SAFE public projects git trees. - safe/jmp/linux-2.6/blob - tools/perf/builtin-report.c

   1 /*
   2  * builtin-report.c
   3  *
   4  * Builtin report command: Analyze the perf.data input file,
   5  * look up and read DSOs and symbol information and display
   6  * a histogram of results, along various sorting keys.
   7  */
   8 #include "builtin.h"
   9
  10 #include "util/util.h"
  11
  12 #include "util/color.h"
  13 #include "util/list.h"
  14 #include "util/cache.h"
  15 #include "util/rbtree.h"
  16 #include "util/symbol.h"
  17 #include "util/string.h"
  18
  19 #include "perf.h"
  20
  21 #include "util/parse-options.h"
  22 #include "util/parse-events.h"
  23
  24 #define SHOW_KERNEL     1
  25 #define SHOW_USER       2
  26 #define SHOW_HV         4
  27
  28 static char             const *input_name = "perf.data";
  29 static char             *vmlinux = NULL;
  30
  31 static char             default_sort_order[] = "comm,dso";
  32 static char             *sort_order = default_sort_order;
  33
  34 static int              input;
  35 static int              show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV;
  36
  37 static int              dump_trace = 0;
  38 #define dprintf(x...)   do { if (dump_trace) printf(x); } while (0)
  39 #define cdprintf(x...)  do { if (dump_trace) color_fprintf(stdout, color, x); } while (0)
  40
  41 static int              verbose;
  42 #define eprintf(x...)   do { if (verbose) fprintf(stderr, x); } while (0)
  43
  44 static int              full_paths;
  45
  46 static unsigned long    page_size;
  47 static unsigned long    mmap_window = 32;
  48
  49 static char             default_parent_pattern[] = "^sys_|^do_page_fault";
  50 static char             *parent_pattern = default_parent_pattern;
  51 static regex_t          parent_regex;
  52
  53 static int              exclude_other = 1;
  54
  55 struct ip_event {
  56         struct perf_event_header header;
  57         u64 ip;
  58         u32 pid, tid;
  59         unsigned char __more_data[];
  60 };
  61
  62 struct ip_callchain {
  63         u64 nr;
  64         u64 ips[0];
  65 };
  66
  67 struct mmap_event {
  68         struct perf_event_header header;
  69         u32 pid, tid;
  70         u64 start;
  71         u64 len;
  72         u64 pgoff;
  73         char filename[PATH_MAX];
  74 };
  75
  76 struct comm_event {
  77         struct perf_event_header header;
  78         u32 pid, tid;
  79         char comm[16];
  80 };
  81
  82 struct fork_event {
  83         struct perf_event_header header;
  84         u32 pid, ppid;
  85 };
  86
  87 struct period_event {
  88         struct perf_event_header header;
  89         u64 time;
  90         u64 id;
  91         u64 sample_period;
  92 };
  93
  94 struct lost_event {
  95         struct perf_event_header header;
  96         u64 id;
  97         u64 lost;
  98 };
  99
 100 typedef union event_union {
 101         struct perf_event_header        header;
 102         struct ip_event                 ip;
 103         struct mmap_event               mmap;
 104         struct comm_event               comm;
 105         struct fork_event               fork;
 106         struct period_event             period;
 107         struct lost_event               lost;
 108 } event_t;
 109
 110 static LIST_HEAD(dsos);
 111 static struct dso *kernel_dso;
 112 static struct dso *vdso;
 113
 114 static void dsos__add(struct dso *dso)
 115 {
 116         list_add_tail(&dso->node, &dsos);
 117 }
 118
 119 static struct dso *dsos__find(const char *name)
 120 {
 121         struct dso *pos;
 122
 123         list_for_each_entry(pos, &dsos, node)
 124                 if (strcmp(pos->name, name) == 0)
 125                         return pos;
 126         return NULL;
 127 }
 128
 129 static struct dso *dsos__findnew(const char *name)
 130 {
 131         struct dso *dso = dsos__find(name);
 132         int nr;
 133
 134         if (dso)
 135                 return dso;
 136
 137         dso = dso__new(name, 0);
 138         if (!dso)
 139                 goto out_delete_dso;
 140
 141         nr = dso__load(dso, NULL, verbose);
 142         if (nr < 0) {
 143                 eprintf("Failed to open: %s\n", name);
 144                 goto out_delete_dso;
 145         }
 146         if (!nr)
 147                 eprintf("No symbols found in: %s, maybe install a debug package?\n", name);
 148
 149         dsos__add(dso);
 150
 151         return dso;
 152
 153 out_delete_dso:
 154         dso__delete(dso);
 155         return NULL;
 156 }
 157
 158 static void dsos__fprintf(FILE *fp)
 159 {
 160         struct dso *pos;
 161
 162         list_for_each_entry(pos, &dsos, node)
 163                 dso__fprintf(pos, fp);
 164 }
 165
 166 static struct symbol *vdso__find_symbol(struct dso *dso, u64 ip)
 167 {
 168         return dso__find_symbol(kernel_dso, ip);
 169 }
 170
 171 static int load_kernel(void)
 172 {
 173         int err;
 174
 175         kernel_dso = dso__new("[kernel]", 0);
 176         if (!kernel_dso)
 177                 return -1;
 178
 179         err = dso__load_kernel(kernel_dso, vmlinux, NULL, verbose);
 180         if (err) {
 181                 dso__delete(kernel_dso);
 182                 kernel_dso = NULL;
 183         } else
 184                 dsos__add(kernel_dso);
 185
 186         vdso = dso__new("[vdso]", 0);
 187         if (!vdso)
 188                 return -1;
 189
 190         vdso->find_symbol = vdso__find_symbol;
 191
 192         dsos__add(vdso);
 193
 194         return err;
 195 }
 196
 197 static char __cwd[PATH_MAX];
 198 static char *cwd = __cwd;
 199 static int cwdlen;
 200
 201 static int strcommon(const char *pathname)
 202 {
 203         int n = 0;
 204
 205         while (pathname[n] == cwd[n] && n < cwdlen)
 206                 ++n;
 207
 208         return n;
 209 }
 210
 211 struct map {
 212         struct list_head node;
 213         u64      start;
 214         u64      end;
 215         u64      pgoff;
 216         u64      (*map_ip)(struct map *, u64);
 217         struct dso       *dso;
 218 };
 219
 220 static u64 map__map_ip(struct map *map, u64 ip)
 221 {
 222         return ip - map->start + map->pgoff;
 223 }
 224
 225 static u64 vdso__map_ip(struct map *map, u64 ip)
 226 {
 227         return ip;
 228 }
 229
 230 static inline int is_anon_memory(const char *filename)
 231 {
 232      return strcmp(filename, "//anon") == 0;
 233 }
 234
 235 static struct map *map__new(struct mmap_event *event)
 236 {
 237         struct map *self = malloc(sizeof(*self));
 238
 239         if (self != NULL) {
 240                 const char *filename = event->filename;
 241                 char newfilename[PATH_MAX];
 242                 int anon;
 243
 244                 if (cwd) {
 245                         int n = strcommon(filename);
 246
 247                         if (n == cwdlen) {
 248                                 snprintf(newfilename, sizeof(newfilename),
 249                                          ".%s", filename + n);
 250                                 filename = newfilename;
 251                         }
 252                 }
 253
 254                 anon = is_anon_memory(filename);
 255
 256                 if (anon) {
 257                         snprintf(newfilename, sizeof(newfilename), "/tmp/perf-%d.map", event->pid);
 258                         filename = newfilename;
 259                 }
 260
 261                 self->start = event->start;
 262                 self->end   = event->start + event->len;
 263                 self->pgoff = event->pgoff;
 264
 265                 self->dso = dsos__findnew(filename);
 266                 if (self->dso == NULL)
 267                         goto out_delete;
 268
 269                 if (self->dso == vdso || anon)
 270                         self->map_ip = vdso__map_ip;
 271                 else
 272                         self->map_ip = map__map_ip;
 273         }
 274         return self;
 275 out_delete:
 276         free(self);
 277         return NULL;
 278 }
 279
 280 static struct map *map__clone(struct map *self)
 281 {
 282         struct map *map = malloc(sizeof(*self));
 283
 284         if (!map)
 285                 return NULL;
 286
 287         memcpy(map, self, sizeof(*self));
 288
 289         return map;
 290 }
 291
 292 static int map__overlap(struct map *l, struct map *r)
 293 {
 294         if (l->start > r->start) {
 295                 struct map *t = l;
 296                 l = r;
 297                 r = t;
 298         }
 299
 300         if (l->end > r->start)
 301                 return 1;
 302
 303         return 0;
 304 }
 305
 306 static size_t map__fprintf(struct map *self, FILE *fp)
 307 {
 308         return fprintf(fp, " %Lx-%Lx %Lx %s\n",
 309                        self->start, self->end, self->pgoff, self->dso->name);
 310 }
 311
 312
 313 struct thread {
 314         struct rb_node   rb_node;
 315         struct list_head maps;
 316         pid_t            pid;
 317         char             *comm;
 318 };
 319
 320 static struct thread *thread__new(pid_t pid)
 321 {
 322         struct thread *self = malloc(sizeof(*self));
 323
 324         if (self != NULL) {
 325                 self->pid = pid;
 326                 self->comm = malloc(32);
 327                 if (self->comm)
 328                         snprintf(self->comm, 32, ":%d", self->pid);
 329                 INIT_LIST_HEAD(&self->maps);
 330         }
 331
 332         return self;
 333 }
 334
 335 static int thread__set_comm(struct thread *self, const char *comm)
 336 {
 337         if (self->comm)
 338                 free(self->comm);
 339         self->comm = strdup(comm);
 340         return self->comm ? 0 : -ENOMEM;
 341 }
 342
 343 static size_t thread__fprintf(struct thread *self, FILE *fp)
 344 {
 345         struct map *pos;
 346         size_t ret = fprintf(fp, "Thread %d %s\n", self->pid, self->comm);
 347
 348         list_for_each_entry(pos, &self->maps, node)
 349                 ret += map__fprintf(pos, fp);
 350
 351         return ret;
 352 }
 353
 354
 355 static struct rb_root threads;
 356 static struct thread *last_match;
 357
 358 static struct thread *threads__findnew(pid_t pid)
 359 {
 360         struct rb_node **p = &threads.rb_node;
 361         struct rb_node *parent = NULL;
 362         struct thread *th;
 363
 364         /*
 365          * Font-end cache - PID lookups come in blocks,
 366          * so most of the time we dont have to look up
 367          * the full rbtree:
 368          */
 369         if (last_match && last_match->pid == pid)
 370                 return last_match;
 371
 372         while (*p != NULL) {
 373                 parent = *p;
 374                 th = rb_entry(parent, struct thread, rb_node);
 375
 376                 if (th->pid == pid) {
 377                         last_match = th;
 378                         return th;
 379                 }
 380
 381                 if (pid < th->pid)
 382                         p = &(*p)->rb_left;
 383                 else
 384                         p = &(*p)->rb_right;
 385         }
 386
 387         th = thread__new(pid);
 388         if (th != NULL) {
 389                 rb_link_node(&th->rb_node, parent, p);
 390                 rb_insert_color(&th->rb_node, &threads);
 391                 last_match = th;
 392         }
 393
 394         return th;
 395 }
 396
 397 static void thread__insert_map(struct thread *self, struct map *map)
 398 {
 399         struct map *pos, *tmp;
 400
 401         list_for_each_entry_safe(pos, tmp, &self->maps, node) {
 402                 if (map__overlap(pos, map)) {
 403                         list_del_init(&pos->node);
 404                         /* XXX leaks dsos */
 405                         free(pos);
 406                 }
 407         }
 408
 409         list_add_tail(&map->node, &self->maps);
 410 }
 411
 412 static int thread__fork(struct thread *self, struct thread *parent)
 413 {
 414         struct map *map;
 415
 416         if (self->comm)
 417                 free(self->comm);
 418         self->comm = strdup(parent->comm);
 419         if (!self->comm)
 420                 return -ENOMEM;
 421
 422         list_for_each_entry(map, &parent->maps, node) {
 423                 struct map *new = map__clone(map);
 424                 if (!new)
 425                         return -ENOMEM;
 426                 thread__insert_map(self, new);
 427         }
 428
 429         return 0;
 430 }
 431
 432 static struct map *thread__find_map(struct thread *self, u64 ip)
 433 {
 434         struct map *pos;
 435
 436         if (self == NULL)
 437                 return NULL;
 438
 439         list_for_each_entry(pos, &self->maps, node)
 440                 if (ip >= pos->start && ip <= pos->end)
 441                         return pos;
 442
 443         return NULL;
 444 }
 445
 446 static size_t threads__fprintf(FILE *fp)
 447 {
 448         size_t ret = 0;
 449         struct rb_node *nd;
 450
 451         for (nd = rb_first(&threads); nd; nd = rb_next(nd)) {
 452                 struct thread *pos = rb_entry(nd, struct thread, rb_node);
 453
 454                 ret += thread__fprintf(pos, fp);
 455         }
 456
 457         return ret;
 458 }
 459
 460 /*
 461  * histogram, sorted on item, collects counts
 462  */
 463
 464 static struct rb_root hist;
 465
 466 struct hist_entry {
 467         struct rb_node   rb_node;
 468
 469         struct thread    *thread;
 470         struct map       *map;
 471         struct dso       *dso;
 472         struct symbol    *sym;
 473         struct symbol    *parent;
 474         u64              ip;
 475         char             level;
 476
 477         u64              count;
 478 };
 479
 480 /*
 481  * configurable sorting bits
 482  */
 483
 484 struct sort_entry {
 485         struct list_head list;
 486
 487         char *header;
 488
 489         int64_t (*cmp)(struct hist_entry *, struct hist_entry *);
 490         int64_t (*collapse)(struct hist_entry *, struct hist_entry *);
 491         size_t  (*print)(FILE *fp, struct hist_entry *);
 492 };
 493
 494 static int64_t cmp_null(void *l, void *r)
 495 {
 496         if (!l && !r)
 497                 return 0;
 498         else if (!l)
 499                 return -1;
 500         else
 501                 return 1;
 502 }
 503
 504 /* --sort pid */
 505
 506 static int64_t
 507 sort__thread_cmp(struct hist_entry *left, struct hist_entry *right)
 508 {
 509         return right->thread->pid - left->thread->pid;
 510 }
 511
 512 static size_t
 513 sort__thread_print(FILE *fp, struct hist_entry *self)
 514 {
 515         return fprintf(fp, "%16s:%5d", self->thread->comm ?: "", self->thread->pid);
 516 }
 517
 518 static struct sort_entry sort_thread = {
 519         .header = "         Command:  Pid",
 520         .cmp    = sort__thread_cmp,
 521         .print  = sort__thread_print,
 522 };
 523
 524 /* --sort comm */
 525
 526 static int64_t
 527 sort__comm_cmp(struct hist_entry *left, struct hist_entry *right)
 528 {
 529         return right->thread->pid - left->thread->pid;
 530 }
 531
 532 static int64_t
 533 sort__comm_collapse(struct hist_entry *left, struct hist_entry *right)
 534 {
 535         char *comm_l = left->thread->comm;
 536         char *comm_r = right->thread->comm;
 537
 538         if (!comm_l || !comm_r)
 539                 return cmp_null(comm_l, comm_r);
 540
 541         return strcmp(comm_l, comm_r);
 542 }
 543
 544 static size_t
 545 sort__comm_print(FILE *fp, struct hist_entry *self)
 546 {
 547         return fprintf(fp, "%16s", self->thread->comm);
 548 }
 549
 550 static struct sort_entry sort_comm = {
 551         .header         = "         Command",
 552         .cmp            = sort__comm_cmp,
 553         .collapse       = sort__comm_collapse,
 554         .print          = sort__comm_print,
 555 };
 556
 557 /* --sort dso */
 558
 559 static int64_t
 560 sort__dso_cmp(struct hist_entry *left, struct hist_entry *right)
 561 {
 562         struct dso *dso_l = left->dso;
 563         struct dso *dso_r = right->dso;
 564
 565         if (!dso_l || !dso_r)
 566                 return cmp_null(dso_l, dso_r);
 567
 568         return strcmp(dso_l->name, dso_r->name);
 569 }
 570
 571 static size_t
 572 sort__dso_print(FILE *fp, struct hist_entry *self)
 573 {
 574         if (self->dso)
 575                 return fprintf(fp, "%-25s", self->dso->name);
 576
 577         return fprintf(fp, "%016llx         ", (u64)self->ip);
 578 }
 579
 580 static struct sort_entry sort_dso = {
 581         .header = "Shared Object            ",
 582         .cmp    = sort__dso_cmp,
 583         .print  = sort__dso_print,
 584 };
 585
 586 /* --sort symbol */
 587
 588 static int64_t
 589 sort__sym_cmp(struct hist_entry *left, struct hist_entry *right)
 590 {
 591         u64 ip_l, ip_r;
 592
 593         if (left->sym == right->sym)
 594                 return 0;
 595
 596         ip_l = left->sym ? left->sym->start : left->ip;
 597         ip_r = right->sym ? right->sym->start : right->ip;
 598
 599         return (int64_t)(ip_r - ip_l);
 600 }
 601
 602 static size_t
 603 sort__sym_print(FILE *fp, struct hist_entry *self)
 604 {
 605         size_t ret = 0;
 606
 607         if (verbose)
 608                 ret += fprintf(fp, "%#018llx  ", (u64)self->ip);
 609
 610         if (self->sym) {
 611                 ret += fprintf(fp, "[%c] %s",
 612                         self->dso == kernel_dso ? 'k' : '.', self->sym->name);
 613         } else {
 614                 ret += fprintf(fp, "%#016llx", (u64)self->ip);
 615         }
 616
 617         return ret;
 618 }
 619
 620 static struct sort_entry sort_sym = {
 621         .header = "Symbol",
 622         .cmp    = sort__sym_cmp,
 623         .print  = sort__sym_print,
 624 };
 625
 626 /* --sort parent */
 627
 628 static int64_t
 629 sort__parent_cmp(struct hist_entry *left, struct hist_entry *right)
 630 {
 631         struct symbol *sym_l = left->parent;
 632         struct symbol *sym_r = right->parent;
 633
 634         if (!sym_l || !sym_r)
 635                 return cmp_null(sym_l, sym_r);
 636
 637         return strcmp(sym_l->name, sym_r->name);
 638 }
 639
 640 static size_t
 641 sort__parent_print(FILE *fp, struct hist_entry *self)
 642 {
 643         size_t ret = 0;
 644
 645         ret += fprintf(fp, "%-20s", self->parent ? self->parent->name : "[other]");
 646
 647         return ret;
 648 }
 649
 650 static struct sort_entry sort_parent = {
 651         .header = "Parent symbol       ",
 652         .cmp    = sort__parent_cmp,
 653         .print  = sort__parent_print,
 654 };
 655
 656 static int sort__need_collapse = 0;
 657 static int sort__has_parent = 0;
 658
 659 struct sort_dimension {
 660         char                    *name;
 661         struct sort_entry       *entry;
 662         int                     taken;
 663 };
 664
 665 static struct sort_dimension sort_dimensions[] = {
 666         { .name = "pid",        .entry = &sort_thread,  },
 667         { .name = "comm",       .entry = &sort_comm,    },
 668         { .name = "dso",        .entry = &sort_dso,     },
 669         { .name = "symbol",     .entry = &sort_sym,     },
 670         { .name = "parent",     .entry = &sort_parent,  },
 671 };
 672
 673 static LIST_HEAD(hist_entry__sort_list);
 674
 675 static int sort_dimension__add(char *tok)
 676 {
 677         int i;
 678
 679         for (i = 0; i < ARRAY_SIZE(sort_dimensions); i++) {
 680                 struct sort_dimension *sd = &sort_dimensions[i];
 681
 682                 if (sd->taken)
 683                         continue;
 684
 685                 if (strncasecmp(tok, sd->name, strlen(tok)))
 686                         continue;
 687
 688                 if (sd->entry->collapse)
 689                         sort__need_collapse = 1;
 690
 691                 if (sd->entry == &sort_parent) {
 692                         int ret = regcomp(&parent_regex, parent_pattern, REG_EXTENDED);
 693                         if (ret) {
 694                                 char err[BUFSIZ];
 695
 696                                 regerror(ret, &parent_regex, err, sizeof(err));
 697                                 fprintf(stderr, "Invalid regex: %s\n%s",
 698                                         parent_pattern, err);
 699                                 exit(-1);
 700                         }
 701                         sort__has_parent = 1;
 702                 }
 703
 704                 list_add_tail(&sd->entry->list, &hist_entry__sort_list);
 705                 sd->taken = 1;
 706
 707                 return 0;
 708         }
 709
 710         return -ESRCH;
 711 }
 712
 713 static int64_t
 714 hist_entry__cmp(struct hist_entry *left, struct hist_entry *right)
 715 {
 716         struct sort_entry *se;
 717         int64_t cmp = 0;
 718
 719         list_for_each_entry(se, &hist_entry__sort_list, list) {
 720                 cmp = se->cmp(left, right);
 721                 if (cmp)
 722                         break;
 723         }
 724
 725         return cmp;
 726 }
 727
 728 static int64_t
 729 hist_entry__collapse(struct hist_entry *left, struct hist_entry *right)
 730 {
 731         struct sort_entry *se;
 732         int64_t cmp = 0;
 733
 734         list_for_each_entry(se, &hist_entry__sort_list, list) {
 735                 int64_t (*f)(struct hist_entry *, struct hist_entry *);
 736
 737                 f = se->collapse ?: se->cmp;
 738
 739                 cmp = f(left, right);
 740                 if (cmp)
 741                         break;
 742         }
 743
 744         return cmp;
 745 }
 746
 747 static size_t
 748 hist_entry__fprintf(FILE *fp, struct hist_entry *self, u64 total_samples)
 749 {
 750         struct sort_entry *se;
 751         size_t ret;
 752
 753         if (exclude_other && !self->parent)
 754                 return 0;
 755
 756         if (total_samples) {
 757                 double percent = self->count * 100.0 / total_samples;
 758                 char *color = PERF_COLOR_NORMAL;
 759
 760                 /*
 761                  * We color high-overhead entries in red, mid-overhead
 762                  * entries in green - and keep the low overhead places
 763                  * normal:
 764                  */
 765                 if (percent >= 5.0) {
 766                         color = PERF_COLOR_RED;
 767                 } else {
 768                         if (percent >= 0.5)
 769                                 color = PERF_COLOR_GREEN;
 770                 }
 771
 772                 ret = color_fprintf(fp, color, "   %6.2f%%",
 773                                 (self->count * 100.0) / total_samples);
 774         } else
 775                 ret = fprintf(fp, "%12Ld ", self->count);
 776
 777         list_for_each_entry(se, &hist_entry__sort_list, list) {
 778                 if (exclude_other && (se == &sort_parent))
 779                         continue;
 780
 781                 fprintf(fp, "  ");
 782                 ret += se->print(fp, self);
 783         }
 784
 785         ret += fprintf(fp, "\n");
 786
 787         return ret;
 788 }
 789
 790 /*
 791  *
 792  */
 793
 794 static struct symbol *
 795 resolve_symbol(struct thread *thread, struct map **mapp,
 796                struct dso **dsop, u64 *ipp)
 797 {
 798         struct dso *dso = dsop ? *dsop : NULL;
 799         struct map *map = mapp ? *mapp : NULL;
 800         u64 ip = *ipp;
 801
 802         if (!thread)
 803                 return NULL;
 804
 805         if (dso)
 806                 goto got_dso;
 807
 808         if (map)
 809                 goto got_map;
 810
 811         map = thread__find_map(thread, ip);
 812         if (map != NULL) {
 813                 if (mapp)
 814                         *mapp = map;
 815 got_map:
 816                 ip = map->map_ip(map, ip);
 817
 818                 dso = map->dso;
 819         } else {
 820                 /*
 821                  * If this is outside of all known maps,
 822                  * and is a negative address, try to look it
 823                  * up in the kernel dso, as it might be a
 824                  * vsyscall (which executes in user-mode):
 825                  */
 826                 if ((long long)ip < 0)
 827                 dso = kernel_dso;
 828         }
 829         dprintf(" ...... dso: %s\n", dso ? dso->name : "<not found>");
 830         dprintf(" ...... map: %Lx -> %Lx\n", *ipp, ip);
 831         *ipp  = ip;
 832
 833         if (dsop)
 834                 *dsop = dso;
 835
 836         if (!dso)
 837                 return NULL;
 838 got_dso:
 839         return dso->find_symbol(dso, ip);
 840 }
 841
 842 static int call__match(struct symbol *sym)
 843 {
 844         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
 845                 return 1;
 846
 847         return 0;
 848 }
 849
 850 /*
 851  * collect histogram counts
 852  */
 853
 854 static int
 855 hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
 856                 struct symbol *sym, u64 ip, struct ip_callchain *chain,
 857                 char level, u64 count)
 858 {
 859         struct rb_node **p = &hist.rb_node;
 860         struct rb_node *parent = NULL;
 861         struct hist_entry *he;
 862         struct hist_entry entry = {
 863                 .thread = thread,
 864                 .map    = map,
 865                 .dso    = dso,
 866                 .sym    = sym,
 867                 .ip     = ip,
 868                 .level  = level,
 869                 .count  = count,
 870                 .parent = NULL,
 871         };
 872         int cmp;
 873
 874         if (sort__has_parent && chain) {
 875                 u64 context = PERF_CONTEXT_MAX;
 876                 int i;
 877
 878                 for (i = 0; i < chain->nr; i++) {
 879                         u64 ip = chain->ips[i];
 880                         struct dso *dso = NULL;
 881                         struct symbol *sym;
 882
 883                         if (ip >= PERF_CONTEXT_MAX) {
 884                                 context = ip;
 885                                 continue;
 886                         }
 887
 888                         switch (context) {
 889                         case PERF_CONTEXT_KERNEL:
 890                                 dso = kernel_dso;
 891                                 break;
 892                         default:
 893                                 break;
 894                         }
 895
 896                         sym = resolve_symbol(thread, NULL, &dso, &ip);
 897
 898                         if (sym && call__match(sym)) {
 899                                 entry.parent = sym;
 900                                 break;
 901                         }
 902                 }
 903         }
 904
 905         while (*p != NULL) {
 906                 parent = *p;
 907                 he = rb_entry(parent, struct hist_entry, rb_node);
 908
 909                 cmp = hist_entry__cmp(&entry, he);
 910
 911                 if (!cmp) {
 912                         he->count += count;
 913                         return 0;
 914                 }
 915
 916                 if (cmp < 0)
 917                         p = &(*p)->rb_left;
 918                 else
 919                         p = &(*p)->rb_right;
 920         }
 921
 922         he = malloc(sizeof(*he));
 923         if (!he)
 924                 return -ENOMEM;
 925         *he = entry;
 926         rb_link_node(&he->rb_node, parent, p);
 927         rb_insert_color(&he->rb_node, &hist);
 928
 929         return 0;
 930 }
 931
 932 static void hist_entry__free(struct hist_entry *he)
 933 {
 934         free(he);
 935 }
 936
 937 /*
 938  * collapse the histogram
 939  */
 940
 941 static struct rb_root collapse_hists;
 942
 943 static void collapse__insert_entry(struct hist_entry *he)
 944 {
 945         struct rb_node **p = &collapse_hists.rb_node;
 946         struct rb_node *parent = NULL;
 947         struct hist_entry *iter;
 948         int64_t cmp;
 949
 950         while (*p != NULL) {
 951                 parent = *p;
 952                 iter = rb_entry(parent, struct hist_entry, rb_node);
 953
 954                 cmp = hist_entry__collapse(iter, he);
 955
 956                 if (!cmp) {
 957                         iter->count += he->count;
 958                         hist_entry__free(he);
 959                         return;
 960                 }
 961
 962                 if (cmp < 0)
 963                         p = &(*p)->rb_left;
 964                 else
 965                         p = &(*p)->rb_right;
 966         }
 967
 968         rb_link_node(&he->rb_node, parent, p);
 969         rb_insert_color(&he->rb_node, &collapse_hists);
 970 }
 971
 972 static void collapse__resort(void)
 973 {
 974         struct rb_node *next;
 975         struct hist_entry *n;
 976
 977         if (!sort__need_collapse)
 978                 return;
 979
 980         next = rb_first(&hist);
 981         while (next) {
 982                 n = rb_entry(next, struct hist_entry, rb_node);
 983                 next = rb_next(&n->rb_node);
 984
 985                 rb_erase(&n->rb_node, &hist);
 986                 collapse__insert_entry(n);
 987         }
 988 }
 989
 990 /*
 991  * reverse the map, sort on count.
 992  */
 993
 994 static struct rb_root output_hists;
 995
 996 static void output__insert_entry(struct hist_entry *he)
 997 {
 998         struct rb_node **p = &output_hists.rb_node;
 999         struct rb_node *parent = NULL;
1000         struct hist_entry *iter;
1001
1002         while (*p != NULL) {
1003                 parent = *p;
1004                 iter = rb_entry(parent, struct hist_entry, rb_node);
1005
1006                 if (he->count > iter->count)
1007                         p = &(*p)->rb_left;
1008                 else
1009                         p = &(*p)->rb_right;
1010         }
1011
1012         rb_link_node(&he->rb_node, parent, p);
1013         rb_insert_color(&he->rb_node, &output_hists);
1014 }
1015
1016 static void output__resort(void)
1017 {
1018         struct rb_node *next;
1019         struct hist_entry *n;
1020         struct rb_root *tree = &hist;
1021
1022         if (sort__need_collapse)
1023                 tree = &collapse_hists;
1024
1025         next = rb_first(tree);
1026
1027         while (next) {
1028                 n = rb_entry(next, struct hist_entry, rb_node);
1029                 next = rb_next(&n->rb_node);
1030
1031                 rb_erase(&n->rb_node, tree);
1032                 output__insert_entry(n);
1033         }
1034 }
1035
1036 static size_t output__fprintf(FILE *fp, u64 total_samples)
1037 {
1038         struct hist_entry *pos;
1039         struct sort_entry *se;
1040         struct rb_node *nd;
1041         size_t ret = 0;
1042
1043         fprintf(fp, "\n");
1044         fprintf(fp, "#\n");
1045         fprintf(fp, "# (%Ld samples)\n", (u64)total_samples);
1046         fprintf(fp, "#\n");
1047
1048         fprintf(fp, "# Overhead");
1049         list_for_each_entry(se, &hist_entry__sort_list, list) {
1050                 if (exclude_other && (se == &sort_parent))
1051                         continue;
1052                 fprintf(fp, "  %s", se->header);
1053         }
1054         fprintf(fp, "\n");
1055
1056         fprintf(fp, "# ........");
1057         list_for_each_entry(se, &hist_entry__sort_list, list) {
1058                 int i;
1059
1060                 if (exclude_other && (se == &sort_parent))
1061                         continue;
1062
1063                 fprintf(fp, "  ");
1064                 for (i = 0; i < strlen(se->header); i++)
1065                         fprintf(fp, ".");
1066         }
1067         fprintf(fp, "\n");
1068
1069         fprintf(fp, "#\n");
1070
1071         for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
1072                 pos = rb_entry(nd, struct hist_entry, rb_node);
1073                 ret += hist_entry__fprintf(fp, pos, total_samples);
1074         }
1075
1076         if (sort_order == default_sort_order &&
1077                         parent_pattern == default_parent_pattern) {
1078                 fprintf(fp, "#\n");
1079                 fprintf(fp, "# (For more details, try: perf report --sort comm,dso,symbol)\n");
1080                 fprintf(fp, "#\n");
1081         }
1082         fprintf(fp, "\n");
1083
1084         return ret;
1085 }
1086
1087 static void register_idle_thread(void)
1088 {
1089         struct thread *thread = threads__findnew(0);
1090
1091         if (thread == NULL ||
1092                         thread__set_comm(thread, "[idle]")) {
1093                 fprintf(stderr, "problem inserting idle task.\n");
1094                 exit(-1);
1095         }
1096 }
1097
1098 static unsigned long total = 0,
1099                      total_mmap = 0,
1100                      total_comm = 0,
1101                      total_fork = 0,
1102                      total_unknown = 0,
1103                      total_lost = 0;
1104
1105 static int validate_chain(struct ip_callchain *chain, event_t *event)
1106 {
1107         unsigned int chain_size;
1108
1109         chain_size = event->header.size;
1110         chain_size -= (unsigned long)&event->ip.__more_data - (unsigned long)event;
1111
1112         if (chain->nr*sizeof(u64) > chain_size)
1113                 return -1;
1114
1115         return 0;
1116 }
1117
1118 static int
1119 process_overflow_event(event_t *event, unsigned long offset, unsigned long head)
1120 {
1121         char level;
1122         int show = 0;
1123         struct dso *dso = NULL;
1124         struct thread *thread = threads__findnew(event->ip.pid);
1125         u64 ip = event->ip.ip;
1126         u64 period = 1;
1127         struct map *map = NULL;
1128         void *more_data = event->ip.__more_data;
1129         struct ip_callchain *chain = NULL;
1130
1131         if (event->header.type & PERF_SAMPLE_PERIOD) {
1132                 period = *(u64 *)more_data;
1133                 more_data += sizeof(u64);
1134         }
1135
1136         dprintf("%p [%p]: PERF_EVENT (IP, %d): %d: %p period: %Ld\n",
1137                 (void *)(offset + head),
1138                 (void *)(long)(event->header.size),
1139                 event->header.misc,
1140                 event->ip.pid,
1141                 (void *)(long)ip,
1142                 (long long)period);
1143
1144         if (event->header.type & PERF_SAMPLE_CALLCHAIN) {
1145                 int i;
1146
1147                 chain = (void *)more_data;
1148
1149                 dprintf("... chain: nr:%Lu\n", chain->nr);
1150
1151                 if (validate_chain(chain, event) < 0) {
1152                         eprintf("call-chain problem with event, skipping it.\n");
1153                         return 0;
1154                 }
1155
1156                 if (dump_trace) {
1157                         for (i = 0; i < chain->nr; i++)
1158                                 dprintf("..... %2d: %016Lx\n", i, chain->ips[i]);
1159                 }
1160         }
1161
1162         dprintf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1163
1164         if (thread == NULL) {
1165                 eprintf("problem processing %d event, skipping it.\n",
1166                         event->header.type);
1167                 return -1;
1168         }
1169
1170         if (event->header.misc & PERF_EVENT_MISC_KERNEL) {
1171                 show = SHOW_KERNEL;
1172                 level = 'k';
1173
1174                 dso = kernel_dso;
1175
1176                 dprintf(" ...... dso: %s\n", dso->name);
1177
1178         } else if (event->header.misc & PERF_EVENT_MISC_USER) {
1179
1180                 show = SHOW_USER;
1181                 level = '.';
1182
1183         } else {
1184                 show = SHOW_HV;
1185                 level = 'H';
1186                 dprintf(" ...... dso: [hypervisor]\n");
1187         }
1188
1189         if (show & show_mask) {
1190                 struct symbol *sym = resolve_symbol(thread, &map, &dso, &ip);
1191
1192                 if (hist_entry__add(thread, map, dso, sym, ip, chain, level, period)) {
1193                         eprintf("problem incrementing symbol count, skipping event\n");
1194                         return -1;
1195                 }
1196         }
1197         total += period;
1198
1199         return 0;
1200 }
1201
1202 static int
1203 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
1204 {
1205         struct thread *thread = threads__findnew(event->mmap.pid);
1206         struct map *map = map__new(&event->mmap);
1207
1208         dprintf("%p [%p]: PERF_EVENT_MMAP %d: [%p(%p) @ %p]: %s\n",
1209                 (void *)(offset + head),
1210                 (void *)(long)(event->header.size),
1211                 event->mmap.pid,
1212                 (void *)(long)event->mmap.start,
1213                 (void *)(long)event->mmap.len,
1214                 (void *)(long)event->mmap.pgoff,
1215                 event->mmap.filename);
1216
1217         if (thread == NULL || map == NULL) {
1218                 dprintf("problem processing PERF_EVENT_MMAP, skipping event.\n");
1219                 return 0;
1220         }
1221
1222         thread__insert_map(thread, map);
1223         total_mmap++;
1224
1225         return 0;
1226 }
1227
1228 static int
1229 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
1230 {
1231         struct thread *thread = threads__findnew(event->comm.pid);
1232
1233         dprintf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
1234                 (void *)(offset + head),
1235                 (void *)(long)(event->header.size),
1236                 event->comm.comm, event->comm.pid);
1237
1238         if (thread == NULL ||
1239             thread__set_comm(thread, event->comm.comm)) {
1240                 dprintf("problem processing PERF_EVENT_COMM, skipping event.\n");
1241                 return -1;
1242         }
1243         total_comm++;
1244
1245         return 0;
1246 }
1247
1248 static int
1249 process_fork_event(event_t *event, unsigned long offset, unsigned long head)
1250 {
1251         struct thread *thread = threads__findnew(event->fork.pid);
1252         struct thread *parent = threads__findnew(event->fork.ppid);
1253
1254         dprintf("%p [%p]: PERF_EVENT_FORK: %d:%d\n",
1255                 (void *)(offset + head),
1256                 (void *)(long)(event->header.size),
1257                 event->fork.pid, event->fork.ppid);
1258
1259         if (!thread || !parent || thread__fork(thread, parent)) {
1260                 dprintf("problem processing PERF_EVENT_FORK, skipping event.\n");
1261                 return -1;
1262         }
1263         total_fork++;
1264
1265         return 0;
1266 }
1267
1268 static int
1269 process_period_event(event_t *event, unsigned long offset, unsigned long head)
1270 {
1271         dprintf("%p [%p]: PERF_EVENT_PERIOD: time:%Ld, id:%Ld: period:%Ld\n",
1272                 (void *)(offset + head),
1273                 (void *)(long)(event->header.size),
1274                 event->period.time,
1275                 event->period.id,
1276                 event->period.sample_period);
1277
1278         return 0;
1279 }
1280
1281 static int
1282 process_lost_event(event_t *event, unsigned long offset, unsigned long head)
1283 {
1284         dprintf("%p [%p]: PERF_EVENT_LOST: id:%Ld: lost:%Ld\n",
1285                 (void *)(offset + head),
1286                 (void *)(long)(event->header.size),
1287                 event->lost.id,
1288                 event->lost.lost);
1289
1290         total_lost += event->lost.lost;
1291
1292         return 0;
1293 }
1294
1295 static void trace_event(event_t *event)
1296 {
1297         unsigned char *raw_event = (void *)event;
1298         char *color = PERF_COLOR_BLUE;
1299         int i, j;
1300
1301         if (!dump_trace)
1302                 return;
1303
1304         dprintf(".");
1305         cdprintf("\n. ... raw event: size %d bytes\n", event->header.size);
1306
1307         for (i = 0; i < event->header.size; i++) {
1308                 if ((i & 15) == 0) {
1309                         dprintf(".");
1310                         cdprintf("  %04x: ", i);
1311                 }
1312
1313                 cdprintf(" %02x", raw_event[i]);
1314
1315                 if (((i & 15) == 15) || i == event->header.size-1) {
1316                         cdprintf("  ");
1317                         for (j = 0; j < 15-(i & 15); j++)
1318                                 cdprintf("   ");
1319                         for (j = 0; j < (i & 15); j++) {
1320                                 if (isprint(raw_event[i-15+j]))
1321                                         cdprintf("%c", raw_event[i-15+j]);
1322                                 else
1323                                         cdprintf(".");
1324                         }
1325                         cdprintf("\n");
1326                 }
1327         }
1328         dprintf(".\n");
1329 }
1330
1331 static int
1332 process_event(event_t *event, unsigned long offset, unsigned long head)
1333 {
1334         trace_event(event);
1335
1336         if (event->header.misc & PERF_EVENT_MISC_OVERFLOW)
1337                 return process_overflow_event(event, offset, head);
1338
1339         switch (event->header.type) {
1340         case PERF_EVENT_MMAP:
1341                 return process_mmap_event(event, offset, head);
1342
1343         case PERF_EVENT_COMM:
1344                 return process_comm_event(event, offset, head);
1345
1346         case PERF_EVENT_FORK:
1347                 return process_fork_event(event, offset, head);
1348
1349         case PERF_EVENT_PERIOD:
1350                 return process_period_event(event, offset, head);
1351
1352         case PERF_EVENT_LOST:
1353                 return process_lost_event(event, offset, head);
1354
1355         /*
1356          * We dont process them right now but they are fine:
1357          */
1358
1359         case PERF_EVENT_THROTTLE:
1360         case PERF_EVENT_UNTHROTTLE:
1361                 return 0;
1362
1363         default:
1364                 return -1;
1365         }
1366
1367         return 0;
1368 }
1369
1370 static struct perf_file_header          file_header;
1371
1372 static int __cmd_report(void)
1373 {
1374         int ret, rc = EXIT_FAILURE;
1375         unsigned long offset = 0;
1376         unsigned long head = sizeof(file_header);
1377         struct stat stat;
1378         event_t *event;
1379         uint32_t size;
1380         char *buf;
1381
1382         register_idle_thread();
1383
1384         input = open(input_name, O_RDONLY);
1385         if (input < 0) {
1386                 fprintf(stderr, " failed to open file: %s", input_name);
1387                 if (!strcmp(input_name, "perf.data"))
1388                         fprintf(stderr, "  (try 'perf record' first)");
1389                 fprintf(stderr, "\n");
1390                 exit(-1);
1391         }
1392
1393         ret = fstat(input, &stat);
1394         if (ret < 0) {
1395                 perror("failed to stat file");
1396                 exit(-1);
1397         }
1398
1399         if (!stat.st_size) {
1400                 fprintf(stderr, "zero-sized file, nothing to do!\n");
1401                 exit(0);
1402         }
1403
1404         if (read(input, &file_header, sizeof(file_header)) == -1) {
1405                 perror("failed to read file headers");
1406                 exit(-1);
1407         }
1408
1409         if (sort__has_parent &&
1410             !(file_header.sample_type & PERF_SAMPLE_CALLCHAIN)) {
1411                 fprintf(stderr, "selected --sort parent, but no callchain data\n");
1412                 exit(-1);
1413         }
1414
1415         if (load_kernel() < 0) {
1416                 perror("failed to load kernel symbols");
1417                 return EXIT_FAILURE;
1418         }
1419
1420         if (!full_paths) {
1421                 if (getcwd(__cwd, sizeof(__cwd)) == NULL) {
1422                         perror("failed to get the current directory");
1423                         return EXIT_FAILURE;
1424                 }
1425                 cwdlen = strlen(cwd);
1426         } else {
1427                 cwd = NULL;
1428                 cwdlen = 0;
1429         }
1430 remap:
1431         buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1432                            MAP_SHARED, input, offset);
1433         if (buf == MAP_FAILED) {
1434                 perror("failed to mmap file");
1435                 exit(-1);
1436         }
1437
1438 more:
1439         event = (event_t *)(buf + head);
1440
1441         size = event->header.size;
1442         if (!size)
1443                 size = 8;
1444
1445         if (head + event->header.size >= page_size * mmap_window) {
1446                 unsigned long shift = page_size * (head / page_size);
1447                 int ret;
1448
1449                 ret = munmap(buf, page_size * mmap_window);
1450                 assert(ret == 0);
1451
1452                 offset += shift;
1453                 head -= shift;
1454                 goto remap;
1455         }
1456
1457         size = event->header.size;
1458
1459         dprintf("\n%p [%p]: event: %d\n",
1460                         (void *)(offset + head),
1461                         (void *)(long)event->header.size,
1462                         event->header.type);
1463
1464         if (!size || process_event(event, offset, head) < 0) {
1465
1466                 dprintf("%p [%p]: skipping unknown header type: %d\n",
1467                         (void *)(offset + head),
1468                         (void *)(long)(event->header.size),
1469                         event->header.type);
1470
1471                 total_unknown++;
1472
1473                 /*
1474                  * assume we lost track of the stream, check alignment, and
1475                  * increment a single u64 in the hope to catch on again 'soon'.
1476                  */
1477
1478                 if (unlikely(head & 7))
1479                         head &= ~7ULL;
1480
1481                 size = 8;
1482         }
1483
1484         head += size;
1485
1486         if (offset + head >= sizeof(file_header) + file_header.data_size)
1487                 goto done;
1488
1489         if (offset + head < stat.st_size)
1490                 goto more;
1491
1492 done:
1493         rc = EXIT_SUCCESS;
1494         close(input);
1495
1496         dprintf("      IP events: %10ld\n", total);
1497         dprintf("    mmap events: %10ld\n", total_mmap);
1498         dprintf("    comm events: %10ld\n", total_comm);
1499         dprintf("    fork events: %10ld\n", total_fork);
1500         dprintf("    lost events: %10ld\n", total_lost);
1501         dprintf(" unknown events: %10ld\n", total_unknown);
1502
1503         if (dump_trace)
1504                 return 0;
1505
1506         if (verbose >= 3)
1507                 threads__fprintf(stdout);
1508
1509         if (verbose >= 2)
1510                 dsos__fprintf(stdout);
1511
1512         collapse__resort();
1513         output__resort();
1514         output__fprintf(stdout, total);
1515
1516         return rc;
1517 }
1518
1519 static const char * const report_usage[] = {
1520         "perf report [<options>] <command>",
1521         NULL
1522 };
1523
1524 static const struct option options[] = {
1525         OPT_STRING('i', "input", &input_name, "file",
1526                     "input file name"),
1527         OPT_BOOLEAN('v', "verbose", &verbose,
1528                     "be more verbose (show symbol address, etc)"),
1529         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1530                     "dump raw trace in ASCII"),
1531         OPT_STRING('k', "vmlinux", &vmlinux, "file", "vmlinux pathname"),
1532         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1533                    "sort by key(s): pid, comm, dso, symbol, parent"),
1534         OPT_BOOLEAN('P', "full-paths", &full_paths,
1535                     "Don't shorten the pathnames taking into account the cwd"),
1536         OPT_STRING('p', "parent", &parent_pattern, "regex",
1537                    "regex filter to identify parent, see: '--sort parent'"),
1538         OPT_BOOLEAN('x', "exclude-other", &exclude_other,
1539                     "Only display entries with parent-match"),
1540         OPT_END()
1541 };
1542
1543 static void setup_sorting(void)
1544 {
1545         char *tmp, *tok, *str = strdup(sort_order);
1546
1547         for (tok = strtok_r(str, ", ", &tmp);
1548                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1549                 if (sort_dimension__add(tok) < 0) {
1550                         error("Unknown --sort key: `%s'", tok);
1551                         usage_with_options(report_usage, options);
1552                 }
1553         }
1554
1555         free(str);
1556 }
1557
1558 int cmd_report(int argc, const char **argv, const char *prefix)
1559 {
1560         symbol__init();
1561
1562         page_size = getpagesize();
1563
1564         argc = parse_options(argc, argv, options, report_usage, 0);
1565
1566         setup_sorting();
1567
1568         if (parent_pattern != default_parent_pattern)
1569                 sort_dimension__add("parent");
1570         else
1571                 exclude_other = 0;
1572
1573         /*
1574          * Any (unrecognized) arguments left?
1575          */
1576         if (argc)
1577                 usage_with_options(report_usage, options);
1578
1579         setup_pager();
1580
1581         return __cmd_report();
1582 }