git.oblomov.eu Git - 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 "util/list.h"
  14 #include "util/cache.h"
  15 #include "util/rbtree.h"
  16 #include "util/symbol.h"
  17 #include "util/string.h"
  18
  19 #include "perf.h"
  20
  21 #include "util/parse-options.h"
  22 #include "util/parse-events.h"
  23
  24 #define SHOW_KERNEL     1
  25 #define SHOW_USER       2
  26 #define SHOW_HV         4
  27
  28 static char             const *input_name = "perf.data";
  29 static char             *vmlinux = NULL;
  30
  31 static char             default_sort_order[] = "comm,dso";
  32 static char             *sort_order = default_sort_order;
  33
  34 static int              input;
  35 static int              show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV;
  36
  37 static int              dump_trace = 0;
  38 #define dprintf(x...)   do { if (dump_trace) printf(x); } while (0)
  39 #define cdprintf(x...)  do { if (dump_trace) color_fprintf(stdout, color, x); } while (0)
  40
  41 static int              verbose;
  42 #define eprintf(x...)   do { if (verbose) fprintf(stderr, x); } while (0)
  43
  44 static int              full_paths;
  45
  46 static unsigned long    page_size;
  47 static unsigned long    mmap_window = 32;
  48
  49 static char             default_parent_pattern[] = "^sys_|^do_page_fault";
  50 static char             *parent_pattern = default_parent_pattern;
  51 static regex_t          parent_regex;
  52
  53 static int              exclude_other = 1;
  54
  55 struct ip_event {
  56         struct perf_event_header header;
  57         u64 ip;
  58         u32 pid, tid;
  59         unsigned char __more_data[];
  60 };
  61
  62 struct ip_callchain {
  63         u64 nr;
  64         u64 ips[0];
  65 };
  66
  67 struct mmap_event {
  68         struct perf_event_header header;
  69         u32 pid, tid;
  70         u64 start;
  71         u64 len;
  72         u64 pgoff;
  73         char filename[PATH_MAX];
  74 };
  75
  76 struct comm_event {
  77         struct perf_event_header header;
  78         u32 pid, tid;
  79         char comm[16];
  80 };
  81
  82 struct fork_event {
  83         struct perf_event_header header;
  84         u32 pid, ppid;
  85 };
  86
  87 struct period_event {
  88         struct perf_event_header header;
  89         u64 time;
  90         u64 id;
  91         u64 sample_period;
  92 };
  93
  94 struct lost_event {
  95         struct perf_event_header header;
  96         u64 id;
  97         u64 lost;
  98 };
  99
 100 typedef union event_union {
 101         struct perf_event_header        header;
 102         struct ip_event                 ip;
 103         struct mmap_event               mmap;
 104         struct comm_event               comm;
 105         struct fork_event               fork;
 106         struct period_event             period;
 107         struct lost_event               lost;
 108 } event_t;
 109
 110 static LIST_HEAD(dsos);
 111 static struct dso *kernel_dso;
 112 static struct dso *vdso;
 113
 114 static void dsos__add(struct dso *dso)
 115 {
 116         list_add_tail(&dso->node, &dsos);
 117 }
 118
 119 static struct dso *dsos__find(const char *name)
 120 {
 121         struct dso *pos;
 122
 123         list_for_each_entry(pos, &dsos, node)
 124                 if (strcmp(pos->name, name) == 0)
 125                         return pos;
 126         return NULL;
 127 }
 128
 129 static struct dso *dsos__findnew(const char *name)
 130 {
 131         struct dso *dso = dsos__find(name);
 132         int nr;
 133
 134         if (dso)
 135                 return dso;
 136
 137         dso = dso__new(name, 0);
 138         if (!dso)
 139                 goto out_delete_dso;
 140
 141         nr = dso__load(dso, NULL, verbose);
 142         if (nr < 0) {
 143                 eprintf("Failed to open: %s\n", name);
 144                 goto out_delete_dso;
 145         }
 146         if (!nr)
 147                 eprintf("No symbols found in: %s, maybe install a debug package?\n", name);
 148
 149         dsos__add(dso);
 150
 151         return dso;
 152
 153 out_delete_dso:
 154         dso__delete(dso);
 155         return NULL;
 156 }
 157
 158 static void dsos__fprintf(FILE *fp)
 159 {
 160         struct dso *pos;
 161
 162         list_for_each_entry(pos, &dsos, node)
 163                 dso__fprintf(pos, fp);
 164 }
 165
 166 static struct symbol *vdso__find_symbol(struct dso *dso, u64 ip)
 167 {
 168         return dso__find_symbol(kernel_dso, ip);
 169 }
 170
 171 static int load_kernel(void)
 172 {
 173         int err;
 174
 175         kernel_dso = dso__new("[kernel]", 0);
 176         if (!kernel_dso)
 177                 return -1;
 178
 179         err = dso__load_kernel(kernel_dso, vmlinux, NULL, verbose);
 180         if (err) {
 181                 dso__delete(kernel_dso);
 182                 kernel_dso = NULL;
 183         } else
 184                 dsos__add(kernel_dso);
 185
 186         vdso = dso__new("[vdso]", 0);
 187         if (!vdso)
 188                 return -1;
 189
 190         vdso->find_symbol = vdso__find_symbol;
 191
 192         dsos__add(vdso);
 193
 194         return err;
 195 }
 196
 197 static char __cwd[PATH_MAX];
 198 static char *cwd = __cwd;
 199 static int cwdlen;
 200
 201 static int strcommon(const char *pathname)
 202 {
 203         int n = 0;
 204
 205         while (pathname[n] == cwd[n] && n < cwdlen)
 206                 ++n;
 207
 208         return n;
 209 }
 210
 211 struct map {
 212         struct list_head node;
 213         u64      start;
 214         u64      end;
 215         u64      pgoff;
 216         u64      (*map_ip)(struct map *, u64);
 217         struct dso       *dso;
 218 };
 219
 220 static u64 map__map_ip(struct map *map, u64 ip)
 221 {
 222         return ip - map->start + map->pgoff;
 223 }
 224
 225 static u64 vdso__map_ip(struct map *map, u64 ip)
 226 {
 227         return ip;
 228 }
 229
 230 static inline int is_anon_memory(const char *filename)
 231 {
 232      return strcmp(filename, "//anon") == 0;
 233 }
 234
 235 static struct map *map__new(struct mmap_event *event)
 236 {
 237         struct map *self = malloc(sizeof(*self));
 238
 239         if (self != NULL) {
 240                 const char *filename = event->filename;
 241                 char newfilename[PATH_MAX];
 242                 int anon;
 243
 244                 if (cwd) {
 245                         int n = strcommon(filename);
 246
 247                         if (n == cwdlen) {
 248                                 snprintf(newfilename, sizeof(newfilename),
 249                                          ".%s", filename + n);
 250                                 filename = newfilename;
 251                         }
 252                 }
 253
 254                 anon = is_anon_memory(filename);
 255
 256                 if (anon) {
 257                         snprintf(newfilename, sizeof(newfilename), "/tmp/perf-%d.map", event->pid);
 258                         filename = newfilename;
 259                 }
 260
 261                 self->start = event->start;
 262                 self->end   = event->start + event->len;
 263                 self->pgoff = event->pgoff;
 264
 265                 self->dso = dsos__findnew(filename);
 266                 if (self->dso == NULL)
 267                         goto out_delete;
 268
 269                 if (self->dso == vdso || anon)
 270                         self->map_ip = vdso__map_ip;
 271                 else
 272                         self->map_ip = map__map_ip;
 273         }
 274         return self;
 275 out_delete:
 276         free(self);
 277         return NULL;
 278 }
 279
 280 static struct map *map__clone(struct map *self)
 281 {
 282         struct map *map = malloc(sizeof(*self));
 283
 284         if (!map)
 285                 return NULL;
 286
 287         memcpy(map, self, sizeof(*self));
 288
 289         return map;
 290 }
 291
 292 static int map__overlap(struct map *l, struct map *r)
 293 {
 294         if (l->start > r->start) {
 295                 struct map *t = l;
 296                 l = r;
 297                 r = t;
 298         }
 299
 300         if (l->end > r->start)
 301                 return 1;
 302
 303         return 0;
 304 }
 305
 306 static size_t map__fprintf(struct map *self, FILE *fp)
 307 {
 308         return fprintf(fp, " %Lx-%Lx %Lx %s\n",
 309                        self->start, self->end, self->pgoff, self->dso->name);
 310 }
 311
 312
 313 struct thread {
 314         struct rb_node   rb_node;
 315         struct list_head maps;
 316         pid_t            pid;
 317         char             *comm;
 318 };
 319
 320 static struct thread *thread__new(pid_t pid)
 321 {
 322         struct thread *self = malloc(sizeof(*self));
 323
 324         if (self != NULL) {
 325                 self->pid = pid;
 326                 self->comm = malloc(32);
 327                 if (self->comm)
 328                         snprintf(self->comm, 32, ":%d", self->pid);
 329                 INIT_LIST_HEAD(&self->maps);
 330         }
 331
 332         return self;
 333 }
 334
 335 static int thread__set_comm(struct thread *self, const char *comm)
 336 {
 337         if (self->comm)
 338                 free(self->comm);
 339         self->comm = strdup(comm);
 340         return self->comm ? 0 : -ENOMEM;
 341 }
 342
 343 static size_t thread__fprintf(struct thread *self, FILE *fp)
 344 {
 345         struct map *pos;
 346         size_t ret = fprintf(fp, "Thread %d %s\n", self->pid, self->comm);
 347
 348         list_for_each_entry(pos, &self->maps, node)
 349                 ret += map__fprintf(pos, fp);
 350
 351         return ret;
 352 }
 353
 354
 355 static struct rb_root threads;
 356 static struct thread *last_match;
 357
 358 static struct thread *threads__findnew(pid_t pid)
 359 {
 360         struct rb_node **p = &threads.rb_node;
 361         struct rb_node *parent = NULL;
 362         struct thread *th;
 363
 364         /*
 365          * Font-end cache - PID lookups come in blocks,
 366          * so most of the time we dont have to look up
 367          * the full rbtree:
 368          */
 369         if (last_match && last_match->pid == pid)
 370                 return last_match;
 371
 372         while (*p != NULL) {
 373                 parent = *p;
 374                 th = rb_entry(parent, struct thread, rb_node);
 375
 376                 if (th->pid == pid) {
 377                         last_match = th;
 378                         return th;
 379                 }
 380
 381                 if (pid < th->pid)
 382                         p = &(*p)->rb_left;
 383                 else
 384                         p = &(*p)->rb_right;
 385         }
 386
 387         th = thread__new(pid);
 388         if (th != NULL) {
 389                 rb_link_node(&th->rb_node, parent, p);
 390                 rb_insert_color(&th->rb_node, &threads);
 391                 last_match = th;
 392         }
 393
 394         return th;
 395 }
 396
 397 static void thread__insert_map(struct thread *self, struct map *map)
 398 {
 399         struct map *pos, *tmp;
 400
 401         list_for_each_entry_safe(pos, tmp, &self->maps, node) {
 402                 if (map__overlap(pos, map)) {
 403                         list_del_init(&pos->node);
 404                         /* XXX leaks dsos */
 405                         free(pos);
 406                 }
 407         }
 408
 409         list_add_tail(&map->node, &self->maps);
 410 }
 411
 412 static int thread__fork(struct thread *self, struct thread *parent)
 413 {
 414         struct map *map;
 415
 416         if (self->comm)
 417                 free(self->comm);
 418         self->comm = strdup(parent->comm);
 419         if (!self->comm)
 420                 return -ENOMEM;
 421
 422         list_for_each_entry(map, &parent->maps, node) {
 423                 struct map *new = map__clone(map);
 424                 if (!new)
 425                         return -ENOMEM;
 426                 thread__insert_map(self, new);
 427         }
 428
 429         return 0;
 430 }
 431
 432 static struct map *thread__find_map(struct thread *self, u64 ip)
 433 {
 434         struct map *pos;
 435
 436         if (self == NULL)
 437                 return NULL;
 438
 439         list_for_each_entry(pos, &self->maps, node)
 440                 if (ip >= pos->start && ip <= pos->end)
 441                         return pos;
 442
 443         return NULL;
 444 }
 445
 446 static size_t threads__fprintf(FILE *fp)
 447 {
 448         size_t ret = 0;
 449         struct rb_node *nd;
 450
 451         for (nd = rb_first(&threads); nd; nd = rb_next(nd)) {
 452                 struct thread *pos = rb_entry(nd, struct thread, rb_node);
 453
 454                 ret += thread__fprintf(pos, fp);
 455         }
 456
 457         return ret;
 458 }
 459
 460 /*
 461  * histogram, sorted on item, collects counts
 462  */
 463
 464 static struct rb_root hist;
 465
 466 struct hist_entry {
 467         struct rb_node   rb_node;
 468
 469         struct thread    *thread;
 470         struct map       *map;
 471         struct dso       *dso;
 472         struct symbol    *sym;
 473         struct symbol    *parent;
 474         u64              ip;
 475         char             level;
 476
 477         u64              count;
 478 };
 479
 480 /*
 481  * configurable sorting bits
 482  */
 483
 484 struct sort_entry {
 485         struct list_head list;
 486
 487         char *header;
 488
 489         int64_t (*cmp)(struct hist_entry *, struct hist_entry *);
 490         int64_t (*collapse)(struct hist_entry *, struct hist_entry *);
 491         size_t  (*print)(FILE *fp, struct hist_entry *);
 492 };
 493
 494 static int64_t cmp_null(void *l, void *r)
 495 {
 496         if (!l && !r)
 497                 return 0;
 498         else if (!l)
 499                 return -1;
 500         else
 501                 return 1;
 502 }
 503
 504 /* --sort pid */
 505
 506 static int64_t
 507 sort__thread_cmp(struct hist_entry *left, struct hist_entry *right)
 508 {
 509         return right->thread->pid - left->thread->pid;
 510 }
 511
 512 static size_t
 513 sort__thread_print(FILE *fp, struct hist_entry *self)
 514 {
 515         return fprintf(fp, "%16s:%5d", self->thread->comm ?: "", self->thread->pid);
 516 }
 517
 518 static struct sort_entry sort_thread = {
 519         .header = "         Command:  Pid",
 520         .cmp    = sort__thread_cmp,
 521         .print  = sort__thread_print,
 522 };
 523
 524 /* --sort comm */
 525
 526 static int64_t
 527 sort__comm_cmp(struct hist_entry *left, struct hist_entry *right)
 528 {
 529         return right->thread->pid - left->thread->pid;
 530 }
 531
 532 static int64_t
 533 sort__comm_collapse(struct hist_entry *left, struct hist_entry *right)
 534 {
 535         char *comm_l = left->thread->comm;
 536         char *comm_r = right->thread->comm;
 537
 538         if (!comm_l || !comm_r)
 539                 return cmp_null(comm_l, comm_r);
 540
 541         return strcmp(comm_l, comm_r);
 542 }
 543
 544 static size_t
 545 sort__comm_print(FILE *fp, struct hist_entry *self)
 546 {
 547         return fprintf(fp, "%16s", self->thread->comm);
 548 }
 549
 550 static struct sort_entry sort_comm = {
 551         .header         = "         Command",
 552         .cmp            = sort__comm_cmp,
 553         .collapse       = sort__comm_collapse,
 554         .print          = sort__comm_print,
 555 };
 556
 557 /* --sort dso */
 558
 559 static int64_t
 560 sort__dso_cmp(struct hist_entry *left, struct hist_entry *right)
 561 {
 562         struct dso *dso_l = left->dso;
 563         struct dso *dso_r = right->dso;
 564
 565         if (!dso_l || !dso_r)
 566                 return cmp_null(dso_l, dso_r);
 567
 568         return strcmp(dso_l->name, dso_r->name);
 569 }
 570
 571 static size_t
 572 sort__dso_print(FILE *fp, struct hist_entry *self)
 573 {
 574         if (self->dso)
 575                 return fprintf(fp, "%-25s", self->dso->name);
 576
 577         return fprintf(fp, "%016llx         ", (u64)self->ip);
 578 }
 579
 580 static struct sort_entry sort_dso = {
 581         .header = "Shared Object            ",
 582         .cmp    = sort__dso_cmp,
 583         .print  = sort__dso_print,
 584 };
 585
 586 /* --sort symbol */
 587
 588 static int64_t
 589 sort__sym_cmp(struct hist_entry *left, struct hist_entry *right)
 590 {
 591         u64 ip_l, ip_r;
 592
 593         if (left->sym == right->sym)
 594                 return 0;
 595
 596         ip_l = left->sym ? left->sym->start : left->ip;
 597         ip_r = right->sym ? right->sym->start : right->ip;
 598
 599         return (int64_t)(ip_r - ip_l);
 600 }
 601
 602 static size_t
 603 sort__sym_print(FILE *fp, struct hist_entry *self)
 604 {
 605         size_t ret = 0;
 606
 607         if (verbose)
 608                 ret += fprintf(fp, "%#018llx  ", (u64)self->ip);
 609
 610         if (self->sym) {
 611                 ret += fprintf(fp, "[%c] %s",
 612                         self->dso == kernel_dso ? 'k' : '.', self->sym->name);
 613         } else {
 614                 ret += fprintf(fp, "%#016llx", (u64)self->ip);
 615         }
 616
 617         return ret;
 618 }
 619
 620 static struct sort_entry sort_sym = {
 621         .header = "Symbol",
 622         .cmp    = sort__sym_cmp,
 623         .print  = sort__sym_print,
 624 };
 625
 626 /* --sort parent */
 627
 628 static int64_t
 629 sort__parent_cmp(struct hist_entry *left, struct hist_entry *right)
 630 {
 631         struct symbol *sym_l = left->parent;
 632         struct symbol *sym_r = right->parent;
 633
 634         if (!sym_l || !sym_r)
 635                 return cmp_null(sym_l, sym_r);
 636
 637         return strcmp(sym_l->name, sym_r->name);
 638 }
 639
 640 static size_t
 641 sort__parent_print(FILE *fp, struct hist_entry *self)
 642 {
 643         size_t ret = 0;
 644
 645         ret += fprintf(fp, "%-20s", self->parent ? self->parent->name : "[other]");
 646
 647         return ret;
 648 }
 649
 650 static struct sort_entry sort_parent = {
 651         .header = "Parent symbol       ",
 652         .cmp    = sort__parent_cmp,
 653         .print  = sort__parent_print,
 654 };
 655
 656 static int sort__need_collapse = 0;
 657 static int sort__has_parent = 0;
 658
 659 struct sort_dimension {
 660         char                    *name;
 661         struct sort_entry       *entry;
 662         int                     taken;
 663 };
 664
 665 static struct sort_dimension sort_dimensions[] = {
 666         { .name = "pid",        .entry = &sort_thread,  },
 667         { .name = "comm",       .entry = &sort_comm,    },
 668         { .name = "dso",        .entry = &sort_dso,     },
 669         { .name = "symbol",     .entry = &sort_sym,     },
 670         { .name = "parent",     .entry = &sort_parent,  },
 671 };
 672
 673 static LIST_HEAD(hist_entry__sort_list);
 674
 675 static int sort_dimension__add(char *tok)
 676 {
 677         int i;
 678
 679         for (i = 0; i < ARRAY_SIZE(sort_dimensions); i++) {
 680                 struct sort_dimension *sd = &sort_dimensions[i];
 681
 682                 if (sd->taken)
 683                         continue;
 684
 685                 if (strncasecmp(tok, sd->name, strlen(tok)))
 686                         continue;
 687
 688                 if (sd->entry->collapse)
 689                         sort__need_collapse = 1;
 690
 691                 if (sd->entry == &sort_parent) {
 692                         int ret = regcomp(&parent_regex, parent_pattern, REG_EXTENDED);
 693                         if (ret) {
 694                                 char err[BUFSIZ];
 695
 696                                 regerror(ret, &parent_regex, err, sizeof(err));
 697                                 fprintf(stderr, "Invalid regex: %s\n%s",
 698                                         parent_pattern, err);
 699                                 exit(-1);
 700                         }
 701                         sort__has_parent = 1;
 702                 }
 703
 704                 list_add_tail(&sd->entry->list, &hist_entry__sort_list);
 705                 sd->taken = 1;
 706
 707                 return 0;
 708         }
 709
 710         return -ESRCH;
 711 }
 712
 713 static int64_t
 714 hist_entry__cmp(struct hist_entry *left, struct hist_entry *right)
 715 {
 716         struct sort_entry *se;
 717         int64_t cmp = 0;
 718
 719         list_for_each_entry(se, &hist_entry__sort_list, list) {
 720                 cmp = se->cmp(left, right);
 721                 if (cmp)
 722                         break;
 723         }
 724
 725         return cmp;
 726 }
 727
 728 static int64_t
 729 hist_entry__collapse(struct hist_entry *left, struct hist_entry *right)
 730 {
 731         struct sort_entry *se;
 732         int64_t cmp = 0;
 733
 734         list_for_each_entry(se, &hist_entry__sort_list, list) {
 735                 int64_t (*f)(struct hist_entry *, struct hist_entry *);
 736
 737                 f = se->collapse ?: se->cmp;
 738
 739                 cmp = f(left, right);
 740                 if (cmp)
 741                         break;
 742         }
 743
 744         return cmp;
 745 }
 746
 747 static size_t
 748 hist_entry__fprintf(FILE *fp, struct hist_entry *self, u64 total_samples)
 749 {
 750         struct sort_entry *se;
 751         size_t ret;
 752
 753         if (exclude_other && !self->parent)
 754                 return 0;
 755
 756         if (total_samples) {
 757                 double percent = self->count * 100.0 / total_samples;
 758                 char *color = PERF_COLOR_NORMAL;
 759
 760                 /*
 761                  * We color high-overhead entries in red, mid-overhead
 762                  * entries in green - and keep the low overhead places
 763                  * normal:
 764                  */
 765                 if (percent >= 5.0) {
 766                         color = PERF_COLOR_RED;
 767                 } else {
 768                         if (percent >= 0.5)
 769                                 color = PERF_COLOR_GREEN;
 770                 }
 771
 772                 ret = color_fprintf(fp, color, "   %6.2f%%",
 773                                 (self->count * 100.0) / total_samples);
 774         } else
 775                 ret = fprintf(fp, "%12Ld ", self->count);
 776
 777         list_for_each_entry(se, &hist_entry__sort_list, list) {
 778                 if (exclude_other && (se == &sort_parent))
 779                         continue;
 780
 781                 fprintf(fp, "  ");
 782                 ret += se->print(fp, self);
 783         }
 784
 785         ret += fprintf(fp, "\n");
 786
 787         return ret;
 788 }
 789
 790 /*
 791  *
 792  */
 793
 794 static struct symbol *
 795 resolve_symbol(struct thread *thread, struct map **mapp,
 796                struct dso **dsop, u64 *ipp)
 797 {
 798         struct dso *dso = dsop ? *dsop : NULL;
 799         struct map *map = mapp ? *mapp : NULL;
 800         uint64_t ip = *ipp;
 801
 802         if (!thread)
 803                 return NULL;
 804
 805         if (dso)
 806                 goto got_dso;
 807
 808         if (map)
 809                 goto got_map;
 810
 811         map = thread__find_map(thread, ip);
 812         if (map != NULL) {
 813                 if (mapp)
 814                         *mapp = map;
 815 got_map:
 816                 ip = map->map_ip(map, ip);
 817                 *ipp  = ip;
 818
 819                 dso = map->dso;
 820         } else {
 821                 /*
 822                  * If this is outside of all known maps,
 823                  * and is a negative address, try to look it
 824                  * up in the kernel dso, as it might be a
 825                  * vsyscall (which executes in user-mode):
 826                  */
 827                 if ((long long)ip < 0)
 828                 dso = kernel_dso;
 829         }
 830         dprintf(" ...... dso: %s\n", dso ? dso->name : "<not found>");
 831
 832         if (dsop)
 833                 *dsop = dso;
 834
 835         if (!dso)
 836                 return NULL;
 837 got_dso:
 838         return dso->find_symbol(dso, ip);
 839 }
 840
 841 static int call__match(struct symbol *sym)
 842 {
 843         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
 844                 return 1;
 845
 846         return 0;
 847 }
 848
 849 /*
 850  * collect histogram counts
 851  */
 852
 853 static int
 854 hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
 855                 struct symbol *sym, u64 ip, struct ip_callchain *chain,
 856                 char level, u64 count)
 857 {
 858         struct rb_node **p = &hist.rb_node;
 859         struct rb_node *parent = NULL;
 860         struct hist_entry *he;
 861         struct hist_entry entry = {
 862                 .thread = thread,
 863                 .map    = map,
 864                 .dso    = dso,
 865                 .sym    = sym,
 866                 .ip     = ip,
 867                 .level  = level,
 868                 .count  = count,
 869                 .parent = NULL,
 870         };
 871         int cmp;
 872
 873         if (sort__has_parent && chain) {
 874                 u64 context = PERF_CONTEXT_MAX;
 875                 int i;
 876
 877                 for (i = 0; i < chain->nr; i++) {
 878                         u64 ip = chain->ips[i];
 879                         struct dso *dso = NULL;
 880                         struct symbol *sym;
 881
 882                         if (ip >= PERF_CONTEXT_MAX) {
 883                                 context = ip;
 884                                 continue;
 885                         }
 886
 887                         switch (context) {
 888                         case PERF_CONTEXT_KERNEL:
 889                                 dso = kernel_dso;
 890                                 break;
 891                         default:
 892                                 break;
 893                         }
 894
 895                         sym = resolve_symbol(thread, NULL, &dso, &ip);
 896
 897                         if (sym && call__match(sym)) {
 898                                 entry.parent = sym;
 899                                 break;
 900                         }
 901                 }
 902         }
 903
 904         while (*p != NULL) {
 905                 parent = *p;
 906                 he = rb_entry(parent, struct hist_entry, rb_node);
 907
 908                 cmp = hist_entry__cmp(&entry, he);
 909
 910                 if (!cmp) {
 911                         he->count += count;
 912                         return 0;
 913                 }
 914
 915                 if (cmp < 0)
 916                         p = &(*p)->rb_left;
 917                 else
 918                         p = &(*p)->rb_right;
 919         }
 920
 921         he = malloc(sizeof(*he));
 922         if (!he)
 923                 return -ENOMEM;
 924         *he = entry;
 925         rb_link_node(&he->rb_node, parent, p);
 926         rb_insert_color(&he->rb_node, &hist);
 927
 928         return 0;
 929 }
 930
 931 static void hist_entry__free(struct hist_entry *he)
 932 {
 933         free(he);
 934 }
 935
 936 /*
 937  * collapse the histogram
 938  */
 939
 940 static struct rb_root collapse_hists;
 941
 942 static void collapse__insert_entry(struct hist_entry *he)
 943 {
 944         struct rb_node **p = &collapse_hists.rb_node;
 945         struct rb_node *parent = NULL;
 946         struct hist_entry *iter;
 947         int64_t cmp;
 948
 949         while (*p != NULL) {
 950                 parent = *p;
 951                 iter = rb_entry(parent, struct hist_entry, rb_node);
 952
 953                 cmp = hist_entry__collapse(iter, he);
 954
 955                 if (!cmp) {
 956                         iter->count += he->count;
 957                         hist_entry__free(he);
 958                         return;
 959                 }
 960
 961                 if (cmp < 0)
 962                         p = &(*p)->rb_left;
 963                 else
 964                         p = &(*p)->rb_right;
 965         }
 966
 967         rb_link_node(&he->rb_node, parent, p);
 968         rb_insert_color(&he->rb_node, &collapse_hists);
 969 }
 970
 971 static void collapse__resort(void)
 972 {
 973         struct rb_node *next;
 974         struct hist_entry *n;
 975
 976         if (!sort__need_collapse)
 977                 return;
 978
 979         next = rb_first(&hist);
 980         while (next) {
 981                 n = rb_entry(next, struct hist_entry, rb_node);
 982                 next = rb_next(&n->rb_node);
 983
 984                 rb_erase(&n->rb_node, &hist);
 985                 collapse__insert_entry(n);
 986         }
 987 }
 988
 989 /*
 990  * reverse the map, sort on count.
 991  */
 992
 993 static struct rb_root output_hists;
 994
 995 static void output__insert_entry(struct hist_entry *he)
 996 {
 997         struct rb_node **p = &output_hists.rb_node;
 998         struct rb_node *parent = NULL;
 999         struct hist_entry *iter;
1000
1001         while (*p != NULL) {
1002                 parent = *p;
1003                 iter = rb_entry(parent, struct hist_entry, rb_node);
1004
1005                 if (he->count > iter->count)
1006                         p = &(*p)->rb_left;
1007                 else
1008                         p = &(*p)->rb_right;
1009         }
1010
1011         rb_link_node(&he->rb_node, parent, p);
1012         rb_insert_color(&he->rb_node, &output_hists);
1013 }
1014
1015 static void output__resort(void)
1016 {
1017         struct rb_node *next;
1018         struct hist_entry *n;
1019         struct rb_root *tree = &hist;
1020
1021         if (sort__need_collapse)
1022                 tree = &collapse_hists;
1023
1024         next = rb_first(tree);
1025
1026         while (next) {
1027                 n = rb_entry(next, struct hist_entry, rb_node);
1028                 next = rb_next(&n->rb_node);
1029
1030                 rb_erase(&n->rb_node, tree);
1031                 output__insert_entry(n);
1032         }
1033 }
1034
1035 static size_t output__fprintf(FILE *fp, u64 total_samples)
1036 {
1037         struct hist_entry *pos;
1038         struct sort_entry *se;
1039         struct rb_node *nd;
1040         size_t ret = 0;
1041
1042         fprintf(fp, "\n");
1043         fprintf(fp, "#\n");
1044         fprintf(fp, "# (%Ld samples)\n", (u64)total_samples);
1045         fprintf(fp, "#\n");
1046
1047         fprintf(fp, "# Overhead");
1048         list_for_each_entry(se, &hist_entry__sort_list, list) {
1049                 if (exclude_other && (se == &sort_parent))
1050                         continue;
1051                 fprintf(fp, "  %s", se->header);
1052         }
1053         fprintf(fp, "\n");
1054
1055         fprintf(fp, "# ........");
1056         list_for_each_entry(se, &hist_entry__sort_list, list) {
1057                 int i;
1058
1059                 if (exclude_other && (se == &sort_parent))
1060                         continue;
1061
1062                 fprintf(fp, "  ");
1063                 for (i = 0; i < strlen(se->header); i++)
1064                         fprintf(fp, ".");
1065         }
1066         fprintf(fp, "\n");
1067
1068         fprintf(fp, "#\n");
1069
1070         for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
1071                 pos = rb_entry(nd, struct hist_entry, rb_node);
1072                 ret += hist_entry__fprintf(fp, pos, total_samples);
1073         }
1074
1075         if (sort_order == default_sort_order &&
1076                         parent_pattern == default_parent_pattern) {
1077                 fprintf(fp, "#\n");
1078                 fprintf(fp, "# (For more details, try: perf report --sort comm,dso,symbol)\n");
1079                 fprintf(fp, "#\n");
1080         }
1081         fprintf(fp, "\n");
1082
1083         return ret;
1084 }
1085
1086 static void register_idle_thread(void)
1087 {
1088         struct thread *thread = threads__findnew(0);
1089
1090         if (thread == NULL ||
1091                         thread__set_comm(thread, "[idle]")) {
1092                 fprintf(stderr, "problem inserting idle task.\n");
1093                 exit(-1);
1094         }
1095 }
1096
1097 static unsigned long total = 0,
1098                      total_mmap = 0,
1099                      total_comm = 0,
1100                      total_fork = 0,
1101                      total_unknown = 0,
1102                      total_lost = 0;
1103
1104 static int validate_chain(struct ip_callchain *chain, event_t *event)
1105 {
1106         unsigned int chain_size;
1107
1108         chain_size = event->header.size;
1109         chain_size -= (unsigned long)&event->ip.__more_data - (unsigned long)event;
1110
1111         if (chain->nr*sizeof(u64) > chain_size)
1112                 return -1;
1113
1114         return 0;
1115 }
1116
1117 static int
1118 process_overflow_event(event_t *event, unsigned long offset, unsigned long head)
1119 {
1120         char level;
1121         int show = 0;
1122         struct dso *dso = NULL;
1123         struct thread *thread = threads__findnew(event->ip.pid);
1124         u64 ip = event->ip.ip;
1125         u64 period = 1;
1126         struct map *map = NULL;
1127         void *more_data = event->ip.__more_data;
1128         struct ip_callchain *chain = NULL;
1129
1130         if (event->header.type & PERF_SAMPLE_PERIOD) {
1131                 period = *(u64 *)more_data;
1132                 more_data += sizeof(u64);
1133         }
1134
1135         dprintf("%p [%p]: PERF_EVENT (IP, %d): %d: %p period: %Ld\n",
1136                 (void *)(offset + head),
1137                 (void *)(long)(event->header.size),
1138                 event->header.misc,
1139                 event->ip.pid,
1140                 (void *)(long)ip,
1141                 (long long)period);
1142
1143         if (event->header.type & PERF_SAMPLE_CALLCHAIN) {
1144                 int i;
1145
1146                 chain = (void *)more_data;
1147
1148                 dprintf("... chain: nr:%Lu\n", chain->nr);
1149
1150                 if (validate_chain(chain, event) < 0) {
1151                         eprintf("call-chain problem with event, skipping it.\n");
1152                         return 0;
1153                 }
1154
1155                 if (dump_trace) {
1156                         for (i = 0; i < chain->nr; i++)
1157                                 dprintf("..... %2d: %016Lx\n", i, chain->ips[i]);
1158                 }
1159         }
1160
1161         dprintf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1162
1163         if (thread == NULL) {
1164                 eprintf("problem processing %d event, skipping it.\n",
1165                         event->header.type);
1166                 return -1;
1167         }
1168
1169         if (event->header.misc & PERF_EVENT_MISC_KERNEL) {
1170                 show = SHOW_KERNEL;
1171                 level = 'k';
1172
1173                 dso = kernel_dso;
1174
1175                 dprintf(" ...... dso: %s\n", dso->name);
1176
1177         } else if (event->header.misc & PERF_EVENT_MISC_USER) {
1178
1179                 show = SHOW_USER;
1180                 level = '.';
1181
1182         } else {
1183                 show = SHOW_HV;
1184                 level = 'H';
1185                 dprintf(" ...... dso: [hypervisor]\n");
1186         }
1187
1188         if (show & show_mask) {
1189                 struct symbol *sym = resolve_symbol(thread, &map, &dso, &ip);
1190
1191                 if (hist_entry__add(thread, map, dso, sym, ip, chain, level, period)) {
1192                         eprintf("problem incrementing symbol count, skipping event\n");
1193                         return -1;
1194                 }
1195         }
1196         total += period;
1197
1198         return 0;
1199 }
1200
1201 static int
1202 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
1203 {
1204         struct thread *thread = threads__findnew(event->mmap.pid);
1205         struct map *map = map__new(&event->mmap);
1206
1207         dprintf("%p [%p]: PERF_EVENT_MMAP %d: [%p(%p) @ %p]: %s\n",
1208                 (void *)(offset + head),
1209                 (void *)(long)(event->header.size),
1210                 event->mmap.pid,
1211                 (void *)(long)event->mmap.start,
1212                 (void *)(long)event->mmap.len,
1213                 (void *)(long)event->mmap.pgoff,
1214                 event->mmap.filename);
1215
1216         if (thread == NULL || map == NULL) {
1217                 dprintf("problem processing PERF_EVENT_MMAP, skipping event.\n");
1218                 return 0;
1219         }
1220
1221         thread__insert_map(thread, map);
1222         total_mmap++;
1223
1224         return 0;
1225 }
1226
1227 static int
1228 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
1229 {
1230         struct thread *thread = threads__findnew(event->comm.pid);
1231
1232         dprintf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
1233                 (void *)(offset + head),
1234                 (void *)(long)(event->header.size),
1235                 event->comm.comm, event->comm.pid);
1236
1237         if (thread == NULL ||
1238             thread__set_comm(thread, event->comm.comm)) {
1239                 dprintf("problem processing PERF_EVENT_COMM, skipping event.\n");
1240                 return -1;
1241         }
1242         total_comm++;
1243
1244         return 0;
1245 }
1246
1247 static int
1248 process_fork_event(event_t *event, unsigned long offset, unsigned long head)
1249 {
1250         struct thread *thread = threads__findnew(event->fork.pid);
1251         struct thread *parent = threads__findnew(event->fork.ppid);
1252
1253         dprintf("%p [%p]: PERF_EVENT_FORK: %d:%d\n",
1254                 (void *)(offset + head),
1255                 (void *)(long)(event->header.size),
1256                 event->fork.pid, event->fork.ppid);
1257
1258         if (!thread || !parent || thread__fork(thread, parent)) {
1259                 dprintf("problem processing PERF_EVENT_FORK, skipping event.\n");
1260                 return -1;
1261         }
1262         total_fork++;
1263
1264         return 0;
1265 }
1266
1267 static int
1268 process_period_event(event_t *event, unsigned long offset, unsigned long head)
1269 {
1270         dprintf("%p [%p]: PERF_EVENT_PERIOD: time:%Ld, id:%Ld: period:%Ld\n",
1271                 (void *)(offset + head),
1272                 (void *)(long)(event->header.size),
1273                 event->period.time,
1274                 event->period.id,
1275                 event->period.sample_period);
1276
1277         return 0;
1278 }
1279
1280 static int
1281 process_lost_event(event_t *event, unsigned long offset, unsigned long head)
1282 {
1283         dprintf("%p [%p]: PERF_EVENT_LOST: id:%Ld: lost:%Ld\n",
1284                 (void *)(offset + head),
1285                 (void *)(long)(event->header.size),
1286                 event->lost.id,
1287                 event->lost.lost);
1288
1289         total_lost += event->lost.lost;
1290
1291         return 0;
1292 }
1293
1294 static void trace_event(event_t *event)
1295 {
1296         unsigned char *raw_event = (void *)event;
1297         char *color = PERF_COLOR_BLUE;
1298         int i, j;
1299
1300         if (!dump_trace)
1301                 return;
1302
1303         dprintf(".");
1304         cdprintf("\n. ... raw event: size %d bytes\n", event->header.size);
1305
1306         for (i = 0; i < event->header.size; i++) {
1307                 if ((i & 15) == 0) {
1308                         dprintf(".");
1309                         cdprintf("  %04x: ", i);
1310                 }
1311
1312                 cdprintf(" %02x", raw_event[i]);
1313
1314                 if (((i & 15) == 15) || i == event->header.size-1) {
1315                         cdprintf("  ");
1316                         for (j = 0; j < 15-(i & 15); j++)
1317                                 cdprintf("   ");
1318                         for (j = 0; j < (i & 15); j++) {
1319                                 if (isprint(raw_event[i-15+j]))
1320                                         cdprintf("%c", raw_event[i-15+j]);
1321                                 else
1322                                         cdprintf(".");
1323                         }
1324                         cdprintf("\n");
1325                 }
1326         }
1327         dprintf(".\n");
1328 }
1329
1330 static int
1331 process_event(event_t *event, unsigned long offset, unsigned long head)
1332 {
1333         trace_event(event);
1334
1335         if (event->header.misc & PERF_EVENT_MISC_OVERFLOW)
1336                 return process_overflow_event(event, offset, head);
1337
1338         switch (event->header.type) {
1339         case PERF_EVENT_MMAP:
1340                 return process_mmap_event(event, offset, head);
1341
1342         case PERF_EVENT_COMM:
1343                 return process_comm_event(event, offset, head);
1344
1345         case PERF_EVENT_FORK:
1346                 return process_fork_event(event, offset, head);
1347
1348         case PERF_EVENT_PERIOD:
1349                 return process_period_event(event, offset, head);
1350
1351         case PERF_EVENT_LOST:
1352                 return process_lost_event(event, offset, head);
1353
1354         /*
1355          * We dont process them right now but they are fine:
1356          */
1357
1358         case PERF_EVENT_THROTTLE:
1359         case PERF_EVENT_UNTHROTTLE:
1360                 return 0;
1361
1362         default:
1363                 return -1;
1364         }
1365
1366         return 0;
1367 }
1368
1369 static struct perf_file_header          file_header;
1370
1371 static int __cmd_report(void)
1372 {
1373         int ret, rc = EXIT_FAILURE;
1374         unsigned long offset = 0;
1375         unsigned long head = sizeof(file_header);
1376         struct stat stat;
1377         event_t *event;
1378         uint32_t size;
1379         char *buf;
1380
1381         register_idle_thread();
1382
1383         input = open(input_name, O_RDONLY);
1384         if (input < 0) {
1385                 fprintf(stderr, " failed to open file: %s", input_name);
1386                 if (!strcmp(input_name, "perf.data"))
1387                         fprintf(stderr, "  (try 'perf record' first)");
1388                 fprintf(stderr, "\n");
1389                 exit(-1);
1390         }
1391
1392         ret = fstat(input, &stat);
1393         if (ret < 0) {
1394                 perror("failed to stat file");
1395                 exit(-1);
1396         }
1397
1398         if (!stat.st_size) {
1399                 fprintf(stderr, "zero-sized file, nothing to do!\n");
1400                 exit(0);
1401         }
1402
1403         if (read(input, &file_header, sizeof(file_header)) == -1) {
1404                 perror("failed to read file headers");
1405                 exit(-1);
1406         }
1407
1408         if (sort__has_parent &&
1409             !(file_header.sample_type & PERF_SAMPLE_CALLCHAIN)) {
1410                 fprintf(stderr, "selected --sort parent, but no callchain data\n");
1411                 exit(-1);
1412         }
1413
1414         if (load_kernel() < 0) {
1415                 perror("failed to load kernel symbols");
1416                 return EXIT_FAILURE;
1417         }
1418
1419         if (!full_paths) {
1420                 if (getcwd(__cwd, sizeof(__cwd)) == NULL) {
1421                         perror("failed to get the current directory");
1422                         return EXIT_FAILURE;
1423                 }
1424                 cwdlen = strlen(cwd);
1425         } else {
1426                 cwd = NULL;
1427                 cwdlen = 0;
1428         }
1429 remap:
1430         buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1431                            MAP_SHARED, input, offset);
1432         if (buf == MAP_FAILED) {
1433                 perror("failed to mmap file");
1434                 exit(-1);
1435         }
1436
1437 more:
1438         event = (event_t *)(buf + head);
1439
1440         size = event->header.size;
1441         if (!size)
1442                 size = 8;
1443
1444         if (head + event->header.size >= page_size * mmap_window) {
1445                 unsigned long shift = page_size * (head / page_size);
1446                 int ret;
1447
1448                 ret = munmap(buf, page_size * mmap_window);
1449                 assert(ret == 0);
1450
1451                 offset += shift;
1452                 head -= shift;
1453                 goto remap;
1454         }
1455
1456         size = event->header.size;
1457
1458         dprintf("\n%p [%p]: event: %d\n",
1459                         (void *)(offset + head),
1460                         (void *)(long)event->header.size,
1461                         event->header.type);
1462
1463         if (!size || process_event(event, offset, head) < 0) {
1464
1465                 dprintf("%p [%p]: skipping unknown header type: %d\n",
1466                         (void *)(offset + head),
1467                         (void *)(long)(event->header.size),
1468                         event->header.type);
1469
1470                 total_unknown++;
1471
1472                 /*
1473                  * assume we lost track of the stream, check alignment, and
1474                  * increment a single u64 in the hope to catch on again 'soon'.
1475                  */
1476
1477                 if (unlikely(head & 7))
1478                         head &= ~7ULL;
1479
1480                 size = 8;
1481         }
1482
1483         head += size;
1484
1485         if (offset + head >= sizeof(file_header) + file_header.data_size)
1486                 goto done;
1487
1488         if (offset + head < stat.st_size)
1489                 goto more;
1490
1491 done:
1492         rc = EXIT_SUCCESS;
1493         close(input);
1494
1495         dprintf("      IP events: %10ld\n", total);
1496         dprintf("    mmap events: %10ld\n", total_mmap);
1497         dprintf("    comm events: %10ld\n", total_comm);
1498         dprintf("    fork events: %10ld\n", total_fork);
1499         dprintf("    lost events: %10ld\n", total_lost);
1500         dprintf(" unknown events: %10ld\n", total_unknown);
1501
1502         if (dump_trace)
1503                 return 0;
1504
1505         if (verbose >= 3)
1506                 threads__fprintf(stdout);
1507
1508         if (verbose >= 2)
1509                 dsos__fprintf(stdout);
1510
1511         collapse__resort();
1512         output__resort();
1513         output__fprintf(stdout, total);
1514
1515         return rc;
1516 }
1517
1518 static const char * const report_usage[] = {
1519         "perf report [<options>] <command>",
1520         NULL
1521 };
1522
1523 static const struct option options[] = {
1524         OPT_STRING('i', "input", &input_name, "file",
1525                     "input file name"),
1526         OPT_BOOLEAN('v', "verbose", &verbose,
1527                     "be more verbose (show symbol address, etc)"),
1528         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1529                     "dump raw trace in ASCII"),
1530         OPT_STRING('k', "vmlinux", &vmlinux, "file", "vmlinux pathname"),
1531         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1532                    "sort by key(s): pid, comm, dso, symbol, parent"),
1533         OPT_BOOLEAN('P', "full-paths", &full_paths,
1534                     "Don't shorten the pathnames taking into account the cwd"),
1535         OPT_STRING('p', "parent", &parent_pattern, "regex",
1536                    "regex filter to identify parent, see: '--sort parent'"),
1537         OPT_BOOLEAN('x', "exclude-other", &exclude_other,
1538                     "Only display entries with parent-match"),
1539         OPT_END()
1540 };
1541
1542 static void setup_sorting(void)
1543 {
1544         char *tmp, *tok, *str = strdup(sort_order);
1545
1546         for (tok = strtok_r(str, ", ", &tmp);
1547                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1548                 if (sort_dimension__add(tok) < 0) {
1549                         error("Unknown --sort key: `%s'", tok);
1550                         usage_with_options(report_usage, options);
1551                 }
1552         }
1553
1554         free(str);
1555 }
1556
1557 int cmd_report(int argc, const char **argv, const char *prefix)
1558 {
1559         symbol__init();
1560
1561         page_size = getpagesize();
1562
1563         argc = parse_options(argc, argv, options, report_usage, 0);
1564
1565         setup_sorting();
1566
1567         if (parent_pattern != default_parent_pattern)
1568                 sort_dimension__add("parent");
1569         else
1570                 exclude_other = 0;
1571
1572         /*
1573          * Any (unrecognized) arguments left?
1574          */
1575         if (argc)
1576                 usage_with_options(report_usage, options);
1577
1578         setup_pager();
1579
1580         return __cmd_report();
1581 }