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 #include "util/header.h"
  21
  22 #include "util/parse-options.h"
  23 #include "util/parse-events.h"
  24
  25 #define SHOW_KERNEL     1
  26 #define SHOW_USER       2
  27 #define SHOW_HV         4
  28
  29 static char             const *input_name = "perf.data";
  30 static char             *vmlinux = NULL;
  31
  32 static char             default_sort_order[] = "comm,dso";
  33 static char             *sort_order = default_sort_order;
  34
  35 static int              input;
  36 static int              show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV;
  37
  38 static int              dump_trace = 0;
  39 #define dprintf(x...)   do { if (dump_trace) printf(x); } while (0)
  40 #define cdprintf(x...)  do { if (dump_trace) color_fprintf(stdout, color, x); } while (0)
  41
  42 static int              verbose;
  43 #define eprintf(x...)   do { if (verbose) fprintf(stderr, x); } while (0)
  44
  45 static int              full_paths;
  46
  47 static unsigned long    page_size;
  48 static unsigned long    mmap_window = 32;
  49
  50 static char             default_parent_pattern[] = "^sys_|^do_page_fault";
  51 static char             *parent_pattern = default_parent_pattern;
  52 static regex_t          parent_regex;
  53
  54 static int              exclude_other = 1;
  55
  56 static u64              sample_type;
  57
  58 struct ip_event {
  59         struct perf_event_header header;
  60         u64 ip;
  61         u32 pid, tid;
  62         unsigned char __more_data[];
  63 };
  64
  65 struct ip_callchain {
  66         u64 nr;
  67         u64 ips[0];
  68 };
  69
  70 struct mmap_event {
  71         struct perf_event_header header;
  72         u32 pid, tid;
  73         u64 start;
  74         u64 len;
  75         u64 pgoff;
  76         char filename[PATH_MAX];
  77 };
  78
  79 struct comm_event {
  80         struct perf_event_header header;
  81         u32 pid, tid;
  82         char comm[16];
  83 };
  84
  85 struct fork_event {
  86         struct perf_event_header header;
  87         u32 pid, ppid;
  88 };
  89
  90 struct period_event {
  91         struct perf_event_header header;
  92         u64 time;
  93         u64 id;
  94         u64 sample_period;
  95 };
  96
  97 struct lost_event {
  98         struct perf_event_header header;
  99         u64 id;
 100         u64 lost;
 101 };
 102
 103 typedef union event_union {
 104         struct perf_event_header        header;
 105         struct ip_event                 ip;
 106         struct mmap_event               mmap;
 107         struct comm_event               comm;
 108         struct fork_event               fork;
 109         struct period_event             period;
 110         struct lost_event               lost;
 111 } event_t;
 112
 113 static LIST_HEAD(dsos);
 114 static struct dso *kernel_dso;
 115 static struct dso *vdso;
 116
 117 static void dsos__add(struct dso *dso)
 118 {
 119         list_add_tail(&dso->node, &dsos);
 120 }
 121
 122 static struct dso *dsos__find(const char *name)
 123 {
 124         struct dso *pos;
 125
 126         list_for_each_entry(pos, &dsos, node)
 127                 if (strcmp(pos->name, name) == 0)
 128                         return pos;
 129         return NULL;
 130 }
 131
 132 static struct dso *dsos__findnew(const char *name)
 133 {
 134         struct dso *dso = dsos__find(name);
 135         int nr;
 136
 137         if (dso)
 138                 return dso;
 139
 140         dso = dso__new(name, 0);
 141         if (!dso)
 142                 goto out_delete_dso;
 143
 144         nr = dso__load(dso, NULL, verbose);
 145         if (nr < 0) {
 146                 eprintf("Failed to open: %s\n", name);
 147                 goto out_delete_dso;
 148         }
 149         if (!nr)
 150                 eprintf("No symbols found in: %s, maybe install a debug package?\n", name);
 151
 152         dsos__add(dso);
 153
 154         return dso;
 155
 156 out_delete_dso:
 157         dso__delete(dso);
 158         return NULL;
 159 }
 160
 161 static void dsos__fprintf(FILE *fp)
 162 {
 163         struct dso *pos;
 164
 165         list_for_each_entry(pos, &dsos, node)
 166                 dso__fprintf(pos, fp);
 167 }
 168
 169 static struct symbol *vdso__find_symbol(struct dso *dso, u64 ip)
 170 {
 171         return dso__find_symbol(kernel_dso, ip);
 172 }
 173
 174 static int load_kernel(void)
 175 {
 176         int err;
 177
 178         kernel_dso = dso__new("[kernel]", 0);
 179         if (!kernel_dso)
 180                 return -1;
 181
 182         err = dso__load_kernel(kernel_dso, vmlinux, NULL, verbose);
 183         if (err) {
 184                 dso__delete(kernel_dso);
 185                 kernel_dso = NULL;
 186         } else
 187                 dsos__add(kernel_dso);
 188
 189         vdso = dso__new("[vdso]", 0);
 190         if (!vdso)
 191                 return -1;
 192
 193         vdso->find_symbol = vdso__find_symbol;
 194
 195         dsos__add(vdso);
 196
 197         return err;
 198 }
 199
 200 static char __cwd[PATH_MAX];
 201 static char *cwd = __cwd;
 202 static int cwdlen;
 203
 204 static int strcommon(const char *pathname)
 205 {
 206         int n = 0;
 207
 208         while (pathname[n] == cwd[n] && n < cwdlen)
 209                 ++n;
 210
 211         return n;
 212 }
 213
 214 struct map {
 215         struct list_head node;
 216         u64      start;
 217         u64      end;
 218         u64      pgoff;
 219         u64      (*map_ip)(struct map *, u64);
 220         struct dso       *dso;
 221 };
 222
 223 static u64 map__map_ip(struct map *map, u64 ip)
 224 {
 225         return ip - map->start + map->pgoff;
 226 }
 227
 228 static u64 vdso__map_ip(struct map *map, u64 ip)
 229 {
 230         return ip;
 231 }
 232
 233 static inline int is_anon_memory(const char *filename)
 234 {
 235      return strcmp(filename, "//anon") == 0;
 236 }
 237
 238 static struct map *map__new(struct mmap_event *event)
 239 {
 240         struct map *self = malloc(sizeof(*self));
 241
 242         if (self != NULL) {
 243                 const char *filename = event->filename;
 244                 char newfilename[PATH_MAX];
 245                 int anon;
 246
 247                 if (cwd) {
 248                         int n = strcommon(filename);
 249
 250                         if (n == cwdlen) {
 251                                 snprintf(newfilename, sizeof(newfilename),
 252                                          ".%s", filename + n);
 253                                 filename = newfilename;
 254                         }
 255                 }
 256
 257                 anon = is_anon_memory(filename);
 258
 259                 if (anon) {
 260                         snprintf(newfilename, sizeof(newfilename), "/tmp/perf-%d.map", event->pid);
 261                         filename = newfilename;
 262                 }
 263
 264                 self->start = event->start;
 265                 self->end   = event->start + event->len;
 266                 self->pgoff = event->pgoff;
 267
 268                 self->dso = dsos__findnew(filename);
 269                 if (self->dso == NULL)
 270                         goto out_delete;
 271
 272                 if (self->dso == vdso || anon)
 273                         self->map_ip = vdso__map_ip;
 274                 else
 275                         self->map_ip = map__map_ip;
 276         }
 277         return self;
 278 out_delete:
 279         free(self);
 280         return NULL;
 281 }
 282
 283 static struct map *map__clone(struct map *self)
 284 {
 285         struct map *map = malloc(sizeof(*self));
 286
 287         if (!map)
 288                 return NULL;
 289
 290         memcpy(map, self, sizeof(*self));
 291
 292         return map;
 293 }
 294
 295 static int map__overlap(struct map *l, struct map *r)
 296 {
 297         if (l->start > r->start) {
 298                 struct map *t = l;
 299                 l = r;
 300                 r = t;
 301         }
 302
 303         if (l->end > r->start)
 304                 return 1;
 305
 306         return 0;
 307 }
 308
 309 static size_t map__fprintf(struct map *self, FILE *fp)
 310 {
 311         return fprintf(fp, " %Lx-%Lx %Lx %s\n",
 312                        self->start, self->end, self->pgoff, self->dso->name);
 313 }
 314
 315
 316 struct thread {
 317         struct rb_node   rb_node;
 318         struct list_head maps;
 319         pid_t            pid;
 320         char             *comm;
 321 };
 322
 323 static struct thread *thread__new(pid_t pid)
 324 {
 325         struct thread *self = malloc(sizeof(*self));
 326
 327         if (self != NULL) {
 328                 self->pid = pid;
 329                 self->comm = malloc(32);
 330                 if (self->comm)
 331                         snprintf(self->comm, 32, ":%d", self->pid);
 332                 INIT_LIST_HEAD(&self->maps);
 333         }
 334
 335         return self;
 336 }
 337
 338 static int thread__set_comm(struct thread *self, const char *comm)
 339 {
 340         if (self->comm)
 341                 free(self->comm);
 342         self->comm = strdup(comm);
 343         return self->comm ? 0 : -ENOMEM;
 344 }
 345
 346 static size_t thread__fprintf(struct thread *self, FILE *fp)
 347 {
 348         struct map *pos;
 349         size_t ret = fprintf(fp, "Thread %d %s\n", self->pid, self->comm);
 350
 351         list_for_each_entry(pos, &self->maps, node)
 352                 ret += map__fprintf(pos, fp);
 353
 354         return ret;
 355 }
 356
 357
 358 static struct rb_root threads;
 359 static struct thread *last_match;
 360
 361 static struct thread *threads__findnew(pid_t pid)
 362 {
 363         struct rb_node **p = &threads.rb_node;
 364         struct rb_node *parent = NULL;
 365         struct thread *th;
 366
 367         /*
 368          * Font-end cache - PID lookups come in blocks,
 369          * so most of the time we dont have to look up
 370          * the full rbtree:
 371          */
 372         if (last_match && last_match->pid == pid)
 373                 return last_match;
 374
 375         while (*p != NULL) {
 376                 parent = *p;
 377                 th = rb_entry(parent, struct thread, rb_node);
 378
 379                 if (th->pid == pid) {
 380                         last_match = th;
 381                         return th;
 382                 }
 383
 384                 if (pid < th->pid)
 385                         p = &(*p)->rb_left;
 386                 else
 387                         p = &(*p)->rb_right;
 388         }
 389
 390         th = thread__new(pid);
 391         if (th != NULL) {
 392                 rb_link_node(&th->rb_node, parent, p);
 393                 rb_insert_color(&th->rb_node, &threads);
 394                 last_match = th;
 395         }
 396
 397         return th;
 398 }
 399
 400 static void thread__insert_map(struct thread *self, struct map *map)
 401 {
 402         struct map *pos, *tmp;
 403
 404         list_for_each_entry_safe(pos, tmp, &self->maps, node) {
 405                 if (map__overlap(pos, map)) {
 406                         if (verbose >= 2) {
 407                                 printf("overlapping maps:\n");
 408                                 map__fprintf(map, stdout);
 409                                 map__fprintf(pos, stdout);
 410                         }
 411
 412                         if (map->start <= pos->start && map->end > pos->start)
 413                                 pos->start = map->end;
 414
 415                         if (map->end >= pos->end && map->start < pos->end)
 416                                 pos->end = map->start;
 417
 418                         if (verbose >= 2) {
 419                                 printf("after collision:\n");
 420                                 map__fprintf(pos, stdout);
 421                         }
 422
 423                         if (pos->start >= pos->end) {
 424                                 list_del_init(&pos->node);
 425                                 free(pos);
 426                         }
 427                 }
 428         }
 429
 430         list_add_tail(&map->node, &self->maps);
 431 }
 432
 433 static int thread__fork(struct thread *self, struct thread *parent)
 434 {
 435         struct map *map;
 436
 437         if (self->comm)
 438                 free(self->comm);
 439         self->comm = strdup(parent->comm);
 440         if (!self->comm)
 441                 return -ENOMEM;
 442
 443         list_for_each_entry(map, &parent->maps, node) {
 444                 struct map *new = map__clone(map);
 445                 if (!new)
 446                         return -ENOMEM;
 447                 thread__insert_map(self, new);
 448         }
 449
 450         return 0;
 451 }
 452
 453 static struct map *thread__find_map(struct thread *self, u64 ip)
 454 {
 455         struct map *pos;
 456
 457         if (self == NULL)
 458                 return NULL;
 459
 460         list_for_each_entry(pos, &self->maps, node)
 461                 if (ip >= pos->start && ip <= pos->end)
 462                         return pos;
 463
 464         return NULL;
 465 }
 466
 467 static size_t threads__fprintf(FILE *fp)
 468 {
 469         size_t ret = 0;
 470         struct rb_node *nd;
 471
 472         for (nd = rb_first(&threads); nd; nd = rb_next(nd)) {
 473                 struct thread *pos = rb_entry(nd, struct thread, rb_node);
 474
 475                 ret += thread__fprintf(pos, fp);
 476         }
 477
 478         return ret;
 479 }
 480
 481 /*
 482  * histogram, sorted on item, collects counts
 483  */
 484
 485 static struct rb_root hist;
 486
 487 struct hist_entry {
 488         struct rb_node   rb_node;
 489
 490         struct thread    *thread;
 491         struct map       *map;
 492         struct dso       *dso;
 493         struct symbol    *sym;
 494         struct symbol    *parent;
 495         u64              ip;
 496         char             level;
 497
 498         u64              count;
 499 };
 500
 501 /*
 502  * configurable sorting bits
 503  */
 504
 505 struct sort_entry {
 506         struct list_head list;
 507
 508         char *header;
 509
 510         int64_t (*cmp)(struct hist_entry *, struct hist_entry *);
 511         int64_t (*collapse)(struct hist_entry *, struct hist_entry *);
 512         size_t  (*print)(FILE *fp, struct hist_entry *);
 513 };
 514
 515 static int64_t cmp_null(void *l, void *r)
 516 {
 517         if (!l && !r)
 518                 return 0;
 519         else if (!l)
 520                 return -1;
 521         else
 522                 return 1;
 523 }
 524
 525 /* --sort pid */
 526
 527 static int64_t
 528 sort__thread_cmp(struct hist_entry *left, struct hist_entry *right)
 529 {
 530         return right->thread->pid - left->thread->pid;
 531 }
 532
 533 static size_t
 534 sort__thread_print(FILE *fp, struct hist_entry *self)
 535 {
 536         return fprintf(fp, "%16s:%5d", self->thread->comm ?: "", self->thread->pid);
 537 }
 538
 539 static struct sort_entry sort_thread = {
 540         .header = "         Command:  Pid",
 541         .cmp    = sort__thread_cmp,
 542         .print  = sort__thread_print,
 543 };
 544
 545 /* --sort comm */
 546
 547 static int64_t
 548 sort__comm_cmp(struct hist_entry *left, struct hist_entry *right)
 549 {
 550         return right->thread->pid - left->thread->pid;
 551 }
 552
 553 static int64_t
 554 sort__comm_collapse(struct hist_entry *left, struct hist_entry *right)
 555 {
 556         char *comm_l = left->thread->comm;
 557         char *comm_r = right->thread->comm;
 558
 559         if (!comm_l || !comm_r)
 560                 return cmp_null(comm_l, comm_r);
 561
 562         return strcmp(comm_l, comm_r);
 563 }
 564
 565 static size_t
 566 sort__comm_print(FILE *fp, struct hist_entry *self)
 567 {
 568         return fprintf(fp, "%16s", self->thread->comm);
 569 }
 570
 571 static struct sort_entry sort_comm = {
 572         .header         = "         Command",
 573         .cmp            = sort__comm_cmp,
 574         .collapse       = sort__comm_collapse,
 575         .print          = sort__comm_print,
 576 };
 577
 578 /* --sort dso */
 579
 580 static int64_t
 581 sort__dso_cmp(struct hist_entry *left, struct hist_entry *right)
 582 {
 583         struct dso *dso_l = left->dso;
 584         struct dso *dso_r = right->dso;
 585
 586         if (!dso_l || !dso_r)
 587                 return cmp_null(dso_l, dso_r);
 588
 589         return strcmp(dso_l->name, dso_r->name);
 590 }
 591
 592 static size_t
 593 sort__dso_print(FILE *fp, struct hist_entry *self)
 594 {
 595         if (self->dso)
 596                 return fprintf(fp, "%-25s", self->dso->name);
 597
 598         return fprintf(fp, "%016llx         ", (u64)self->ip);
 599 }
 600
 601 static struct sort_entry sort_dso = {
 602         .header = "Shared Object            ",
 603         .cmp    = sort__dso_cmp,
 604         .print  = sort__dso_print,
 605 };
 606
 607 /* --sort symbol */
 608
 609 static int64_t
 610 sort__sym_cmp(struct hist_entry *left, struct hist_entry *right)
 611 {
 612         u64 ip_l, ip_r;
 613
 614         if (left->sym == right->sym)
 615                 return 0;
 616
 617         ip_l = left->sym ? left->sym->start : left->ip;
 618         ip_r = right->sym ? right->sym->start : right->ip;
 619
 620         return (int64_t)(ip_r - ip_l);
 621 }
 622
 623 static size_t
 624 sort__sym_print(FILE *fp, struct hist_entry *self)
 625 {
 626         size_t ret = 0;
 627
 628         if (verbose)
 629                 ret += fprintf(fp, "%#018llx  ", (u64)self->ip);
 630
 631         if (self->sym) {
 632                 ret += fprintf(fp, "[%c] %s",
 633                         self->dso == kernel_dso ? 'k' : '.', self->sym->name);
 634         } else {
 635                 ret += fprintf(fp, "%#016llx", (u64)self->ip);
 636         }
 637
 638         return ret;
 639 }
 640
 641 static struct sort_entry sort_sym = {
 642         .header = "Symbol",
 643         .cmp    = sort__sym_cmp,
 644         .print  = sort__sym_print,
 645 };
 646
 647 /* --sort parent */
 648
 649 static int64_t
 650 sort__parent_cmp(struct hist_entry *left, struct hist_entry *right)
 651 {
 652         struct symbol *sym_l = left->parent;
 653         struct symbol *sym_r = right->parent;
 654
 655         if (!sym_l || !sym_r)
 656                 return cmp_null(sym_l, sym_r);
 657
 658         return strcmp(sym_l->name, sym_r->name);
 659 }
 660
 661 static size_t
 662 sort__parent_print(FILE *fp, struct hist_entry *self)
 663 {
 664         size_t ret = 0;
 665
 666         ret += fprintf(fp, "%-20s", self->parent ? self->parent->name : "[other]");
 667
 668         return ret;
 669 }
 670
 671 static struct sort_entry sort_parent = {
 672         .header = "Parent symbol       ",
 673         .cmp    = sort__parent_cmp,
 674         .print  = sort__parent_print,
 675 };
 676
 677 static int sort__need_collapse = 0;
 678 static int sort__has_parent = 0;
 679
 680 struct sort_dimension {
 681         char                    *name;
 682         struct sort_entry       *entry;
 683         int                     taken;
 684 };
 685
 686 static struct sort_dimension sort_dimensions[] = {
 687         { .name = "pid",        .entry = &sort_thread,  },
 688         { .name = "comm",       .entry = &sort_comm,    },
 689         { .name = "dso",        .entry = &sort_dso,     },
 690         { .name = "symbol",     .entry = &sort_sym,     },
 691         { .name = "parent",     .entry = &sort_parent,  },
 692 };
 693
 694 static LIST_HEAD(hist_entry__sort_list);
 695
 696 static int sort_dimension__add(char *tok)
 697 {
 698         int i;
 699
 700         for (i = 0; i < ARRAY_SIZE(sort_dimensions); i++) {
 701                 struct sort_dimension *sd = &sort_dimensions[i];
 702
 703                 if (sd->taken)
 704                         continue;
 705
 706                 if (strncasecmp(tok, sd->name, strlen(tok)))
 707                         continue;
 708
 709                 if (sd->entry->collapse)
 710                         sort__need_collapse = 1;
 711
 712                 if (sd->entry == &sort_parent) {
 713                         int ret = regcomp(&parent_regex, parent_pattern, REG_EXTENDED);
 714                         if (ret) {
 715                                 char err[BUFSIZ];
 716
 717                                 regerror(ret, &parent_regex, err, sizeof(err));
 718                                 fprintf(stderr, "Invalid regex: %s\n%s",
 719                                         parent_pattern, err);
 720                                 exit(-1);
 721                         }
 722                         sort__has_parent = 1;
 723                 }
 724
 725                 list_add_tail(&sd->entry->list, &hist_entry__sort_list);
 726                 sd->taken = 1;
 727
 728                 return 0;
 729         }
 730
 731         return -ESRCH;
 732 }
 733
 734 static int64_t
 735 hist_entry__cmp(struct hist_entry *left, struct hist_entry *right)
 736 {
 737         struct sort_entry *se;
 738         int64_t cmp = 0;
 739
 740         list_for_each_entry(se, &hist_entry__sort_list, list) {
 741                 cmp = se->cmp(left, right);
 742                 if (cmp)
 743                         break;
 744         }
 745
 746         return cmp;
 747 }
 748
 749 static int64_t
 750 hist_entry__collapse(struct hist_entry *left, struct hist_entry *right)
 751 {
 752         struct sort_entry *se;
 753         int64_t cmp = 0;
 754
 755         list_for_each_entry(se, &hist_entry__sort_list, list) {
 756                 int64_t (*f)(struct hist_entry *, struct hist_entry *);
 757
 758                 f = se->collapse ?: se->cmp;
 759
 760                 cmp = f(left, right);
 761                 if (cmp)
 762                         break;
 763         }
 764
 765         return cmp;
 766 }
 767
 768 static size_t
 769 hist_entry__fprintf(FILE *fp, struct hist_entry *self, u64 total_samples)
 770 {
 771         struct sort_entry *se;
 772         size_t ret;
 773
 774         if (exclude_other && !self->parent)
 775                 return 0;
 776
 777         if (total_samples) {
 778                 double percent = self->count * 100.0 / total_samples;
 779                 char *color = PERF_COLOR_NORMAL;
 780
 781                 /*
 782                  * We color high-overhead entries in red, mid-overhead
 783                  * entries in green - and keep the low overhead places
 784                  * normal:
 785                  */
 786                 if (percent >= 5.0) {
 787                         color = PERF_COLOR_RED;
 788                 } else {
 789                         if (percent >= 0.5)
 790                                 color = PERF_COLOR_GREEN;
 791                 }
 792
 793                 ret = color_fprintf(fp, color, "   %6.2f%%",
 794                                 (self->count * 100.0) / total_samples);
 795         } else
 796                 ret = fprintf(fp, "%12Ld ", self->count);
 797
 798         list_for_each_entry(se, &hist_entry__sort_list, list) {
 799                 if (exclude_other && (se == &sort_parent))
 800                         continue;
 801
 802                 fprintf(fp, "  ");
 803                 ret += se->print(fp, self);
 804         }
 805
 806         ret += fprintf(fp, "\n");
 807
 808         return ret;
 809 }
 810
 811 /*
 812  *
 813  */
 814
 815 static struct symbol *
 816 resolve_symbol(struct thread *thread, struct map **mapp,
 817                struct dso **dsop, u64 *ipp)
 818 {
 819         struct dso *dso = dsop ? *dsop : NULL;
 820         struct map *map = mapp ? *mapp : NULL;
 821         u64 ip = *ipp;
 822
 823         if (!thread)
 824                 return NULL;
 825
 826         if (dso)
 827                 goto got_dso;
 828
 829         if (map)
 830                 goto got_map;
 831
 832         map = thread__find_map(thread, ip);
 833         if (map != NULL) {
 834                 if (mapp)
 835                         *mapp = map;
 836 got_map:
 837                 ip = map->map_ip(map, ip);
 838
 839                 dso = map->dso;
 840         } else {
 841                 /*
 842                  * If this is outside of all known maps,
 843                  * and is a negative address, try to look it
 844                  * up in the kernel dso, as it might be a
 845                  * vsyscall (which executes in user-mode):
 846                  */
 847                 if ((long long)ip < 0)
 848                 dso = kernel_dso;
 849         }
 850         dprintf(" ...... dso: %s\n", dso ? dso->name : "<not found>");
 851         dprintf(" ...... map: %Lx -> %Lx\n", *ipp, ip);
 852         *ipp  = ip;
 853
 854         if (dsop)
 855                 *dsop = dso;
 856
 857         if (!dso)
 858                 return NULL;
 859 got_dso:
 860         return dso->find_symbol(dso, ip);
 861 }
 862
 863 static int call__match(struct symbol *sym)
 864 {
 865         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
 866                 return 1;
 867
 868         return 0;
 869 }
 870
 871 /*
 872  * collect histogram counts
 873  */
 874
 875 static int
 876 hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
 877                 struct symbol *sym, u64 ip, struct ip_callchain *chain,
 878                 char level, u64 count)
 879 {
 880         struct rb_node **p = &hist.rb_node;
 881         struct rb_node *parent = NULL;
 882         struct hist_entry *he;
 883         struct hist_entry entry = {
 884                 .thread = thread,
 885                 .map    = map,
 886                 .dso    = dso,
 887                 .sym    = sym,
 888                 .ip     = ip,
 889                 .level  = level,
 890                 .count  = count,
 891                 .parent = NULL,
 892         };
 893         int cmp;
 894
 895         if (sort__has_parent && chain) {
 896                 u64 context = PERF_CONTEXT_MAX;
 897                 int i;
 898
 899                 for (i = 0; i < chain->nr; i++) {
 900                         u64 ip = chain->ips[i];
 901                         struct dso *dso = NULL;
 902                         struct symbol *sym;
 903
 904                         if (ip >= PERF_CONTEXT_MAX) {
 905                                 context = ip;
 906                                 continue;
 907                         }
 908
 909                         switch (context) {
 910                         case PERF_CONTEXT_KERNEL:
 911                                 dso = kernel_dso;
 912                                 break;
 913                         default:
 914                                 break;
 915                         }
 916
 917                         sym = resolve_symbol(thread, NULL, &dso, &ip);
 918
 919                         if (sym && call__match(sym)) {
 920                                 entry.parent = sym;
 921                                 break;
 922                         }
 923                 }
 924         }
 925
 926         while (*p != NULL) {
 927                 parent = *p;
 928                 he = rb_entry(parent, struct hist_entry, rb_node);
 929
 930                 cmp = hist_entry__cmp(&entry, he);
 931
 932                 if (!cmp) {
 933                         he->count += count;
 934                         return 0;
 935                 }
 936
 937                 if (cmp < 0)
 938                         p = &(*p)->rb_left;
 939                 else
 940                         p = &(*p)->rb_right;
 941         }
 942
 943         he = malloc(sizeof(*he));
 944         if (!he)
 945                 return -ENOMEM;
 946         *he = entry;
 947         rb_link_node(&he->rb_node, parent, p);
 948         rb_insert_color(&he->rb_node, &hist);
 949
 950         return 0;
 951 }
 952
 953 static void hist_entry__free(struct hist_entry *he)
 954 {
 955         free(he);
 956 }
 957
 958 /*
 959  * collapse the histogram
 960  */
 961
 962 static struct rb_root collapse_hists;
 963
 964 static void collapse__insert_entry(struct hist_entry *he)
 965 {
 966         struct rb_node **p = &collapse_hists.rb_node;
 967         struct rb_node *parent = NULL;
 968         struct hist_entry *iter;
 969         int64_t cmp;
 970
 971         while (*p != NULL) {
 972                 parent = *p;
 973                 iter = rb_entry(parent, struct hist_entry, rb_node);
 974
 975                 cmp = hist_entry__collapse(iter, he);
 976
 977                 if (!cmp) {
 978                         iter->count += he->count;
 979                         hist_entry__free(he);
 980                         return;
 981                 }
 982
 983                 if (cmp < 0)
 984                         p = &(*p)->rb_left;
 985                 else
 986                         p = &(*p)->rb_right;
 987         }
 988
 989         rb_link_node(&he->rb_node, parent, p);
 990         rb_insert_color(&he->rb_node, &collapse_hists);
 991 }
 992
 993 static void collapse__resort(void)
 994 {
 995         struct rb_node *next;
 996         struct hist_entry *n;
 997
 998         if (!sort__need_collapse)
 999                 return;
1000
1001         next = rb_first(&hist);
1002         while (next) {
1003                 n = rb_entry(next, struct hist_entry, rb_node);
1004                 next = rb_next(&n->rb_node);
1005
1006                 rb_erase(&n->rb_node, &hist);
1007                 collapse__insert_entry(n);
1008         }
1009 }
1010
1011 /*
1012  * reverse the map, sort on count.
1013  */
1014
1015 static struct rb_root output_hists;
1016
1017 static void output__insert_entry(struct hist_entry *he)
1018 {
1019         struct rb_node **p = &output_hists.rb_node;
1020         struct rb_node *parent = NULL;
1021         struct hist_entry *iter;
1022
1023         while (*p != NULL) {
1024                 parent = *p;
1025                 iter = rb_entry(parent, struct hist_entry, rb_node);
1026
1027                 if (he->count > iter->count)
1028                         p = &(*p)->rb_left;
1029                 else
1030                         p = &(*p)->rb_right;
1031         }
1032
1033         rb_link_node(&he->rb_node, parent, p);
1034         rb_insert_color(&he->rb_node, &output_hists);
1035 }
1036
1037 static void output__resort(void)
1038 {
1039         struct rb_node *next;
1040         struct hist_entry *n;
1041         struct rb_root *tree = &hist;
1042
1043         if (sort__need_collapse)
1044                 tree = &collapse_hists;
1045
1046         next = rb_first(tree);
1047
1048         while (next) {
1049                 n = rb_entry(next, struct hist_entry, rb_node);
1050                 next = rb_next(&n->rb_node);
1051
1052                 rb_erase(&n->rb_node, tree);
1053                 output__insert_entry(n);
1054         }
1055 }
1056
1057 static size_t output__fprintf(FILE *fp, u64 total_samples)
1058 {
1059         struct hist_entry *pos;
1060         struct sort_entry *se;
1061         struct rb_node *nd;
1062         size_t ret = 0;
1063
1064         fprintf(fp, "\n");
1065         fprintf(fp, "#\n");
1066         fprintf(fp, "# (%Ld samples)\n", (u64)total_samples);
1067         fprintf(fp, "#\n");
1068
1069         fprintf(fp, "# Overhead");
1070         list_for_each_entry(se, &hist_entry__sort_list, list) {
1071                 if (exclude_other && (se == &sort_parent))
1072                         continue;
1073                 fprintf(fp, "  %s", se->header);
1074         }
1075         fprintf(fp, "\n");
1076
1077         fprintf(fp, "# ........");
1078         list_for_each_entry(se, &hist_entry__sort_list, list) {
1079                 int i;
1080
1081                 if (exclude_other && (se == &sort_parent))
1082                         continue;
1083
1084                 fprintf(fp, "  ");
1085                 for (i = 0; i < strlen(se->header); i++)
1086                         fprintf(fp, ".");
1087         }
1088         fprintf(fp, "\n");
1089
1090         fprintf(fp, "#\n");
1091
1092         for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
1093                 pos = rb_entry(nd, struct hist_entry, rb_node);
1094                 ret += hist_entry__fprintf(fp, pos, total_samples);
1095         }
1096
1097         if (sort_order == default_sort_order &&
1098                         parent_pattern == default_parent_pattern) {
1099                 fprintf(fp, "#\n");
1100                 fprintf(fp, "# (For more details, try: perf report --sort comm,dso,symbol)\n");
1101                 fprintf(fp, "#\n");
1102         }
1103         fprintf(fp, "\n");
1104
1105         return ret;
1106 }
1107
1108 static void register_idle_thread(void)
1109 {
1110         struct thread *thread = threads__findnew(0);
1111
1112         if (thread == NULL ||
1113                         thread__set_comm(thread, "[idle]")) {
1114                 fprintf(stderr, "problem inserting idle task.\n");
1115                 exit(-1);
1116         }
1117 }
1118
1119 static unsigned long total = 0,
1120                      total_mmap = 0,
1121                      total_comm = 0,
1122                      total_fork = 0,
1123                      total_unknown = 0,
1124                      total_lost = 0;
1125
1126 static int validate_chain(struct ip_callchain *chain, event_t *event)
1127 {
1128         unsigned int chain_size;
1129
1130         chain_size = event->header.size;
1131         chain_size -= (unsigned long)&event->ip.__more_data - (unsigned long)event;
1132
1133         if (chain->nr*sizeof(u64) > chain_size)
1134                 return -1;
1135
1136         return 0;
1137 }
1138
1139 static int
1140 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
1141 {
1142         char level;
1143         int show = 0;
1144         struct dso *dso = NULL;
1145         struct thread *thread = threads__findnew(event->ip.pid);
1146         u64 ip = event->ip.ip;
1147         u64 period = 1;
1148         struct map *map = NULL;
1149         void *more_data = event->ip.__more_data;
1150         struct ip_callchain *chain = NULL;
1151
1152         if (sample_type & PERF_SAMPLE_PERIOD) {
1153                 period = *(u64 *)more_data;
1154                 more_data += sizeof(u64);
1155         }
1156
1157         dprintf("%p [%p]: PERF_EVENT_SAMPLE (IP, %d): %d: %p period: %Ld\n",
1158                 (void *)(offset + head),
1159                 (void *)(long)(event->header.size),
1160                 event->header.misc,
1161                 event->ip.pid,
1162                 (void *)(long)ip,
1163                 (long long)period);
1164
1165         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
1166                 int i;
1167
1168                 chain = (void *)more_data;
1169
1170                 dprintf("... chain: nr:%Lu\n", chain->nr);
1171
1172                 if (validate_chain(chain, event) < 0) {
1173                         eprintf("call-chain problem with event, skipping it.\n");
1174                         return 0;
1175                 }
1176
1177                 if (dump_trace) {
1178                         for (i = 0; i < chain->nr; i++)
1179                                 dprintf("..... %2d: %016Lx\n", i, chain->ips[i]);
1180                 }
1181         }
1182
1183         dprintf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1184
1185         if (thread == NULL) {
1186                 eprintf("problem processing %d event, skipping it.\n",
1187                         event->header.type);
1188                 return -1;
1189         }
1190
1191         if (event->header.misc & PERF_EVENT_MISC_KERNEL) {
1192                 show = SHOW_KERNEL;
1193                 level = 'k';
1194
1195                 dso = kernel_dso;
1196
1197                 dprintf(" ...... dso: %s\n", dso->name);
1198
1199         } else if (event->header.misc & PERF_EVENT_MISC_USER) {
1200
1201                 show = SHOW_USER;
1202                 level = '.';
1203
1204         } else {
1205                 show = SHOW_HV;
1206                 level = 'H';
1207                 dprintf(" ...... dso: [hypervisor]\n");
1208         }
1209
1210         if (show & show_mask) {
1211                 struct symbol *sym = resolve_symbol(thread, &map, &dso, &ip);
1212
1213                 if (hist_entry__add(thread, map, dso, sym, ip, chain, level, period)) {
1214                         eprintf("problem incrementing symbol count, skipping event\n");
1215                         return -1;
1216                 }
1217         }
1218         total += period;
1219
1220         return 0;
1221 }
1222
1223 static int
1224 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
1225 {
1226         struct thread *thread = threads__findnew(event->mmap.pid);
1227         struct map *map = map__new(&event->mmap);
1228
1229         dprintf("%p [%p]: PERF_EVENT_MMAP %d: [%p(%p) @ %p]: %s\n",
1230                 (void *)(offset + head),
1231                 (void *)(long)(event->header.size),
1232                 event->mmap.pid,
1233                 (void *)(long)event->mmap.start,
1234                 (void *)(long)event->mmap.len,
1235                 (void *)(long)event->mmap.pgoff,
1236                 event->mmap.filename);
1237
1238         if (thread == NULL || map == NULL) {
1239                 dprintf("problem processing PERF_EVENT_MMAP, skipping event.\n");
1240                 return 0;
1241         }
1242
1243         thread__insert_map(thread, map);
1244         total_mmap++;
1245
1246         return 0;
1247 }
1248
1249 static int
1250 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
1251 {
1252         struct thread *thread = threads__findnew(event->comm.pid);
1253
1254         dprintf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
1255                 (void *)(offset + head),
1256                 (void *)(long)(event->header.size),
1257                 event->comm.comm, event->comm.pid);
1258
1259         if (thread == NULL ||
1260             thread__set_comm(thread, event->comm.comm)) {
1261                 dprintf("problem processing PERF_EVENT_COMM, skipping event.\n");
1262                 return -1;
1263         }
1264         total_comm++;
1265
1266         return 0;
1267 }
1268
1269 static int
1270 process_fork_event(event_t *event, unsigned long offset, unsigned long head)
1271 {
1272         struct thread *thread = threads__findnew(event->fork.pid);
1273         struct thread *parent = threads__findnew(event->fork.ppid);
1274
1275         dprintf("%p [%p]: PERF_EVENT_FORK: %d:%d\n",
1276                 (void *)(offset + head),
1277                 (void *)(long)(event->header.size),
1278                 event->fork.pid, event->fork.ppid);
1279
1280         if (!thread || !parent || thread__fork(thread, parent)) {
1281                 dprintf("problem processing PERF_EVENT_FORK, skipping event.\n");
1282                 return -1;
1283         }
1284         total_fork++;
1285
1286         return 0;
1287 }
1288
1289 static int
1290 process_period_event(event_t *event, unsigned long offset, unsigned long head)
1291 {
1292         dprintf("%p [%p]: PERF_EVENT_PERIOD: time:%Ld, id:%Ld: period:%Ld\n",
1293                 (void *)(offset + head),
1294                 (void *)(long)(event->header.size),
1295                 event->period.time,
1296                 event->period.id,
1297                 event->period.sample_period);
1298
1299         return 0;
1300 }
1301
1302 static int
1303 process_lost_event(event_t *event, unsigned long offset, unsigned long head)
1304 {
1305         dprintf("%p [%p]: PERF_EVENT_LOST: id:%Ld: lost:%Ld\n",
1306                 (void *)(offset + head),
1307                 (void *)(long)(event->header.size),
1308                 event->lost.id,
1309                 event->lost.lost);
1310
1311         total_lost += event->lost.lost;
1312
1313         return 0;
1314 }
1315
1316 static void trace_event(event_t *event)
1317 {
1318         unsigned char *raw_event = (void *)event;
1319         char *color = PERF_COLOR_BLUE;
1320         int i, j;
1321
1322         if (!dump_trace)
1323                 return;
1324
1325         dprintf(".");
1326         cdprintf("\n. ... raw event: size %d bytes\n", event->header.size);
1327
1328         for (i = 0; i < event->header.size; i++) {
1329                 if ((i & 15) == 0) {
1330                         dprintf(".");
1331                         cdprintf("  %04x: ", i);
1332                 }
1333
1334                 cdprintf(" %02x", raw_event[i]);
1335
1336                 if (((i & 15) == 15) || i == event->header.size-1) {
1337                         cdprintf("  ");
1338                         for (j = 0; j < 15-(i & 15); j++)
1339                                 cdprintf("   ");
1340                         for (j = 0; j < (i & 15); j++) {
1341                                 if (isprint(raw_event[i-15+j]))
1342                                         cdprintf("%c", raw_event[i-15+j]);
1343                                 else
1344                                         cdprintf(".");
1345                         }
1346                         cdprintf("\n");
1347                 }
1348         }
1349         dprintf(".\n");
1350 }
1351
1352 static int
1353 process_event(event_t *event, unsigned long offset, unsigned long head)
1354 {
1355         trace_event(event);
1356
1357         switch (event->header.type) {
1358         case PERF_EVENT_SAMPLE:
1359                 return process_sample_event(event, offset, head);
1360
1361         case PERF_EVENT_MMAP:
1362                 return process_mmap_event(event, offset, head);
1363
1364         case PERF_EVENT_COMM:
1365                 return process_comm_event(event, offset, head);
1366
1367         case PERF_EVENT_FORK:
1368                 return process_fork_event(event, offset, head);
1369
1370         case PERF_EVENT_PERIOD:
1371                 return process_period_event(event, offset, head);
1372
1373         case PERF_EVENT_LOST:
1374                 return process_lost_event(event, offset, head);
1375
1376         /*
1377          * We dont process them right now but they are fine:
1378          */
1379
1380         case PERF_EVENT_THROTTLE:
1381         case PERF_EVENT_UNTHROTTLE:
1382                 return 0;
1383
1384         default:
1385                 return -1;
1386         }
1387
1388         return 0;
1389 }
1390
1391 static struct perf_header       *header;
1392
1393 static u64 perf_header__sample_type(void)
1394 {
1395         u64 sample_type = 0;
1396         int i;
1397
1398         for (i = 0; i < header->attrs; i++) {
1399                 struct perf_header_attr *attr = header->attr[i];
1400
1401                 if (!sample_type)
1402                         sample_type = attr->attr.sample_type;
1403                 else if (sample_type != attr->attr.sample_type)
1404                         die("non matching sample_type");
1405         }
1406
1407         return sample_type;
1408 }
1409
1410 static int __cmd_report(void)
1411 {
1412         int ret, rc = EXIT_FAILURE;
1413         unsigned long offset = 0;
1414         unsigned long head, shift;
1415         struct stat stat;
1416         event_t *event;
1417         uint32_t size;
1418         char *buf;
1419
1420         register_idle_thread();
1421
1422         input = open(input_name, O_RDONLY);
1423         if (input < 0) {
1424                 fprintf(stderr, " failed to open file: %s", input_name);
1425                 if (!strcmp(input_name, "perf.data"))
1426                         fprintf(stderr, "  (try 'perf record' first)");
1427                 fprintf(stderr, "\n");
1428                 exit(-1);
1429         }
1430
1431         ret = fstat(input, &stat);
1432         if (ret < 0) {
1433                 perror("failed to stat file");
1434                 exit(-1);
1435         }
1436
1437         if (!stat.st_size) {
1438                 fprintf(stderr, "zero-sized file, nothing to do!\n");
1439                 exit(0);
1440         }
1441
1442         header = perf_header__read(input);
1443         head = header->data_offset;
1444
1445         sample_type = perf_header__sample_type();
1446
1447         if (sort__has_parent && !(sample_type & PERF_SAMPLE_CALLCHAIN)) {
1448                 fprintf(stderr, "selected --sort parent, but no callchain data\n");
1449                 exit(-1);
1450         }
1451
1452         if (load_kernel() < 0) {
1453                 perror("failed to load kernel symbols");
1454                 return EXIT_FAILURE;
1455         }
1456
1457         if (!full_paths) {
1458                 if (getcwd(__cwd, sizeof(__cwd)) == NULL) {
1459                         perror("failed to get the current directory");
1460                         return EXIT_FAILURE;
1461                 }
1462                 cwdlen = strlen(cwd);
1463         } else {
1464                 cwd = NULL;
1465                 cwdlen = 0;
1466         }
1467
1468         shift = page_size * (head / page_size);
1469         offset += shift;
1470         head -= shift;
1471
1472 remap:
1473         buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1474                            MAP_SHARED, input, offset);
1475         if (buf == MAP_FAILED) {
1476                 perror("failed to mmap file");
1477                 exit(-1);
1478         }
1479
1480 more:
1481         event = (event_t *)(buf + head);
1482
1483         size = event->header.size;
1484         if (!size)
1485                 size = 8;
1486
1487         if (head + event->header.size >= page_size * mmap_window) {
1488                 int ret;
1489
1490                 shift = page_size * (head / page_size);
1491
1492                 ret = munmap(buf, page_size * mmap_window);
1493                 assert(ret == 0);
1494
1495                 offset += shift;
1496                 head -= shift;
1497                 goto remap;
1498         }
1499
1500         size = event->header.size;
1501
1502         dprintf("\n%p [%p]: event: %d\n",
1503                         (void *)(offset + head),
1504                         (void *)(long)event->header.size,
1505                         event->header.type);
1506
1507         if (!size || process_event(event, offset, head) < 0) {
1508
1509                 dprintf("%p [%p]: skipping unknown header type: %d\n",
1510                         (void *)(offset + head),
1511                         (void *)(long)(event->header.size),
1512                         event->header.type);
1513
1514                 total_unknown++;
1515
1516                 /*
1517                  * assume we lost track of the stream, check alignment, and
1518                  * increment a single u64 in the hope to catch on again 'soon'.
1519                  */
1520
1521                 if (unlikely(head & 7))
1522                         head &= ~7ULL;
1523
1524                 size = 8;
1525         }
1526
1527         head += size;
1528
1529         if (offset + head >= header->data_offset + header->data_size)
1530                 goto done;
1531
1532         if (offset + head < stat.st_size)
1533                 goto more;
1534
1535 done:
1536         rc = EXIT_SUCCESS;
1537         close(input);
1538
1539         dprintf("      IP events: %10ld\n", total);
1540         dprintf("    mmap events: %10ld\n", total_mmap);
1541         dprintf("    comm events: %10ld\n", total_comm);
1542         dprintf("    fork events: %10ld\n", total_fork);
1543         dprintf("    lost events: %10ld\n", total_lost);
1544         dprintf(" unknown events: %10ld\n", total_unknown);
1545
1546         if (dump_trace)
1547                 return 0;
1548
1549         if (verbose >= 3)
1550                 threads__fprintf(stdout);
1551
1552         if (verbose >= 2)
1553                 dsos__fprintf(stdout);
1554
1555         collapse__resort();
1556         output__resort();
1557         output__fprintf(stdout, total);
1558
1559         return rc;
1560 }
1561
1562 static const char * const report_usage[] = {
1563         "perf report [<options>] <command>",
1564         NULL
1565 };
1566
1567 static const struct option options[] = {
1568         OPT_STRING('i', "input", &input_name, "file",
1569                     "input file name"),
1570         OPT_BOOLEAN('v', "verbose", &verbose,
1571                     "be more verbose (show symbol address, etc)"),
1572         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1573                     "dump raw trace in ASCII"),
1574         OPT_STRING('k', "vmlinux", &vmlinux, "file", "vmlinux pathname"),
1575         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1576                    "sort by key(s): pid, comm, dso, symbol, parent"),
1577         OPT_BOOLEAN('P', "full-paths", &full_paths,
1578                     "Don't shorten the pathnames taking into account the cwd"),
1579         OPT_STRING('p', "parent", &parent_pattern, "regex",
1580                    "regex filter to identify parent, see: '--sort parent'"),
1581         OPT_BOOLEAN('x', "exclude-other", &exclude_other,
1582                     "Only display entries with parent-match"),
1583         OPT_END()
1584 };
1585
1586 static void setup_sorting(void)
1587 {
1588         char *tmp, *tok, *str = strdup(sort_order);
1589
1590         for (tok = strtok_r(str, ", ", &tmp);
1591                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1592                 if (sort_dimension__add(tok) < 0) {
1593                         error("Unknown --sort key: `%s'", tok);
1594                         usage_with_options(report_usage, options);
1595                 }
1596         }
1597
1598         free(str);
1599 }
1600
1601 int cmd_report(int argc, const char **argv, const char *prefix)
1602 {
1603         symbol__init();
1604
1605         page_size = getpagesize();
1606
1607         argc = parse_options(argc, argv, options, report_usage, 0);
1608
1609         setup_sorting();
1610
1611         if (parent_pattern != default_parent_pattern)
1612                 sort_dimension__add("parent");
1613         else
1614                 exclude_other = 0;
1615
1616         /*
1617          * Any (unrecognized) arguments left?
1618          */
1619         if (argc)
1620                 usage_with_options(report_usage, options);
1621
1622         setup_pager();
1623
1624         return __cmd_report();
1625 }