_MEM_EVENTS__LOAD_STORE); 3288 /* 3289 * The load and store operations are required, use the event 3290 * PERF_MEM_EVENTS__LOAD_STORE if it is supported. 3291 */ 3292 if (e->tag) { 3293 e->record = true; 3294 rec_argv[i++] = "-W"; 3295 } else { 3296 e = perf_pmu__mem_events_ptr(pmu, PERF_MEM_EVENTS__LOAD); 3297 e->record = true; 3298 3299 e = perf_pmu__mem_events_ptr(pmu, PERF_MEM_EVENTS__STORE); 3300 e->record = true; 3301 } 3302 } 3303 3304 e = perf_pmu__mem_events_ptr(pmu, PERF_MEM_EVENTS__LOAD); 3305 if (e->record) 3306 rec_argv[i++] = "-W"; 3307 3308 rec_argv[i++] = "-d"; 3309 rec_argv[i++] = "--phys-data"; 3310 rec_argv[i++] = "--sample-cpu"; 3311 3312 ret = perf_mem_events__record_args(rec_argv, &i); 3313 if (ret) 3314 goto out; 3315 3316 if (all_user) 3317 rec_argv[i++] = "--all-user"; 3318 3319 if (all_kernel) 3320 rec_argv[i++] = "--all-kernel"; 3321 3322 for (j = 0; j < argc; j++, i++) 3323 rec_argv[i] = argv[j]; 3324 3325 if (verbose > 0) { 3326 pr_debug("calling: "); 3327 3328 j = 0; 3329 3330 while (rec_argv[j]) { 3331 pr_debug("%s ", rec_argv[j]); 3332 j++; 3333 } 3334 pr_debug("\n"); 3335 } 3336 3337 ret = cmd_record(i, rec_argv); 3338out: 3339 free(rec_argv); 3340 return ret; 3341} 3342 3343int cmd_c2c(int argc, const char **argv) 3344{ 3345 argc = parse_options(argc, argv, c2c_options, c2c_usage, 3346 PARSE_OPT_STOP_AT_NON_OPTION); 3347 3348 if (!argc) 3349 usage_with_options(c2c_usage, c2c_options); 3350 3351 if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) { 3352 return perf_c2c__record(argc, argv); 3353 } else if (strlen(argv[0]) > 2 && strstarts("report", argv[0])) { 3354 return perf_c2c__report(argc, argv); 3355 } else { 3356 usage_with_options(c2c_usage, c2c_options); 3357 } 3358 3359 return 0; 3360} 3361