tools/perf/builtin-c2c.c

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