SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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
  30 static char             default_sort_order[] = "comm,symbol";
  31 static char             *sort_order = default_sort_order;
  32
  33 static int              input;
  34 static int              show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV;
  35
  36 static int              dump_trace = 0;
  37 #define dprintf(x...)   do { if (dump_trace) printf(x); } while (0)
  38
  39
  40 static int              full_paths;
  41
  42 static int              print_line;
  43
  44 static unsigned long    page_size;
  45 static unsigned long    mmap_window = 32;
  46
  47
  48 struct sym_ext {
  49         struct rb_node  node;
  50         double          percent;
  51         char            *path;
  52 };
  53
  54 struct map {
  55         struct list_head node;
  56         u64      start;
  57         u64      end;
  58         u64      pgoff;
  59         u64      (*map_ip)(struct map *, u64);
  60         struct dso       *dso;
  61 };
  62
  63 static u64 map__map_ip(struct map *map, u64 ip)
  64 {
  65         return ip - map->start + map->pgoff;
  66 }
  67
  68 static u64 vdso__map_ip(struct map *map __used, u64 ip)
  69 {
  70         return ip;
  71 }
  72
  73 static struct map *map__new(struct mmap_event *event)
  74 {
  75         struct map *self = malloc(sizeof(*self));
  76
  77         if (self != NULL) {
  78                 const char *filename = event->filename;
  79
  80                 self->start = event->start;
  81                 self->end   = event->start + event->len;
  82                 self->pgoff = event->pgoff;
  83
  84                 self->dso = dsos__findnew(filename);
  85                 if (self->dso == NULL)
  86                         goto out_delete;
  87
  88                 if (self->dso == vdso)
  89                         self->map_ip = vdso__map_ip;
  90                 else
  91                         self->map_ip = map__map_ip;
  92         }
  93         return self;
  94 out_delete:
  95         free(self);
  96         return NULL;
  97 }
  98
  99 static struct map *map__clone(struct map *self)
 100 {
 101         struct map *map = malloc(sizeof(*self));
 102
 103         if (!map)
 104                 return NULL;
 105
 106         memcpy(map, self, sizeof(*self));
 107
 108         return map;
 109 }
 110
 111 static int map__overlap(struct map *l, struct map *r)
 112 {
 113         if (l->start > r->start) {
 114                 struct map *t = l;
 115                 l = r;
 116                 r = t;
 117         }
 118
 119         if (l->end > r->start)
 120                 return 1;
 121
 122         return 0;
 123 }
 124
 125 static size_t map__fprintf(struct map *self, FILE *fp)
 126 {
 127         return fprintf(fp, " %Lx-%Lx %Lx %s\n",
 128                        self->start, self->end, self->pgoff, self->dso->name);
 129 }
 130
 131
 132 struct thread {
 133         struct rb_node   rb_node;
 134         struct list_head maps;
 135         pid_t            pid;
 136         char             *comm;
 137 };
 138
 139 static struct thread *thread__new(pid_t pid)
 140 {
 141         struct thread *self = malloc(sizeof(*self));
 142
 143         if (self != NULL) {
 144                 self->pid = pid;
 145                 self->comm = malloc(32);
 146                 if (self->comm)
 147                         snprintf(self->comm, 32, ":%d", self->pid);
 148                 INIT_LIST_HEAD(&self->maps);
 149         }
 150
 151         return self;
 152 }
 153
 154 static int thread__set_comm(struct thread *self, const char *comm)
 155 {
 156         if (self->comm)
 157                 free(self->comm);
 158         self->comm = strdup(comm);
 159         return self->comm ? 0 : -ENOMEM;
 160 }
 161
 162 static size_t thread__fprintf(struct thread *self, FILE *fp)
 163 {
 164         struct map *pos;
 165         size_t ret = fprintf(fp, "Thread %d %s\n", self->pid, self->comm);
 166
 167         list_for_each_entry(pos, &self->maps, node)
 168                 ret += map__fprintf(pos, fp);
 169
 170         return ret;
 171 }
 172
 173
 174 static struct rb_root threads;
 175 static struct thread *last_match;
 176
 177 static struct thread *threads__findnew(pid_t pid)
 178 {
 179         struct rb_node **p = &threads.rb_node;
 180         struct rb_node *parent = NULL;
 181         struct thread *th;
 182
 183         /*
 184          * Font-end cache - PID lookups come in blocks,
 185          * so most of the time we dont have to look up
 186          * the full rbtree:
 187          */
 188         if (last_match && last_match->pid == pid)
 189                 return last_match;
 190
 191         while (*p != NULL) {
 192                 parent = *p;
 193                 th = rb_entry(parent, struct thread, rb_node);
 194
 195                 if (th->pid == pid) {
 196                         last_match = th;
 197                         return th;
 198                 }
 199
 200                 if (pid < th->pid)
 201                         p = &(*p)->rb_left;
 202                 else
 203                         p = &(*p)->rb_right;
 204         }
 205
 206         th = thread__new(pid);
 207         if (th != NULL) {
 208                 rb_link_node(&th->rb_node, parent, p);
 209                 rb_insert_color(&th->rb_node, &threads);
 210                 last_match = th;
 211         }
 212
 213         return th;
 214 }
 215
 216 static void thread__insert_map(struct thread *self, struct map *map)
 217 {
 218         struct map *pos, *tmp;
 219
 220         list_for_each_entry_safe(pos, tmp, &self->maps, node) {
 221                 if (map__overlap(pos, map)) {
 222                         list_del_init(&pos->node);
 223                         /* XXX leaks dsos */
 224                         free(pos);
 225                 }
 226         }
 227
 228         list_add_tail(&map->node, &self->maps);
 229 }
 230
 231 static int thread__fork(struct thread *self, struct thread *parent)
 232 {
 233         struct map *map;
 234
 235         if (self->comm)
 236                 free(self->comm);
 237         self->comm = strdup(parent->comm);
 238         if (!self->comm)
 239                 return -ENOMEM;
 240
 241         list_for_each_entry(map, &parent->maps, node) {
 242                 struct map *new = map__clone(map);
 243                 if (!new)
 244                         return -ENOMEM;
 245                 thread__insert_map(self, new);
 246         }
 247
 248         return 0;
 249 }
 250
 251 static struct map *thread__find_map(struct thread *self, u64 ip)
 252 {
 253         struct map *pos;
 254
 255         if (self == NULL)
 256                 return NULL;
 257
 258         list_for_each_entry(pos, &self->maps, node)
 259                 if (ip >= pos->start && ip <= pos->end)
 260                         return pos;
 261
 262         return NULL;
 263 }
 264
 265 static size_t threads__fprintf(FILE *fp)
 266 {
 267         size_t ret = 0;
 268         struct rb_node *nd;
 269
 270         for (nd = rb_first(&threads); nd; nd = rb_next(nd)) {
 271                 struct thread *pos = rb_entry(nd, struct thread, rb_node);
 272
 273                 ret += thread__fprintf(pos, fp);
 274         }
 275
 276         return ret;
 277 }
 278
 279 /*
 280  * histogram, sorted on item, collects counts
 281  */
 282
 283 static struct rb_root hist;
 284
 285 struct hist_entry {
 286         struct rb_node   rb_node;
 287
 288         struct thread    *thread;
 289         struct map       *map;
 290         struct dso       *dso;
 291         struct symbol    *sym;
 292         u64      ip;
 293         char             level;
 294
 295         uint32_t         count;
 296 };
 297
 298 /*
 299  * configurable sorting bits
 300  */
 301
 302 struct sort_entry {
 303         struct list_head list;
 304
 305         char *header;
 306
 307         int64_t (*cmp)(struct hist_entry *, struct hist_entry *);
 308         int64_t (*collapse)(struct hist_entry *, struct hist_entry *);
 309         size_t  (*print)(FILE *fp, struct hist_entry *);
 310 };
 311
 312 /* --sort pid */
 313
 314 static int64_t
 315 sort__thread_cmp(struct hist_entry *left, struct hist_entry *right)
 316 {
 317         return right->thread->pid - left->thread->pid;
 318 }
 319
 320 static size_t
 321 sort__thread_print(FILE *fp, struct hist_entry *self)
 322 {
 323         return fprintf(fp, "%16s:%5d", self->thread->comm ?: "", self->thread->pid);
 324 }
 325
 326 static struct sort_entry sort_thread = {
 327         .header = "         Command:  Pid",
 328         .cmp    = sort__thread_cmp,
 329         .print  = sort__thread_print,
 330 };
 331
 332 /* --sort comm */
 333
 334 static int64_t
 335 sort__comm_cmp(struct hist_entry *left, struct hist_entry *right)
 336 {
 337         return right->thread->pid - left->thread->pid;
 338 }
 339
 340 static int64_t
 341 sort__comm_collapse(struct hist_entry *left, struct hist_entry *right)
 342 {
 343         char *comm_l = left->thread->comm;
 344         char *comm_r = right->thread->comm;
 345
 346         if (!comm_l || !comm_r) {
 347                 if (!comm_l && !comm_r)
 348                         return 0;
 349                 else if (!comm_l)
 350                         return -1;
 351                 else
 352                         return 1;
 353         }
 354
 355         return strcmp(comm_l, comm_r);
 356 }
 357
 358 static size_t
 359 sort__comm_print(FILE *fp, struct hist_entry *self)
 360 {
 361         return fprintf(fp, "%16s", self->thread->comm);
 362 }
 363
 364 static struct sort_entry sort_comm = {
 365         .header         = "         Command",
 366         .cmp            = sort__comm_cmp,
 367         .collapse       = sort__comm_collapse,
 368         .print          = sort__comm_print,
 369 };
 370
 371 /* --sort dso */
 372
 373 static int64_t
 374 sort__dso_cmp(struct hist_entry *left, struct hist_entry *right)
 375 {
 376         struct dso *dso_l = left->dso;
 377         struct dso *dso_r = right->dso;
 378
 379         if (!dso_l || !dso_r) {
 380                 if (!dso_l && !dso_r)
 381                         return 0;
 382                 else if (!dso_l)
 383                         return -1;
 384                 else
 385                         return 1;
 386         }
 387
 388         return strcmp(dso_l->name, dso_r->name);
 389 }
 390
 391 static size_t
 392 sort__dso_print(FILE *fp, struct hist_entry *self)
 393 {
 394         if (self->dso)
 395                 return fprintf(fp, "%-25s", self->dso->name);
 396
 397         return fprintf(fp, "%016llx         ", (u64)self->ip);
 398 }
 399
 400 static struct sort_entry sort_dso = {
 401         .header = "Shared Object            ",
 402         .cmp    = sort__dso_cmp,
 403         .print  = sort__dso_print,
 404 };
 405
 406 /* --sort symbol */
 407
 408 static int64_t
 409 sort__sym_cmp(struct hist_entry *left, struct hist_entry *right)
 410 {
 411         u64 ip_l, ip_r;
 412
 413         if (left->sym == right->sym)
 414                 return 0;
 415
 416         ip_l = left->sym ? left->sym->start : left->ip;
 417         ip_r = right->sym ? right->sym->start : right->ip;
 418
 419         return (int64_t)(ip_r - ip_l);
 420 }
 421
 422 static size_t
 423 sort__sym_print(FILE *fp, struct hist_entry *self)
 424 {
 425         size_t ret = 0;
 426
 427         if (verbose)
 428                 ret += fprintf(fp, "%#018llx  ", (u64)self->ip);
 429
 430         if (self->sym) {
 431                 ret += fprintf(fp, "[%c] %s",
 432                         self->dso == kernel_dso ? 'k' : '.', self->sym->name);
 433         } else {
 434                 ret += fprintf(fp, "%#016llx", (u64)self->ip);
 435         }
 436
 437         return ret;
 438 }
 439
 440 static struct sort_entry sort_sym = {
 441         .header = "Symbol",
 442         .cmp    = sort__sym_cmp,
 443         .print  = sort__sym_print,
 444 };
 445
 446 static int sort__need_collapse = 0;
 447
 448 struct sort_dimension {
 449         char                    *name;
 450         struct sort_entry       *entry;
 451         int                     taken;
 452 };
 453
 454 static struct sort_dimension sort_dimensions[] = {
 455         { .name = "pid",        .entry = &sort_thread,  },
 456         { .name = "comm",       .entry = &sort_comm,    },
 457         { .name = "dso",        .entry = &sort_dso,     },
 458         { .name = "symbol",     .entry = &sort_sym,     },
 459 };
 460
 461 static LIST_HEAD(hist_entry__sort_list);
 462
 463 static int sort_dimension__add(char *tok)
 464 {
 465         unsigned int i;
 466
 467         for (i = 0; i < ARRAY_SIZE(sort_dimensions); i++) {
 468                 struct sort_dimension *sd = &sort_dimensions[i];
 469
 470                 if (sd->taken)
 471                         continue;
 472
 473                 if (strncasecmp(tok, sd->name, strlen(tok)))
 474                         continue;
 475
 476                 if (sd->entry->collapse)
 477                         sort__need_collapse = 1;
 478
 479                 list_add_tail(&sd->entry->list, &hist_entry__sort_list);
 480                 sd->taken = 1;
 481
 482                 return 0;
 483         }
 484
 485         return -ESRCH;
 486 }
 487
 488 static int64_t
 489 hist_entry__cmp(struct hist_entry *left, struct hist_entry *right)
 490 {
 491         struct sort_entry *se;
 492         int64_t cmp = 0;
 493
 494         list_for_each_entry(se, &hist_entry__sort_list, list) {
 495                 cmp = se->cmp(left, right);
 496                 if (cmp)
 497                         break;
 498         }
 499
 500         return cmp;
 501 }
 502
 503 static int64_t
 504 hist_entry__collapse(struct hist_entry *left, struct hist_entry *right)
 505 {
 506         struct sort_entry *se;
 507         int64_t cmp = 0;
 508
 509         list_for_each_entry(se, &hist_entry__sort_list, list) {
 510                 int64_t (*f)(struct hist_entry *, struct hist_entry *);
 511
 512                 f = se->collapse ?: se->cmp;
 513
 514                 cmp = f(left, right);
 515                 if (cmp)
 516                         break;
 517         }
 518
 519         return cmp;
 520 }
 521
 522 /*
 523  * collect histogram counts
 524  */
 525 static void hist_hit(struct hist_entry *he, u64 ip)
 526 {
 527         unsigned int sym_size, offset;
 528         struct symbol *sym = he->sym;
 529
 530         he->count++;
 531
 532         if (!sym || !sym->hist)
 533                 return;
 534
 535         sym_size = sym->end - sym->start;
 536         offset = ip - sym->start;
 537
 538         if (offset >= sym_size)
 539                 return;
 540
 541         sym->hist_sum++;
 542         sym->hist[offset]++;
 543
 544         if (verbose >= 3)
 545                 printf("%p %s: count++ [ip: %p, %08Lx] => %Ld\n",
 546                         (void *)(unsigned long)he->sym->start,
 547                         he->sym->name,
 548                         (void *)(unsigned long)ip, ip - he->sym->start,
 549                         sym->hist[offset]);
 550 }
 551
 552 static int
 553 hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
 554                 struct symbol *sym, u64 ip, char level)
 555 {
 556         struct rb_node **p = &hist.rb_node;
 557         struct rb_node *parent = NULL;
 558         struct hist_entry *he;
 559         struct hist_entry entry = {
 560                 .thread = thread,
 561                 .map    = map,
 562                 .dso    = dso,
 563                 .sym    = sym,
 564                 .ip     = ip,
 565                 .level  = level,
 566                 .count  = 1,
 567         };
 568         int cmp;
 569
 570         while (*p != NULL) {
 571                 parent = *p;
 572                 he = rb_entry(parent, struct hist_entry, rb_node);
 573
 574                 cmp = hist_entry__cmp(&entry, he);
 575
 576                 if (!cmp) {
 577                         hist_hit(he, ip);
 578
 579                         return 0;
 580                 }
 581
 582                 if (cmp < 0)
 583                         p = &(*p)->rb_left;
 584                 else
 585                         p = &(*p)->rb_right;
 586         }
 587
 588         he = malloc(sizeof(*he));
 589         if (!he)
 590                 return -ENOMEM;
 591         *he = entry;
 592         rb_link_node(&he->rb_node, parent, p);
 593         rb_insert_color(&he->rb_node, &hist);
 594
 595         return 0;
 596 }
 597
 598 static void hist_entry__free(struct hist_entry *he)
 599 {
 600         free(he);
 601 }
 602
 603 /*
 604  * collapse the histogram
 605  */
 606
 607 static struct rb_root collapse_hists;
 608
 609 static void collapse__insert_entry(struct hist_entry *he)
 610 {
 611         struct rb_node **p = &collapse_hists.rb_node;
 612         struct rb_node *parent = NULL;
 613         struct hist_entry *iter;
 614         int64_t cmp;
 615
 616         while (*p != NULL) {
 617                 parent = *p;
 618                 iter = rb_entry(parent, struct hist_entry, rb_node);
 619
 620                 cmp = hist_entry__collapse(iter, he);
 621
 622                 if (!cmp) {
 623                         iter->count += he->count;
 624                         hist_entry__free(he);
 625                         return;
 626                 }
 627
 628                 if (cmp < 0)
 629                         p = &(*p)->rb_left;
 630                 else
 631                         p = &(*p)->rb_right;
 632         }
 633
 634         rb_link_node(&he->rb_node, parent, p);
 635         rb_insert_color(&he->rb_node, &collapse_hists);
 636 }
 637
 638 static void collapse__resort(void)
 639 {
 640         struct rb_node *next;
 641         struct hist_entry *n;
 642
 643         if (!sort__need_collapse)
 644                 return;
 645
 646         next = rb_first(&hist);
 647         while (next) {
 648                 n = rb_entry(next, struct hist_entry, rb_node);
 649                 next = rb_next(&n->rb_node);
 650
 651                 rb_erase(&n->rb_node, &hist);
 652                 collapse__insert_entry(n);
 653         }
 654 }
 655
 656 /*
 657  * reverse the map, sort on count.
 658  */
 659
 660 static struct rb_root output_hists;
 661
 662 static void output__insert_entry(struct hist_entry *he)
 663 {
 664         struct rb_node **p = &output_hists.rb_node;
 665         struct rb_node *parent = NULL;
 666         struct hist_entry *iter;
 667
 668         while (*p != NULL) {
 669                 parent = *p;
 670                 iter = rb_entry(parent, struct hist_entry, rb_node);
 671
 672                 if (he->count > iter->count)
 673                         p = &(*p)->rb_left;
 674                 else
 675                         p = &(*p)->rb_right;
 676         }
 677
 678         rb_link_node(&he->rb_node, parent, p);
 679         rb_insert_color(&he->rb_node, &output_hists);
 680 }
 681
 682 static void output__resort(void)
 683 {
 684         struct rb_node *next;
 685         struct hist_entry *n;
 686         struct rb_root *tree = &hist;
 687
 688         if (sort__need_collapse)
 689                 tree = &collapse_hists;
 690
 691         next = rb_first(tree);
 692
 693         while (next) {
 694                 n = rb_entry(next, struct hist_entry, rb_node);
 695                 next = rb_next(&n->rb_node);
 696
 697                 rb_erase(&n->rb_node, tree);
 698                 output__insert_entry(n);
 699         }
 700 }
 701
 702 static void register_idle_thread(void)
 703 {
 704         struct thread *thread = threads__findnew(0);
 705
 706         if (thread == NULL ||
 707                         thread__set_comm(thread, "[idle]")) {
 708                 fprintf(stderr, "problem inserting idle task.\n");
 709                 exit(-1);
 710         }
 711 }
 712
 713 static unsigned long total = 0,
 714                      total_mmap = 0,
 715                      total_comm = 0,
 716                      total_fork = 0,
 717                      total_unknown = 0;
 718
 719 static int
 720 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
 721 {
 722         char level;
 723         int show = 0;
 724         struct dso *dso = NULL;
 725         struct thread *thread = threads__findnew(event->ip.pid);
 726         u64 ip = event->ip.ip;
 727         struct map *map = NULL;
 728
 729         dprintf("%p [%p]: PERF_EVENT (IP, %d): %d: %p\n",
 730                 (void *)(offset + head),
 731                 (void *)(long)(event->header.size),
 732                 event->header.misc,
 733                 event->ip.pid,
 734                 (void *)(long)ip);
 735
 736         dprintf(" ... thread: %s:%d\n", thread->comm, thread->pid);
 737
 738         if (thread == NULL) {
 739                 fprintf(stderr, "problem processing %d event, skipping it.\n",
 740                         event->header.type);
 741                 return -1;
 742         }
 743
 744         if (event->header.misc & PERF_EVENT_MISC_KERNEL) {
 745                 show = SHOW_KERNEL;
 746                 level = 'k';
 747
 748                 dso = kernel_dso;
 749
 750                 dprintf(" ...... dso: %s\n", dso->name);
 751
 752         } else if (event->header.misc & PERF_EVENT_MISC_USER) {
 753
 754                 show = SHOW_USER;
 755                 level = '.';
 756
 757                 map = thread__find_map(thread, ip);
 758                 if (map != NULL) {
 759                         ip = map->map_ip(map, ip);
 760                         dso = map->dso;
 761                 } else {
 762                         /*
 763                          * If this is outside of all known maps,
 764                          * and is a negative address, try to look it
 765                          * up in the kernel dso, as it might be a
 766                          * vsyscall (which executes in user-mode):
 767                          */
 768                         if ((long long)ip < 0)
 769                                 dso = kernel_dso;
 770                 }
 771                 dprintf(" ...... dso: %s\n", dso ? dso->name : "<not found>");
 772
 773         } else {
 774                 show = SHOW_HV;
 775                 level = 'H';
 776                 dprintf(" ...... dso: [hypervisor]\n");
 777         }
 778
 779         if (show & show_mask) {
 780                 struct symbol *sym = NULL;
 781
 782                 if (dso)
 783                         sym = dso->find_symbol(dso, ip);
 784
 785                 if (hist_entry__add(thread, map, dso, sym, ip, level)) {
 786                         fprintf(stderr,
 787                 "problem incrementing symbol count, skipping event\n");
 788                         return -1;
 789                 }
 790         }
 791         total++;
 792
 793         return 0;
 794 }
 795
 796 static int
 797 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
 798 {
 799         struct thread *thread = threads__findnew(event->mmap.pid);
 800         struct map *map = map__new(&event->mmap);
 801
 802         dprintf("%p [%p]: PERF_EVENT_MMAP %d: [%p(%p) @ %p]: %s\n",
 803                 (void *)(offset + head),
 804                 (void *)(long)(event->header.size),
 805                 event->mmap.pid,
 806                 (void *)(long)event->mmap.start,
 807                 (void *)(long)event->mmap.len,
 808                 (void *)(long)event->mmap.pgoff,
 809                 event->mmap.filename);
 810
 811         if (thread == NULL || map == NULL) {
 812                 dprintf("problem processing PERF_EVENT_MMAP, skipping event.\n");
 813                 return 0;
 814         }
 815
 816         thread__insert_map(thread, map);
 817         total_mmap++;
 818
 819         return 0;
 820 }
 821
 822 static int
 823 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
 824 {
 825         struct thread *thread = threads__findnew(event->comm.pid);
 826
 827         dprintf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
 828                 (void *)(offset + head),
 829                 (void *)(long)(event->header.size),
 830                 event->comm.comm, event->comm.pid);
 831
 832         if (thread == NULL ||
 833             thread__set_comm(thread, event->comm.comm)) {
 834                 dprintf("problem processing PERF_EVENT_COMM, skipping event.\n");
 835                 return -1;
 836         }
 837         total_comm++;
 838
 839         return 0;
 840 }
 841
 842 static int
 843 process_fork_event(event_t *event, unsigned long offset, unsigned long head)
 844 {
 845         struct thread *thread = threads__findnew(event->fork.pid);
 846         struct thread *parent = threads__findnew(event->fork.ppid);
 847
 848         dprintf("%p [%p]: PERF_EVENT_FORK: %d:%d\n",
 849                 (void *)(offset + head),
 850                 (void *)(long)(event->header.size),
 851                 event->fork.pid, event->fork.ppid);
 852
 853         if (!thread || !parent || thread__fork(thread, parent)) {
 854                 dprintf("problem processing PERF_EVENT_FORK, skipping event.\n");
 855                 return -1;
 856         }
 857         total_fork++;
 858
 859         return 0;
 860 }
 861
 862 static int
 863 process_event(event_t *event, unsigned long offset, unsigned long head)
 864 {
 865         switch (event->header.type) {
 866         case PERF_EVENT_SAMPLE:
 867                 return process_sample_event(event, offset, head);
 868
 869         case PERF_EVENT_MMAP:
 870                 return process_mmap_event(event, offset, head);
 871
 872         case PERF_EVENT_COMM:
 873                 return process_comm_event(event, offset, head);
 874
 875         case PERF_EVENT_FORK:
 876                 return process_fork_event(event, offset, head);
 877         /*
 878          * We dont process them right now but they are fine:
 879          */
 880
 881         case PERF_EVENT_THROTTLE:
 882         case PERF_EVENT_UNTHROTTLE:
 883                 return 0;
 884
 885         default:
 886                 return -1;
 887         }
 888
 889         return 0;
 890 }
 891
 892 static int
 893 parse_line(FILE *file, struct symbol *sym, u64 start, u64 len)
 894 {
 895         char *line = NULL, *tmp, *tmp2;
 896         static const char *prev_line;
 897         static const char *prev_color;
 898         unsigned int offset;
 899         size_t line_len;
 900         s64 line_ip;
 901         int ret;
 902         char *c;
 903
 904         if (getline(&line, &line_len, file) < 0)
 905                 return -1;
 906         if (!line)
 907                 return -1;
 908
 909         c = strchr(line, '\n');
 910         if (c)
 911                 *c = 0;
 912
 913         line_ip = -1;
 914         offset = 0;
 915         ret = -2;
 916
 917         /*
 918          * Strip leading spaces:
 919          */
 920         tmp = line;
 921         while (*tmp) {
 922                 if (*tmp != ' ')
 923                         break;
 924                 tmp++;
 925         }
 926
 927         if (*tmp) {
 928                 /*
 929                  * Parse hexa addresses followed by ':'
 930                  */
 931                 line_ip = strtoull(tmp, &tmp2, 16);
 932                 if (*tmp2 != ':')
 933                         line_ip = -1;
 934         }
 935
 936         if (line_ip != -1) {
 937                 const char *path = NULL;
 938                 unsigned int hits = 0;
 939                 double percent = 0.0;
 940                 char *color;
 941                 struct sym_ext *sym_ext = sym->priv;
 942
 943                 offset = line_ip - start;
 944                 if (offset < len)
 945                         hits = sym->hist[offset];
 946
 947                 if (offset < len && sym_ext) {
 948                         path = sym_ext[offset].path;
 949                         percent = sym_ext[offset].percent;
 950                 } else if (sym->hist_sum)
 951                         percent = 100.0 * hits / sym->hist_sum;
 952
 953                 color = get_percent_color(percent);
 954
 955                 /*
 956                  * Also color the filename and line if needed, with
 957                  * the same color than the percentage. Don't print it
 958                  * twice for close colored ip with the same filename:line
 959                  */
 960                 if (path) {
 961                         if (!prev_line || strcmp(prev_line, path)
 962                                        || color != prev_color) {
 963                                 color_fprintf(stdout, color, " %s", path);
 964                                 prev_line = path;
 965                                 prev_color = color;
 966                         }
 967                 }
 968
 969                 color_fprintf(stdout, color, " %7.2f", percent);
 970                 printf(" :      ");
 971                 color_fprintf(stdout, PERF_COLOR_BLUE, "%s\n", line);
 972         } else {
 973                 if (!*line)
 974                         printf("         :\n");
 975                 else
 976                         printf("         :      %s\n", line);
 977         }
 978
 979         return 0;
 980 }
 981
 982 static struct rb_root root_sym_ext;
 983
 984 static void insert_source_line(struct sym_ext *sym_ext)
 985 {
 986         struct sym_ext *iter;
 987         struct rb_node **p = &root_sym_ext.rb_node;
 988         struct rb_node *parent = NULL;
 989
 990         while (*p != NULL) {
 991                 parent = *p;
 992                 iter = rb_entry(parent, struct sym_ext, node);
 993
 994                 if (sym_ext->percent > iter->percent)
 995                         p = &(*p)->rb_left;
 996                 else
 997                         p = &(*p)->rb_right;
 998         }
 999
1000         rb_link_node(&sym_ext->node, parent, p);
1001         rb_insert_color(&sym_ext->node, &root_sym_ext);
1002 }
1003
1004 static void free_source_line(struct symbol *sym, int len)
1005 {
1006         struct sym_ext *sym_ext = sym->priv;
1007         int i;
1008
1009         if (!sym_ext)
1010                 return;
1011
1012         for (i = 0; i < len; i++)
1013                 free(sym_ext[i].path);
1014         free(sym_ext);
1015
1016         sym->priv = NULL;
1017         root_sym_ext = RB_ROOT;
1018 }
1019
1020 /* Get the filename:line for the colored entries */
1021 static void
1022 get_source_line(struct symbol *sym, u64 start, int len, char *filename)
1023 {
1024         int i;
1025         char cmd[PATH_MAX * 2];
1026         struct sym_ext *sym_ext;
1027
1028         if (!sym->hist_sum)
1029                 return;
1030
1031         sym->priv = calloc(len, sizeof(struct sym_ext));
1032         if (!sym->priv)
1033                 return;
1034
1035         sym_ext = sym->priv;
1036
1037         for (i = 0; i < len; i++) {
1038                 char *path = NULL;
1039                 size_t line_len;
1040                 u64 offset;
1041                 FILE *fp;
1042
1043                 sym_ext[i].percent = 100.0 * sym->hist[i] / sym->hist_sum;
1044                 if (sym_ext[i].percent <= 0.5)
1045                         continue;
1046
1047                 offset = start + i;
1048                 sprintf(cmd, "addr2line -e %s %016llx", filename, offset);
1049                 fp = popen(cmd, "r");
1050                 if (!fp)
1051                         continue;
1052
1053                 if (getline(&path, &line_len, fp) < 0 || !line_len)
1054                         goto next;
1055
1056                 sym_ext[i].path = malloc(sizeof(char) * line_len + 1);
1057                 if (!sym_ext[i].path)
1058                         goto next;
1059
1060                 strcpy(sym_ext[i].path, path);
1061                 insert_source_line(&sym_ext[i]);
1062
1063         next:
1064                 pclose(fp);
1065         }
1066 }
1067
1068 static void print_summary(char *filename)
1069 {
1070         struct sym_ext *sym_ext;
1071         struct rb_node *node;
1072
1073         printf("\nSorted summary for file %s\n", filename);
1074         printf("----------------------------------------------\n\n");
1075
1076         if (RB_EMPTY_ROOT(&root_sym_ext)) {
1077                 printf(" Nothing higher than %1.1f%%\n", MIN_GREEN);
1078                 return;
1079         }
1080
1081         node = rb_first(&root_sym_ext);
1082         while (node) {
1083                 double percent;
1084                 char *color;
1085                 char *path;
1086
1087                 sym_ext = rb_entry(node, struct sym_ext, node);
1088                 percent = sym_ext->percent;
1089                 color = get_percent_color(percent);
1090                 path = sym_ext->path;
1091
1092                 color_fprintf(stdout, color, " %7.2f %s", percent, path);
1093                 node = rb_next(node);
1094         }
1095 }
1096
1097 static void annotate_sym(struct dso *dso, struct symbol *sym)
1098 {
1099         char *filename = dso->name, *d_filename;
1100         u64 start, end, len;
1101         char command[PATH_MAX*2];
1102         FILE *file;
1103
1104         if (!filename)
1105                 return;
1106         if (sym->module)
1107                 filename = sym->module->path;
1108         else if (dso == kernel_dso)
1109                 filename = vmlinux;
1110
1111         start = sym->obj_start;
1112         if (!start)
1113                 start = sym->start;
1114         if (full_paths)
1115                 d_filename = filename;
1116         else
1117                 d_filename = basename(filename);
1118
1119         end = start + sym->end - sym->start + 1;
1120         len = sym->end - sym->start;
1121
1122         if (print_line) {
1123                 get_source_line(sym, start, len, filename);
1124                 print_summary(filename);
1125         }
1126
1127         printf("\n\n------------------------------------------------\n");
1128         printf(" Percent |      Source code & Disassembly of %s\n", d_filename);
1129         printf("------------------------------------------------\n");
1130
1131         if (verbose >= 2)
1132                 printf("annotating [%p] %30s : [%p] %30s\n", dso, dso->name, sym, sym->name);
1133
1134         sprintf(command, "objdump --start-address=0x%016Lx --stop-address=0x%016Lx -dS %s|grep -v %s",
1135                         (u64)start, (u64)end, filename, filename);
1136
1137         if (verbose >= 3)
1138                 printf("doing: %s\n", command);
1139
1140         file = popen(command, "r");
1141         if (!file)
1142                 return;
1143
1144         while (!feof(file)) {
1145                 if (parse_line(file, sym, start, len) < 0)
1146                         break;
1147         }
1148
1149         pclose(file);
1150         if (print_line)
1151                 free_source_line(sym, len);
1152 }
1153
1154 static void find_annotations(void)
1155 {
1156         struct rb_node *nd;
1157         struct dso *dso;
1158         int count = 0;
1159
1160         list_for_each_entry(dso, &dsos, node) {
1161
1162                 for (nd = rb_first(&dso->syms); nd; nd = rb_next(nd)) {
1163                         struct symbol *sym = rb_entry(nd, struct symbol, rb_node);
1164
1165                         if (sym->hist) {
1166                                 annotate_sym(dso, sym);
1167                                 count++;
1168                         }
1169                 }
1170         }
1171
1172         if (!count)
1173                 printf(" Error: symbol '%s' not present amongst the samples.\n", sym_hist_filter);
1174 }
1175
1176 static int __cmd_annotate(void)
1177 {
1178         int ret, rc = EXIT_FAILURE;
1179         unsigned long offset = 0;
1180         unsigned long head = 0;
1181         struct stat stat;
1182         event_t *event;
1183         uint32_t size;
1184         char *buf;
1185
1186         register_idle_thread();
1187
1188         input = open(input_name, O_RDONLY);
1189         if (input < 0) {
1190                 perror("failed to open file");
1191                 exit(-1);
1192         }
1193
1194         ret = fstat(input, &stat);
1195         if (ret < 0) {
1196                 perror("failed to stat file");
1197                 exit(-1);
1198         }
1199
1200         if (!stat.st_size) {
1201                 fprintf(stderr, "zero-sized file, nothing to do!\n");
1202                 exit(0);
1203         }
1204
1205         if (load_kernel() < 0) {
1206                 perror("failed to load kernel symbols");
1207                 return EXIT_FAILURE;
1208         }
1209
1210 remap:
1211         buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1212                            MAP_SHARED, input, offset);
1213         if (buf == MAP_FAILED) {
1214                 perror("failed to mmap file");
1215                 exit(-1);
1216         }
1217
1218 more:
1219         event = (event_t *)(buf + head);
1220
1221         size = event->header.size;
1222         if (!size)
1223                 size = 8;
1224
1225         if (head + event->header.size >= page_size * mmap_window) {
1226                 unsigned long shift = page_size * (head / page_size);
1227                 int ret;
1228
1229                 ret = munmap(buf, page_size * mmap_window);
1230                 assert(ret == 0);
1231
1232                 offset += shift;
1233                 head -= shift;
1234                 goto remap;
1235         }
1236
1237         size = event->header.size;
1238
1239         dprintf("%p [%p]: event: %d\n",
1240                         (void *)(offset + head),
1241                         (void *)(long)event->header.size,
1242                         event->header.type);
1243
1244         if (!size || process_event(event, offset, head) < 0) {
1245
1246                 dprintf("%p [%p]: skipping unknown header type: %d\n",
1247                         (void *)(offset + head),
1248                         (void *)(long)(event->header.size),
1249                         event->header.type);
1250
1251                 total_unknown++;
1252
1253                 /*
1254                  * assume we lost track of the stream, check alignment, and
1255                  * increment a single u64 in the hope to catch on again 'soon'.
1256                  */
1257
1258                 if (unlikely(head & 7))
1259                         head &= ~7ULL;
1260
1261                 size = 8;
1262         }
1263
1264         head += size;
1265
1266         if (offset + head < (unsigned long)stat.st_size)
1267                 goto more;
1268
1269         rc = EXIT_SUCCESS;
1270         close(input);
1271
1272         dprintf("      IP events: %10ld\n", total);
1273         dprintf("    mmap events: %10ld\n", total_mmap);
1274         dprintf("    comm events: %10ld\n", total_comm);
1275         dprintf("    fork events: %10ld\n", total_fork);
1276         dprintf(" unknown events: %10ld\n", total_unknown);
1277
1278         if (dump_trace)
1279                 return 0;
1280
1281         if (verbose >= 3)
1282                 threads__fprintf(stdout);
1283
1284         if (verbose >= 2)
1285                 dsos__fprintf(stdout);
1286
1287         collapse__resort();
1288         output__resort();
1289
1290         find_annotations();
1291
1292         return rc;
1293 }
1294
1295 static const char * const annotate_usage[] = {
1296         "perf annotate [<options>] <command>",
1297         NULL
1298 };
1299
1300 static const struct option options[] = {
1301         OPT_STRING('i', "input", &input_name, "file",
1302                     "input file name"),
1303         OPT_STRING('s', "symbol", &sym_hist_filter, "symbol",
1304                     "symbol to annotate"),
1305         OPT_BOOLEAN('v', "verbose", &verbose,
1306                     "be more verbose (show symbol address, etc)"),
1307         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1308                     "dump raw trace in ASCII"),
1309         OPT_STRING('k', "vmlinux", &vmlinux, "file", "vmlinux pathname"),
1310         OPT_BOOLEAN('m', "modules", &modules,
1311                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
1312         OPT_BOOLEAN('l', "print-line", &print_line,
1313                     "print matching source lines (may be slow)"),
1314         OPT_BOOLEAN('P', "full-paths", &full_paths,
1315                     "Don't shorten the displayed pathnames"),
1316         OPT_END()
1317 };
1318
1319 static void setup_sorting(void)
1320 {
1321         char *tmp, *tok, *str = strdup(sort_order);
1322
1323         for (tok = strtok_r(str, ", ", &tmp);
1324                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1325                 if (sort_dimension__add(tok) < 0) {
1326                         error("Unknown --sort key: `%s'", tok);
1327                         usage_with_options(annotate_usage, options);
1328                 }
1329         }
1330
1331         free(str);
1332 }
1333
1334 int cmd_annotate(int argc, const char **argv, const char *prefix __used)
1335 {
1336         symbol__init();
1337
1338         page_size = getpagesize();
1339
1340         argc = parse_options(argc, argv, options, annotate_usage, 0);
1341
1342         setup_sorting();
1343
1344         if (argc) {
1345                 /*
1346                  * Special case: if there's an argument left then assume tha
1347                  * it's a symbol filter:
1348                  */
1349                 if (argc > 1)
1350                         usage_with_options(annotate_usage, options);
1351
1352                 sym_hist_filter = argv[0];
1353         }
1354
1355         if (!sym_hist_filter)
1356                 usage_with_options(annotate_usage, options);
1357
1358         setup_pager();
1359
1360         return __cmd_annotate();
1361 }