tools/perf/builtin-c2c.c

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