tools/perf/builtin-annotate.c

   1 /*
   2  * builtin-annotate.c
   3  *
   4  * Builtin annotate 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 <linux/list.h>
  14 #include "util/cache.h"
  15 #include <linux/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 = "vmlinux";
  30
  31 static char             default_sort_order[] = "comm,symbol";
  32 static char             *sort_order = default_sort_order;
  33
  34 static int              force;
  35 static int              input;
  36 static int              show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV;
  37
  38 static int              dump_trace = 0;
  39 #define dprintf(x...)   do { if (dump_trace) printf(x); } while (0)
  40
  41 static int              verbose;
  42
  43 static int              modules;
  44
  45 static int              full_paths;
  46
  47 static int              print_line;
  48
  49 static unsigned long    page_size;
  50 static unsigned long    mmap_window = 32;
  51
  52 struct ip_event {
  53         struct perf_event_header header;
  54         u64 ip;
  55         u32 pid, tid;
  56 };
  57
  58 struct mmap_event {
  59         struct perf_event_header header;
  60         u32 pid, tid;
  61         u64 start;
  62         u64 len;
  63         u64 pgoff;
  64         char filename[PATH_MAX];
  65 };
  66
  67 struct comm_event {
  68         struct perf_event_header header;
  69         u32 pid, tid;
  70         char comm[16];
  71 };
  72
  73 struct fork_event {
  74         struct perf_event_header header;
  75         u32 pid, ppid;
  76 };
  77
  78 typedef union event_union {
  79         struct perf_event_header        header;
  80         struct ip_event                 ip;
  81         struct mmap_event               mmap;
  82         struct comm_event               comm;
  83         struct fork_event               fork;
  84 } event_t;
  85
  86
  87 struct sym_ext {
  88         struct rb_node  node;
  89         double          percent;
  90         char            *path;
  91 };
  92
  93 static LIST_HEAD(dsos);
  94 static struct dso *kernel_dso;
  95 static struct dso *vdso;
  96
  97
  98 static void dsos__add(struct dso *dso)
  99 {
 100         list_add_tail(&dso->node, &dsos);
 101 }
 102
 103 static struct dso *dsos__find(const char *name)
 104 {
 105         struct dso *pos;
 106
 107         list_for_each_entry(pos, &dsos, node)
 108                 if (strcmp(pos->name, name) == 0)
 109                         return pos;
 110         return NULL;
 111 }
 112
 113 static struct dso *dsos__findnew(const char *name)
 114 {
 115         struct dso *dso = dsos__find(name);
 116         int nr;
 117
 118         if (dso)
 119                 return dso;
 120
 121         dso = dso__new(name, 0);
 122         if (!dso)
 123                 goto out_delete_dso;
 124
 125         nr = dso__load(dso, NULL, verbose);
 126         if (nr < 0) {
 127                 if (verbose)
 128                         fprintf(stderr, "Failed to open: %s\n", name);
 129                 goto out_delete_dso;
 130         }
 131         if (!nr && verbose) {
 132                 fprintf(stderr,
 133                 "No symbols found in: %s, maybe install a debug package?\n",
 134                                 name);
 135         }
 136
 137         dsos__add(dso);
 138
 139         return dso;
 140
 141 out_delete_dso:
 142         dso__delete(dso);
 143         return NULL;
 144 }
 145
 146 static void dsos__fprintf(FILE *fp)
 147 {
 148         struct dso *pos;
 149
 150         list_for_each_entry(pos, &dsos, node)
 151                 dso__fprintf(pos, fp);
 152 }
 153
 154 static struct symbol *vdso__find_symbol(struct dso *dso, u64 ip)
 155 {
 156         return dso__find_symbol(dso, ip);
 157 }
 158
 159 static int load_kernel(void)
 160 {
 161         int err;
 162
 163         kernel_dso = dso__new("[kernel]", 0);
 164         if (!kernel_dso)
 165                 return -1;
 166
 167         err = dso__load_kernel(kernel_dso, vmlinux, NULL, verbose, modules);
 168         if (err <= 0) {
 169                 dso__delete(kernel_dso);
 170                 kernel_dso = NULL;
 171         } else
 172                 dsos__add(kernel_dso);
 173
 174         vdso = dso__new("[vdso]", 0);
 175         if (!vdso)
 176                 return -1;
 177
 178         vdso->find_symbol = vdso__find_symbol;
 179
 180         dsos__add(vdso);
 181
 182         return err;
 183 }
 184
 185 struct map {
 186         struct list_head node;
 187         u64      start;
 188         u64      end;
 189         u64      pgoff;
 190         u64      (*map_ip)(struct map *, u64);
 191         struct dso       *dso;
 192 };
 193
 194 static u64 map__map_ip(struct map *map, u64 ip)
 195 {
 196         return ip - map->start + map->pgoff;
 197 }
 198
 199 static u64 vdso__map_ip(struct map *map __used, u64 ip)
 200 {
 201         return ip;
 202 }
 203
 204 static struct map *map__new(struct mmap_event *event)
 205 {
 206         struct map *self = malloc(sizeof(*self));
 207
 208         if (self != NULL) {
 209                 const char *filename = event->filename;
 210
 211                 self->start = event->start;
 212                 self->end   = event->start + event->len;
 213                 self->pgoff = event->pgoff;
 214
 215                 self->dso = dsos__findnew(filename);
 216                 if (self->dso == NULL)
 217                         goto out_delete;
 218
 219                 if (self->dso == vdso)
 220                         self->map_ip = vdso__map_ip;
 221                 else
 222                         self->map_ip = map__map_ip;
 223         }
 224         return self;
 225 out_delete:
 226         free(self);
 227         return NULL;
 228 }
 229
 230 static struct map *map__clone(struct map *self)
 231 {
 232         struct map *map = malloc(sizeof(*self));
 233
 234         if (!map)
 235                 return NULL;
 236
 237         memcpy(map, self, sizeof(*self));
 238
 239         return map;
 240 }
 241
 242 static int map__overlap(struct map *l, struct map *r)
 243 {
 244         if (l->start > r->start) {
 245                 struct map *t = l;
 246                 l = r;
 247                 r = t;
 248         }
 249
 250         if (l->end > r->start)
 251                 return 1;
 252
 253         return 0;
 254 }
 255
 256 static size_t map__fprintf(struct map *self, FILE *fp)
 257 {
 258         return fprintf(fp, " %Lx-%Lx %Lx %s\n",
 259                        self->start, self->end, self->pgoff, self->dso->name);
 260 }
 261
 262
 263 struct thread {
 264         struct rb_node   rb_node;
 265         struct list_head maps;
 266         pid_t            pid;
 267         char             *comm;
 268 };
 269
 270 static struct thread *thread__new(pid_t pid)
 271 {
 272         struct thread *self = malloc(sizeof(*self));
 273
 274         if (self != NULL) {
 275                 self->pid = pid;
 276                 self->comm = malloc(32);
 277                 if (self->comm)
 278                         snprintf(self->comm, 32, ":%d", self->pid);
 279                 INIT_LIST_HEAD(&self->maps);
 280         }
 281
 282         return self;
 283 }
 284
 285 static int thread__set_comm(struct thread *self, const char *comm)
 286 {
 287         if (self->comm)
 288                 free(self->comm);
 289         self->comm = strdup(comm);
 290         return self->comm ? 0 : -ENOMEM;
 291 }
 292
 293 static size_t thread__fprintf(struct thread *self, FILE *fp)
 294 {
 295         struct map *pos;
 296         size_t ret = fprintf(fp, "Thread %d %s\n", self->pid, self->comm);
 297
 298         list_for_each_entry(pos, &self->maps, node)
 299                 ret += map__fprintf(pos, fp);
 300
 301         return ret;
 302 }
 303
 304
 305 static struct rb_root threads;
 306 static struct thread *last_match;
 307
 308 static struct thread *threads__findnew(pid_t pid)
 309 {
 310         struct rb_node **p = &threads.rb_node;
 311         struct rb_node *parent = NULL;
 312         struct thread *th;
 313
 314         /*
 315          * Font-end cache - PID lookups come in blocks,
 316          * so most of the time we dont have to look up
 317          * the full rbtree:
 318          */
 319         if (last_match && last_match->pid == pid)
 320                 return last_match;
 321
 322         while (*p != NULL) {
 323                 parent = *p;
 324                 th = rb_entry(parent, struct thread, rb_node);
 325
 326                 if (th->pid == pid) {
 327                         last_match = th;
 328                         return th;
 329                 }
 330
 331                 if (pid < th->pid)
 332                         p = &(*p)->rb_left;
 333                 else
 334                         p = &(*p)->rb_right;
 335         }
 336
 337         th = thread__new(pid);
 338         if (th != NULL) {
 339                 rb_link_node(&th->rb_node, parent, p);
 340                 rb_insert_color(&th->rb_node, &threads);
 341                 last_match = th;
 342         }
 343
 344         return th;
 345 }
 346
 347 static void thread__insert_map(struct thread *self, struct map *map)
 348 {
 349         struct map *pos, *tmp;
 350
 351         list_for_each_entry_safe(pos, tmp, &self->maps, node) {
 352                 if (map__overlap(pos, map)) {
 353                         list_del_init(&pos->node);
 354                         /* XXX leaks dsos */
 355                         free(pos);
 356                 }
 357         }
 358
 359         list_add_tail(&map->node, &self->maps);
 360 }
 361
 362 static int thread__fork(struct thread *self, struct thread *parent)
 363 {
 364         struct map *map;
 365
 366         if (self->comm)
 367                 free(self->comm);
 368         self->comm = strdup(parent->comm);
 369         if (!self->comm)
 370                 return -ENOMEM;
 371
 372         list_for_each_entry(map, &parent->maps, node) {
 373                 struct map *new = map__clone(map);
 374                 if (!new)
 375                         return -ENOMEM;
 376                 thread__insert_map(self, new);
 377         }
 378
 379         return 0;
 380 }
 381
 382 static struct map *thread__find_map(struct thread *self, u64 ip)
 383 {
 384         struct map *pos;
 385
 386         if (self == NULL)
 387                 return NULL;
 388
 389         list_for_each_entry(pos, &self->maps, node)
 390                 if (ip >= pos->start && ip <= pos->end)
 391                         return pos;
 392
 393         return NULL;
 394 }
 395
 396 static size_t threads__fprintf(FILE *fp)
 397 {
 398         size_t ret = 0;
 399         struct rb_node *nd;
 400
 401         for (nd = rb_first(&threads); nd; nd = rb_next(nd)) {
 402                 struct thread *pos = rb_entry(nd, struct thread, rb_node);
 403
 404                 ret += thread__fprintf(pos, fp);
 405         }
 406
 407         return ret;
 408 }
 409
 410 /*
 411  * histogram, sorted on item, collects counts
 412  */
 413
 414 static struct rb_root hist;
 415
 416 struct hist_entry {
 417         struct rb_node   rb_node;
 418
 419         struct thread    *thread;
 420         struct map       *map;
 421         struct dso       *dso;
 422         struct symbol    *sym;
 423         u64      ip;
 424         char             level;
 425
 426         uint32_t         count;
 427 };
 428
 429 /*
 430  * configurable sorting bits
 431  */
 432
 433 struct sort_entry {
 434         struct list_head list;
 435
 436         char *header;
 437
 438         int64_t (*cmp)(struct hist_entry *, struct hist_entry *);
 439         int64_t (*collapse)(struct hist_entry *, struct hist_entry *);
 440         size_t  (*print)(FILE *fp, struct hist_entry *);
 441 };
 442
 443 /* --sort pid */
 444
 445 static int64_t
 446 sort__thread_cmp(struct hist_entry *left, struct hist_entry *right)
 447 {
 448         return right->thread->pid - left->thread->pid;
 449 }
 450
 451 static size_t
 452 sort__thread_print(FILE *fp, struct hist_entry *self)
 453 {
 454         return fprintf(fp, "%16s:%5d", self->thread->comm ?: "", self->thread->pid);
 455 }
 456
 457 static struct sort_entry sort_thread = {
 458         .header = "         Command:  Pid",
 459         .cmp    = sort__thread_cmp,
 460         .print  = sort__thread_print,
 461 };
 462
 463 /* --sort comm */
 464
 465 static int64_t
 466 sort__comm_cmp(struct hist_entry *left, struct hist_entry *right)
 467 {
 468         return right->thread->pid - left->thread->pid;
 469 }
 470
 471 static int64_t
 472 sort__comm_collapse(struct hist_entry *left, struct hist_entry *right)
 473 {
 474         char *comm_l = left->thread->comm;
 475         char *comm_r = right->thread->comm;
 476
 477         if (!comm_l || !comm_r) {
 478                 if (!comm_l && !comm_r)
 479                         return 0;
 480                 else if (!comm_l)
 481                         return -1;
 482                 else
 483                         return 1;
 484         }
 485
 486         return strcmp(comm_l, comm_r);
 487 }
 488
 489 static size_t
 490 sort__comm_print(FILE *fp, struct hist_entry *self)
 491 {
 492         return fprintf(fp, "%16s", self->thread->comm);
 493 }
 494
 495 static struct sort_entry sort_comm = {
 496         .header         = "         Command",
 497         .cmp            = sort__comm_cmp,
 498         .collapse       = sort__comm_collapse,
 499         .print          = sort__comm_print,
 500 };
 501
 502 /* --sort dso */
 503
 504 static int64_t
 505 sort__dso_cmp(struct hist_entry *left, struct hist_entry *right)
 506 {
 507         struct dso *dso_l = left->dso;
 508         struct dso *dso_r = right->dso;
 509
 510         if (!dso_l || !dso_r) {
 511                 if (!dso_l && !dso_r)
 512                         return 0;
 513                 else if (!dso_l)
 514                         return -1;
 515                 else
 516                         return 1;
 517         }
 518
 519         return strcmp(dso_l->name, dso_r->name);
 520 }
 521
 522 static size_t
 523 sort__dso_print(FILE *fp, struct hist_entry *self)
 524 {
 525         if (self->dso)
 526                 return fprintf(fp, "%-25s", self->dso->name);
 527
 528         return fprintf(fp, "%016llx         ", (u64)self->ip);
 529 }
 530
 531 static struct sort_entry sort_dso = {
 532         .header = "Shared Object            ",
 533         .cmp    = sort__dso_cmp,
 534         .print  = sort__dso_print,
 535 };
 536
 537 /* --sort symbol */
 538
 539 static int64_t
 540 sort__sym_cmp(struct hist_entry *left, struct hist_entry *right)
 541 {
 542         u64 ip_l, ip_r;
 543
 544         if (left->sym == right->sym)
 545                 return 0;
 546
 547         ip_l = left->sym ? left->sym->start : left->ip;
 548         ip_r = right->sym ? right->sym->start : right->ip;
 549
 550         return (int64_t)(ip_r - ip_l);
 551 }
 552
 553 static size_t
 554 sort__sym_print(FILE *fp, struct hist_entry *self)
 555 {
 556         size_t ret = 0;
 557
 558         if (verbose)
 559                 ret += fprintf(fp, "%#018llx  ", (u64)self->ip);
 560
 561         if (self->sym) {
 562                 ret += fprintf(fp, "[%c] %s",
 563                         self->dso == kernel_dso ? 'k' : '.', self->sym->name);
 564         } else {
 565                 ret += fprintf(fp, "%#016llx", (u64)self->ip);
 566         }
 567
 568         return ret;
 569 }
 570
 571 static struct sort_entry sort_sym = {
 572         .header = "Symbol",
 573         .cmp    = sort__sym_cmp,
 574         .print  = sort__sym_print,
 575 };
 576
 577 static int sort__need_collapse = 0;
 578
 579 struct sort_dimension {
 580         char                    *name;
 581         struct sort_entry       *entry;
 582         int                     taken;
 583 };
 584
 585 static struct sort_dimension sort_dimensions[] = {
 586         { .name = "pid",        .entry = &sort_thread,  },
 587         { .name = "comm",       .entry = &sort_comm,    },
 588         { .name = "dso",        .entry = &sort_dso,     },
 589         { .name = "symbol",     .entry = &sort_sym,     },
 590 };
 591
 592 static LIST_HEAD(hist_entry__sort_list);
 593
 594 static int sort_dimension__add(char *tok)
 595 {
 596         unsigned int i;
 597
 598         for (i = 0; i < ARRAY_SIZE(sort_dimensions); i++) {
 599                 struct sort_dimension *sd = &sort_dimensions[i];
 600
 601                 if (sd->taken)
 602                         continue;
 603
 604                 if (strncasecmp(tok, sd->name, strlen(tok)))
 605                         continue;
 606
 607                 if (sd->entry->collapse)
 608                         sort__need_collapse = 1;
 609
 610                 list_add_tail(&sd->entry->list, &hist_entry__sort_list);
 611                 sd->taken = 1;
 612
 613                 return 0;
 614         }
 615
 616         return -ESRCH;
 617 }
 618
 619 static int64_t
 620 hist_entry__cmp(struct hist_entry *left, struct hist_entry *right)
 621 {
 622         struct sort_entry *se;
 623         int64_t cmp = 0;
 624
 625         list_for_each_entry(se, &hist_entry__sort_list, list) {
 626                 cmp = se->cmp(left, right);
 627                 if (cmp)
 628                         break;
 629         }
 630
 631         return cmp;
 632 }
 633
 634 static int64_t
 635 hist_entry__collapse(struct hist_entry *left, struct hist_entry *right)
 636 {
 637         struct sort_entry *se;
 638         int64_t cmp = 0;
 639
 640         list_for_each_entry(se, &hist_entry__sort_list, list) {
 641                 int64_t (*f)(struct hist_entry *, struct hist_entry *);
 642
 643                 f = se->collapse ?: se->cmp;
 644
 645                 cmp = f(left, right);
 646                 if (cmp)
 647                         break;
 648         }
 649
 650         return cmp;
 651 }
 652
 653 /*
 654  * collect histogram counts
 655  */
 656 static void hist_hit(struct hist_entry *he, u64 ip)
 657 {
 658         unsigned int sym_size, offset;
 659         struct symbol *sym = he->sym;
 660
 661         he->count++;
 662
 663         if (!sym || !sym->hist)
 664                 return;
 665
 666         sym_size = sym->end - sym->start;
 667         offset = ip - sym->start;
 668
 669         if (offset >= sym_size)
 670                 return;
 671
 672         sym->hist_sum++;
 673         sym->hist[offset]++;
 674
 675         if (verbose >= 3)
 676                 printf("%p %s: count++ [ip: %p, %08Lx] => %Ld\n",
 677                         (void *)(unsigned long)he->sym->start,
 678                         he->sym->name,
 679                         (void *)(unsigned long)ip, ip - he->sym->start,
 680                         sym->hist[offset]);
 681 }
 682
 683 static int
 684 hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
 685                 struct symbol *sym, u64 ip, char level)
 686 {
 687         struct rb_node **p = &hist.rb_node;
 688         struct rb_node *parent = NULL;
 689         struct hist_entry *he;
 690         struct hist_entry entry = {
 691                 .thread = thread,
 692                 .map    = map,
 693                 .dso    = dso,
 694                 .sym    = sym,
 695                 .ip     = ip,
 696                 .level  = level,
 697                 .count  = 1,
 698         };
 699         int cmp;
 700
 701         while (*p != NULL) {
 702                 parent = *p;
 703                 he = rb_entry(parent, struct hist_entry, rb_node);
 704
 705                 cmp = hist_entry__cmp(&entry, he);
 706
 707                 if (!cmp) {
 708                         hist_hit(he, ip);
 709
 710                         return 0;
 711                 }
 712
 713                 if (cmp < 0)
 714                         p = &(*p)->rb_left;
 715                 else
 716                         p = &(*p)->rb_right;
 717         }
 718
 719         he = malloc(sizeof(*he));
 720         if (!he)
 721                 return -ENOMEM;
 722         *he = entry;
 723         rb_link_node(&he->rb_node, parent, p);
 724         rb_insert_color(&he->rb_node, &hist);
 725
 726         return 0;
 727 }
 728
 729 static void hist_entry__free(struct hist_entry *he)
 730 {
 731         free(he);
 732 }
 733
 734 /*
 735  * collapse the histogram
 736  */
 737
 738 static struct rb_root collapse_hists;
 739
 740 static void collapse__insert_entry(struct hist_entry *he)
 741 {
 742         struct rb_node **p = &collapse_hists.rb_node;
 743         struct rb_node *parent = NULL;
 744         struct hist_entry *iter;
 745         int64_t cmp;
 746
 747         while (*p != NULL) {
 748                 parent = *p;
 749                 iter = rb_entry(parent, struct hist_entry, rb_node);
 750
 751                 cmp = hist_entry__collapse(iter, he);
 752
 753                 if (!cmp) {
 754                         iter->count += he->count;
 755                         hist_entry__free(he);
 756                         return;
 757                 }
 758
 759                 if (cmp < 0)
 760                         p = &(*p)->rb_left;
 761                 else
 762                         p = &(*p)->rb_right;
 763         }
 764
 765         rb_link_node(&he->rb_node, parent, p);
 766         rb_insert_color(&he->rb_node, &collapse_hists);
 767 }
 768
 769 static void collapse__resort(void)
 770 {
 771         struct rb_node *next;
 772         struct hist_entry *n;
 773
 774         if (!sort__need_collapse)
 775                 return;
 776
 777         next = rb_first(&hist);
 778         while (next) {
 779                 n = rb_entry(next, struct hist_entry, rb_node);
 780                 next = rb_next(&n->rb_node);
 781
 782                 rb_erase(&n->rb_node, &hist);
 783                 collapse__insert_entry(n);
 784         }
 785 }
 786
 787 /*
 788  * reverse the map, sort on count.
 789  */
 790
 791 static struct rb_root output_hists;
 792
 793 static void output__insert_entry(struct hist_entry *he)
 794 {
 795         struct rb_node **p = &output_hists.rb_node;
 796         struct rb_node *parent = NULL;
 797         struct hist_entry *iter;
 798
 799         while (*p != NULL) {
 800                 parent = *p;
 801                 iter = rb_entry(parent, struct hist_entry, rb_node);
 802
 803                 if (he->count > iter->count)
 804                         p = &(*p)->rb_left;
 805                 else
 806                         p = &(*p)->rb_right;
 807         }
 808
 809         rb_link_node(&he->rb_node, parent, p);
 810         rb_insert_color(&he->rb_node, &output_hists);
 811 }
 812
 813 static void output__resort(void)
 814 {
 815         struct rb_node *next;
 816         struct hist_entry *n;
 817         struct rb_root *tree = &hist;
 818
 819         if (sort__need_collapse)
 820                 tree = &collapse_hists;
 821
 822         next = rb_first(tree);
 823
 824         while (next) {
 825                 n = rb_entry(next, struct hist_entry, rb_node);
 826                 next = rb_next(&n->rb_node);
 827
 828                 rb_erase(&n->rb_node, tree);
 829                 output__insert_entry(n);
 830         }
 831 }
 832
 833 static void register_idle_thread(void)
 834 {
 835         struct thread *thread = threads__findnew(0);
 836
 837         if (thread == NULL ||
 838                         thread__set_comm(thread, "[idle]")) {
 839                 fprintf(stderr, "problem inserting idle task.\n");
 840                 exit(-1);
 841         }
 842 }
 843
 844 static unsigned long total = 0,
 845                      total_mmap = 0,
 846                      total_comm = 0,
 847                      total_fork = 0,
 848                      total_unknown = 0;
 849
 850 static int
 851 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
 852 {
 853         char level;
 854         int show = 0;
 855         struct dso *dso = NULL;
 856         struct thread *thread = threads__findnew(event->ip.pid);
 857         u64 ip = event->ip.ip;
 858         struct map *map = NULL;
 859
 860         dprintf("%p [%p]: PERF_EVENT (IP, %d): %d: %p\n",
 861                 (void *)(offset + head),
 862                 (void *)(long)(event->header.size),
 863                 event->header.misc,
 864                 event->ip.pid,
 865                 (void *)(long)ip);
 866
 867         dprintf(" ... thread: %s:%d\n", thread->comm, thread->pid);
 868
 869         if (thread == NULL) {
 870                 fprintf(stderr, "problem processing %d event, skipping it.\n",
 871                         event->header.type);
 872                 return -1;
 873         }
 874
 875         if (event->header.misc & PERF_EVENT_MISC_KERNEL) {
 876                 show = SHOW_KERNEL;
 877                 level = 'k';
 878
 879                 dso = kernel_dso;
 880
 881                 dprintf(" ...... dso: %s\n", dso->name);
 882
 883         } else if (event->header.misc & PERF_EVENT_MISC_USER) {
 884
 885                 show = SHOW_USER;
 886                 level = '.';
 887
 888                 map = thread__find_map(thread, ip);
 889                 if (map != NULL) {
 890                         ip = map->map_ip(map, ip);
 891                         dso = map->dso;
 892                 } else {
 893                         /*
 894                          * If this is outside of all known maps,
 895                          * and is a negative address, try to look it
 896                          * up in the kernel dso, as it might be a
 897                          * vsyscall (which executes in user-mode):
 898                          */
 899                         if ((long long)ip < 0)
 900                                 dso = kernel_dso;
 901                 }
 902                 dprintf(" ...... dso: %s\n", dso ? dso->name : "<not found>");
 903
 904         } else {
 905                 show = SHOW_HV;
 906                 level = 'H';
 907                 dprintf(" ...... dso: [hypervisor]\n");
 908         }
 909
 910         if (show & show_mask) {
 911                 struct symbol *sym = NULL;
 912
 913                 if (dso)
 914                         sym = dso->find_symbol(dso, ip);
 915
 916                 if (hist_entry__add(thread, map, dso, sym, ip, level)) {
 917                         fprintf(stderr,
 918                 "problem incrementing symbol count, skipping event\n");
 919                         return -1;
 920                 }
 921         }
 922         total++;
 923
 924         return 0;
 925 }
 926
 927 static int
 928 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
 929 {
 930         struct thread *thread = threads__findnew(event->mmap.pid);
 931         struct map *map = map__new(&event->mmap);
 932
 933         dprintf("%p [%p]: PERF_EVENT_MMAP %d: [%p(%p) @ %p]: %s\n",
 934                 (void *)(offset + head),
 935                 (void *)(long)(event->header.size),
 936                 event->mmap.pid,
 937                 (void *)(long)event->mmap.start,
 938                 (void *)(long)event->mmap.len,
 939                 (void *)(long)event->mmap.pgoff,
 940                 event->mmap.filename);
 941
 942         if (thread == NULL || map == NULL) {
 943                 dprintf("problem processing PERF_EVENT_MMAP, skipping event.\n");
 944                 return 0;
 945         }
 946
 947         thread__insert_map(thread, map);
 948         total_mmap++;
 949
 950         return 0;
 951 }
 952
 953 static int
 954 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
 955 {
 956         struct thread *thread = threads__findnew(event->comm.pid);
 957
 958         dprintf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
 959                 (void *)(offset + head),
 960                 (void *)(long)(event->header.size),
 961                 event->comm.comm, event->comm.pid);
 962
 963         if (thread == NULL ||
 964             thread__set_comm(thread, event->comm.comm)) {
 965                 dprintf("problem processing PERF_EVENT_COMM, skipping event.\n");
 966                 return -1;
 967         }
 968         total_comm++;
 969
 970         return 0;
 971 }
 972
 973 static int
 974 process_fork_event(event_t *event, unsigned long offset, unsigned long head)
 975 {
 976         struct thread *thread = threads__findnew(event->fork.pid);
 977         struct thread *parent = threads__findnew(event->fork.ppid);
 978
 979         dprintf("%p [%p]: PERF_EVENT_FORK: %d:%d\n",
 980                 (void *)(offset + head),
 981                 (void *)(long)(event->header.size),
 982                 event->fork.pid, event->fork.ppid);
 983
 984         /*
 985          * A thread clone will have the same PID for both
 986          * parent and child.
 987          */
 988         if (thread == parent)
 989                 return 0;
 990
 991         if (!thread || !parent || thread__fork(thread, parent)) {
 992                 dprintf("problem processing PERF_EVENT_FORK, skipping event.\n");
 993                 return -1;
 994         }
 995         total_fork++;
 996
 997         return 0;
 998 }
 999
1000 static int
1001 process_event(event_t *event, unsigned long offset, unsigned long head)
1002 {
1003         switch (event->header.type) {
1004         case PERF_EVENT_SAMPLE:
1005                 return process_sample_event(event, offset, head);
1006
1007         case PERF_EVENT_MMAP:
1008                 return process_mmap_event(event, offset, head);
1009
1010         case PERF_EVENT_COMM:
1011                 return process_comm_event(event, offset, head);
1012
1013         case PERF_EVENT_FORK:
1014                 return process_fork_event(event, offset, head);
1015         /*
1016          * We dont process them right now but they are fine:
1017          */
1018
1019         case PERF_EVENT_THROTTLE:
1020         case PERF_EVENT_UNTHROTTLE:
1021                 return 0;
1022
1023         default:
1024                 return -1;
1025         }
1026
1027         return 0;
1028 }
1029
1030 static int
1031 parse_line(FILE *file, struct symbol *sym, u64 start, u64 len)
1032 {
1033         char *line = NULL, *tmp, *tmp2;
1034         static const char *prev_line;
1035         static const char *prev_color;
1036         unsigned int offset;
1037         size_t line_len;
1038         s64 line_ip;
1039         int ret;
1040         char *c;
1041
1042         if (getline(&line, &line_len, file) < 0)
1043                 return -1;
1044         if (!line)
1045                 return -1;
1046
1047         c = strchr(line, '\n');
1048         if (c)
1049                 *c = 0;
1050
1051         line_ip = -1;
1052         offset = 0;
1053         ret = -2;
1054
1055         /*
1056          * Strip leading spaces:
1057          */
1058         tmp = line;
1059         while (*tmp) {
1060                 if (*tmp != ' ')
1061                         break;
1062                 tmp++;
1063         }
1064
1065         if (*tmp) {
1066                 /*
1067                  * Parse hexa addresses followed by ':'
1068                  */
1069                 line_ip = strtoull(tmp, &tmp2, 16);
1070                 if (*tmp2 != ':')
1071                         line_ip = -1;
1072         }
1073
1074         if (line_ip != -1) {
1075                 const char *path = NULL;
1076                 unsigned int hits = 0;
1077                 double percent = 0.0;
1078                 char *color;
1079                 struct sym_ext *sym_ext = sym->priv;
1080
1081                 offset = line_ip - start;
1082                 if (offset < len)
1083                         hits = sym->hist[offset];
1084
1085                 if (offset < len && sym_ext) {
1086                         path = sym_ext[offset].path;
1087                         percent = sym_ext[offset].percent;
1088                 } else if (sym->hist_sum)
1089                         percent = 100.0 * hits / sym->hist_sum;
1090
1091                 color = get_percent_color(percent);
1092
1093                 /*
1094                  * Also color the filename and line if needed, with
1095                  * the same color than the percentage. Don't print it
1096                  * twice for close colored ip with the same filename:line
1097                  */
1098                 if (path) {
1099                         if (!prev_line || strcmp(prev_line, path)
1100                                        || color != prev_color) {
1101                                 color_fprintf(stdout, color, " %s", path);
1102                                 prev_line = path;
1103                                 prev_color = color;
1104                         }
1105                 }
1106
1107                 color_fprintf(stdout, color, " %7.2f", percent);
1108                 printf(" :      ");
1109                 color_fprintf(stdout, PERF_COLOR_BLUE, "%s\n", line);
1110         } else {
1111                 if (!*line)
1112                         printf("         :\n");
1113                 else
1114                         printf("         :      %s\n", line);
1115         }
1116
1117         return 0;
1118 }
1119
1120 static struct rb_root root_sym_ext;
1121
1122 static void insert_source_line(struct sym_ext *sym_ext)
1123 {
1124         struct sym_ext *iter;
1125         struct rb_node **p = &root_sym_ext.rb_node;
1126         struct rb_node *parent = NULL;
1127
1128         while (*p != NULL) {
1129                 parent = *p;
1130                 iter = rb_entry(parent, struct sym_ext, node);
1131
1132                 if (sym_ext->percent > iter->percent)
1133                         p = &(*p)->rb_left;
1134                 else
1135                         p = &(*p)->rb_right;
1136         }
1137
1138         rb_link_node(&sym_ext->node, parent, p);
1139         rb_insert_color(&sym_ext->node, &root_sym_ext);
1140 }
1141
1142 static void free_source_line(struct symbol *sym, int len)
1143 {
1144         struct sym_ext *sym_ext = sym->priv;
1145         int i;
1146
1147         if (!sym_ext)
1148                 return;
1149
1150         for (i = 0; i < len; i++)
1151                 free(sym_ext[i].path);
1152         free(sym_ext);
1153
1154         sym->priv = NULL;
1155         root_sym_ext = RB_ROOT;
1156 }
1157
1158 /* Get the filename:line for the colored entries */
1159 static void
1160 get_source_line(struct symbol *sym, u64 start, int len, char *filename)
1161 {
1162         int i;
1163         char cmd[PATH_MAX * 2];
1164         struct sym_ext *sym_ext;
1165
1166         if (!sym->hist_sum)
1167                 return;
1168
1169         sym->priv = calloc(len, sizeof(struct sym_ext));
1170         if (!sym->priv)
1171                 return;
1172
1173         sym_ext = sym->priv;
1174
1175         for (i = 0; i < len; i++) {
1176                 char *path = NULL;
1177                 size_t line_len;
1178                 u64 offset;
1179                 FILE *fp;
1180
1181                 sym_ext[i].percent = 100.0 * sym->hist[i] / sym->hist_sum;
1182                 if (sym_ext[i].percent <= 0.5)
1183                         continue;
1184
1185                 offset = start + i;
1186                 sprintf(cmd, "addr2line -e %s %016llx", filename, offset);
1187                 fp = popen(cmd, "r");
1188                 if (!fp)
1189                         continue;
1190
1191                 if (getline(&path, &line_len, fp) < 0 || !line_len)
1192                         goto next;
1193
1194                 sym_ext[i].path = malloc(sizeof(char) * line_len + 1);
1195                 if (!sym_ext[i].path)
1196                         goto next;
1197
1198                 strcpy(sym_ext[i].path, path);
1199                 insert_source_line(&sym_ext[i]);
1200
1201         next:
1202                 pclose(fp);
1203         }
1204 }
1205
1206 static void print_summary(char *filename)
1207 {
1208         struct sym_ext *sym_ext;
1209         struct rb_node *node;
1210
1211         printf("\nSorted summary for file %s\n", filename);
1212         printf("----------------------------------------------\n\n");
1213
1214         if (RB_EMPTY_ROOT(&root_sym_ext)) {
1215                 printf(" Nothing higher than %1.1f%%\n", MIN_GREEN);
1216                 return;
1217         }
1218
1219         node = rb_first(&root_sym_ext);
1220         while (node) {
1221                 double percent;
1222                 char *color;
1223                 char *path;
1224
1225                 sym_ext = rb_entry(node, struct sym_ext, node);
1226                 percent = sym_ext->percent;
1227                 color = get_percent_color(percent);
1228                 path = sym_ext->path;
1229
1230                 color_fprintf(stdout, color, " %7.2f %s", percent, path);
1231                 node = rb_next(node);
1232         }
1233 }
1234
1235 static void annotate_sym(struct dso *dso, struct symbol *sym)
1236 {
1237         char *filename = dso->name, *d_filename;
1238         u64 start, end, len;
1239         char command[PATH_MAX*2];
1240         FILE *file;
1241
1242         if (!filename)
1243                 return;
1244         if (sym->module)
1245                 filename = sym->module->path;
1246         else if (dso == kernel_dso)
1247                 filename = vmlinux;
1248
1249         start = sym->obj_start;
1250         if (!start)
1251                 start = sym->start;
1252         if (full_paths)
1253                 d_filename = filename;
1254         else
1255                 d_filename = basename(filename);
1256
1257         end = start + sym->end - sym->start + 1;
1258         len = sym->end - sym->start;
1259
1260         if (print_line) {
1261                 get_source_line(sym, start, len, filename);
1262                 print_summary(filename);
1263         }
1264
1265         printf("\n\n------------------------------------------------\n");
1266         printf(" Percent |      Source code & Disassembly of %s\n", d_filename);
1267         printf("------------------------------------------------\n");
1268
1269         if (verbose >= 2)
1270                 printf("annotating [%p] %30s : [%p] %30s\n", dso, dso->name, sym, sym->name);
1271
1272         sprintf(command, "objdump --start-address=0x%016Lx --stop-address=0x%016Lx -dS %s|grep -v %s",
1273                         (u64)start, (u64)end, filename, filename);
1274
1275         if (verbose >= 3)
1276                 printf("doing: %s\n", command);
1277
1278         file = popen(command, "r");
1279         if (!file)
1280                 return;
1281
1282         while (!feof(file)) {
1283                 if (parse_line(file, sym, start, len) < 0)
1284                         break;
1285         }
1286
1287         pclose(file);
1288         if (print_line)
1289                 free_source_line(sym, len);
1290 }
1291
1292 static void find_annotations(void)
1293 {
1294         struct rb_node *nd;
1295         struct dso *dso;
1296         int count = 0;
1297
1298         list_for_each_entry(dso, &dsos, node) {
1299
1300                 for (nd = rb_first(&dso->syms); nd; nd = rb_next(nd)) {
1301                         struct symbol *sym = rb_entry(nd, struct symbol, rb_node);
1302
1303                         if (sym->hist) {
1304                                 annotate_sym(dso, sym);
1305                                 count++;
1306                         }
1307                 }
1308         }
1309
1310         if (!count)
1311                 printf(" Error: symbol '%s' not present amongst the samples.\n", sym_hist_filter);
1312 }
1313
1314 static int __cmd_annotate(void)
1315 {
1316         int ret, rc = EXIT_FAILURE;
1317         unsigned long offset = 0;
1318         unsigned long head = 0;
1319         struct stat stat;
1320         event_t *event;
1321         uint32_t size;
1322         char *buf;
1323
1324         register_idle_thread();
1325
1326         input = open(input_name, O_RDONLY);
1327         if (input < 0) {
1328                 perror("failed to open file");
1329                 exit(-1);
1330         }
1331
1332         ret = fstat(input, &stat);
1333         if (ret < 0) {
1334                 perror("failed to stat file");
1335                 exit(-1);
1336         }
1337
1338         if (!force && (stat.st_uid != geteuid())) {
1339                 fprintf(stderr, "file: %s not owned by current user\n", input_name);
1340                 exit(-1);
1341         }
1342
1343         if (!stat.st_size) {
1344                 fprintf(stderr, "zero-sized file, nothing to do!\n");
1345                 exit(0);
1346         }
1347
1348         if (load_kernel() < 0) {
1349                 perror("failed to load kernel symbols");
1350                 return EXIT_FAILURE;
1351         }
1352
1353 remap:
1354         buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1355                            MAP_SHARED, input, offset);
1356         if (buf == MAP_FAILED) {
1357                 perror("failed to mmap file");
1358                 exit(-1);
1359         }
1360
1361 more:
1362         event = (event_t *)(buf + head);
1363
1364         size = event->header.size;
1365         if (!size)
1366                 size = 8;
1367
1368         if (head + event->header.size >= page_size * mmap_window) {
1369                 unsigned long shift = page_size * (head / page_size);
1370                 int ret;
1371
1372                 ret = munmap(buf, page_size * mmap_window);
1373                 assert(ret == 0);
1374
1375                 offset += shift;
1376                 head -= shift;
1377                 goto remap;
1378         }
1379
1380         size = event->header.size;
1381
1382         dprintf("%p [%p]: event: %d\n",
1383                         (void *)(offset + head),
1384                         (void *)(long)event->header.size,
1385                         event->header.type);
1386
1387         if (!size || process_event(event, offset, head) < 0) {
1388
1389                 dprintf("%p [%p]: skipping unknown header type: %d\n",
1390                         (void *)(offset + head),
1391                         (void *)(long)(event->header.size),
1392                         event->header.type);
1393
1394                 total_unknown++;
1395
1396                 /*
1397                  * assume we lost track of the stream, check alignment, and
1398                  * increment a single u64 in the hope to catch on again 'soon'.
1399                  */
1400
1401                 if (unlikely(head & 7))
1402                         head &= ~7ULL;
1403
1404                 size = 8;
1405         }
1406
1407         head += size;
1408
1409         if (offset + head < (unsigned long)stat.st_size)
1410                 goto more;
1411
1412         rc = EXIT_SUCCESS;
1413         close(input);
1414
1415         dprintf("      IP events: %10ld\n", total);
1416         dprintf("    mmap events: %10ld\n", total_mmap);
1417         dprintf("    comm events: %10ld\n", total_comm);
1418         dprintf("    fork events: %10ld\n", total_fork);
1419         dprintf(" unknown events: %10ld\n", total_unknown);
1420
1421         if (dump_trace)
1422                 return 0;
1423
1424         if (verbose >= 3)
1425                 threads__fprintf(stdout);
1426
1427         if (verbose >= 2)
1428                 dsos__fprintf(stdout);
1429
1430         collapse__resort();
1431         output__resort();
1432
1433         find_annotations();
1434
1435         return rc;
1436 }
1437
1438 static const char * const annotate_usage[] = {
1439         "perf annotate [<options>] <command>",
1440         NULL
1441 };
1442
1443 static const struct option options[] = {
1444         OPT_STRING('i', "input", &input_name, "file",
1445                     "input file name"),
1446         OPT_STRING('s', "symbol", &sym_hist_filter, "symbol",
1447                     "symbol to annotate"),
1448         OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
1449         OPT_BOOLEAN('v', "verbose", &verbose,
1450                     "be more verbose (show symbol address, etc)"),
1451         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1452                     "dump raw trace in ASCII"),
1453         OPT_STRING('k', "vmlinux", &vmlinux, "file", "vmlinux pathname"),
1454         OPT_BOOLEAN('m', "modules", &modules,
1455                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
1456         OPT_BOOLEAN('l', "print-line", &print_line,
1457                     "print matching source lines (may be slow)"),
1458         OPT_BOOLEAN('P', "full-paths", &full_paths,
1459                     "Don't shorten the displayed pathnames"),
1460         OPT_END()
1461 };
1462
1463 static void setup_sorting(void)
1464 {
1465         char *tmp, *tok, *str = strdup(sort_order);
1466
1467         for (tok = strtok_r(str, ", ", &tmp);
1468                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1469                 if (sort_dimension__add(tok) < 0) {
1470                         error("Unknown --sort key: `%s'", tok);
1471                         usage_with_options(annotate_usage, options);
1472                 }
1473         }
1474
1475         free(str);
1476 }
1477
1478 int cmd_annotate(int argc, const char **argv, const char *prefix __used)
1479 {
1480         symbol__init();
1481
1482         page_size = getpagesize();
1483
1484         argc = parse_options(argc, argv, options, annotate_usage, 0);
1485
1486         setup_sorting();
1487
1488         if (argc) {
1489                 /*
1490                  * Special case: if there's an argument left then assume tha
1491                  * it's a symbol filter:
1492                  */
1493                 if (argc > 1)
1494                         usage_with_options(annotate_usage, options);
1495
1496                 sym_hist_filter = argv[0];
1497         }
1498
1499         if (!sym_hist_filter)
1500                 usage_with_options(annotate_usage, options);
1501
1502         setup_pager();
1503
1504         return __cmd_annotate();
1505 }