tools/perf/builtin-c2c.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * This is rewrite of original c2c tool introduced in here:
   4  *   http://lwn.net/Articles/588866/
   5  *
   6  * The original tool was changed to fit in current perf state.
   7  *
   8  * Original authors:
   9  *   Don Zickus <dzickus@redhat.com>
  10  *   Dick Fowles <fowles@inreach.com>
  11  *   Joe Mario <jmario@redhat.com>
  12  */
  13 #include <errno.h>
  14 #include <inttypes.h>
  15 #include <linux/compiler.h>
  16 #include <linux/kernel.h>
  17 #include <linux/stringify.h>
  18 #include <asm/bug.h>
  19 #include <sys/param.h>
  20 #include "util.h"
  21 #include "debug.h"
  22 #include "builtin.h"
  23 #include <subcmd/parse-options.h>
  24 #include "mem-events.h"
  25 #include "session.h"
  26 #include "hist.h"
  27 #include "sort.h"
  28 #include "tool.h"
  29 #include "data.h"
  30 #include "sort.h"
  31 #include "event.h"
  32 #include "evlist.h"
  33 #include "evsel.h"
  34 #include <asm/bug.h>
  35 #include "ui/browsers/hists.h"
  36 #include "evlist.h"
  37 #include "thread.h"
  38
  39 struct c2c_hists {
  40         struct hists            hists;
  41         struct perf_hpp_list    list;
  42         struct c2c_stats        stats;
  43 };
  44
  45 struct compute_stats {
  46         struct stats             lcl_hitm;
  47         struct stats             rmt_hitm;
  48         struct stats             load;
  49 };
  50
  51 struct c2c_hist_entry {
  52         struct c2c_hists        *hists;
  53         struct c2c_stats         stats;
  54         unsigned long           *cpuset;
  55         struct c2c_stats        *node_stats;
  56         unsigned int             cacheline_idx;
  57
  58         struct compute_stats     cstats;
  59
  60         /*
  61          * must be at the end,
  62          * because of its callchain dynamic entry
  63          */
  64         struct hist_entry       he;
  65 };
  66
  67 static char const *coalesce_default = "pid,iaddr";
  68
  69 struct perf_c2c {
  70         struct perf_tool        tool;
  71         struct c2c_hists        hists;
  72
  73         unsigned long           **nodes;
  74         int                      nodes_cnt;
  75         int                      cpus_cnt;
  76         int                     *cpu2node;
  77         int                      node_info;
  78
  79         bool                     show_src;
  80         bool                     show_all;
  81         bool                     use_stdio;
  82         bool                     stats_only;
  83         bool                     symbol_full;
  84
  85         /* HITM shared clines stats */
  86         struct c2c_stats        hitm_stats;
  87         int                     shared_clines;
  88
  89         int                      display;
  90
  91         const char              *coalesce;
  92         char                    *cl_sort;
  93         char                    *cl_resort;
  94         char                    *cl_output;
  95 };
  96
  97 enum {
  98         DISPLAY_LCL,
  99         DISPLAY_RMT,
 100         DISPLAY_TOT,
 101         DISPLAY_MAX,
 102 };
 103
 104 static const char *display_str[DISPLAY_MAX] = {
 105         [DISPLAY_LCL] = "Local",
 106         [DISPLAY_RMT] = "Remote",
 107         [DISPLAY_TOT] = "Total",
 108 };
 109
 110 static const struct option c2c_options[] = {
 111         OPT_INCR('v', "verbose", &verbose, "be more verbose (show counter open errors, etc)"),
 112         OPT_END()
 113 };
 114
 115 static struct perf_c2c c2c;
 116
 117 static void *c2c_he_zalloc(size_t size)
 118 {
 119         struct c2c_hist_entry *c2c_he;
 120
 121         c2c_he = zalloc(size + sizeof(*c2c_he));
 122         if (!c2c_he)
 123                 return NULL;
 124
 125         c2c_he->cpuset = bitmap_alloc(c2c.cpus_cnt);
 126         if (!c2c_he->cpuset)
 127                 return NULL;
 128
 129         c2c_he->node_stats = zalloc(c2c.nodes_cnt * sizeof(*c2c_he->node_stats));
 130         if (!c2c_he->node_stats)
 131                 return NULL;
 132
 133         init_stats(&c2c_he->cstats.lcl_hitm);
 134         init_stats(&c2c_he->cstats.rmt_hitm);
 135         init_stats(&c2c_he->cstats.load);
 136
 137         return &c2c_he->he;
 138 }
 139
 140 static void c2c_he_free(void *he)
 141 {
 142         struct c2c_hist_entry *c2c_he;
 143
 144         c2c_he = container_of(he, struct c2c_hist_entry, he);
 145         if (c2c_he->hists) {
 146                 hists__delete_entries(&c2c_he->hists->hists);
 147                 free(c2c_he->hists);
 148         }
 149
 150         free(c2c_he->cpuset);
 151         free(c2c_he->node_stats);
 152         free(c2c_he);
 153 }
 154
 155 static struct hist_entry_ops c2c_entry_ops = {
 156         .new    = c2c_he_zalloc,
 157         .free   = c2c_he_free,
 158 };
 159
 160 static int c2c_hists__init(struct c2c_hists *hists,
 161                            const char *sort,
 162                            int nr_header_lines);
 163
 164 static struct c2c_hists*
 165 he__get_c2c_hists(struct hist_entry *he,
 166                   const char *sort,
 167                   int nr_header_lines)
 168 {
 169         struct c2c_hist_entry *c2c_he;
 170         struct c2c_hists *hists;
 171         int ret;
 172
 173         c2c_he = container_of(he, struct c2c_hist_entry, he);
 174         if (c2c_he->hists)
 175                 return c2c_he->hists;
 176
 177         hists = c2c_he->hists = zalloc(sizeof(*hists));
 178         if (!hists)
 179                 return NULL;
 180
 181         ret = c2c_hists__init(hists, sort, nr_header_lines);
 182         if (ret) {
 183                 free(hists);
 184                 return NULL;
 185         }
 186
 187         return hists;
 188 }
 189
 190 static void c2c_he__set_cpu(struct c2c_hist_entry *c2c_he,
 191                             struct perf_sample *sample)
 192 {
 193         if (WARN_ONCE(sample->cpu == (unsigned int) -1,
 194                       "WARNING: no sample cpu value"))
 195                 return;
 196
 197         set_bit(sample->cpu, c2c_he->cpuset);
 198 }
 199
 200 static void compute_stats(struct c2c_hist_entry *c2c_he,
 201                           struct c2c_stats *stats,
 202                           u64 weight)
 203 {
 204         struct compute_stats *cstats = &c2c_he->cstats;
 205
 206         if (stats->rmt_hitm)
 207                 update_stats(&cstats->rmt_hitm, weight);
 208         else if (stats->lcl_hitm)
 209                 update_stats(&cstats->lcl_hitm, weight);
 210         else if (stats->load)
 211                 update_stats(&cstats->load, weight);
 212 }
 213
 214 static int process_sample_event(struct perf_tool *tool __maybe_unused,
 215                                 union perf_event *event,
 216                                 struct perf_sample *sample,
 217                                 struct perf_evsel *evsel,
 218                                 struct machine *machine)
 219 {
 220         struct c2c_hists *c2c_hists = &c2c.hists;
 221         struct c2c_hist_entry *c2c_he;
 222         struct c2c_stats stats = { .nr_entries = 0, };
 223         struct hist_entry *he;
 224         struct addr_location al;
 225         struct mem_info *mi, *mi_dup;
 226         int ret;
 227
 228         if (machine__resolve(machine, &al, sample) < 0) {
 229                 pr_debug("problem processing %d event, skipping it.\n",
 230                          event->header.type);
 231                 return -1;
 232         }
 233
 234         ret = sample__resolve_callchain(sample, &callchain_cursor, NULL,
 235                                         evsel, &al, sysctl_perf_event_max_stack);
 236         if (ret)
 237                 goto out;
 238
 239         mi = sample__resolve_mem(sample, &al);
 240         if (mi == NULL)
 241                 return -ENOMEM;
 242
 243         mi_dup = memdup(mi, sizeof(*mi));
 244         if (!mi_dup)
 245                 goto free_mi;
 246
 247         c2c_decode_stats(&stats, mi);
 248
 249         he = hists__add_entry_ops(&c2c_hists->hists, &c2c_entry_ops,
 250                                   &al, NULL, NULL, mi,
 251                                   sample, true);
 252         if (he == NULL)
 253                 goto free_mi_dup;
 254
 255         c2c_he = container_of(he, struct c2c_hist_entry, he);
 256         c2c_add_stats(&c2c_he->stats, &stats);
 257         c2c_add_stats(&c2c_hists->stats, &stats);
 258
 259         c2c_he__set_cpu(c2c_he, sample);
 260
 261         hists__inc_nr_samples(&c2c_hists->hists, he->filtered);
 262         ret = hist_entry__append_callchain(he, sample);
 263
 264         if (!ret) {
 265                 /*
 266                  * There's already been warning about missing
 267                  * sample's cpu value. Let's account all to
 268                  * node 0 in this case, without any further
 269                  * warning.
 270                  *
 271                  * Doing node stats only for single callchain data.
 272                  */
 273                 int cpu = sample->cpu == (unsigned int) -1 ? 0 : sample->cpu;
 274                 int node = c2c.cpu2node[cpu];
 275
 276                 mi = mi_dup;
 277
 278                 mi_dup = memdup(mi, sizeof(*mi));
 279                 if (!mi_dup)
 280                         goto free_mi;
 281
 282                 c2c_hists = he__get_c2c_hists(he, c2c.cl_sort, 2);
 283                 if (!c2c_hists)
 284                         goto free_mi_dup;
 285
 286                 he = hists__add_entry_ops(&c2c_hists->hists, &c2c_entry_ops,
 287                                           &al, NULL, NULL, mi,
 288                                           sample, true);
 289                 if (he == NULL)
 290                         goto free_mi_dup;
 291
 292                 c2c_he = container_of(he, struct c2c_hist_entry, he);
 293                 c2c_add_stats(&c2c_he->stats, &stats);
 294                 c2c_add_stats(&c2c_hists->stats, &stats);
 295                 c2c_add_stats(&c2c_he->node_stats[node], &stats);
 296
 297                 compute_stats(c2c_he, &stats, sample->weight);
 298
 299                 c2c_he__set_cpu(c2c_he, sample);
 300
 301                 hists__inc_nr_samples(&c2c_hists->hists, he->filtered);
 302                 ret = hist_entry__append_callchain(he, sample);
 303         }
 304
 305 out:
 306         addr_location__put(&al);
 307         return ret;
 308
 309 free_mi_dup:
 310         free(mi_dup);
 311 free_mi:
 312         free(mi);
 313         ret = -ENOMEM;
 314         goto out;
 315 }
 316
 317 static struct perf_c2c c2c = {
 318         .tool = {
 319                 .sample         = process_sample_event,
 320                 .mmap           = perf_event__process_mmap,
 321                 .mmap2          = perf_event__process_mmap2,
 322                 .comm           = perf_event__process_comm,
 323                 .exit           = perf_event__process_exit,
 324                 .fork           = perf_event__process_fork,
 325                 .lost           = perf_event__process_lost,
 326                 .ordered_events = true,
 327                 .ordering_requires_timestamps = true,
 328         },
 329 };
 330
 331 static const char * const c2c_usage[] = {
 332         "perf c2c {record|report}",
 333         NULL
 334 };
 335
 336 static const char * const __usage_report[] = {
 337         "perf c2c report",
 338         NULL
 339 };
 340
 341 static const char * const *report_c2c_usage = __usage_report;
 342
 343 #define C2C_HEADER_MAX 2
 344
 345 struct c2c_header {
 346         struct {
 347                 const char *text;
 348                 int         span;
 349         } line[C2C_HEADER_MAX];
 350 };
 351
 352 struct c2c_dimension {
 353         struct c2c_header        header;
 354         const char              *name;
 355         int                      width;
 356         struct sort_entry       *se;
 357
 358         int64_t (*cmp)(struct perf_hpp_fmt *fmt,
 359                        struct hist_entry *, struct hist_entry *);
 360         int   (*entry)(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 361                        struct hist_entry *he);
 362         int   (*color)(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 363                        struct hist_entry *he);
 364 };
 365
 366 struct c2c_fmt {
 367         struct perf_hpp_fmt      fmt;
 368         struct c2c_dimension    *dim;
 369 };
 370
 371 #define SYMBOL_WIDTH 30
 372
 373 static struct c2c_dimension dim_symbol;
 374 static struct c2c_dimension dim_srcline;
 375
 376 static int symbol_width(struct hists *hists, struct sort_entry *se)
 377 {
 378         int width = hists__col_len(hists, se->se_width_idx);
 379
 380         if (!c2c.symbol_full)
 381                 width = MIN(width, SYMBOL_WIDTH);
 382
 383         return width;
 384 }
 385
 386 static int c2c_width(struct perf_hpp_fmt *fmt,
 387                      struct perf_hpp *hpp __maybe_unused,
 388                      struct hists *hists)
 389 {
 390         struct c2c_fmt *c2c_fmt;
 391         struct c2c_dimension *dim;
 392
 393         c2c_fmt = container_of(fmt, struct c2c_fmt, fmt);
 394         dim = c2c_fmt->dim;
 395
 396         if (dim == &dim_symbol || dim == &dim_srcline)
 397                 return symbol_width(hists, dim->se);
 398
 399         return dim->se ? hists__col_len(hists, dim->se->se_width_idx) :
 400                          c2c_fmt->dim->width;
 401 }
 402
 403 static int c2c_header(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 404                       struct hists *hists, int line, int *span)
 405 {
 406         struct perf_hpp_list *hpp_list = hists->hpp_list;
 407         struct c2c_fmt *c2c_fmt;
 408         struct c2c_dimension *dim;
 409         const char *text = NULL;
 410         int width = c2c_width(fmt, hpp, hists);
 411
 412         c2c_fmt = container_of(fmt, struct c2c_fmt, fmt);
 413         dim = c2c_fmt->dim;
 414
 415         if (dim->se) {
 416                 text = dim->header.line[line].text;
 417                 /* Use the last line from sort_entry if not defined. */
 418                 if (!text && (line == hpp_list->nr_header_lines - 1))
 419                         text = dim->se->se_header;
 420         } else {
 421                 text = dim->header.line[line].text;
 422
 423                 if (*span) {
 424                         (*span)--;
 425                         return 0;
 426                 } else {
 427                         *span = dim->header.line[line].span;
 428                 }
 429         }
 430
 431         if (text == NULL)
 432                 text = "";
 433
 434         return scnprintf(hpp->buf, hpp->size, "%*s", width, text);
 435 }
 436
 437 #define HEX_STR(__s, __v)                               \
 438 ({                                                      \
 439         scnprintf(__s, sizeof(__s), "0x%" PRIx64, __v); \
 440         __s;                                            \
 441 })
 442
 443 static int64_t
 444 dcacheline_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 445                struct hist_entry *left, struct hist_entry *right)
 446 {
 447         return sort__dcacheline_cmp(left, right);
 448 }
 449
 450 static int dcacheline_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 451                             struct hist_entry *he)
 452 {
 453         uint64_t addr = 0;
 454         int width = c2c_width(fmt, hpp, he->hists);
 455         char buf[20];
 456
 457         if (he->mem_info)
 458                 addr = cl_address(he->mem_info->daddr.addr);
 459
 460         return scnprintf(hpp->buf, hpp->size, "%*s", width, HEX_STR(buf, addr));
 461 }
 462
 463 static int offset_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 464                         struct hist_entry *he)
 465 {
 466         uint64_t addr = 0;
 467         int width = c2c_width(fmt, hpp, he->hists);
 468         char buf[20];
 469
 470         if (he->mem_info)
 471                 addr = cl_offset(he->mem_info->daddr.al_addr);
 472
 473         return scnprintf(hpp->buf, hpp->size, "%*s", width, HEX_STR(buf, addr));
 474 }
 475
 476 static int64_t
 477 offset_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 478            struct hist_entry *left, struct hist_entry *right)
 479 {
 480         uint64_t l = 0, r = 0;
 481
 482         if (left->mem_info)
 483                 l = cl_offset(left->mem_info->daddr.addr);
 484         if (right->mem_info)
 485                 r = cl_offset(right->mem_info->daddr.addr);
 486
 487         return (int64_t)(r - l);
 488 }
 489
 490 static int
 491 iaddr_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 492             struct hist_entry *he)
 493 {
 494         uint64_t addr = 0;
 495         int width = c2c_width(fmt, hpp, he->hists);
 496         char buf[20];
 497
 498         if (he->mem_info)
 499                 addr = he->mem_info->iaddr.addr;
 500
 501         return scnprintf(hpp->buf, hpp->size, "%*s", width, HEX_STR(buf, addr));
 502 }
 503
 504 static int64_t
 505 iaddr_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 506           struct hist_entry *left, struct hist_entry *right)
 507 {
 508         return sort__iaddr_cmp(left, right);
 509 }
 510
 511 static int
 512 tot_hitm_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 513                struct hist_entry *he)
 514 {
 515         struct c2c_hist_entry *c2c_he;
 516         int width = c2c_width(fmt, hpp, he->hists);
 517         unsigned int tot_hitm;
 518
 519         c2c_he = container_of(he, struct c2c_hist_entry, he);
 520         tot_hitm = c2c_he->stats.lcl_hitm + c2c_he->stats.rmt_hitm;
 521
 522         return scnprintf(hpp->buf, hpp->size, "%*u", width, tot_hitm);
 523 }
 524
 525 static int64_t
 526 tot_hitm_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 527              struct hist_entry *left, struct hist_entry *right)
 528 {
 529         struct c2c_hist_entry *c2c_left;
 530         struct c2c_hist_entry *c2c_right;
 531         unsigned int tot_hitm_left;
 532         unsigned int tot_hitm_right;
 533
 534         c2c_left  = container_of(left, struct c2c_hist_entry, he);
 535         c2c_right = container_of(right, struct c2c_hist_entry, he);
 536
 537         tot_hitm_left  = c2c_left->stats.lcl_hitm + c2c_left->stats.rmt_hitm;
 538         tot_hitm_right = c2c_right->stats.lcl_hitm + c2c_right->stats.rmt_hitm;
 539
 540         return tot_hitm_left - tot_hitm_right;
 541 }
 542
 543 #define STAT_FN_ENTRY(__f)                                      \
 544 static int                                                      \
 545 __f ## _entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,   \
 546               struct hist_entry *he)                            \
 547 {                                                               \
 548         struct c2c_hist_entry *c2c_he;                          \
 549         int width = c2c_width(fmt, hpp, he->hists);             \
 550                                                                 \
 551         c2c_he = container_of(he, struct c2c_hist_entry, he);   \
 552         return scnprintf(hpp->buf, hpp->size, "%*u", width,     \
 553                          c2c_he->stats.__f);                    \
 554 }
 555
 556 #define STAT_FN_CMP(__f)                                                \
 557 static int64_t                                                          \
 558 __f ## _cmp(struct perf_hpp_fmt *fmt __maybe_unused,                    \
 559             struct hist_entry *left, struct hist_entry *right)          \
 560 {                                                                       \
 561         struct c2c_hist_entry *c2c_left, *c2c_right;                    \
 562                                                                         \
 563         c2c_left  = container_of(left, struct c2c_hist_entry, he);      \
 564         c2c_right = container_of(right, struct c2c_hist_entry, he);     \
 565         return c2c_left->stats.__f - c2c_right->stats.__f;              \
 566 }
 567
 568 #define STAT_FN(__f)            \
 569         STAT_FN_ENTRY(__f)      \
 570         STAT_FN_CMP(__f)
 571
 572 STAT_FN(rmt_hitm)
 573 STAT_FN(lcl_hitm)
 574 STAT_FN(store)
 575 STAT_FN(st_l1hit)
 576 STAT_FN(st_l1miss)
 577 STAT_FN(ld_fbhit)
 578 STAT_FN(ld_l1hit)
 579 STAT_FN(ld_l2hit)
 580 STAT_FN(ld_llchit)
 581 STAT_FN(rmt_hit)
 582
 583 static uint64_t llc_miss(struct c2c_stats *stats)
 584 {
 585         uint64_t llcmiss;
 586
 587         llcmiss = stats->lcl_dram +
 588                   stats->rmt_dram +
 589                   stats->rmt_hitm +
 590                   stats->rmt_hit;
 591
 592         return llcmiss;
 593 }
 594
 595 static int
 596 ld_llcmiss_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 597                  struct hist_entry *he)
 598 {
 599         struct c2c_hist_entry *c2c_he;
 600         int width = c2c_width(fmt, hpp, he->hists);
 601
 602         c2c_he = container_of(he, struct c2c_hist_entry, he);
 603
 604         return scnprintf(hpp->buf, hpp->size, "%*lu", width,
 605                          llc_miss(&c2c_he->stats));
 606 }
 607
 608 static int64_t
 609 ld_llcmiss_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 610                struct hist_entry *left, struct hist_entry *right)
 611 {
 612         struct c2c_hist_entry *c2c_left;
 613         struct c2c_hist_entry *c2c_right;
 614
 615         c2c_left  = container_of(left, struct c2c_hist_entry, he);
 616         c2c_right = container_of(right, struct c2c_hist_entry, he);
 617
 618         return llc_miss(&c2c_left->stats) - llc_miss(&c2c_right->stats);
 619 }
 620
 621 static uint64_t total_records(struct c2c_stats *stats)
 622 {
 623         uint64_t lclmiss, ldcnt, total;
 624
 625         lclmiss  = stats->lcl_dram +
 626                    stats->rmt_dram +
 627                    stats->rmt_hitm +
 628                    stats->rmt_hit;
 629
 630         ldcnt    = lclmiss +
 631                    stats->ld_fbhit +
 632                    stats->ld_l1hit +
 633                    stats->ld_l2hit +
 634                    stats->ld_llchit +
 635                    stats->lcl_hitm;
 636
 637         total    = ldcnt +
 638                    stats->st_l1hit +
 639                    stats->st_l1miss;
 640
 641         return total;
 642 }
 643
 644 static int
 645 tot_recs_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 646                 struct hist_entry *he)
 647 {
 648         struct c2c_hist_entry *c2c_he;
 649         int width = c2c_width(fmt, hpp, he->hists);
 650         uint64_t tot_recs;
 651
 652         c2c_he = container_of(he, struct c2c_hist_entry, he);
 653         tot_recs = total_records(&c2c_he->stats);
 654
 655         return scnprintf(hpp->buf, hpp->size, "%*" PRIu64, width, tot_recs);
 656 }
 657
 658 static int64_t
 659 tot_recs_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 660              struct hist_entry *left, struct hist_entry *right)
 661 {
 662         struct c2c_hist_entry *c2c_left;
 663         struct c2c_hist_entry *c2c_right;
 664         uint64_t tot_recs_left;
 665         uint64_t tot_recs_right;
 666
 667         c2c_left  = container_of(left, struct c2c_hist_entry, he);
 668         c2c_right = container_of(right, struct c2c_hist_entry, he);
 669
 670         tot_recs_left  = total_records(&c2c_left->stats);
 671         tot_recs_right = total_records(&c2c_right->stats);
 672
 673         return tot_recs_left - tot_recs_right;
 674 }
 675
 676 static uint64_t total_loads(struct c2c_stats *stats)
 677 {
 678         uint64_t lclmiss, ldcnt;
 679
 680         lclmiss  = stats->lcl_dram +
 681                    stats->rmt_dram +
 682                    stats->rmt_hitm +
 683                    stats->rmt_hit;
 684
 685         ldcnt    = lclmiss +
 686                    stats->ld_fbhit +
 687                    stats->ld_l1hit +
 688                    stats->ld_l2hit +
 689                    stats->ld_llchit +
 690                    stats->lcl_hitm;
 691
 692         return ldcnt;
 693 }
 694
 695 static int
 696 tot_loads_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 697                 struct hist_entry *he)
 698 {
 699         struct c2c_hist_entry *c2c_he;
 700         int width = c2c_width(fmt, hpp, he->hists);
 701         uint64_t tot_recs;
 702
 703         c2c_he = container_of(he, struct c2c_hist_entry, he);
 704         tot_recs = total_loads(&c2c_he->stats);
 705
 706         return scnprintf(hpp->buf, hpp->size, "%*" PRIu64, width, tot_recs);
 707 }
 708
 709 static int64_t
 710 tot_loads_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 711               struct hist_entry *left, struct hist_entry *right)
 712 {
 713         struct c2c_hist_entry *c2c_left;
 714         struct c2c_hist_entry *c2c_right;
 715         uint64_t tot_recs_left;
 716         uint64_t tot_recs_right;
 717
 718         c2c_left  = container_of(left, struct c2c_hist_entry, he);
 719         c2c_right = container_of(right, struct c2c_hist_entry, he);
 720
 721         tot_recs_left  = total_loads(&c2c_left->stats);
 722         tot_recs_right = total_loads(&c2c_right->stats);
 723
 724         return tot_recs_left - tot_recs_right;
 725 }
 726
 727 typedef double (get_percent_cb)(struct c2c_hist_entry *);
 728
 729 static int
 730 percent_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 731               struct hist_entry *he, get_percent_cb get_percent)
 732 {
 733         struct c2c_hist_entry *c2c_he;
 734         int width = c2c_width(fmt, hpp, he->hists);
 735         double per;
 736
 737         c2c_he = container_of(he, struct c2c_hist_entry, he);
 738         per = get_percent(c2c_he);
 739
 740 #ifdef HAVE_SLANG_SUPPORT
 741         if (use_browser)
 742                 return __hpp__slsmg_color_printf(hpp, "%*.2f%%", width - 1, per);
 743 #endif
 744         return hpp_color_scnprintf(hpp, "%*.2f%%", width - 1, per);
 745 }
 746
 747 static double percent_hitm(struct c2c_hist_entry *c2c_he)
 748 {
 749         struct c2c_hists *hists;
 750         struct c2c_stats *stats;
 751         struct c2c_stats *total;
 752         int tot = 0, st = 0;
 753         double p;
 754
 755         hists = container_of(c2c_he->he.hists, struct c2c_hists, hists);
 756         stats = &c2c_he->stats;
 757         total = &hists->stats;
 758
 759         switch (c2c.display) {
 760         case DISPLAY_RMT:
 761                 st  = stats->rmt_hitm;
 762                 tot = total->rmt_hitm;
 763                 break;
 764         case DISPLAY_LCL:
 765                 st  = stats->lcl_hitm;
 766                 tot = total->lcl_hitm;
 767                 break;
 768         case DISPLAY_TOT:
 769                 st  = stats->tot_hitm;
 770                 tot = total->tot_hitm;
 771         default:
 772                 break;
 773         }
 774
 775         p = tot ? (double) st / tot : 0;
 776
 777         return 100 * p;
 778 }
 779
 780 #define PERC_STR(__s, __v)                              \
 781 ({                                                      \
 782         scnprintf(__s, sizeof(__s), "%.2F%%", __v);     \
 783         __s;                                            \
 784 })
 785
 786 static int
 787 percent_hitm_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 788                    struct hist_entry *he)
 789 {
 790         struct c2c_hist_entry *c2c_he;
 791         int width = c2c_width(fmt, hpp, he->hists);
 792         char buf[10];
 793         double per;
 794
 795         c2c_he = container_of(he, struct c2c_hist_entry, he);
 796         per = percent_hitm(c2c_he);
 797         return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
 798 }
 799
 800 static int
 801 percent_hitm_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 802                    struct hist_entry *he)
 803 {
 804         return percent_color(fmt, hpp, he, percent_hitm);
 805 }
 806
 807 static int64_t
 808 percent_hitm_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 809                  struct hist_entry *left, struct hist_entry *right)
 810 {
 811         struct c2c_hist_entry *c2c_left;
 812         struct c2c_hist_entry *c2c_right;
 813         double per_left;
 814         double per_right;
 815
 816         c2c_left  = container_of(left, struct c2c_hist_entry, he);
 817         c2c_right = container_of(right, struct c2c_hist_entry, he);
 818
 819         per_left  = percent_hitm(c2c_left);
 820         per_right = percent_hitm(c2c_right);
 821
 822         return per_left - per_right;
 823 }
 824
 825 static struct c2c_stats *he_stats(struct hist_entry *he)
 826 {
 827         struct c2c_hist_entry *c2c_he;
 828
 829         c2c_he = container_of(he, struct c2c_hist_entry, he);
 830         return &c2c_he->stats;
 831 }
 832
 833 static struct c2c_stats *total_stats(struct hist_entry *he)
 834 {
 835         struct c2c_hists *hists;
 836
 837         hists = container_of(he->hists, struct c2c_hists, hists);
 838         return &hists->stats;
 839 }
 840
 841 static double percent(int st, int tot)
 842 {
 843         return tot ? 100. * (double) st / (double) tot : 0;
 844 }
 845
 846 #define PERCENT(__h, __f) percent(he_stats(__h)->__f, total_stats(__h)->__f)
 847
 848 #define PERCENT_FN(__f)                                                         \
 849 static double percent_ ## __f(struct c2c_hist_entry *c2c_he)                    \
 850 {                                                                               \
 851         struct c2c_hists *hists;                                                \
 852                                                                                 \
 853         hists = container_of(c2c_he->he.hists, struct c2c_hists, hists);        \
 854         return percent(c2c_he->stats.__f, hists->stats.__f);                    \
 855 }
 856
 857 PERCENT_FN(rmt_hitm)
 858 PERCENT_FN(lcl_hitm)
 859 PERCENT_FN(st_l1hit)
 860 PERCENT_FN(st_l1miss)
 861
 862 static int
 863 percent_rmt_hitm_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 864                        struct hist_entry *he)
 865 {
 866         int width = c2c_width(fmt, hpp, he->hists);
 867         double per = PERCENT(he, rmt_hitm);
 868         char buf[10];
 869
 870         return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
 871 }
 872
 873 static int
 874 percent_rmt_hitm_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 875                        struct hist_entry *he)
 876 {
 877         return percent_color(fmt, hpp, he, percent_rmt_hitm);
 878 }
 879
 880 static int64_t
 881 percent_rmt_hitm_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 882                      struct hist_entry *left, struct hist_entry *right)
 883 {
 884         double per_left;
 885         double per_right;
 886
 887         per_left  = PERCENT(left, lcl_hitm);
 888         per_right = PERCENT(right, lcl_hitm);
 889
 890         return per_left - per_right;
 891 }
 892
 893 static int
 894 percent_lcl_hitm_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 895                        struct hist_entry *he)
 896 {
 897         int width = c2c_width(fmt, hpp, he->hists);
 898         double per = PERCENT(he, lcl_hitm);
 899         char buf[10];
 900
 901         return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
 902 }
 903
 904 static int
 905 percent_lcl_hitm_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 906                        struct hist_entry *he)
 907 {
 908         return percent_color(fmt, hpp, he, percent_lcl_hitm);
 909 }
 910
 911 static int64_t
 912 percent_lcl_hitm_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 913                      struct hist_entry *left, struct hist_entry *right)
 914 {
 915         double per_left;
 916         double per_right;
 917
 918         per_left  = PERCENT(left, lcl_hitm);
 919         per_right = PERCENT(right, lcl_hitm);
 920
 921         return per_left - per_right;
 922 }
 923
 924 static int
 925 percent_stores_l1hit_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 926                            struct hist_entry *he)
 927 {
 928         int width = c2c_width(fmt, hpp, he->hists);
 929         double per = PERCENT(he, st_l1hit);
 930         char buf[10];
 931
 932         return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
 933 }
 934
 935 static int
 936 percent_stores_l1hit_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 937                            struct hist_entry *he)
 938 {
 939         return percent_color(fmt, hpp, he, percent_st_l1hit);
 940 }
 941
 942 static int64_t
 943 percent_stores_l1hit_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 944                         struct hist_entry *left, struct hist_entry *right)
 945 {
 946         double per_left;
 947         double per_right;
 948
 949         per_left  = PERCENT(left, st_l1hit);
 950         per_right = PERCENT(right, st_l1hit);
 951
 952         return per_left - per_right;
 953 }
 954
 955 static int
 956 percent_stores_l1miss_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 957                            struct hist_entry *he)
 958 {
 959         int width = c2c_width(fmt, hpp, he->hists);
 960         double per = PERCENT(he, st_l1miss);
 961         char buf[10];
 962
 963         return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
 964 }
 965
 966 static int
 967 percent_stores_l1miss_color(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 968                             struct hist_entry *he)
 969 {
 970         return percent_color(fmt, hpp, he, percent_st_l1miss);
 971 }
 972
 973 static int64_t
 974 percent_stores_l1miss_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 975                           struct hist_entry *left, struct hist_entry *right)
 976 {
 977         double per_left;
 978         double per_right;
 979
 980         per_left  = PERCENT(left, st_l1miss);
 981         per_right = PERCENT(right, st_l1miss);
 982
 983         return per_left - per_right;
 984 }
 985
 986 STAT_FN(lcl_dram)
 987 STAT_FN(rmt_dram)
 988
 989 static int
 990 pid_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 991           struct hist_entry *he)
 992 {
 993         int width = c2c_width(fmt, hpp, he->hists);
 994
 995         return scnprintf(hpp->buf, hpp->size, "%*d", width, he->thread->pid_);
 996 }
 997
 998 static int64_t
 999 pid_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
1000         struct hist_entry *left, struct hist_entry *right)
1001 {
1002         return left->thread->pid_ - right->thread->pid_;
1003 }
1004
1005 static int64_t
1006 empty_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
1007           struct hist_entry *left __maybe_unused,
1008           struct hist_entry *right __maybe_unused)
1009 {
1010         return 0;
1011 }
1012
1013 static int
1014 node_entry(struct perf_hpp_fmt *fmt __maybe_unused, struct perf_hpp *hpp,
1015            struct hist_entry *he)
1016 {
1017         struct c2c_hist_entry *c2c_he;
1018         bool first = true;
1019         int node;
1020         int ret = 0;
1021
1022         c2c_he = container_of(he, struct c2c_hist_entry, he);
1023
1024         for (node = 0; node < c2c.nodes_cnt; node++) {
1025                 DECLARE_BITMAP(set, c2c.cpus_cnt);
1026
1027                 bitmap_zero(set, c2c.cpus_cnt);
1028                 bitmap_and(set, c2c_he->cpuset, c2c.nodes[node], c2c.cpus_cnt);
1029
1030                 if (!bitmap_weight(set, c2c.cpus_cnt)) {
1031                         if (c2c.node_info == 1) {
1032                                 ret = scnprintf(hpp->buf, hpp->size, "%21s", " ");
1033                                 advance_hpp(hpp, ret);
1034                         }
1035                         continue;
1036                 }
1037
1038                 if (!first) {
1039                         ret = scnprintf(hpp->buf, hpp->size, " ");
1040                         advance_hpp(hpp, ret);
1041                 }
1042
1043                 switch (c2c.node_info) {
1044                 case 0:
1045                         ret = scnprintf(hpp->buf, hpp->size, "%2d", node);
1046                         advance_hpp(hpp, ret);
1047                         break;
1048                 case 1:
1049                 {
1050                         int num = bitmap_weight(c2c_he->cpuset, c2c.cpus_cnt);
1051                         struct c2c_stats *stats = &c2c_he->node_stats[node];
1052
1053                         ret = scnprintf(hpp->buf, hpp->size, "%2d{%2d ", node, num);
1054                         advance_hpp(hpp, ret);
1055
1056                 #define DISPLAY_HITM(__h)                                               \
1057                         if (c2c_he->stats.__h> 0) {                                     \
1058                                 ret = scnprintf(hpp->buf, hpp->size, "%5.1f%% ",        \
1059                                                 percent(stats->__h, c2c_he->stats.__h));\
1060                         } else {                                                        \
1061                                 ret = scnprintf(hpp->buf, hpp->size, "%6s ", "n/a");    \
1062                         }
1063
1064                         switch (c2c.display) {
1065                         case DISPLAY_RMT:
1066                                 DISPLAY_HITM(rmt_hitm);
1067                                 break;
1068                         case DISPLAY_LCL:
1069                                 DISPLAY_HITM(lcl_hitm);
1070                                 break;
1071                         case DISPLAY_TOT:
1072                                 DISPLAY_HITM(tot_hitm);
1073                         default:
1074                                 break;
1075                         }
1076
1077                 #undef DISPLAY_HITM
1078
1079                         advance_hpp(hpp, ret);
1080
1081                         if (c2c_he->stats.store > 0) {
1082                                 ret = scnprintf(hpp->buf, hpp->size, "%5.1f%%}",
1083                                                 percent(stats->store, c2c_he->stats.store));
1084                         } else {
1085                                 ret = scnprintf(hpp->buf, hpp->size, "%6s}", "n/a");
1086                         }
1087
1088                         advance_hpp(hpp, ret);
1089                         break;
1090                 }
1091                 case 2:
1092                         ret = scnprintf(hpp->buf, hpp->size, "%2d{", node);
1093                         advance_hpp(hpp, ret);
1094
1095                         ret = bitmap_scnprintf(set, c2c.cpus_cnt, hpp->buf, hpp->size);
1096                         advance_hpp(hpp, ret);
1097
1098                         ret = scnprintf(hpp->buf, hpp->size, "}");
1099                         advance_hpp(hpp, ret);
1100                         break;
1101                 default:
1102                         break;
1103                 }
1104
1105                 first = false;
1106         }
1107
1108         return 0;
1109 }
1110
1111 static int
1112 mean_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
1113            struct hist_entry *he, double mean)
1114 {
1115         int width = c2c_width(fmt, hpp, he->hists);
1116         char buf[10];
1117
1118         scnprintf(buf, 10, "%6.0f", mean);
1119         return scnprintf(hpp->buf, hpp->size, "%*s", width, buf);
1120 }
1121
1122 #define MEAN_ENTRY(__func, __val)                                               \
1123 static int                                                                      \
1124 __func(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp, struct hist_entry *he)   \
1125 {                                                                               \
1126         struct c2c_hist_entry *c2c_he;                                          \
1127         c2c_he = container_of(he, struct c2c_hist_entry, he);                   \
1128         return mean_entry(fmt, hpp, he, avg_stats(&c2c_he->cstats.__val));      \
1129 }
1130
1131 MEAN_ENTRY(mean_rmt_entry,  rmt_hitm);
1132 MEAN_ENTRY(mean_lcl_entry,  lcl_hitm);
1133 MEAN_ENTRY(mean_load_entry, load);
1134
1135 static int
1136 cpucnt_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
1137              struct hist_entry *he)
1138 {
1139         struct c2c_hist_entry *c2c_he;
1140         int width = c2c_width(fmt, hpp, he->hists);
1141         char buf[10];
1142
1143         c2c_he = container_of(he, struct c2c_hist_entry, he);
1144
1145         scnprintf(buf, 10, "%d", bitmap_weight(c2c_he->cpuset, c2c.cpus_cnt));
1146         return scnprintf(hpp->buf, hpp->size, "%*s", width, buf);
1147 }
1148
1149 static int
1150 cl_idx_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
1151              struct hist_entry *he)
1152 {
1153         struct c2c_hist_entry *c2c_he;
1154         int width = c2c_width(fmt, hpp, he->hists);
1155         char buf[10];
1156
1157         c2c_he = container_of(he, struct c2c_hist_entry, he);
1158
1159         scnprintf(buf, 10, "%u", c2c_he->cacheline_idx);
1160         return scnprintf(hpp->buf, hpp->size, "%*s", width, buf);
1161 }
1162
1163 static int
1164 cl_idx_empty_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
1165                    struct hist_entry *he)
1166 {
1167         int width = c2c_width(fmt, hpp, he->hists);
1168
1169         return scnprintf(hpp->buf, hpp->size, "%*s", width, "");
1170 }
1171
1172 #define HEADER_LOW(__h)                 \
1173         {                               \
1174                 .line[1] = {            \
1175                         .text = __h,    \
1176                 },                      \
1177         }
1178
1179 #define HEADER_BOTH(__h0, __h1)         \
1180         {                               \
1181                 .line[0] = {            \
1182                         .text = __h0,   \
1183                 },                      \
1184                 .line[1] = {            \
1185                         .text = __h1,   \
1186                 },                      \
1187         }
1188
1189 #define HEADER_SPAN(__h0, __h1, __s)    \
1190         {                               \
1191                 .line[0] = {            \
1192                         .text = __h0,   \
1193                         .span = __s,    \
1194                 },                      \
1195                 .line[1] = {            \
1196                         .text = __h1,   \
1197                 },                      \
1198         }
1199
1200 #define HEADER_SPAN_LOW(__h)            \
1201         {                               \
1202                 .line[1] = {            \
1203                         .text = __h,    \
1204                 },                      \
1205         }
1206
1207 static struct c2c_dimension dim_dcacheline = {
1208         .header         = HEADER_LOW("Cacheline"),
1209         .name           = "dcacheline",
1210         .cmp            = dcacheline_cmp,
1211         .entry          = dcacheline_entry,
1212         .width          = 18,
1213 };
1214
1215 static struct c2c_header header_offset_tui = HEADER_LOW("Off");
1216
1217 static struct c2c_dimension dim_offset = {
1218         .header         = HEADER_BOTH("Data address", "Offset"),
1219         .name           = "offset",
1220         .cmp            = offset_cmp,
1221         .entry          = offset_entry,
1222         .width          = 18,
1223 };
1224
1225 static struct c2c_dimension dim_iaddr = {
1226         .header         = HEADER_LOW("Code address"),
1227         .name           = "iaddr",
1228         .cmp            = iaddr_cmp,
1229         .entry          = iaddr_entry,
1230         .width          = 18,
1231 };
1232
1233 static struct c2c_dimension dim_tot_hitm = {
1234         .header         = HEADER_SPAN("----- LLC Load Hitm -----", "Total", 2),
1235         .name           = "tot_hitm",
1236         .cmp            = tot_hitm_cmp,
1237         .entry          = tot_hitm_entry,
1238         .width          = 7,
1239 };
1240
1241 static struct c2c_dimension dim_lcl_hitm = {
1242         .header         = HEADER_SPAN_LOW("Lcl"),
1243         .name           = "lcl_hitm",
1244         .cmp            = lcl_hitm_cmp,
1245         .entry          = lcl_hitm_entry,
1246         .width          = 7,
1247 };
1248
1249 static struct c2c_dimension dim_rmt_hitm = {
1250         .header         = HEADER_SPAN_LOW("Rmt"),
1251         .name           = "rmt_hitm",
1252         .cmp            = rmt_hitm_cmp,
1253         .entry          = rmt_hitm_entry,
1254         .width          = 7,
1255 };
1256
1257 static struct c2c_dimension dim_cl_rmt_hitm = {
1258         .header         = HEADER_SPAN("----- HITM -----", "Rmt", 1),
1259         .name           = "cl_rmt_hitm",
1260         .cmp            = rmt_hitm_cmp,
1261         .entry          = rmt_hitm_entry,
1262         .width          = 7,
1263 };
1264
1265 static struct c2c_dimension dim_cl_lcl_hitm = {
1266         .header         = HEADER_SPAN_LOW("Lcl"),
1267         .name           = "cl_lcl_hitm",
1268         .cmp            = lcl_hitm_cmp,
1269         .entry          = lcl_hitm_entry,
1270         .width          = 7,
1271 };
1272
1273 static struct c2c_dimension dim_stores = {
1274         .header         = HEADER_SPAN("---- Store Reference ----", "Total", 2),
1275         .name           = "stores",
1276         .cmp            = store_cmp,
1277         .entry          = store_entry,
1278         .width          = 7,
1279 };
1280
1281 static struct c2c_dimension dim_stores_l1hit = {
1282         .header         = HEADER_SPAN_LOW("L1Hit"),
1283         .name           = "stores_l1hit",
1284         .cmp            = st_l1hit_cmp,
1285         .entry          = st_l1hit_entry,
1286         .width          = 7,
1287 };
1288
1289 static struct c2c_dimension dim_stores_l1miss = {
1290         .header         = HEADER_SPAN_LOW("L1Miss"),
1291         .name           = "stores_l1miss",
1292         .cmp            = st_l1miss_cmp,
1293         .entry          = st_l1miss_entry,
1294         .width          = 7,
1295 };
1296
1297 static struct c2c_dimension dim_cl_stores_l1hit = {
1298         .header         = HEADER_SPAN("-- Store Refs --", "L1 Hit", 1),
1299         .name           = "cl_stores_l1hit",
1300         .cmp            = st_l1hit_cmp,
1301         .entry          = st_l1hit_entry,
1302         .width          = 7,
1303 };
1304
1305 static struct c2c_dimension dim_cl_stores_l1miss = {
1306         .header         = HEADER_SPAN_LOW("L1 Miss"),
1307         .name           = "cl_stores_l1miss",
1308         .cmp            = st_l1miss_cmp,
1309         .entry          = st_l1miss_entry,
1310         .width          = 7,
1311 };
1312
1313 static struct c2c_dimension dim_ld_fbhit = {
1314         .header         = HEADER_SPAN("----- Core Load Hit -----", "FB", 2),
1315         .name           = "ld_fbhit",
1316         .cmp            = ld_fbhit_cmp,
1317         .entry          = ld_fbhit_entry,
1318         .width          = 7,
1319 };
1320
1321 static struct c2c_dimension dim_ld_l1hit = {
1322         .header         = HEADER_SPAN_LOW("L1"),
1323         .name           = "ld_l1hit",
1324         .cmp            = ld_l1hit_cmp,
1325         .entry          = ld_l1hit_entry,
1326         .width          = 7,
1327 };
1328
1329 static struct c2c_dimension dim_ld_l2hit = {
1330         .header         = HEADER_SPAN_LOW("L2"),
1331         .name           = "ld_l2hit",
1332         .cmp            = ld_l2hit_cmp,
1333         .entry          = ld_l2hit_entry,
1334         .width          = 7,
1335 };
1336
1337 static struct c2c_dimension dim_ld_llchit = {
1338         .header         = HEADER_SPAN("-- LLC Load Hit --", "Llc", 1),
1339         .name           = "ld_lclhit",
1340         .cmp            = ld_llchit_cmp,
1341         .entry          = ld_llchit_entry,
1342         .width          = 8,
1343 };
1344
1345 static struct c2c_dimension dim_ld_rmthit = {
1346         .header         = HEADER_SPAN_LOW("Rmt"),
1347         .name           = "ld_rmthit",
1348         .cmp            = rmt_hit_cmp,
1349         .entry          = rmt_hit_entry,
1350         .width          = 8,
1351 };
1352
1353 static struct c2c_dimension dim_ld_llcmiss = {
1354         .header         = HEADER_BOTH("LLC", "Ld Miss"),
1355         .name           = "ld_llcmiss",
1356         .cmp            = ld_llcmiss_cmp,
1357         .entry          = ld_llcmiss_entry,
1358         .width          = 7,
1359 };
1360
1361 static struct c2c_dimension dim_tot_recs = {
1362         .header         = HEADER_BOTH("Total", "records"),
1363         .name           = "tot_recs",
1364         .cmp            = tot_recs_cmp,
1365         .entry          = tot_recs_entry,
1366         .width          = 7,
1367 };
1368
1369 static struct c2c_dimension dim_tot_loads = {
1370         .header         = HEADER_BOTH("Total", "Loads"),
1371         .name           = "tot_loads",
1372         .cmp            = tot_loads_cmp,
1373         .entry          = tot_loads_entry,
1374         .width          = 7,
1375 };
1376
1377 static struct c2c_header percent_hitm_header[] = {
1378         [DISPLAY_LCL] = HEADER_BOTH("Lcl", "Hitm"),
1379         [DISPLAY_RMT] = HEADER_BOTH("Rmt", "Hitm"),
1380         [DISPLAY_TOT] = HEADER_BOTH("Tot", "Hitm"),
1381 };
1382
1383 static struct c2c_dimension dim_percent_hitm = {
1384         .name           = "percent_hitm",
1385         .cmp            = percent_hitm_cmp,
1386         .entry          = percent_hitm_entry,
1387         .color          = percent_hitm_color,
1388         .width          = 7,
1389 };
1390
1391 static struct c2c_dimension dim_percent_rmt_hitm = {
1392         .header         = HEADER_SPAN("----- HITM -----", "Rmt", 1),
1393         .name           = "percent_rmt_hitm",
1394         .cmp            = percent_rmt_hitm_cmp,
1395         .entry          = percent_rmt_hitm_entry,
1396         .color          = percent_rmt_hitm_color,
1397         .width          = 7,
1398 };
1399
1400 static struct c2c_dimension dim_percent_lcl_hitm = {
1401         .header         = HEADER_SPAN_LOW("Lcl"),
1402         .name           = "percent_lcl_hitm",
1403         .cmp            = percent_lcl_hitm_cmp,
1404         .entry          = percent_lcl_hitm_entry,
1405         .color          = percent_lcl_hitm_color,
1406         .width          = 7,
1407 };
1408
1409 static struct c2c_dimension dim_percent_stores_l1hit = {
1410         .header         = HEADER_SPAN("-- Store Refs --", "L1 Hit", 1),
1411         .name           = "percent_stores_l1hit",
1412         .cmp            = percent_stores_l1hit_cmp,
1413         .entry          = percent_stores_l1hit_entry,
1414         .color          = percent_stores_l1hit_color,
1415         .width          = 7,
1416 };
1417
1418 static struct c2c_dimension dim_percent_stores_l1miss = {
1419         .header         = HEADER_SPAN_LOW("L1 Miss"),
1420         .name           = "percent_stores_l1miss",
1421         .cmp            = percent_stores_l1miss_cmp,
1422         .entry          = percent_stores_l1miss_entry,
1423         .color          = percent_stores_l1miss_color,
1424         .width          = 7,
1425 };
1426
1427 static struct c2c_dimension dim_dram_lcl = {
1428         .header         = HEADER_SPAN("--- Load Dram ----", "Lcl", 1),
1429         .name           = "dram_lcl",
1430         .cmp            = lcl_dram_cmp,
1431         .entry          = lcl_dram_entry,
1432         .width          = 8,
1433 };
1434
1435 static struct c2c_dimension dim_dram_rmt = {
1436         .header         = HEADER_SPAN_LOW("Rmt"),
1437         .name           = "dram_rmt",
1438         .cmp            = rmt_dram_cmp,
1439         .entry          = rmt_dram_entry,
1440         .width          = 8,
1441 };
1442
1443 static struct c2c_dimension dim_pid = {
1444         .header         = HEADER_LOW("Pid"),
1445         .name           = "pid",
1446         .cmp            = pid_cmp,
1447         .entry          = pid_entry,
1448         .width          = 7,
1449 };
1450
1451 static struct c2c_dimension dim_tid = {
1452         .header         = HEADER_LOW("Tid"),
1453         .name           = "tid",
1454         .se             = &sort_thread,
1455 };
1456
1457 static struct c2c_dimension dim_symbol = {
1458         .name           = "symbol",
1459         .se             = &sort_sym,
1460 };
1461
1462 static struct c2c_dimension dim_dso = {
1463         .header         = HEADER_BOTH("Shared", "Object"),
1464         .name           = "dso",
1465         .se             = &sort_dso,
1466 };
1467
1468 static struct c2c_header header_node[3] = {
1469         HEADER_LOW("Node"),
1470         HEADER_LOW("Node{cpus %hitms %stores}"),
1471         HEADER_LOW("Node{cpu list}"),
1472 };
1473
1474 static struct c2c_dimension dim_node = {
1475         .name           = "node",
1476         .cmp            = empty_cmp,
1477         .entry          = node_entry,
1478         .width          = 4,
1479 };
1480
1481 static struct c2c_dimension dim_mean_rmt = {
1482         .header         = HEADER_SPAN("---------- cycles ----------", "rmt hitm", 2),
1483         .name           = "mean_rmt",
1484         .cmp            = empty_cmp,
1485         .entry          = mean_rmt_entry,
1486         .width          = 8,
1487 };
1488
1489 static struct c2c_dimension dim_mean_lcl = {
1490         .header         = HEADER_SPAN_LOW("lcl hitm"),
1491         .name           = "mean_lcl",
1492         .cmp            = empty_cmp,
1493         .entry          = mean_lcl_entry,
1494         .width          = 8,
1495 };
1496
1497 static struct c2c_dimension dim_mean_load = {
1498         .header         = HEADER_SPAN_LOW("load"),
1499         .name           = "mean_load",
1500         .cmp            = empty_cmp,
1501         .entry          = mean_load_entry,
1502         .width          = 8,
1503 };
1504
1505 static struct c2c_dimension dim_cpucnt = {
1506         .header         = HEADER_BOTH("cpu", "cnt"),
1507         .name           = "cpucnt",
1508         .cmp            = empty_cmp,
1509         .entry          = cpucnt_entry,
1510         .width          = 8,
1511 };
1512
1513 static struct c2c_dimension dim_srcline = {
1514         .name           = "cl_srcline",
1515         .se             = &sort_srcline,
1516 };
1517
1518 static struct c2c_dimension dim_dcacheline_idx = {
1519         .header         = HEADER_LOW("Index"),
1520         .name           = "cl_idx",
1521         .cmp            = empty_cmp,
1522         .entry          = cl_idx_entry,
1523         .width          = 5,
1524 };
1525
1526 static struct c2c_dimension dim_dcacheline_num = {
1527         .header         = HEADER_LOW("Num"),
1528         .name           = "cl_num",
1529         .cmp            = empty_cmp,
1530         .entry          = cl_idx_entry,
1531         .width          = 5,
1532 };
1533
1534 static struct c2c_dimension dim_dcacheline_num_empty = {
1535         .header         = HEADER_LOW("Num"),
1536         .name           = "cl_num_empty",
1537         .cmp            = empty_cmp,
1538         .entry          = cl_idx_empty_entry,
1539         .width          = 5,
1540 };
1541
1542 static struct c2c_dimension *dimensions[] = {
1543         &dim_dcacheline,
1544         &dim_offset,
1545         &dim_iaddr,
1546         &dim_tot_hitm,
1547         &dim_lcl_hitm,
1548         &dim_rmt_hitm,
1549         &dim_cl_lcl_hitm,
1550         &dim_cl_rmt_hitm,
1551         &dim_stores,
1552         &dim_stores_l1hit,
1553         &dim_stores_l1miss,
1554         &dim_cl_stores_l1hit,
1555         &dim_cl_stores_l1miss,
1556         &dim_ld_fbhit,
1557         &dim_ld_l1hit,
1558         &dim_ld_l2hit,
1559         &dim_ld_llchit,
1560         &dim_ld_rmthit,
1561         &dim_ld_llcmiss,
1562         &dim_tot_recs,
1563         &dim_tot_loads,
1564         &dim_percent_hitm,
1565         &dim_percent_rmt_hitm,
1566         &dim_percent_lcl_hitm,
1567         &dim_percent_stores_l1hit,
1568         &dim_percent_stores_l1miss,
1569         &dim_dram_lcl,
1570         &dim_dram_rmt,
1571         &dim_pid,
1572         &dim_tid,
1573         &dim_symbol,
1574         &dim_dso,
1575         &dim_node,
1576         &dim_mean_rmt,
1577         &dim_mean_lcl,
1578         &dim_mean_load,
1579         &dim_cpucnt,
1580         &dim_srcline,
1581         &dim_dcacheline_idx,
1582         &dim_dcacheline_num,
1583         &dim_dcacheline_num_empty,
1584         NULL,
1585 };
1586
1587 static void fmt_free(struct perf_hpp_fmt *fmt)
1588 {
1589         struct c2c_fmt *c2c_fmt;
1590
1591         c2c_fmt = container_of(fmt, struct c2c_fmt, fmt);
1592         free(c2c_fmt);
1593 }
1594
1595 static bool fmt_equal(struct perf_hpp_fmt *a, struct perf_hpp_fmt *b)
1596 {
1597         struct c2c_fmt *c2c_a = container_of(a, struct c2c_fmt, fmt);
1598         struct c2c_fmt *c2c_b = container_of(b, struct c2c_fmt, fmt);
1599
1600         return c2c_a->dim == c2c_b->dim;
1601 }
1602
1603 static struct c2c_dimension *get_dimension(const char *name)
1604 {
1605         unsigned int i;
1606
1607         for (i = 0; dimensions[i]; i++) {
1608                 struct c2c_dimension *dim = dimensions[i];
1609
1610                 if (!strcmp(dim->name, name))
1611                         return dim;
1612         };
1613
1614         return NULL;
1615 }
1616
1617 static int c2c_se_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
1618                         struct hist_entry *he)
1619 {
1620         struct c2c_fmt *c2c_fmt = container_of(fmt, struct c2c_fmt, fmt);
1621         struct c2c_dimension *dim = c2c_fmt->dim;
1622         size_t len = fmt->user_len;
1623
1624         if (!len) {
1625                 len = hists__col_len(he->hists, dim->se->se_width_idx);
1626
1627                 if (dim == &dim_symbol || dim == &dim_srcline)
1628                         len = symbol_width(he->hists, dim->se);
1629         }
1630
1631         return dim->se->se_snprintf(he, hpp->buf, hpp->size, len);
1632 }
1633
1634 static int64_t c2c_se_cmp(struct perf_hpp_fmt *fmt,
1635                           struct hist_entry *a, struct hist_entry *b)
1636 {
1637         struct c2c_fmt *c2c_fmt = container_of(fmt, struct c2c_fmt, fmt);
1638         struct c2c_dimension *dim = c2c_fmt->dim;
1639
1640         return dim->se->se_cmp(a, b);
1641 }
1642
1643 static int64_t c2c_se_collapse(struct perf_hpp_fmt *fmt,
1644                                struct hist_entry *a, struct hist_entry *b)
1645 {
1646         struct c2c_fmt *c2c_fmt = container_of(fmt, struct c2c_fmt, fmt);
1647         struct c2c_dimension *dim = c2c_fmt->dim;
1648         int64_t (*collapse_fn)(struct hist_entry *, struct hist_entry *);
1649
1650         collapse_fn = dim->se->se_collapse ?: dim->se->se_cmp;
1651         return collapse_fn(a, b);
1652 }
1653
1654 static struct c2c_fmt *get_format(const char *name)
1655 {
1656         struct c2c_dimension *dim = get_dimension(name);
1657         struct c2c_fmt *c2c_fmt;
1658         struct perf_hpp_fmt *fmt;
1659
1660         if (!dim)
1661                 return NULL;
1662
1663         c2c_fmt = zalloc(sizeof(*c2c_fmt));
1664         if (!c2c_fmt)
1665                 return NULL;
1666
1667         c2c_fmt->dim = dim;
1668
1669         fmt = &c2c_fmt->fmt;
1670         INIT_LIST_HEAD(&fmt->list);
1671         INIT_LIST_HEAD(&fmt->sort_list);
1672
1673         fmt->cmp        = dim->se ? c2c_se_cmp   : dim->cmp;
1674         fmt->sort       = dim->se ? c2c_se_cmp   : dim->cmp;
1675         fmt->color      = dim->se ? NULL         : dim->color;
1676         fmt->entry      = dim->se ? c2c_se_entry : dim->entry;
1677         fmt->header     = c2c_header;
1678         fmt->width      = c2c_width;
1679         fmt->collapse   = dim->se ? c2c_se_collapse : dim->cmp;
1680         fmt->equal      = fmt_equal;
1681         fmt->free       = fmt_free;
1682
1683         return c2c_fmt;
1684 }
1685
1686 static int c2c_hists__init_output(struct perf_hpp_list *hpp_list, char *name)
1687 {
1688         struct c2c_fmt *c2c_fmt = get_format(name);
1689
1690         if (!c2c_fmt) {
1691                 reset_dimensions();
1692                 return output_field_add(hpp_list, name);
1693         }
1694
1695         perf_hpp_list__column_register(hpp_list, &c2c_fmt->fmt);
1696         return 0;
1697 }
1698
1699 static int c2c_hists__init_sort(struct perf_hpp_list *hpp_list, char *name)
1700 {
1701         struct c2c_fmt *c2c_fmt = get_format(name);
1702         struct c2c_dimension *dim;
1703
1704         if (!c2c_fmt) {
1705                 reset_dimensions();
1706                 return sort_dimension__add(hpp_list, name, NULL, 0);
1707         }
1708
1709         dim = c2c_fmt->dim;
1710         if (dim == &dim_dso)
1711                 hpp_list->dso = 1;
1712
1713         perf_hpp_list__register_sort_field(hpp_list, &c2c_fmt->fmt);
1714         return 0;
1715 }
1716
1717 #define PARSE_LIST(_list, _fn)                                                  \
1718         do {                                                                    \
1719                 char *tmp, *tok;                                                \
1720                 ret = 0;                                                        \
1721                                                                                 \
1722                 if (!_list)                                                     \
1723                         break;                                                  \
1724                                                                                 \
1725                 for (tok = strtok_r((char *)_list, ", ", &tmp);                 \
1726                                 tok; tok = strtok_r(NULL, ", ", &tmp)) {        \
1727                         ret = _fn(hpp_list, tok);                               \
1728                         if (ret == -EINVAL) {                                   \
1729                                 pr_err("Invalid --fields key: `%s'", tok);      \
1730                                 break;                                          \
1731                         } else if (ret == -ESRCH) {                             \
1732                                 pr_err("Unknown --fields key: `%s'", tok);      \
1733                                 break;                                          \
1734                         }                                                       \
1735                 }                                                               \
1736         } while (0)
1737
1738 static int hpp_list__parse(struct perf_hpp_list *hpp_list,
1739                            const char *output_,
1740                            const char *sort_)
1741 {
1742         char *output = output_ ? strdup(output_) : NULL;
1743         char *sort   = sort_   ? strdup(sort_) : NULL;
1744         int ret;
1745
1746         PARSE_LIST(output, c2c_hists__init_output);
1747         PARSE_LIST(sort,   c2c_hists__init_sort);
1748
1749         /* copy sort keys to output fields */
1750         perf_hpp__setup_output_field(hpp_list);
1751
1752         /*
1753          * We dont need other sorting keys other than those
1754          * we already specified. It also really slows down
1755          * the processing a lot with big number of output
1756          * fields, so switching this off for c2c.
1757          */
1758
1759 #if 0
1760         /* and then copy output fields to sort keys */
1761         perf_hpp__append_sort_keys(&hists->list);
1762 #endif
1763
1764         free(output);
1765         free(sort);
1766         return ret;
1767 }
1768
1769 static int c2c_hists__init(struct c2c_hists *hists,
1770                            const char *sort,
1771                            int nr_header_lines)
1772 {
1773         __hists__init(&hists->hists, &hists->list);
1774
1775         /*
1776          * Initialize only with sort fields, we need to resort
1777          * later anyway, and that's where we add output fields
1778          * as well.
1779          */
1780         perf_hpp_list__init(&hists->list);
1781
1782         /* Overload number of header lines.*/
1783         hists->list.nr_header_lines = nr_header_lines;
1784
1785         return hpp_list__parse(&hists->list, NULL, sort);
1786 }
1787
1788 static int c2c_hists__reinit(struct c2c_hists *c2c_hists,
1789                              const char *output,
1790                              const char *sort)
1791 {
1792         perf_hpp__reset_output_field(&c2c_hists->list);
1793         return hpp_list__parse(&c2c_hists->list, output, sort);
1794 }
1795
1796 #define DISPLAY_LINE_LIMIT  0.0005
1797
1798 static bool he__display(struct hist_entry *he, struct c2c_stats *stats)
1799 {
1800         struct c2c_hist_entry *c2c_he;
1801         double ld_dist;
1802
1803         if (c2c.show_all)
1804                 return true;
1805
1806         c2c_he = container_of(he, struct c2c_hist_entry, he);
1807
1808 #define FILTER_HITM(__h)                                                \
1809         if (stats->__h) {                                               \
1810                 ld_dist = ((double)c2c_he->stats.__h / stats->__h);     \
1811                 if (ld_dist < DISPLAY_LINE_LIMIT)                       \
1812                         he->filtered = HIST_FILTER__C2C;                \
1813         } else {                                                        \
1814                 he->filtered = HIST_FILTER__C2C;                        \
1815         }
1816
1817         switch (c2c.display) {
1818         case DISPLAY_LCL:
1819                 FILTER_HITM(lcl_hitm);
1820                 break;
1821         case DISPLAY_RMT:
1822                 FILTER_HITM(rmt_hitm);
1823                 break;
1824         case DISPLAY_TOT:
1825                 FILTER_HITM(tot_hitm);
1826         default:
1827                 break;
1828         };
1829
1830 #undef FILTER_HITM
1831
1832         return he->filtered == 0;
1833 }
1834
1835 static inline int valid_hitm_or_store(struct hist_entry *he)
1836 {
1837         struct c2c_hist_entry *c2c_he;
1838         bool has_hitm;
1839
1840         c2c_he = container_of(he, struct c2c_hist_entry, he);
1841         has_hitm = c2c.display == DISPLAY_TOT ? c2c_he->stats.tot_hitm :
1842                    c2c.display == DISPLAY_LCL ? c2c_he->stats.lcl_hitm :
1843                                                 c2c_he->stats.rmt_hitm;
1844         return has_hitm || c2c_he->stats.store;
1845 }
1846
1847 static void calc_width(struct hist_entry *he)
1848 {
1849         struct c2c_hists *c2c_hists;
1850
1851         c2c_hists = container_of(he->hists, struct c2c_hists, hists);
1852         hists__calc_col_len(&c2c_hists->hists, he);
1853 }
1854
1855 static int filter_cb(struct hist_entry *he)
1856 {
1857         if (c2c.show_src && !he->srcline)
1858                 he->srcline = hist_entry__get_srcline(he);
1859
1860         calc_width(he);
1861
1862         if (!valid_hitm_or_store(he))
1863                 he->filtered = HIST_FILTER__C2C;
1864
1865         return 0;
1866 }
1867
1868 static int resort_cl_cb(struct hist_entry *he)
1869 {
1870         struct c2c_hist_entry *c2c_he;
1871         struct c2c_hists *c2c_hists;
1872         bool display = he__display(he, &c2c.hitm_stats);
1873
1874         c2c_he = container_of(he, struct c2c_hist_entry, he);
1875         c2c_hists = c2c_he->hists;
1876
1877         calc_width(he);
1878
1879         if (display && c2c_hists) {
1880                 static unsigned int idx;
1881
1882                 c2c_he->cacheline_idx = idx++;
1883
1884                 c2c_hists__reinit(c2c_hists, c2c.cl_output, c2c.cl_resort);
1885
1886                 hists__collapse_resort(&c2c_hists->hists, NULL);
1887                 hists__output_resort_cb(&c2c_hists->hists, NULL, filter_cb);
1888         }
1889
1890         return 0;
1891 }
1892
1893 static void setup_nodes_header(void)
1894 {
1895         dim_node.header = header_node[c2c.node_info];
1896 }
1897
1898 static int setup_nodes(struct perf_session *session)
1899 {
1900         struct numa_node *n;
1901         unsigned long **nodes;
1902         int node, cpu;
1903         int *cpu2node;
1904
1905         if (c2c.node_info > 2)
1906                 c2c.node_info = 2;
1907
1908         c2c.nodes_cnt = session->header.env.nr_numa_nodes;
1909         c2c.cpus_cnt  = session->header.env.nr_cpus_online;
1910
1911         n = session->header.env.numa_nodes;
1912         if (!n)
1913                 return -EINVAL;
1914
1915         nodes = zalloc(sizeof(unsigned long *) * c2c.nodes_cnt);
1916         if (!nodes)
1917                 return -ENOMEM;
1918
1919         c2c.nodes = nodes;
1920
1921         cpu2node = zalloc(sizeof(int) * c2c.cpus_cnt);
1922         if (!cpu2node)
1923                 return -ENOMEM;
1924
1925         for (cpu = 0; cpu < c2c.cpus_cnt; cpu++)
1926                 cpu2node[cpu] = -1;
1927
1928         c2c.cpu2node = cpu2node;
1929
1930         for (node = 0; node < c2c.nodes_cnt; node++) {
1931                 struct cpu_map *map = n[node].map;
1932                 unsigned long *set;
1933
1934                 set = bitmap_alloc(c2c.cpus_cnt);
1935                 if (!set)
1936                         return -ENOMEM;
1937
1938                 for (cpu = 0; cpu < map->nr; cpu++) {
1939                         set_bit(map->map[cpu], set);
1940
1941                         if (WARN_ONCE(cpu2node[map->map[cpu]] != -1, "node/cpu topology bug"))
1942                                 return -EINVAL;
1943
1944                         cpu2node[map->map[cpu]] = node;
1945                 }
1946
1947                 nodes[node] = set;
1948         }
1949
1950         setup_nodes_header();
1951         return 0;
1952 }
1953
1954 #define HAS_HITMS(__h) ((__h)->stats.lcl_hitm || (__h)->stats.rmt_hitm)
1955
1956 static int resort_hitm_cb(struct hist_entry *he)
1957 {
1958         struct c2c_hist_entry *c2c_he;
1959         c2c_he = container_of(he, struct c2c_hist_entry, he);
1960
1961         if (HAS_HITMS(c2c_he)) {
1962                 c2c.shared_clines++;
1963                 c2c_add_stats(&c2c.hitm_stats, &c2c_he->stats);
1964         }
1965
1966         return 0;
1967 }
1968
1969 static int hists__iterate_cb(struct hists *hists, hists__resort_cb_t cb)
1970 {
1971         struct rb_node *next = rb_first(&hists->entries);
1972         int ret = 0;
1973
1974         while (next) {
1975                 struct hist_entry *he;
1976
1977                 he = rb_entry(next, struct hist_entry, rb_node);
1978                 ret = cb(he);
1979                 if (ret)
1980                         break;
1981                 next = rb_next(&he->rb_node);
1982         }
1983
1984         return ret;
1985 }
1986
1987 static void print_c2c__display_stats(FILE *out)
1988 {
1989         int llc_misses;
1990         struct c2c_stats *stats = &c2c.hists.stats;
1991
1992         llc_misses = stats->lcl_dram +
1993                      stats->rmt_dram +
1994                      stats->rmt_hit +
1995                      stats->rmt_hitm;
1996
1997         fprintf(out, "=================================================\n");
1998         fprintf(out, "            Trace Event Information              \n");
1999         fprintf(out, "=================================================\n");
2000         fprintf(out, "  Total records                     : %10d\n", stats->nr_entries);
2001         fprintf(out, "  Locked Load/Store Operations      : %10d\n", stats->locks);
2002         fprintf(out, "  Load Operations                   : %10d\n", stats->load);
2003         fprintf(out, "  Loads - uncacheable               : %10d\n", stats->ld_uncache);
2004         fprintf(out, "  Loads - IO                        : %10d\n", stats->ld_io);
2005         fprintf(out, "  Loads - Miss                      : %10d\n", stats->ld_miss);
2006         fprintf(out, "  Loads - no mapping                : %10d\n", stats->ld_noadrs);
2007         fprintf(out, "  Load Fill Buffer Hit              : %10d\n", stats->ld_fbhit);
2008         fprintf(out, "  Load L1D hit                      : %10d\n", stats->ld_l1hit);
2009         fprintf(out, "  Load L2D hit                      : %10d\n", stats->ld_l2hit);
2010         fprintf(out, "  Load LLC hit                      : %10d\n", stats->ld_llchit + stats->lcl_hitm);
2011         fprintf(out, "  Load Local HITM                   : %10d\n", stats->lcl_hitm);
2012         fprintf(out, "  Load Remote HITM                  : %10d\n", stats->rmt_hitm);
2013         fprintf(out, "  Load Remote HIT                   : %10d\n", stats->rmt_hit);
2014         fprintf(out, "  Load Local DRAM                   : %10d\n", stats->lcl_dram);
2015         fprintf(out, "  Load Remote DRAM                  : %10d\n", stats->rmt_dram);
2016         fprintf(out, "  Load MESI State Exclusive         : %10d\n", stats->ld_excl);
2017         fprintf(out, "  Load MESI State Shared            : %10d\n", stats->ld_shared);
2018         fprintf(out, "  Load LLC Misses                   : %10d\n", llc_misses);
2019         fprintf(out, "  LLC Misses to Local DRAM          : %10.1f%%\n", ((double)stats->lcl_dram/(double)llc_misses) * 100.);
2020         fprintf(out, "  LLC Misses to Remote DRAM         : %10.1f%%\n", ((double)stats->rmt_dram/(double)llc_misses) * 100.);
2021         fprintf(out, "  LLC Misses to Remote cache (HIT)  : %10.1f%%\n", ((double)stats->rmt_hit /(double)llc_misses) * 100.);
2022         fprintf(out, "  LLC Misses to Remote cache (HITM) : %10.1f%%\n", ((double)stats->rmt_hitm/(double)llc_misses) * 100.);
2023         fprintf(out, "  Store Operations                  : %10d\n", stats->store);
2024         fprintf(out, "  Store - uncacheable               : %10d\n", stats->st_uncache);
2025         fprintf(out, "  Store - no mapping                : %10d\n", stats->st_noadrs);
2026         fprintf(out, "  Store L1D Hit                     : %10d\n", stats->st_l1hit);
2027         fprintf(out, "  Store L1D Miss                    : %10d\n", stats->st_l1miss);
2028         fprintf(out, "  No Page Map Rejects               : %10d\n", stats->nomap);
2029         fprintf(out, "  Unable to parse data source       : %10d\n", stats->noparse);
2030 }
2031
2032 static void print_shared_cacheline_info(FILE *out)
2033 {
2034         struct c2c_stats *stats = &c2c.hitm_stats;
2035         int hitm_cnt = stats->lcl_hitm + stats->rmt_hitm;
2036
2037         fprintf(out, "=================================================\n");
2038         fprintf(out, "    Global Shared Cache Line Event Information   \n");
2039         fprintf(out, "=================================================\n");
2040         fprintf(out, "  Total Shared Cache Lines          : %10d\n", c2c.shared_clines);
2041         fprintf(out, "  Load HITs on shared lines         : %10d\n", stats->load);
2042         fprintf(out, "  Fill Buffer Hits on shared lines  : %10d\n", stats->ld_fbhit);
2043         fprintf(out, "  L1D hits on shared lines          : %10d\n", stats->ld_l1hit);
2044         fprintf(out, "  L2D hits on shared lines          : %10d\n", stats->ld_l2hit);
2045         fprintf(out, "  LLC hits on shared lines          : %10d\n", stats->ld_llchit + stats->lcl_hitm);
2046         fprintf(out, "  Locked Access on shared lines     : %10d\n", stats->locks);
2047         fprintf(out, "  Store HITs on shared lines        : %10d\n", stats->store);
2048         fprintf(out, "  Store L1D hits on shared lines    : %10d\n", stats->st_l1hit);
2049         fprintf(out, "  Total Merged records              : %10d\n", hitm_cnt + stats->store);
2050 }
2051
2052 static void print_cacheline(struct c2c_hists *c2c_hists,
2053                             struct hist_entry *he_cl,
2054                             struct perf_hpp_list *hpp_list,
2055                             FILE *out)
2056 {
2057         char bf[1000];
2058         struct perf_hpp hpp = {
2059                 .buf            = bf,
2060                 .size           = 1000,
2061         };
2062         static bool once;
2063
2064         if (!once) {
2065                 hists__fprintf_headers(&c2c_hists->hists, out);
2066                 once = true;
2067         } else {
2068                 fprintf(out, "\n");
2069         }
2070
2071         fprintf(out, "  -------------------------------------------------------------\n");
2072         __hist_entry__snprintf(he_cl, &hpp, hpp_list);
2073         fprintf(out, "%s\n", bf);
2074         fprintf(out, "  -------------------------------------------------------------\n");
2075
2076         hists__fprintf(&c2c_hists->hists, false, 0, 0, 0, out, true);
2077 }
2078
2079 static void print_pareto(FILE *out)
2080 {
2081         struct perf_hpp_list hpp_list;
2082         struct rb_node *nd;
2083         int ret;
2084
2085         perf_hpp_list__init(&hpp_list);
2086         ret = hpp_list__parse(&hpp_list,
2087                                 "cl_num,"
2088                                 "cl_rmt_hitm,"
2089                                 "cl_lcl_hitm,"
2090                                 "cl_stores_l1hit,"
2091                                 "cl_stores_l1miss,"
2092                                 "dcacheline",
2093                                 NULL);
2094
2095         if (WARN_ONCE(ret, "failed to setup sort entries\n"))
2096                 return;
2097
2098         nd = rb_first(&c2c.hists.hists.entries);
2099
2100         for (; nd; nd = rb_next(nd)) {
2101                 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
2102                 struct c2c_hist_entry *c2c_he;
2103
2104                 if (he->filtered)
2105                         continue;
2106
2107                 c2c_he = container_of(he, struct c2c_hist_entry, he);
2108                 print_cacheline(c2c_he->hists, he, &hpp_list, out);
2109         }
2110 }
2111
2112 static void print_c2c_info(FILE *out, struct perf_session *session)
2113 {
2114         struct perf_evlist *evlist = session->evlist;
2115         struct perf_evsel *evsel;
2116         bool first = true;
2117
2118         fprintf(out, "=================================================\n");
2119         fprintf(out, "                 c2c details                     \n");
2120         fprintf(out, "=================================================\n");
2121
2122         evlist__for_each_entry(evlist, evsel) {
2123                 fprintf(out, "%-36s: %s\n", first ? "  Events" : "",
2124                         perf_evsel__name(evsel));
2125                 first = false;
2126         }
2127         fprintf(out, "  Cachelines sort on                : %s HITMs\n",
2128                 display_str[c2c.display]);
2129         fprintf(out, "  Cacheline data grouping           : %s\n", c2c.cl_sort);
2130 }
2131
2132 static void perf_c2c__hists_fprintf(FILE *out, struct perf_session *session)
2133 {
2134         setup_pager();
2135
2136         print_c2c__display_stats(out);
2137         fprintf(out, "\n");
2138         print_shared_cacheline_info(out);
2139         fprintf(out, "\n");
2140         print_c2c_info(out, session);
2141
2142         if (c2c.stats_only)
2143                 return;
2144
2145         fprintf(out, "\n");
2146         fprintf(out, "=================================================\n");
2147         fprintf(out, "           Shared Data Cache Line Table          \n");
2148         fprintf(out, "=================================================\n");
2149         fprintf(out, "#\n");
2150
2151         hists__fprintf(&c2c.hists.hists, true, 0, 0, 0, stdout, false);
2152
2153         fprintf(out, "\n");
2154         fprintf(out, "=================================================\n");
2155         fprintf(out, "      Shared Cache Line Distribution Pareto      \n");
2156         fprintf(out, "=================================================\n");
2157         fprintf(out, "#\n");
2158
2159         print_pareto(out);
2160 }
2161
2162 #ifdef HAVE_SLANG_SUPPORT
2163 static void c2c_browser__update_nr_entries(struct hist_browser *hb)
2164 {
2165         u64 nr_entries = 0;
2166         struct rb_node *nd = rb_first(&hb->hists->entries);
2167
2168         while (nd) {
2169                 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
2170
2171                 if (!he->filtered)
2172                         nr_entries++;
2173
2174                 nd = rb_next(nd);
2175         }
2176
2177         hb->nr_non_filtered_entries = nr_entries;
2178 }
2179
2180 struct c2c_cacheline_browser {
2181         struct hist_browser      hb;
2182         struct hist_entry       *he;
2183 };
2184
2185 static int
2186 perf_c2c_cacheline_browser__title(struct hist_browser *browser,
2187                                   char *bf, size_t size)
2188 {
2189         struct c2c_cacheline_browser *cl_browser;
2190         struct hist_entry *he;
2191         uint64_t addr = 0;
2192
2193         cl_browser = container_of(browser, struct c2c_cacheline_browser, hb);
2194         he = cl_browser->he;
2195
2196         if (he->mem_info)
2197                 addr = cl_address(he->mem_info->daddr.addr);
2198
2199         scnprintf(bf, size, "Cacheline 0x%lx", addr);
2200         return 0;
2201 }
2202
2203 static struct c2c_cacheline_browser*
2204 c2c_cacheline_browser__new(struct hists *hists, struct hist_entry *he)
2205 {
2206         struct c2c_cacheline_browser *browser;
2207
2208         browser = zalloc(sizeof(*browser));
2209         if (browser) {
2210                 hist_browser__init(&browser->hb, hists);
2211                 browser->hb.c2c_filter  = true;
2212                 browser->hb.title       = perf_c2c_cacheline_browser__title;
2213                 browser->he             = he;
2214         }
2215
2216         return browser;
2217 }
2218
2219 static int perf_c2c__browse_cacheline(struct hist_entry *he)
2220 {
2221         struct c2c_hist_entry *c2c_he;
2222         struct c2c_hists *c2c_hists;
2223         struct c2c_cacheline_browser *cl_browser;
2224         struct hist_browser *browser;
2225         int key = -1;
2226         const char help[] =
2227         " ENTER         Togle callchains (if present) \n"
2228         " n             Togle Node details info \n"
2229         " s             Togle full lenght of symbol and source line columns \n"
2230         " q             Return back to cacheline list \n";
2231
2232         /* Display compact version first. */
2233         c2c.symbol_full = false;
2234
2235         c2c_he = container_of(he, struct c2c_hist_entry, he);
2236         c2c_hists = c2c_he->hists;
2237
2238         cl_browser = c2c_cacheline_browser__new(&c2c_hists->hists, he);
2239         if (cl_browser == NULL)
2240                 return -1;
2241
2242         browser = &cl_browser->hb;
2243
2244         /* reset abort key so that it can get Ctrl-C as a key */
2245         SLang_reset_tty();
2246         SLang_init_tty(0, 0, 0);
2247
2248         c2c_browser__update_nr_entries(browser);
2249
2250         while (1) {
2251                 key = hist_browser__run(browser, "? - help");
2252
2253                 switch (key) {
2254                 case 's':
2255                         c2c.symbol_full = !c2c.symbol_full;
2256                         break;
2257                 case 'n':
2258                         c2c.node_info = (c2c.node_info + 1) % 3;
2259                         setup_nodes_header();
2260                         break;
2261                 case 'q':
2262                         goto out;
2263                 case '?':
2264                         ui_browser__help_window(&browser->b, help);
2265                         break;
2266                 default:
2267                         break;
2268                 }
2269         }
2270
2271 out:
2272         free(cl_browser);
2273         return 0;
2274 }
2275
2276 static int perf_c2c_browser__title(struct hist_browser *browser,
2277                                    char *bf, size_t size)
2278 {
2279         scnprintf(bf, size,
2280                   "Shared Data Cache Line Table     "
2281                   "(%lu entries, sorted on %s HITMs)",
2282                   browser->nr_non_filtered_entries,
2283                   display_str[c2c.display]);
2284         return 0;
2285 }
2286
2287 static struct hist_browser*
2288 perf_c2c_browser__new(struct hists *hists)
2289 {
2290         struct hist_browser *browser = hist_browser__new(hists);
2291
2292         if (browser) {
2293                 browser->title = perf_c2c_browser__title;
2294                 browser->c2c_filter = true;
2295         }
2296
2297         return browser;
2298 }
2299
2300 static int perf_c2c__hists_browse(struct hists *hists)
2301 {
2302         struct hist_browser *browser;
2303         int key = -1;
2304         const char help[] =
2305         " d             Display cacheline details \n"
2306         " ENTER         Togle callchains (if present) \n"
2307         " q             Quit \n";
2308
2309         browser = perf_c2c_browser__new(hists);
2310         if (browser == NULL)
2311                 return -1;
2312
2313         /* reset abort key so that it can get Ctrl-C as a key */
2314         SLang_reset_tty();
2315         SLang_init_tty(0, 0, 0);
2316
2317         c2c_browser__update_nr_entries(browser);
2318
2319         while (1) {
2320                 key = hist_browser__run(browser, "? - help");
2321
2322                 switch (key) {
2323                 case 'q':
2324                         goto out;
2325                 case 'd':
2326                         perf_c2c__browse_cacheline(browser->he_selection);
2327                         break;
2328                 case '?':
2329                         ui_browser__help_window(&browser->b, help);
2330                         break;
2331                 default:
2332                         break;
2333                 }
2334         }
2335
2336 out:
2337         hist_browser__delete(browser);
2338         return 0;
2339 }
2340
2341 static void perf_c2c_display(struct perf_session *session)
2342 {
2343         if (use_browser == 0)
2344                 perf_c2c__hists_fprintf(stdout, session);
2345         else
2346                 perf_c2c__hists_browse(&c2c.hists.hists);
2347 }
2348 #else
2349 static void perf_c2c_display(struct perf_session *session)
2350 {
2351         use_browser = 0;
2352         perf_c2c__hists_fprintf(stdout, session);
2353 }
2354 #endif /* HAVE_SLANG_SUPPORT */
2355
2356 static void ui_quirks(void)
2357 {
2358         if (!c2c.use_stdio) {
2359                 dim_offset.width  = 5;
2360                 dim_offset.header = header_offset_tui;
2361         }
2362
2363         dim_percent_hitm.header = percent_hitm_header[c2c.display];
2364 }
2365
2366 #define CALLCHAIN_DEFAULT_OPT  "graph,0.5,caller,function,percent"
2367
2368 const char callchain_help[] = "Display call graph (stack chain/backtrace):\n\n"
2369                                 CALLCHAIN_REPORT_HELP
2370                                 "\n\t\t\t\tDefault: " CALLCHAIN_DEFAULT_OPT;
2371
2372 static int
2373 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
2374 {
2375         struct callchain_param *callchain = opt->value;
2376
2377         callchain->enabled = !unset;
2378         /*
2379          * --no-call-graph
2380          */
2381         if (unset) {
2382                 symbol_conf.use_callchain = false;
2383                 callchain->mode = CHAIN_NONE;
2384                 return 0;
2385         }
2386
2387         return parse_callchain_report_opt(arg);
2388 }
2389
2390 static int setup_callchain(struct perf_evlist *evlist)
2391 {
2392         u64 sample_type = perf_evlist__combined_sample_type(evlist);
2393         enum perf_call_graph_mode mode = CALLCHAIN_NONE;
2394
2395         if ((sample_type & PERF_SAMPLE_REGS_USER) &&
2396             (sample_type & PERF_SAMPLE_STACK_USER))
2397                 mode = CALLCHAIN_DWARF;
2398         else if (sample_type & PERF_SAMPLE_BRANCH_STACK)
2399                 mode = CALLCHAIN_LBR;
2400         else if (sample_type & PERF_SAMPLE_CALLCHAIN)
2401                 mode = CALLCHAIN_FP;
2402
2403         if (!callchain_param.enabled &&
2404             callchain_param.mode != CHAIN_NONE &&
2405             mode != CALLCHAIN_NONE) {
2406                 symbol_conf.use_callchain = true;
2407                 if (callchain_register_param(&callchain_param) < 0) {
2408                         ui__error("Can't register callchain params.\n");
2409                         return -EINVAL;
2410                 }
2411         }
2412
2413         callchain_param.record_mode = mode;
2414         callchain_param.min_percent = 0;
2415         return 0;
2416 }
2417
2418 static int setup_display(const char *str)
2419 {
2420         const char *display = str ?: "tot";
2421
2422         if (!strcmp(display, "tot"))
2423                 c2c.display = DISPLAY_TOT;
2424         else if (!strcmp(display, "rmt"))
2425                 c2c.display = DISPLAY_RMT;
2426         else if (!strcmp(display, "lcl"))
2427                 c2c.display = DISPLAY_LCL;
2428         else {
2429                 pr_err("failed: unknown display type: %s\n", str);
2430                 return -1;
2431         }
2432
2433         return 0;
2434 }
2435
2436 #define for_each_token(__tok, __buf, __sep, __tmp)              \
2437         for (__tok = strtok_r(__buf, __sep, &__tmp); __tok;     \
2438              __tok = strtok_r(NULL,  __sep, &__tmp))
2439
2440 static int build_cl_output(char *cl_sort, bool no_source)
2441 {
2442         char *tok, *tmp, *buf = strdup(cl_sort);
2443         bool add_pid   = false;
2444         bool add_tid   = false;
2445         bool add_iaddr = false;
2446         bool add_sym   = false;
2447         bool add_dso   = false;
2448         bool add_src   = false;
2449
2450         if (!buf)
2451                 return -ENOMEM;
2452
2453         for_each_token(tok, buf, ",", tmp) {
2454                 if (!strcmp(tok, "tid")) {
2455                         add_tid = true;
2456                 } else if (!strcmp(tok, "pid")) {
2457                         add_pid = true;
2458                 } else if (!strcmp(tok, "iaddr")) {
2459                         add_iaddr = true;
2460                         add_sym   = true;
2461                         add_dso   = true;
2462                         add_src   = no_source ? false : true;
2463                 } else if (!strcmp(tok, "dso")) {
2464                         add_dso = true;
2465                 } else if (strcmp(tok, "offset")) {
2466                         pr_err("unrecognized sort token: %s\n", tok);
2467                         return -EINVAL;
2468                 }
2469         }
2470
2471         if (asprintf(&c2c.cl_output,
2472                 "%s%s%s%s%s%s%s%s%s%s",
2473                 c2c.use_stdio ? "cl_num_empty," : "",
2474                 "percent_rmt_hitm,"
2475                 "percent_lcl_hitm,"
2476                 "percent_stores_l1hit,"
2477                 "percent_stores_l1miss,"
2478                 "offset,",
2479                 add_pid   ? "pid," : "",
2480                 add_tid   ? "tid," : "",
2481                 add_iaddr ? "iaddr," : "",
2482                 "mean_rmt,"
2483                 "mean_lcl,"
2484                 "mean_load,"
2485                 "tot_recs,"
2486                 "cpucnt,",
2487                 add_sym ? "symbol," : "",
2488                 add_dso ? "dso," : "",
2489                 add_src ? "cl_srcline," : "",
2490                 "node") < 0)
2491                 return -ENOMEM;
2492
2493         c2c.show_src = add_src;
2494
2495         free(buf);
2496         return 0;
2497 }
2498
2499 static int setup_coalesce(const char *coalesce, bool no_source)
2500 {
2501         const char *c = coalesce ?: coalesce_default;
2502
2503         if (asprintf(&c2c.cl_sort, "offset,%s", c) < 0)
2504                 return -ENOMEM;
2505
2506         if (build_cl_output(c2c.cl_sort, no_source))
2507                 return -1;
2508
2509         if (asprintf(&c2c.cl_resort, "offset,%s",
2510                      c2c.display == DISPLAY_TOT ?
2511                      "tot_hitm" :
2512                      c2c.display == DISPLAY_RMT ?
2513                      "rmt_hitm,lcl_hitm" :
2514                      "lcl_hitm,rmt_hitm") < 0)
2515                 return -ENOMEM;
2516
2517         pr_debug("coalesce sort   fields: %s\n", c2c.cl_sort);
2518         pr_debug("coalesce resort fields: %s\n", c2c.cl_resort);
2519         pr_debug("coalesce output fields: %s\n", c2c.cl_output);
2520         return 0;
2521 }
2522
2523 static int perf_c2c__report(int argc, const char **argv)
2524 {
2525         struct perf_session *session;
2526         struct ui_progress prog;
2527         struct perf_data_file file = {
2528                 .mode = PERF_DATA_MODE_READ,
2529         };
2530         char callchain_default_opt[] = CALLCHAIN_DEFAULT_OPT;
2531         const char *display = NULL;
2532         const char *coalesce = NULL;
2533         bool no_source = false;
2534         const struct option options[] = {
2535         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
2536                    "file", "vmlinux pathname"),
2537         OPT_STRING('i', "input", &input_name, "file",
2538                    "the input file to process"),
2539         OPT_INCR('N', "node-info", &c2c.node_info,
2540                  "show extra node info in report (repeat for more info)"),
2541 #ifdef HAVE_SLANG_SUPPORT
2542         OPT_BOOLEAN(0, "stdio", &c2c.use_stdio, "Use the stdio interface"),
2543 #endif
2544         OPT_BOOLEAN(0, "stats", &c2c.stats_only,
2545                     "Display only statistic tables (implies --stdio)"),
2546         OPT_BOOLEAN(0, "full-symbols", &c2c.symbol_full,
2547                     "Display full length of symbols"),
2548         OPT_BOOLEAN(0, "no-source", &no_source,
2549                     "Do not display Source Line column"),
2550         OPT_BOOLEAN(0, "show-all", &c2c.show_all,
2551                     "Show all captured HITM lines."),
2552         OPT_CALLBACK_DEFAULT('g', "call-graph", &callchain_param,
2553                              "print_type,threshold[,print_limit],order,sort_key[,branch],value",
2554                              callchain_help, &parse_callchain_opt,
2555                              callchain_default_opt),
2556         OPT_STRING('d', "display", &display, "Switch HITM output type", "lcl,rmt"),
2557         OPT_STRING('c', "coalesce", &coalesce, "coalesce fields",
2558                    "coalesce fields: pid,tid,iaddr,dso"),
2559         OPT_BOOLEAN('f', "force", &symbol_conf.force, "don't complain, do it"),
2560         OPT_PARENT(c2c_options),
2561         OPT_END()
2562         };
2563         int err = 0;
2564
2565         argc = parse_options(argc, argv, options, report_c2c_usage,
2566                              PARSE_OPT_STOP_AT_NON_OPTION);
2567         if (argc)
2568                 usage_with_options(report_c2c_usage, options);
2569
2570         if (c2c.stats_only)
2571                 c2c.use_stdio = true;
2572
2573         if (!input_name || !strlen(input_name))
2574                 input_name = "perf.data";
2575
2576         file.path  = input_name;
2577         file.force = symbol_conf.force;
2578
2579         err = setup_display(display);
2580         if (err)
2581                 goto out;
2582
2583         err = setup_coalesce(coalesce, no_source);
2584         if (err) {
2585                 pr_debug("Failed to initialize hists\n");
2586                 goto out;
2587         }
2588
2589         err = c2c_hists__init(&c2c.hists, "dcacheline", 2);
2590         if (err) {
2591                 pr_debug("Failed to initialize hists\n");
2592                 goto out;
2593         }
2594
2595         session = perf_session__new(&file, 0, &c2c.tool);
2596         if (session == NULL) {
2597                 pr_debug("No memory for session\n");
2598                 goto out;
2599         }
2600
2601         err = setup_nodes(session);
2602         if (err) {
2603                 pr_err("Failed setup nodes\n");
2604                 goto out;
2605         }
2606
2607         err = setup_callchain(session->evlist);
2608         if (err)
2609                 goto out_session;
2610
2611         if (symbol__init(&session->header.env) < 0)
2612                 goto out_session;
2613
2614         /* No pipe support at the moment. */
2615         if (perf_data_file__is_pipe(session->file)) {
2616                 pr_debug("No pipe support at the moment.\n");
2617                 goto out_session;
2618         }
2619
2620         if (c2c.use_stdio)
2621                 use_browser = 0;
2622         else
2623                 use_browser = 1;
2624
2625         setup_browser(false);
2626
2627         err = perf_session__process_events(session);
2628         if (err) {
2629                 pr_err("failed to process sample\n");
2630                 goto out_session;
2631         }
2632
2633         c2c_hists__reinit(&c2c.hists,
2634                         "cl_idx,"
2635                         "dcacheline,"
2636                         "tot_recs,"
2637                         "percent_hitm,"
2638                         "tot_hitm,lcl_hitm,rmt_hitm,"
2639                         "stores,stores_l1hit,stores_l1miss,"
2640                         "dram_lcl,dram_rmt,"
2641                         "ld_llcmiss,"
2642                         "tot_loads,"
2643                         "ld_fbhit,ld_l1hit,ld_l2hit,"
2644                         "ld_lclhit,ld_rmthit",
2645                         c2c.display == DISPLAY_TOT ? "tot_hitm" :
2646                         c2c.display == DISPLAY_LCL ? "lcl_hitm" : "rmt_hitm"
2647                         );
2648
2649         ui_progress__init(&prog, c2c.hists.hists.nr_entries, "Sorting...");
2650
2651         hists__collapse_resort(&c2c.hists.hists, NULL);
2652         hists__output_resort_cb(&c2c.hists.hists, &prog, resort_hitm_cb);
2653         hists__iterate_cb(&c2c.hists.hists, resort_cl_cb);
2654
2655         ui_progress__finish();
2656
2657         ui_quirks();
2658
2659         perf_c2c_display(session);
2660
2661 out_session:
2662         perf_session__delete(session);
2663 out:
2664         return err;
2665 }
2666
2667 static int parse_record_events(const struct option *opt,
2668                                const char *str, int unset __maybe_unused)
2669 {
2670         bool *event_set = (bool *) opt->value;
2671
2672         *event_set = true;
2673         return perf_mem_events__parse(str);
2674 }
2675
2676
2677 static const char * const __usage_record[] = {
2678         "perf c2c record [<options>] [<command>]",
2679         "perf c2c record [<options>] -- <command> [<options>]",
2680         NULL
2681 };
2682
2683 static const char * const *record_mem_usage = __usage_record;
2684
2685 static int perf_c2c__record(int argc, const char **argv)
2686 {
2687         int rec_argc, i = 0, j;
2688         const char **rec_argv;
2689         int ret;
2690         bool all_user = false, all_kernel = false;
2691         bool event_set = false;
2692         struct option options[] = {
2693         OPT_CALLBACK('e', "event", &event_set, "event",
2694                      "event selector. Use 'perf mem record -e list' to list available events",
2695                      parse_record_events),
2696         OPT_BOOLEAN('u', "all-user", &all_user, "collect only user level data"),
2697         OPT_BOOLEAN('k', "all-kernel", &all_kernel, "collect only kernel level data"),
2698         OPT_UINTEGER('l', "ldlat", &perf_mem_events__loads_ldlat, "setup mem-loads latency"),
2699         OPT_PARENT(c2c_options),
2700         OPT_END()
2701         };
2702
2703         if (perf_mem_events__init()) {
2704                 pr_err("failed: memory events not supported\n");
2705                 return -1;
2706         }
2707
2708         argc = parse_options(argc, argv, options, record_mem_usage,
2709                              PARSE_OPT_KEEP_UNKNOWN);
2710
2711         rec_argc = argc + 10; /* max number of arguments */
2712         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2713         if (!rec_argv)
2714                 return -1;
2715
2716         rec_argv[i++] = "record";
2717
2718         if (!event_set) {
2719                 perf_mem_events[PERF_MEM_EVENTS__LOAD].record  = true;
2720                 perf_mem_events[PERF_MEM_EVENTS__STORE].record = true;
2721         }
2722
2723         if (perf_mem_events[PERF_MEM_EVENTS__LOAD].record)
2724                 rec_argv[i++] = "-W";
2725
2726         rec_argv[i++] = "-d";
2727         rec_argv[i++] = "--sample-cpu";
2728
2729         for (j = 0; j < PERF_MEM_EVENTS__MAX; j++) {
2730                 if (!perf_mem_events[j].record)
2731                         continue;
2732
2733                 if (!perf_mem_events[j].supported) {
2734                         pr_err("failed: event '%s' not supported\n",
2735                                perf_mem_events[j].name);
2736                         return -1;
2737                 }
2738
2739                 rec_argv[i++] = "-e";
2740                 rec_argv[i++] = perf_mem_events__name(j);
2741         };
2742
2743         if (all_user)
2744                 rec_argv[i++] = "--all-user";
2745
2746         if (all_kernel)
2747                 rec_argv[i++] = "--all-kernel";
2748
2749         for (j = 0; j < argc; j++, i++)
2750                 rec_argv[i] = argv[j];
2751
2752         if (verbose > 0) {
2753                 pr_debug("calling: ");
2754
2755                 j = 0;
2756
2757                 while (rec_argv[j]) {
2758                         pr_debug("%s ", rec_argv[j]);
2759                         j++;
2760                 }
2761                 pr_debug("\n");
2762         }
2763
2764         ret = cmd_record(i, rec_argv);
2765         free(rec_argv);
2766         return ret;
2767 }
2768
2769 int cmd_c2c(int argc, const char **argv)
2770 {
2771         argc = parse_options(argc, argv, c2c_options, c2c_usage,
2772                              PARSE_OPT_STOP_AT_NON_OPTION);
2773
2774         if (!argc)
2775                 usage_with_options(c2c_usage, c2c_options);
2776
2777         if (!strncmp(argv[0], "rec", 3)) {
2778                 return perf_c2c__record(argc, argv);
2779         } else if (!strncmp(argv[0], "rep", 3)) {
2780                 return perf_c2c__report(argc, argv);
2781         } else {
2782                 usage_with_options(c2c_usage, c2c_options);
2783         }
2784
2785         return 0;
2786 }