tools/perf/builtin-report.c

   1 /*
   2  * builtin-report.c
   3  *
   4  * Builtin report command: Analyze the perf.data input file,
   5  * look up and read DSOs and symbol information and display
   6  * a histogram of results, along various sorting keys.
   7  */
   8 #include "builtin.h"
   9
  10 #include "util/util.h"
  11
  12 #include "util/color.h"
  13 #include <linux/list.h>
  14 #include "util/cache.h"
  15 #include <linux/rbtree.h>
  16 #include "util/symbol.h"
  17 #include "util/string.h"
  18 #include "util/callchain.h"
  19 #include "util/strlist.h"
  20 #include "util/values.h"
  21
  22 #include "perf.h"
  23 #include "util/debug.h"
  24 #include "util/header.h"
  25
  26 #include "util/parse-options.h"
  27 #include "util/parse-events.h"
  28
  29 #include "util/thread.h"
  30 #include "util/sort.h"
  31
  32 static char             const *input_name = "perf.data";
  33
  34 static char             *dso_list_str, *comm_list_str, *sym_list_str,
  35                         *col_width_list_str;
  36 static struct strlist   *dso_list, *comm_list, *sym_list;
  37
  38 static int              force;
  39 static int              input;
  40 static int              show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV;
  41
  42 static int              full_paths;
  43 static int              show_nr_samples;
  44
  45 static int              show_threads;
  46 static struct perf_read_values  show_threads_values;
  47
  48 static char             default_pretty_printing_style[] = "normal";
  49 static char             *pretty_printing_style = default_pretty_printing_style;
  50
  51 static unsigned long    page_size;
  52 static unsigned long    mmap_window = 32;
  53
  54 static int              exclude_other = 1;
  55
  56 static char             callchain_default_opt[] = "fractal,0.5";
  57
  58 static int              callchain;
  59
  60 static char             __cwd[PATH_MAX];
  61 static char             *cwd = __cwd;
  62 static int              cwdlen;
  63
  64 static struct rb_root   threads;
  65 static struct thread    *last_match;
  66
  67 static struct perf_header *header;
  68
  69 static
  70 struct callchain_param  callchain_param = {
  71         .mode   = CHAIN_GRAPH_REL,
  72         .min_percent = 0.5
  73 };
  74
  75 static u64              sample_type;
  76
  77 static struct rb_root hist;
  78
  79 static int64_t
  80 hist_entry__cmp(struct hist_entry *left, struct hist_entry *right)
  81 {
  82         struct sort_entry *se;
  83         int64_t cmp = 0;
  84
  85         list_for_each_entry(se, &hist_entry__sort_list, list) {
  86                 cmp = se->cmp(left, right);
  87                 if (cmp)
  88                         break;
  89         }
  90
  91         return cmp;
  92 }
  93
  94 static int64_t
  95 hist_entry__collapse(struct hist_entry *left, struct hist_entry *right)
  96 {
  97         struct sort_entry *se;
  98         int64_t cmp = 0;
  99
 100         list_for_each_entry(se, &hist_entry__sort_list, list) {
 101                 int64_t (*f)(struct hist_entry *, struct hist_entry *);
 102
 103                 f = se->collapse ?: se->cmp;
 104
 105                 cmp = f(left, right);
 106                 if (cmp)
 107                         break;
 108         }
 109
 110         return cmp;
 111 }
 112
 113 static size_t ipchain__fprintf_graph_line(FILE *fp, int depth, int depth_mask)
 114 {
 115         int i;
 116         size_t ret = 0;
 117
 118         ret += fprintf(fp, "%s", "                ");
 119
 120         for (i = 0; i < depth; i++)
 121                 if (depth_mask & (1 << i))
 122                         ret += fprintf(fp, "|          ");
 123                 else
 124                         ret += fprintf(fp, "           ");
 125
 126         ret += fprintf(fp, "\n");
 127
 128         return ret;
 129 }
 130 static size_t
 131 ipchain__fprintf_graph(FILE *fp, struct callchain_list *chain, int depth,
 132                        int depth_mask, int count, u64 total_samples,
 133                        int hits)
 134 {
 135         int i;
 136         size_t ret = 0;
 137
 138         ret += fprintf(fp, "%s", "                ");
 139         for (i = 0; i < depth; i++) {
 140                 if (depth_mask & (1 << i))
 141                         ret += fprintf(fp, "|");
 142                 else
 143                         ret += fprintf(fp, " ");
 144                 if (!count && i == depth - 1) {
 145                         double percent;
 146
 147                         percent = hits * 100.0 / total_samples;
 148                         ret += percent_color_fprintf(fp, "--%2.2f%%-- ", percent);
 149                 } else
 150                         ret += fprintf(fp, "%s", "          ");
 151         }
 152         if (chain->sym)
 153                 ret += fprintf(fp, "%s\n", chain->sym->name);
 154         else
 155                 ret += fprintf(fp, "%p\n", (void *)(long)chain->ip);
 156
 157         return ret;
 158 }
 159
 160 static struct symbol *rem_sq_bracket;
 161 static struct callchain_list rem_hits;
 162
 163 static void init_rem_hits(void)
 164 {
 165         rem_sq_bracket = malloc(sizeof(*rem_sq_bracket) + 6);
 166         if (!rem_sq_bracket) {
 167                 fprintf(stderr, "Not enough memory to display remaining hits\n");
 168                 return;
 169         }
 170
 171         strcpy(rem_sq_bracket->name, "[...]");
 172         rem_hits.sym = rem_sq_bracket;
 173 }
 174
 175 static size_t
 176 callchain__fprintf_graph(FILE *fp, struct callchain_node *self,
 177                         u64 total_samples, int depth, int depth_mask)
 178 {
 179         struct rb_node *node, *next;
 180         struct callchain_node *child;
 181         struct callchain_list *chain;
 182         int new_depth_mask = depth_mask;
 183         u64 new_total;
 184         u64 remaining;
 185         size_t ret = 0;
 186         int i;
 187
 188         if (callchain_param.mode == CHAIN_GRAPH_REL)
 189                 new_total = self->children_hit;
 190         else
 191                 new_total = total_samples;
 192
 193         remaining = new_total;
 194
 195         node = rb_first(&self->rb_root);
 196         while (node) {
 197                 u64 cumul;
 198
 199                 child = rb_entry(node, struct callchain_node, rb_node);
 200                 cumul = cumul_hits(child);
 201                 remaining -= cumul;
 202
 203                 /*
 204                  * The depth mask manages the output of pipes that show
 205                  * the depth. We don't want to keep the pipes of the current
 206                  * level for the last child of this depth.
 207                  * Except if we have remaining filtered hits. They will
 208                  * supersede the last child
 209                  */
 210                 next = rb_next(node);
 211                 if (!next && (callchain_param.mode != CHAIN_GRAPH_REL || !remaining))
 212                         new_depth_mask &= ~(1 << (depth - 1));
 213
 214                 /*
 215                  * But we keep the older depth mask for the line seperator
 216                  * to keep the level link until we reach the last child
 217                  */
 218                 ret += ipchain__fprintf_graph_line(fp, depth, depth_mask);
 219                 i = 0;
 220                 list_for_each_entry(chain, &child->val, list) {
 221                         if (chain->ip >= PERF_CONTEXT_MAX)
 222                                 continue;
 223                         ret += ipchain__fprintf_graph(fp, chain, depth,
 224                                                       new_depth_mask, i++,
 225                                                       new_total,
 226                                                       cumul);
 227                 }
 228                 ret += callchain__fprintf_graph(fp, child, new_total,
 229                                                 depth + 1,
 230                                                 new_depth_mask | (1 << depth));
 231                 node = next;
 232         }
 233
 234         if (callchain_param.mode == CHAIN_GRAPH_REL &&
 235                 remaining && remaining != new_total) {
 236
 237                 if (!rem_sq_bracket)
 238                         return ret;
 239
 240                 new_depth_mask &= ~(1 << (depth - 1));
 241
 242                 ret += ipchain__fprintf_graph(fp, &rem_hits, depth,
 243                                               new_depth_mask, 0, new_total,
 244                                               remaining);
 245         }
 246
 247         return ret;
 248 }
 249
 250 static size_t
 251 callchain__fprintf_flat(FILE *fp, struct callchain_node *self,
 252                         u64 total_samples)
 253 {
 254         struct callchain_list *chain;
 255         size_t ret = 0;
 256
 257         if (!self)
 258                 return 0;
 259
 260         ret += callchain__fprintf_flat(fp, self->parent, total_samples);
 261
 262
 263         list_for_each_entry(chain, &self->val, list) {
 264                 if (chain->ip >= PERF_CONTEXT_MAX)
 265                         continue;
 266                 if (chain->sym)
 267                         ret += fprintf(fp, "                %s\n", chain->sym->name);
 268                 else
 269                         ret += fprintf(fp, "                %p\n",
 270                                         (void *)(long)chain->ip);
 271         }
 272
 273         return ret;
 274 }
 275
 276 static size_t
 277 hist_entry_callchain__fprintf(FILE *fp, struct hist_entry *self,
 278                               u64 total_samples)
 279 {
 280         struct rb_node *rb_node;
 281         struct callchain_node *chain;
 282         size_t ret = 0;
 283
 284         rb_node = rb_first(&self->sorted_chain);
 285         while (rb_node) {
 286                 double percent;
 287
 288                 chain = rb_entry(rb_node, struct callchain_node, rb_node);
 289                 percent = chain->hit * 100.0 / total_samples;
 290                 switch (callchain_param.mode) {
 291                 case CHAIN_FLAT:
 292                         ret += percent_color_fprintf(fp, "           %6.2f%%\n",
 293                                                      percent);
 294                         ret += callchain__fprintf_flat(fp, chain, total_samples);
 295                         break;
 296                 case CHAIN_GRAPH_ABS: /* Falldown */
 297                 case CHAIN_GRAPH_REL:
 298                         ret += callchain__fprintf_graph(fp, chain,
 299                                                         total_samples, 1, 1);
 300                 case CHAIN_NONE:
 301                 default:
 302                         break;
 303                 }
 304                 ret += fprintf(fp, "\n");
 305                 rb_node = rb_next(rb_node);
 306         }
 307
 308         return ret;
 309 }
 310
 311
 312 static size_t
 313 hist_entry__fprintf(FILE *fp, struct hist_entry *self, u64 total_samples)
 314 {
 315         struct sort_entry *se;
 316         size_t ret;
 317
 318         if (exclude_other && !self->parent)
 319                 return 0;
 320
 321         if (total_samples)
 322                 ret = percent_color_fprintf(fp,
 323                                             field_sep ? "%.2f" : "   %6.2f%%",
 324                                         (self->count * 100.0) / total_samples);
 325         else
 326                 ret = fprintf(fp, field_sep ? "%lld" : "%12lld ", self->count);
 327
 328         if (show_nr_samples) {
 329                 if (field_sep)
 330                         fprintf(fp, "%c%lld", *field_sep, self->count);
 331                 else
 332                         fprintf(fp, "%11lld", self->count);
 333         }
 334
 335         list_for_each_entry(se, &hist_entry__sort_list, list) {
 336                 if (se->elide)
 337                         continue;
 338
 339                 fprintf(fp, "%s", field_sep ?: "  ");
 340                 ret += se->print(fp, self, se->width ? *se->width : 0);
 341         }
 342
 343         ret += fprintf(fp, "\n");
 344
 345         if (callchain)
 346                 hist_entry_callchain__fprintf(fp, self, total_samples);
 347
 348         return ret;
 349 }
 350
 351 /*
 352  *
 353  */
 354
 355 static void dso__calc_col_width(struct dso *self)
 356 {
 357         if (!col_width_list_str && !field_sep &&
 358             (!dso_list || strlist__has_entry(dso_list, self->name))) {
 359                 unsigned int slen = strlen(self->name);
 360                 if (slen > dsos__col_width)
 361                         dsos__col_width = slen;
 362         }
 363
 364         self->slen_calculated = 1;
 365 }
 366
 367 static void thread__comm_adjust(struct thread *self)
 368 {
 369         char *comm = self->comm;
 370
 371         if (!col_width_list_str && !field_sep &&
 372             (!comm_list || strlist__has_entry(comm_list, comm))) {
 373                 unsigned int slen = strlen(comm);
 374
 375                 if (slen > comms__col_width) {
 376                         comms__col_width = slen;
 377                         threads__col_width = slen + 6;
 378                 }
 379         }
 380 }
 381
 382 static int thread__set_comm_adjust(struct thread *self, const char *comm)
 383 {
 384         int ret = thread__set_comm(self, comm);
 385
 386         if (ret)
 387                 return ret;
 388
 389         thread__comm_adjust(self);
 390
 391         return 0;
 392 }
 393
 394
 395 static struct symbol *
 396 resolve_symbol(struct thread *thread, struct map **mapp,
 397                struct dso **dsop, u64 *ipp)
 398 {
 399         struct dso *dso = dsop ? *dsop : NULL;
 400         struct map *map = mapp ? *mapp : NULL;
 401         u64 ip = *ipp;
 402
 403         if (!thread)
 404                 return NULL;
 405
 406         if (dso)
 407                 goto got_dso;
 408
 409         if (map)
 410                 goto got_map;
 411
 412         map = thread__find_map(thread, ip);
 413         if (map != NULL) {
 414                 /*
 415                  * We have to do this here as we may have a dso
 416                  * with no symbol hit that has a name longer than
 417                  * the ones with symbols sampled.
 418                  */
 419                 if (!sort_dso.elide && !map->dso->slen_calculated)
 420                         dso__calc_col_width(map->dso);
 421
 422                 if (mapp)
 423                         *mapp = map;
 424 got_map:
 425                 ip = map->map_ip(map, ip);
 426
 427                 dso = map->dso;
 428         } else {
 429                 /*
 430                  * If this is outside of all known maps,
 431                  * and is a negative address, try to look it
 432                  * up in the kernel dso, as it might be a
 433                  * vsyscall (which executes in user-mode):
 434                  */
 435                 if ((long long)ip < 0)
 436                 dso = kernel_dso;
 437         }
 438         dump_printf(" ...... dso: %s\n", dso ? dso->name : "<not found>");
 439         dump_printf(" ...... map: %Lx -> %Lx\n", *ipp, ip);
 440         *ipp  = ip;
 441
 442         if (dsop)
 443                 *dsop = dso;
 444
 445         if (!dso)
 446                 return NULL;
 447 got_dso:
 448         return dso->find_symbol(dso, ip);
 449 }
 450
 451 static int call__match(struct symbol *sym)
 452 {
 453         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
 454                 return 1;
 455
 456         return 0;
 457 }
 458
 459 static struct symbol **
 460 resolve_callchain(struct thread *thread, struct map *map __used,
 461                     struct ip_callchain *chain, struct hist_entry *entry)
 462 {
 463         u64 context = PERF_CONTEXT_MAX;
 464         struct symbol **syms = NULL;
 465         unsigned int i;
 466
 467         if (callchain) {
 468                 syms = calloc(chain->nr, sizeof(*syms));
 469                 if (!syms) {
 470                         fprintf(stderr, "Can't allocate memory for symbols\n");
 471                         exit(-1);
 472                 }
 473         }
 474
 475         for (i = 0; i < chain->nr; i++) {
 476                 u64 ip = chain->ips[i];
 477                 struct dso *dso = NULL;
 478                 struct symbol *sym;
 479
 480                 if (ip >= PERF_CONTEXT_MAX) {
 481                         context = ip;
 482                         continue;
 483                 }
 484
 485                 switch (context) {
 486                 case PERF_CONTEXT_HV:
 487                         dso = hypervisor_dso;
 488                         break;
 489                 case PERF_CONTEXT_KERNEL:
 490                         dso = kernel_dso;
 491                         break;
 492                 default:
 493                         break;
 494                 }
 495
 496                 sym = resolve_symbol(thread, NULL, &dso, &ip);
 497
 498                 if (sym) {
 499                         if (sort__has_parent && call__match(sym) &&
 500                             !entry->parent)
 501                                 entry->parent = sym;
 502                         if (!callchain)
 503                                 break;
 504                         syms[i] = sym;
 505                 }
 506         }
 507
 508         return syms;
 509 }
 510
 511 /*
 512  * collect histogram counts
 513  */
 514
 515 static int
 516 hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
 517                 struct symbol *sym, u64 ip, struct ip_callchain *chain,
 518                 char level, u64 count)
 519 {
 520         struct rb_node **p = &hist.rb_node;
 521         struct rb_node *parent = NULL;
 522         struct hist_entry *he;
 523         struct symbol **syms = NULL;
 524         struct hist_entry entry = {
 525                 .thread = thread,
 526                 .map    = map,
 527                 .dso    = dso,
 528                 .sym    = sym,
 529                 .ip     = ip,
 530                 .level  = level,
 531                 .count  = count,
 532                 .parent = NULL,
 533                 .sorted_chain = RB_ROOT
 534         };
 535         int cmp;
 536
 537         if ((sort__has_parent || callchain) && chain)
 538                 syms = resolve_callchain(thread, map, chain, &entry);
 539
 540         while (*p != NULL) {
 541                 parent = *p;
 542                 he = rb_entry(parent, struct hist_entry, rb_node);
 543
 544                 cmp = hist_entry__cmp(&entry, he);
 545
 546                 if (!cmp) {
 547                         he->count += count;
 548                         if (callchain) {
 549                                 append_chain(&he->callchain, chain, syms);
 550                                 free(syms);
 551                         }
 552                         return 0;
 553                 }
 554
 555                 if (cmp < 0)
 556                         p = &(*p)->rb_left;
 557                 else
 558                         p = &(*p)->rb_right;
 559         }
 560
 561         he = malloc(sizeof(*he));
 562         if (!he)
 563                 return -ENOMEM;
 564         *he = entry;
 565         if (callchain) {
 566                 callchain_init(&he->callchain);
 567                 append_chain(&he->callchain, chain, syms);
 568                 free(syms);
 569         }
 570         rb_link_node(&he->rb_node, parent, p);
 571         rb_insert_color(&he->rb_node, &hist);
 572
 573         return 0;
 574 }
 575
 576 static void hist_entry__free(struct hist_entry *he)
 577 {
 578         free(he);
 579 }
 580
 581 /*
 582  * collapse the histogram
 583  */
 584
 585 static struct rb_root collapse_hists;
 586
 587 static void collapse__insert_entry(struct hist_entry *he)
 588 {
 589         struct rb_node **p = &collapse_hists.rb_node;
 590         struct rb_node *parent = NULL;
 591         struct hist_entry *iter;
 592         int64_t cmp;
 593
 594         while (*p != NULL) {
 595                 parent = *p;
 596                 iter = rb_entry(parent, struct hist_entry, rb_node);
 597
 598                 cmp = hist_entry__collapse(iter, he);
 599
 600                 if (!cmp) {
 601                         iter->count += he->count;
 602                         hist_entry__free(he);
 603                         return;
 604                 }
 605
 606                 if (cmp < 0)
 607                         p = &(*p)->rb_left;
 608                 else
 609                         p = &(*p)->rb_right;
 610         }
 611
 612         rb_link_node(&he->rb_node, parent, p);
 613         rb_insert_color(&he->rb_node, &collapse_hists);
 614 }
 615
 616 static void collapse__resort(void)
 617 {
 618         struct rb_node *next;
 619         struct hist_entry *n;
 620
 621         if (!sort__need_collapse)
 622                 return;
 623
 624         next = rb_first(&hist);
 625         while (next) {
 626                 n = rb_entry(next, struct hist_entry, rb_node);
 627                 next = rb_next(&n->rb_node);
 628
 629                 rb_erase(&n->rb_node, &hist);
 630                 collapse__insert_entry(n);
 631         }
 632 }
 633
 634 /*
 635  * reverse the map, sort on count.
 636  */
 637
 638 static struct rb_root output_hists;
 639
 640 static void output__insert_entry(struct hist_entry *he, u64 min_callchain_hits)
 641 {
 642         struct rb_node **p = &output_hists.rb_node;
 643         struct rb_node *parent = NULL;
 644         struct hist_entry *iter;
 645
 646         if (callchain)
 647                 callchain_param.sort(&he->sorted_chain, &he->callchain,
 648                                       min_callchain_hits, &callchain_param);
 649
 650         while (*p != NULL) {
 651                 parent = *p;
 652                 iter = rb_entry(parent, struct hist_entry, rb_node);
 653
 654                 if (he->count > iter->count)
 655                         p = &(*p)->rb_left;
 656                 else
 657                         p = &(*p)->rb_right;
 658         }
 659
 660         rb_link_node(&he->rb_node, parent, p);
 661         rb_insert_color(&he->rb_node, &output_hists);
 662 }
 663
 664 static void output__resort(u64 total_samples)
 665 {
 666         struct rb_node *next;
 667         struct hist_entry *n;
 668         struct rb_root *tree = &hist;
 669         u64 min_callchain_hits;
 670
 671         min_callchain_hits = total_samples * (callchain_param.min_percent / 100);
 672
 673         if (sort__need_collapse)
 674                 tree = &collapse_hists;
 675
 676         next = rb_first(tree);
 677
 678         while (next) {
 679                 n = rb_entry(next, struct hist_entry, rb_node);
 680                 next = rb_next(&n->rb_node);
 681
 682                 rb_erase(&n->rb_node, tree);
 683                 output__insert_entry(n, min_callchain_hits);
 684         }
 685 }
 686
 687 static size_t output__fprintf(FILE *fp, u64 total_samples)
 688 {
 689         struct hist_entry *pos;
 690         struct sort_entry *se;
 691         struct rb_node *nd;
 692         size_t ret = 0;
 693         unsigned int width;
 694         char *col_width = col_width_list_str;
 695         int raw_printing_style;
 696
 697         raw_printing_style = !strcmp(pretty_printing_style, "raw");
 698
 699         init_rem_hits();
 700
 701         fprintf(fp, "# Samples: %Ld\n", (u64)total_samples);
 702         fprintf(fp, "#\n");
 703
 704         fprintf(fp, "# Overhead");
 705         if (show_nr_samples) {
 706                 if (field_sep)
 707                         fprintf(fp, "%cSamples", *field_sep);
 708                 else
 709                         fputs("  Samples  ", fp);
 710         }
 711         list_for_each_entry(se, &hist_entry__sort_list, list) {
 712                 if (se->elide)
 713                         continue;
 714                 if (field_sep) {
 715                         fprintf(fp, "%c%s", *field_sep, se->header);
 716                         continue;
 717                 }
 718                 width = strlen(se->header);
 719                 if (se->width) {
 720                         if (col_width_list_str) {
 721                                 if (col_width) {
 722                                         *se->width = atoi(col_width);
 723                                         col_width = strchr(col_width, ',');
 724                                         if (col_width)
 725                                                 ++col_width;
 726                                 }
 727                         }
 728                         width = *se->width = max(*se->width, width);
 729                 }
 730                 fprintf(fp, "  %*s", width, se->header);
 731         }
 732         fprintf(fp, "\n");
 733
 734         if (field_sep)
 735                 goto print_entries;
 736
 737         fprintf(fp, "# ........");
 738         if (show_nr_samples)
 739                 fprintf(fp, " ..........");
 740         list_for_each_entry(se, &hist_entry__sort_list, list) {
 741                 unsigned int i;
 742
 743                 if (se->elide)
 744                         continue;
 745
 746                 fprintf(fp, "  ");
 747                 if (se->width)
 748                         width = *se->width;
 749                 else
 750                         width = strlen(se->header);
 751                 for (i = 0; i < width; i++)
 752                         fprintf(fp, ".");
 753         }
 754         fprintf(fp, "\n");
 755
 756         fprintf(fp, "#\n");
 757
 758 print_entries:
 759         for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
 760                 pos = rb_entry(nd, struct hist_entry, rb_node);
 761                 ret += hist_entry__fprintf(fp, pos, total_samples);
 762         }
 763
 764         if (sort_order == default_sort_order &&
 765                         parent_pattern == default_parent_pattern) {
 766                 fprintf(fp, "#\n");
 767                 fprintf(fp, "# (For a higher level overview, try: perf report --sort comm,dso)\n");
 768                 fprintf(fp, "#\n");
 769         }
 770         fprintf(fp, "\n");
 771
 772         free(rem_sq_bracket);
 773
 774         if (show_threads)
 775                 perf_read_values_display(fp, &show_threads_values,
 776                                          raw_printing_style);
 777
 778         return ret;
 779 }
 780
 781 static unsigned long total = 0,
 782                      total_mmap = 0,
 783                      total_comm = 0,
 784                      total_fork = 0,
 785                      total_unknown = 0,
 786                      total_lost = 0;
 787
 788 static int validate_chain(struct ip_callchain *chain, event_t *event)
 789 {
 790         unsigned int chain_size;
 791
 792         chain_size = event->header.size;
 793         chain_size -= (unsigned long)&event->ip.__more_data - (unsigned long)event;
 794
 795         if (chain->nr*sizeof(u64) > chain_size)
 796                 return -1;
 797
 798         return 0;
 799 }
 800
 801 static int
 802 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
 803 {
 804         char level;
 805         int show = 0;
 806         struct dso *dso = NULL;
 807         struct thread *thread;
 808         u64 ip = event->ip.ip;
 809         u64 period = 1;
 810         struct map *map = NULL;
 811         void *more_data = event->ip.__more_data;
 812         struct ip_callchain *chain = NULL;
 813         int cpumode;
 814
 815         thread = threads__findnew(event->ip.pid, &threads, &last_match);
 816
 817         if (sample_type & PERF_SAMPLE_PERIOD) {
 818                 period = *(u64 *)more_data;
 819                 more_data += sizeof(u64);
 820         }
 821
 822         dump_printf("%p [%p]: PERF_RECORD_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
 823                 (void *)(offset + head),
 824                 (void *)(long)(event->header.size),
 825                 event->header.misc,
 826                 event->ip.pid, event->ip.tid,
 827                 (void *)(long)ip,
 828                 (long long)period);
 829
 830         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
 831                 unsigned int i;
 832
 833                 chain = (void *)more_data;
 834
 835                 dump_printf("... chain: nr:%Lu\n", chain->nr);
 836
 837                 if (validate_chain(chain, event) < 0) {
 838                         eprintf("call-chain problem with event, skipping it.\n");
 839                         return 0;
 840                 }
 841
 842                 if (dump_trace) {
 843                         for (i = 0; i < chain->nr; i++)
 844                                 dump_printf("..... %2d: %016Lx\n", i, chain->ips[i]);
 845                 }
 846         }
 847
 848         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
 849
 850         if (thread == NULL) {
 851                 eprintf("problem processing %d event, skipping it.\n",
 852                         event->header.type);
 853                 return -1;
 854         }
 855
 856         if (comm_list && !strlist__has_entry(comm_list, thread->comm))
 857                 return 0;
 858
 859         cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 860
 861         if (cpumode == PERF_RECORD_MISC_KERNEL) {
 862                 show = SHOW_KERNEL;
 863                 level = 'k';
 864
 865                 dso = kernel_dso;
 866
 867                 dump_printf(" ...... dso: %s\n", dso->name);
 868
 869         } else if (cpumode == PERF_RECORD_MISC_USER) {
 870
 871                 show = SHOW_USER;
 872                 level = '.';
 873
 874         } else {
 875                 show = SHOW_HV;
 876                 level = 'H';
 877
 878                 dso = hypervisor_dso;
 879
 880                 dump_printf(" ...... dso: [hypervisor]\n");
 881         }
 882
 883         if (show & show_mask) {
 884                 struct symbol *sym = resolve_symbol(thread, &map, &dso, &ip);
 885
 886                 if (dso_list && (!dso || !dso->name ||
 887                                  !strlist__has_entry(dso_list, dso->name)))
 888                         return 0;
 889
 890                 if (sym_list && (!sym || !strlist__has_entry(sym_list, sym->name)))
 891                         return 0;
 892
 893                 if (hist_entry__add(thread, map, dso, sym, ip, chain, level, period)) {
 894                         eprintf("problem incrementing symbol count, skipping event\n");
 895                         return -1;
 896                 }
 897         }
 898         total += period;
 899
 900         return 0;
 901 }
 902
 903 static int
 904 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
 905 {
 906         struct thread *thread;
 907         struct map *map = map__new(&event->mmap, cwd, cwdlen);
 908
 909         thread = threads__findnew(event->mmap.pid, &threads, &last_match);
 910
 911         dump_printf("%p [%p]: PERF_RECORD_MMAP %d/%d: [%p(%p) @ %p]: %s\n",
 912                 (void *)(offset + head),
 913                 (void *)(long)(event->header.size),
 914                 event->mmap.pid,
 915                 event->mmap.tid,
 916                 (void *)(long)event->mmap.start,
 917                 (void *)(long)event->mmap.len,
 918                 (void *)(long)event->mmap.pgoff,
 919                 event->mmap.filename);
 920
 921         if (thread == NULL || map == NULL) {
 922                 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
 923                 return 0;
 924         }
 925
 926         thread__insert_map(thread, map);
 927         total_mmap++;
 928
 929         return 0;
 930 }
 931
 932 static int
 933 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
 934 {
 935         struct thread *thread;
 936
 937         thread = threads__findnew(event->comm.pid, &threads, &last_match);
 938
 939         dump_printf("%p [%p]: PERF_RECORD_COMM: %s:%d\n",
 940                 (void *)(offset + head),
 941                 (void *)(long)(event->header.size),
 942                 event->comm.comm, event->comm.pid);
 943
 944         if (thread == NULL ||
 945             thread__set_comm_adjust(thread, event->comm.comm)) {
 946                 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
 947                 return -1;
 948         }
 949         total_comm++;
 950
 951         return 0;
 952 }
 953
 954 static int
 955 process_task_event(event_t *event, unsigned long offset, unsigned long head)
 956 {
 957         struct thread *thread;
 958         struct thread *parent;
 959
 960         thread = threads__findnew(event->fork.pid, &threads, &last_match);
 961         parent = threads__findnew(event->fork.ppid, &threads, &last_match);
 962
 963         dump_printf("%p [%p]: PERF_RECORD_%s: (%d:%d):(%d:%d)\n",
 964                 (void *)(offset + head),
 965                 (void *)(long)(event->header.size),
 966                 event->header.type == PERF_RECORD_FORK ? "FORK" : "EXIT",
 967                 event->fork.pid, event->fork.tid,
 968                 event->fork.ppid, event->fork.ptid);
 969
 970         /*
 971          * A thread clone will have the same PID for both
 972          * parent and child.
 973          */
 974         if (thread == parent)
 975                 return 0;
 976
 977         if (event->header.type == PERF_RECORD_EXIT)
 978                 return 0;
 979
 980         if (!thread || !parent || thread__fork(thread, parent)) {
 981                 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
 982                 return -1;
 983         }
 984         total_fork++;
 985
 986         return 0;
 987 }
 988
 989 static int
 990 process_lost_event(event_t *event, unsigned long offset, unsigned long head)
 991 {
 992         dump_printf("%p [%p]: PERF_RECORD_LOST: id:%Ld: lost:%Ld\n",
 993                 (void *)(offset + head),
 994                 (void *)(long)(event->header.size),
 995                 event->lost.id,
 996                 event->lost.lost);
 997
 998         total_lost += event->lost.lost;
 999
1000         return 0;
1001 }
1002
1003 static int
1004 process_read_event(event_t *event, unsigned long offset, unsigned long head)
1005 {
1006         struct perf_event_attr *attr;
1007
1008         attr = perf_header__find_attr(event->read.id, header);
1009
1010         if (show_threads) {
1011                 const char *name = attr ? __event_name(attr->type, attr->config)
1012                                    : "unknown";
1013                 perf_read_values_add_value(&show_threads_values,
1014                                            event->read.pid, event->read.tid,
1015                                            event->read.id,
1016                                            name,
1017                                            event->read.value);
1018         }
1019
1020         dump_printf("%p [%p]: PERF_RECORD_READ: %d %d %s %Lu\n",
1021                         (void *)(offset + head),
1022                         (void *)(long)(event->header.size),
1023                         event->read.pid,
1024                         event->read.tid,
1025                         attr ? __event_name(attr->type, attr->config)
1026                              : "FAIL",
1027                         event->read.value);
1028
1029         return 0;
1030 }
1031
1032 static int
1033 process_event(event_t *event, unsigned long offset, unsigned long head)
1034 {
1035         trace_event(event);
1036
1037         switch (event->header.type) {
1038         case PERF_RECORD_SAMPLE:
1039                 return process_sample_event(event, offset, head);
1040
1041         case PERF_RECORD_MMAP:
1042                 return process_mmap_event(event, offset, head);
1043
1044         case PERF_RECORD_COMM:
1045                 return process_comm_event(event, offset, head);
1046
1047         case PERF_RECORD_FORK:
1048         case PERF_RECORD_EXIT:
1049                 return process_task_event(event, offset, head);
1050
1051         case PERF_RECORD_LOST:
1052                 return process_lost_event(event, offset, head);
1053
1054         case PERF_RECORD_READ:
1055                 return process_read_event(event, offset, head);
1056
1057         /*
1058          * We dont process them right now but they are fine:
1059          */
1060
1061         case PERF_RECORD_THROTTLE:
1062         case PERF_RECORD_UNTHROTTLE:
1063                 return 0;
1064
1065         default:
1066                 return -1;
1067         }
1068
1069         return 0;
1070 }
1071
1072 static int __cmd_report(void)
1073 {
1074         int ret, rc = EXIT_FAILURE;
1075         unsigned long offset = 0;
1076         unsigned long head, shift;
1077         struct stat input_stat;
1078         struct thread *idle;
1079         event_t *event;
1080         uint32_t size;
1081         char *buf;
1082
1083         idle = register_idle_thread(&threads, &last_match);
1084         thread__comm_adjust(idle);
1085
1086         if (show_threads)
1087                 perf_read_values_init(&show_threads_values);
1088
1089         input = open(input_name, O_RDONLY);
1090         if (input < 0) {
1091                 fprintf(stderr, " failed to open file: %s", input_name);
1092                 if (!strcmp(input_name, "perf.data"))
1093                         fprintf(stderr, "  (try 'perf record' first)");
1094                 fprintf(stderr, "\n");
1095                 exit(-1);
1096         }
1097
1098         ret = fstat(input, &input_stat);
1099         if (ret < 0) {
1100                 perror("failed to stat file");
1101                 exit(-1);
1102         }
1103
1104         if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
1105                 fprintf(stderr, "file: %s not owned by current user or root\n", input_name);
1106                 exit(-1);
1107         }
1108
1109         if (!input_stat.st_size) {
1110                 fprintf(stderr, "zero-sized file, nothing to do!\n");
1111                 exit(0);
1112         }
1113
1114         header = perf_header__read(input);
1115         head = header->data_offset;
1116
1117         sample_type = perf_header__sample_type(header);
1118
1119         if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
1120                 if (sort__has_parent) {
1121                         fprintf(stderr, "selected --sort parent, but no"
1122                                         " callchain data. Did you call"
1123                                         " perf record without -g?\n");
1124                         exit(-1);
1125                 }
1126                 if (callchain) {
1127                         fprintf(stderr, "selected -g but no callchain data."
1128                                         " Did you call perf record without"
1129                                         " -g?\n");
1130                         exit(-1);
1131                 }
1132         } else if (callchain_param.mode != CHAIN_NONE && !callchain) {
1133                         callchain = 1;
1134                         if (register_callchain_param(&callchain_param) < 0) {
1135                                 fprintf(stderr, "Can't register callchain"
1136                                                 " params\n");
1137                                 exit(-1);
1138                         }
1139         }
1140
1141         if (load_kernel() < 0) {
1142                 perror("failed to load kernel symbols");
1143                 return EXIT_FAILURE;
1144         }
1145
1146         if (!full_paths) {
1147                 if (getcwd(__cwd, sizeof(__cwd)) == NULL) {
1148                         perror("failed to get the current directory");
1149                         return EXIT_FAILURE;
1150                 }
1151                 cwdlen = strlen(cwd);
1152         } else {
1153                 cwd = NULL;
1154                 cwdlen = 0;
1155         }
1156
1157         shift = page_size * (head / page_size);
1158         offset += shift;
1159         head -= shift;
1160
1161 remap:
1162         buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1163                            MAP_SHARED, input, offset);
1164         if (buf == MAP_FAILED) {
1165                 perror("failed to mmap file");
1166                 exit(-1);
1167         }
1168
1169 more:
1170         event = (event_t *)(buf + head);
1171
1172         size = event->header.size;
1173         if (!size)
1174                 size = 8;
1175
1176         if (head + event->header.size >= page_size * mmap_window) {
1177                 int munmap_ret;
1178
1179                 shift = page_size * (head / page_size);
1180
1181                 munmap_ret = munmap(buf, page_size * mmap_window);
1182                 assert(munmap_ret == 0);
1183
1184                 offset += shift;
1185                 head -= shift;
1186                 goto remap;
1187         }
1188
1189         size = event->header.size;
1190
1191         dump_printf("\n%p [%p]: event: %d\n",
1192                         (void *)(offset + head),
1193                         (void *)(long)event->header.size,
1194                         event->header.type);
1195
1196         if (!size || process_event(event, offset, head) < 0) {
1197
1198                 dump_printf("%p [%p]: skipping unknown header type: %d\n",
1199                         (void *)(offset + head),
1200                         (void *)(long)(event->header.size),
1201                         event->header.type);
1202
1203                 total_unknown++;
1204
1205                 /*
1206                  * assume we lost track of the stream, check alignment, and
1207                  * increment a single u64 in the hope to catch on again 'soon'.
1208                  */
1209
1210                 if (unlikely(head & 7))
1211                         head &= ~7ULL;
1212
1213                 size = 8;
1214         }
1215
1216         head += size;
1217
1218         if (offset + head >= header->data_offset + header->data_size)
1219                 goto done;
1220
1221         if (offset + head < (unsigned long)input_stat.st_size)
1222                 goto more;
1223
1224 done:
1225         rc = EXIT_SUCCESS;
1226         close(input);
1227
1228         dump_printf("      IP events: %10ld\n", total);
1229         dump_printf("    mmap events: %10ld\n", total_mmap);
1230         dump_printf("    comm events: %10ld\n", total_comm);
1231         dump_printf("    fork events: %10ld\n", total_fork);
1232         dump_printf("    lost events: %10ld\n", total_lost);
1233         dump_printf(" unknown events: %10ld\n", total_unknown);
1234
1235         if (dump_trace)
1236                 return 0;
1237
1238         if (verbose >= 3)
1239                 threads__fprintf(stdout, &threads);
1240
1241         if (verbose >= 2)
1242                 dsos__fprintf(stdout);
1243
1244         collapse__resort();
1245         output__resort(total);
1246         output__fprintf(stdout, total);
1247
1248         if (show_threads)
1249                 perf_read_values_destroy(&show_threads_values);
1250
1251         return rc;
1252 }
1253
1254 static int
1255 parse_callchain_opt(const struct option *opt __used, const char *arg,
1256                     int unset __used)
1257 {
1258         char *tok;
1259         char *endptr;
1260
1261         callchain = 1;
1262
1263         if (!arg)
1264                 return 0;
1265
1266         tok = strtok((char *)arg, ",");
1267         if (!tok)
1268                 return -1;
1269
1270         /* get the output mode */
1271         if (!strncmp(tok, "graph", strlen(arg)))
1272                 callchain_param.mode = CHAIN_GRAPH_ABS;
1273
1274         else if (!strncmp(tok, "flat", strlen(arg)))
1275                 callchain_param.mode = CHAIN_FLAT;
1276
1277         else if (!strncmp(tok, "fractal", strlen(arg)))
1278                 callchain_param.mode = CHAIN_GRAPH_REL;
1279
1280         else if (!strncmp(tok, "none", strlen(arg))) {
1281                 callchain_param.mode = CHAIN_NONE;
1282                 callchain = 0;
1283
1284                 return 0;
1285         }
1286
1287         else
1288                 return -1;
1289
1290         /* get the min percentage */
1291         tok = strtok(NULL, ",");
1292         if (!tok)
1293                 goto setup;
1294
1295         callchain_param.min_percent = strtod(tok, &endptr);
1296         if (tok == endptr)
1297                 return -1;
1298
1299 setup:
1300         if (register_callchain_param(&callchain_param) < 0) {
1301                 fprintf(stderr, "Can't register callchain params\n");
1302                 return -1;
1303         }
1304         return 0;
1305 }
1306
1307 //static const char * const report_usage[] = {
1308 const char * const report_usage[] = {
1309         "perf report [<options>] <command>",
1310         NULL
1311 };
1312
1313 static const struct option options[] = {
1314         OPT_STRING('i', "input", &input_name, "file",
1315                     "input file name"),
1316         OPT_BOOLEAN('v', "verbose", &verbose,
1317                     "be more verbose (show symbol address, etc)"),
1318         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1319                     "dump raw trace in ASCII"),
1320         OPT_STRING('k', "vmlinux", &vmlinux_name, "file", "vmlinux pathname"),
1321         OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
1322         OPT_BOOLEAN('m', "modules", &modules,
1323                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
1324         OPT_BOOLEAN('n', "show-nr-samples", &show_nr_samples,
1325                     "Show a column with the number of samples"),
1326         OPT_BOOLEAN('T', "threads", &show_threads,
1327                     "Show per-thread event counters"),
1328         OPT_STRING(0, "pretty", &pretty_printing_style, "key",
1329                    "pretty printing style key: normal raw"),
1330         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1331                    "sort by key(s): pid, comm, dso, symbol, parent"),
1332         OPT_BOOLEAN('P', "full-paths", &full_paths,
1333                     "Don't shorten the pathnames taking into account the cwd"),
1334         OPT_STRING('p', "parent", &parent_pattern, "regex",
1335                    "regex filter to identify parent, see: '--sort parent'"),
1336         OPT_BOOLEAN('x', "exclude-other", &exclude_other,
1337                     "Only display entries with parent-match"),
1338         OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent",
1339                      "Display callchains using output_type and min percent threshold. "
1340                      "Default: fractal,0.5", &parse_callchain_opt, callchain_default_opt),
1341         OPT_STRING('d', "dsos", &dso_list_str, "dso[,dso...]",
1342                    "only consider symbols in these dsos"),
1343         OPT_STRING('C', "comms", &comm_list_str, "comm[,comm...]",
1344                    "only consider symbols in these comms"),
1345         OPT_STRING('S', "symbols", &sym_list_str, "symbol[,symbol...]",
1346                    "only consider these symbols"),
1347         OPT_STRING('w', "column-widths", &col_width_list_str,
1348                    "width[,width...]",
1349                    "don't try to adjust column width, use these fixed values"),
1350         OPT_STRING('t', "field-separator", &field_sep, "separator",
1351                    "separator for columns, no spaces will be added between "
1352                    "columns '.' is reserved."),
1353         OPT_END()
1354 };
1355
1356 static void setup_sorting(void)
1357 {
1358         char *tmp, *tok, *str = strdup(sort_order);
1359
1360         for (tok = strtok_r(str, ", ", &tmp);
1361                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1362                 if (sort_dimension__add(tok) < 0) {
1363                         error("Unknown --sort key: `%s'", tok);
1364                         usage_with_options(report_usage, options);
1365                 }
1366         }
1367
1368         free(str);
1369 }
1370
1371 static void setup_list(struct strlist **list, const char *list_str,
1372                        struct sort_entry *se, const char *list_name,
1373                        FILE *fp)
1374 {
1375         if (list_str) {
1376                 *list = strlist__new(true, list_str);
1377                 if (!*list) {
1378                         fprintf(stderr, "problems parsing %s list\n",
1379                                 list_name);
1380                         exit(129);
1381                 }
1382                 if (strlist__nr_entries(*list) == 1) {
1383                         fprintf(fp, "# %s: %s\n", list_name,
1384                                 strlist__entry(*list, 0)->s);
1385                         se->elide = true;
1386                 }
1387         }
1388 }
1389
1390 int cmd_report(int argc, const char **argv, const char *prefix __used)
1391 {
1392         symbol__init();
1393
1394         page_size = getpagesize();
1395
1396         argc = parse_options(argc, argv, options, report_usage, 0);
1397
1398         setup_sorting();
1399
1400         if (parent_pattern != default_parent_pattern) {
1401                 sort_dimension__add("parent");
1402                 sort_parent.elide = 1;
1403         } else
1404                 exclude_other = 0;
1405
1406         /*
1407          * Any (unrecognized) arguments left?
1408          */
1409         if (argc)
1410                 usage_with_options(report_usage, options);
1411
1412         setup_pager();
1413
1414         setup_list(&dso_list, dso_list_str, &sort_dso, "dso", stdout);
1415         setup_list(&comm_list, comm_list_str, &sort_comm, "comm", stdout);
1416         setup_list(&sym_list, sym_list_str, &sort_sym, "symbol", stdout);
1417
1418         if (field_sep && *field_sep == '.') {
1419                 fputs("'.' is the only non valid --field-separator argument\n",
1420                       stderr);
1421                 exit(129);
1422         }
1423
1424         return __cmd_report();
1425 }