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             *parent_pattern = "^sys_|^do_page_fault";
  50 static regex_t          parent_regex;
  51
  52 struct ip_event {
  53         struct perf_event_header header;
  54         __u64 ip;
  55         __u32 pid, tid;
  56         unsigned char __more_data[];
  57 };
  58
  59 struct mmap_event {
  60         struct perf_event_header header;
  61         __u32 pid, tid;
  62         __u64 start;
  63         __u64 len;
  64         __u64 pgoff;
  65         char filename[PATH_MAX];
  66 };
  67
  68 struct comm_event {
  69         struct perf_event_header header;
  70         __u32 pid, tid;
  71         char comm[16];
  72 };
  73
  74 struct fork_event {
  75         struct perf_event_header header;
  76         __u32 pid, ppid;
  77 };
  78
  79 struct period_event {
  80         struct perf_event_header header;
  81         __u64 time;
  82         __u64 id;
  83         __u64 sample_period;
  84 };
  85
  86 struct lost_event {
  87         struct perf_event_header header;
  88         __u64 id;
  89         __u64 lost;
  90 };
  91
  92 typedef union event_union {
  93         struct perf_event_header        header;
  94         struct ip_event                 ip;
  95         struct mmap_event               mmap;
  96         struct comm_event               comm;
  97         struct fork_event               fork;
  98         struct period_event             period;
  99         struct lost_event               lost;
 100 } event_t;
 101
 102 static LIST_HEAD(dsos);
 103 static struct dso *kernel_dso;
 104 static struct dso *vdso;
 105
 106 static void dsos__add(struct dso *dso)
 107 {
 108         list_add_tail(&dso->node, &dsos);
 109 }
 110
 111 static struct dso *dsos__find(const char *name)
 112 {
 113         struct dso *pos;
 114
 115         list_for_each_entry(pos, &dsos, node)
 116                 if (strcmp(pos->name, name) == 0)
 117                         return pos;
 118         return NULL;
 119 }
 120
 121 static struct dso *dsos__findnew(const char *name)
 122 {
 123         struct dso *dso = dsos__find(name);
 124         int nr;
 125
 126         if (dso)
 127                 return dso;
 128
 129         dso = dso__new(name, 0);
 130         if (!dso)
 131                 goto out_delete_dso;
 132
 133         nr = dso__load(dso, NULL, verbose);
 134         if (nr < 0) {
 135                 eprintf("Failed to open: %s\n", name);
 136                 goto out_delete_dso;
 137         }
 138         if (!nr)
 139                 eprintf("No symbols found in: %s, maybe install a debug package?\n", name);
 140
 141         dsos__add(dso);
 142
 143         return dso;
 144
 145 out_delete_dso:
 146         dso__delete(dso);
 147         return NULL;
 148 }
 149
 150 static void dsos__fprintf(FILE *fp)
 151 {
 152         struct dso *pos;
 153
 154         list_for_each_entry(pos, &dsos, node)
 155                 dso__fprintf(pos, fp);
 156 }
 157
 158 static struct symbol *vdso__find_symbol(struct dso *dso, __u64 ip)
 159 {
 160         return dso__find_symbol(kernel_dso, ip);
 161 }
 162
 163 static int load_kernel(void)
 164 {
 165         int err;
 166
 167         kernel_dso = dso__new("[kernel]", 0);
 168         if (!kernel_dso)
 169                 return -1;
 170
 171         err = dso__load_kernel(kernel_dso, vmlinux, NULL, verbose);
 172         if (err) {
 173                 dso__delete(kernel_dso);
 174                 kernel_dso = NULL;
 175         } else
 176                 dsos__add(kernel_dso);
 177
 178         vdso = dso__new("[vdso]", 0);
 179         if (!vdso)
 180                 return -1;
 181
 182         vdso->find_symbol = vdso__find_symbol;
 183
 184         dsos__add(vdso);
 185
 186         return err;
 187 }
 188
 189 static char __cwd[PATH_MAX];
 190 static char *cwd = __cwd;
 191 static int cwdlen;
 192
 193 static int strcommon(const char *pathname)
 194 {
 195         int n = 0;
 196
 197         while (pathname[n] == cwd[n] && n < cwdlen)
 198                 ++n;
 199
 200         return n;
 201 }
 202
 203 struct map {
 204         struct list_head node;
 205         __u64    start;
 206         __u64    end;
 207         __u64    pgoff;
 208         __u64    (*map_ip)(struct map *, __u64);
 209         struct dso       *dso;
 210 };
 211
 212 static __u64 map__map_ip(struct map *map, __u64 ip)
 213 {
 214         return ip - map->start + map->pgoff;
 215 }
 216
 217 static __u64 vdso__map_ip(struct map *map, __u64 ip)
 218 {
 219         return ip;
 220 }
 221
 222 static inline int is_anon_memory(const char *filename)
 223 {
 224      return strcmp(filename, "//anon") == 0;
 225 }
 226
 227 static struct map *map__new(struct mmap_event *event)
 228 {
 229         struct map *self = malloc(sizeof(*self));
 230
 231         if (self != NULL) {
 232                 const char *filename = event->filename;
 233                 char newfilename[PATH_MAX];
 234                 int anon;
 235
 236                 if (cwd) {
 237                         int n = strcommon(filename);
 238
 239                         if (n == cwdlen) {
 240                                 snprintf(newfilename, sizeof(newfilename),
 241                                          ".%s", filename + n);
 242                                 filename = newfilename;
 243                         }
 244                 }
 245
 246                 anon = is_anon_memory(filename);
 247
 248                 if (anon) {
 249                         snprintf(newfilename, sizeof(newfilename), "/tmp/perf-%d.map", event->pid);
 250                         filename = newfilename;
 251                 }
 252
 253                 self->start = event->start;
 254                 self->end   = event->start + event->len;
 255                 self->pgoff = event->pgoff;
 256
 257                 self->dso = dsos__findnew(filename);
 258                 if (self->dso == NULL)
 259                         goto out_delete;
 260
 261                 if (self->dso == vdso || anon)
 262                         self->map_ip = vdso__map_ip;
 263                 else
 264                         self->map_ip = map__map_ip;
 265         }
 266         return self;
 267 out_delete:
 268         free(self);
 269         return NULL;
 270 }
 271
 272 static struct map *map__clone(struct map *self)
 273 {
 274         struct map *map = malloc(sizeof(*self));
 275
 276         if (!map)
 277                 return NULL;
 278
 279         memcpy(map, self, sizeof(*self));
 280
 281         return map;
 282 }
 283
 284 static int map__overlap(struct map *l, struct map *r)
 285 {
 286         if (l->start > r->start) {
 287                 struct map *t = l;
 288                 l = r;
 289                 r = t;
 290         }
 291
 292         if (l->end > r->start)
 293                 return 1;
 294
 295         return 0;
 296 }
 297
 298 static size_t map__fprintf(struct map *self, FILE *fp)
 299 {
 300         return fprintf(fp, " %Lx-%Lx %Lx %s\n",
 301                        self->start, self->end, self->pgoff, self->dso->name);
 302 }
 303
 304
 305 struct thread {
 306         struct rb_node   rb_node;
 307         struct list_head maps;
 308         pid_t            pid;
 309         char             *comm;
 310 };
 311
 312 static struct thread *thread__new(pid_t pid)
 313 {
 314         struct thread *self = malloc(sizeof(*self));
 315
 316         if (self != NULL) {
 317                 self->pid = pid;
 318                 self->comm = malloc(32);
 319                 if (self->comm)
 320                         snprintf(self->comm, 32, ":%d", self->pid);
 321                 INIT_LIST_HEAD(&self->maps);
 322         }
 323
 324         return self;
 325 }
 326
 327 static int thread__set_comm(struct thread *self, const char *comm)
 328 {
 329         if (self->comm)
 330                 free(self->comm);
 331         self->comm = strdup(comm);
 332         return self->comm ? 0 : -ENOMEM;
 333 }
 334
 335 static size_t thread__fprintf(struct thread *self, FILE *fp)
 336 {
 337         struct map *pos;
 338         size_t ret = fprintf(fp, "Thread %d %s\n", self->pid, self->comm);
 339
 340         list_for_each_entry(pos, &self->maps, node)
 341                 ret += map__fprintf(pos, fp);
 342
 343         return ret;
 344 }
 345
 346
 347 static struct rb_root threads;
 348 static struct thread *last_match;
 349
 350 static struct thread *threads__findnew(pid_t pid)
 351 {
 352         struct rb_node **p = &threads.rb_node;
 353         struct rb_node *parent = NULL;
 354         struct thread *th;
 355
 356         /*
 357          * Font-end cache - PID lookups come in blocks,
 358          * so most of the time we dont have to look up
 359          * the full rbtree:
 360          */
 361         if (last_match && last_match->pid == pid)
 362                 return last_match;
 363
 364         while (*p != NULL) {
 365                 parent = *p;
 366                 th = rb_entry(parent, struct thread, rb_node);
 367
 368                 if (th->pid == pid) {
 369                         last_match = th;
 370                         return th;
 371                 }
 372
 373                 if (pid < th->pid)
 374                         p = &(*p)->rb_left;
 375                 else
 376                         p = &(*p)->rb_right;
 377         }
 378
 379         th = thread__new(pid);
 380         if (th != NULL) {
 381                 rb_link_node(&th->rb_node, parent, p);
 382                 rb_insert_color(&th->rb_node, &threads);
 383                 last_match = th;
 384         }
 385
 386         return th;
 387 }
 388
 389 static void thread__insert_map(struct thread *self, struct map *map)
 390 {
 391         struct map *pos, *tmp;
 392
 393         list_for_each_entry_safe(pos, tmp, &self->maps, node) {
 394                 if (map__overlap(pos, map)) {
 395                         list_del_init(&pos->node);
 396                         /* XXX leaks dsos */
 397                         free(pos);
 398                 }
 399         }
 400
 401         list_add_tail(&map->node, &self->maps);
 402 }
 403
 404 static int thread__fork(struct thread *self, struct thread *parent)
 405 {
 406         struct map *map;
 407
 408         if (self->comm)
 409                 free(self->comm);
 410         self->comm = strdup(parent->comm);
 411         if (!self->comm)
 412                 return -ENOMEM;
 413
 414         list_for_each_entry(map, &parent->maps, node) {
 415                 struct map *new = map__clone(map);
 416                 if (!new)
 417                         return -ENOMEM;
 418                 thread__insert_map(self, new);
 419         }
 420
 421         return 0;
 422 }
 423
 424 static struct map *thread__find_map(struct thread *self, __u64 ip)
 425 {
 426         struct map *pos;
 427
 428         if (self == NULL)
 429                 return NULL;
 430
 431         list_for_each_entry(pos, &self->maps, node)
 432                 if (ip >= pos->start && ip <= pos->end)
 433                         return pos;
 434
 435         return NULL;
 436 }
 437
 438 static size_t threads__fprintf(FILE *fp)
 439 {
 440         size_t ret = 0;
 441         struct rb_node *nd;
 442
 443         for (nd = rb_first(&threads); nd; nd = rb_next(nd)) {
 444                 struct thread *pos = rb_entry(nd, struct thread, rb_node);
 445
 446                 ret += thread__fprintf(pos, fp);
 447         }
 448
 449         return ret;
 450 }
 451
 452 /*
 453  * histogram, sorted on item, collects counts
 454  */
 455
 456 static struct rb_root hist;
 457
 458 struct hist_entry {
 459         struct rb_node   rb_node;
 460
 461         struct thread    *thread;
 462         struct map       *map;
 463         struct dso       *dso;
 464         struct symbol    *sym;
 465         struct symbol    *parent;
 466         __u64            ip;
 467         char             level;
 468
 469         __u64            count;
 470 };
 471
 472 /*
 473  * configurable sorting bits
 474  */
 475
 476 struct sort_entry {
 477         struct list_head list;
 478
 479         char *header;
 480
 481         int64_t (*cmp)(struct hist_entry *, struct hist_entry *);
 482         int64_t (*collapse)(struct hist_entry *, struct hist_entry *);
 483         size_t  (*print)(FILE *fp, struct hist_entry *);
 484 };
 485
 486 static int64_t cmp_null(void *l, void *r)
 487 {
 488         if (!l && !r)
 489                 return 0;
 490         else if (!l)
 491                 return -1;
 492         else
 493                 return 1;
 494 }
 495
 496 /* --sort pid */
 497
 498 static int64_t
 499 sort__thread_cmp(struct hist_entry *left, struct hist_entry *right)
 500 {
 501         return right->thread->pid - left->thread->pid;
 502 }
 503
 504 static size_t
 505 sort__thread_print(FILE *fp, struct hist_entry *self)
 506 {
 507         return fprintf(fp, "%16s:%5d", self->thread->comm ?: "", self->thread->pid);
 508 }
 509
 510 static struct sort_entry sort_thread = {
 511         .header = "         Command:  Pid",
 512         .cmp    = sort__thread_cmp,
 513         .print  = sort__thread_print,
 514 };
 515
 516 /* --sort comm */
 517
 518 static int64_t
 519 sort__comm_cmp(struct hist_entry *left, struct hist_entry *right)
 520 {
 521         return right->thread->pid - left->thread->pid;
 522 }
 523
 524 static int64_t
 525 sort__comm_collapse(struct hist_entry *left, struct hist_entry *right)
 526 {
 527         char *comm_l = left->thread->comm;
 528         char *comm_r = right->thread->comm;
 529
 530         if (!comm_l || !comm_r)
 531                 return cmp_null(comm_l, comm_r);
 532
 533         return strcmp(comm_l, comm_r);
 534 }
 535
 536 static size_t
 537 sort__comm_print(FILE *fp, struct hist_entry *self)
 538 {
 539         return fprintf(fp, "%16s", self->thread->comm);
 540 }
 541
 542 static struct sort_entry sort_comm = {
 543         .header         = "         Command",
 544         .cmp            = sort__comm_cmp,
 545         .collapse       = sort__comm_collapse,
 546         .print          = sort__comm_print,
 547 };
 548
 549 /* --sort dso */
 550
 551 static int64_t
 552 sort__dso_cmp(struct hist_entry *left, struct hist_entry *right)
 553 {
 554         struct dso *dso_l = left->dso;
 555         struct dso *dso_r = right->dso;
 556
 557         if (!dso_l || !dso_r)
 558                 return cmp_null(dso_l, dso_r);
 559
 560         return strcmp(dso_l->name, dso_r->name);
 561 }
 562
 563 static size_t
 564 sort__dso_print(FILE *fp, struct hist_entry *self)
 565 {
 566         if (self->dso)
 567                 return fprintf(fp, "%-25s", self->dso->name);
 568
 569         return fprintf(fp, "%016llx         ", (__u64)self->ip);
 570 }
 571
 572 static struct sort_entry sort_dso = {
 573         .header = "Shared Object            ",
 574         .cmp    = sort__dso_cmp,
 575         .print  = sort__dso_print,
 576 };
 577
 578 /* --sort symbol */
 579
 580 static int64_t
 581 sort__sym_cmp(struct hist_entry *left, struct hist_entry *right)
 582 {
 583         __u64 ip_l, ip_r;
 584
 585         if (left->sym == right->sym)
 586                 return 0;
 587
 588         ip_l = left->sym ? left->sym->start : left->ip;
 589         ip_r = right->sym ? right->sym->start : right->ip;
 590
 591         return (int64_t)(ip_r - ip_l);
 592 }
 593
 594 static size_t
 595 sort__sym_print(FILE *fp, struct hist_entry *self)
 596 {
 597         size_t ret = 0;
 598
 599         if (verbose)
 600                 ret += fprintf(fp, "%#018llx  ", (__u64)self->ip);
 601
 602         if (self->sym) {
 603                 ret += fprintf(fp, "[%c] %s",
 604                         self->dso == kernel_dso ? 'k' : '.', self->sym->name);
 605         } else {
 606                 ret += fprintf(fp, "%#016llx", (__u64)self->ip);
 607         }
 608
 609         return ret;
 610 }
 611
 612 static struct sort_entry sort_sym = {
 613         .header = "Symbol",
 614         .cmp    = sort__sym_cmp,
 615         .print  = sort__sym_print,
 616 };
 617
 618 /* --sort parent */
 619
 620 static int64_t
 621 sort__parent_cmp(struct hist_entry *left, struct hist_entry *right)
 622 {
 623         struct symbol *sym_l = left->parent;
 624         struct symbol *sym_r = right->parent;
 625
 626         if (!sym_l || !sym_r)
 627                 return cmp_null(sym_l, sym_r);
 628
 629         return strcmp(sym_l->name, sym_r->name);
 630 }
 631
 632 static size_t
 633 sort__parent_print(FILE *fp, struct hist_entry *self)
 634 {
 635         size_t ret = 0;
 636
 637         ret += fprintf(fp, "%-20s", self->parent ? self->parent->name : "[other]");
 638
 639         return ret;
 640 }
 641
 642 static struct sort_entry sort_parent = {
 643         .header = "Parent symbol       ",
 644         .cmp    = sort__parent_cmp,
 645         .print  = sort__parent_print,
 646 };
 647
 648 static int sort__need_collapse = 0;
 649 static int sort__has_parent = 0;
 650
 651 struct sort_dimension {
 652         char                    *name;
 653         struct sort_entry       *entry;
 654         int                     taken;
 655 };
 656
 657 static struct sort_dimension sort_dimensions[] = {
 658         { .name = "pid",        .entry = &sort_thread,  },
 659         { .name = "comm",       .entry = &sort_comm,    },
 660         { .name = "dso",        .entry = &sort_dso,     },
 661         { .name = "symbol",     .entry = &sort_sym,     },
 662         { .name = "parent",     .entry = &sort_parent,  },
 663 };
 664
 665 static LIST_HEAD(hist_entry__sort_list);
 666
 667 static int sort_dimension__add(char *tok)
 668 {
 669         int i;
 670
 671         for (i = 0; i < ARRAY_SIZE(sort_dimensions); i++) {
 672                 struct sort_dimension *sd = &sort_dimensions[i];
 673
 674                 if (sd->taken)
 675                         continue;
 676
 677                 if (strncasecmp(tok, sd->name, strlen(tok)))
 678                         continue;
 679
 680                 if (sd->entry->collapse)
 681                         sort__need_collapse = 1;
 682
 683                 if (sd->entry == &sort_parent) {
 684                         int ret = regcomp(&parent_regex, parent_pattern, REG_EXTENDED);
 685                         if (ret) {
 686                                 char err[BUFSIZ];
 687
 688                                 regerror(ret, &parent_regex, err, sizeof(err));
 689                                 fprintf(stderr, "Invalid regex: %s\n%s",
 690                                         parent_pattern, err);
 691                                 exit(-1);
 692                         }
 693                         sort__has_parent = 1;
 694                 }
 695
 696                 list_add_tail(&sd->entry->list, &hist_entry__sort_list);
 697                 sd->taken = 1;
 698
 699                 return 0;
 700         }
 701
 702         return -ESRCH;
 703 }
 704
 705 static int64_t
 706 hist_entry__cmp(struct hist_entry *left, struct hist_entry *right)
 707 {
 708         struct sort_entry *se;
 709         int64_t cmp = 0;
 710
 711         list_for_each_entry(se, &hist_entry__sort_list, list) {
 712                 cmp = se->cmp(left, right);
 713                 if (cmp)
 714                         break;
 715         }
 716
 717         return cmp;
 718 }
 719
 720 static int64_t
 721 hist_entry__collapse(struct hist_entry *left, struct hist_entry *right)
 722 {
 723         struct sort_entry *se;
 724         int64_t cmp = 0;
 725
 726         list_for_each_entry(se, &hist_entry__sort_list, list) {
 727                 int64_t (*f)(struct hist_entry *, struct hist_entry *);
 728
 729                 f = se->collapse ?: se->cmp;
 730
 731                 cmp = f(left, right);
 732                 if (cmp)
 733                         break;
 734         }
 735
 736         return cmp;
 737 }
 738
 739 static size_t
 740 hist_entry__fprintf(FILE *fp, struct hist_entry *self, __u64 total_samples)
 741 {
 742         struct sort_entry *se;
 743         size_t ret;
 744
 745         if (total_samples) {
 746                 double percent = self->count * 100.0 / total_samples;
 747                 char *color = PERF_COLOR_NORMAL;
 748
 749                 /*
 750                  * We color high-overhead entries in red, mid-overhead
 751                  * entries in green - and keep the low overhead places
 752                  * normal:
 753                  */
 754                 if (percent >= 5.0) {
 755                         color = PERF_COLOR_RED;
 756                 } else {
 757                         if (percent >= 0.5)
 758                                 color = PERF_COLOR_GREEN;
 759                 }
 760
 761                 ret = color_fprintf(fp, color, "   %6.2f%%",
 762                                 (self->count * 100.0) / total_samples);
 763         } else
 764                 ret = fprintf(fp, "%12Ld ", self->count);
 765
 766         list_for_each_entry(se, &hist_entry__sort_list, list) {
 767                 fprintf(fp, "  ");
 768                 ret += se->print(fp, self);
 769         }
 770
 771         ret += fprintf(fp, "\n");
 772
 773         return ret;
 774 }
 775
 776 /*
 777  *
 778  */
 779
 780 static struct symbol *
 781 resolve_symbol(struct thread *thread, struct map **mapp,
 782                struct dso **dsop, __u64 *ipp)
 783 {
 784         struct dso *dso = dsop ? *dsop : NULL;
 785         struct map *map = mapp ? *mapp : NULL;
 786         uint64_t ip = *ipp;
 787
 788         if (!thread)
 789                 return NULL;
 790
 791         if (dso)
 792                 goto got_dso;
 793
 794         if (map)
 795                 goto got_map;
 796
 797         map = thread__find_map(thread, ip);
 798         if (map != NULL) {
 799                 if (mapp)
 800                         *mapp = map;
 801 got_map:
 802                 ip = map->map_ip(map, ip);
 803                 *ipp  = ip;
 804
 805                 dso = map->dso;
 806         } else {
 807                 /*
 808                  * If this is outside of all known maps,
 809                  * and is a negative address, try to look it
 810                  * up in the kernel dso, as it might be a
 811                  * vsyscall (which executes in user-mode):
 812                  */
 813                 if ((long long)ip < 0)
 814                 dso = kernel_dso;
 815         }
 816         dprintf(" ...... dso: %s\n", dso ? dso->name : "<not found>");
 817
 818         if (dsop)
 819                 *dsop = dso;
 820
 821         if (!dso)
 822                 return NULL;
 823 got_dso:
 824         return dso->find_symbol(dso, ip);
 825 }
 826
 827 static struct symbol *call__match(struct symbol *sym)
 828 {
 829         if (!sym)
 830                 return NULL;
 831
 832         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
 833                 return sym;
 834
 835         return NULL;
 836 }
 837
 838 /*
 839  * collect histogram counts
 840  */
 841
 842 static int
 843 hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
 844                 struct symbol *sym, __u64 ip, struct perf_callchain_entry *chain,
 845                 char level, __u64 count)
 846 {
 847         struct rb_node **p = &hist.rb_node;
 848         struct rb_node *parent = NULL;
 849         struct hist_entry *he;
 850         struct hist_entry entry = {
 851                 .thread = thread,
 852                 .map    = map,
 853                 .dso    = dso,
 854                 .sym    = sym,
 855                 .ip     = ip,
 856                 .level  = level,
 857                 .count  = count,
 858         };
 859         int cmp;
 860
 861         if (sort__has_parent && chain) {
 862                 int i, nr = chain->hv;
 863                 struct symbol *sym;
 864                 struct dso *dso;
 865                 __u64 ip;
 866
 867                 for (i = 0; i < chain->kernel; i++) {
 868                         ip = chain->ip[nr + i];
 869                         dso = kernel_dso;
 870                         sym = resolve_symbol(thread, NULL, &dso, &ip);
 871                         entry.parent = call__match(sym);
 872                         if (entry.parent)
 873                                 goto got_parent;
 874                 }
 875                 nr += i;
 876
 877                 for (i = 0; i < chain->user; i++) {
 878                         ip = chain->ip[nr + i];
 879                         sym = resolve_symbol(thread, NULL, NULL, &ip);
 880                         entry.parent = call__match(sym);
 881                         if (entry.parent)
 882                                 goto got_parent;
 883                 }
 884                 nr += i;
 885         }
 886 got_parent:
 887
 888         while (*p != NULL) {
 889                 parent = *p;
 890                 he = rb_entry(parent, struct hist_entry, rb_node);
 891
 892                 cmp = hist_entry__cmp(&entry, he);
 893
 894                 if (!cmp) {
 895                         he->count += count;
 896                         return 0;
 897                 }
 898
 899                 if (cmp < 0)
 900                         p = &(*p)->rb_left;
 901                 else
 902                         p = &(*p)->rb_right;
 903         }
 904
 905         he = malloc(sizeof(*he));
 906         if (!he)
 907                 return -ENOMEM;
 908         *he = entry;
 909         rb_link_node(&he->rb_node, parent, p);
 910         rb_insert_color(&he->rb_node, &hist);
 911
 912         return 0;
 913 }
 914
 915 static void hist_entry__free(struct hist_entry *he)
 916 {
 917         free(he);
 918 }
 919
 920 /*
 921  * collapse the histogram
 922  */
 923
 924 static struct rb_root collapse_hists;
 925
 926 static void collapse__insert_entry(struct hist_entry *he)
 927 {
 928         struct rb_node **p = &collapse_hists.rb_node;
 929         struct rb_node *parent = NULL;
 930         struct hist_entry *iter;
 931         int64_t cmp;
 932
 933         while (*p != NULL) {
 934                 parent = *p;
 935                 iter = rb_entry(parent, struct hist_entry, rb_node);
 936
 937                 cmp = hist_entry__collapse(iter, he);
 938
 939                 if (!cmp) {
 940                         iter->count += he->count;
 941                         hist_entry__free(he);
 942                         return;
 943                 }
 944
 945                 if (cmp < 0)
 946                         p = &(*p)->rb_left;
 947                 else
 948                         p = &(*p)->rb_right;
 949         }
 950
 951         rb_link_node(&he->rb_node, parent, p);
 952         rb_insert_color(&he->rb_node, &collapse_hists);
 953 }
 954
 955 static void collapse__resort(void)
 956 {
 957         struct rb_node *next;
 958         struct hist_entry *n;
 959
 960         if (!sort__need_collapse)
 961                 return;
 962
 963         next = rb_first(&hist);
 964         while (next) {
 965                 n = rb_entry(next, struct hist_entry, rb_node);
 966                 next = rb_next(&n->rb_node);
 967
 968                 rb_erase(&n->rb_node, &hist);
 969                 collapse__insert_entry(n);
 970         }
 971 }
 972
 973 /*
 974  * reverse the map, sort on count.
 975  */
 976
 977 static struct rb_root output_hists;
 978
 979 static void output__insert_entry(struct hist_entry *he)
 980 {
 981         struct rb_node **p = &output_hists.rb_node;
 982         struct rb_node *parent = NULL;
 983         struct hist_entry *iter;
 984
 985         while (*p != NULL) {
 986                 parent = *p;
 987                 iter = rb_entry(parent, struct hist_entry, rb_node);
 988
 989                 if (he->count > iter->count)
 990                         p = &(*p)->rb_left;
 991                 else
 992                         p = &(*p)->rb_right;
 993         }
 994
 995         rb_link_node(&he->rb_node, parent, p);
 996         rb_insert_color(&he->rb_node, &output_hists);
 997 }
 998
 999 static void output__resort(void)
1000 {
1001         struct rb_node *next;
1002         struct hist_entry *n;
1003         struct rb_root *tree = &hist;
1004
1005         if (sort__need_collapse)
1006                 tree = &collapse_hists;
1007
1008         next = rb_first(tree);
1009
1010         while (next) {
1011                 n = rb_entry(next, struct hist_entry, rb_node);
1012                 next = rb_next(&n->rb_node);
1013
1014                 rb_erase(&n->rb_node, tree);
1015                 output__insert_entry(n);
1016         }
1017 }
1018
1019 static size_t output__fprintf(FILE *fp, __u64 total_samples)
1020 {
1021         struct hist_entry *pos;
1022         struct sort_entry *se;
1023         struct rb_node *nd;
1024         size_t ret = 0;
1025
1026         fprintf(fp, "\n");
1027         fprintf(fp, "#\n");
1028         fprintf(fp, "# (%Ld samples)\n", (__u64)total_samples);
1029         fprintf(fp, "#\n");
1030
1031         fprintf(fp, "# Overhead");
1032         list_for_each_entry(se, &hist_entry__sort_list, list)
1033                 fprintf(fp, "  %s", se->header);
1034         fprintf(fp, "\n");
1035
1036         fprintf(fp, "# ........");
1037         list_for_each_entry(se, &hist_entry__sort_list, list) {
1038                 int i;
1039
1040                 fprintf(fp, "  ");
1041                 for (i = 0; i < strlen(se->header); i++)
1042                         fprintf(fp, ".");
1043         }
1044         fprintf(fp, "\n");
1045
1046         fprintf(fp, "#\n");
1047
1048         for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
1049                 pos = rb_entry(nd, struct hist_entry, rb_node);
1050                 ret += hist_entry__fprintf(fp, pos, total_samples);
1051         }
1052
1053         if (!strcmp(sort_order, default_sort_order)) {
1054                 fprintf(fp, "#\n");
1055                 fprintf(fp, "# (For more details, try: perf report --sort comm,dso,symbol)\n");
1056                 fprintf(fp, "#\n");
1057         }
1058         fprintf(fp, "\n");
1059
1060         return ret;
1061 }
1062
1063 static void register_idle_thread(void)
1064 {
1065         struct thread *thread = threads__findnew(0);
1066
1067         if (thread == NULL ||
1068                         thread__set_comm(thread, "[idle]")) {
1069                 fprintf(stderr, "problem inserting idle task.\n");
1070                 exit(-1);
1071         }
1072 }
1073
1074 static unsigned long total = 0,
1075                      total_mmap = 0,
1076                      total_comm = 0,
1077                      total_fork = 0,
1078                      total_unknown = 0,
1079                      total_lost = 0;
1080
1081 static int validate_chain(struct perf_callchain_entry *chain, event_t *event)
1082 {
1083         unsigned int chain_size;
1084
1085         if (chain->nr > MAX_STACK_DEPTH)
1086                 return -1;
1087         if (chain->hv > MAX_STACK_DEPTH)
1088                 return -1;
1089         if (chain->kernel > MAX_STACK_DEPTH)
1090                 return -1;
1091         if (chain->user > MAX_STACK_DEPTH)
1092                 return -1;
1093         if (chain->hv + chain->kernel + chain->user != chain->nr)
1094                 return -1;
1095
1096         chain_size = event->header.size;
1097         chain_size -= (unsigned long)&event->ip.__more_data - (unsigned long)event;
1098
1099         if (chain->nr*sizeof(__u64) > chain_size)
1100                 return -1;
1101
1102         return 0;
1103 }
1104
1105 static int
1106 process_overflow_event(event_t *event, unsigned long offset, unsigned long head)
1107 {
1108         char level;
1109         int show = 0;
1110         struct dso *dso = NULL;
1111         struct thread *thread = threads__findnew(event->ip.pid);
1112         __u64 ip = event->ip.ip;
1113         __u64 period = 1;
1114         struct map *map = NULL;
1115         void *more_data = event->ip.__more_data;
1116         struct perf_callchain_entry *chain = NULL;
1117
1118         if (event->header.type & PERF_SAMPLE_PERIOD) {
1119                 period = *(__u64 *)more_data;
1120                 more_data += sizeof(__u64);
1121         }
1122
1123         dprintf("%p [%p]: PERF_EVENT (IP, %d): %d: %p period: %Ld\n",
1124                 (void *)(offset + head),
1125                 (void *)(long)(event->header.size),
1126                 event->header.misc,
1127                 event->ip.pid,
1128                 (void *)(long)ip,
1129                 (long long)period);
1130
1131         if (event->header.type & PERF_SAMPLE_CALLCHAIN) {
1132                 int i;
1133
1134                 chain = (void *)more_data;
1135
1136                 dprintf("... chain: u:%d, k:%d, nr:%d\n",
1137                         chain->user,
1138                         chain->kernel,
1139                         chain->nr);
1140
1141                 if (validate_chain(chain, event) < 0) {
1142                         eprintf("call-chain problem with event, skipping it.\n");
1143                         return 0;
1144                 }
1145
1146                 if (dump_trace) {
1147                         for (i = 0; i < chain->nr; i++)
1148                                 dprintf("..... %2d: %016Lx\n", i, chain->ip[i]);
1149                 }
1150         }
1151
1152         dprintf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1153
1154         if (thread == NULL) {
1155                 eprintf("problem processing %d event, skipping it.\n",
1156                         event->header.type);
1157                 return -1;
1158         }
1159
1160         if (event->header.misc & PERF_EVENT_MISC_KERNEL) {
1161                 show = SHOW_KERNEL;
1162                 level = 'k';
1163
1164                 dso = kernel_dso;
1165
1166                 dprintf(" ...... dso: %s\n", dso->name);
1167
1168         } else if (event->header.misc & PERF_EVENT_MISC_USER) {
1169
1170                 show = SHOW_USER;
1171                 level = '.';
1172
1173         } else {
1174                 show = SHOW_HV;
1175                 level = 'H';
1176                 dprintf(" ...... dso: [hypervisor]\n");
1177         }
1178
1179         if (show & show_mask) {
1180                 struct symbol *sym = resolve_symbol(thread, &map, &dso, &ip);
1181
1182                 if (hist_entry__add(thread, map, dso, sym, ip, chain, level, period)) {
1183                         eprintf("problem incrementing symbol count, skipping event\n");
1184                         return -1;
1185                 }
1186         }
1187         total += period;
1188
1189         return 0;
1190 }
1191
1192 static int
1193 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
1194 {
1195         struct thread *thread = threads__findnew(event->mmap.pid);
1196         struct map *map = map__new(&event->mmap);
1197
1198         dprintf("%p [%p]: PERF_EVENT_MMAP %d: [%p(%p) @ %p]: %s\n",
1199                 (void *)(offset + head),
1200                 (void *)(long)(event->header.size),
1201                 event->mmap.pid,
1202                 (void *)(long)event->mmap.start,
1203                 (void *)(long)event->mmap.len,
1204                 (void *)(long)event->mmap.pgoff,
1205                 event->mmap.filename);
1206
1207         if (thread == NULL || map == NULL) {
1208                 dprintf("problem processing PERF_EVENT_MMAP, skipping event.\n");
1209                 return 0;
1210         }
1211
1212         thread__insert_map(thread, map);
1213         total_mmap++;
1214
1215         return 0;
1216 }
1217
1218 static int
1219 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
1220 {
1221         struct thread *thread = threads__findnew(event->comm.pid);
1222
1223         dprintf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
1224                 (void *)(offset + head),
1225                 (void *)(long)(event->header.size),
1226                 event->comm.comm, event->comm.pid);
1227
1228         if (thread == NULL ||
1229             thread__set_comm(thread, event->comm.comm)) {
1230                 dprintf("problem processing PERF_EVENT_COMM, skipping event.\n");
1231                 return -1;
1232         }
1233         total_comm++;
1234
1235         return 0;
1236 }
1237
1238 static int
1239 process_fork_event(event_t *event, unsigned long offset, unsigned long head)
1240 {
1241         struct thread *thread = threads__findnew(event->fork.pid);
1242         struct thread *parent = threads__findnew(event->fork.ppid);
1243
1244         dprintf("%p [%p]: PERF_EVENT_FORK: %d:%d\n",
1245                 (void *)(offset + head),
1246                 (void *)(long)(event->header.size),
1247                 event->fork.pid, event->fork.ppid);
1248
1249         if (!thread || !parent || thread__fork(thread, parent)) {
1250                 dprintf("problem processing PERF_EVENT_FORK, skipping event.\n");
1251                 return -1;
1252         }
1253         total_fork++;
1254
1255         return 0;
1256 }
1257
1258 static int
1259 process_period_event(event_t *event, unsigned long offset, unsigned long head)
1260 {
1261         dprintf("%p [%p]: PERF_EVENT_PERIOD: time:%Ld, id:%Ld: period:%Ld\n",
1262                 (void *)(offset + head),
1263                 (void *)(long)(event->header.size),
1264                 event->period.time,
1265                 event->period.id,
1266                 event->period.sample_period);
1267
1268         return 0;
1269 }
1270
1271 static int
1272 process_lost_event(event_t *event, unsigned long offset, unsigned long head)
1273 {
1274         dprintf("%p [%p]: PERF_EVENT_LOST: id:%Ld: lost:%Ld\n",
1275                 (void *)(offset + head),
1276                 (void *)(long)(event->header.size),
1277                 event->lost.id,
1278                 event->lost.lost);
1279
1280         total_lost += event->lost.lost;
1281
1282         return 0;
1283 }
1284
1285 static void trace_event(event_t *event)
1286 {
1287         unsigned char *raw_event = (void *)event;
1288         char *color = PERF_COLOR_BLUE;
1289         int i, j;
1290
1291         if (!dump_trace)
1292                 return;
1293
1294         dprintf(".");
1295         cdprintf("\n. ... raw event: size %d bytes\n", event->header.size);
1296
1297         for (i = 0; i < event->header.size; i++) {
1298                 if ((i & 15) == 0) {
1299                         dprintf(".");
1300                         cdprintf("  %04x: ", i);
1301                 }
1302
1303                 cdprintf(" %02x", raw_event[i]);
1304
1305                 if (((i & 15) == 15) || i == event->header.size-1) {
1306                         cdprintf("  ");
1307                         for (j = 0; j < 15-(i & 15); j++)
1308                                 cdprintf("   ");
1309                         for (j = 0; j < (i & 15); j++) {
1310                                 if (isprint(raw_event[i-15+j]))
1311                                         cdprintf("%c", raw_event[i-15+j]);
1312                                 else
1313                                         cdprintf(".");
1314                         }
1315                         cdprintf("\n");
1316                 }
1317         }
1318         dprintf(".\n");
1319 }
1320
1321 static int
1322 process_event(event_t *event, unsigned long offset, unsigned long head)
1323 {
1324         trace_event(event);
1325
1326         if (event->header.misc & PERF_EVENT_MISC_OVERFLOW)
1327                 return process_overflow_event(event, offset, head);
1328
1329         switch (event->header.type) {
1330         case PERF_EVENT_MMAP:
1331                 return process_mmap_event(event, offset, head);
1332
1333         case PERF_EVENT_COMM:
1334                 return process_comm_event(event, offset, head);
1335
1336         case PERF_EVENT_FORK:
1337                 return process_fork_event(event, offset, head);
1338
1339         case PERF_EVENT_PERIOD:
1340                 return process_period_event(event, offset, head);
1341
1342         case PERF_EVENT_LOST:
1343                 return process_lost_event(event, offset, head);
1344
1345         /*
1346          * We dont process them right now but they are fine:
1347          */
1348
1349         case PERF_EVENT_THROTTLE:
1350         case PERF_EVENT_UNTHROTTLE:
1351                 return 0;
1352
1353         default:
1354                 return -1;
1355         }
1356
1357         return 0;
1358 }
1359
1360 static int __cmd_report(void)
1361 {
1362         int ret, rc = EXIT_FAILURE;
1363         unsigned long offset = 0;
1364         unsigned long head = 0;
1365         struct stat stat;
1366         event_t *event;
1367         uint32_t size;
1368         char *buf;
1369
1370         register_idle_thread();
1371
1372         input = open(input_name, O_RDONLY);
1373         if (input < 0) {
1374                 fprintf(stderr, " failed to open file: %s", input_name);
1375                 if (!strcmp(input_name, "perf.data"))
1376                         fprintf(stderr, "  (try 'perf record' first)");
1377                 fprintf(stderr, "\n");
1378                 exit(-1);
1379         }
1380
1381         ret = fstat(input, &stat);
1382         if (ret < 0) {
1383                 perror("failed to stat file");
1384                 exit(-1);
1385         }
1386
1387         if (!stat.st_size) {
1388                 fprintf(stderr, "zero-sized file, nothing to do!\n");
1389                 exit(0);
1390         }
1391
1392         if (load_kernel() < 0) {
1393                 perror("failed to load kernel symbols");
1394                 return EXIT_FAILURE;
1395         }
1396
1397         if (!full_paths) {
1398                 if (getcwd(__cwd, sizeof(__cwd)) == NULL) {
1399                         perror("failed to get the current directory");
1400                         return EXIT_FAILURE;
1401                 }
1402                 cwdlen = strlen(cwd);
1403         } else {
1404                 cwd = NULL;
1405                 cwdlen = 0;
1406         }
1407 remap:
1408         buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1409                            MAP_SHARED, input, offset);
1410         if (buf == MAP_FAILED) {
1411                 perror("failed to mmap file");
1412                 exit(-1);
1413         }
1414
1415 more:
1416         event = (event_t *)(buf + head);
1417
1418         size = event->header.size;
1419         if (!size)
1420                 size = 8;
1421
1422         if (head + event->header.size >= page_size * mmap_window) {
1423                 unsigned long shift = page_size * (head / page_size);
1424                 int ret;
1425
1426                 ret = munmap(buf, page_size * mmap_window);
1427                 assert(ret == 0);
1428
1429                 offset += shift;
1430                 head -= shift;
1431                 goto remap;
1432         }
1433
1434         size = event->header.size;
1435
1436         dprintf("\n%p [%p]: event: %d\n",
1437                         (void *)(offset + head),
1438                         (void *)(long)event->header.size,
1439                         event->header.type);
1440
1441         if (!size || process_event(event, offset, head) < 0) {
1442
1443                 dprintf("%p [%p]: skipping unknown header type: %d\n",
1444                         (void *)(offset + head),
1445                         (void *)(long)(event->header.size),
1446                         event->header.type);
1447
1448                 total_unknown++;
1449
1450                 /*
1451                  * assume we lost track of the stream, check alignment, and
1452                  * increment a single u64 in the hope to catch on again 'soon'.
1453                  */
1454
1455                 if (unlikely(head & 7))
1456                         head &= ~7ULL;
1457
1458                 size = 8;
1459         }
1460
1461         head += size;
1462
1463         if (offset + head < stat.st_size)
1464                 goto more;
1465
1466         rc = EXIT_SUCCESS;
1467         close(input);
1468
1469         dprintf("      IP events: %10ld\n", total);
1470         dprintf("    mmap events: %10ld\n", total_mmap);
1471         dprintf("    comm events: %10ld\n", total_comm);
1472         dprintf("    fork events: %10ld\n", total_fork);
1473         dprintf("    lost events: %10ld\n", total_lost);
1474         dprintf(" unknown events: %10ld\n", total_unknown);
1475
1476         if (dump_trace)
1477                 return 0;
1478
1479         if (verbose >= 3)
1480                 threads__fprintf(stdout);
1481
1482         if (verbose >= 2)
1483                 dsos__fprintf(stdout);
1484
1485         collapse__resort();
1486         output__resort();
1487         output__fprintf(stdout, total);
1488
1489         return rc;
1490 }
1491
1492 static const char * const report_usage[] = {
1493         "perf report [<options>] <command>",
1494         NULL
1495 };
1496
1497 static const struct option options[] = {
1498         OPT_STRING('i', "input", &input_name, "file",
1499                     "input file name"),
1500         OPT_BOOLEAN('v', "verbose", &verbose,
1501                     "be more verbose (show symbol address, etc)"),
1502         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1503                     "dump raw trace in ASCII"),
1504         OPT_STRING('k', "vmlinux", &vmlinux, "file", "vmlinux pathname"),
1505         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1506                    "sort by key(s): pid, comm, dso, symbol, parent"),
1507         OPT_BOOLEAN('P', "full-paths", &full_paths,
1508                     "Don't shorten the pathnames taking into account the cwd"),
1509         OPT_STRING('p', "parent", &parent_pattern, "regex",
1510                    "regex filter to identify parent, see: '--sort parent'"),
1511         OPT_END()
1512 };
1513
1514 static void setup_sorting(void)
1515 {
1516         char *tmp, *tok, *str = strdup(sort_order);
1517
1518         for (tok = strtok_r(str, ", ", &tmp);
1519                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1520                 if (sort_dimension__add(tok) < 0) {
1521                         error("Unknown --sort key: `%s'", tok);
1522                         usage_with_options(report_usage, options);
1523                 }
1524         }
1525
1526         free(str);
1527 }
1528
1529 int cmd_report(int argc, const char **argv, const char *prefix)
1530 {
1531         symbol__init();
1532
1533         page_size = getpagesize();
1534
1535         argc = parse_options(argc, argv, options, report_usage, 0);
1536
1537         setup_sorting();
1538
1539         /*
1540          * Any (unrecognized) arguments left?
1541          */
1542         if (argc)
1543                 usage_with_options(report_usage, options);
1544
1545         setup_pager();
1546
1547         return __cmd_report();
1548 }