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 <linux/list.h>
  14 #include "util/cache.h"
  15 #include <linux/rbtree.h>
  16 #include "util/symbol.h"
  17 #include "util/string.h"
  18 #include "util/callchain.h"
  19 #include "util/strlist.h"
  20
  21 #include "perf.h"
  22 #include "util/header.h"
  23
  24 #include "util/parse-options.h"
  25 #include "util/parse-events.h"
  26
  27 #define SHOW_KERNEL     1
  28 #define SHOW_USER       2
  29 #define SHOW_HV         4
  30
  31 static char             const *input_name = "perf.data";
  32 static char             *vmlinux = NULL;
  33
  34 static char             default_sort_order[] = "comm,dso,symbol";
  35 static char             *sort_order = default_sort_order;
  36 static char             *dso_list_str, *comm_list_str, *sym_list_str,
  37                         *col_width_list_str;
  38 static struct strlist   *dso_list, *comm_list, *sym_list;
  39 static char             *field_sep;
  40
  41 static int              input;
  42 static int              show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV;
  43
  44 static int              dump_trace = 0;
  45 #define dprintf(x...)   do { if (dump_trace) printf(x); } while (0)
  46 #define cdprintf(x...)  do { if (dump_trace) color_fprintf(stdout, color, x); } while (0)
  47
  48 static int              verbose;
  49 #define eprintf(x...)   do { if (verbose) fprintf(stderr, x); } while (0)
  50
  51 static int              modules;
  52
  53 static int              full_paths;
  54 static int              show_nr_samples;
  55
  56 static unsigned long    page_size;
  57 static unsigned long    mmap_window = 32;
  58
  59 static char             default_parent_pattern[] = "^sys_|^do_page_fault";
  60 static char             *parent_pattern = default_parent_pattern;
  61 static regex_t          parent_regex;
  62
  63 static int              exclude_other = 1;
  64
  65 static char             callchain_default_opt[] = "fractal,0.5";
  66
  67 static int              callchain;
  68
  69 static
  70 struct callchain_param  callchain_param = {
  71         .mode   = CHAIN_GRAPH_ABS,
  72         .min_percent = 0.5
  73 };
  74
  75 static u64              sample_type;
  76
  77 struct ip_event {
  78         struct perf_event_header header;
  79         u64 ip;
  80         u32 pid, tid;
  81         unsigned char __more_data[];
  82 };
  83
  84 struct mmap_event {
  85         struct perf_event_header header;
  86         u32 pid, tid;
  87         u64 start;
  88         u64 len;
  89         u64 pgoff;
  90         char filename[PATH_MAX];
  91 };
  92
  93 struct comm_event {
  94         struct perf_event_header header;
  95         u32 pid, tid;
  96         char comm[16];
  97 };
  98
  99 struct fork_event {
 100         struct perf_event_header header;
 101         u32 pid, ppid;
 102         u32 tid, ptid;
 103 };
 104
 105 struct lost_event {
 106         struct perf_event_header header;
 107         u64 id;
 108         u64 lost;
 109 };
 110
 111 struct read_event {
 112         struct perf_event_header header;
 113         u32 pid,tid;
 114         u64 value;
 115         u64 format[3];
 116 };
 117
 118 typedef union event_union {
 119         struct perf_event_header        header;
 120         struct ip_event                 ip;
 121         struct mmap_event               mmap;
 122         struct comm_event               comm;
 123         struct fork_event               fork;
 124         struct lost_event               lost;
 125         struct read_event               read;
 126 } event_t;
 127
 128 static int repsep_fprintf(FILE *fp, const char *fmt, ...)
 129 {
 130         int n;
 131         va_list ap;
 132
 133         va_start(ap, fmt);
 134         if (!field_sep)
 135                 n = vfprintf(fp, fmt, ap);
 136         else {
 137                 char *bf = NULL;
 138                 n = vasprintf(&bf, fmt, ap);
 139                 if (n > 0) {
 140                         char *sep = bf;
 141                         while (1) {
 142                                 sep = strchr(sep, *field_sep);
 143                                 if (sep == NULL)
 144                                         break;
 145                                 *sep = '.';
 146                         }
 147                 }
 148                 fputs(bf, fp);
 149                 free(bf);
 150         }
 151         va_end(ap);
 152         return n;
 153 }
 154
 155 static LIST_HEAD(dsos);
 156 static struct dso *kernel_dso;
 157 static struct dso *vdso;
 158 static struct dso *hypervisor_dso;
 159
 160 static void dsos__add(struct dso *dso)
 161 {
 162         list_add_tail(&dso->node, &dsos);
 163 }
 164
 165 static struct dso *dsos__find(const char *name)
 166 {
 167         struct dso *pos;
 168
 169         list_for_each_entry(pos, &dsos, node)
 170                 if (strcmp(pos->name, name) == 0)
 171                         return pos;
 172         return NULL;
 173 }
 174
 175 static struct dso *dsos__findnew(const char *name)
 176 {
 177         struct dso *dso = dsos__find(name);
 178         int nr;
 179
 180         if (dso)
 181                 return dso;
 182
 183         dso = dso__new(name, 0);
 184         if (!dso)
 185                 goto out_delete_dso;
 186
 187         nr = dso__load(dso, NULL, verbose);
 188         if (nr < 0) {
 189                 eprintf("Failed to open: %s\n", name);
 190                 goto out_delete_dso;
 191         }
 192         if (!nr)
 193                 eprintf("No symbols found in: %s, maybe install a debug package?\n", name);
 194
 195         dsos__add(dso);
 196
 197         return dso;
 198
 199 out_delete_dso:
 200         dso__delete(dso);
 201         return NULL;
 202 }
 203
 204 static void dsos__fprintf(FILE *fp)
 205 {
 206         struct dso *pos;
 207
 208         list_for_each_entry(pos, &dsos, node)
 209                 dso__fprintf(pos, fp);
 210 }
 211
 212 static struct symbol *vdso__find_symbol(struct dso *dso, u64 ip)
 213 {
 214         return dso__find_symbol(dso, ip);
 215 }
 216
 217 static int load_kernel(void)
 218 {
 219         int err;
 220
 221         kernel_dso = dso__new("[kernel]", 0);
 222         if (!kernel_dso)
 223                 return -1;
 224
 225         err = dso__load_kernel(kernel_dso, vmlinux, NULL, verbose, modules);
 226         if (err <= 0) {
 227                 dso__delete(kernel_dso);
 228                 kernel_dso = NULL;
 229         } else
 230                 dsos__add(kernel_dso);
 231
 232         vdso = dso__new("[vdso]", 0);
 233         if (!vdso)
 234                 return -1;
 235
 236         vdso->find_symbol = vdso__find_symbol;
 237
 238         dsos__add(vdso);
 239
 240         hypervisor_dso = dso__new("[hypervisor]", 0);
 241         if (!hypervisor_dso)
 242                 return -1;
 243         dsos__add(hypervisor_dso);
 244
 245         return err;
 246 }
 247
 248 static char __cwd[PATH_MAX];
 249 static char *cwd = __cwd;
 250 static int cwdlen;
 251
 252 static int strcommon(const char *pathname)
 253 {
 254         int n = 0;
 255
 256         while (n < cwdlen && pathname[n] == cwd[n])
 257                 ++n;
 258
 259         return n;
 260 }
 261
 262 struct map {
 263         struct list_head node;
 264         u64      start;
 265         u64      end;
 266         u64      pgoff;
 267         u64      (*map_ip)(struct map *, u64);
 268         struct dso       *dso;
 269 };
 270
 271 static u64 map__map_ip(struct map *map, u64 ip)
 272 {
 273         return ip - map->start + map->pgoff;
 274 }
 275
 276 static u64 vdso__map_ip(struct map *map __used, u64 ip)
 277 {
 278         return ip;
 279 }
 280
 281 static inline int is_anon_memory(const char *filename)
 282 {
 283         return strcmp(filename, "//anon") == 0;
 284 }
 285
 286 static struct map *map__new(struct mmap_event *event)
 287 {
 288         struct map *self = malloc(sizeof(*self));
 289
 290         if (self != NULL) {
 291                 const char *filename = event->filename;
 292                 char newfilename[PATH_MAX];
 293                 int anon;
 294
 295                 if (cwd) {
 296                         int n = strcommon(filename);
 297
 298                         if (n == cwdlen) {
 299                                 snprintf(newfilename, sizeof(newfilename),
 300                                          ".%s", filename + n);
 301                                 filename = newfilename;
 302                         }
 303                 }
 304
 305                 anon = is_anon_memory(filename);
 306
 307                 if (anon) {
 308                         snprintf(newfilename, sizeof(newfilename), "/tmp/perf-%d.map", event->pid);
 309                         filename = newfilename;
 310                 }
 311
 312                 self->start = event->start;
 313                 self->end   = event->start + event->len;
 314                 self->pgoff = event->pgoff;
 315
 316                 self->dso = dsos__findnew(filename);
 317                 if (self->dso == NULL)
 318                         goto out_delete;
 319
 320                 if (self->dso == vdso || anon)
 321                         self->map_ip = vdso__map_ip;
 322                 else
 323                         self->map_ip = map__map_ip;
 324         }
 325         return self;
 326 out_delete:
 327         free(self);
 328         return NULL;
 329 }
 330
 331 static struct map *map__clone(struct map *self)
 332 {
 333         struct map *map = malloc(sizeof(*self));
 334
 335         if (!map)
 336                 return NULL;
 337
 338         memcpy(map, self, sizeof(*self));
 339
 340         return map;
 341 }
 342
 343 static int map__overlap(struct map *l, struct map *r)
 344 {
 345         if (l->start > r->start) {
 346                 struct map *t = l;
 347                 l = r;
 348                 r = t;
 349         }
 350
 351         if (l->end > r->start)
 352                 return 1;
 353
 354         return 0;
 355 }
 356
 357 static size_t map__fprintf(struct map *self, FILE *fp)
 358 {
 359         return fprintf(fp, " %Lx-%Lx %Lx %s\n",
 360                        self->start, self->end, self->pgoff, self->dso->name);
 361 }
 362
 363
 364 struct thread {
 365         struct rb_node   rb_node;
 366         struct list_head maps;
 367         pid_t            pid;
 368         char             *comm;
 369 };
 370
 371 static struct thread *thread__new(pid_t pid)
 372 {
 373         struct thread *self = malloc(sizeof(*self));
 374
 375         if (self != NULL) {
 376                 self->pid = pid;
 377                 self->comm = malloc(32);
 378                 if (self->comm)
 379                         snprintf(self->comm, 32, ":%d", self->pid);
 380                 INIT_LIST_HEAD(&self->maps);
 381         }
 382
 383         return self;
 384 }
 385
 386 static unsigned int dsos__col_width,
 387                     comms__col_width,
 388                     threads__col_width;
 389
 390 static int thread__set_comm(struct thread *self, const char *comm)
 391 {
 392         if (self->comm)
 393                 free(self->comm);
 394         self->comm = strdup(comm);
 395         if (!self->comm)
 396                 return -ENOMEM;
 397
 398         if (!col_width_list_str && !field_sep &&
 399             (!comm_list || strlist__has_entry(comm_list, comm))) {
 400                 unsigned int slen = strlen(comm);
 401                 if (slen > comms__col_width) {
 402                         comms__col_width = slen;
 403                         threads__col_width = slen + 6;
 404                 }
 405         }
 406
 407         return 0;
 408 }
 409
 410 static size_t thread__fprintf(struct thread *self, FILE *fp)
 411 {
 412         struct map *pos;
 413         size_t ret = fprintf(fp, "Thread %d %s\n", self->pid, self->comm);
 414
 415         list_for_each_entry(pos, &self->maps, node)
 416                 ret += map__fprintf(pos, fp);
 417
 418         return ret;
 419 }
 420
 421
 422 static struct rb_root threads;
 423 static struct thread *last_match;
 424
 425 static struct thread *threads__findnew(pid_t pid)
 426 {
 427         struct rb_node **p = &threads.rb_node;
 428         struct rb_node *parent = NULL;
 429         struct thread *th;
 430
 431         /*
 432          * Font-end cache - PID lookups come in blocks,
 433          * so most of the time we dont have to look up
 434          * the full rbtree:
 435          */
 436         if (last_match && last_match->pid == pid)
 437                 return last_match;
 438
 439         while (*p != NULL) {
 440                 parent = *p;
 441                 th = rb_entry(parent, struct thread, rb_node);
 442
 443                 if (th->pid == pid) {
 444                         last_match = th;
 445                         return th;
 446                 }
 447
 448                 if (pid < th->pid)
 449                         p = &(*p)->rb_left;
 450                 else
 451                         p = &(*p)->rb_right;
 452         }
 453
 454         th = thread__new(pid);
 455         if (th != NULL) {
 456                 rb_link_node(&th->rb_node, parent, p);
 457                 rb_insert_color(&th->rb_node, &threads);
 458                 last_match = th;
 459         }
 460
 461         return th;
 462 }
 463
 464 static void thread__insert_map(struct thread *self, struct map *map)
 465 {
 466         struct map *pos, *tmp;
 467
 468         list_for_each_entry_safe(pos, tmp, &self->maps, node) {
 469                 if (map__overlap(pos, map)) {
 470                         if (verbose >= 2) {
 471                                 printf("overlapping maps:\n");
 472                                 map__fprintf(map, stdout);
 473                                 map__fprintf(pos, stdout);
 474                         }
 475
 476                         if (map->start <= pos->start && map->end > pos->start)
 477                                 pos->start = map->end;
 478
 479                         if (map->end >= pos->end && map->start < pos->end)
 480                                 pos->end = map->start;
 481
 482                         if (verbose >= 2) {
 483                                 printf("after collision:\n");
 484                                 map__fprintf(pos, stdout);
 485                         }
 486
 487                         if (pos->start >= pos->end) {
 488                                 list_del_init(&pos->node);
 489                                 free(pos);
 490                         }
 491                 }
 492         }
 493
 494         list_add_tail(&map->node, &self->maps);
 495 }
 496
 497 static int thread__fork(struct thread *self, struct thread *parent)
 498 {
 499         struct map *map;
 500
 501         if (self->comm)
 502                 free(self->comm);
 503         self->comm = strdup(parent->comm);
 504         if (!self->comm)
 505                 return -ENOMEM;
 506
 507         list_for_each_entry(map, &parent->maps, node) {
 508                 struct map *new = map__clone(map);
 509                 if (!new)
 510                         return -ENOMEM;
 511                 thread__insert_map(self, new);
 512         }
 513
 514         return 0;
 515 }
 516
 517 static struct map *thread__find_map(struct thread *self, u64 ip)
 518 {
 519         struct map *pos;
 520
 521         if (self == NULL)
 522                 return NULL;
 523
 524         list_for_each_entry(pos, &self->maps, node)
 525                 if (ip >= pos->start && ip <= pos->end)
 526                         return pos;
 527
 528         return NULL;
 529 }
 530
 531 static size_t threads__fprintf(FILE *fp)
 532 {
 533         size_t ret = 0;
 534         struct rb_node *nd;
 535
 536         for (nd = rb_first(&threads); nd; nd = rb_next(nd)) {
 537                 struct thread *pos = rb_entry(nd, struct thread, rb_node);
 538
 539                 ret += thread__fprintf(pos, fp);
 540         }
 541
 542         return ret;
 543 }
 544
 545 /*
 546  * histogram, sorted on item, collects counts
 547  */
 548
 549 static struct rb_root hist;
 550
 551 struct hist_entry {
 552         struct rb_node          rb_node;
 553
 554         struct thread           *thread;
 555         struct map              *map;
 556         struct dso              *dso;
 557         struct symbol           *sym;
 558         struct symbol           *parent;
 559         u64                     ip;
 560         char                    level;
 561         struct callchain_node   callchain;
 562         struct rb_root          sorted_chain;
 563
 564         u64                     count;
 565 };
 566
 567 /*
 568  * configurable sorting bits
 569  */
 570
 571 struct sort_entry {
 572         struct list_head list;
 573
 574         char *header;
 575
 576         int64_t (*cmp)(struct hist_entry *, struct hist_entry *);
 577         int64_t (*collapse)(struct hist_entry *, struct hist_entry *);
 578         size_t  (*print)(FILE *fp, struct hist_entry *, unsigned int width);
 579         unsigned int *width;
 580         bool    elide;
 581 };
 582
 583 static int64_t cmp_null(void *l, void *r)
 584 {
 585         if (!l && !r)
 586                 return 0;
 587         else if (!l)
 588                 return -1;
 589         else
 590                 return 1;
 591 }
 592
 593 /* --sort pid */
 594
 595 static int64_t
 596 sort__thread_cmp(struct hist_entry *left, struct hist_entry *right)
 597 {
 598         return right->thread->pid - left->thread->pid;
 599 }
 600
 601 static size_t
 602 sort__thread_print(FILE *fp, struct hist_entry *self, unsigned int width)
 603 {
 604         return repsep_fprintf(fp, "%*s:%5d", width - 6,
 605                               self->thread->comm ?: "", self->thread->pid);
 606 }
 607
 608 static struct sort_entry sort_thread = {
 609         .header = "Command:  Pid",
 610         .cmp    = sort__thread_cmp,
 611         .print  = sort__thread_print,
 612         .width  = &threads__col_width,
 613 };
 614
 615 /* --sort comm */
 616
 617 static int64_t
 618 sort__comm_cmp(struct hist_entry *left, struct hist_entry *right)
 619 {
 620         return right->thread->pid - left->thread->pid;
 621 }
 622
 623 static int64_t
 624 sort__comm_collapse(struct hist_entry *left, struct hist_entry *right)
 625 {
 626         char *comm_l = left->thread->comm;
 627         char *comm_r = right->thread->comm;
 628
 629         if (!comm_l || !comm_r)
 630                 return cmp_null(comm_l, comm_r);
 631
 632         return strcmp(comm_l, comm_r);
 633 }
 634
 635 static size_t
 636 sort__comm_print(FILE *fp, struct hist_entry *self, unsigned int width)
 637 {
 638         return repsep_fprintf(fp, "%*s", width, self->thread->comm);
 639 }
 640
 641 static struct sort_entry sort_comm = {
 642         .header         = "Command",
 643         .cmp            = sort__comm_cmp,
 644         .collapse       = sort__comm_collapse,
 645         .print          = sort__comm_print,
 646         .width          = &comms__col_width,
 647 };
 648
 649 /* --sort dso */
 650
 651 static int64_t
 652 sort__dso_cmp(struct hist_entry *left, struct hist_entry *right)
 653 {
 654         struct dso *dso_l = left->dso;
 655         struct dso *dso_r = right->dso;
 656
 657         if (!dso_l || !dso_r)
 658                 return cmp_null(dso_l, dso_r);
 659
 660         return strcmp(dso_l->name, dso_r->name);
 661 }
 662
 663 static size_t
 664 sort__dso_print(FILE *fp, struct hist_entry *self, unsigned int width)
 665 {
 666         if (self->dso)
 667                 return repsep_fprintf(fp, "%-*s", width, self->dso->name);
 668
 669         return repsep_fprintf(fp, "%*llx", width, (u64)self->ip);
 670 }
 671
 672 static struct sort_entry sort_dso = {
 673         .header = "Shared Object",
 674         .cmp    = sort__dso_cmp,
 675         .print  = sort__dso_print,
 676         .width  = &dsos__col_width,
 677 };
 678
 679 /* --sort symbol */
 680
 681 static int64_t
 682 sort__sym_cmp(struct hist_entry *left, struct hist_entry *right)
 683 {
 684         u64 ip_l, ip_r;
 685
 686         if (left->sym == right->sym)
 687                 return 0;
 688
 689         ip_l = left->sym ? left->sym->start : left->ip;
 690         ip_r = right->sym ? right->sym->start : right->ip;
 691
 692         return (int64_t)(ip_r - ip_l);
 693 }
 694
 695 static size_t
 696 sort__sym_print(FILE *fp, struct hist_entry *self, unsigned int width __used)
 697 {
 698         size_t ret = 0;
 699
 700         if (verbose)
 701                 ret += repsep_fprintf(fp, "%#018llx  ", (u64)self->ip);
 702
 703         ret += repsep_fprintf(fp, "[%c] ", self->level);
 704         if (self->sym) {
 705                 ret += repsep_fprintf(fp, "%s", self->sym->name);
 706
 707                 if (self->sym->module)
 708                         ret += repsep_fprintf(fp, "\t[%s]",
 709                                              self->sym->module->name);
 710         } else {
 711                 ret += repsep_fprintf(fp, "%#016llx", (u64)self->ip);
 712         }
 713
 714         return ret;
 715 }
 716
 717 static struct sort_entry sort_sym = {
 718         .header = "Symbol",
 719         .cmp    = sort__sym_cmp,
 720         .print  = sort__sym_print,
 721 };
 722
 723 /* --sort parent */
 724
 725 static int64_t
 726 sort__parent_cmp(struct hist_entry *left, struct hist_entry *right)
 727 {
 728         struct symbol *sym_l = left->parent;
 729         struct symbol *sym_r = right->parent;
 730
 731         if (!sym_l || !sym_r)
 732                 return cmp_null(sym_l, sym_r);
 733
 734         return strcmp(sym_l->name, sym_r->name);
 735 }
 736
 737 static size_t
 738 sort__parent_print(FILE *fp, struct hist_entry *self, unsigned int width)
 739 {
 740         return repsep_fprintf(fp, "%-*s", width,
 741                               self->parent ? self->parent->name : "[other]");
 742 }
 743
 744 static unsigned int parent_symbol__col_width;
 745
 746 static struct sort_entry sort_parent = {
 747         .header = "Parent symbol",
 748         .cmp    = sort__parent_cmp,
 749         .print  = sort__parent_print,
 750         .width  = &parent_symbol__col_width,
 751 };
 752
 753 static int sort__need_collapse = 0;
 754 static int sort__has_parent = 0;
 755
 756 struct sort_dimension {
 757         char                    *name;
 758         struct sort_entry       *entry;
 759         int                     taken;
 760 };
 761
 762 static struct sort_dimension sort_dimensions[] = {
 763         { .name = "pid",        .entry = &sort_thread,  },
 764         { .name = "comm",       .entry = &sort_comm,    },
 765         { .name = "dso",        .entry = &sort_dso,     },
 766         { .name = "symbol",     .entry = &sort_sym,     },
 767         { .name = "parent",     .entry = &sort_parent,  },
 768 };
 769
 770 static LIST_HEAD(hist_entry__sort_list);
 771
 772 static int sort_dimension__add(char *tok)
 773 {
 774         unsigned int i;
 775
 776         for (i = 0; i < ARRAY_SIZE(sort_dimensions); i++) {
 777                 struct sort_dimension *sd = &sort_dimensions[i];
 778
 779                 if (sd->taken)
 780                         continue;
 781
 782                 if (strncasecmp(tok, sd->name, strlen(tok)))
 783                         continue;
 784
 785                 if (sd->entry->collapse)
 786                         sort__need_collapse = 1;
 787
 788                 if (sd->entry == &sort_parent) {
 789                         int ret = regcomp(&parent_regex, parent_pattern, REG_EXTENDED);
 790                         if (ret) {
 791                                 char err[BUFSIZ];
 792
 793                                 regerror(ret, &parent_regex, err, sizeof(err));
 794                                 fprintf(stderr, "Invalid regex: %s\n%s",
 795                                         parent_pattern, err);
 796                                 exit(-1);
 797                         }
 798                         sort__has_parent = 1;
 799                 }
 800
 801                 list_add_tail(&sd->entry->list, &hist_entry__sort_list);
 802                 sd->taken = 1;
 803
 804                 return 0;
 805         }
 806
 807         return -ESRCH;
 808 }
 809
 810 static int64_t
 811 hist_entry__cmp(struct hist_entry *left, struct hist_entry *right)
 812 {
 813         struct sort_entry *se;
 814         int64_t cmp = 0;
 815
 816         list_for_each_entry(se, &hist_entry__sort_list, list) {
 817                 cmp = se->cmp(left, right);
 818                 if (cmp)
 819                         break;
 820         }
 821
 822         return cmp;
 823 }
 824
 825 static int64_t
 826 hist_entry__collapse(struct hist_entry *left, struct hist_entry *right)
 827 {
 828         struct sort_entry *se;
 829         int64_t cmp = 0;
 830
 831         list_for_each_entry(se, &hist_entry__sort_list, list) {
 832                 int64_t (*f)(struct hist_entry *, struct hist_entry *);
 833
 834                 f = se->collapse ?: se->cmp;
 835
 836                 cmp = f(left, right);
 837                 if (cmp)
 838                         break;
 839         }
 840
 841         return cmp;
 842 }
 843
 844 static size_t ipchain__fprintf_graph_line(FILE *fp, int depth, int depth_mask)
 845 {
 846         int i;
 847         size_t ret = 0;
 848
 849         ret += fprintf(fp, "%s", "                ");
 850
 851         for (i = 0; i < depth; i++)
 852                 if (depth_mask & (1 << i))
 853                         ret += fprintf(fp, "|          ");
 854                 else
 855                         ret += fprintf(fp, "           ");
 856
 857         ret += fprintf(fp, "\n");
 858
 859         return ret;
 860 }
 861 static size_t
 862 ipchain__fprintf_graph(FILE *fp, struct callchain_list *chain, int depth,
 863                        int depth_mask, int count, u64 total_samples,
 864                        int hits)
 865 {
 866         int i;
 867         size_t ret = 0;
 868
 869         ret += fprintf(fp, "%s", "                ");
 870         for (i = 0; i < depth; i++) {
 871                 if (depth_mask & (1 << i))
 872                         ret += fprintf(fp, "|");
 873                 else
 874                         ret += fprintf(fp, " ");
 875                 if (!count && i == depth - 1) {
 876                         double percent;
 877
 878                         percent = hits * 100.0 / total_samples;
 879                         ret += percent_color_fprintf(fp, "--%2.2f%%-- ", percent);
 880                 } else
 881                         ret += fprintf(fp, "%s", "          ");
 882         }
 883         if (chain->sym)
 884                 ret += fprintf(fp, "%s\n", chain->sym->name);
 885         else
 886                 ret += fprintf(fp, "%p\n", (void *)(long)chain->ip);
 887
 888         return ret;
 889 }
 890
 891 static size_t
 892 callchain__fprintf_graph(FILE *fp, struct callchain_node *self,
 893                         u64 total_samples, int depth, int depth_mask)
 894 {
 895         struct rb_node *node, *next;
 896         struct callchain_node *child;
 897         struct callchain_list *chain;
 898         int new_depth_mask = depth_mask;
 899         u64 new_total;
 900         size_t ret = 0;
 901         int i;
 902
 903         if (callchain_param.mode == CHAIN_GRAPH_REL)
 904                 new_total = self->cumul_hit;
 905         else
 906                 new_total = total_samples;
 907
 908         node = rb_first(&self->rb_root);
 909         while (node) {
 910                 child = rb_entry(node, struct callchain_node, rb_node);
 911
 912                 /*
 913                  * The depth mask manages the output of pipes that show
 914                  * the depth. We don't want to keep the pipes of the current
 915                  * level for the last child of this depth
 916                  */
 917                 next = rb_next(node);
 918                 if (!next)
 919                         new_depth_mask &= ~(1 << (depth - 1));
 920
 921                 /*
 922                  * But we keep the older depth mask for the line seperator
 923                  * to keep the level link until we reach the last child
 924                  */
 925                 ret += ipchain__fprintf_graph_line(fp, depth, depth_mask);
 926                 i = 0;
 927                 list_for_each_entry(chain, &child->val, list) {
 928                         if (chain->ip >= PERF_CONTEXT_MAX)
 929                                 continue;
 930                         ret += ipchain__fprintf_graph(fp, chain, depth,
 931                                                       new_depth_mask, i++,
 932                                                       new_total,
 933                                                       child->cumul_hit);
 934                 }
 935                 ret += callchain__fprintf_graph(fp, child, new_total,
 936                                                 depth + 1,
 937                                                 new_depth_mask | (1 << depth));
 938                 node = next;
 939         }
 940
 941         return ret;
 942 }
 943
 944 static size_t
 945 callchain__fprintf_flat(FILE *fp, struct callchain_node *self,
 946                         u64 total_samples)
 947 {
 948         struct callchain_list *chain;
 949         size_t ret = 0;
 950
 951         if (!self)
 952                 return 0;
 953
 954         ret += callchain__fprintf_flat(fp, self->parent, total_samples);
 955
 956
 957         list_for_each_entry(chain, &self->val, list) {
 958                 if (chain->ip >= PERF_CONTEXT_MAX)
 959                         continue;
 960                 if (chain->sym)
 961                         ret += fprintf(fp, "                %s\n", chain->sym->name);
 962                 else
 963                         ret += fprintf(fp, "                %p\n",
 964                                         (void *)(long)chain->ip);
 965         }
 966
 967         return ret;
 968 }
 969
 970 static size_t
 971 hist_entry_callchain__fprintf(FILE *fp, struct hist_entry *self,
 972                               u64 total_samples)
 973 {
 974         struct rb_node *rb_node;
 975         struct callchain_node *chain;
 976         size_t ret = 0;
 977
 978         rb_node = rb_first(&self->sorted_chain);
 979         while (rb_node) {
 980                 double percent;
 981
 982                 chain = rb_entry(rb_node, struct callchain_node, rb_node);
 983                 percent = chain->hit * 100.0 / total_samples;
 984                 switch (callchain_param.mode) {
 985                 case CHAIN_FLAT:
 986                         ret += percent_color_fprintf(fp, "           %6.2f%%\n",
 987                                                      percent);
 988                         ret += callchain__fprintf_flat(fp, chain, total_samples);
 989                         break;
 990                 case CHAIN_GRAPH_ABS: /* Falldown */
 991                 case CHAIN_GRAPH_REL:
 992                         ret += callchain__fprintf_graph(fp, chain,
 993                                                         total_samples, 1, 1);
 994                 default:
 995                         break;
 996                 }
 997                 ret += fprintf(fp, "\n");
 998                 rb_node = rb_next(rb_node);
 999         }
1000
1001         return ret;
1002 }
1003
1004
1005 static size_t
1006 hist_entry__fprintf(FILE *fp, struct hist_entry *self, u64 total_samples)
1007 {
1008         struct sort_entry *se;
1009         size_t ret;
1010
1011         if (exclude_other && !self->parent)
1012                 return 0;
1013
1014         if (total_samples)
1015                 ret = percent_color_fprintf(fp,
1016                                             field_sep ? "%.2f" : "   %6.2f%%",
1017                                         (self->count * 100.0) / total_samples);
1018         else
1019                 ret = fprintf(fp, field_sep ? "%lld" : "%12lld ", self->count);
1020
1021         if (show_nr_samples) {
1022                 if (field_sep)
1023                         fprintf(fp, "%c%lld", *field_sep, self->count);
1024                 else
1025                         fprintf(fp, "%11lld", self->count);
1026         }
1027
1028         list_for_each_entry(se, &hist_entry__sort_list, list) {
1029                 if (se->elide)
1030                         continue;
1031
1032                 fprintf(fp, "%s", field_sep ?: "  ");
1033                 ret += se->print(fp, self, se->width ? *se->width : 0);
1034         }
1035
1036         ret += fprintf(fp, "\n");
1037
1038         if (callchain)
1039                 hist_entry_callchain__fprintf(fp, self, total_samples);
1040
1041         return ret;
1042 }
1043
1044 /*
1045  *
1046  */
1047
1048 static void dso__calc_col_width(struct dso *self)
1049 {
1050         if (!col_width_list_str && !field_sep &&
1051             (!dso_list || strlist__has_entry(dso_list, self->name))) {
1052                 unsigned int slen = strlen(self->name);
1053                 if (slen > dsos__col_width)
1054                         dsos__col_width = slen;
1055         }
1056
1057         self->slen_calculated = 1;
1058 }
1059
1060 static struct symbol *
1061 resolve_symbol(struct thread *thread, struct map **mapp,
1062                struct dso **dsop, u64 *ipp)
1063 {
1064         struct dso *dso = dsop ? *dsop : NULL;
1065         struct map *map = mapp ? *mapp : NULL;
1066         u64 ip = *ipp;
1067
1068         if (!thread)
1069                 return NULL;
1070
1071         if (dso)
1072                 goto got_dso;
1073
1074         if (map)
1075                 goto got_map;
1076
1077         map = thread__find_map(thread, ip);
1078         if (map != NULL) {
1079                 /*
1080                  * We have to do this here as we may have a dso
1081                  * with no symbol hit that has a name longer than
1082                  * the ones with symbols sampled.
1083                  */
1084                 if (!sort_dso.elide && !map->dso->slen_calculated)
1085                         dso__calc_col_width(map->dso);
1086
1087                 if (mapp)
1088                         *mapp = map;
1089 got_map:
1090                 ip = map->map_ip(map, ip);
1091
1092                 dso = map->dso;
1093         } else {
1094                 /*
1095                  * If this is outside of all known maps,
1096                  * and is a negative address, try to look it
1097                  * up in the kernel dso, as it might be a
1098                  * vsyscall (which executes in user-mode):
1099                  */
1100                 if ((long long)ip < 0)
1101                 dso = kernel_dso;
1102         }
1103         dprintf(" ...... dso: %s\n", dso ? dso->name : "<not found>");
1104         dprintf(" ...... map: %Lx -> %Lx\n", *ipp, ip);
1105         *ipp  = ip;
1106
1107         if (dsop)
1108                 *dsop = dso;
1109
1110         if (!dso)
1111                 return NULL;
1112 got_dso:
1113         return dso->find_symbol(dso, ip);
1114 }
1115
1116 static int call__match(struct symbol *sym)
1117 {
1118         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
1119                 return 1;
1120
1121         return 0;
1122 }
1123
1124 static struct symbol **
1125 resolve_callchain(struct thread *thread, struct map *map __used,
1126                     struct ip_callchain *chain, struct hist_entry *entry)
1127 {
1128         u64 context = PERF_CONTEXT_MAX;
1129         struct symbol **syms = NULL;
1130         unsigned int i;
1131
1132         if (callchain) {
1133                 syms = calloc(chain->nr, sizeof(*syms));
1134                 if (!syms) {
1135                         fprintf(stderr, "Can't allocate memory for symbols\n");
1136                         exit(-1);
1137                 }
1138         }
1139
1140         for (i = 0; i < chain->nr; i++) {
1141                 u64 ip = chain->ips[i];
1142                 struct dso *dso = NULL;
1143                 struct symbol *sym;
1144
1145                 if (ip >= PERF_CONTEXT_MAX) {
1146                         context = ip;
1147                         continue;
1148                 }
1149
1150                 switch (context) {
1151                 case PERF_CONTEXT_HV:
1152                         dso = hypervisor_dso;
1153                         break;
1154                 case PERF_CONTEXT_KERNEL:
1155                         dso = kernel_dso;
1156                         break;
1157                 default:
1158                         break;
1159                 }
1160
1161                 sym = resolve_symbol(thread, NULL, &dso, &ip);
1162
1163                 if (sym) {
1164                         if (sort__has_parent && call__match(sym) &&
1165                             !entry->parent)
1166                                 entry->parent = sym;
1167                         if (!callchain)
1168                                 break;
1169                         syms[i] = sym;
1170                 }
1171         }
1172
1173         return syms;
1174 }
1175
1176 /*
1177  * collect histogram counts
1178  */
1179
1180 static int
1181 hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
1182                 struct symbol *sym, u64 ip, struct ip_callchain *chain,
1183                 char level, u64 count)
1184 {
1185         struct rb_node **p = &hist.rb_node;
1186         struct rb_node *parent = NULL;
1187         struct hist_entry *he;
1188         struct symbol **syms = NULL;
1189         struct hist_entry entry = {
1190                 .thread = thread,
1191                 .map    = map,
1192                 .dso    = dso,
1193                 .sym    = sym,
1194                 .ip     = ip,
1195                 .level  = level,
1196                 .count  = count,
1197                 .parent = NULL,
1198                 .sorted_chain = RB_ROOT
1199         };
1200         int cmp;
1201
1202         if ((sort__has_parent || callchain) && chain)
1203                 syms = resolve_callchain(thread, map, chain, &entry);
1204
1205         while (*p != NULL) {
1206                 parent = *p;
1207                 he = rb_entry(parent, struct hist_entry, rb_node);
1208
1209                 cmp = hist_entry__cmp(&entry, he);
1210
1211                 if (!cmp) {
1212                         he->count += count;
1213                         if (callchain) {
1214                                 append_chain(&he->callchain, chain, syms);
1215                                 free(syms);
1216                         }
1217                         return 0;
1218                 }
1219
1220                 if (cmp < 0)
1221                         p = &(*p)->rb_left;
1222                 else
1223                         p = &(*p)->rb_right;
1224         }
1225
1226         he = malloc(sizeof(*he));
1227         if (!he)
1228                 return -ENOMEM;
1229         *he = entry;
1230         if (callchain) {
1231                 callchain_init(&he->callchain);
1232                 append_chain(&he->callchain, chain, syms);
1233                 free(syms);
1234         }
1235         rb_link_node(&he->rb_node, parent, p);
1236         rb_insert_color(&he->rb_node, &hist);
1237
1238         return 0;
1239 }
1240
1241 static void hist_entry__free(struct hist_entry *he)
1242 {
1243         free(he);
1244 }
1245
1246 /*
1247  * collapse the histogram
1248  */
1249
1250 static struct rb_root collapse_hists;
1251
1252 static void collapse__insert_entry(struct hist_entry *he)
1253 {
1254         struct rb_node **p = &collapse_hists.rb_node;
1255         struct rb_node *parent = NULL;
1256         struct hist_entry *iter;
1257         int64_t cmp;
1258
1259         while (*p != NULL) {
1260                 parent = *p;
1261                 iter = rb_entry(parent, struct hist_entry, rb_node);
1262
1263                 cmp = hist_entry__collapse(iter, he);
1264
1265                 if (!cmp) {
1266                         iter->count += he->count;
1267                         hist_entry__free(he);
1268                         return;
1269                 }
1270
1271                 if (cmp < 0)
1272                         p = &(*p)->rb_left;
1273                 else
1274                         p = &(*p)->rb_right;
1275         }
1276
1277         rb_link_node(&he->rb_node, parent, p);
1278         rb_insert_color(&he->rb_node, &collapse_hists);
1279 }
1280
1281 static void collapse__resort(void)
1282 {
1283         struct rb_node *next;
1284         struct hist_entry *n;
1285
1286         if (!sort__need_collapse)
1287                 return;
1288
1289         next = rb_first(&hist);
1290         while (next) {
1291                 n = rb_entry(next, struct hist_entry, rb_node);
1292                 next = rb_next(&n->rb_node);
1293
1294                 rb_erase(&n->rb_node, &hist);
1295                 collapse__insert_entry(n);
1296         }
1297 }
1298
1299 /*
1300  * reverse the map, sort on count.
1301  */
1302
1303 static struct rb_root output_hists;
1304
1305 static void output__insert_entry(struct hist_entry *he, u64 min_callchain_hits)
1306 {
1307         struct rb_node **p = &output_hists.rb_node;
1308         struct rb_node *parent = NULL;
1309         struct hist_entry *iter;
1310
1311         if (callchain)
1312                 callchain_param.sort(&he->sorted_chain, &he->callchain,
1313                                       min_callchain_hits, &callchain_param);
1314
1315         while (*p != NULL) {
1316                 parent = *p;
1317                 iter = rb_entry(parent, struct hist_entry, rb_node);
1318
1319                 if (he->count > iter->count)
1320                         p = &(*p)->rb_left;
1321                 else
1322                         p = &(*p)->rb_right;
1323         }
1324
1325         rb_link_node(&he->rb_node, parent, p);
1326         rb_insert_color(&he->rb_node, &output_hists);
1327 }
1328
1329 static void output__resort(u64 total_samples)
1330 {
1331         struct rb_node *next;
1332         struct hist_entry *n;
1333         struct rb_root *tree = &hist;
1334         u64 min_callchain_hits;
1335
1336         min_callchain_hits = total_samples * (callchain_param.min_percent / 100);
1337
1338         if (sort__need_collapse)
1339                 tree = &collapse_hists;
1340
1341         next = rb_first(tree);
1342
1343         while (next) {
1344                 n = rb_entry(next, struct hist_entry, rb_node);
1345                 next = rb_next(&n->rb_node);
1346
1347                 rb_erase(&n->rb_node, tree);
1348                 output__insert_entry(n, min_callchain_hits);
1349         }
1350 }
1351
1352 static size_t output__fprintf(FILE *fp, u64 total_samples)
1353 {
1354         struct hist_entry *pos;
1355         struct sort_entry *se;
1356         struct rb_node *nd;
1357         size_t ret = 0;
1358         unsigned int width;
1359         char *col_width = col_width_list_str;
1360
1361         fprintf(fp, "# Samples: %Ld\n", (u64)total_samples);
1362         fprintf(fp, "#\n");
1363
1364         fprintf(fp, "# Overhead");
1365         if (show_nr_samples) {
1366                 if (field_sep)
1367                         fprintf(fp, "%cSamples", *field_sep);
1368                 else
1369                         fputs("  Samples  ", fp);
1370         }
1371         list_for_each_entry(se, &hist_entry__sort_list, list) {
1372                 if (se->elide)
1373                         continue;
1374                 if (field_sep) {
1375                         fprintf(fp, "%c%s", *field_sep, se->header);
1376                         continue;
1377                 }
1378                 width = strlen(se->header);
1379                 if (se->width) {
1380                         if (col_width_list_str) {
1381                                 if (col_width) {
1382                                         *se->width = atoi(col_width);
1383                                         col_width = strchr(col_width, ',');
1384                                         if (col_width)
1385                                                 ++col_width;
1386                                 }
1387                         }
1388                         width = *se->width = max(*se->width, width);
1389                 }
1390                 fprintf(fp, "  %*s", width, se->header);
1391         }
1392         fprintf(fp, "\n");
1393
1394         if (field_sep)
1395                 goto print_entries;
1396
1397         fprintf(fp, "# ........");
1398         if (show_nr_samples)
1399                 fprintf(fp, " ..........");
1400         list_for_each_entry(se, &hist_entry__sort_list, list) {
1401                 unsigned int i;
1402
1403                 if (se->elide)
1404                         continue;
1405
1406                 fprintf(fp, "  ");
1407                 if (se->width)
1408                         width = *se->width;
1409                 else
1410                         width = strlen(se->header);
1411                 for (i = 0; i < width; i++)
1412                         fprintf(fp, ".");
1413         }
1414         fprintf(fp, "\n");
1415
1416         fprintf(fp, "#\n");
1417
1418 print_entries:
1419         for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
1420                 pos = rb_entry(nd, struct hist_entry, rb_node);
1421                 ret += hist_entry__fprintf(fp, pos, total_samples);
1422         }
1423
1424         if (sort_order == default_sort_order &&
1425                         parent_pattern == default_parent_pattern) {
1426                 fprintf(fp, "#\n");
1427                 fprintf(fp, "# (For a higher level overview, try: perf report --sort comm,dso)\n");
1428                 fprintf(fp, "#\n");
1429         }
1430         fprintf(fp, "\n");
1431
1432         return ret;
1433 }
1434
1435 static void register_idle_thread(void)
1436 {
1437         struct thread *thread = threads__findnew(0);
1438
1439         if (thread == NULL ||
1440                         thread__set_comm(thread, "[idle]")) {
1441                 fprintf(stderr, "problem inserting idle task.\n");
1442                 exit(-1);
1443         }
1444 }
1445
1446 static unsigned long total = 0,
1447                      total_mmap = 0,
1448                      total_comm = 0,
1449                      total_fork = 0,
1450                      total_unknown = 0,
1451                      total_lost = 0;
1452
1453 static int validate_chain(struct ip_callchain *chain, event_t *event)
1454 {
1455         unsigned int chain_size;
1456
1457         chain_size = event->header.size;
1458         chain_size -= (unsigned long)&event->ip.__more_data - (unsigned long)event;
1459
1460         if (chain->nr*sizeof(u64) > chain_size)
1461                 return -1;
1462
1463         return 0;
1464 }
1465
1466 static int
1467 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
1468 {
1469         char level;
1470         int show = 0;
1471         struct dso *dso = NULL;
1472         struct thread *thread = threads__findnew(event->ip.pid);
1473         u64 ip = event->ip.ip;
1474         u64 period = 1;
1475         struct map *map = NULL;
1476         void *more_data = event->ip.__more_data;
1477         struct ip_callchain *chain = NULL;
1478         int cpumode;
1479
1480         if (sample_type & PERF_SAMPLE_PERIOD) {
1481                 period = *(u64 *)more_data;
1482                 more_data += sizeof(u64);
1483         }
1484
1485         dprintf("%p [%p]: PERF_EVENT_SAMPLE (IP, %d): %d: %p period: %Ld\n",
1486                 (void *)(offset + head),
1487                 (void *)(long)(event->header.size),
1488                 event->header.misc,
1489                 event->ip.pid,
1490                 (void *)(long)ip,
1491                 (long long)period);
1492
1493         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
1494                 unsigned int i;
1495
1496                 chain = (void *)more_data;
1497
1498                 dprintf("... chain: nr:%Lu\n", chain->nr);
1499
1500                 if (validate_chain(chain, event) < 0) {
1501                         eprintf("call-chain problem with event, skipping it.\n");
1502                         return 0;
1503                 }
1504
1505                 if (dump_trace) {
1506                         for (i = 0; i < chain->nr; i++)
1507                                 dprintf("..... %2d: %016Lx\n", i, chain->ips[i]);
1508                 }
1509         }
1510
1511         dprintf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1512
1513         if (thread == NULL) {
1514                 eprintf("problem processing %d event, skipping it.\n",
1515                         event->header.type);
1516                 return -1;
1517         }
1518
1519         if (comm_list && !strlist__has_entry(comm_list, thread->comm))
1520                 return 0;
1521
1522         cpumode = event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK;
1523
1524         if (cpumode == PERF_EVENT_MISC_KERNEL) {
1525                 show = SHOW_KERNEL;
1526                 level = 'k';
1527
1528                 dso = kernel_dso;
1529
1530                 dprintf(" ...... dso: %s\n", dso->name);
1531
1532         } else if (cpumode == PERF_EVENT_MISC_USER) {
1533
1534                 show = SHOW_USER;
1535                 level = '.';
1536
1537         } else {
1538                 show = SHOW_HV;
1539                 level = 'H';
1540
1541                 dso = hypervisor_dso;
1542
1543                 dprintf(" ...... dso: [hypervisor]\n");
1544         }
1545
1546         if (show & show_mask) {
1547                 struct symbol *sym = resolve_symbol(thread, &map, &dso, &ip);
1548
1549                 if (dso_list && dso && dso->name && !strlist__has_entry(dso_list, dso->name))
1550                         return 0;
1551
1552                 if (sym_list && sym && !strlist__has_entry(sym_list, sym->name))
1553                         return 0;
1554
1555                 if (hist_entry__add(thread, map, dso, sym, ip, chain, level, period)) {
1556                         eprintf("problem incrementing symbol count, skipping event\n");
1557                         return -1;
1558                 }
1559         }
1560         total += period;
1561
1562         return 0;
1563 }
1564
1565 static int
1566 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
1567 {
1568         struct thread *thread = threads__findnew(event->mmap.pid);
1569         struct map *map = map__new(&event->mmap);
1570
1571         dprintf("%p [%p]: PERF_EVENT_MMAP %d: [%p(%p) @ %p]: %s\n",
1572                 (void *)(offset + head),
1573                 (void *)(long)(event->header.size),
1574                 event->mmap.pid,
1575                 (void *)(long)event->mmap.start,
1576                 (void *)(long)event->mmap.len,
1577                 (void *)(long)event->mmap.pgoff,
1578                 event->mmap.filename);
1579
1580         if (thread == NULL || map == NULL) {
1581                 dprintf("problem processing PERF_EVENT_MMAP, skipping event.\n");
1582                 return 0;
1583         }
1584
1585         thread__insert_map(thread, map);
1586         total_mmap++;
1587
1588         return 0;
1589 }
1590
1591 static int
1592 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
1593 {
1594         struct thread *thread = threads__findnew(event->comm.pid);
1595
1596         dprintf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
1597                 (void *)(offset + head),
1598                 (void *)(long)(event->header.size),
1599                 event->comm.comm, event->comm.pid);
1600
1601         if (thread == NULL ||
1602             thread__set_comm(thread, event->comm.comm)) {
1603                 dprintf("problem processing PERF_EVENT_COMM, skipping event.\n");
1604                 return -1;
1605         }
1606         total_comm++;
1607
1608         return 0;
1609 }
1610
1611 static int
1612 process_task_event(event_t *event, unsigned long offset, unsigned long head)
1613 {
1614         struct thread *thread = threads__findnew(event->fork.pid);
1615         struct thread *parent = threads__findnew(event->fork.ppid);
1616
1617         dprintf("%p [%p]: PERF_EVENT_%s: (%d:%d):(%d:%d)\n",
1618                 (void *)(offset + head),
1619                 (void *)(long)(event->header.size),
1620                 event->header.type == PERF_EVENT_FORK ? "FORK" : "EXIT",
1621                 event->fork.pid, event->fork.tid,
1622                 event->fork.ppid, event->fork.ptid);
1623
1624         /*
1625          * A thread clone will have the same PID for both
1626          * parent and child.
1627          */
1628         if (thread == parent)
1629                 return 0;
1630
1631         if (event->header.type == PERF_EVENT_EXIT)
1632                 return 0;
1633
1634         if (!thread || !parent || thread__fork(thread, parent)) {
1635                 dprintf("problem processing PERF_EVENT_FORK, skipping event.\n");
1636                 return -1;
1637         }
1638         total_fork++;
1639
1640         return 0;
1641 }
1642
1643 static int
1644 process_lost_event(event_t *event, unsigned long offset, unsigned long head)
1645 {
1646         dprintf("%p [%p]: PERF_EVENT_LOST: id:%Ld: lost:%Ld\n",
1647                 (void *)(offset + head),
1648                 (void *)(long)(event->header.size),
1649                 event->lost.id,
1650                 event->lost.lost);
1651
1652         total_lost += event->lost.lost;
1653
1654         return 0;
1655 }
1656
1657 static void trace_event(event_t *event)
1658 {
1659         unsigned char *raw_event = (void *)event;
1660         char *color = PERF_COLOR_BLUE;
1661         int i, j;
1662
1663         if (!dump_trace)
1664                 return;
1665
1666         dprintf(".");
1667         cdprintf("\n. ... raw event: size %d bytes\n", event->header.size);
1668
1669         for (i = 0; i < event->header.size; i++) {
1670                 if ((i & 15) == 0) {
1671                         dprintf(".");
1672                         cdprintf("  %04x: ", i);
1673                 }
1674
1675                 cdprintf(" %02x", raw_event[i]);
1676
1677                 if (((i & 15) == 15) || i == event->header.size-1) {
1678                         cdprintf("  ");
1679                         for (j = 0; j < 15-(i & 15); j++)
1680                                 cdprintf("   ");
1681                         for (j = 0; j < (i & 15); j++) {
1682                                 if (isprint(raw_event[i-15+j]))
1683                                         cdprintf("%c", raw_event[i-15+j]);
1684                                 else
1685                                         cdprintf(".");
1686                         }
1687                         cdprintf("\n");
1688                 }
1689         }
1690         dprintf(".\n");
1691 }
1692
1693 static int
1694 process_read_event(event_t *event, unsigned long offset, unsigned long head)
1695 {
1696         dprintf("%p [%p]: PERF_EVENT_READ: %d %d %Lu\n",
1697                         (void *)(offset + head),
1698                         (void *)(long)(event->header.size),
1699                         event->read.pid,
1700                         event->read.tid,
1701                         event->read.value);
1702
1703         return 0;
1704 }
1705
1706 static int
1707 process_event(event_t *event, unsigned long offset, unsigned long head)
1708 {
1709         trace_event(event);
1710
1711         switch (event->header.type) {
1712         case PERF_EVENT_SAMPLE:
1713                 return process_sample_event(event, offset, head);
1714
1715         case PERF_EVENT_MMAP:
1716                 return process_mmap_event(event, offset, head);
1717
1718         case PERF_EVENT_COMM:
1719                 return process_comm_event(event, offset, head);
1720
1721         case PERF_EVENT_FORK:
1722         case PERF_EVENT_EXIT:
1723                 return process_task_event(event, offset, head);
1724
1725         case PERF_EVENT_LOST:
1726                 return process_lost_event(event, offset, head);
1727
1728         case PERF_EVENT_READ:
1729                 return process_read_event(event, offset, head);
1730
1731         /*
1732          * We dont process them right now but they are fine:
1733          */
1734
1735         case PERF_EVENT_THROTTLE:
1736         case PERF_EVENT_UNTHROTTLE:
1737                 return 0;
1738
1739         default:
1740                 return -1;
1741         }
1742
1743         return 0;
1744 }
1745
1746 static struct perf_header       *header;
1747
1748 static u64 perf_header__sample_type(void)
1749 {
1750         u64 sample_type = 0;
1751         int i;
1752
1753         for (i = 0; i < header->attrs; i++) {
1754                 struct perf_header_attr *attr = header->attr[i];
1755
1756                 if (!sample_type)
1757                         sample_type = attr->attr.sample_type;
1758                 else if (sample_type != attr->attr.sample_type)
1759                         die("non matching sample_type");
1760         }
1761
1762         return sample_type;
1763 }
1764
1765 static int __cmd_report(void)
1766 {
1767         int ret, rc = EXIT_FAILURE;
1768         unsigned long offset = 0;
1769         unsigned long head, shift;
1770         struct stat stat;
1771         event_t *event;
1772         uint32_t size;
1773         char *buf;
1774
1775         register_idle_thread();
1776
1777         input = open(input_name, O_RDONLY);
1778         if (input < 0) {
1779                 fprintf(stderr, " failed to open file: %s", input_name);
1780                 if (!strcmp(input_name, "perf.data"))
1781                         fprintf(stderr, "  (try 'perf record' first)");
1782                 fprintf(stderr, "\n");
1783                 exit(-1);
1784         }
1785
1786         ret = fstat(input, &stat);
1787         if (ret < 0) {
1788                 perror("failed to stat file");
1789                 exit(-1);
1790         }
1791
1792         if (!stat.st_size) {
1793                 fprintf(stderr, "zero-sized file, nothing to do!\n");
1794                 exit(0);
1795         }
1796
1797         header = perf_header__read(input);
1798         head = header->data_offset;
1799
1800         sample_type = perf_header__sample_type();
1801
1802         if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
1803                 if (sort__has_parent) {
1804                         fprintf(stderr, "selected --sort parent, but no"
1805                                         " callchain data. Did you call"
1806                                         " perf record without -g?\n");
1807                         exit(-1);
1808                 }
1809                 if (callchain) {
1810                         fprintf(stderr, "selected -c but no callchain data."
1811                                         " Did you call perf record without"
1812                                         " -g?\n");
1813                         exit(-1);
1814                 }
1815         }
1816
1817         if (load_kernel() < 0) {
1818                 perror("failed to load kernel symbols");
1819                 return EXIT_FAILURE;
1820         }
1821
1822         if (!full_paths) {
1823                 if (getcwd(__cwd, sizeof(__cwd)) == NULL) {
1824                         perror("failed to get the current directory");
1825                         return EXIT_FAILURE;
1826                 }
1827                 cwdlen = strlen(cwd);
1828         } else {
1829                 cwd = NULL;
1830                 cwdlen = 0;
1831         }
1832
1833         shift = page_size * (head / page_size);
1834         offset += shift;
1835         head -= shift;
1836
1837 remap:
1838         buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1839                            MAP_SHARED, input, offset);
1840         if (buf == MAP_FAILED) {
1841                 perror("failed to mmap file");
1842                 exit(-1);
1843         }
1844
1845 more:
1846         event = (event_t *)(buf + head);
1847
1848         size = event->header.size;
1849         if (!size)
1850                 size = 8;
1851
1852         if (head + event->header.size >= page_size * mmap_window) {
1853                 int ret;
1854
1855                 shift = page_size * (head / page_size);
1856
1857                 ret = munmap(buf, page_size * mmap_window);
1858                 assert(ret == 0);
1859
1860                 offset += shift;
1861                 head -= shift;
1862                 goto remap;
1863         }
1864
1865         size = event->header.size;
1866
1867         dprintf("\n%p [%p]: event: %d\n",
1868                         (void *)(offset + head),
1869                         (void *)(long)event->header.size,
1870                         event->header.type);
1871
1872         if (!size || process_event(event, offset, head) < 0) {
1873
1874                 dprintf("%p [%p]: skipping unknown header type: %d\n",
1875                         (void *)(offset + head),
1876                         (void *)(long)(event->header.size),
1877                         event->header.type);
1878
1879                 total_unknown++;
1880
1881                 /*
1882                  * assume we lost track of the stream, check alignment, and
1883                  * increment a single u64 in the hope to catch on again 'soon'.
1884                  */
1885
1886                 if (unlikely(head & 7))
1887                         head &= ~7ULL;
1888
1889                 size = 8;
1890         }
1891
1892         head += size;
1893
1894         if (offset + head >= header->data_offset + header->data_size)
1895                 goto done;
1896
1897         if (offset + head < (unsigned long)stat.st_size)
1898                 goto more;
1899
1900 done:
1901         rc = EXIT_SUCCESS;
1902         close(input);
1903
1904         dprintf("      IP events: %10ld\n", total);
1905         dprintf("    mmap events: %10ld\n", total_mmap);
1906         dprintf("    comm events: %10ld\n", total_comm);
1907         dprintf("    fork events: %10ld\n", total_fork);
1908         dprintf("    lost events: %10ld\n", total_lost);
1909         dprintf(" unknown events: %10ld\n", total_unknown);
1910
1911         if (dump_trace)
1912                 return 0;
1913
1914         if (verbose >= 3)
1915                 threads__fprintf(stdout);
1916
1917         if (verbose >= 2)
1918                 dsos__fprintf(stdout);
1919
1920         collapse__resort();
1921         output__resort(total);
1922         output__fprintf(stdout, total);
1923
1924         return rc;
1925 }
1926
1927 static int
1928 parse_callchain_opt(const struct option *opt __used, const char *arg,
1929                     int unset __used)
1930 {
1931         char *tok;
1932         char *endptr;
1933
1934         callchain = 1;
1935
1936         if (!arg)
1937                 return 0;
1938
1939         tok = strtok((char *)arg, ",");
1940         if (!tok)
1941                 return -1;
1942
1943         /* get the output mode */
1944         if (!strncmp(tok, "graph", strlen(arg)))
1945                 callchain_param.mode = CHAIN_GRAPH_ABS;
1946
1947         else if (!strncmp(tok, "flat", strlen(arg)))
1948                 callchain_param.mode = CHAIN_FLAT;
1949
1950         else if (!strncmp(tok, "fractal", strlen(arg)))
1951                 callchain_param.mode = CHAIN_GRAPH_REL;
1952
1953         else
1954                 return -1;
1955
1956         /* get the min percentage */
1957         tok = strtok(NULL, ",");
1958         if (!tok)
1959                 goto setup;
1960
1961         callchain_param.min_percent = strtod(tok, &endptr);
1962         if (tok == endptr)
1963                 return -1;
1964
1965 setup:
1966         if (register_callchain_param(&callchain_param) < 0) {
1967                 fprintf(stderr, "Can't register callchain params\n");
1968                 return -1;
1969         }
1970         return 0;
1971 }
1972
1973 static const char * const report_usage[] = {
1974         "perf report [<options>] <command>",
1975         NULL
1976 };
1977
1978 static const struct option options[] = {
1979         OPT_STRING('i', "input", &input_name, "file",
1980                     "input file name"),
1981         OPT_BOOLEAN('v', "verbose", &verbose,
1982                     "be more verbose (show symbol address, etc)"),
1983         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1984                     "dump raw trace in ASCII"),
1985         OPT_STRING('k', "vmlinux", &vmlinux, "file", "vmlinux pathname"),
1986         OPT_BOOLEAN('m', "modules", &modules,
1987                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
1988         OPT_BOOLEAN('n', "show-nr-samples", &show_nr_samples,
1989                     "Show a column with the number of samples"),
1990         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1991                    "sort by key(s): pid, comm, dso, symbol, parent"),
1992         OPT_BOOLEAN('P', "full-paths", &full_paths,
1993                     "Don't shorten the pathnames taking into account the cwd"),
1994         OPT_STRING('p', "parent", &parent_pattern, "regex",
1995                    "regex filter to identify parent, see: '--sort parent'"),
1996         OPT_BOOLEAN('x', "exclude-other", &exclude_other,
1997                     "Only display entries with parent-match"),
1998         OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent",
1999                      "Display callchains using output_type and min percent threshold. "
2000                      "Default: fractal,0.5", &parse_callchain_opt, callchain_default_opt),
2001         OPT_STRING('d', "dsos", &dso_list_str, "dso[,dso...]",
2002                    "only consider symbols in these dsos"),
2003         OPT_STRING('C', "comms", &comm_list_str, "comm[,comm...]",
2004                    "only consider symbols in these comms"),
2005         OPT_STRING('S', "symbols", &sym_list_str, "symbol[,symbol...]",
2006                    "only consider these symbols"),
2007         OPT_STRING('w', "column-widths", &col_width_list_str,
2008                    "width[,width...]",
2009                    "don't try to adjust column width, use these fixed values"),
2010         OPT_STRING('t', "field-separator", &field_sep, "separator",
2011                    "separator for columns, no spaces will be added between "
2012                    "columns '.' is reserved."),
2013         OPT_END()
2014 };
2015
2016 static void setup_sorting(void)
2017 {
2018         char *tmp, *tok, *str = strdup(sort_order);
2019
2020         for (tok = strtok_r(str, ", ", &tmp);
2021                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
2022                 if (sort_dimension__add(tok) < 0) {
2023                         error("Unknown --sort key: `%s'", tok);
2024                         usage_with_options(report_usage, options);
2025                 }
2026         }
2027
2028         free(str);
2029 }
2030
2031 static void setup_list(struct strlist **list, const char *list_str,
2032                        struct sort_entry *se, const char *list_name,
2033                        FILE *fp)
2034 {
2035         if (list_str) {
2036                 *list = strlist__new(true, list_str);
2037                 if (!*list) {
2038                         fprintf(stderr, "problems parsing %s list\n",
2039                                 list_name);
2040                         exit(129);
2041                 }
2042                 if (strlist__nr_entries(*list) == 1) {
2043                         fprintf(fp, "# %s: %s\n", list_name,
2044                                 strlist__entry(*list, 0)->s);
2045                         se->elide = true;
2046                 }
2047         }
2048 }
2049
2050 int cmd_report(int argc, const char **argv, const char *prefix __used)
2051 {
2052         symbol__init();
2053
2054         page_size = getpagesize();
2055
2056         argc = parse_options(argc, argv, options, report_usage, 0);
2057
2058         setup_sorting();
2059
2060         if (parent_pattern != default_parent_pattern) {
2061                 sort_dimension__add("parent");
2062                 sort_parent.elide = 1;
2063         } else
2064                 exclude_other = 0;
2065
2066         /*
2067          * Any (unrecognized) arguments left?
2068          */
2069         if (argc)
2070                 usage_with_options(report_usage, options);
2071
2072         setup_pager();
2073
2074         setup_list(&dso_list, dso_list_str, &sort_dso, "dso", stdout);
2075         setup_list(&comm_list, comm_list_str, &sort_comm, "comm", stdout);
2076         setup_list(&sym_list, sym_list_str, &sort_sym, "symbol", stdout);
2077
2078         if (field_sep && *field_sep == '.') {
2079                 fputs("'.' is the only non valid --field-separator argument\n",
2080                       stderr);
2081                 exit(129);
2082         }
2083
2084         return __cmd_report();
2085 }