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