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 mmap_event {
  63         struct perf_event_header header;
  64         __u32 pid, tid;
  65         __u64 start;
  66         __u64 len;
  67         __u64 pgoff;
  68         char filename[PATH_MAX];
  69 };
  70
  71 struct comm_event {
  72         struct perf_event_header header;
  73         __u32 pid, tid;
  74         char comm[16];
  75 };
  76
  77 struct fork_event {
  78         struct perf_event_header header;
  79         __u32 pid, ppid;
  80 };
  81
  82 struct period_event {
  83         struct perf_event_header header;
  84         __u64 time;
  85         __u64 id;
  86         __u64 sample_period;
  87 };
  88
  89 struct lost_event {
  90         struct perf_event_header header;
  91         __u64 id;
  92         __u64 lost;
  93 };
  94
  95 typedef union event_union {
  96         struct perf_event_header        header;
  97         struct ip_event                 ip;
  98         struct mmap_event               mmap;
  99         struct comm_event               comm;
 100         struct fork_event               fork;
 101         struct period_event             period;
 102         struct lost_event               lost;
 103 } event_t;
 104
 105 static LIST_HEAD(dsos);
 106 static struct dso *kernel_dso;
 107 static struct dso *vdso;
 108
 109 static void dsos__add(struct dso *dso)
 110 {
 111         list_add_tail(&dso->node, &dsos);
 112 }
 113
 114 static struct dso *dsos__find(const char *name)
 115 {
 116         struct dso *pos;
 117
 118         list_for_each_entry(pos, &dsos, node)
 119                 if (strcmp(pos->name, name) == 0)
 120                         return pos;
 121         return NULL;
 122 }
 123
 124 static struct dso *dsos__findnew(const char *name)
 125 {
 126         struct dso *dso = dsos__find(name);
 127         int nr;
 128
 129         if (dso)
 130                 return dso;
 131
 132         dso = dso__new(name, 0);
 133         if (!dso)
 134                 goto out_delete_dso;
 135
 136         nr = dso__load(dso, NULL, verbose);
 137         if (nr < 0) {
 138                 eprintf("Failed to open: %s\n", name);
 139                 goto out_delete_dso;
 140         }
 141         if (!nr)
 142                 eprintf("No symbols found in: %s, maybe install a debug package?\n", name);
 143
 144         dsos__add(dso);
 145
 146         return dso;
 147
 148 out_delete_dso:
 149         dso__delete(dso);
 150         return NULL;
 151 }
 152
 153 static void dsos__fprintf(FILE *fp)
 154 {
 155         struct dso *pos;
 156
 157         list_for_each_entry(pos, &dsos, node)
 158                 dso__fprintf(pos, fp);
 159 }
 160
 161 static struct symbol *vdso__find_symbol(struct dso *dso, __u64 ip)
 162 {
 163         return dso__find_symbol(kernel_dso, ip);
 164 }
 165
 166 static int load_kernel(void)
 167 {
 168         int err;
 169
 170         kernel_dso = dso__new("[kernel]", 0);
 171         if (!kernel_dso)
 172                 return -1;
 173
 174         err = dso__load_kernel(kernel_dso, vmlinux, NULL, verbose);
 175         if (err) {
 176                 dso__delete(kernel_dso);
 177                 kernel_dso = NULL;
 178         } else
 179                 dsos__add(kernel_dso);
 180
 181         vdso = dso__new("[vdso]", 0);
 182         if (!vdso)
 183                 return -1;
 184
 185         vdso->find_symbol = vdso__find_symbol;
 186
 187         dsos__add(vdso);
 188
 189         return err;
 190 }
 191
 192 static char __cwd[PATH_MAX];
 193 static char *cwd = __cwd;
 194 static int cwdlen;
 195
 196 static int strcommon(const char *pathname)
 197 {
 198         int n = 0;
 199
 200         while (pathname[n] == cwd[n] && n < cwdlen)
 201                 ++n;
 202
 203         return n;
 204 }
 205
 206 struct map {
 207         struct list_head node;
 208         __u64    start;
 209         __u64    end;
 210         __u64    pgoff;
 211         __u64    (*map_ip)(struct map *, __u64);
 212         struct dso       *dso;
 213 };
 214
 215 static __u64 map__map_ip(struct map *map, __u64 ip)
 216 {
 217         return ip - map->start + map->pgoff;
 218 }
 219
 220 static __u64 vdso__map_ip(struct map *map, __u64 ip)
 221 {
 222         return ip;
 223 }
 224
 225 static inline int is_anon_memory(const char *filename)
 226 {
 227      return strcmp(filename, "//anon") == 0;
 228 }
 229
 230 static struct map *map__new(struct mmap_event *event)
 231 {
 232         struct map *self = malloc(sizeof(*self));
 233
 234         if (self != NULL) {
 235                 const char *filename = event->filename;
 236                 char newfilename[PATH_MAX];
 237                 int anon;
 238
 239                 if (cwd) {
 240                         int n = strcommon(filename);
 241
 242                         if (n == cwdlen) {
 243                                 snprintf(newfilename, sizeof(newfilename),
 244                                          ".%s", filename + n);
 245                                 filename = newfilename;
 246                         }
 247                 }
 248
 249                 anon = is_anon_memory(filename);
 250
 251                 if (anon) {
 252                         snprintf(newfilename, sizeof(newfilename), "/tmp/perf-%d.map", event->pid);
 253                         filename = newfilename;
 254                 }
 255
 256                 self->start = event->start;
 257                 self->end   = event->start + event->len;
 258                 self->pgoff = event->pgoff;
 259
 260                 self->dso = dsos__findnew(filename);
 261                 if (self->dso == NULL)
 262                         goto out_delete;
 263
 264                 if (self->dso == vdso || anon)
 265                         self->map_ip = vdso__map_ip;
 266                 else
 267                         self->map_ip = map__map_ip;
 268         }
 269         return self;
 270 out_delete:
 271         free(self);
 272         return NULL;
 273 }
 274
 275 static struct map *map__clone(struct map *self)
 276 {
 277         struct map *map = malloc(sizeof(*self));
 278
 279         if (!map)
 280                 return NULL;
 281
 282         memcpy(map, self, sizeof(*self));
 283
 284         return map;
 285 }
 286
 287 static int map__overlap(struct map *l, struct map *r)
 288 {
 289         if (l->start > r->start) {
 290                 struct map *t = l;
 291                 l = r;
 292                 r = t;
 293         }
 294
 295         if (l->end > r->start)
 296                 return 1;
 297
 298         return 0;
 299 }
 300
 301 static size_t map__fprintf(struct map *self, FILE *fp)
 302 {
 303         return fprintf(fp, " %Lx-%Lx %Lx %s\n",
 304                        self->start, self->end, self->pgoff, self->dso->name);
 305 }
 306
 307
 308 struct thread {
 309         struct rb_node   rb_node;
 310         struct list_head maps;
 311         pid_t            pid;
 312         char             *comm;
 313 };
 314
 315 static struct thread *thread__new(pid_t pid)
 316 {
 317         struct thread *self = malloc(sizeof(*self));
 318
 319         if (self != NULL) {
 320                 self->pid = pid;
 321                 self->comm = malloc(32);
 322                 if (self->comm)
 323                         snprintf(self->comm, 32, ":%d", self->pid);
 324                 INIT_LIST_HEAD(&self->maps);
 325         }
 326
 327         return self;
 328 }
 329
 330 static int thread__set_comm(struct thread *self, const char *comm)
 331 {
 332         if (self->comm)
 333                 free(self->comm);
 334         self->comm = strdup(comm);
 335         return self->comm ? 0 : -ENOMEM;
 336 }
 337
 338 static size_t thread__fprintf(struct thread *self, FILE *fp)
 339 {
 340         struct map *pos;
 341         size_t ret = fprintf(fp, "Thread %d %s\n", self->pid, self->comm);
 342
 343         list_for_each_entry(pos, &self->maps, node)
 344                 ret += map__fprintf(pos, fp);
 345
 346         return ret;
 347 }
 348
 349
 350 static struct rb_root threads;
 351 static struct thread *last_match;
 352
 353 static struct thread *threads__findnew(pid_t pid)
 354 {
 355         struct rb_node **p = &threads.rb_node;
 356         struct rb_node *parent = NULL;
 357         struct thread *th;
 358
 359         /*
 360          * Font-end cache - PID lookups come in blocks,
 361          * so most of the time we dont have to look up
 362          * the full rbtree:
 363          */
 364         if (last_match && last_match->pid == pid)
 365                 return last_match;
 366
 367         while (*p != NULL) {
 368                 parent = *p;
 369                 th = rb_entry(parent, struct thread, rb_node);
 370
 371                 if (th->pid == pid) {
 372                         last_match = th;
 373                         return th;
 374                 }
 375
 376                 if (pid < th->pid)
 377                         p = &(*p)->rb_left;
 378                 else
 379                         p = &(*p)->rb_right;
 380         }
 381
 382         th = thread__new(pid);
 383         if (th != NULL) {
 384                 rb_link_node(&th->rb_node, parent, p);
 385                 rb_insert_color(&th->rb_node, &threads);
 386                 last_match = th;
 387         }
 388
 389         return th;
 390 }
 391
 392 static void thread__insert_map(struct thread *self, struct map *map)
 393 {
 394         struct map *pos, *tmp;
 395
 396         list_for_each_entry_safe(pos, tmp, &self->maps, node) {
 397                 if (map__overlap(pos, map)) {
 398                         list_del_init(&pos->node);
 399                         /* XXX leaks dsos */
 400                         free(pos);
 401                 }
 402         }
 403
 404         list_add_tail(&map->node, &self->maps);
 405 }
 406
 407 static int thread__fork(struct thread *self, struct thread *parent)
 408 {
 409         struct map *map;
 410
 411         if (self->comm)
 412                 free(self->comm);
 413         self->comm = strdup(parent->comm);
 414         if (!self->comm)
 415                 return -ENOMEM;
 416
 417         list_for_each_entry(map, &parent->maps, node) {
 418                 struct map *new = map__clone(map);
 419                 if (!new)
 420                         return -ENOMEM;
 421                 thread__insert_map(self, new);
 422         }
 423
 424         return 0;
 425 }
 426
 427 static struct map *thread__find_map(struct thread *self, __u64 ip)
 428 {
 429         struct map *pos;
 430
 431         if (self == NULL)
 432                 return NULL;
 433
 434         list_for_each_entry(pos, &self->maps, node)
 435                 if (ip >= pos->start && ip <= pos->end)
 436                         return pos;
 437
 438         return NULL;
 439 }
 440
 441 static size_t threads__fprintf(FILE *fp)
 442 {
 443         size_t ret = 0;
 444         struct rb_node *nd;
 445
 446         for (nd = rb_first(&threads); nd; nd = rb_next(nd)) {
 447                 struct thread *pos = rb_entry(nd, struct thread, rb_node);
 448
 449                 ret += thread__fprintf(pos, fp);
 450         }
 451
 452         return ret;
 453 }
 454
 455 /*
 456  * histogram, sorted on item, collects counts
 457  */
 458
 459 static struct rb_root hist;
 460
 461 struct hist_entry {
 462         struct rb_node   rb_node;
 463
 464         struct thread    *thread;
 465         struct map       *map;
 466         struct dso       *dso;
 467         struct symbol    *sym;
 468         struct symbol    *parent;
 469         __u64            ip;
 470         char             level;
 471
 472         __u64            count;
 473 };
 474
 475 /*
 476  * configurable sorting bits
 477  */
 478
 479 struct sort_entry {
 480         struct list_head list;
 481
 482         char *header;
 483
 484         int64_t (*cmp)(struct hist_entry *, struct hist_entry *);
 485         int64_t (*collapse)(struct hist_entry *, struct hist_entry *);
 486         size_t  (*print)(FILE *fp, struct hist_entry *);
 487 };
 488
 489 static int64_t cmp_null(void *l, void *r)
 490 {
 491         if (!l && !r)
 492                 return 0;
 493         else if (!l)
 494                 return -1;
 495         else
 496                 return 1;
 497 }
 498
 499 /* --sort pid */
 500
 501 static int64_t
 502 sort__thread_cmp(struct hist_entry *left, struct hist_entry *right)
 503 {
 504         return right->thread->pid - left->thread->pid;
 505 }
 506
 507 static size_t
 508 sort__thread_print(FILE *fp, struct hist_entry *self)
 509 {
 510         return fprintf(fp, "%16s:%5d", self->thread->comm ?: "", self->thread->pid);
 511 }
 512
 513 static struct sort_entry sort_thread = {
 514         .header = "         Command:  Pid",
 515         .cmp    = sort__thread_cmp,
 516         .print  = sort__thread_print,
 517 };
 518
 519 /* --sort comm */
 520
 521 static int64_t
 522 sort__comm_cmp(struct hist_entry *left, struct hist_entry *right)
 523 {
 524         return right->thread->pid - left->thread->pid;
 525 }
 526
 527 static int64_t
 528 sort__comm_collapse(struct hist_entry *left, struct hist_entry *right)
 529 {
 530         char *comm_l = left->thread->comm;
 531         char *comm_r = right->thread->comm;
 532
 533         if (!comm_l || !comm_r)
 534                 return cmp_null(comm_l, comm_r);
 535
 536         return strcmp(comm_l, comm_r);
 537 }
 538
 539 static size_t
 540 sort__comm_print(FILE *fp, struct hist_entry *self)
 541 {
 542         return fprintf(fp, "%16s", self->thread->comm);
 543 }
 544
 545 static struct sort_entry sort_comm = {
 546         .header         = "         Command",
 547         .cmp            = sort__comm_cmp,
 548         .collapse       = sort__comm_collapse,
 549         .print          = sort__comm_print,
 550 };
 551
 552 /* --sort dso */
 553
 554 static int64_t
 555 sort__dso_cmp(struct hist_entry *left, struct hist_entry *right)
 556 {
 557         struct dso *dso_l = left->dso;
 558         struct dso *dso_r = right->dso;
 559
 560         if (!dso_l || !dso_r)
 561                 return cmp_null(dso_l, dso_r);
 562
 563         return strcmp(dso_l->name, dso_r->name);
 564 }
 565
 566 static size_t
 567 sort__dso_print(FILE *fp, struct hist_entry *self)
 568 {
 569         if (self->dso)
 570                 return fprintf(fp, "%-25s", self->dso->name);
 571
 572         return fprintf(fp, "%016llx         ", (__u64)self->ip);
 573 }
 574
 575 static struct sort_entry sort_dso = {
 576         .header = "Shared Object            ",
 577         .cmp    = sort__dso_cmp,
 578         .print  = sort__dso_print,
 579 };
 580
 581 /* --sort symbol */
 582
 583 static int64_t
 584 sort__sym_cmp(struct hist_entry *left, struct hist_entry *right)
 585 {
 586         __u64 ip_l, ip_r;
 587
 588         if (left->sym == right->sym)
 589                 return 0;
 590
 591         ip_l = left->sym ? left->sym->start : left->ip;
 592         ip_r = right->sym ? right->sym->start : right->ip;
 593
 594         return (int64_t)(ip_r - ip_l);
 595 }
 596
 597 static size_t
 598 sort__sym_print(FILE *fp, struct hist_entry *self)
 599 {
 600         size_t ret = 0;
 601
 602         if (verbose)
 603                 ret += fprintf(fp, "%#018llx  ", (__u64)self->ip);
 604
 605         if (self->sym) {
 606                 ret += fprintf(fp, "[%c] %s",
 607                         self->dso == kernel_dso ? 'k' : '.', self->sym->name);
 608         } else {
 609                 ret += fprintf(fp, "%#016llx", (__u64)self->ip);
 610         }
 611
 612         return ret;
 613 }
 614
 615 static struct sort_entry sort_sym = {
 616         .header = "Symbol",
 617         .cmp    = sort__sym_cmp,
 618         .print  = sort__sym_print,
 619 };
 620
 621 /* --sort parent */
 622
 623 static int64_t
 624 sort__parent_cmp(struct hist_entry *left, struct hist_entry *right)
 625 {
 626         struct symbol *sym_l = left->parent;
 627         struct symbol *sym_r = right->parent;
 628
 629         if (!sym_l || !sym_r)
 630                 return cmp_null(sym_l, sym_r);
 631
 632         return strcmp(sym_l->name, sym_r->name);
 633 }
 634
 635 static size_t
 636 sort__parent_print(FILE *fp, struct hist_entry *self)
 637 {
 638         size_t ret = 0;
 639
 640         ret += fprintf(fp, "%-20s", self->parent ? self->parent->name : "[other]");
 641
 642         return ret;
 643 }
 644
 645 static struct sort_entry sort_parent = {
 646         .header = "Parent symbol       ",
 647         .cmp    = sort__parent_cmp,
 648         .print  = sort__parent_print,
 649 };
 650
 651 static int sort__need_collapse = 0;
 652 static int sort__has_parent = 0;
 653
 654 struct sort_dimension {
 655         char                    *name;
 656         struct sort_entry       *entry;
 657         int                     taken;
 658 };
 659
 660 static struct sort_dimension sort_dimensions[] = {
 661         { .name = "pid",        .entry = &sort_thread,  },
 662         { .name = "comm",       .entry = &sort_comm,    },
 663         { .name = "dso",        .entry = &sort_dso,     },
 664         { .name = "symbol",     .entry = &sort_sym,     },
 665         { .name = "parent",     .entry = &sort_parent,  },
 666 };
 667
 668 static LIST_HEAD(hist_entry__sort_list);
 669
 670 static int sort_dimension__add(char *tok)
 671 {
 672         int i;
 673
 674         for (i = 0; i < ARRAY_SIZE(sort_dimensions); i++) {
 675                 struct sort_dimension *sd = &sort_dimensions[i];
 676
 677                 if (sd->taken)
 678                         continue;
 679
 680                 if (strncasecmp(tok, sd->name, strlen(tok)))
 681                         continue;
 682
 683                 if (sd->entry->collapse)
 684                         sort__need_collapse = 1;
 685
 686                 if (sd->entry == &sort_parent) {
 687                         int ret = regcomp(&parent_regex, parent_pattern, REG_EXTENDED);
 688                         if (ret) {
 689                                 char err[BUFSIZ];
 690
 691                                 regerror(ret, &parent_regex, err, sizeof(err));
 692                                 fprintf(stderr, "Invalid regex: %s\n%s",
 693                                         parent_pattern, err);
 694                                 exit(-1);
 695                         }
 696                         sort__has_parent = 1;
 697                 }
 698
 699                 list_add_tail(&sd->entry->list, &hist_entry__sort_list);
 700                 sd->taken = 1;
 701
 702                 return 0;
 703         }
 704
 705         return -ESRCH;
 706 }
 707
 708 static int64_t
 709 hist_entry__cmp(struct hist_entry *left, struct hist_entry *right)
 710 {
 711         struct sort_entry *se;
 712         int64_t cmp = 0;
 713
 714         list_for_each_entry(se, &hist_entry__sort_list, list) {
 715                 cmp = se->cmp(left, right);
 716                 if (cmp)
 717                         break;
 718         }
 719
 720         return cmp;
 721 }
 722
 723 static int64_t
 724 hist_entry__collapse(struct hist_entry *left, struct hist_entry *right)
 725 {
 726         struct sort_entry *se;
 727         int64_t cmp = 0;
 728
 729         list_for_each_entry(se, &hist_entry__sort_list, list) {
 730                 int64_t (*f)(struct hist_entry *, struct hist_entry *);
 731
 732                 f = se->collapse ?: se->cmp;
 733
 734                 cmp = f(left, right);
 735                 if (cmp)
 736                         break;
 737         }
 738
 739         return cmp;
 740 }
 741
 742 static size_t
 743 hist_entry__fprintf(FILE *fp, struct hist_entry *self, __u64 total_samples)
 744 {
 745         struct sort_entry *se;
 746         size_t ret;
 747
 748         if (exclude_other && !self->parent)
 749                 return 0;
 750
 751         if (total_samples) {
 752                 double percent = self->count * 100.0 / total_samples;
 753                 char *color = PERF_COLOR_NORMAL;
 754
 755                 /*
 756                  * We color high-overhead entries in red, mid-overhead
 757                  * entries in green - and keep the low overhead places
 758                  * normal:
 759                  */
 760                 if (percent >= 5.0) {
 761                         color = PERF_COLOR_RED;
 762                 } else {
 763                         if (percent >= 0.5)
 764                                 color = PERF_COLOR_GREEN;
 765                 }
 766
 767                 ret = color_fprintf(fp, color, "   %6.2f%%",
 768                                 (self->count * 100.0) / total_samples);
 769         } else
 770                 ret = fprintf(fp, "%12Ld ", self->count);
 771
 772         list_for_each_entry(se, &hist_entry__sort_list, list) {
 773                 if (exclude_other && (se == &sort_parent))
 774                         continue;
 775
 776                 fprintf(fp, "  ");
 777                 ret += se->print(fp, self);
 778         }
 779
 780         ret += fprintf(fp, "\n");
 781
 782         return ret;
 783 }
 784
 785 /*
 786  *
 787  */
 788
 789 static struct symbol *
 790 resolve_symbol(struct thread *thread, struct map **mapp,
 791                struct dso **dsop, __u64 *ipp)
 792 {
 793         struct dso *dso = dsop ? *dsop : NULL;
 794         struct map *map = mapp ? *mapp : NULL;
 795         uint64_t ip = *ipp;
 796
 797         if (!thread)
 798                 return NULL;
 799
 800         if (dso)
 801                 goto got_dso;
 802
 803         if (map)
 804                 goto got_map;
 805
 806         map = thread__find_map(thread, ip);
 807         if (map != NULL) {
 808                 if (mapp)
 809                         *mapp = map;
 810 got_map:
 811                 ip = map->map_ip(map, ip);
 812                 *ipp  = ip;
 813
 814                 dso = map->dso;
 815         } else {
 816                 /*
 817                  * If this is outside of all known maps,
 818                  * and is a negative address, try to look it
 819                  * up in the kernel dso, as it might be a
 820                  * vsyscall (which executes in user-mode):
 821                  */
 822                 if ((long long)ip < 0)
 823                 dso = kernel_dso;
 824         }
 825         dprintf(" ...... dso: %s\n", dso ? dso->name : "<not found>");
 826
 827         if (dsop)
 828                 *dsop = dso;
 829
 830         if (!dso)
 831                 return NULL;
 832 got_dso:
 833         return dso->find_symbol(dso, ip);
 834 }
 835
 836 static struct symbol *call__match(struct symbol *sym)
 837 {
 838         if (!sym)
 839                 return NULL;
 840
 841         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
 842                 return sym;
 843
 844         return NULL;
 845 }
 846
 847 /*
 848  * collect histogram counts
 849  */
 850
 851 static int
 852 hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
 853                 struct symbol *sym, __u64 ip, struct perf_callchain_entry *chain,
 854                 char level, __u64 count)
 855 {
 856         struct rb_node **p = &hist.rb_node;
 857         struct rb_node *parent = NULL;
 858         struct hist_entry *he;
 859         struct hist_entry entry = {
 860                 .thread = thread,
 861                 .map    = map,
 862                 .dso    = dso,
 863                 .sym    = sym,
 864                 .ip     = ip,
 865                 .level  = level,
 866                 .count  = count,
 867                 .parent = NULL,
 868         };
 869         int cmp;
 870
 871         if (sort__has_parent && chain) {
 872                 int i, nr = chain->hv;
 873                 struct symbol *sym;
 874                 struct dso *dso;
 875                 __u64 ip;
 876
 877                 for (i = 0; i < chain->kernel; i++) {
 878                         ip = chain->ip[nr + i];
 879                         dso = kernel_dso;
 880                         sym = resolve_symbol(thread, NULL, &dso, &ip);
 881                         entry.parent = call__match(sym);
 882                         if (entry.parent)
 883                                 goto got_parent;
 884                 }
 885                 nr += i;
 886
 887                 for (i = 0; i < chain->user; i++) {
 888                         ip = chain->ip[nr + i];
 889                         sym = resolve_symbol(thread, NULL, NULL, &ip);
 890                         entry.parent = call__match(sym);
 891                         if (entry.parent)
 892                                 goto got_parent;
 893                 }
 894                 nr += i;
 895         }
 896 got_parent:
 897
 898         while (*p != NULL) {
 899                 parent = *p;
 900                 he = rb_entry(parent, struct hist_entry, rb_node);
 901
 902                 cmp = hist_entry__cmp(&entry, he);
 903
 904                 if (!cmp) {
 905                         he->count += count;
 906                         return 0;
 907                 }
 908
 909                 if (cmp < 0)
 910                         p = &(*p)->rb_left;
 911                 else
 912                         p = &(*p)->rb_right;
 913         }
 914
 915         he = malloc(sizeof(*he));
 916         if (!he)
 917                 return -ENOMEM;
 918         *he = entry;
 919         rb_link_node(&he->rb_node, parent, p);
 920         rb_insert_color(&he->rb_node, &hist);
 921
 922         return 0;
 923 }
 924
 925 static void hist_entry__free(struct hist_entry *he)
 926 {
 927         free(he);
 928 }
 929
 930 /*
 931  * collapse the histogram
 932  */
 933
 934 static struct rb_root collapse_hists;
 935
 936 static void collapse__insert_entry(struct hist_entry *he)
 937 {
 938         struct rb_node **p = &collapse_hists.rb_node;
 939         struct rb_node *parent = NULL;
 940         struct hist_entry *iter;
 941         int64_t cmp;
 942
 943         while (*p != NULL) {
 944                 parent = *p;
 945                 iter = rb_entry(parent, struct hist_entry, rb_node);
 946
 947                 cmp = hist_entry__collapse(iter, he);
 948
 949                 if (!cmp) {
 950                         iter->count += he->count;
 951                         hist_entry__free(he);
 952                         return;
 953                 }
 954
 955                 if (cmp < 0)
 956                         p = &(*p)->rb_left;
 957                 else
 958                         p = &(*p)->rb_right;
 959         }
 960
 961         rb_link_node(&he->rb_node, parent, p);
 962         rb_insert_color(&he->rb_node, &collapse_hists);
 963 }
 964
 965 static void collapse__resort(void)
 966 {
 967         struct rb_node *next;
 968         struct hist_entry *n;
 969
 970         if (!sort__need_collapse)
 971                 return;
 972
 973         next = rb_first(&hist);
 974         while (next) {
 975                 n = rb_entry(next, struct hist_entry, rb_node);
 976                 next = rb_next(&n->rb_node);
 977
 978                 rb_erase(&n->rb_node, &hist);
 979                 collapse__insert_entry(n);
 980         }
 981 }
 982
 983 /*
 984  * reverse the map, sort on count.
 985  */
 986
 987 static struct rb_root output_hists;
 988
 989 static void output__insert_entry(struct hist_entry *he)
 990 {
 991         struct rb_node **p = &output_hists.rb_node;
 992         struct rb_node *parent = NULL;
 993         struct hist_entry *iter;
 994
 995         while (*p != NULL) {
 996                 parent = *p;
 997                 iter = rb_entry(parent, struct hist_entry, rb_node);
 998
 999                 if (he->count > iter->count)
1000                         p = &(*p)->rb_left;
1001                 else
1002                         p = &(*p)->rb_right;
1003         }
1004
1005         rb_link_node(&he->rb_node, parent, p);
1006         rb_insert_color(&he->rb_node, &output_hists);
1007 }
1008
1009 static void output__resort(void)
1010 {
1011         struct rb_node *next;
1012         struct hist_entry *n;
1013         struct rb_root *tree = &hist;
1014
1015         if (sort__need_collapse)
1016                 tree = &collapse_hists;
1017
1018         next = rb_first(tree);
1019
1020         while (next) {
1021                 n = rb_entry(next, struct hist_entry, rb_node);
1022                 next = rb_next(&n->rb_node);
1023
1024                 rb_erase(&n->rb_node, tree);
1025                 output__insert_entry(n);
1026         }
1027 }
1028
1029 static size_t output__fprintf(FILE *fp, __u64 total_samples)
1030 {
1031         struct hist_entry *pos;
1032         struct sort_entry *se;
1033         struct rb_node *nd;
1034         size_t ret = 0;
1035
1036         fprintf(fp, "\n");
1037         fprintf(fp, "#\n");
1038         fprintf(fp, "# (%Ld samples)\n", (__u64)total_samples);
1039         fprintf(fp, "#\n");
1040
1041         fprintf(fp, "# Overhead");
1042         list_for_each_entry(se, &hist_entry__sort_list, list) {
1043                 if (exclude_other && (se == &sort_parent))
1044                         continue;
1045                 fprintf(fp, "  %s", se->header);
1046         }
1047         fprintf(fp, "\n");
1048
1049         fprintf(fp, "# ........");
1050         list_for_each_entry(se, &hist_entry__sort_list, list) {
1051                 int i;
1052
1053                 if (exclude_other && (se == &sort_parent))
1054                         continue;
1055
1056                 fprintf(fp, "  ");
1057                 for (i = 0; i < strlen(se->header); i++)
1058                         fprintf(fp, ".");
1059         }
1060         fprintf(fp, "\n");
1061
1062         fprintf(fp, "#\n");
1063
1064         for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
1065                 pos = rb_entry(nd, struct hist_entry, rb_node);
1066                 ret += hist_entry__fprintf(fp, pos, total_samples);
1067         }
1068
1069         if (sort_order == default_sort_order &&
1070                         parent_pattern == default_parent_pattern) {
1071                 fprintf(fp, "#\n");
1072                 fprintf(fp, "# (For more details, try: perf report --sort comm,dso,symbol)\n");
1073                 fprintf(fp, "#\n");
1074         }
1075         fprintf(fp, "\n");
1076
1077         return ret;
1078 }
1079
1080 static void register_idle_thread(void)
1081 {
1082         struct thread *thread = threads__findnew(0);
1083
1084         if (thread == NULL ||
1085                         thread__set_comm(thread, "[idle]")) {
1086                 fprintf(stderr, "problem inserting idle task.\n");
1087                 exit(-1);
1088         }
1089 }
1090
1091 static unsigned long total = 0,
1092                      total_mmap = 0,
1093                      total_comm = 0,
1094                      total_fork = 0,
1095                      total_unknown = 0,
1096                      total_lost = 0;
1097
1098 static int validate_chain(struct perf_callchain_entry *chain, event_t *event)
1099 {
1100         unsigned int chain_size;
1101
1102         if (chain->nr > MAX_STACK_DEPTH)
1103                 return -1;
1104         if (chain->hv > MAX_STACK_DEPTH)
1105                 return -1;
1106         if (chain->kernel > MAX_STACK_DEPTH)
1107                 return -1;
1108         if (chain->user > MAX_STACK_DEPTH)
1109                 return -1;
1110         if (chain->hv + chain->kernel + chain->user != chain->nr)
1111                 return -1;
1112
1113         chain_size = event->header.size;
1114         chain_size -= (unsigned long)&event->ip.__more_data - (unsigned long)event;
1115
1116         if (chain->nr*sizeof(__u64) > chain_size)
1117                 return -1;
1118
1119         return 0;
1120 }
1121
1122 static int
1123 process_overflow_event(event_t *event, unsigned long offset, unsigned long head)
1124 {
1125         char level;
1126         int show = 0;
1127         struct dso *dso = NULL;
1128         struct thread *thread = threads__findnew(event->ip.pid);
1129         __u64 ip = event->ip.ip;
1130         __u64 period = 1;
1131         struct map *map = NULL;
1132         void *more_data = event->ip.__more_data;
1133         struct perf_callchain_entry *chain = NULL;
1134
1135         if (event->header.type & PERF_SAMPLE_PERIOD) {
1136                 period = *(__u64 *)more_data;
1137                 more_data += sizeof(__u64);
1138         }
1139
1140         dprintf("%p [%p]: PERF_EVENT (IP, %d): %d: %p period: %Ld\n",
1141                 (void *)(offset + head),
1142                 (void *)(long)(event->header.size),
1143                 event->header.misc,
1144                 event->ip.pid,
1145                 (void *)(long)ip,
1146                 (long long)period);
1147
1148         if (event->header.type & PERF_SAMPLE_CALLCHAIN) {
1149                 int i;
1150
1151                 chain = (void *)more_data;
1152
1153                 dprintf("... chain: u:%d, k:%d, nr:%d\n",
1154                         chain->user,
1155                         chain->kernel,
1156                         chain->nr);
1157
1158                 if (validate_chain(chain, event) < 0) {
1159                         eprintf("call-chain problem with event, skipping it.\n");
1160                         return 0;
1161                 }
1162
1163                 if (dump_trace) {
1164                         for (i = 0; i < chain->nr; i++)
1165                                 dprintf("..... %2d: %016Lx\n", i, chain->ip[i]);
1166                 }
1167         }
1168
1169         dprintf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1170
1171         if (thread == NULL) {
1172                 eprintf("problem processing %d event, skipping it.\n",
1173                         event->header.type);
1174                 return -1;
1175         }
1176
1177         if (event->header.misc & PERF_EVENT_MISC_KERNEL) {
1178                 show = SHOW_KERNEL;
1179                 level = 'k';
1180
1181                 dso = kernel_dso;
1182
1183                 dprintf(" ...... dso: %s\n", dso->name);
1184
1185         } else if (event->header.misc & PERF_EVENT_MISC_USER) {
1186
1187                 show = SHOW_USER;
1188                 level = '.';
1189
1190         } else {
1191                 show = SHOW_HV;
1192                 level = 'H';
1193                 dprintf(" ...... dso: [hypervisor]\n");
1194         }
1195
1196         if (show & show_mask) {
1197                 struct symbol *sym = resolve_symbol(thread, &map, &dso, &ip);
1198
1199                 if (hist_entry__add(thread, map, dso, sym, ip, chain, level, period)) {
1200                         eprintf("problem incrementing symbol count, skipping event\n");
1201                         return -1;
1202                 }
1203         }
1204         total += period;
1205
1206         return 0;
1207 }
1208
1209 static int
1210 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
1211 {
1212         struct thread *thread = threads__findnew(event->mmap.pid);
1213         struct map *map = map__new(&event->mmap);
1214
1215         dprintf("%p [%p]: PERF_EVENT_MMAP %d: [%p(%p) @ %p]: %s\n",
1216                 (void *)(offset + head),
1217                 (void *)(long)(event->header.size),
1218                 event->mmap.pid,
1219                 (void *)(long)event->mmap.start,
1220                 (void *)(long)event->mmap.len,
1221                 (void *)(long)event->mmap.pgoff,
1222                 event->mmap.filename);
1223
1224         if (thread == NULL || map == NULL) {
1225                 dprintf("problem processing PERF_EVENT_MMAP, skipping event.\n");
1226                 return 0;
1227         }
1228
1229         thread__insert_map(thread, map);
1230         total_mmap++;
1231
1232         return 0;
1233 }
1234
1235 static int
1236 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
1237 {
1238         struct thread *thread = threads__findnew(event->comm.pid);
1239
1240         dprintf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
1241                 (void *)(offset + head),
1242                 (void *)(long)(event->header.size),
1243                 event->comm.comm, event->comm.pid);
1244
1245         if (thread == NULL ||
1246             thread__set_comm(thread, event->comm.comm)) {
1247                 dprintf("problem processing PERF_EVENT_COMM, skipping event.\n");
1248                 return -1;
1249         }
1250         total_comm++;
1251
1252         return 0;
1253 }
1254
1255 static int
1256 process_fork_event(event_t *event, unsigned long offset, unsigned long head)
1257 {
1258         struct thread *thread = threads__findnew(event->fork.pid);
1259         struct thread *parent = threads__findnew(event->fork.ppid);
1260
1261         dprintf("%p [%p]: PERF_EVENT_FORK: %d:%d\n",
1262                 (void *)(offset + head),
1263                 (void *)(long)(event->header.size),
1264                 event->fork.pid, event->fork.ppid);
1265
1266         if (!thread || !parent || thread__fork(thread, parent)) {
1267                 dprintf("problem processing PERF_EVENT_FORK, skipping event.\n");
1268                 return -1;
1269         }
1270         total_fork++;
1271
1272         return 0;
1273 }
1274
1275 static int
1276 process_period_event(event_t *event, unsigned long offset, unsigned long head)
1277 {
1278         dprintf("%p [%p]: PERF_EVENT_PERIOD: time:%Ld, id:%Ld: period:%Ld\n",
1279                 (void *)(offset + head),
1280                 (void *)(long)(event->header.size),
1281                 event->period.time,
1282                 event->period.id,
1283                 event->period.sample_period);
1284
1285         return 0;
1286 }
1287
1288 static int
1289 process_lost_event(event_t *event, unsigned long offset, unsigned long head)
1290 {
1291         dprintf("%p [%p]: PERF_EVENT_LOST: id:%Ld: lost:%Ld\n",
1292                 (void *)(offset + head),
1293                 (void *)(long)(event->header.size),
1294                 event->lost.id,
1295                 event->lost.lost);
1296
1297         total_lost += event->lost.lost;
1298
1299         return 0;
1300 }
1301
1302 static void trace_event(event_t *event)
1303 {
1304         unsigned char *raw_event = (void *)event;
1305         char *color = PERF_COLOR_BLUE;
1306         int i, j;
1307
1308         if (!dump_trace)
1309                 return;
1310
1311         dprintf(".");
1312         cdprintf("\n. ... raw event: size %d bytes\n", event->header.size);
1313
1314         for (i = 0; i < event->header.size; i++) {
1315                 if ((i & 15) == 0) {
1316                         dprintf(".");
1317                         cdprintf("  %04x: ", i);
1318                 }
1319
1320                 cdprintf(" %02x", raw_event[i]);
1321
1322                 if (((i & 15) == 15) || i == event->header.size-1) {
1323                         cdprintf("  ");
1324                         for (j = 0; j < 15-(i & 15); j++)
1325                                 cdprintf("   ");
1326                         for (j = 0; j < (i & 15); j++) {
1327                                 if (isprint(raw_event[i-15+j]))
1328                                         cdprintf("%c", raw_event[i-15+j]);
1329                                 else
1330                                         cdprintf(".");
1331                         }
1332                         cdprintf("\n");
1333                 }
1334         }
1335         dprintf(".\n");
1336 }
1337
1338 static int
1339 process_event(event_t *event, unsigned long offset, unsigned long head)
1340 {
1341         trace_event(event);
1342
1343         if (event->header.misc & PERF_EVENT_MISC_OVERFLOW)
1344                 return process_overflow_event(event, offset, head);
1345
1346         switch (event->header.type) {
1347         case PERF_EVENT_MMAP:
1348                 return process_mmap_event(event, offset, head);
1349
1350         case PERF_EVENT_COMM:
1351                 return process_comm_event(event, offset, head);
1352
1353         case PERF_EVENT_FORK:
1354                 return process_fork_event(event, offset, head);
1355
1356         case PERF_EVENT_PERIOD:
1357                 return process_period_event(event, offset, head);
1358
1359         case PERF_EVENT_LOST:
1360                 return process_lost_event(event, offset, head);
1361
1362         /*
1363          * We dont process them right now but they are fine:
1364          */
1365
1366         case PERF_EVENT_THROTTLE:
1367         case PERF_EVENT_UNTHROTTLE:
1368                 return 0;
1369
1370         default:
1371                 return -1;
1372         }
1373
1374         return 0;
1375 }
1376
1377 static int __cmd_report(void)
1378 {
1379         int ret, rc = EXIT_FAILURE;
1380         unsigned long offset = 0;
1381         unsigned long head = 0;
1382         struct stat stat;
1383         event_t *event;
1384         uint32_t size;
1385         char *buf;
1386
1387         register_idle_thread();
1388
1389         input = open(input_name, O_RDONLY);
1390         if (input < 0) {
1391                 fprintf(stderr, " failed to open file: %s", input_name);
1392                 if (!strcmp(input_name, "perf.data"))
1393                         fprintf(stderr, "  (try 'perf record' first)");
1394                 fprintf(stderr, "\n");
1395                 exit(-1);
1396         }
1397
1398         ret = fstat(input, &stat);
1399         if (ret < 0) {
1400                 perror("failed to stat file");
1401                 exit(-1);
1402         }
1403
1404         if (!stat.st_size) {
1405                 fprintf(stderr, "zero-sized file, nothing to do!\n");
1406                 exit(0);
1407         }
1408
1409         if (load_kernel() < 0) {
1410                 perror("failed to load kernel symbols");
1411                 return EXIT_FAILURE;
1412         }
1413
1414         if (!full_paths) {
1415                 if (getcwd(__cwd, sizeof(__cwd)) == NULL) {
1416                         perror("failed to get the current directory");
1417                         return EXIT_FAILURE;
1418                 }
1419                 cwdlen = strlen(cwd);
1420         } else {
1421                 cwd = NULL;
1422                 cwdlen = 0;
1423         }
1424 remap:
1425         buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1426                            MAP_SHARED, input, offset);
1427         if (buf == MAP_FAILED) {
1428                 perror("failed to mmap file");
1429                 exit(-1);
1430         }
1431
1432 more:
1433         event = (event_t *)(buf + head);
1434
1435         size = event->header.size;
1436         if (!size)
1437                 size = 8;
1438
1439         if (head + event->header.size >= page_size * mmap_window) {
1440                 unsigned long shift = page_size * (head / page_size);
1441                 int ret;
1442
1443                 ret = munmap(buf, page_size * mmap_window);
1444                 assert(ret == 0);
1445
1446                 offset += shift;
1447                 head -= shift;
1448                 goto remap;
1449         }
1450
1451         size = event->header.size;
1452
1453         dprintf("\n%p [%p]: event: %d\n",
1454                         (void *)(offset + head),
1455                         (void *)(long)event->header.size,
1456                         event->header.type);
1457
1458         if (!size || process_event(event, offset, head) < 0) {
1459
1460                 dprintf("%p [%p]: skipping unknown header type: %d\n",
1461                         (void *)(offset + head),
1462                         (void *)(long)(event->header.size),
1463                         event->header.type);
1464
1465                 total_unknown++;
1466
1467                 /*
1468                  * assume we lost track of the stream, check alignment, and
1469                  * increment a single u64 in the hope to catch on again 'soon'.
1470                  */
1471
1472                 if (unlikely(head & 7))
1473                         head &= ~7ULL;
1474
1475                 size = 8;
1476         }
1477
1478         head += size;
1479
1480         if (offset + head < stat.st_size)
1481                 goto more;
1482
1483         rc = EXIT_SUCCESS;
1484         close(input);
1485
1486         dprintf("      IP events: %10ld\n", total);
1487         dprintf("    mmap events: %10ld\n", total_mmap);
1488         dprintf("    comm events: %10ld\n", total_comm);
1489         dprintf("    fork events: %10ld\n", total_fork);
1490         dprintf("    lost events: %10ld\n", total_lost);
1491         dprintf(" unknown events: %10ld\n", total_unknown);
1492
1493         if (dump_trace)
1494                 return 0;
1495
1496         if (verbose >= 3)
1497                 threads__fprintf(stdout);
1498
1499         if (verbose >= 2)
1500                 dsos__fprintf(stdout);
1501
1502         collapse__resort();
1503         output__resort();
1504         output__fprintf(stdout, total);
1505
1506         return rc;
1507 }
1508
1509 static const char * const report_usage[] = {
1510         "perf report [<options>] <command>",
1511         NULL
1512 };
1513
1514 static const struct option options[] = {
1515         OPT_STRING('i', "input", &input_name, "file",
1516                     "input file name"),
1517         OPT_BOOLEAN('v', "verbose", &verbose,
1518                     "be more verbose (show symbol address, etc)"),
1519         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1520                     "dump raw trace in ASCII"),
1521         OPT_STRING('k', "vmlinux", &vmlinux, "file", "vmlinux pathname"),
1522         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1523                    "sort by key(s): pid, comm, dso, symbol, parent"),
1524         OPT_BOOLEAN('P', "full-paths", &full_paths,
1525                     "Don't shorten the pathnames taking into account the cwd"),
1526         OPT_STRING('p', "parent", &parent_pattern, "regex",
1527                    "regex filter to identify parent, see: '--sort parent'"),
1528         OPT_BOOLEAN('x', "exclude-other", &exclude_other,
1529                     "Only display entries with parent-match"),
1530         OPT_END()
1531 };
1532
1533 static void setup_sorting(void)
1534 {
1535         char *tmp, *tok, *str = strdup(sort_order);
1536
1537         for (tok = strtok_r(str, ", ", &tmp);
1538                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1539                 if (sort_dimension__add(tok) < 0) {
1540                         error("Unknown --sort key: `%s'", tok);
1541                         usage_with_options(report_usage, options);
1542                 }
1543         }
1544
1545         free(str);
1546 }
1547
1548 int cmd_report(int argc, const char **argv, const char *prefix)
1549 {
1550         symbol__init();
1551
1552         page_size = getpagesize();
1553
1554         argc = parse_options(argc, argv, options, report_usage, 0);
1555
1556         setup_sorting();
1557
1558         if (parent_pattern != default_parent_pattern)
1559                 sort_dimension__add("parent");
1560         else
1561                 exclude_other = 0;
1562
1563         /*
1564          * Any (unrecognized) arguments left?
1565          */
1566         if (argc)
1567                 usage_with_options(report_usage, options);
1568
1569         setup_pager();
1570
1571         return __cmd_report();
1572 }