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