SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 **resolve_callchain(struct thread *thread, struct map *map,
 411                                          struct ip_callchain *chain,
 412                                          struct symbol **parent)
 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 && !*parent && call__match(sym))
 448                                 *parent = sym;
 449                         if (!callchain)
 450                                 break;
 451                         syms[i] = sym;
 452                 }
 453         }
 454
 455         return syms;
 456 }
 457
 458 /*
 459  * collect histogram counts
 460  */
 461
 462 static int
 463 hist_entry__add(struct thread *thread, struct map *map,
 464                 struct symbol *sym, u64 ip, struct ip_callchain *chain,
 465                 char level, u64 count)
 466 {
 467         struct symbol **syms = NULL, *parent = NULL;
 468         bool hit;
 469         struct hist_entry *he;
 470
 471         if ((sort__has_parent || callchain) && chain)
 472                 syms = resolve_callchain(thread, map, chain, &parent);
 473
 474         he = __hist_entry__add(thread, map, sym, parent,
 475                                ip, count, level, &hit);
 476         if (he == NULL)
 477                 return -ENOMEM;
 478
 479         if (hit)
 480                 he->count += count;
 481
 482         if (callchain) {
 483                 if (!hit)
 484                         callchain_init(&he->callchain);
 485                 append_chain(&he->callchain, chain, syms);
 486                 free(syms);
 487         }
 488
 489         return 0;
 490 }
 491
 492 static size_t output__fprintf(FILE *fp, u64 total_samples)
 493 {
 494         struct hist_entry *pos;
 495         struct sort_entry *se;
 496         struct rb_node *nd;
 497         size_t ret = 0;
 498         unsigned int width;
 499         char *col_width = col_width_list_str;
 500         int raw_printing_style;
 501
 502         raw_printing_style = !strcmp(pretty_printing_style, "raw");
 503
 504         init_rem_hits();
 505
 506         fprintf(fp, "# Samples: %Ld\n", (u64)total_samples);
 507         fprintf(fp, "#\n");
 508
 509         fprintf(fp, "# Overhead");
 510         if (show_nr_samples) {
 511                 if (field_sep)
 512                         fprintf(fp, "%cSamples", *field_sep);
 513                 else
 514                         fputs("  Samples  ", fp);
 515         }
 516         list_for_each_entry(se, &hist_entry__sort_list, list) {
 517                 if (se->elide)
 518                         continue;
 519                 if (field_sep) {
 520                         fprintf(fp, "%c%s", *field_sep, se->header);
 521                         continue;
 522                 }
 523                 width = strlen(se->header);
 524                 if (se->width) {
 525                         if (col_width_list_str) {
 526                                 if (col_width) {
 527                                         *se->width = atoi(col_width);
 528                                         col_width = strchr(col_width, ',');
 529                                         if (col_width)
 530                                                 ++col_width;
 531                                 }
 532                         }
 533                         width = *se->width = max(*se->width, width);
 534                 }
 535                 fprintf(fp, "  %*s", width, se->header);
 536         }
 537         fprintf(fp, "\n");
 538
 539         if (field_sep)
 540                 goto print_entries;
 541
 542         fprintf(fp, "# ........");
 543         if (show_nr_samples)
 544                 fprintf(fp, " ..........");
 545         list_for_each_entry(se, &hist_entry__sort_list, list) {
 546                 unsigned int i;
 547
 548                 if (se->elide)
 549                         continue;
 550
 551                 fprintf(fp, "  ");
 552                 if (se->width)
 553                         width = *se->width;
 554                 else
 555                         width = strlen(se->header);
 556                 for (i = 0; i < width; i++)
 557                         fprintf(fp, ".");
 558         }
 559         fprintf(fp, "\n");
 560
 561         fprintf(fp, "#\n");
 562
 563 print_entries:
 564         for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
 565                 pos = rb_entry(nd, struct hist_entry, rb_node);
 566                 ret += hist_entry__fprintf(fp, pos, total_samples);
 567         }
 568
 569         if (sort_order == default_sort_order &&
 570                         parent_pattern == default_parent_pattern) {
 571                 fprintf(fp, "#\n");
 572                 fprintf(fp, "# (For a higher level overview, try: perf report --sort comm,dso)\n");
 573                 fprintf(fp, "#\n");
 574         }
 575         fprintf(fp, "\n");
 576
 577         free(rem_sq_bracket);
 578
 579         if (show_threads)
 580                 perf_read_values_display(fp, &show_threads_values,
 581                                          raw_printing_style);
 582
 583         return ret;
 584 }
 585
 586 static int validate_chain(struct ip_callchain *chain, event_t *event)
 587 {
 588         unsigned int chain_size;
 589
 590         chain_size = event->header.size;
 591         chain_size -= (unsigned long)&event->ip.__more_data - (unsigned long)event;
 592
 593         if (chain->nr*sizeof(u64) > chain_size)
 594                 return -1;
 595
 596         return 0;
 597 }
 598
 599 static int
 600 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
 601 {
 602         char level;
 603         int show = 0;
 604         struct symbol *sym = NULL;
 605         struct thread *thread;
 606         u64 ip = event->ip.ip;
 607         u64 period = 1;
 608         struct map *map = NULL;
 609         void *more_data = event->ip.__more_data;
 610         struct ip_callchain *chain = NULL;
 611         int cpumode;
 612
 613         thread = threads__findnew(event->ip.pid, &threads, &last_match);
 614
 615         if (sample_type & PERF_SAMPLE_PERIOD) {
 616                 period = *(u64 *)more_data;
 617                 more_data += sizeof(u64);
 618         }
 619
 620         dump_printf("%p [%p]: PERF_RECORD_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
 621                 (void *)(offset + head),
 622                 (void *)(long)(event->header.size),
 623                 event->header.misc,
 624                 event->ip.pid, event->ip.tid,
 625                 (void *)(long)ip,
 626                 (long long)period);
 627
 628         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
 629                 unsigned int i;
 630
 631                 chain = (void *)more_data;
 632
 633                 dump_printf("... chain: nr:%Lu\n", chain->nr);
 634
 635                 if (validate_chain(chain, event) < 0) {
 636                         eprintf("call-chain problem with event, skipping it.\n");
 637                         return 0;
 638                 }
 639
 640                 if (dump_trace) {
 641                         for (i = 0; i < chain->nr; i++)
 642                                 dump_printf("..... %2d: %016Lx\n", i, chain->ips[i]);
 643                 }
 644         }
 645
 646         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
 647
 648         if (thread == NULL) {
 649                 eprintf("problem processing %d event, skipping it.\n",
 650                         event->header.type);
 651                 return -1;
 652         }
 653
 654         if (comm_list && !strlist__has_entry(comm_list, thread->comm))
 655                 return 0;
 656
 657         cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 658
 659         if (cpumode == PERF_RECORD_MISC_KERNEL) {
 660                 show = SHOW_KERNEL;
 661                 level = 'k';
 662
 663                 sym = kernel_maps__find_symbol(ip, &map);
 664                 dump_printf(" ...... dso: %s\n",
 665                             map ? map->dso->long_name : "<not found>");
 666         } else if (cpumode == PERF_RECORD_MISC_USER) {
 667
 668                 show = SHOW_USER;
 669                 level = '.';
 670                 sym = resolve_symbol(thread, &map, &ip);
 671
 672         } else {
 673                 show = SHOW_HV;
 674                 level = 'H';
 675
 676                 dump_printf(" ...... dso: [hypervisor]\n");
 677         }
 678
 679         if (show & show_mask) {
 680                 if (dso_list &&
 681                     (!map || !map->dso ||
 682                      !(strlist__has_entry(dso_list, map->dso->short_name) ||
 683                        (map->dso->short_name != map->dso->long_name &&
 684                         strlist__has_entry(dso_list, map->dso->long_name)))))
 685                         return 0;
 686
 687                 if (sym_list && sym && !strlist__has_entry(sym_list, sym->name))
 688                         return 0;
 689
 690                 if (hist_entry__add(thread, map, sym, ip,
 691                                     chain, level, period)) {
 692                         eprintf("problem incrementing symbol count, skipping event\n");
 693                         return -1;
 694                 }
 695         }
 696         total += period;
 697
 698         return 0;
 699 }
 700
 701 static int
 702 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
 703 {
 704         struct thread *thread;
 705         struct map *map = map__new(&event->mmap, cwd, cwdlen);
 706
 707         thread = threads__findnew(event->mmap.pid, &threads, &last_match);
 708
 709         dump_printf("%p [%p]: PERF_RECORD_MMAP %d/%d: [%p(%p) @ %p]: %s\n",
 710                 (void *)(offset + head),
 711                 (void *)(long)(event->header.size),
 712                 event->mmap.pid,
 713                 event->mmap.tid,
 714                 (void *)(long)event->mmap.start,
 715                 (void *)(long)event->mmap.len,
 716                 (void *)(long)event->mmap.pgoff,
 717                 event->mmap.filename);
 718
 719         if (thread == NULL || map == NULL) {
 720                 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
 721                 return 0;
 722         }
 723
 724         thread__insert_map(thread, map);
 725         total_mmap++;
 726
 727         return 0;
 728 }
 729
 730 static int
 731 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
 732 {
 733         struct thread *thread;
 734
 735         thread = threads__findnew(event->comm.pid, &threads, &last_match);
 736
 737         dump_printf("%p [%p]: PERF_RECORD_COMM: %s:%d\n",
 738                 (void *)(offset + head),
 739                 (void *)(long)(event->header.size),
 740                 event->comm.comm, event->comm.pid);
 741
 742         if (thread == NULL ||
 743             thread__set_comm_adjust(thread, event->comm.comm)) {
 744                 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
 745                 return -1;
 746         }
 747         total_comm++;
 748
 749         return 0;
 750 }
 751
 752 static int
 753 process_task_event(event_t *event, unsigned long offset, unsigned long head)
 754 {
 755         struct thread *thread;
 756         struct thread *parent;
 757
 758         thread = threads__findnew(event->fork.pid, &threads, &last_match);
 759         parent = threads__findnew(event->fork.ppid, &threads, &last_match);
 760
 761         dump_printf("%p [%p]: PERF_RECORD_%s: (%d:%d):(%d:%d)\n",
 762                 (void *)(offset + head),
 763                 (void *)(long)(event->header.size),
 764                 event->header.type == PERF_RECORD_FORK ? "FORK" : "EXIT",
 765                 event->fork.pid, event->fork.tid,
 766                 event->fork.ppid, event->fork.ptid);
 767
 768         /*
 769          * A thread clone will have the same PID for both
 770          * parent and child.
 771          */
 772         if (thread == parent)
 773                 return 0;
 774
 775         if (event->header.type == PERF_RECORD_EXIT)
 776                 return 0;
 777
 778         if (!thread || !parent || thread__fork(thread, parent)) {
 779                 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
 780                 return -1;
 781         }
 782         total_fork++;
 783
 784         return 0;
 785 }
 786
 787 static int
 788 process_lost_event(event_t *event, unsigned long offset, unsigned long head)
 789 {
 790         dump_printf("%p [%p]: PERF_RECORD_LOST: id:%Ld: lost:%Ld\n",
 791                 (void *)(offset + head),
 792                 (void *)(long)(event->header.size),
 793                 event->lost.id,
 794                 event->lost.lost);
 795
 796         total_lost += event->lost.lost;
 797
 798         return 0;
 799 }
 800
 801 static int
 802 process_read_event(event_t *event, unsigned long offset, unsigned long head)
 803 {
 804         struct perf_event_attr *attr;
 805
 806         attr = perf_header__find_attr(event->read.id, header);
 807
 808         if (show_threads) {
 809                 const char *name = attr ? __event_name(attr->type, attr->config)
 810                                    : "unknown";
 811                 perf_read_values_add_value(&show_threads_values,
 812                                            event->read.pid, event->read.tid,
 813                                            event->read.id,
 814                                            name,
 815                                            event->read.value);
 816         }
 817
 818         dump_printf("%p [%p]: PERF_RECORD_READ: %d %d %s %Lu\n",
 819                         (void *)(offset + head),
 820                         (void *)(long)(event->header.size),
 821                         event->read.pid,
 822                         event->read.tid,
 823                         attr ? __event_name(attr->type, attr->config)
 824                              : "FAIL",
 825                         event->read.value);
 826
 827         return 0;
 828 }
 829
 830 static int
 831 process_event(event_t *event, unsigned long offset, unsigned long head)
 832 {
 833         trace_event(event);
 834
 835         switch (event->header.type) {
 836         case PERF_RECORD_SAMPLE:
 837                 return process_sample_event(event, offset, head);
 838
 839         case PERF_RECORD_MMAP:
 840                 return process_mmap_event(event, offset, head);
 841
 842         case PERF_RECORD_COMM:
 843                 return process_comm_event(event, offset, head);
 844
 845         case PERF_RECORD_FORK:
 846         case PERF_RECORD_EXIT:
 847                 return process_task_event(event, offset, head);
 848
 849         case PERF_RECORD_LOST:
 850                 return process_lost_event(event, offset, head);
 851
 852         case PERF_RECORD_READ:
 853                 return process_read_event(event, offset, head);
 854
 855         /*
 856          * We dont process them right now but they are fine:
 857          */
 858
 859         case PERF_RECORD_THROTTLE:
 860         case PERF_RECORD_UNTHROTTLE:
 861                 return 0;
 862
 863         default:
 864                 return -1;
 865         }
 866
 867         return 0;
 868 }
 869
 870 static int __cmd_report(void)
 871 {
 872         int ret, rc = EXIT_FAILURE;
 873         unsigned long offset = 0;
 874         unsigned long head, shift;
 875         struct stat input_stat;
 876         struct thread *idle;
 877         event_t *event;
 878         uint32_t size;
 879         char *buf;
 880
 881         idle = register_idle_thread(&threads, &last_match);
 882         thread__comm_adjust(idle);
 883
 884         if (show_threads)
 885                 perf_read_values_init(&show_threads_values);
 886
 887         input = open(input_name, O_RDONLY);
 888         if (input < 0) {
 889                 fprintf(stderr, " failed to open file: %s", input_name);
 890                 if (!strcmp(input_name, "perf.data"))
 891                         fprintf(stderr, "  (try 'perf record' first)");
 892                 fprintf(stderr, "\n");
 893                 exit(-1);
 894         }
 895
 896         ret = fstat(input, &input_stat);
 897         if (ret < 0) {
 898                 perror("failed to stat file");
 899                 exit(-1);
 900         }
 901
 902         if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
 903                 fprintf(stderr, "file: %s not owned by current user or root\n", input_name);
 904                 exit(-1);
 905         }
 906
 907         if (!input_stat.st_size) {
 908                 fprintf(stderr, "zero-sized file, nothing to do!\n");
 909                 exit(0);
 910         }
 911
 912         header = perf_header__read(input);
 913         head = header->data_offset;
 914
 915         sample_type = perf_header__sample_type(header);
 916
 917         if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
 918                 if (sort__has_parent) {
 919                         fprintf(stderr, "selected --sort parent, but no"
 920                                         " callchain data. Did you call"
 921                                         " perf record without -g?\n");
 922                         exit(-1);
 923                 }
 924                 if (callchain) {
 925                         fprintf(stderr, "selected -g but no callchain data."
 926                                         " Did you call perf record without"
 927                                         " -g?\n");
 928                         exit(-1);
 929                 }
 930         } else if (callchain_param.mode != CHAIN_NONE && !callchain) {
 931                         callchain = 1;
 932                         if (register_callchain_param(&callchain_param) < 0) {
 933                                 fprintf(stderr, "Can't register callchain"
 934                                                 " params\n");
 935                                 exit(-1);
 936                         }
 937         }
 938
 939         if (load_kernel() < 0) {
 940                 perror("failed to load kernel symbols");
 941                 return EXIT_FAILURE;
 942         }
 943
 944         if (!full_paths) {
 945                 if (getcwd(__cwd, sizeof(__cwd)) == NULL) {
 946                         perror("failed to get the current directory");
 947                         return EXIT_FAILURE;
 948                 }
 949                 cwdlen = strlen(cwd);
 950         } else {
 951                 cwd = NULL;
 952                 cwdlen = 0;
 953         }
 954
 955         shift = page_size * (head / page_size);
 956         offset += shift;
 957         head -= shift;
 958
 959 remap:
 960         buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
 961                            MAP_SHARED, input, offset);
 962         if (buf == MAP_FAILED) {
 963                 perror("failed to mmap file");
 964                 exit(-1);
 965         }
 966
 967 more:
 968         event = (event_t *)(buf + head);
 969
 970         size = event->header.size;
 971         if (!size)
 972                 size = 8;
 973
 974         if (head + event->header.size >= page_size * mmap_window) {
 975                 int munmap_ret;
 976
 977                 shift = page_size * (head / page_size);
 978
 979                 munmap_ret = munmap(buf, page_size * mmap_window);
 980                 assert(munmap_ret == 0);
 981
 982                 offset += shift;
 983                 head -= shift;
 984                 goto remap;
 985         }
 986
 987         size = event->header.size;
 988
 989         dump_printf("\n%p [%p]: event: %d\n",
 990                         (void *)(offset + head),
 991                         (void *)(long)event->header.size,
 992                         event->header.type);
 993
 994         if (!size || process_event(event, offset, head) < 0) {
 995
 996                 dump_printf("%p [%p]: skipping unknown header type: %d\n",
 997                         (void *)(offset + head),
 998                         (void *)(long)(event->header.size),
 999                         event->header.type);
1000
1001                 total_unknown++;
1002
1003                 /*
1004                  * assume we lost track of the stream, check alignment, and
1005                  * increment a single u64 in the hope to catch on again 'soon'.
1006                  */
1007
1008                 if (unlikely(head & 7))
1009                         head &= ~7ULL;
1010
1011                 size = 8;
1012         }
1013
1014         head += size;
1015
1016         if (offset + head >= header->data_offset + header->data_size)
1017                 goto done;
1018
1019         if (offset + head < (unsigned long)input_stat.st_size)
1020                 goto more;
1021
1022 done:
1023         rc = EXIT_SUCCESS;
1024         close(input);
1025
1026         dump_printf("      IP events: %10ld\n", total);
1027         dump_printf("    mmap events: %10ld\n", total_mmap);
1028         dump_printf("    comm events: %10ld\n", total_comm);
1029         dump_printf("    fork events: %10ld\n", total_fork);
1030         dump_printf("    lost events: %10ld\n", total_lost);
1031         dump_printf(" unknown events: %10ld\n", total_unknown);
1032
1033         if (dump_trace)
1034                 return 0;
1035
1036         if (verbose >= 3)
1037                 threads__fprintf(stdout, &threads);
1038
1039         if (verbose >= 2)
1040                 dsos__fprintf(stdout);
1041
1042         collapse__resort();
1043         output__resort(total);
1044         output__fprintf(stdout, total);
1045
1046         if (show_threads)
1047                 perf_read_values_destroy(&show_threads_values);
1048
1049         return rc;
1050 }
1051
1052 static int
1053 parse_callchain_opt(const struct option *opt __used, const char *arg,
1054                     int unset __used)
1055 {
1056         char *tok;
1057         char *endptr;
1058
1059         callchain = 1;
1060
1061         if (!arg)
1062                 return 0;
1063
1064         tok = strtok((char *)arg, ",");
1065         if (!tok)
1066                 return -1;
1067
1068         /* get the output mode */
1069         if (!strncmp(tok, "graph", strlen(arg)))
1070                 callchain_param.mode = CHAIN_GRAPH_ABS;
1071
1072         else if (!strncmp(tok, "flat", strlen(arg)))
1073                 callchain_param.mode = CHAIN_FLAT;
1074
1075         else if (!strncmp(tok, "fractal", strlen(arg)))
1076                 callchain_param.mode = CHAIN_GRAPH_REL;
1077
1078         else if (!strncmp(tok, "none", strlen(arg))) {
1079                 callchain_param.mode = CHAIN_NONE;
1080                 callchain = 0;
1081
1082                 return 0;
1083         }
1084
1085         else
1086                 return -1;
1087
1088         /* get the min percentage */
1089         tok = strtok(NULL, ",");
1090         if (!tok)
1091                 goto setup;
1092
1093         callchain_param.min_percent = strtod(tok, &endptr);
1094         if (tok == endptr)
1095                 return -1;
1096
1097 setup:
1098         if (register_callchain_param(&callchain_param) < 0) {
1099                 fprintf(stderr, "Can't register callchain params\n");
1100                 return -1;
1101         }
1102         return 0;
1103 }
1104
1105 //static const char * const report_usage[] = {
1106 const char * const report_usage[] = {
1107         "perf report [<options>] <command>",
1108         NULL
1109 };
1110
1111 static const struct option options[] = {
1112         OPT_STRING('i', "input", &input_name, "file",
1113                     "input file name"),
1114         OPT_BOOLEAN('v', "verbose", &verbose,
1115                     "be more verbose (show symbol address, etc)"),
1116         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1117                     "dump raw trace in ASCII"),
1118         OPT_STRING('k', "vmlinux", &vmlinux_name, "file", "vmlinux pathname"),
1119         OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
1120         OPT_BOOLEAN('m', "modules", &modules,
1121                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
1122         OPT_BOOLEAN('n', "show-nr-samples", &show_nr_samples,
1123                     "Show a column with the number of samples"),
1124         OPT_BOOLEAN('T', "threads", &show_threads,
1125                     "Show per-thread event counters"),
1126         OPT_STRING(0, "pretty", &pretty_printing_style, "key",
1127                    "pretty printing style key: normal raw"),
1128         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1129                    "sort by key(s): pid, comm, dso, symbol, parent"),
1130         OPT_BOOLEAN('P', "full-paths", &full_paths,
1131                     "Don't shorten the pathnames taking into account the cwd"),
1132         OPT_STRING('p', "parent", &parent_pattern, "regex",
1133                    "regex filter to identify parent, see: '--sort parent'"),
1134         OPT_BOOLEAN('x', "exclude-other", &exclude_other,
1135                     "Only display entries with parent-match"),
1136         OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent",
1137                      "Display callchains using output_type and min percent threshold. "
1138                      "Default: fractal,0.5", &parse_callchain_opt, callchain_default_opt),
1139         OPT_STRING('d', "dsos", &dso_list_str, "dso[,dso...]",
1140                    "only consider symbols in these dsos"),
1141         OPT_STRING('C', "comms", &comm_list_str, "comm[,comm...]",
1142                    "only consider symbols in these comms"),
1143         OPT_STRING('S', "symbols", &sym_list_str, "symbol[,symbol...]",
1144                    "only consider these symbols"),
1145         OPT_STRING('w', "column-widths", &col_width_list_str,
1146                    "width[,width...]",
1147                    "don't try to adjust column width, use these fixed values"),
1148         OPT_STRING('t', "field-separator", &field_sep, "separator",
1149                    "separator for columns, no spaces will be added between "
1150                    "columns '.' is reserved."),
1151         OPT_END()
1152 };
1153
1154 static void setup_sorting(void)
1155 {
1156         char *tmp, *tok, *str = strdup(sort_order);
1157
1158         for (tok = strtok_r(str, ", ", &tmp);
1159                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1160                 if (sort_dimension__add(tok) < 0) {
1161                         error("Unknown --sort key: `%s'", tok);
1162                         usage_with_options(report_usage, options);
1163                 }
1164         }
1165
1166         free(str);
1167 }
1168
1169 static void setup_list(struct strlist **list, const char *list_str,
1170                        struct sort_entry *se, const char *list_name,
1171                        FILE *fp)
1172 {
1173         if (list_str) {
1174                 *list = strlist__new(true, list_str);
1175                 if (!*list) {
1176                         fprintf(stderr, "problems parsing %s list\n",
1177                                 list_name);
1178                         exit(129);
1179                 }
1180                 if (strlist__nr_entries(*list) == 1) {
1181                         fprintf(fp, "# %s: %s\n", list_name,
1182                                 strlist__entry(*list, 0)->s);
1183                         se->elide = true;
1184                 }
1185         }
1186 }
1187
1188 int cmd_report(int argc, const char **argv, const char *prefix __used)
1189 {
1190         symbol__init();
1191
1192         page_size = getpagesize();
1193
1194         argc = parse_options(argc, argv, options, report_usage, 0);
1195
1196         setup_sorting();
1197
1198         if (parent_pattern != default_parent_pattern) {
1199                 sort_dimension__add("parent");
1200                 sort_parent.elide = 1;
1201         } else
1202                 exclude_other = 0;
1203
1204         /*
1205          * Any (unrecognized) arguments left?
1206          */
1207         if (argc)
1208                 usage_with_options(report_usage, options);
1209
1210         setup_pager();
1211
1212         setup_list(&dso_list, dso_list_str, &sort_dso, "dso", stdout);
1213         setup_list(&comm_list, comm_list_str, &sort_comm, "comm", stdout);
1214         setup_list(&sym_list, sym_list_str, &sort_sym, "symbol", stdout);
1215
1216         if (field_sep && *field_sep == '.') {
1217                 fputs("'.' is the only non valid --field-separator argument\n",
1218                       stderr);
1219                 exit(129);
1220         }
1221
1222         return __cmd_report();
1223 }