tools/perf/builtin-c2c.c

   1 /*
   2  * This is rewrite of original c2c tool introduced in here:
   3  *   http://lwn.net/Articles/588866/
   4  *
   5  * The original tool was changed to fit in current perf state.
   6  *
   7  * Original authors:
   8  *   Don Zickus <dzickus@redhat.com>
   9  *   Dick Fowles <fowles@inreach.com>
  10  *   Joe Mario <jmario@redhat.com>
  11  */
  12 #include <linux/compiler.h>
  13 #include <linux/kernel.h>
  14 #include <linux/stringify.h>
  15 #include <asm/bug.h>
  16 #include "util.h"
  17 #include "debug.h"
  18 #include "builtin.h"
  19 #include <subcmd/parse-options.h>
  20 #include "mem-events.h"
  21 #include "session.h"
  22 #include "hist.h"
  23 #include "sort.h"
  24 #include "tool.h"
  25 #include "data.h"
  26 #include "sort.h"
  27 #include "evlist.h"
  28 #include "evsel.h"
  29 #include <asm/bug.h>
  30 #include "ui/browsers/hists.h"
  31 #include "evlist.h"
  32
  33 struct c2c_hists {
  34         struct hists            hists;
  35         struct perf_hpp_list    list;
  36         struct c2c_stats        stats;
  37 };
  38
  39 struct compute_stats {
  40         struct stats             lcl_hitm;
  41         struct stats             rmt_hitm;
  42         struct stats             load;
  43 };
  44
  45 struct c2c_hist_entry {
  46         struct c2c_hists        *hists;
  47         struct c2c_stats         stats;
  48         unsigned long           *cpuset;
  49         struct c2c_stats        *node_stats;
  50         unsigned int             cacheline_idx;
  51
  52         struct compute_stats     cstats;
  53
  54         /*
  55          * must be at the end,
  56          * because of its callchain dynamic entry
  57          */
  58         struct hist_entry       he;
  59 };
  60
  61 static char const *coalesce_default = "pid,tid,iaddr";
  62
  63 struct perf_c2c {
  64         struct perf_tool        tool;
  65         struct c2c_hists        hists;
  66
  67         unsigned long           **nodes;
  68         int                      nodes_cnt;
  69         int                      cpus_cnt;
  70         int                     *cpu2node;
  71         int                      node_info;
  72
  73         bool                     show_src;
  74         bool                     show_all;
  75         bool                     use_stdio;
  76         bool                     stats_only;
  77         bool                     symbol_full;
  78
  79         /* HITM shared clines stats */
  80         struct c2c_stats        hitm_stats;
  81         int                     shared_clines;
  82
  83         int                      display;
  84
  85         const char              *coalesce;
  86         char                    *cl_sort;
  87         char                    *cl_resort;
  88         char                    *cl_output;
  89 };
  90
  91 enum {
  92         DISPLAY_LCL,
  93         DISPLAY_RMT,
  94         DISPLAY_TOT,
  95         DISPLAY_MAX,
  96 };
  97
  98 static const char *display_str[DISPLAY_MAX] = {
  99         [DISPLAY_LCL] = "Local",
 100         [DISPLAY_RMT] = "Remote",
 101         [DISPLAY_TOT] = "Total",
 102 };
 103
 104 static const struct option c2c_options[] = {
 105         OPT_INCR('v', "verbose", &verbose, "be more verbose (show counter open errors, etc)"),
 106         OPT_END()
 107 };
 108
 109 static struct perf_c2c c2c;
 110
 111 static void *c2c_he_zalloc(size_t size)
 112 {
 113         struct c2c_hist_entry *c2c_he;
 114
 115         c2c_he = zalloc(size + sizeof(*c2c_he));
 116         if (!c2c_he)
 117                 return NULL;
 118
 119         c2c_he->cpuset = bitmap_alloc(c2c.cpus_cnt);
 120         if (!c2c_he->cpuset)
 121                 return NULL;
 122
 123         c2c_he->node_stats = zalloc(c2c.nodes_cnt * sizeof(*c2c_he->node_stats));
 124         if (!c2c_he->node_stats)
 125                 return NULL;
 126
 127         init_stats(&c2c_he->cstats.lcl_hitm);
 128         init_stats(&c2c_he->cstats.rmt_hitm);
 129         init_stats(&c2c_he->cstats.load);
 130
 131         return &c2c_he->he;
 132 }
 133
 134 static void c2c_he_free(void *he)
 135 {
 136         struct c2c_hist_entry *c2c_he;
 137
 138         c2c_he = container_of(he, struct c2c_hist_entry, he);
 139         if (c2c_he->hists) {
 140                 hists__delete_entries(&c2c_he->hists->hists);
 141                 free(c2c_he->hists);
 142         }
 143
 144         free(c2c_he->cpuset);
 145         free(c2c_he->node_stats);
 146         free(c2c_he);
 147 }
 148
 149 static struct hist_entry_ops c2c_entry_ops = {
 150         .new    = c2c_he_zalloc,
 151         .free   = c2c_he_free,
 152 };
 153
 154 static int c2c_hists__init(struct c2c_hists *hists,
 155                            const char *sort,
 156                            int nr_header_lines);
 157
 158 static struct c2c_hists*
 159 he__get_c2c_hists(struct hist_entry *he,
 160                   const char *sort,
 161                   int nr_header_lines)
 162 {
 163         struct c2c_hist_entry *c2c_he;
 164         struct c2c_hists *hists;
 165         int ret;
 166
 167         c2c_he = container_of(he, struct c2c_hist_entry, he);
 168         if (c2c_he->hists)
 169                 return c2c_he->hists;
 170
 171         hists = c2c_he->hists = zalloc(sizeof(*hists));
 172         if (!hists)
 173                 return NULL;
 174
 175         ret = c2c_hists__init(hists, sort, nr_header_lines);
 176         if (ret) {
 177                 free(hists);
 178                 return NULL;
 179         }
 180
 181         return hists;
 182 }
 183
 184 static void c2c_he__set_cpu(struct c2c_hist_entry *c2c_he,
 185                             struct perf_sample *sample)
 186 {
 187         if (WARN_ONCE(sample->cpu == (unsigned int) -1,
 188                       "WARNING: no sample cpu value"))
 189                 return;
 190
 191         set_bit(sample->cpu, c2c_he->cpuset);
 192 }
 193
 194 static void compute_stats(struct c2c_hist_entry *c2c_he,
 195                           struct c2c_stats *stats,
 196                           u64 weight)
 197 {
 198         struct compute_stats *cstats = &c2c_he->cstats;
 199
 200         if (stats->rmt_hitm)
 201                 update_stats(&cstats->rmt_hitm, weight);
 202         else if (stats->lcl_hitm)
 203                 update_stats(&cstats->lcl_hitm, weight);
 204         else if (stats->load)
 205                 update_stats(&cstats->load, weight);
 206 }
 207
 208 static int process_sample_event(struct perf_tool *tool __maybe_unused,
 209                                 union perf_event *event,
 210                                 struct perf_sample *sample,
 211                                 struct perf_evsel *evsel __maybe_unused,
 212                                 struct machine *machine)
 213 {
 214         struct c2c_hists *c2c_hists = &c2c.hists;
 215         struct c2c_hist_entry *c2c_he;
 216         struct c2c_stats stats = { .nr_entries = 0, };
 217         struct hist_entry *he;
 218         struct addr_location al;
 219         struct mem_info *mi, *mi_dup;
 220         int ret;
 221
 222         if (machine__resolve(machine, &al, sample) < 0) {
 223                 pr_debug("problem processing %d event, skipping it.\n",
 224                          event->header.type);
 225                 return -1;
 226         }
 227
 228         ret = sample__resolve_callchain(sample, &callchain_cursor, NULL,
 229                                         evsel, &al, sysctl_perf_event_max_stack);
 230         if (ret)
 231                 goto out;
 232
 233         mi = sample__resolve_mem(sample, &al);
 234         if (mi == NULL)
 235                 return -ENOMEM;
 236
 237         mi_dup = memdup(mi, sizeof(*mi));
 238         if (!mi_dup)
 239                 goto free_mi;
 240
 241         c2c_decode_stats(&stats, mi);
 242
 243         he = hists__add_entry_ops(&c2c_hists->hists, &c2c_entry_ops,
 244                                   &al, NULL, NULL, mi,
 245                                   sample, true);
 246         if (he == NULL)
 247                 goto free_mi_dup;
 248
 249         c2c_he = container_of(he, struct c2c_hist_entry, he);
 250         c2c_add_stats(&c2c_he->stats, &stats);
 251         c2c_add_stats(&c2c_hists->stats, &stats);
 252
 253         c2c_he__set_cpu(c2c_he, sample);
 254
 255         hists__inc_nr_samples(&c2c_hists->hists, he->filtered);
 256         ret = hist_entry__append_callchain(he, sample);
 257
 258         if (!ret) {
 259                 /*
 260                  * There's already been warning about missing
 261                  * sample's cpu value. Let's account all to
 262                  * node 0 in this case, without any further
 263                  * warning.
 264                  *
 265                  * Doing node stats only for single callchain data.
 266                  */
 267                 int cpu = sample->cpu == (unsigned int) -1 ? 0 : sample->cpu;
 268                 int node = c2c.cpu2node[cpu];
 269
 270                 mi = mi_dup;
 271
 272                 mi_dup = memdup(mi, sizeof(*mi));
 273                 if (!mi_dup)
 274                         goto free_mi;
 275
 276                 c2c_hists = he__get_c2c_hists(he, c2c.cl_sort, 2);
 277                 if (!c2c_hists)
 278                         goto free_mi_dup;
 279
 280                 he = hists__add_entry_ops(&c2c_hists->hists, &c2c_entry_ops,
 281                                           &al, NULL, NULL, mi,
 282                                           sample, true);
 283                 if (he == NULL)
 284                         goto free_mi_dup;
 285
 286                 c2c_he = container_of(he, struct c2c_hist_entry, he);
 287                 c2c_add_stats(&c2c_he->stats, &stats);
 288                 c2c_add_stats(&c2c_hists->stats, &stats);
 289                 c2c_add_stats(&c2c_he->node_stats[node], &stats);
 290
 291                 compute_stats(c2c_he, &stats, sample->weight);
 292
 293                 c2c_he__set_cpu(c2c_he, sample);
 294
 295                 hists__inc_nr_samples(&c2c_hists->hists, he->filtered);
 296                 ret = hist_entry__append_callchain(he, sample);
 297         }
 298
 299 out:
 300         addr_location__put(&al);
 301         return ret;
 302
 303 free_mi_dup:
 304         free(mi_dup);
 305 free_mi:
 306         free(mi);
 307         ret = -ENOMEM;
 308         goto out;
 309 }
 310
 311 static struct perf_c2c c2c = {
 312         .tool = {
 313                 .sample         = process_sample_event,
 314                 .mmap           = perf_event__process_mmap,
 315                 .mmap2          = perf_event__process_mmap2,
 316                 .comm           = perf_event__process_comm,
 317                 .exit           = perf_event__process_exit,
 318                 .fork           = perf_event__process_fork,
 319                 .lost           = perf_event__process_lost,
 320                 .ordered_events = true,
 321                 .ordering_requires_timestamps = true,
 322         },
 323 };
 324
 325 static const char * const c2c_usage[] = {
 326         "perf c2c {record|report}",
 327         NULL
 328 };
 329
 330 static const char * const __usage_report[] = {
 331         "perf c2c report",
 332         NULL
 333 };
 334
 335 static const char * const *report_c2c_usage = __usage_report;
 336
 337 #define C2C_HEADER_MAX 2
 338
 339 struct c2c_header {
 340         struct {
 341                 const char *text;
 342                 int         span;
 343         } line[C2C_HEADER_MAX];
 344 };
 345
 346 struct c2c_dimension {
 347         struct c2c_header        header;
 348         const char              *name;
 349         int                      width;
 350         struct sort_entry       *se;
 351
 352         int64_t (*cmp)(struct perf_hpp_fmt *fmt,
 353                        struct hist_entry *, struct hist_entry *);
 354         int   (*entry)(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 355                        struct hist_entry *he);
 356         int   (*color)(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 357                        struct hist_entry *he);
 358 };
 359
 360 struct c2c_fmt {
 361         struct perf_hpp_fmt      fmt;
 362         struct c2c_dimension    *dim;
 363 };
 364
 365 #define SYMBOL_WIDTH 30
 366
 367 static struct c2c_dimension dim_symbol;
 368 static struct c2c_dimension dim_srcline;
 369
 370 static int symbol_width(struct hists *hists, struct sort_entry *se)
 371 {
 372         int width = hists__col_len(hists, se->se_width_idx);
 373
 374         if (!c2c.symbol_full)
 375                 width = MIN(width, SYMBOL_WIDTH);
 376
 377         return width;
 378 }
 379
 380 static int c2c_width(struct perf_hpp_fmt *fmt,
 381                      struct perf_hpp *hpp __maybe_unused,
 382                      struct hists *hists __maybe_unused)
 383 {
 384         struct c2c_fmt *c2c_fmt;
 385         struct c2c_dimension *dim;
 386
 387         c2c_fmt = container_of(fmt, struct c2c_fmt, fmt);
 388         dim = c2c_fmt->dim;
 389
 390         if (dim == &dim_symbol || dim == &dim_srcline)
 391                 return symbol_width(hists, dim->se);
 392
 393         return dim->se ? hists__col_len(hists, dim->se->se_width_idx) :
 394                          c2c_fmt->dim->width;
 395 }
 396
 397 static int c2c_header(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 398                       struct hists *hists, int line, int *span)
 399 {
 400         struct perf_hpp_list *hpp_list = hists->hpp_list;
 401         struct c2c_fmt *c2c_fmt;
 402         struct c2c_dimension *dim;
 403         const char *text = NULL;
 404         int width = c2c_width(fmt, hpp, hists);
 405
 406         c2c_fmt = container_of(fmt, struct c2c_fmt, fmt);
 407         dim = c2c_fmt->dim;
 408
 409         if (dim->se) {
 410                 text = dim->header.line[line].text;
 411                 /* Use the last line from sort_entry if not defined. */
 412                 if (!text && (line == hpp_list->nr_header_lines - 1))
 413                         text = dim->se->se_header;
 414         } else {
 415                 text = dim->header.line[line].text;
 416
 417                 if (*span) {
 418                         (*span)--;
 419                         return 0;
 420                 } else {
 421                         *span = dim->header.line[line].span;
 422                 }
 423         }
 424
 425         if (text == NULL)
 426                 text = "";
 427
 428         return scnprintf(hpp->buf, hpp->size, "%*s", width, text);
 429 }
 430
 431 #define HEX_STR(__s, __v)                               \
 432 ({                                                      \
 433         scnprintf(__s, sizeof(__s), "0x%" PRIx64, __v); \
 434         __s;                                            \
 435 })
 436
 437 static int64_t
 438 dcacheline_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 439                struct hist_entry *left, struct hist_entry *right)
 440 {
 441         return sort__dcacheline_cmp(left, right);
 442 }
 443
 444 static int dcacheline_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 445                             struct hist_entry *he)
 446 {
 447         uint64_t addr = 0;
 448         int width = c2c_width(fmt, hpp, he->hists);
 449         char buf[20];
 450
 451         if (he->mem_info)
 452                 addr = cl_address(he->mem_info->daddr.addr);
 453
 454         return scnprintf(hpp->buf, hpp->size, "%*s", width, HEX_STR(buf, addr));
 455 }
 456
 457 static int offset_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 458                         struct hist_entry *he)
 459 {
 460         uint64_t addr = 0;
 461         int width = c2c_width(fmt, hpp, he->hists);
 462         char buf[20];
 463
 464         if (he->mem_info)
 465                 addr = cl_offset(he->mem_info->daddr.al_addr);
 466
 467         return scnprintf(hpp->buf, hpp->size, "%*s", width, HEX_STR(buf, addr));
 468 }
 469
 470 static int64_t
 471 offset_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 472            struct hist_entry *left, struct hist_entry *right)
 473 {
 474         uint64_t l = 0, r = 0;
 475
 476         if (left->mem_info)
 477                 l = cl_offset(left->mem_info->daddr.addr);
 478         if (right->mem_info)
 479                 r = cl_offset(right->mem_info->daddr.addr);
 480
 481         return (int64_t)(r - l);
 482 }
 483
 484 static int
 485 iaddr_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 486             struct hist_entry *he)
 487 {
 488         uint64_t addr = 0;
 489         int width = c2c_width(fmt, hpp, he->hists);
 490         char buf[20];
 491
 492         if (he->mem_info)
 493                 addr = he->mem_info->iaddr.addr;
 494
 495         return scnprintf(hpp->buf, hpp->size, "%*s", width, HEX_STR(buf, addr));
 496 }
 497
 498 static int64_t
 499 iaddr_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 500           struct hist_entry *left, struct hist_entry *right)
 501 {
 502         return sort__iaddr_cmp(left, right);
 503 }
 504
 505 static int
 506 tot_hitm_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 507                struct hist_entry *he)
 508 {
 509         struct c2c_hist_entry *c2c_he;
 510         int width = c2c_width(fmt, hpp, he->hists);
 511         unsigned int tot_hitm;
 512
 513         c2c_he = container_of(he, struct c2c_hist_entry, he);
 514         tot_hitm = c2c_he->stats.lcl_hitm + c2c_he->stats.rmt_hitm;
 515
 516         return scnprintf(hpp->buf, hpp->size, "%*u", width, tot_hitm);
 517 }
 518
 519 static int64_t
 520 tot_hitm_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 521              struct hist_entry *left, struct hist_entry *right)
 522 {
 523         struct c2c_hist_entry *c2c_left;
 524         struct c2c_hist_entry *c2c_right;
 525         unsigned int tot_hitm_left;
 526         unsigned int tot_hitm_right;
 527
 528         c2c_left  = container_of(left, struct c2c_hist_entry, he);
 529         c2c_right = container_of(right, struct c2c_hist_entry, he);
 530
 531         tot_hitm_left  = c2c_left->stats.lcl_hitm + c2c_left->stats.rmt_hitm;
 532         tot_hitm_right = c2c_right->stats.lcl_hitm + c2c_right->stats.rmt_hitm;
 533
 534         return tot_hitm_left - tot_hitm_right;
 535 }
 536
 537 #define STAT_FN_ENTRY(__f)                                      \
 538 static int                                                      \
 539 __f ## _entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,   \
 540               struct hist_entry *he)                            \
 541 {                                                               \
 542         struct c2c_hist_entry *c2c_he;                          \
 543         int width = c2c_width(fmt, hpp, he->hists);             \
 544                                                                 \
 545         c2c_he = container_of(he, struct c2c_hist_entry, he);   \
 546         return scnprintf(hpp->buf, hpp->size, "%*u", width,     \
 547                          c2c_he->stats.__f);                    \
 548 }
 549
 550 #define STAT_FN_CMP(__f)                                                \
 551 static int64_t                                                          \
 552 __f ## _cmp(struct perf_hpp_fmt *fmt __maybe_unused,                    \
 553             struct hist_entry *left, struct hist_entry *right)          \
 554 {                                                                       \
 555         struct c2c_hist_entry *c2c_left, *c2c_right;                    \
 556                                                                         \
 557         c2c_left  = container_of(left, struct c2c_hist_entry, he);      \
 558         c2c_right = container_of(right, struct c2c_hist_entry, he);     \
 559         return c2c_left->stats.__f - c2c_right->stats.__f;              \
 560 }
 561
 562 #define STAT_FN(__f)            \
 563         STAT_FN_ENTRY(__f)      \
 564         STAT_FN_CMP(__f)
 565
 566 STAT_FN(rmt_hitm)
 567 STAT_FN(lcl_hitm)
 568 STAT_FN(store)
 569 STAT_FN(st_l1hit)
 570 STAT_FN(st_l1miss)
 571 STAT_FN(ld_fbhit)
 572 STAT_FN(ld_l1hit)
 573 STAT_FN(ld_l2hit)
 574 STAT_FN(ld_llchit)
 575 STAT_FN(rmt_hit)
 576
 577 static uint64_t llc_miss(struct c2c_stats *stats)
 578 {
 579         uint64_t llcmiss;
 580
 581         llcmiss = stats->lcl_dram +
 582                   stats->rmt_dram +
 583                   stats->rmt_hitm +
 584                   stats->rmt_hit;
 585
 586         return llcmiss;
 587 }
 588
 589 static int
 590 ld_llcmiss_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 591                  struct hist_entry *he)
 592 {
 593         struct c2c_hist_entry *c2c_he;
 594         int width = c2c_width(fmt, hpp, he->hists);
 595
 596         c2c_he = container_of(he, struct c2c_hist_entry, he);
 597
 598         return scnprintf(hpp->buf, hpp->size, "%*lu", width,
 599                          llc_miss(&c2c_he->stats));
 600 }
 601
 602 static int64_t
 603 ld_llcmiss_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 604                struct hist_entry *left, struct hist_entry *right)
 605 {
 606         struct c2c_hist_entry *c2c_left;
 607         struct c2c_hist_entry *c2c_right;
 608
 609         c2c_left  = container_of(left, struct c2c_hist_entry, he);
 610         c2c_right = container_of(right, struct c2c_hist_entry, he);
 611
 612         return llc_miss(&c2c_left->stats) - llc_miss(&c2c_right->stats);
 613 }
 614
 615 static uint64_t total_records(struct c2c_stats *stats)
 616 {
 617         uint64_t lclmiss, ldcnt, total;
 618
 619         lclmiss  = stats->lcl_dram +
 620                    stats->rmt_dram +
 621                    stats->rmt_hitm +
 622                    stats->rmt_hit;
 623
 624         ldcnt    = lclmiss +
 625                    stats->ld_fbhit +
 626                    stats->ld_l1hit +
 627                    stats->ld_l2hit +
 628                    stats->ld_llchit +
 629                    stats->lcl_hitm;
 630
 631         total    = ldcnt +
 632                    stats->st_l1hit +
 633                    stats->st_l1miss;
 634
 635         return total;
 636 }
 637
 638 static int
 639 tot_recs_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 640                 struct hist_entry *he)
 641 {
 642         struct c2c_hist_entry *c2c_he;
 643         int width = c2c_width(fmt, hpp, he->hists);
 644         uint64_t tot_recs;
 645
 646         c2c_he = container_of(he, struct c2c_hist_entry, he);
 647         tot_recs = total_records(&c2c_he->stats);
 648
 649         return scnprintf(hpp->buf, hpp->size, "%*" PRIu64, width, tot_recs);
 650 }
 651
 652 static int64_t
 653 tot_recs_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 654              struct hist_entry *left, struct hist_entry *right)
 655 {
 656         struct c2c_hist_entry *c2c_left;
 657         struct c2c_hist_entry *c2c_right;
 658         uint64_t tot_recs_left;
 659         uint64_t tot_recs_right;
 660
 661         c2c_left  = container_of(left, struct c2c_hist_entry, he);
 662         c2c_right = container_of(right, struct c2c_hist_entry, he);
 663
 664         tot_recs_left  = total_records(&c2c_left->stats);
 665         tot_recs_right = total_records(&c2c_right->stats);
 666
 667         return tot_recs_left - tot_recs_right;
 668 }
 669
 670 static uint64_t total_loads(struct c2c_stats *stats)
 671 {
 672         uint64_t lclmiss, ldcnt;
 673
 674         lclmiss  = stats->lcl_dram +
 675                    stats->rmt_dram +
 676                    stats->rmt_hitm +
 677                    stats->rmt_hit;
 678
 679         ldcnt    = lclmiss +
 680                    stats->ld_fbhit +
 681                    stats->ld_l1hit +
 682                    stats->ld_l2hit +
 683                    stats->ld_llchit +
 684                    stats->lcl_hitm;
 685
 686         return ldcnt;
 687 }
 688
 689 static int
 690 tot_loads_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 691                 struct hist_entry *he)
 692 {
 693         struct c2c_hist_entry *c2c_he;
 694         int width = c2c_width(fmt, hpp, he->hists);
 695         uint64_t tot_recs;
 696
 697         c2c_he = container_of(he, struct c2c_hist_entry, he);
 698         tot_recs = total_loads(&c2c_he->stats);
 699
 700         return scnprintf(hpp->buf, hpp->size, "%*" PRIu64, width, tot_recs);
 701 }
 702
 703 static int64_t
 704 tot_loads_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 705               struct hist_entry *left, struct hist_entry *right)
 706 {
 707         struct c2c_hist_entry *c2c_left;
 708         struct c2c_hist_entry *c2c_right;
 709         uint64_t tot_recs_left;
 710         uint64_t tot_recs_right;
 711
 712         c2c_left  = container_of(left, struct c2c_hist_entry, he);
 713         c2c_right = container_of(right, struct c2c_hist_entry, he);
 714
 715         tot_recs_left  = total_loads(&c2c_left->stats);
 716         tot_recs_right = total_loads(&c2c_right->stats);
 717
 718         return tot_recs_left - tot_recs_right;
 719 }
 720
 721 typedef double (get_percent_cb)(struct c2c_hist_entry *);
 722
 723 static int
 724 percent_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 725               struct hist_entry *he, get_percent_cb get_percent)
 726 {
 727         struct c2c_hist_entry *c2c_he;
 728         int width = c2c_width(fmt, hpp, he->hists);
 729         double per;
 730
 731         c2c_he = container_of(he, struct c2c_hist_entry, he);
 732         per = get_percent(c2c_he);
 733
 734 #ifdef HAVE_SLANG_SUPPORT
 735         if (use_browser)
 736                 return __hpp__slsmg_color_printf(hpp, "%*.2f%%", width - 1, per);
 737 #endif
 738         return hpp_color_scnprintf(hpp, "%*.2f%%", width - 1, per);
 739 }
 740
 741 static double percent_hitm(struct c2c_hist_entry *c2c_he)
 742 {
 743         struct c2c_hists *hists;
 744         struct c2c_stats *stats;
 745         struct c2c_stats *total;
 746         int tot = 0, st = 0;
 747         double p;
 748
 749         hists = container_of(c2c_he->he.hists, struct c2c_hists, hists);
 750         stats = &c2c_he->stats;
 751         total = &hists->stats;
 752
 753         switch (c2c.display) {
 754         case DISPLAY_RMT:
 755                 st  = stats->rmt_hitm;
 756                 tot = total->rmt_hitm;
 757                 break;
 758         case DISPLAY_LCL:
 759                 st  = stats->lcl_hitm;
 760                 tot = total->lcl_hitm;
 761                 break;
 762         case DISPLAY_TOT:
 763                 st  = stats->tot_hitm;
 764                 tot = total->tot_hitm;
 765         default:
 766                 break;
 767         }
 768
 769         p = tot ? (double) st / tot : 0;
 770
 771         return 100 * p;
 772 }
 773
 774 #define PERC_STR(__s, __v)                              \
 775 ({                                                      \
 776         scnprintf(__s, sizeof(__s), "%.2F%%", __v);     \
 777         __s;                                            \
 778 })
 779
 780 static int
 781 percent_hitm_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 782                    struct hist_entry *he)
 783 {
 784         struct c2c_hist_entry *c2c_he;
 785         int width = c2c_width(fmt, hpp, he->hists);
 786         char buf[10];
 787         double per;
 788
 789         c2c_he = container_of(he, struct c2c_hist_entry, he);
 790         per = percent_hitm(c2c_he);
 791         return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
 792 }
 793
 794 static int
 795 percent_hitm_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 796                    struct hist_entry *he)
 797 {
 798         return percent_color(fmt, hpp, he, percent_hitm);
 799 }
 800
 801 static int64_t
 802 percent_hitm_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 803                  struct hist_entry *left, struct hist_entry *right)
 804 {
 805         struct c2c_hist_entry *c2c_left;
 806         struct c2c_hist_entry *c2c_right;
 807         double per_left;
 808         double per_right;
 809
 810         c2c_left  = container_of(left, struct c2c_hist_entry, he);
 811         c2c_right = container_of(right, struct c2c_hist_entry, he);
 812
 813         per_left  = percent_hitm(c2c_left);
 814         per_right = percent_hitm(c2c_right);
 815
 816         return per_left - per_right;
 817 }
 818
 819 static struct c2c_stats *he_stats(struct hist_entry *he)
 820 {
 821         struct c2c_hist_entry *c2c_he;
 822
 823         c2c_he = container_of(he, struct c2c_hist_entry, he);
 824         return &c2c_he->stats;
 825 }
 826
 827 static struct c2c_stats *total_stats(struct hist_entry *he)
 828 {
 829         struct c2c_hists *hists;
 830
 831         hists = container_of(he->hists, struct c2c_hists, hists);
 832         return &hists->stats;
 833 }
 834
 835 static double percent(int st, int tot)
 836 {
 837         return tot ? 100. * (double) st / (double) tot : 0;
 838 }
 839
 840 #define PERCENT(__h, __f) percent(he_stats(__h)->__f, total_stats(__h)->__f)
 841
 842 #define PERCENT_FN(__f)                                                         \
 843 static double percent_ ## __f(struct c2c_hist_entry *c2c_he)                    \
 844 {                                                                               \
 845         struct c2c_hists *hists;                                                \
 846                                                                                 \
 847         hists = container_of(c2c_he->he.hists, struct c2c_hists, hists);        \
 848         return percent(c2c_he->stats.__f, hists->stats.__f);                    \
 849 }
 850
 851 PERCENT_FN(rmt_hitm)
 852 PERCENT_FN(lcl_hitm)
 853 PERCENT_FN(st_l1hit)
 854 PERCENT_FN(st_l1miss)
 855
 856 static int
 857 percent_rmt_hitm_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 858                        struct hist_entry *he)
 859 {
 860         int width = c2c_width(fmt, hpp, he->hists);
 861         double per = PERCENT(he, rmt_hitm);
 862         char buf[10];
 863
 864         return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
 865 }
 866
 867 static int
 868 percent_rmt_hitm_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 869                        struct hist_entry *he)
 870 {
 871         return percent_color(fmt, hpp, he, percent_rmt_hitm);
 872 }
 873
 874 static int64_t
 875 percent_rmt_hitm_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 876                      struct hist_entry *left, struct hist_entry *right)
 877 {
 878         double per_left;
 879         double per_right;
 880
 881         per_left  = PERCENT(left, lcl_hitm);
 882         per_right = PERCENT(right, lcl_hitm);
 883
 884         return per_left - per_right;
 885 }
 886
 887 static int
 888 percent_lcl_hitm_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 889                        struct hist_entry *he)
 890 {
 891         int width = c2c_width(fmt, hpp, he->hists);
 892         double per = PERCENT(he, lcl_hitm);
 893         char buf[10];
 894
 895         return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
 896 }
 897
 898 static int
 899 percent_lcl_hitm_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 900                        struct hist_entry *he)
 901 {
 902         return percent_color(fmt, hpp, he, percent_lcl_hitm);
 903 }
 904
 905 static int64_t
 906 percent_lcl_hitm_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 907                      struct hist_entry *left, struct hist_entry *right)
 908 {
 909         double per_left;
 910         double per_right;
 911
 912         per_left  = PERCENT(left, lcl_hitm);
 913         per_right = PERCENT(right, lcl_hitm);
 914
 915         return per_left - per_right;
 916 }
 917
 918 static int
 919 percent_stores_l1hit_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 920                            struct hist_entry *he)
 921 {
 922         int width = c2c_width(fmt, hpp, he->hists);
 923         double per = PERCENT(he, st_l1hit);
 924         char buf[10];
 925
 926         return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
 927 }
 928
 929 static int
 930 percent_stores_l1hit_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 931                            struct hist_entry *he)
 932 {
 933         return percent_color(fmt, hpp, he, percent_st_l1hit);
 934 }
 935
 936 static int64_t
 937 percent_stores_l1hit_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 938                         struct hist_entry *left, struct hist_entry *right)
 939 {
 940         double per_left;
 941         double per_right;
 942
 943         per_left  = PERCENT(left, st_l1hit);
 944         per_right = PERCENT(right, st_l1hit);
 945
 946         return per_left - per_right;
 947 }
 948
 949 static int
 950 percent_stores_l1miss_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 951                            struct hist_entry *he)
 952 {
 953         int width = c2c_width(fmt, hpp, he->hists);
 954         double per = PERCENT(he, st_l1miss);
 955         char buf[10];
 956
 957         return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
 958 }
 959
 960 static int
 961 percent_stores_l1miss_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 962                             struct hist_entry *he)
 963 {
 964         return percent_color(fmt, hpp, he, percent_st_l1miss);
 965 }
 966
 967 static int64_t
 968 percent_stores_l1miss_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 969                           struct hist_entry *left, struct hist_entry *right)
 970 {
 971         double per_left;
 972         double per_right;
 973
 974         per_left  = PERCENT(left, st_l1miss);
 975         per_right = PERCENT(right, st_l1miss);
 976
 977         return per_left - per_right;
 978 }
 979
 980 STAT_FN(lcl_dram)
 981 STAT_FN(rmt_dram)
 982
 983 static int
 984 pid_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 985           struct hist_entry *he)
 986 {
 987         int width = c2c_width(fmt, hpp, he->hists);
 988
 989         return scnprintf(hpp->buf, hpp->size, "%*d", width, he->thread->pid_);
 990 }
 991
 992 static int64_t
 993 pid_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 994         struct hist_entry *left, struct hist_entry *right)
 995 {
 996         return left->thread->pid_ - right->thread->pid_;
 997 }
 998
 999 static int64_t
1000 empty_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
1001           struct hist_entry *left __maybe_unused,
1002           struct hist_entry *right __maybe_unused)
1003 {
1004         return 0;
1005 }
1006
1007 static int
1008 node_entry(struct perf_hpp_fmt *fmt __maybe_unused, struct perf_hpp *hpp,
1009            struct hist_entry *he)
1010 {
1011         struct c2c_hist_entry *c2c_he;
1012         bool first = true;
1013         int node;
1014         int ret = 0;
1015
1016         c2c_he = container_of(he, struct c2c_hist_entry, he);
1017
1018         for (node = 0; node < c2c.nodes_cnt; node++) {
1019                 DECLARE_BITMAP(set, c2c.cpus_cnt);
1020
1021                 bitmap_zero(set, c2c.cpus_cnt);
1022                 bitmap_and(set, c2c_he->cpuset, c2c.nodes[node], c2c.cpus_cnt);
1023
1024                 if (!bitmap_weight(set, c2c.cpus_cnt)) {
1025                         if (c2c.node_info == 1) {
1026                                 ret = scnprintf(hpp->buf, hpp->size, "%21s", " ");
1027                                 advance_hpp(hpp, ret);
1028                         }
1029                         continue;
1030                 }
1031
1032                 if (!first) {
1033                         ret = scnprintf(hpp->buf, hpp->size, " ");
1034                         advance_hpp(hpp, ret);
1035                 }
1036
1037                 switch (c2c.node_info) {
1038                 case 0:
1039                         ret = scnprintf(hpp->buf, hpp->size, "%2d", node);
1040                         advance_hpp(hpp, ret);
1041                         break;
1042                 case 1:
1043                 {
1044                         int num = bitmap_weight(c2c_he->cpuset, c2c.cpus_cnt);
1045                         struct c2c_stats *stats = &c2c_he->node_stats[node];
1046
1047                         ret = scnprintf(hpp->buf, hpp->size, "%2d{%2d ", node, num);
1048                         advance_hpp(hpp, ret);
1049
1050                 #define DISPLAY_HITM(__h)                                               \
1051                         if (c2c_he->stats.__h> 0) {                                     \
1052                                 ret = scnprintf(hpp->buf, hpp->size, "%5.1f%% ",        \
1053                                                 percent(stats->__h, c2c_he->stats.__h));\
1054                         } else {                                                        \
1055                                 ret = scnprintf(hpp->buf, hpp->size, "%6s ", "n/a");    \
1056                         }
1057
1058                         switch (c2c.display) {
1059                         case DISPLAY_RMT:
1060                                 DISPLAY_HITM(rmt_hitm);
1061                                 break;
1062                         case DISPLAY_LCL:
1063                                 DISPLAY_HITM(lcl_hitm);
1064                                 break;
1065                         case DISPLAY_TOT:
1066                                 DISPLAY_HITM(tot_hitm);
1067                         default:
1068                                 break;
1069                         }
1070
1071                 #undef DISPLAY_HITM
1072
1073                         advance_hpp(hpp, ret);
1074
1075                         if (c2c_he->stats.store > 0) {
1076                                 ret = scnprintf(hpp->buf, hpp->size, "%5.1f%%}",
1077                                                 percent(stats->store, c2c_he->stats.store));
1078                         } else {
1079                                 ret = scnprintf(hpp->buf, hpp->size, "%6s}", "n/a");
1080                         }
1081
1082                         advance_hpp(hpp, ret);
1083                         break;
1084                 }
1085                 case 2:
1086                         ret = scnprintf(hpp->buf, hpp->size, "%2d{", node);
1087                         advance_hpp(hpp, ret);
1088
1089                         ret = bitmap_scnprintf(set, c2c.cpus_cnt, hpp->buf, hpp->size);
1090                         advance_hpp(hpp, ret);
1091
1092                         ret = scnprintf(hpp->buf, hpp->size, "}");
1093                         advance_hpp(hpp, ret);
1094                         break;
1095                 default:
1096                         break;
1097                 }
1098
1099                 first = false;
1100         }
1101
1102         return 0;
1103 }
1104
1105 static int
1106 mean_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
1107            struct hist_entry *he, double mean)
1108 {
1109         int width = c2c_width(fmt, hpp, he->hists);
1110         char buf[10];
1111
1112         scnprintf(buf, 10, "%6.0f", mean);
1113         return scnprintf(hpp->buf, hpp->size, "%*s", width, buf);
1114 }
1115
1116 #define MEAN_ENTRY(__func, __val)                                               \
1117 static int                                                                      \
1118 __func(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, struct hist_entry *he)   \
1119 {                                                                               \
1120         struct c2c_hist_entry *c2c_he;                                          \
1121         c2c_he = container_of(he, struct c2c_hist_entry, he);                   \
1122         return mean_entry(fmt, hpp, he, avg_stats(&c2c_he->cstats.__val));      \
1123 }
1124
1125 MEAN_ENTRY(mean_rmt_entry,  rmt_hitm);
1126 MEAN_ENTRY(mean_lcl_entry,  lcl_hitm);
1127 MEAN_ENTRY(mean_load_entry, load);
1128
1129 static int
1130 cpucnt_entry(struct perf_hpp_fmt *fmt __maybe_unused, struct perf_hpp *hpp,
1131              struct hist_entry *he)
1132 {
1133         struct c2c_hist_entry *c2c_he;
1134         int width = c2c_width(fmt, hpp, he->hists);
1135         char buf[10];
1136
1137         c2c_he = container_of(he, struct c2c_hist_entry, he);
1138
1139         scnprintf(buf, 10, "%d", bitmap_weight(c2c_he->cpuset, c2c.cpus_cnt));
1140         return scnprintf(hpp->buf, hpp->size, "%*s", width, buf);
1141 }
1142
1143 static int
1144 cl_idx_entry(struct perf_hpp_fmt *fmt __maybe_unused, struct perf_hpp *hpp,
1145              struct hist_entry *he)
1146 {
1147         struct c2c_hist_entry *c2c_he;
1148         int width = c2c_width(fmt, hpp, he->hists);
1149         char buf[10];
1150
1151         c2c_he = container_of(he, struct c2c_hist_entry, he);
1152
1153         scnprintf(buf, 10, "%u", c2c_he->cacheline_idx);
1154         return scnprintf(hpp->buf, hpp->size, "%*s", width, buf);
1155 }
1156
1157 static int
1158 cl_idx_empty_entry(struct perf_hpp_fmt *fmt __maybe_unused, struct perf_hpp *hpp,
1159                    struct hist_entry *he)
1160 {
1161         int width = c2c_width(fmt, hpp, he->hists);
1162
1163         return scnprintf(hpp->buf, hpp->size, "%*s", width, "");
1164 }
1165
1166 #define HEADER_LOW(__h)                 \
1167         {                               \
1168                 .line[1] = {            \
1169                         .text = __h,    \
1170                 },                      \
1171         }
1172
1173 #define HEADER_BOTH(__h0, __h1)         \
1174         {                               \
1175                 .line[0] = {            \
1176                         .text = __h0,   \
1177                 },                      \
1178                 .line[1] = {            \
1179                         .text = __h1,   \
1180                 },                      \
1181         }
1182
1183 #define HEADER_SPAN(__h0, __h1, __s)    \
1184         {                               \
1185                 .line[0] = {            \
1186                         .text = __h0,   \
1187                         .span = __s,    \
1188                 },                      \
1189                 .line[1] = {            \
1190                         .text = __h1,   \
1191                 },                      \
1192         }
1193
1194 #define HEADER_SPAN_LOW(__h)            \
1195         {                               \
1196                 .line[1] = {            \
1197                         .text = __h,    \
1198                 },                      \
1199         }
1200
1201 static struct c2c_dimension dim_dcacheline = {
1202         .header         = HEADER_LOW("Cacheline"),
1203         .name           = "dcacheline",
1204         .cmp            = dcacheline_cmp,
1205         .entry          = dcacheline_entry,
1206         .width          = 18,
1207 };
1208
1209 static struct c2c_header header_offset_tui = HEADER_LOW("Off");
1210
1211 static struct c2c_dimension dim_offset = {
1212         .header         = HEADER_BOTH("Data address", "Offset"),
1213         .name           = "offset",
1214         .cmp            = offset_cmp,
1215         .entry          = offset_entry,
1216         .width          = 18,
1217 };
1218
1219 static struct c2c_dimension dim_iaddr = {
1220         .header         = HEADER_LOW("Code address"),
1221         .name           = "iaddr",
1222         .cmp            = iaddr_cmp,
1223         .entry          = iaddr_entry,
1224         .width          = 18,
1225 };
1226
1227 static struct c2c_dimension dim_tot_hitm = {
1228         .header         = HEADER_SPAN("----- LLC Load Hitm -----", "Total", 2),
1229         .name           = "tot_hitm",
1230         .cmp            = tot_hitm_cmp,
1231         .entry          = tot_hitm_entry,
1232         .width          = 7,
1233 };
1234
1235 static struct c2c_dimension dim_lcl_hitm = {
1236         .header         = HEADER_SPAN_LOW("Lcl"),
1237         .name           = "lcl_hitm",
1238         .cmp            = lcl_hitm_cmp,
1239         .entry          = lcl_hitm_entry,
1240         .width          = 7,
1241 };
1242
1243 static struct c2c_dimension dim_rmt_hitm = {
1244         .header         = HEADER_SPAN_LOW("Rmt"),
1245         .name           = "rmt_hitm",
1246         .cmp            = rmt_hitm_cmp,
1247         .entry          = rmt_hitm_entry,
1248         .width          = 7,
1249 };
1250
1251 static struct c2c_dimension dim_cl_rmt_hitm = {
1252         .header         = HEADER_SPAN("----- HITM -----", "Rmt", 1),
1253         .name           = "cl_rmt_hitm",
1254         .cmp            = rmt_hitm_cmp,
1255         .entry          = rmt_hitm_entry,
1256         .width          = 7,
1257 };
1258
1259 static struct c2c_dimension dim_cl_lcl_hitm = {
1260         .header         = HEADER_SPAN_LOW("Lcl"),
1261         .name           = "cl_lcl_hitm",
1262         .cmp            = lcl_hitm_cmp,
1263         .entry          = lcl_hitm_entry,
1264         .width          = 7,
1265 };
1266
1267 static struct c2c_dimension dim_stores = {
1268         .header         = HEADER_SPAN("---- Store Reference ----", "Total", 2),
1269         .name           = "stores",
1270         .cmp            = store_cmp,
1271         .entry          = store_entry,
1272         .width          = 7,
1273 };
1274
1275 static struct c2c_dimension dim_stores_l1hit = {
1276         .header         = HEADER_SPAN_LOW("L1Hit"),
1277         .name           = "stores_l1hit",
1278         .cmp            = st_l1hit_cmp,
1279         .entry          = st_l1hit_entry,
1280         .width          = 7,
1281 };
1282
1283 static struct c2c_dimension dim_stores_l1miss = {
1284         .header         = HEADER_SPAN_LOW("L1Miss"),
1285         .name           = "stores_l1miss",
1286         .cmp            = st_l1miss_cmp,
1287         .entry          = st_l1miss_entry,
1288         .width          = 7,
1289 };
1290
1291 static struct c2c_dimension dim_cl_stores_l1hit = {
1292         .header         = HEADER_SPAN("-- Store Refs --", "L1 Hit", 1),
1293         .name           = "cl_stores_l1hit",
1294         .cmp            = st_l1hit_cmp,
1295         .entry          = st_l1hit_entry,
1296         .width          = 7,
1297 };
1298
1299 static struct c2c_dimension dim_cl_stores_l1miss = {
1300         .header         = HEADER_SPAN_LOW("L1 Miss"),
1301         .name           = "cl_stores_l1miss",
1302         .cmp            = st_l1miss_cmp,
1303         .entry          = st_l1miss_entry,
1304         .width          = 7,
1305 };
1306
1307 static struct c2c_dimension dim_ld_fbhit = {
1308         .header         = HEADER_SPAN("----- Core Load Hit -----", "FB", 2),
1309         .name           = "ld_fbhit",
1310         .cmp            = ld_fbhit_cmp,
1311         .entry          = ld_fbhit_entry,
1312         .width          = 7,
1313 };
1314
1315 static struct c2c_dimension dim_ld_l1hit = {
1316         .header         = HEADER_SPAN_LOW("L1"),
1317         .name           = "ld_l1hit",
1318         .cmp            = ld_l1hit_cmp,
1319         .entry          = ld_l1hit_entry,
1320         .width          = 7,
1321 };
1322
1323 static struct c2c_dimension dim_ld_l2hit = {
1324         .header         = HEADER_SPAN_LOW("L2"),
1325         .name           = "ld_l2hit",
1326         .cmp            = ld_l2hit_cmp,
1327         .entry          = ld_l2hit_entry,
1328         .width          = 7,
1329 };
1330
1331 static struct c2c_dimension dim_ld_llchit = {
1332         .header         = HEADER_SPAN("-- LLC Load Hit --", "Llc", 1),
1333         .name           = "ld_lclhit",
1334         .cmp            = ld_llchit_cmp,
1335         .entry          = ld_llchit_entry,
1336         .width          = 8,
1337 };
1338
1339 static struct c2c_dimension dim_ld_rmthit = {
1340         .header         = HEADER_SPAN_LOW("Rmt"),
1341         .name           = "ld_rmthit",
1342         .cmp            = rmt_hit_cmp,
1343         .entry          = rmt_hit_entry,
1344         .width          = 8,
1345 };
1346
1347 static struct c2c_dimension dim_ld_llcmiss = {
1348         .header         = HEADER_BOTH("LLC", "Ld Miss"),
1349         .name           = "ld_llcmiss",
1350         .cmp            = ld_llcmiss_cmp,
1351         .entry          = ld_llcmiss_entry,
1352         .width          = 7,
1353 };
1354
1355 static struct c2c_dimension dim_tot_recs = {
1356         .header         = HEADER_BOTH("Total", "records"),
1357         .name           = "tot_recs",
1358         .cmp            = tot_recs_cmp,
1359         .entry          = tot_recs_entry,
1360         .width          = 7,
1361 };
1362
1363 static struct c2c_dimension dim_tot_loads = {
1364         .header         = HEADER_BOTH("Total", "Loads"),
1365         .name           = "tot_loads",
1366         .cmp            = tot_loads_cmp,
1367         .entry          = tot_loads_entry,
1368         .width          = 7,
1369 };
1370
1371 static struct c2c_header percent_hitm_header[] = {
1372         [DISPLAY_LCL] = HEADER_BOTH("Lcl", "Hitm"),
1373         [DISPLAY_RMT] = HEADER_BOTH("Rmt", "Hitm"),
1374         [DISPLAY_TOT] = HEADER_BOTH("Tot", "Hitm"),
1375 };
1376
1377 static struct c2c_dimension dim_percent_hitm = {
1378         .name           = "percent_hitm",
1379         .cmp            = percent_hitm_cmp,
1380         .entry          = percent_hitm_entry,
1381         .color          = percent_hitm_color,
1382         .width          = 7,
1383 };
1384
1385 static struct c2c_dimension dim_percent_rmt_hitm = {
1386         .header         = HEADER_SPAN("----- HITM -----", "Rmt", 1),
1387         .name           = "percent_rmt_hitm",
1388         .cmp            = percent_rmt_hitm_cmp,
1389         .entry          = percent_rmt_hitm_entry,
1390         .color          = percent_rmt_hitm_color,
1391         .width          = 7,
1392 };
1393
1394 static struct c2c_dimension dim_percent_lcl_hitm = {
1395         .header         = HEADER_SPAN_LOW("Lcl"),
1396         .name           = "percent_lcl_hitm",
1397         .cmp            = percent_lcl_hitm_cmp,
1398         .entry          = percent_lcl_hitm_entry,
1399         .color          = percent_lcl_hitm_color,
1400         .width          = 7,
1401 };
1402
1403 static struct c2c_dimension dim_percent_stores_l1hit = {
1404         .header         = HEADER_SPAN("-- Store Refs --", "L1 Hit", 1),
1405         .name           = "percent_stores_l1hit",
1406         .cmp            = percent_stores_l1hit_cmp,
1407         .entry          = percent_stores_l1hit_entry,
1408         .color          = percent_stores_l1hit_color,
1409         .width          = 7,
1410 };
1411
1412 static struct c2c_dimension dim_percent_stores_l1miss = {
1413         .header         = HEADER_SPAN_LOW("L1 Miss"),
1414         .name           = "percent_stores_l1miss",
1415         .cmp            = percent_stores_l1miss_cmp,
1416         .entry          = percent_stores_l1miss_entry,
1417         .color          = percent_stores_l1miss_color,
1418         .width          = 7,
1419 };
1420
1421 static struct c2c_dimension dim_dram_lcl = {
1422         .header         = HEADER_SPAN("--- Load Dram ----", "Lcl", 1),
1423         .name           = "dram_lcl",
1424         .cmp            = lcl_dram_cmp,
1425         .entry          = lcl_dram_entry,
1426         .width          = 8,
1427 };
1428
1429 static struct c2c_dimension dim_dram_rmt = {
1430         .header         = HEADER_SPAN_LOW("Rmt"),
1431         .name           = "dram_rmt",
1432         .cmp            = rmt_dram_cmp,
1433         .entry          = rmt_dram_entry,
1434         .width          = 8,
1435 };
1436
1437 static struct c2c_dimension dim_pid = {
1438         .header         = HEADER_LOW("Pid"),
1439         .name           = "pid",
1440         .cmp            = pid_cmp,
1441         .entry          = pid_entry,
1442         .width          = 7,
1443 };
1444
1445 static struct c2c_dimension dim_tid = {
1446         .header         = HEADER_LOW("Tid"),
1447         .name           = "tid",
1448         .se             = &sort_thread,
1449 };
1450
1451 static struct c2c_dimension dim_symbol = {
1452         .name           = "symbol",
1453         .se             = &sort_sym,
1454 };
1455
1456 static struct c2c_dimension dim_dso = {
1457         .header         = HEADER_BOTH("Shared", "Object"),
1458         .name           = "dso",
1459         .se             = &sort_dso,
1460 };
1461
1462 static struct c2c_header header_node[3] = {
1463         HEADER_LOW("Node"),
1464         HEADER_LOW("Node{cpus %hitms %stores}"),
1465         HEADER_LOW("Node{cpu list}"),
1466 };
1467
1468 static struct c2c_dimension dim_node = {
1469         .name           = "node",
1470         .cmp            = empty_cmp,
1471         .entry          = node_entry,
1472         .width          = 4,
1473 };
1474
1475 static struct c2c_dimension dim_mean_rmt = {
1476         .header         = HEADER_SPAN("---------- cycles ----------", "rmt hitm", 2),
1477         .name           = "mean_rmt",
1478         .cmp            = empty_cmp,
1479         .entry          = mean_rmt_entry,
1480         .width          = 8,
1481 };
1482
1483 static struct c2c_dimension dim_mean_lcl = {
1484         .header         = HEADER_SPAN_LOW("lcl hitm"),
1485         .name           = "mean_lcl",
1486         .cmp            = empty_cmp,
1487         .entry          = mean_lcl_entry,
1488         .width          = 8,
1489 };
1490
1491 static struct c2c_dimension dim_mean_load = {
1492         .header         = HEADER_SPAN_LOW("load"),
1493         .name           = "mean_load",
1494         .cmp            = empty_cmp,
1495         .entry          = mean_load_entry,
1496         .width          = 8,
1497 };
1498
1499 static struct c2c_dimension dim_cpucnt = {
1500         .header         = HEADER_BOTH("cpu", "cnt"),
1501         .name           = "cpucnt",
1502         .cmp            = empty_cmp,
1503         .entry          = cpucnt_entry,
1504         .width          = 8,
1505 };
1506
1507 static struct c2c_dimension dim_srcline = {
1508         .name           = "cl_srcline",
1509         .se             = &sort_srcline,
1510 };
1511
1512 static struct c2c_dimension dim_dcacheline_idx = {
1513         .header         = HEADER_LOW("Index"),
1514         .name           = "cl_idx",
1515         .cmp            = empty_cmp,
1516         .entry          = cl_idx_entry,
1517         .width          = 5,
1518 };
1519
1520 static struct c2c_dimension dim_dcacheline_num = {
1521         .header         = HEADER_LOW("Num"),
1522         .name           = "cl_num",
1523         .cmp            = empty_cmp,
1524         .entry          = cl_idx_entry,
1525         .width          = 5,
1526 };
1527
1528 static struct c2c_dimension dim_dcacheline_num_empty = {
1529         .header         = HEADER_LOW("Num"),
1530         .name           = "cl_num_empty",
1531         .cmp            = empty_cmp,
1532         .entry          = cl_idx_empty_entry,
1533         .width          = 5,
1534 };
1535
1536 static struct c2c_dimension *dimensions[] = {
1537         &dim_dcacheline,
1538         &dim_offset,
1539         &dim_iaddr,
1540         &dim_tot_hitm,
1541         &dim_lcl_hitm,
1542         &dim_rmt_hitm,
1543         &dim_cl_lcl_hitm,
1544         &dim_cl_rmt_hitm,
1545         &dim_stores,
1546         &dim_stores_l1hit,
1547         &dim_stores_l1miss,
1548         &dim_cl_stores_l1hit,
1549         &dim_cl_stores_l1miss,
1550         &dim_ld_fbhit,
1551         &dim_ld_l1hit,
1552         &dim_ld_l2hit,
1553         &dim_ld_llchit,
1554         &dim_ld_rmthit,
1555         &dim_ld_llcmiss,
1556         &dim_tot_recs,
1557         &dim_tot_loads,
1558         &dim_percent_hitm,
1559         &dim_percent_rmt_hitm,
1560         &dim_percent_lcl_hitm,
1561         &dim_percent_stores_l1hit,
1562         &dim_percent_stores_l1miss,
1563         &dim_dram_lcl,
1564         &dim_dram_rmt,
1565         &dim_pid,
1566         &dim_tid,
1567         &dim_symbol,
1568         &dim_dso,
1569         &dim_node,
1570         &dim_mean_rmt,
1571         &dim_mean_lcl,
1572         &dim_mean_load,
1573         &dim_cpucnt,
1574         &dim_srcline,
1575         &dim_dcacheline_idx,
1576         &dim_dcacheline_num,
1577         &dim_dcacheline_num_empty,
1578         NULL,
1579 };
1580
1581 static void fmt_free(struct perf_hpp_fmt *fmt)
1582 {
1583         struct c2c_fmt *c2c_fmt;
1584
1585         c2c_fmt = container_of(fmt, struct c2c_fmt, fmt);
1586         free(c2c_fmt);
1587 }
1588
1589 static bool fmt_equal(struct perf_hpp_fmt *a, struct perf_hpp_fmt *b)
1590 {
1591         struct c2c_fmt *c2c_a = container_of(a, struct c2c_fmt, fmt);
1592         struct c2c_fmt *c2c_b = container_of(b, struct c2c_fmt, fmt);
1593
1594         return c2c_a->dim == c2c_b->dim;
1595 }
1596
1597 static struct c2c_dimension *get_dimension(const char *name)
1598 {
1599         unsigned int i;
1600
1601         for (i = 0; dimensions[i]; i++) {
1602                 struct c2c_dimension *dim = dimensions[i];
1603
1604                 if (!strcmp(dim->name, name))
1605                         return dim;
1606         };
1607
1608         return NULL;
1609 }
1610
1611 static int c2c_se_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
1612                         struct hist_entry *he)
1613 {
1614         struct c2c_fmt *c2c_fmt = container_of(fmt, struct c2c_fmt, fmt);
1615         struct c2c_dimension *dim = c2c_fmt->dim;
1616         size_t len = fmt->user_len;
1617
1618         if (!len) {
1619                 len = hists__col_len(he->hists, dim->se->se_width_idx);
1620
1621                 if (dim == &dim_symbol || dim == &dim_srcline)
1622                         len = symbol_width(he->hists, dim->se);
1623         }
1624
1625         return dim->se->se_snprintf(he, hpp->buf, hpp->size, len);
1626 }
1627
1628 static int64_t c2c_se_cmp(struct perf_hpp_fmt *fmt,
1629                           struct hist_entry *a, struct hist_entry *b)
1630 {
1631         struct c2c_fmt *c2c_fmt = container_of(fmt, struct c2c_fmt, fmt);
1632         struct c2c_dimension *dim = c2c_fmt->dim;
1633
1634         return dim->se->se_cmp(a, b);
1635 }
1636
1637 static int64_t c2c_se_collapse(struct perf_hpp_fmt *fmt,
1638                                struct hist_entry *a, struct hist_entry *b)
1639 {
1640         struct c2c_fmt *c2c_fmt = container_of(fmt, struct c2c_fmt, fmt);
1641         struct c2c_dimension *dim = c2c_fmt->dim;
1642         int64_t (*collapse_fn)(struct hist_entry *, struct hist_entry *);
1643
1644         collapse_fn = dim->se->se_collapse ?: dim->se->se_cmp;
1645         return collapse_fn(a, b);
1646 }
1647
1648 static struct c2c_fmt *get_format(const char *name)
1649 {
1650         struct c2c_dimension *dim = get_dimension(name);
1651         struct c2c_fmt *c2c_fmt;
1652         struct perf_hpp_fmt *fmt;
1653
1654         if (!dim)
1655                 return NULL;
1656
1657         c2c_fmt = zalloc(sizeof(*c2c_fmt));
1658         if (!c2c_fmt)
1659                 return NULL;
1660
1661         c2c_fmt->dim = dim;
1662
1663         fmt = &c2c_fmt->fmt;
1664         INIT_LIST_HEAD(&fmt->list);
1665         INIT_LIST_HEAD(&fmt->sort_list);
1666
1667         fmt->cmp        = dim->se ? c2c_se_cmp   : dim->cmp;
1668         fmt->sort       = dim->se ? c2c_se_cmp   : dim->cmp;
1669         fmt->color      = dim->se ? NULL         : dim->color;
1670         fmt->entry      = dim->se ? c2c_se_entry : dim->entry;
1671         fmt->header     = c2c_header;
1672         fmt->width      = c2c_width;
1673         fmt->collapse   = dim->se ? c2c_se_collapse : dim->cmp;
1674         fmt->equal      = fmt_equal;
1675         fmt->free       = fmt_free;
1676
1677         return c2c_fmt;
1678 }
1679
1680 static int c2c_hists__init_output(struct perf_hpp_list *hpp_list, char *name)
1681 {
1682         struct c2c_fmt *c2c_fmt = get_format(name);
1683
1684         if (!c2c_fmt) {
1685                 reset_dimensions();
1686                 return output_field_add(hpp_list, name);
1687         }
1688
1689         perf_hpp_list__column_register(hpp_list, &c2c_fmt->fmt);
1690         return 0;
1691 }
1692
1693 static int c2c_hists__init_sort(struct perf_hpp_list *hpp_list, char *name)
1694 {
1695         struct c2c_fmt *c2c_fmt = get_format(name);
1696         struct c2c_dimension *dim;
1697
1698         if (!c2c_fmt) {
1699                 reset_dimensions();
1700                 return sort_dimension__add(hpp_list, name, NULL, 0);
1701         }
1702
1703         dim = c2c_fmt->dim;
1704         if (dim == &dim_dso)
1705                 hpp_list->dso = 1;
1706
1707         perf_hpp_list__register_sort_field(hpp_list, &c2c_fmt->fmt);
1708         return 0;
1709 }
1710
1711 #define PARSE_LIST(_list, _fn)                                                  \
1712         do {                                                                    \
1713                 char *tmp, *tok;                                                \
1714                 ret = 0;                                                        \
1715                                                                                 \
1716                 if (!_list)                                                     \
1717                         break;                                                  \
1718                                                                                 \
1719                 for (tok = strtok_r((char *)_list, ", ", &tmp);                 \
1720                                 tok; tok = strtok_r(NULL, ", ", &tmp)) {        \
1721                         ret = _fn(hpp_list, tok);                               \
1722                         if (ret == -EINVAL) {                                   \
1723                                 error("Invalid --fields key: `%s'", tok);       \
1724                                 break;                                          \
1725                         } else if (ret == -ESRCH) {                             \
1726                                 error("Unknown --fields key: `%s'", tok);       \
1727                                 break;                                          \
1728                         }                                                       \
1729                 }                                                               \
1730         } while (0)
1731
1732 static int hpp_list__parse(struct perf_hpp_list *hpp_list,
1733                            const char *output_,
1734                            const char *sort_)
1735 {
1736         char *output = output_ ? strdup(output_) : NULL;
1737         char *sort   = sort_   ? strdup(sort_) : NULL;
1738         int ret;
1739
1740         PARSE_LIST(output, c2c_hists__init_output);
1741         PARSE_LIST(sort,   c2c_hists__init_sort);
1742
1743         /* copy sort keys to output fields */
1744         perf_hpp__setup_output_field(hpp_list);
1745
1746         /*
1747          * We dont need other sorting keys other than those
1748          * we already specified. It also really slows down
1749          * the processing a lot with big number of output
1750          * fields, so switching this off for c2c.
1751          */
1752
1753 #if 0
1754         /* and then copy output fields to sort keys */
1755         perf_hpp__append_sort_keys(&hists->list);
1756 #endif
1757
1758         free(output);
1759         free(sort);
1760         return ret;
1761 }
1762
1763 static int c2c_hists__init(struct c2c_hists *hists,
1764                            const char *sort,
1765                            int nr_header_lines)
1766 {
1767         __hists__init(&hists->hists, &hists->list);
1768
1769         /*
1770          * Initialize only with sort fields, we need to resort
1771          * later anyway, and that's where we add output fields
1772          * as well.
1773          */
1774         perf_hpp_list__init(&hists->list);
1775
1776         /* Overload number of header lines.*/
1777         hists->list.nr_header_lines = nr_header_lines;
1778
1779         return hpp_list__parse(&hists->list, NULL, sort);
1780 }
1781
1782 __maybe_unused
1783 static int c2c_hists__reinit(struct c2c_hists *c2c_hists,
1784                              const char *output,
1785                              const char *sort)
1786 {
1787         perf_hpp__reset_output_field(&c2c_hists->list);
1788         return hpp_list__parse(&c2c_hists->list, output, sort);
1789 }
1790
1791 #define DISPLAY_LINE_LIMIT  0.0005
1792
1793 static bool he__display(struct hist_entry *he, struct c2c_stats *stats)
1794 {
1795         struct c2c_hist_entry *c2c_he;
1796         double ld_dist;
1797
1798         if (c2c.show_all)
1799                 return true;
1800
1801         c2c_he = container_of(he, struct c2c_hist_entry, he);
1802
1803 #define FILTER_HITM(__h)                                                \
1804         if (stats->__h) {                                               \
1805                 ld_dist = ((double)c2c_he->stats.__h / stats->__h);     \
1806                 if (ld_dist < DISPLAY_LINE_LIMIT)                       \
1807                         he->filtered = HIST_FILTER__C2C;                \
1808         } else {                                                        \
1809                 he->filtered = HIST_FILTER__C2C;                        \
1810         }
1811
1812         switch (c2c.display) {
1813         case DISPLAY_LCL:
1814                 FILTER_HITM(lcl_hitm);
1815                 break;
1816         case DISPLAY_RMT:
1817                 FILTER_HITM(rmt_hitm);
1818                 break;
1819         case DISPLAY_TOT:
1820                 FILTER_HITM(tot_hitm);
1821         default:
1822                 break;
1823         };
1824
1825 #undef FILTER_HITM
1826
1827         return he->filtered == 0;
1828 }
1829
1830 static inline int valid_hitm_or_store(struct hist_entry *he)
1831 {
1832         struct c2c_hist_entry *c2c_he;
1833         bool has_hitm;
1834
1835         c2c_he = container_of(he, struct c2c_hist_entry, he);
1836         has_hitm = c2c.display == DISPLAY_TOT ? c2c_he->stats.tot_hitm :
1837                    c2c.display == DISPLAY_LCL ? c2c_he->stats.lcl_hitm :
1838                                                 c2c_he->stats.rmt_hitm;
1839         return has_hitm || c2c_he->stats.store;
1840 }
1841
1842 static void calc_width(struct hist_entry *he)
1843 {
1844         struct c2c_hists *c2c_hists;
1845
1846         c2c_hists = container_of(he->hists, struct c2c_hists, hists);
1847         hists__calc_col_len(&c2c_hists->hists, he);
1848 }
1849
1850 static int filter_cb(struct hist_entry *he)
1851 {
1852         if (c2c.show_src && !he->srcline)
1853                 he->srcline = hist_entry__get_srcline(he);
1854
1855         calc_width(he);
1856
1857         if (!valid_hitm_or_store(he))
1858                 he->filtered = HIST_FILTER__C2C;
1859
1860         return 0;
1861 }
1862
1863 static int resort_cl_cb(struct hist_entry *he)
1864 {
1865         struct c2c_hist_entry *c2c_he;
1866         struct c2c_hists *c2c_hists;
1867         bool display = he__display(he, &c2c.hitm_stats);
1868
1869         c2c_he = container_of(he, struct c2c_hist_entry, he);
1870         c2c_hists = c2c_he->hists;
1871
1872         calc_width(he);
1873
1874         if (display && c2c_hists) {
1875                 static unsigned int idx;
1876
1877                 c2c_he->cacheline_idx = idx++;
1878
1879                 c2c_hists__reinit(c2c_hists, c2c.cl_output, c2c.cl_resort);
1880
1881                 hists__collapse_resort(&c2c_hists->hists, NULL);
1882                 hists__output_resort_cb(&c2c_hists->hists, NULL, filter_cb);
1883         }
1884
1885         return 0;
1886 }
1887
1888 static void setup_nodes_header(void)
1889 {
1890         dim_node.header = header_node[c2c.node_info];
1891 }
1892
1893 static int setup_nodes(struct perf_session *session)
1894 {
1895         struct numa_node *n;
1896         unsigned long **nodes;
1897         int node, cpu;
1898         int *cpu2node;
1899
1900         if (c2c.node_info > 2)
1901                 c2c.node_info = 2;
1902
1903         c2c.nodes_cnt = session->header.env.nr_numa_nodes;
1904         c2c.cpus_cnt  = session->header.env.nr_cpus_online;
1905
1906         n = session->header.env.numa_nodes;
1907         if (!n)
1908                 return -EINVAL;
1909
1910         nodes = zalloc(sizeof(unsigned long *) * c2c.nodes_cnt);
1911         if (!nodes)
1912                 return -ENOMEM;
1913
1914         c2c.nodes = nodes;
1915
1916         cpu2node = zalloc(sizeof(int) * c2c.cpus_cnt);
1917         if (!cpu2node)
1918                 return -ENOMEM;
1919
1920         for (cpu = 0; cpu < c2c.cpus_cnt; cpu++)
1921                 cpu2node[cpu] = -1;
1922
1923         c2c.cpu2node = cpu2node;
1924
1925         for (node = 0; node < c2c.nodes_cnt; node++) {
1926                 struct cpu_map *map = n[node].map;
1927                 unsigned long *set;
1928
1929                 set = bitmap_alloc(c2c.cpus_cnt);
1930                 if (!set)
1931                         return -ENOMEM;
1932
1933                 for (cpu = 0; cpu < map->nr; cpu++) {
1934                         set_bit(map->map[cpu], set);
1935
1936                         if (WARN_ONCE(cpu2node[map->map[cpu]] != -1, "node/cpu topology bug"))
1937                                 return -EINVAL;
1938
1939                         cpu2node[map->map[cpu]] = node;
1940                 }
1941
1942                 nodes[node] = set;
1943         }
1944
1945         setup_nodes_header();
1946         return 0;
1947 }
1948
1949 #define HAS_HITMS(__h) ((__h)->stats.lcl_hitm || (__h)->stats.rmt_hitm)
1950
1951 static int resort_hitm_cb(struct hist_entry *he)
1952 {
1953         struct c2c_hist_entry *c2c_he;
1954         c2c_he = container_of(he, struct c2c_hist_entry, he);
1955
1956         if (HAS_HITMS(c2c_he)) {
1957                 c2c.shared_clines++;
1958                 c2c_add_stats(&c2c.hitm_stats, &c2c_he->stats);
1959         }
1960
1961         return 0;
1962 }
1963
1964 static int hists__iterate_cb(struct hists *hists, hists__resort_cb_t cb)
1965 {
1966         struct rb_node *next = rb_first(&hists->entries);
1967         int ret = 0;
1968
1969         while (next) {
1970                 struct hist_entry *he;
1971
1972                 he = rb_entry(next, struct hist_entry, rb_node);
1973                 ret = cb(he);
1974                 if (ret)
1975                         break;
1976                 next = rb_next(&he->rb_node);
1977         }
1978
1979         return ret;
1980 }
1981
1982 static void print_c2c__display_stats(FILE *out)
1983 {
1984         int llc_misses;
1985         struct c2c_stats *stats = &c2c.hists.stats;
1986
1987         llc_misses = stats->lcl_dram +
1988                      stats->rmt_dram +
1989                      stats->rmt_hit +
1990                      stats->rmt_hitm;
1991
1992         fprintf(out, "=================================================\n");
1993         fprintf(out, "            Trace Event Information              \n");
1994         fprintf(out, "=================================================\n");
1995         fprintf(out, "  Total records                     : %10d\n", stats->nr_entries);
1996         fprintf(out, "  Locked Load/Store Operations      : %10d\n", stats->locks);
1997         fprintf(out, "  Load Operations                   : %10d\n", stats->load);
1998         fprintf(out, "  Loads - uncacheable               : %10d\n", stats->ld_uncache);
1999         fprintf(out, "  Loads - IO                        : %10d\n", stats->ld_io);
2000         fprintf(out, "  Loads - Miss                      : %10d\n", stats->ld_miss);
2001         fprintf(out, "  Loads - no mapping                : %10d\n", stats->ld_noadrs);
2002         fprintf(out, "  Load Fill Buffer Hit              : %10d\n", stats->ld_fbhit);
2003         fprintf(out, "  Load L1D hit                      : %10d\n", stats->ld_l1hit);
2004         fprintf(out, "  Load L2D hit                      : %10d\n", stats->ld_l2hit);
2005         fprintf(out, "  Load LLC hit                      : %10d\n", stats->ld_llchit + stats->lcl_hitm);
2006         fprintf(out, "  Load Local HITM                   : %10d\n", stats->lcl_hitm);
2007         fprintf(out, "  Load Remote HITM                  : %10d\n", stats->rmt_hitm);
2008         fprintf(out, "  Load Remote HIT                   : %10d\n", stats->rmt_hit);
2009         fprintf(out, "  Load Local DRAM                   : %10d\n", stats->lcl_dram);
2010         fprintf(out, "  Load Remote DRAM                  : %10d\n", stats->rmt_dram);
2011         fprintf(out, "  Load MESI State Exclusive         : %10d\n", stats->ld_excl);
2012         fprintf(out, "  Load MESI State Shared            : %10d\n", stats->ld_shared);
2013         fprintf(out, "  Load LLC Misses                   : %10d\n", llc_misses);
2014         fprintf(out, "  LLC Misses to Local DRAM          : %10.1f%%\n", ((double)stats->lcl_dram/(double)llc_misses) * 100.);
2015         fprintf(out, "  LLC Misses to Remote DRAM         : %10.1f%%\n", ((double)stats->rmt_dram/(double)llc_misses) * 100.);
2016         fprintf(out, "  LLC Misses to Remote cache (HIT)  : %10.1f%%\n", ((double)stats->rmt_hit /(double)llc_misses) * 100.);
2017         fprintf(out, "  LLC Misses to Remote cache (HITM) : %10.1f%%\n", ((double)stats->rmt_hitm/(double)llc_misses) * 100.);
2018         fprintf(out, "  Store Operations                  : %10d\n", stats->store);
2019         fprintf(out, "  Store - uncacheable               : %10d\n", stats->st_uncache);
2020         fprintf(out, "  Store - no mapping                : %10d\n", stats->st_noadrs);
2021         fprintf(out, "  Store L1D Hit                     : %10d\n", stats->st_l1hit);
2022         fprintf(out, "  Store L1D Miss                    : %10d\n", stats->st_l1miss);
2023         fprintf(out, "  No Page Map Rejects               : %10d\n", stats->nomap);
2024         fprintf(out, "  Unable to parse data source       : %10d\n", stats->noparse);
2025 }
2026
2027 static void print_shared_cacheline_info(FILE *out)
2028 {
2029         struct c2c_stats *stats = &c2c.hitm_stats;
2030         int hitm_cnt = stats->lcl_hitm + stats->rmt_hitm;
2031
2032         fprintf(out, "=================================================\n");
2033         fprintf(out, "    Global Shared Cache Line Event Information   \n");
2034         fprintf(out, "=================================================\n");
2035         fprintf(out, "  Total Shared Cache Lines          : %10d\n", c2c.shared_clines);
2036         fprintf(out, "  Load HITs on shared lines         : %10d\n", stats->load);
2037         fprintf(out, "  Fill Buffer Hits on shared lines  : %10d\n", stats->ld_fbhit);
2038         fprintf(out, "  L1D hits on shared lines          : %10d\n", stats->ld_l1hit);
2039         fprintf(out, "  L2D hits on shared lines          : %10d\n", stats->ld_l2hit);
2040         fprintf(out, "  LLC hits on shared lines          : %10d\n", stats->ld_llchit + stats->lcl_hitm);
2041         fprintf(out, "  Locked Access on shared lines     : %10d\n", stats->locks);
2042         fprintf(out, "  Store HITs on shared lines        : %10d\n", stats->store);
2043         fprintf(out, "  Store L1D hits on shared lines    : %10d\n", stats->st_l1hit);
2044         fprintf(out, "  Total Merged records              : %10d\n", hitm_cnt + stats->store);
2045 }
2046
2047 static void print_cacheline(struct c2c_hists *c2c_hists,
2048                             struct hist_entry *he_cl,
2049                             struct perf_hpp_list *hpp_list,
2050                             FILE *out)
2051 {
2052         char bf[1000];
2053         struct perf_hpp hpp = {
2054                 .buf            = bf,
2055                 .size           = 1000,
2056         };
2057         static bool once;
2058
2059         if (!once) {
2060                 hists__fprintf_headers(&c2c_hists->hists, out);
2061                 once = true;
2062         } else {
2063                 fprintf(out, "\n");
2064         }
2065
2066         fprintf(out, "  -------------------------------------------------------------\n");
2067         __hist_entry__snprintf(he_cl, &hpp, hpp_list);
2068         fprintf(out, "%s\n", bf);
2069         fprintf(out, "  -------------------------------------------------------------\n");
2070
2071         hists__fprintf(&c2c_hists->hists, false, 0, 0, 0, out, true);
2072 }
2073
2074 static void print_pareto(FILE *out)
2075 {
2076         struct perf_hpp_list hpp_list;
2077         struct rb_node *nd;
2078         int ret;
2079
2080         perf_hpp_list__init(&hpp_list);
2081         ret = hpp_list__parse(&hpp_list,
2082                                 "cl_num,"
2083                                 "cl_rmt_hitm,"
2084                                 "cl_lcl_hitm,"
2085                                 "cl_stores_l1hit,"
2086                                 "cl_stores_l1miss,"
2087                                 "dcacheline",
2088                                 NULL);
2089
2090         if (WARN_ONCE(ret, "failed to setup sort entries\n"))
2091                 return;
2092
2093         nd = rb_first(&c2c.hists.hists.entries);
2094
2095         for (; nd; nd = rb_next(nd)) {
2096                 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
2097                 struct c2c_hist_entry *c2c_he;
2098
2099                 if (he->filtered)
2100                         continue;
2101
2102                 c2c_he = container_of(he, struct c2c_hist_entry, he);
2103                 print_cacheline(c2c_he->hists, he, &hpp_list, out);
2104         }
2105 }
2106
2107 static void print_c2c_info(FILE *out, struct perf_session *session)
2108 {
2109         struct perf_evlist *evlist = session->evlist;
2110         struct perf_evsel *evsel;
2111         bool first = true;
2112
2113         fprintf(out, "=================================================\n");
2114         fprintf(out, "                 c2c details                     \n");
2115         fprintf(out, "=================================================\n");
2116
2117         evlist__for_each_entry(evlist, evsel) {
2118                 fprintf(out, "%-36s: %s\n", first ? "  Events" : "",
2119                         perf_evsel__name(evsel));
2120                 first = false;
2121         }
2122         fprintf(out, "  Cachelines sort on                : %s HITMs\n",
2123                 display_str[c2c.display]);
2124         fprintf(out, "  Cacheline data grouping           : %s\n", c2c.cl_sort);
2125 }
2126
2127 static void perf_c2c__hists_fprintf(FILE *out, struct perf_session *session)
2128 {
2129         setup_pager();
2130
2131         print_c2c__display_stats(out);
2132         fprintf(out, "\n");
2133         print_shared_cacheline_info(out);
2134         fprintf(out, "\n");
2135         print_c2c_info(out, session);
2136
2137         if (c2c.stats_only)
2138                 return;
2139
2140         fprintf(out, "\n");
2141         fprintf(out, "=================================================\n");
2142         fprintf(out, "           Shared Data Cache Line Table          \n");
2143         fprintf(out, "=================================================\n");
2144         fprintf(out, "#\n");
2145
2146         hists__fprintf(&c2c.hists.hists, true, 0, 0, 0, stdout, false);
2147
2148         fprintf(out, "\n");
2149         fprintf(out, "=================================================\n");
2150         fprintf(out, "      Shared Cache Line Distribution Pareto      \n");
2151         fprintf(out, "=================================================\n");
2152         fprintf(out, "#\n");
2153
2154         print_pareto(out);
2155 }
2156
2157 #ifdef HAVE_SLANG_SUPPORT
2158 static void c2c_browser__update_nr_entries(struct hist_browser *hb)
2159 {
2160         u64 nr_entries = 0;
2161         struct rb_node *nd = rb_first(&hb->hists->entries);
2162
2163         while (nd) {
2164                 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
2165
2166                 if (!he->filtered)
2167                         nr_entries++;
2168
2169                 nd = rb_next(nd);
2170         }
2171
2172         hb->nr_non_filtered_entries = nr_entries;
2173 }
2174
2175 struct c2c_cacheline_browser {
2176         struct hist_browser      hb;
2177         struct hist_entry       *he;
2178 };
2179
2180 static int
2181 perf_c2c_cacheline_browser__title(struct hist_browser *browser,
2182                                   char *bf, size_t size)
2183 {
2184         struct c2c_cacheline_browser *cl_browser;
2185         struct hist_entry *he;
2186         uint64_t addr = 0;
2187
2188         cl_browser = container_of(browser, struct c2c_cacheline_browser, hb);
2189         he = cl_browser->he;
2190
2191         if (he->mem_info)
2192                 addr = cl_address(he->mem_info->daddr.addr);
2193
2194         scnprintf(bf, size, "Cacheline 0x%lx", addr);
2195         return 0;
2196 }
2197
2198 static struct c2c_cacheline_browser*
2199 c2c_cacheline_browser__new(struct hists *hists, struct hist_entry *he)
2200 {
2201         struct c2c_cacheline_browser *browser;
2202
2203         browser = zalloc(sizeof(*browser));
2204         if (browser) {
2205                 hist_browser__init(&browser->hb, hists);
2206                 browser->hb.c2c_filter  = true;
2207                 browser->hb.title       = perf_c2c_cacheline_browser__title;
2208                 browser->he             = he;
2209         }
2210
2211         return browser;
2212 }
2213
2214 static int perf_c2c__browse_cacheline(struct hist_entry *he)
2215 {
2216         struct c2c_hist_entry *c2c_he;
2217         struct c2c_hists *c2c_hists;
2218         struct c2c_cacheline_browser *cl_browser;
2219         struct hist_browser *browser;
2220         int key = -1;
2221         const char help[] =
2222         " ENTER         Togle callchains (if present) \n"
2223         " n             Togle Node details info \n"
2224         " s             Togle full lenght of symbol and source line columns \n"
2225         " q             Return back to cacheline list \n";
2226
2227         /* Display compact version first. */
2228         c2c.symbol_full = false;
2229
2230         c2c_he = container_of(he, struct c2c_hist_entry, he);
2231         c2c_hists = c2c_he->hists;
2232
2233         cl_browser = c2c_cacheline_browser__new(&c2c_hists->hists, he);
2234         if (cl_browser == NULL)
2235                 return -1;
2236
2237         browser = &cl_browser->hb;
2238
2239         /* reset abort key so that it can get Ctrl-C as a key */
2240         SLang_reset_tty();
2241         SLang_init_tty(0, 0, 0);
2242
2243         c2c_browser__update_nr_entries(browser);
2244
2245         while (1) {
2246                 key = hist_browser__run(browser, "? - help");
2247
2248                 switch (key) {
2249                 case 's':
2250                         c2c.symbol_full = !c2c.symbol_full;
2251                         break;
2252                 case 'n':
2253                         c2c.node_info = (c2c.node_info + 1) % 3;
2254                         setup_nodes_header();
2255                         break;
2256                 case 'q':
2257                         goto out;
2258                 case '?':
2259                         ui_browser__help_window(&browser->b, help);
2260                         break;
2261                 default:
2262                         break;
2263                 }
2264         }
2265
2266 out:
2267         free(cl_browser);
2268         return 0;
2269 }
2270
2271 static int perf_c2c_browser__title(struct hist_browser *browser,
2272                                    char *bf, size_t size)
2273 {
2274         scnprintf(bf, size,
2275                   "Shared Data Cache Line Table     "
2276                   "(%lu entries, sorted on %s HITMs)",
2277                   browser->nr_non_filtered_entries,
2278                   display_str[c2c.display]);
2279         return 0;
2280 }
2281
2282 static struct hist_browser*
2283 perf_c2c_browser__new(struct hists *hists)
2284 {
2285         struct hist_browser *browser = hist_browser__new(hists);
2286
2287         if (browser) {
2288                 browser->title = perf_c2c_browser__title;
2289                 browser->c2c_filter = true;
2290         }
2291
2292         return browser;
2293 }
2294
2295 static int perf_c2c__hists_browse(struct hists *hists)
2296 {
2297         struct hist_browser *browser;
2298         int key = -1;
2299         const char help[] =
2300         " d             Display cacheline details \n"
2301         " ENTER         Togle callchains (if present) \n"
2302         " q             Quit \n";
2303
2304         browser = perf_c2c_browser__new(hists);
2305         if (browser == NULL)
2306                 return -1;
2307
2308         /* reset abort key so that it can get Ctrl-C as a key */
2309         SLang_reset_tty();
2310         SLang_init_tty(0, 0, 0);
2311
2312         c2c_browser__update_nr_entries(browser);
2313
2314         while (1) {
2315                 key = hist_browser__run(browser, "? - help");
2316
2317                 switch (key) {
2318                 case 'q':
2319                         goto out;
2320                 case 'd':
2321                         perf_c2c__browse_cacheline(browser->he_selection);
2322                         break;
2323                 case '?':
2324                         ui_browser__help_window(&browser->b, help);
2325                         break;
2326                 default:
2327                         break;
2328                 }
2329         }
2330
2331 out:
2332         hist_browser__delete(browser);
2333         return 0;
2334 }
2335
2336 static void perf_c2c_display(struct perf_session *session)
2337 {
2338         if (c2c.use_stdio)
2339                 perf_c2c__hists_fprintf(stdout, session);
2340         else
2341                 perf_c2c__hists_browse(&c2c.hists.hists);
2342 }
2343 #else
2344 static void perf_c2c_display(struct perf_session *session)
2345 {
2346         use_browser = 0;
2347         perf_c2c__hists_fprintf(stdout, session);
2348 }
2349 #endif /* HAVE_SLANG_SUPPORT */
2350
2351 static void ui_quirks(void)
2352 {
2353         if (!c2c.use_stdio) {
2354                 dim_offset.width  = 5;
2355                 dim_offset.header = header_offset_tui;
2356         }
2357
2358         dim_percent_hitm.header = percent_hitm_header[c2c.display];
2359 }
2360
2361 #define CALLCHAIN_DEFAULT_OPT  "graph,0.5,caller,function,percent"
2362
2363 const char callchain_help[] = "Display call graph (stack chain/backtrace):\n\n"
2364                                 CALLCHAIN_REPORT_HELP
2365                                 "\n\t\t\t\tDefault: " CALLCHAIN_DEFAULT_OPT;
2366
2367 static int
2368 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
2369 {
2370         struct callchain_param *callchain = opt->value;
2371
2372         callchain->enabled = !unset;
2373         /*
2374          * --no-call-graph
2375          */
2376         if (unset) {
2377                 symbol_conf.use_callchain = false;
2378                 callchain->mode = CHAIN_NONE;
2379                 return 0;
2380         }
2381
2382         return parse_callchain_report_opt(arg);
2383 }
2384
2385 static int setup_callchain(struct perf_evlist *evlist)
2386 {
2387         u64 sample_type = perf_evlist__combined_sample_type(evlist);
2388         enum perf_call_graph_mode mode = CALLCHAIN_NONE;
2389
2390         if ((sample_type & PERF_SAMPLE_REGS_USER) &&
2391             (sample_type & PERF_SAMPLE_STACK_USER))
2392                 mode = CALLCHAIN_DWARF;
2393         else if (sample_type & PERF_SAMPLE_BRANCH_STACK)
2394                 mode = CALLCHAIN_LBR;
2395         else if (sample_type & PERF_SAMPLE_CALLCHAIN)
2396                 mode = CALLCHAIN_FP;
2397
2398         if (!callchain_param.enabled &&
2399             callchain_param.mode != CHAIN_NONE &&
2400             mode != CALLCHAIN_NONE) {
2401                 symbol_conf.use_callchain = true;
2402                 if (callchain_register_param(&callchain_param) < 0) {
2403                         ui__error("Can't register callchain params.\n");
2404                         return -EINVAL;
2405                 }
2406         }
2407
2408         callchain_param.record_mode = mode;
2409         callchain_param.min_percent = 0;
2410         return 0;
2411 }
2412
2413 static int setup_display(const char *str)
2414 {
2415         const char *display = str ?: "tot";
2416
2417         if (!strcmp(display, "tot"))
2418                 c2c.display = DISPLAY_TOT;
2419         else if (!strcmp(display, "rmt"))
2420                 c2c.display = DISPLAY_RMT;
2421         else if (!strcmp(display, "lcl"))
2422                 c2c.display = DISPLAY_LCL;
2423         else {
2424                 pr_err("failed: unknown display type: %s\n", str);
2425                 return -1;
2426         }
2427
2428         return 0;
2429 }
2430
2431 #define for_each_token(__tok, __buf, __sep, __tmp)              \
2432         for (__tok = strtok_r(__buf, __sep, &__tmp); __tok;     \
2433              __tok = strtok_r(NULL,  __sep, &__tmp))
2434
2435 static int build_cl_output(char *cl_sort, bool no_source)
2436 {
2437         char *tok, *tmp, *buf = strdup(cl_sort);
2438         bool add_pid   = false;
2439         bool add_tid   = false;
2440         bool add_iaddr = false;
2441         bool add_sym   = false;
2442         bool add_dso   = false;
2443         bool add_src   = false;
2444
2445         if (!buf)
2446                 return -ENOMEM;
2447
2448         for_each_token(tok, buf, ",", tmp) {
2449                 if (!strcmp(tok, "tid")) {
2450                         add_tid = true;
2451                 } else if (!strcmp(tok, "pid")) {
2452                         add_pid = true;
2453                 } else if (!strcmp(tok, "iaddr")) {
2454                         add_iaddr = true;
2455                         add_sym   = true;
2456                         add_dso   = true;
2457                         add_src   = no_source ? false : true;
2458                 } else if (!strcmp(tok, "dso")) {
2459                         add_dso = true;
2460                 } else if (strcmp(tok, "offset")) {
2461                         pr_err("unrecognized sort token: %s\n", tok);
2462                         return -EINVAL;
2463                 }
2464         }
2465
2466         if (asprintf(&c2c.cl_output,
2467                 "%s%s%s%s%s%s%s%s%s%s",
2468                 c2c.use_stdio ? "cl_num_empty," : "",
2469                 "percent_rmt_hitm,"
2470                 "percent_lcl_hitm,"
2471                 "percent_stores_l1hit,"
2472                 "percent_stores_l1miss,"
2473                 "offset,",
2474                 add_pid   ? "pid," : "",
2475                 add_tid   ? "tid," : "",
2476                 add_iaddr ? "iaddr," : "",
2477                 "mean_rmt,"
2478                 "mean_lcl,"
2479                 "mean_load,"
2480                 "cpucnt,",
2481                 add_sym ? "symbol," : "",
2482                 add_dso ? "dso," : "",
2483                 add_src ? "cl_srcline," : "",
2484                 "node") < 0)
2485                 return -ENOMEM;
2486
2487         c2c.show_src = add_src;
2488
2489         free(buf);
2490         return 0;
2491 }
2492
2493 static int setup_coalesce(const char *coalesce, bool no_source)
2494 {
2495         const char *c = coalesce ?: coalesce_default;
2496
2497         if (asprintf(&c2c.cl_sort, "offset,%s", c) < 0)
2498                 return -ENOMEM;
2499
2500         if (build_cl_output(c2c.cl_sort, no_source))
2501                 return -1;
2502
2503         if (asprintf(&c2c.cl_resort, "offset,%s",
2504                      c2c.display == DISPLAY_TOT ?
2505                      "tot_hitm" :
2506                      c2c.display == DISPLAY_RMT ?
2507                      "rmt_hitm,lcl_hitm" :
2508                      "lcl_hitm,rmt_hitm") < 0)
2509                 return -ENOMEM;
2510
2511         pr_debug("coalesce sort   fields: %s\n", c2c.cl_sort);
2512         pr_debug("coalesce resort fields: %s\n", c2c.cl_resort);
2513         pr_debug("coalesce output fields: %s\n", c2c.cl_output);
2514         return 0;
2515 }
2516
2517 static int perf_c2c__report(int argc, const char **argv)
2518 {
2519         struct perf_session *session;
2520         struct ui_progress prog;
2521         struct perf_data_file file = {
2522                 .mode = PERF_DATA_MODE_READ,
2523         };
2524         char callchain_default_opt[] = CALLCHAIN_DEFAULT_OPT;
2525         const char *display = NULL;
2526         const char *coalesce = NULL;
2527         bool no_source = false;
2528         const struct option options[] = {
2529         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
2530                    "file", "vmlinux pathname"),
2531         OPT_STRING('i', "input", &input_name, "file",
2532                    "the input file to process"),
2533         OPT_INCR('N', "node-info", &c2c.node_info,
2534                  "show extra node info in report (repeat for more info)"),
2535 #ifdef HAVE_SLANG_SUPPORT
2536         OPT_BOOLEAN(0, "stdio", &c2c.use_stdio, "Use the stdio interface"),
2537 #endif
2538         OPT_BOOLEAN(0, "stats", &c2c.stats_only,
2539                     "Use the stdio interface"),
2540         OPT_BOOLEAN(0, "full-symbols", &c2c.symbol_full,
2541                     "Display full length of symbols"),
2542         OPT_BOOLEAN(0, "no-source", &no_source,
2543                     "Do not display Source Line column"),
2544         OPT_BOOLEAN(0, "show-all", &c2c.show_all,
2545                     "Show all captured HITM lines."),
2546         OPT_CALLBACK_DEFAULT('g', "call-graph", &callchain_param,
2547                              "print_type,threshold[,print_limit],order,sort_key[,branch],value",
2548                              callchain_help, &parse_callchain_opt,
2549                              callchain_default_opt),
2550         OPT_STRING('d', "display", &display, "Switch HITM output type", "lcl,rmt"),
2551         OPT_STRING('c', "coalesce", &coalesce, "coalesce fields",
2552                    "coalesce fields: pid,tid,iaddr,dso"),
2553         OPT_BOOLEAN('f', "force", &symbol_conf.force, "don't complain, do it"),
2554         OPT_PARENT(c2c_options),
2555         OPT_END()
2556         };
2557         int err = 0;
2558
2559         argc = parse_options(argc, argv, options, report_c2c_usage,
2560                              PARSE_OPT_STOP_AT_NON_OPTION);
2561         if (argc)
2562                 usage_with_options(report_c2c_usage, options);
2563
2564         if (c2c.stats_only)
2565                 c2c.use_stdio = true;
2566
2567         if (!input_name || !strlen(input_name))
2568                 input_name = "perf.data";
2569
2570         file.path  = input_name;
2571         file.force = symbol_conf.force;
2572
2573         err = setup_display(display);
2574         if (err)
2575                 goto out;
2576
2577         err = setup_coalesce(coalesce, no_source);
2578         if (err) {
2579                 pr_debug("Failed to initialize hists\n");
2580                 goto out;
2581         }
2582
2583         err = c2c_hists__init(&c2c.hists, "dcacheline", 2);
2584         if (err) {
2585                 pr_debug("Failed to initialize hists\n");
2586                 goto out;
2587         }
2588
2589         session = perf_session__new(&file, 0, &c2c.tool);
2590         if (session == NULL) {
2591                 pr_debug("No memory for session\n");
2592                 goto out;
2593         }
2594
2595         err = setup_nodes(session);
2596         if (err) {
2597                 pr_err("Failed setup nodes\n");
2598                 goto out;
2599         }
2600
2601         err = setup_callchain(session->evlist);
2602         if (err)
2603                 goto out_session;
2604
2605         if (symbol__init(&session->header.env) < 0)
2606                 goto out_session;
2607
2608         /* No pipe support at the moment. */
2609         if (perf_data_file__is_pipe(session->file)) {
2610                 pr_debug("No pipe support at the moment.\n");
2611                 goto out_session;
2612         }
2613
2614         if (c2c.use_stdio)
2615                 use_browser = 0;
2616         else
2617                 use_browser = 1;
2618
2619         setup_browser(false);
2620
2621         err = perf_session__process_events(session);
2622         if (err) {
2623                 pr_err("failed to process sample\n");
2624                 goto out_session;
2625         }
2626
2627         c2c_hists__reinit(&c2c.hists,
2628                         "cl_idx,"
2629                         "dcacheline,"
2630                         "tot_recs,"
2631                         "percent_hitm,"
2632                         "tot_hitm,lcl_hitm,rmt_hitm,"
2633                         "stores,stores_l1hit,stores_l1miss,"
2634                         "dram_lcl,dram_rmt,"
2635                         "ld_llcmiss,"
2636                         "tot_loads,"
2637                         "ld_fbhit,ld_l1hit,ld_l2hit,"
2638                         "ld_lclhit,ld_rmthit",
2639                         c2c.display == DISPLAY_TOT ? "tot_hitm" :
2640                         c2c.display == DISPLAY_LCL ? "lcl_hitm" : "rmt_hitm"
2641                         );
2642
2643         ui_progress__init(&prog, c2c.hists.hists.nr_entries, "Sorting...");
2644
2645         hists__collapse_resort(&c2c.hists.hists, NULL);
2646         hists__output_resort_cb(&c2c.hists.hists, &prog, resort_hitm_cb);
2647         hists__iterate_cb(&c2c.hists.hists, resort_cl_cb);
2648
2649         ui_progress__finish();
2650
2651         ui_quirks();
2652
2653         perf_c2c_display(session);
2654
2655 out_session:
2656         perf_session__delete(session);
2657 out:
2658         return err;
2659 }
2660
2661 static int parse_record_events(const struct option *opt __maybe_unused,
2662                                const char *str, int unset __maybe_unused)
2663 {
2664         bool *event_set = (bool *) opt->value;
2665
2666         *event_set = true;
2667         return perf_mem_events__parse(str);
2668 }
2669
2670
2671 static const char * const __usage_record[] = {
2672         "perf c2c record [<options>] [<command>]",
2673         "perf c2c record [<options>] -- <command> [<options>]",
2674         NULL
2675 };
2676
2677 static const char * const *record_mem_usage = __usage_record;
2678
2679 static int perf_c2c__record(int argc, const char **argv)
2680 {
2681         int rec_argc, i = 0, j;
2682         const char **rec_argv;
2683         int ret;
2684         bool all_user = false, all_kernel = false;
2685         bool event_set = false;
2686         struct option options[] = {
2687         OPT_CALLBACK('e', "event", &event_set, "event",
2688                      "event selector. Use 'perf mem record -e list' to list available events",
2689                      parse_record_events),
2690         OPT_BOOLEAN('u', "all-user", &all_user, "collect only user level data"),
2691         OPT_BOOLEAN('k', "all-kernel", &all_kernel, "collect only kernel level data"),
2692         OPT_UINTEGER('l', "ldlat", &perf_mem_events__loads_ldlat, "setup mem-loads latency"),
2693         OPT_PARENT(c2c_options),
2694         OPT_END()
2695         };
2696
2697         if (perf_mem_events__init()) {
2698                 pr_err("failed: memory events not supported\n");
2699                 return -1;
2700         }
2701
2702         argc = parse_options(argc, argv, options, record_mem_usage,
2703                              PARSE_OPT_KEEP_UNKNOWN);
2704
2705         rec_argc = argc + 10; /* max number of arguments */
2706         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2707         if (!rec_argv)
2708                 return -1;
2709
2710         rec_argv[i++] = "record";
2711
2712         if (!event_set) {
2713                 perf_mem_events[PERF_MEM_EVENTS__LOAD].record  = true;
2714                 perf_mem_events[PERF_MEM_EVENTS__STORE].record = true;
2715         }
2716
2717         if (perf_mem_events[PERF_MEM_EVENTS__LOAD].record)
2718                 rec_argv[i++] = "-W";
2719
2720         rec_argv[i++] = "-d";
2721         rec_argv[i++] = "--sample-cpu";
2722
2723         for (j = 0; j < PERF_MEM_EVENTS__MAX; j++) {
2724                 if (!perf_mem_events[j].record)
2725                         continue;
2726
2727                 if (!perf_mem_events[j].supported) {
2728                         pr_err("failed: event '%s' not supported\n",
2729                                perf_mem_events[j].name);
2730                         return -1;
2731                 }
2732
2733                 rec_argv[i++] = "-e";
2734                 rec_argv[i++] = perf_mem_events__name(j);
2735         };
2736
2737         if (all_user)
2738                 rec_argv[i++] = "--all-user";
2739
2740         if (all_kernel)
2741                 rec_argv[i++] = "--all-kernel";
2742
2743         for (j = 0; j < argc; j++, i++)
2744                 rec_argv[i] = argv[j];
2745
2746         if (verbose > 0) {
2747                 pr_debug("calling: ");
2748
2749                 j = 0;
2750
2751                 while (rec_argv[j]) {
2752                         pr_debug("%s ", rec_argv[j]);
2753                         j++;
2754                 }
2755                 pr_debug("\n");
2756         }
2757
2758         ret = cmd_record(i, rec_argv, NULL);
2759         free(rec_argv);
2760         return ret;
2761 }
2762
2763 int cmd_c2c(int argc, const char **argv, const char *prefix __maybe_unused)
2764 {
2765         argc = parse_options(argc, argv, c2c_options, c2c_usage,
2766                              PARSE_OPT_STOP_AT_NON_OPTION);
2767
2768         if (!argc)
2769                 usage_with_options(c2c_usage, c2c_options);
2770
2771         if (!strncmp(argv[0], "rec", 3)) {
2772                 return perf_c2c__record(argc, argv);
2773         } else if (!strncmp(argv[0], "rep", 3)) {
2774                 return perf_c2c__report(argc, argv);
2775         } else {
2776                 usage_with_options(c2c_usage, c2c_options);
2777         }
2778
2779         return 0;
2780 }