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