tools/perf/util/annotate.c

   1 /*
   2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   3  *
   4  * Parts came from builtin-annotate.c, see those files for further
   5  * copyright notes.
   6  *
   7  * Released under the GPL v2. (and only v2, not any later version)
   8  */
   9
  10 #include <errno.h>
  11 #include <inttypes.h>
  12 #include "util.h"
  13 #include "ui/ui.h"
  14 #include "sort.h"
  15 #include "build-id.h"
  16 #include "color.h"
  17 #include "cache.h"
  18 #include "symbol.h"
  19 #include "debug.h"
  20 #include "annotate.h"
  21 #include "evsel.h"
  22 #include "block-range.h"
  23 #include "string2.h"
  24 #include "arch/common.h"
  25 #include <regex.h>
  26 #include <pthread.h>
  27 #include <linux/bitops.h>
  28 #include <linux/kernel.h>
  29 #include <sys/utsname.h>
  30
  31 #include "sane_ctype.h"
  32
  33 const char      *disassembler_style;
  34 const char      *objdump_path;
  35 static regex_t   file_lineno;
  36
  37 static struct ins_ops *ins__find(struct arch *arch, const char *name);
  38 static void ins__sort(struct arch *arch);
  39 static int disasm_line__parse(char *line, const char **namep, char **rawp);
  40
  41 struct arch {
  42         const char      *name;
  43         struct ins      *instructions;
  44         size_t          nr_instructions;
  45         size_t          nr_instructions_allocated;
  46         struct ins_ops  *(*associate_instruction_ops)(struct arch *arch, const char *name);
  47         bool            sorted_instructions;
  48         bool            initialized;
  49         void            *priv;
  50         unsigned int    model;
  51         unsigned int    family;
  52         int             (*init)(struct arch *arch);
  53         bool            (*ins_is_fused)(struct arch *arch, const char *ins1,
  54                                         const char *ins2);
  55         int             (*cpuid_parse)(struct arch *arch, char *cpuid);
  56         struct          {
  57                 char comment_char;
  58                 char skip_functions_char;
  59         } objdump;
  60 };
  61
  62 static struct ins_ops call_ops;
  63 static struct ins_ops dec_ops;
  64 static struct ins_ops jump_ops;
  65 static struct ins_ops mov_ops;
  66 static struct ins_ops nop_ops;
  67 static struct ins_ops lock_ops;
  68 static struct ins_ops ret_ops;
  69
  70 static int arch__grow_instructions(struct arch *arch)
  71 {
  72         struct ins *new_instructions;
  73         size_t new_nr_allocated;
  74
  75         if (arch->nr_instructions_allocated == 0 && arch->instructions)
  76                 goto grow_from_non_allocated_table;
  77
  78         new_nr_allocated = arch->nr_instructions_allocated + 128;
  79         new_instructions = realloc(arch->instructions, new_nr_allocated * sizeof(struct ins));
  80         if (new_instructions == NULL)
  81                 return -1;
  82
  83 out_update_instructions:
  84         arch->instructions = new_instructions;
  85         arch->nr_instructions_allocated = new_nr_allocated;
  86         return 0;
  87
  88 grow_from_non_allocated_table:
  89         new_nr_allocated = arch->nr_instructions + 128;
  90         new_instructions = calloc(new_nr_allocated, sizeof(struct ins));
  91         if (new_instructions == NULL)
  92                 return -1;
  93
  94         memcpy(new_instructions, arch->instructions, arch->nr_instructions);
  95         goto out_update_instructions;
  96 }
  97
  98 static int arch__associate_ins_ops(struct arch* arch, const char *name, struct ins_ops *ops)
  99 {
 100         struct ins *ins;
 101
 102         if (arch->nr_instructions == arch->nr_instructions_allocated &&
 103             arch__grow_instructions(arch))
 104                 return -1;
 105
 106         ins = &arch->instructions[arch->nr_instructions];
 107         ins->name = strdup(name);
 108         if (!ins->name)
 109                 return -1;
 110
 111         ins->ops  = ops;
 112         arch->nr_instructions++;
 113
 114         ins__sort(arch);
 115         return 0;
 116 }
 117
 118 #include "arch/arm/annotate/instructions.c"
 119 #include "arch/arm64/annotate/instructions.c"
 120 #include "arch/x86/annotate/instructions.c"
 121 #include "arch/powerpc/annotate/instructions.c"
 122 #include "arch/s390/annotate/instructions.c"
 123
 124 static struct arch architectures[] = {
 125         {
 126                 .name = "arm",
 127                 .init = arm__annotate_init,
 128         },
 129         {
 130                 .name = "arm64",
 131                 .init = arm64__annotate_init,
 132         },
 133         {
 134                 .name = "x86",
 135                 .instructions = x86__instructions,
 136                 .nr_instructions = ARRAY_SIZE(x86__instructions),
 137                 .ins_is_fused = x86__ins_is_fused,
 138                 .cpuid_parse = x86__cpuid_parse,
 139                 .objdump =  {
 140                         .comment_char = '#',
 141                 },
 142         },
 143         {
 144                 .name = "powerpc",
 145                 .init = powerpc__annotate_init,
 146         },
 147         {
 148                 .name = "s390",
 149                 .init = s390__annotate_init,
 150                 .objdump =  {
 151                         .comment_char = '#',
 152                 },
 153         },
 154 };
 155
 156 static void ins__delete(struct ins_operands *ops)
 157 {
 158         if (ops == NULL)
 159                 return;
 160         zfree(&ops->source.raw);
 161         zfree(&ops->source.name);
 162         zfree(&ops->target.raw);
 163         zfree(&ops->target.name);
 164 }
 165
 166 static int ins__raw_scnprintf(struct ins *ins, char *bf, size_t size,
 167                               struct ins_operands *ops)
 168 {
 169         return scnprintf(bf, size, "%-6s %s", ins->name, ops->raw);
 170 }
 171
 172 int ins__scnprintf(struct ins *ins, char *bf, size_t size,
 173                   struct ins_operands *ops)
 174 {
 175         if (ins->ops->scnprintf)
 176                 return ins->ops->scnprintf(ins, bf, size, ops);
 177
 178         return ins__raw_scnprintf(ins, bf, size, ops);
 179 }
 180
 181 bool ins__is_fused(struct arch *arch, const char *ins1, const char *ins2)
 182 {
 183         if (!arch || !arch->ins_is_fused)
 184                 return false;
 185
 186         return arch->ins_is_fused(arch, ins1, ins2);
 187 }
 188
 189 static int call__parse(struct arch *arch, struct ins_operands *ops, struct map *map)
 190 {
 191         char *endptr, *tok, *name;
 192
 193         ops->target.addr = strtoull(ops->raw, &endptr, 16);
 194
 195         name = strchr(endptr, '<');
 196         if (name == NULL)
 197                 goto indirect_call;
 198
 199         name++;
 200
 201         if (arch->objdump.skip_functions_char &&
 202             strchr(name, arch->objdump.skip_functions_char))
 203                 return -1;
 204
 205         tok = strchr(name, '>');
 206         if (tok == NULL)
 207                 return -1;
 208
 209         *tok = '\0';
 210         ops->target.name = strdup(name);
 211         *tok = '>';
 212
 213         return ops->target.name == NULL ? -1 : 0;
 214
 215 indirect_call:
 216         tok = strchr(endptr, '*');
 217         if (tok == NULL) {
 218                 struct symbol *sym = map__find_symbol(map, map->map_ip(map, ops->target.addr));
 219                 if (sym != NULL)
 220                         ops->target.name = strdup(sym->name);
 221                 else
 222                         ops->target.addr = 0;
 223                 return 0;
 224         }
 225
 226         ops->target.addr = strtoull(tok + 1, NULL, 16);
 227         return 0;
 228 }
 229
 230 static int call__scnprintf(struct ins *ins, char *bf, size_t size,
 231                            struct ins_operands *ops)
 232 {
 233         if (ops->target.name)
 234                 return scnprintf(bf, size, "%-6s %s", ins->name, ops->target.name);
 235
 236         if (ops->target.addr == 0)
 237                 return ins__raw_scnprintf(ins, bf, size, ops);
 238
 239         return scnprintf(bf, size, "%-6s *%" PRIx64, ins->name, ops->target.addr);
 240 }
 241
 242 static struct ins_ops call_ops = {
 243         .parse     = call__parse,
 244         .scnprintf = call__scnprintf,
 245 };
 246
 247 bool ins__is_call(const struct ins *ins)
 248 {
 249         return ins->ops == &call_ops;
 250 }
 251
 252 static int jump__parse(struct arch *arch __maybe_unused, struct ins_operands *ops, struct map *map __maybe_unused)
 253 {
 254         const char *s = strchr(ops->raw, '+');
 255         const char *c = strchr(ops->raw, ',');
 256
 257         /*
 258          * skip over possible up to 2 operands to get to address, e.g.:
 259          * tbnz  w0, #26, ffff0000083cd190 <security_file_permission+0xd0>
 260          */
 261         if (c++ != NULL) {
 262                 ops->target.addr = strtoull(c, NULL, 16);
 263                 if (!ops->target.addr) {
 264                         c = strchr(c, ',');
 265                         if (c++ != NULL)
 266                                 ops->target.addr = strtoull(c, NULL, 16);
 267                 }
 268         } else {
 269                 ops->target.addr = strtoull(ops->raw, NULL, 16);
 270         }
 271
 272         if (s++ != NULL) {
 273                 ops->target.offset = strtoull(s, NULL, 16);
 274                 ops->target.offset_avail = true;
 275         } else {
 276                 ops->target.offset_avail = false;
 277         }
 278
 279         return 0;
 280 }
 281
 282 static int jump__scnprintf(struct ins *ins, char *bf, size_t size,
 283                            struct ins_operands *ops)
 284 {
 285         const char *c = strchr(ops->raw, ',');
 286
 287         if (!ops->target.addr || ops->target.offset < 0)
 288                 return ins__raw_scnprintf(ins, bf, size, ops);
 289
 290         if (c != NULL) {
 291                 const char *c2 = strchr(c + 1, ',');
 292
 293                 /* check for 3-op insn */
 294                 if (c2 != NULL)
 295                         c = c2;
 296                 c++;
 297
 298                 /* mirror arch objdump's space-after-comma style */
 299                 if (*c == ' ')
 300                         c++;
 301         }
 302
 303         return scnprintf(bf, size, "%-6s %.*s%" PRIx64,
 304                          ins->name, c ? c - ops->raw : 0, ops->raw,
 305                          ops->target.offset);
 306 }
 307
 308 static struct ins_ops jump_ops = {
 309         .parse     = jump__parse,
 310         .scnprintf = jump__scnprintf,
 311 };
 312
 313 bool ins__is_jump(const struct ins *ins)
 314 {
 315         return ins->ops == &jump_ops;
 316 }
 317
 318 static int comment__symbol(char *raw, char *comment, u64 *addrp, char **namep)
 319 {
 320         char *endptr, *name, *t;
 321
 322         if (strstr(raw, "(%rip)") == NULL)
 323                 return 0;
 324
 325         *addrp = strtoull(comment, &endptr, 16);
 326         if (endptr == comment)
 327                 return 0;
 328         name = strchr(endptr, '<');
 329         if (name == NULL)
 330                 return -1;
 331
 332         name++;
 333
 334         t = strchr(name, '>');
 335         if (t == NULL)
 336                 return 0;
 337
 338         *t = '\0';
 339         *namep = strdup(name);
 340         *t = '>';
 341
 342         return 0;
 343 }
 344
 345 static int lock__parse(struct arch *arch, struct ins_operands *ops, struct map *map)
 346 {
 347         ops->locked.ops = zalloc(sizeof(*ops->locked.ops));
 348         if (ops->locked.ops == NULL)
 349                 return 0;
 350
 351         if (disasm_line__parse(ops->raw, &ops->locked.ins.name, &ops->locked.ops->raw) < 0)
 352                 goto out_free_ops;
 353
 354         ops->locked.ins.ops = ins__find(arch, ops->locked.ins.name);
 355
 356         if (ops->locked.ins.ops == NULL)
 357                 goto out_free_ops;
 358
 359         if (ops->locked.ins.ops->parse &&
 360             ops->locked.ins.ops->parse(arch, ops->locked.ops, map) < 0)
 361                 goto out_free_ops;
 362
 363         return 0;
 364
 365 out_free_ops:
 366         zfree(&ops->locked.ops);
 367         return 0;
 368 }
 369
 370 static int lock__scnprintf(struct ins *ins, char *bf, size_t size,
 371                            struct ins_operands *ops)
 372 {
 373         int printed;
 374
 375         if (ops->locked.ins.ops == NULL)
 376                 return ins__raw_scnprintf(ins, bf, size, ops);
 377
 378         printed = scnprintf(bf, size, "%-6s ", ins->name);
 379         return printed + ins__scnprintf(&ops->locked.ins, bf + printed,
 380                                         size - printed, ops->locked.ops);
 381 }
 382
 383 static void lock__delete(struct ins_operands *ops)
 384 {
 385         struct ins *ins = &ops->locked.ins;
 386
 387         if (ins->ops && ins->ops->free)
 388                 ins->ops->free(ops->locked.ops);
 389         else
 390                 ins__delete(ops->locked.ops);
 391
 392         zfree(&ops->locked.ops);
 393         zfree(&ops->target.raw);
 394         zfree(&ops->target.name);
 395 }
 396
 397 static struct ins_ops lock_ops = {
 398         .free      = lock__delete,
 399         .parse     = lock__parse,
 400         .scnprintf = lock__scnprintf,
 401 };
 402
 403 static int mov__parse(struct arch *arch, struct ins_operands *ops, struct map *map __maybe_unused)
 404 {
 405         char *s = strchr(ops->raw, ','), *target, *comment, prev;
 406
 407         if (s == NULL)
 408                 return -1;
 409
 410         *s = '\0';
 411         ops->source.raw = strdup(ops->raw);
 412         *s = ',';
 413
 414         if (ops->source.raw == NULL)
 415                 return -1;
 416
 417         target = ++s;
 418         comment = strchr(s, arch->objdump.comment_char);
 419
 420         if (comment != NULL)
 421                 s = comment - 1;
 422         else
 423                 s = strchr(s, '\0') - 1;
 424
 425         while (s > target && isspace(s[0]))
 426                 --s;
 427         s++;
 428         prev = *s;
 429         *s = '\0';
 430
 431         ops->target.raw = strdup(target);
 432         *s = prev;
 433
 434         if (ops->target.raw == NULL)
 435                 goto out_free_source;
 436
 437         if (comment == NULL)
 438                 return 0;
 439
 440         comment = ltrim(comment);
 441         comment__symbol(ops->source.raw, comment + 1, &ops->source.addr, &ops->source.name);
 442         comment__symbol(ops->target.raw, comment + 1, &ops->target.addr, &ops->target.name);
 443
 444         return 0;
 445
 446 out_free_source:
 447         zfree(&ops->source.raw);
 448         return -1;
 449 }
 450
 451 static int mov__scnprintf(struct ins *ins, char *bf, size_t size,
 452                            struct ins_operands *ops)
 453 {
 454         return scnprintf(bf, size, "%-6s %s,%s", ins->name,
 455                          ops->source.name ?: ops->source.raw,
 456                          ops->target.name ?: ops->target.raw);
 457 }
 458
 459 static struct ins_ops mov_ops = {
 460         .parse     = mov__parse,
 461         .scnprintf = mov__scnprintf,
 462 };
 463
 464 static int dec__parse(struct arch *arch __maybe_unused, struct ins_operands *ops, struct map *map __maybe_unused)
 465 {
 466         char *target, *comment, *s, prev;
 467
 468         target = s = ops->raw;
 469
 470         while (s[0] != '\0' && !isspace(s[0]))
 471                 ++s;
 472         prev = *s;
 473         *s = '\0';
 474
 475         ops->target.raw = strdup(target);
 476         *s = prev;
 477
 478         if (ops->target.raw == NULL)
 479                 return -1;
 480
 481         comment = strchr(s, arch->objdump.comment_char);
 482         if (comment == NULL)
 483                 return 0;
 484
 485         comment = ltrim(comment);
 486         comment__symbol(ops->target.raw, comment + 1, &ops->target.addr, &ops->target.name);
 487
 488         return 0;
 489 }
 490
 491 static int dec__scnprintf(struct ins *ins, char *bf, size_t size,
 492                            struct ins_operands *ops)
 493 {
 494         return scnprintf(bf, size, "%-6s %s", ins->name,
 495                          ops->target.name ?: ops->target.raw);
 496 }
 497
 498 static struct ins_ops dec_ops = {
 499         .parse     = dec__parse,
 500         .scnprintf = dec__scnprintf,
 501 };
 502
 503 static int nop__scnprintf(struct ins *ins __maybe_unused, char *bf, size_t size,
 504                           struct ins_operands *ops __maybe_unused)
 505 {
 506         return scnprintf(bf, size, "%-6s", "nop");
 507 }
 508
 509 static struct ins_ops nop_ops = {
 510         .scnprintf = nop__scnprintf,
 511 };
 512
 513 static struct ins_ops ret_ops = {
 514         .scnprintf = ins__raw_scnprintf,
 515 };
 516
 517 bool ins__is_ret(const struct ins *ins)
 518 {
 519         return ins->ops == &ret_ops;
 520 }
 521
 522 bool ins__is_lock(const struct ins *ins)
 523 {
 524         return ins->ops == &lock_ops;
 525 }
 526
 527 static int ins__key_cmp(const void *name, const void *insp)
 528 {
 529         const struct ins *ins = insp;
 530
 531         return strcmp(name, ins->name);
 532 }
 533
 534 static int ins__cmp(const void *a, const void *b)
 535 {
 536         const struct ins *ia = a;
 537         const struct ins *ib = b;
 538
 539         return strcmp(ia->name, ib->name);
 540 }
 541
 542 static void ins__sort(struct arch *arch)
 543 {
 544         const int nmemb = arch->nr_instructions;
 545
 546         qsort(arch->instructions, nmemb, sizeof(struct ins), ins__cmp);
 547 }
 548
 549 static struct ins_ops *__ins__find(struct arch *arch, const char *name)
 550 {
 551         struct ins *ins;
 552         const int nmemb = arch->nr_instructions;
 553
 554         if (!arch->sorted_instructions) {
 555                 ins__sort(arch);
 556                 arch->sorted_instructions = true;
 557         }
 558
 559         ins = bsearch(name, arch->instructions, nmemb, sizeof(struct ins), ins__key_cmp);
 560         return ins ? ins->ops : NULL;
 561 }
 562
 563 static struct ins_ops *ins__find(struct arch *arch, const char *name)
 564 {
 565         struct ins_ops *ops = __ins__find(arch, name);
 566
 567         if (!ops && arch->associate_instruction_ops)
 568                 ops = arch->associate_instruction_ops(arch, name);
 569
 570         return ops;
 571 }
 572
 573 static int arch__key_cmp(const void *name, const void *archp)
 574 {
 575         const struct arch *arch = archp;
 576
 577         return strcmp(name, arch->name);
 578 }
 579
 580 static int arch__cmp(const void *a, const void *b)
 581 {
 582         const struct arch *aa = a;
 583         const struct arch *ab = b;
 584
 585         return strcmp(aa->name, ab->name);
 586 }
 587
 588 static void arch__sort(void)
 589 {
 590         const int nmemb = ARRAY_SIZE(architectures);
 591
 592         qsort(architectures, nmemb, sizeof(struct arch), arch__cmp);
 593 }
 594
 595 static struct arch *arch__find(const char *name)
 596 {
 597         const int nmemb = ARRAY_SIZE(architectures);
 598         static bool sorted;
 599
 600         if (!sorted) {
 601                 arch__sort();
 602                 sorted = true;
 603         }
 604
 605         return bsearch(name, architectures, nmemb, sizeof(struct arch), arch__key_cmp);
 606 }
 607
 608 int symbol__alloc_hist(struct symbol *sym)
 609 {
 610         struct annotation *notes = symbol__annotation(sym);
 611         size_t size = symbol__size(sym);
 612         size_t sizeof_sym_hist;
 613
 614         /*
 615          * Add buffer of one element for zero length symbol.
 616          * When sample is taken from first instruction of
 617          * zero length symbol, perf still resolves it and
 618          * shows symbol name in perf report and allows to
 619          * annotate it.
 620          */
 621         if (size == 0)
 622                 size = 1;
 623
 624         /* Check for overflow when calculating sizeof_sym_hist */
 625         if (size > (SIZE_MAX - sizeof(struct sym_hist)) / sizeof(struct sym_hist_entry))
 626                 return -1;
 627
 628         sizeof_sym_hist = (sizeof(struct sym_hist) + size * sizeof(struct sym_hist_entry));
 629
 630         /* Check for overflow in zalloc argument */
 631         if (sizeof_sym_hist > (SIZE_MAX - sizeof(*notes->src))
 632                                 / symbol_conf.nr_events)
 633                 return -1;
 634
 635         notes->src = zalloc(sizeof(*notes->src) + symbol_conf.nr_events * sizeof_sym_hist);
 636         if (notes->src == NULL)
 637                 return -1;
 638         notes->src->sizeof_sym_hist = sizeof_sym_hist;
 639         notes->src->nr_histograms   = symbol_conf.nr_events;
 640         INIT_LIST_HEAD(&notes->src->source);
 641         return 0;
 642 }
 643
 644 /* The cycles histogram is lazily allocated. */
 645 static int symbol__alloc_hist_cycles(struct symbol *sym)
 646 {
 647         struct annotation *notes = symbol__annotation(sym);
 648         const size_t size = symbol__size(sym);
 649
 650         notes->src->cycles_hist = calloc(size, sizeof(struct cyc_hist));
 651         if (notes->src->cycles_hist == NULL)
 652                 return -1;
 653         return 0;
 654 }
 655
 656 void symbol__annotate_zero_histograms(struct symbol *sym)
 657 {
 658         struct annotation *notes = symbol__annotation(sym);
 659
 660         pthread_mutex_lock(&notes->lock);
 661         if (notes->src != NULL) {
 662                 memset(notes->src->histograms, 0,
 663                        notes->src->nr_histograms * notes->src->sizeof_sym_hist);
 664                 if (notes->src->cycles_hist)
 665                         memset(notes->src->cycles_hist, 0,
 666                                 symbol__size(sym) * sizeof(struct cyc_hist));
 667         }
 668         pthread_mutex_unlock(&notes->lock);
 669 }
 670
 671 static int __symbol__account_cycles(struct annotation *notes,
 672                                     u64 start,
 673                                     unsigned offset, unsigned cycles,
 674                                     unsigned have_start)
 675 {
 676         struct cyc_hist *ch;
 677
 678         ch = notes->src->cycles_hist;
 679         /*
 680          * For now we can only account one basic block per
 681          * final jump. But multiple could be overlapping.
 682          * Always account the longest one. So when
 683          * a shorter one has been already seen throw it away.
 684          *
 685          * We separately always account the full cycles.
 686          */
 687         ch[offset].num_aggr++;
 688         ch[offset].cycles_aggr += cycles;
 689
 690         if (!have_start && ch[offset].have_start)
 691                 return 0;
 692         if (ch[offset].num) {
 693                 if (have_start && (!ch[offset].have_start ||
 694                                    ch[offset].start > start)) {
 695                         ch[offset].have_start = 0;
 696                         ch[offset].cycles = 0;
 697                         ch[offset].num = 0;
 698                         if (ch[offset].reset < 0xffff)
 699                                 ch[offset].reset++;
 700                 } else if (have_start &&
 701                            ch[offset].start < start)
 702                         return 0;
 703         }
 704         ch[offset].have_start = have_start;
 705         ch[offset].start = start;
 706         ch[offset].cycles += cycles;
 707         ch[offset].num++;
 708         return 0;
 709 }
 710
 711 static int __symbol__inc_addr_samples(struct symbol *sym, struct map *map,
 712                                       struct annotation *notes, int evidx, u64 addr,
 713                                       struct perf_sample *sample)
 714 {
 715         unsigned offset;
 716         struct sym_hist *h;
 717
 718         pr_debug3("%s: addr=%#" PRIx64 "\n", __func__, map->unmap_ip(map, addr));
 719
 720         if ((addr < sym->start || addr >= sym->end) &&
 721             (addr != sym->end || sym->start != sym->end)) {
 722                 pr_debug("%s(%d): ERANGE! sym->name=%s, start=%#" PRIx64 ", addr=%#" PRIx64 ", end=%#" PRIx64 "\n",
 723                        __func__, __LINE__, sym->name, sym->start, addr, sym->end);
 724                 return -ERANGE;
 725         }
 726
 727         offset = addr - sym->start;
 728         h = annotation__histogram(notes, evidx);
 729         h->nr_samples++;
 730         h->addr[offset].nr_samples++;
 731         h->period += sample->period;
 732         h->addr[offset].period += sample->period;
 733
 734         pr_debug3("%#" PRIx64 " %s: period++ [addr: %#" PRIx64 ", %#" PRIx64
 735                   ", evidx=%d] => nr_samples: %" PRIu64 ", period: %" PRIu64 "\n",
 736                   sym->start, sym->name, addr, addr - sym->start, evidx,
 737                   h->addr[offset].nr_samples, h->addr[offset].period);
 738         return 0;
 739 }
 740
 741 static struct annotation *symbol__get_annotation(struct symbol *sym, bool cycles)
 742 {
 743         struct annotation *notes = symbol__annotation(sym);
 744
 745         if (notes->src == NULL) {
 746                 if (symbol__alloc_hist(sym) < 0)
 747                         return NULL;
 748         }
 749         if (!notes->src->cycles_hist && cycles) {
 750                 if (symbol__alloc_hist_cycles(sym) < 0)
 751                         return NULL;
 752         }
 753         return notes;
 754 }
 755
 756 static int symbol__inc_addr_samples(struct symbol *sym, struct map *map,
 757                                     int evidx, u64 addr,
 758                                     struct perf_sample *sample)
 759 {
 760         struct annotation *notes;
 761
 762         if (sym == NULL)
 763                 return 0;
 764         notes = symbol__get_annotation(sym, false);
 765         if (notes == NULL)
 766                 return -ENOMEM;
 767         return __symbol__inc_addr_samples(sym, map, notes, evidx, addr, sample);
 768 }
 769
 770 static int symbol__account_cycles(u64 addr, u64 start,
 771                                   struct symbol *sym, unsigned cycles)
 772 {
 773         struct annotation *notes;
 774         unsigned offset;
 775
 776         if (sym == NULL)
 777                 return 0;
 778         notes = symbol__get_annotation(sym, true);
 779         if (notes == NULL)
 780                 return -ENOMEM;
 781         if (addr < sym->start || addr >= sym->end)
 782                 return -ERANGE;
 783
 784         if (start) {
 785                 if (start < sym->start || start >= sym->end)
 786                         return -ERANGE;
 787                 if (start >= addr)
 788                         start = 0;
 789         }
 790         offset = addr - sym->start;
 791         return __symbol__account_cycles(notes,
 792                                         start ? start - sym->start : 0,
 793                                         offset, cycles,
 794                                         !!start);
 795 }
 796
 797 int addr_map_symbol__account_cycles(struct addr_map_symbol *ams,
 798                                     struct addr_map_symbol *start,
 799                                     unsigned cycles)
 800 {
 801         u64 saddr = 0;
 802         int err;
 803
 804         if (!cycles)
 805                 return 0;
 806
 807         /*
 808          * Only set start when IPC can be computed. We can only
 809          * compute it when the basic block is completely in a single
 810          * function.
 811          * Special case the case when the jump is elsewhere, but
 812          * it starts on the function start.
 813          */
 814         if (start &&
 815                 (start->sym == ams->sym ||
 816                  (ams->sym &&
 817                    start->addr == ams->sym->start + ams->map->start)))
 818                 saddr = start->al_addr;
 819         if (saddr == 0)
 820                 pr_debug2("BB with bad start: addr %"PRIx64" start %"PRIx64" sym %"PRIx64" saddr %"PRIx64"\n",
 821                         ams->addr,
 822                         start ? start->addr : 0,
 823                         ams->sym ? ams->sym->start + ams->map->start : 0,
 824                         saddr);
 825         err = symbol__account_cycles(ams->al_addr, saddr, ams->sym, cycles);
 826         if (err)
 827                 pr_debug2("account_cycles failed %d\n", err);
 828         return err;
 829 }
 830
 831 int addr_map_symbol__inc_samples(struct addr_map_symbol *ams, struct perf_sample *sample,
 832                                  int evidx)
 833 {
 834         return symbol__inc_addr_samples(ams->sym, ams->map, evidx, ams->al_addr, sample);
 835 }
 836
 837 int hist_entry__inc_addr_samples(struct hist_entry *he, struct perf_sample *sample,
 838                                  int evidx, u64 ip)
 839 {
 840         return symbol__inc_addr_samples(he->ms.sym, he->ms.map, evidx, ip, sample);
 841 }
 842
 843 static void disasm_line__init_ins(struct disasm_line *dl, struct arch *arch, struct map *map)
 844 {
 845         dl->ins.ops = ins__find(arch, dl->ins.name);
 846
 847         if (!dl->ins.ops)
 848                 return;
 849
 850         if (dl->ins.ops->parse && dl->ins.ops->parse(arch, &dl->ops, map) < 0)
 851                 dl->ins.ops = NULL;
 852 }
 853
 854 static int disasm_line__parse(char *line, const char **namep, char **rawp)
 855 {
 856         char tmp, *name = ltrim(line);
 857
 858         if (name[0] == '\0')
 859                 return -1;
 860
 861         *rawp = name + 1;
 862
 863         while ((*rawp)[0] != '\0' && !isspace((*rawp)[0]))
 864                 ++*rawp;
 865
 866         tmp = (*rawp)[0];
 867         (*rawp)[0] = '\0';
 868         *namep = strdup(name);
 869
 870         if (*namep == NULL)
 871                 goto out;
 872
 873         (*rawp)[0] = tmp;
 874         *rawp = ltrim(*rawp);
 875
 876         return 0;
 877
 878 out:
 879         return -1;
 880 }
 881
 882 static struct disasm_line *disasm_line__new(s64 offset, char *line,
 883                                             size_t privsize, int line_nr,
 884                                             struct arch *arch,
 885                                             struct map *map)
 886 {
 887         struct disasm_line *dl = zalloc(sizeof(*dl) + privsize);
 888
 889         if (dl != NULL) {
 890                 dl->offset = offset;
 891                 dl->line = strdup(line);
 892                 dl->line_nr = line_nr;
 893                 if (dl->line == NULL)
 894                         goto out_delete;
 895
 896                 if (offset != -1) {
 897                         if (disasm_line__parse(dl->line, &dl->ins.name, &dl->ops.raw) < 0)
 898                                 goto out_free_line;
 899
 900                         disasm_line__init_ins(dl, arch, map);
 901                 }
 902         }
 903
 904         return dl;
 905
 906 out_free_line:
 907         zfree(&dl->line);
 908 out_delete:
 909         free(dl);
 910         return NULL;
 911 }
 912
 913 void disasm_line__free(struct disasm_line *dl)
 914 {
 915         zfree(&dl->line);
 916         if (dl->ins.ops && dl->ins.ops->free)
 917                 dl->ins.ops->free(&dl->ops);
 918         else
 919                 ins__delete(&dl->ops);
 920         free((void *)dl->ins.name);
 921         dl->ins.name = NULL;
 922         free(dl);
 923 }
 924
 925 int disasm_line__scnprintf(struct disasm_line *dl, char *bf, size_t size, bool raw)
 926 {
 927         if (raw || !dl->ins.ops)
 928                 return scnprintf(bf, size, "%-6s %s", dl->ins.name, dl->ops.raw);
 929
 930         return ins__scnprintf(&dl->ins, bf, size, &dl->ops);
 931 }
 932
 933 static void disasm__add(struct list_head *head, struct disasm_line *line)
 934 {
 935         list_add_tail(&line->node, head);
 936 }
 937
 938 struct disasm_line *disasm__get_next_ip_line(struct list_head *head, struct disasm_line *pos)
 939 {
 940         list_for_each_entry_continue(pos, head, node)
 941                 if (pos->offset >= 0)
 942                         return pos;
 943
 944         return NULL;
 945 }
 946
 947 double disasm__calc_percent(struct annotation *notes, int evidx, s64 offset,
 948                             s64 end, const char **path, struct sym_hist_entry *sample)
 949 {
 950         struct source_line *src_line = notes->src->lines;
 951         double percent = 0.0;
 952
 953         sample->nr_samples = sample->period = 0;
 954
 955         if (src_line) {
 956                 size_t sizeof_src_line = sizeof(*src_line) +
 957                                 sizeof(src_line->samples) * (src_line->nr_pcnt - 1);
 958
 959                 while (offset < end) {
 960                         src_line = (void *)notes->src->lines +
 961                                         (sizeof_src_line * offset);
 962
 963                         if (*path == NULL)
 964                                 *path = src_line->path;
 965
 966                         percent += src_line->samples[evidx].percent;
 967                         sample->nr_samples += src_line->samples[evidx].nr;
 968                         offset++;
 969                 }
 970         } else {
 971                 struct sym_hist *h = annotation__histogram(notes, evidx);
 972                 unsigned int hits = 0;
 973                 u64 period = 0;
 974
 975                 while (offset < end) {
 976                         hits   += h->addr[offset].nr_samples;
 977                         period += h->addr[offset].period;
 978                         ++offset;
 979                 }
 980
 981                 if (h->nr_samples) {
 982                         sample->period     = period;
 983                         sample->nr_samples = hits;
 984                         percent = 100.0 * hits / h->nr_samples;
 985                 }
 986         }
 987
 988         return percent;
 989 }
 990
 991 static const char *annotate__address_color(struct block_range *br)
 992 {
 993         double cov = block_range__coverage(br);
 994
 995         if (cov >= 0) {
 996                 /* mark red for >75% coverage */
 997                 if (cov > 0.75)
 998                         return PERF_COLOR_RED;
 999
1000                 /* mark dull for <1% coverage */
1001                 if (cov < 0.01)
1002                         return PERF_COLOR_NORMAL;
1003         }
1004
1005         return PERF_COLOR_MAGENTA;
1006 }
1007
1008 static const char *annotate__asm_color(struct block_range *br)
1009 {
1010         double cov = block_range__coverage(br);
1011
1012         if (cov >= 0) {
1013                 /* mark dull for <1% coverage */
1014                 if (cov < 0.01)
1015                         return PERF_COLOR_NORMAL;
1016         }
1017
1018         return PERF_COLOR_BLUE;
1019 }
1020
1021 static void annotate__branch_printf(struct block_range *br, u64 addr)
1022 {
1023         bool emit_comment = true;
1024
1025         if (!br)
1026                 return;
1027
1028 #if 1
1029         if (br->is_target && br->start == addr) {
1030                 struct block_range *branch = br;
1031                 double p;
1032
1033                 /*
1034                  * Find matching branch to our target.
1035                  */
1036                 while (!branch->is_branch)
1037                         branch = block_range__next(branch);
1038
1039                 p = 100 *(double)br->entry / branch->coverage;
1040
1041                 if (p > 0.1) {
1042                         if (emit_comment) {
1043                                 emit_comment = false;
1044                                 printf("\t#");
1045                         }
1046
1047                         /*
1048                          * The percentage of coverage joined at this target in relation
1049                          * to the next branch.
1050                          */
1051                         printf(" +%.2f%%", p);
1052                 }
1053         }
1054 #endif
1055         if (br->is_branch && br->end == addr) {
1056                 double p = 100*(double)br->taken / br->coverage;
1057
1058                 if (p > 0.1) {
1059                         if (emit_comment) {
1060                                 emit_comment = false;
1061                                 printf("\t#");
1062                         }
1063
1064                         /*
1065                          * The percentage of coverage leaving at this branch, and
1066                          * its prediction ratio.
1067                          */
1068                         printf(" -%.2f%% (p:%.2f%%)", p, 100*(double)br->pred  / br->taken);
1069                 }
1070         }
1071 }
1072
1073
1074 static int disasm_line__print(struct disasm_line *dl, struct symbol *sym, u64 start,
1075                       struct perf_evsel *evsel, u64 len, int min_pcnt, int printed,
1076                       int max_lines, struct disasm_line *queue)
1077 {
1078         static const char *prev_line;
1079         static const char *prev_color;
1080
1081         if (dl->offset != -1) {
1082                 const char *path = NULL;
1083                 double percent, max_percent = 0.0;
1084                 double *ppercents = &percent;
1085                 struct sym_hist_entry sample;
1086                 struct sym_hist_entry *psamples = &sample;
1087                 int i, nr_percent = 1;
1088                 const char *color;
1089                 struct annotation *notes = symbol__annotation(sym);
1090                 s64 offset = dl->offset;
1091                 const u64 addr = start + offset;
1092                 struct disasm_line *next;
1093                 struct block_range *br;
1094
1095                 next = disasm__get_next_ip_line(&notes->src->source, dl);
1096
1097                 if (perf_evsel__is_group_event(evsel)) {
1098                         nr_percent = evsel->nr_members;
1099                         ppercents = calloc(nr_percent, sizeof(double));
1100                         psamples = calloc(nr_percent, sizeof(struct sym_hist_entry));
1101                         if (ppercents == NULL || psamples == NULL) {
1102                                 return -1;
1103                         }
1104                 }
1105
1106                 for (i = 0; i < nr_percent; i++) {
1107                         percent = disasm__calc_percent(notes,
1108                                         notes->src->lines ? i : evsel->idx + i,
1109                                         offset,
1110                                         next ? next->offset : (s64) len,
1111                                         &path, &sample);
1112
1113                         ppercents[i] = percent;
1114                         psamples[i] = sample;
1115                         if (percent > max_percent)
1116                                 max_percent = percent;
1117                 }
1118
1119                 if (max_percent < min_pcnt)
1120                         return -1;
1121
1122                 if (max_lines && printed >= max_lines)
1123                         return 1;
1124
1125                 if (queue != NULL) {
1126                         list_for_each_entry_from(queue, &notes->src->source, node) {
1127                                 if (queue == dl)
1128                                         break;
1129                                 disasm_line__print(queue, sym, start, evsel, len,
1130                                                     0, 0, 1, NULL);
1131                         }
1132                 }
1133
1134                 color = get_percent_color(max_percent);
1135
1136                 /*
1137                  * Also color the filename and line if needed, with
1138                  * the same color than the percentage. Don't print it
1139                  * twice for close colored addr with the same filename:line
1140                  */
1141                 if (path) {
1142                         if (!prev_line || strcmp(prev_line, path)
1143                                        || color != prev_color) {
1144                                 color_fprintf(stdout, color, " %s", path);
1145                                 prev_line = path;
1146                                 prev_color = color;
1147                         }
1148                 }
1149
1150                 for (i = 0; i < nr_percent; i++) {
1151                         percent = ppercents[i];
1152                         sample = psamples[i];
1153                         color = get_percent_color(percent);
1154
1155                         if (symbol_conf.show_total_period)
1156                                 color_fprintf(stdout, color, " %11" PRIu64,
1157                                               sample.period);
1158                         else if (symbol_conf.show_nr_samples)
1159                                 color_fprintf(stdout, color, " %7" PRIu64,
1160                                               sample.nr_samples);
1161                         else
1162                                 color_fprintf(stdout, color, " %7.2f", percent);
1163                 }
1164
1165                 printf(" :      ");
1166
1167                 br = block_range__find(addr);
1168                 color_fprintf(stdout, annotate__address_color(br), "  %" PRIx64 ":", addr);
1169                 color_fprintf(stdout, annotate__asm_color(br), "%s", dl->line);
1170                 annotate__branch_printf(br, addr);
1171                 printf("\n");
1172
1173                 if (ppercents != &percent)
1174                         free(ppercents);
1175
1176                 if (psamples != &sample)
1177                         free(psamples);
1178
1179         } else if (max_lines && printed >= max_lines)
1180                 return 1;
1181         else {
1182                 int width = symbol_conf.show_total_period ? 12 : 8;
1183
1184                 if (queue)
1185                         return -1;
1186
1187                 if (perf_evsel__is_group_event(evsel))
1188                         width *= evsel->nr_members;
1189
1190                 if (!*dl->line)
1191                         printf(" %*s:\n", width, " ");
1192                 else
1193                         printf(" %*s:   %s\n", width, " ", dl->line);
1194         }
1195
1196         return 0;
1197 }
1198
1199 /*
1200  * symbol__parse_objdump_line() parses objdump output (with -d --no-show-raw)
1201  * which looks like following
1202  *
1203  *  0000000000415500 <_init>:
1204  *    415500:       sub    $0x8,%rsp
1205  *    415504:       mov    0x2f5ad5(%rip),%rax        # 70afe0 <_DYNAMIC+0x2f8>
1206  *    41550b:       test   %rax,%rax
1207  *    41550e:       je     415515 <_init+0x15>
1208  *    415510:       callq  416e70 <__gmon_start__@plt>
1209  *    415515:       add    $0x8,%rsp
1210  *    415519:       retq
1211  *
1212  * it will be parsed and saved into struct disasm_line as
1213  *  <offset>       <name>  <ops.raw>
1214  *
1215  * The offset will be a relative offset from the start of the symbol and -1
1216  * means that it's not a disassembly line so should be treated differently.
1217  * The ops.raw part will be parsed further according to type of the instruction.
1218  */
1219 static int symbol__parse_objdump_line(struct symbol *sym, struct map *map,
1220                                       struct arch *arch,
1221                                       FILE *file, size_t privsize,
1222                                       int *line_nr)
1223 {
1224         struct annotation *notes = symbol__annotation(sym);
1225         struct disasm_line *dl;
1226         char *line = NULL, *parsed_line, *tmp, *tmp2;
1227         size_t line_len;
1228         s64 line_ip, offset = -1;
1229         regmatch_t match[2];
1230
1231         if (getline(&line, &line_len, file) < 0)
1232                 return -1;
1233
1234         if (!line)
1235                 return -1;
1236
1237         line_ip = -1;
1238         parsed_line = rtrim(line);
1239
1240         /* /filename:linenr ? Save line number and ignore. */
1241         if (regexec(&file_lineno, parsed_line, 2, match, 0) == 0) {
1242                 *line_nr = atoi(parsed_line + match[1].rm_so);
1243                 return 0;
1244         }
1245
1246         tmp = ltrim(parsed_line);
1247         if (*tmp) {
1248                 /*
1249                  * Parse hexa addresses followed by ':'
1250                  */
1251                 line_ip = strtoull(tmp, &tmp2, 16);
1252                 if (*tmp2 != ':' || tmp == tmp2 || tmp2[1] == '\0')
1253                         line_ip = -1;
1254         }
1255
1256         if (line_ip != -1) {
1257                 u64 start = map__rip_2objdump(map, sym->start),
1258                     end = map__rip_2objdump(map, sym->end);
1259
1260                 offset = line_ip - start;
1261                 if ((u64)line_ip < start || (u64)line_ip >= end)
1262                         offset = -1;
1263                 else
1264                         parsed_line = tmp2 + 1;
1265         }
1266
1267         dl = disasm_line__new(offset, parsed_line, privsize, *line_nr, arch, map);
1268         free(line);
1269         (*line_nr)++;
1270
1271         if (dl == NULL)
1272                 return -1;
1273
1274         if (!disasm_line__has_offset(dl)) {
1275                 dl->ops.target.offset = dl->ops.target.addr -
1276                                         map__rip_2objdump(map, sym->start);
1277                 dl->ops.target.offset_avail = true;
1278         }
1279
1280         /* kcore has no symbols, so add the call target name */
1281         if (dl->ins.ops && ins__is_call(&dl->ins) && !dl->ops.target.name) {
1282                 struct addr_map_symbol target = {
1283                         .map = map,
1284                         .addr = dl->ops.target.addr,
1285                 };
1286
1287                 if (!map_groups__find_ams(&target) &&
1288                     target.sym->start == target.al_addr)
1289                         dl->ops.target.name = strdup(target.sym->name);
1290         }
1291
1292         disasm__add(&notes->src->source, dl);
1293
1294         return 0;
1295 }
1296
1297 static __attribute__((constructor)) void symbol__init_regexpr(void)
1298 {
1299         regcomp(&file_lineno, "^/[^:]+:([0-9]+)", REG_EXTENDED);
1300 }
1301
1302 static void delete_last_nop(struct symbol *sym)
1303 {
1304         struct annotation *notes = symbol__annotation(sym);
1305         struct list_head *list = &notes->src->source;
1306         struct disasm_line *dl;
1307
1308         while (!list_empty(list)) {
1309                 dl = list_entry(list->prev, struct disasm_line, node);
1310
1311                 if (dl->ins.ops) {
1312                         if (dl->ins.ops != &nop_ops)
1313                                 return;
1314                 } else {
1315                         if (!strstr(dl->line, " nop ") &&
1316                             !strstr(dl->line, " nopl ") &&
1317                             !strstr(dl->line, " nopw "))
1318                                 return;
1319                 }
1320
1321                 list_del(&dl->node);
1322                 disasm_line__free(dl);
1323         }
1324 }
1325
1326 int symbol__strerror_disassemble(struct symbol *sym __maybe_unused, struct map *map,
1327                               int errnum, char *buf, size_t buflen)
1328 {
1329         struct dso *dso = map->dso;
1330
1331         BUG_ON(buflen == 0);
1332
1333         if (errnum >= 0) {
1334                 str_error_r(errnum, buf, buflen);
1335                 return 0;
1336         }
1337
1338         switch (errnum) {
1339         case SYMBOL_ANNOTATE_ERRNO__NO_VMLINUX: {
1340                 char bf[SBUILD_ID_SIZE + 15] = " with build id ";
1341                 char *build_id_msg = NULL;
1342
1343                 if (dso->has_build_id) {
1344                         build_id__sprintf(dso->build_id,
1345                                           sizeof(dso->build_id), bf + 15);
1346                         build_id_msg = bf;
1347                 }
1348                 scnprintf(buf, buflen,
1349                           "No vmlinux file%s\nwas found in the path.\n\n"
1350                           "Note that annotation using /proc/kcore requires CAP_SYS_RAWIO capability.\n\n"
1351                           "Please use:\n\n"
1352                           "  perf buildid-cache -vu vmlinux\n\n"
1353                           "or:\n\n"
1354                           "  --vmlinux vmlinux\n", build_id_msg ?: "");
1355         }
1356                 break;
1357         default:
1358                 scnprintf(buf, buflen, "Internal error: Invalid %d error code\n", errnum);
1359                 break;
1360         }
1361
1362         return 0;
1363 }
1364
1365 static int dso__disassemble_filename(struct dso *dso, char *filename, size_t filename_size)
1366 {
1367         char linkname[PATH_MAX];
1368         char *build_id_filename;
1369         char *build_id_path = NULL;
1370         char *pos;
1371
1372         if (dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
1373             !dso__is_kcore(dso))
1374                 return SYMBOL_ANNOTATE_ERRNO__NO_VMLINUX;
1375
1376         build_id_filename = dso__build_id_filename(dso, NULL, 0, false);
1377         if (build_id_filename) {
1378                 __symbol__join_symfs(filename, filename_size, build_id_filename);
1379                 free(build_id_filename);
1380         } else {
1381                 if (dso->has_build_id)
1382                         return ENOMEM;
1383                 goto fallback;
1384         }
1385
1386         build_id_path = strdup(filename);
1387         if (!build_id_path)
1388                 return -1;
1389
1390         /*
1391          * old style build-id cache has name of XX/XXXXXXX.. while
1392          * new style has XX/XXXXXXX../{elf,kallsyms,vdso}.
1393          * extract the build-id part of dirname in the new style only.
1394          */
1395         pos = strrchr(build_id_path, '/');
1396         if (pos && strlen(pos) < SBUILD_ID_SIZE - 2)
1397                 dirname(build_id_path);
1398
1399         if (dso__is_kcore(dso) ||
1400             readlink(build_id_path, linkname, sizeof(linkname)) < 0 ||
1401             strstr(linkname, DSO__NAME_KALLSYMS) ||
1402             access(filename, R_OK)) {
1403 fallback:
1404                 /*
1405                  * If we don't have build-ids or the build-id file isn't in the
1406                  * cache, or is just a kallsyms file, well, lets hope that this
1407                  * DSO is the same as when 'perf record' ran.
1408                  */
1409                 __symbol__join_symfs(filename, filename_size, dso->long_name);
1410         }
1411
1412         free(build_id_path);
1413         return 0;
1414 }
1415
1416 static const char *annotate__norm_arch(const char *arch_name)
1417 {
1418         struct utsname uts;
1419
1420         if (!arch_name) { /* Assume we are annotating locally. */
1421                 if (uname(&uts) < 0)
1422                         return NULL;
1423                 arch_name = uts.machine;
1424         }
1425         return normalize_arch((char *)arch_name);
1426 }
1427
1428 int symbol__disassemble(struct symbol *sym, struct map *map,
1429                         const char *arch_name, size_t privsize,
1430                         struct arch **parch, char *cpuid)
1431 {
1432         struct dso *dso = map->dso;
1433         char *command;
1434         struct arch *arch = NULL;
1435         FILE *file;
1436         char symfs_filename[PATH_MAX];
1437         struct kcore_extract kce;
1438         bool delete_extract = false;
1439         int stdout_fd[2];
1440         int lineno = 0;
1441         int nline;
1442         pid_t pid;
1443         int err = dso__disassemble_filename(dso, symfs_filename, sizeof(symfs_filename));
1444
1445         if (err)
1446                 return err;
1447
1448         arch_name = annotate__norm_arch(arch_name);
1449         if (!arch_name)
1450                 return -1;
1451
1452         arch = arch__find(arch_name);
1453         if (arch == NULL)
1454                 return -ENOTSUP;
1455
1456         if (parch)
1457                 *parch = arch;
1458
1459         if (arch->init) {
1460                 err = arch->init(arch);
1461                 if (err) {
1462                         pr_err("%s: failed to initialize %s arch priv area\n", __func__, arch->name);
1463                         return err;
1464                 }
1465         }
1466
1467         if (arch->cpuid_parse && cpuid)
1468                 arch->cpuid_parse(arch, cpuid);
1469
1470         pr_debug("%s: filename=%s, sym=%s, start=%#" PRIx64 ", end=%#" PRIx64 "\n", __func__,
1471                  symfs_filename, sym->name, map->unmap_ip(map, sym->start),
1472                  map->unmap_ip(map, sym->end));
1473
1474         pr_debug("annotating [%p] %30s : [%p] %30s\n",
1475                  dso, dso->long_name, sym, sym->name);
1476
1477         if (dso__is_kcore(dso)) {
1478                 kce.kcore_filename = symfs_filename;
1479                 kce.addr = map__rip_2objdump(map, sym->start);
1480                 kce.offs = sym->start;
1481                 kce.len = sym->end - sym->start;
1482                 if (!kcore_extract__create(&kce)) {
1483                         delete_extract = true;
1484                         strlcpy(symfs_filename, kce.extract_filename,
1485                                 sizeof(symfs_filename));
1486                 }
1487         } else if (dso__needs_decompress(dso)) {
1488                 char tmp[KMOD_DECOMP_LEN];
1489
1490                 if (dso__decompress_kmodule_path(dso, symfs_filename,
1491                                                  tmp, sizeof(tmp)) < 0)
1492                         goto out;
1493
1494                 strcpy(symfs_filename, tmp);
1495         }
1496
1497         err = asprintf(&command,
1498                  "%s %s%s --start-address=0x%016" PRIx64
1499                  " --stop-address=0x%016" PRIx64
1500                  " -l -d %s %s -C \"%s\" 2>/dev/null|grep -v \"%s:\"|expand",
1501                  objdump_path ? objdump_path : "objdump",
1502                  disassembler_style ? "-M " : "",
1503                  disassembler_style ? disassembler_style : "",
1504                  map__rip_2objdump(map, sym->start),
1505                  map__rip_2objdump(map, sym->end),
1506                  symbol_conf.annotate_asm_raw ? "" : "--no-show-raw",
1507                  symbol_conf.annotate_src ? "-S" : "",
1508                  symfs_filename, symfs_filename);
1509
1510         if (err < 0) {
1511                 pr_err("Failure allocating memory for the command to run\n");
1512                 goto out_remove_tmp;
1513         }
1514
1515         pr_debug("Executing: %s\n", command);
1516
1517         err = -1;
1518         if (pipe(stdout_fd) < 0) {
1519                 pr_err("Failure creating the pipe to run %s\n", command);
1520                 goto out_free_command;
1521         }
1522
1523         pid = fork();
1524         if (pid < 0) {
1525                 pr_err("Failure forking to run %s\n", command);
1526                 goto out_close_stdout;
1527         }
1528
1529         if (pid == 0) {
1530                 close(stdout_fd[0]);
1531                 dup2(stdout_fd[1], 1);
1532                 close(stdout_fd[1]);
1533                 execl("/bin/sh", "sh", "-c", command, NULL);
1534                 perror(command);
1535                 exit(-1);
1536         }
1537
1538         close(stdout_fd[1]);
1539
1540         file = fdopen(stdout_fd[0], "r");
1541         if (!file) {
1542                 pr_err("Failure creating FILE stream for %s\n", command);
1543                 /*
1544                  * If we were using debug info should retry with
1545                  * original binary.
1546                  */
1547                 goto out_free_command;
1548         }
1549
1550         nline = 0;
1551         while (!feof(file)) {
1552                 /*
1553                  * The source code line number (lineno) needs to be kept in
1554                  * accross calls to symbol__parse_objdump_line(), so that it
1555                  * can associate it with the instructions till the next one.
1556                  * See disasm_line__new() and struct disasm_line::line_nr.
1557                  */
1558                 if (symbol__parse_objdump_line(sym, map, arch, file, privsize,
1559                             &lineno) < 0)
1560                         break;
1561                 nline++;
1562         }
1563
1564         if (nline == 0)
1565                 pr_err("No output from %s\n", command);
1566
1567         /*
1568          * kallsyms does not have symbol sizes so there may a nop at the end.
1569          * Remove it.
1570          */
1571         if (dso__is_kcore(dso))
1572                 delete_last_nop(sym);
1573
1574         fclose(file);
1575         err = 0;
1576 out_free_command:
1577         free(command);
1578 out_remove_tmp:
1579         close(stdout_fd[0]);
1580
1581         if (dso__needs_decompress(dso))
1582                 unlink(symfs_filename);
1583
1584         if (delete_extract)
1585                 kcore_extract__delete(&kce);
1586 out:
1587         return err;
1588
1589 out_close_stdout:
1590         close(stdout_fd[1]);
1591         goto out_free_command;
1592 }
1593
1594 static void insert_source_line(struct rb_root *root, struct source_line *src_line)
1595 {
1596         struct source_line *iter;
1597         struct rb_node **p = &root->rb_node;
1598         struct rb_node *parent = NULL;
1599         int i, ret;
1600
1601         while (*p != NULL) {
1602                 parent = *p;
1603                 iter = rb_entry(parent, struct source_line, node);
1604
1605                 ret = strcmp(iter->path, src_line->path);
1606                 if (ret == 0) {
1607                         for (i = 0; i < src_line->nr_pcnt; i++)
1608                                 iter->samples[i].percent_sum += src_line->samples[i].percent;
1609                         return;
1610                 }
1611
1612                 if (ret < 0)
1613                         p = &(*p)->rb_left;
1614                 else
1615                         p = &(*p)->rb_right;
1616         }
1617
1618         for (i = 0; i < src_line->nr_pcnt; i++)
1619                 src_line->samples[i].percent_sum = src_line->samples[i].percent;
1620
1621         rb_link_node(&src_line->node, parent, p);
1622         rb_insert_color(&src_line->node, root);
1623 }
1624
1625 static int cmp_source_line(struct source_line *a, struct source_line *b)
1626 {
1627         int i;
1628
1629         for (i = 0; i < a->nr_pcnt; i++) {
1630                 if (a->samples[i].percent_sum == b->samples[i].percent_sum)
1631                         continue;
1632                 return a->samples[i].percent_sum > b->samples[i].percent_sum;
1633         }
1634
1635         return 0;
1636 }
1637
1638 static void __resort_source_line(struct rb_root *root, struct source_line *src_line)
1639 {
1640         struct source_line *iter;
1641         struct rb_node **p = &root->rb_node;
1642         struct rb_node *parent = NULL;
1643
1644         while (*p != NULL) {
1645                 parent = *p;
1646                 iter = rb_entry(parent, struct source_line, node);
1647
1648                 if (cmp_source_line(src_line, iter))
1649                         p = &(*p)->rb_left;
1650                 else
1651                         p = &(*p)->rb_right;
1652         }
1653
1654         rb_link_node(&src_line->node, parent, p);
1655         rb_insert_color(&src_line->node, root);
1656 }
1657
1658 static void resort_source_line(struct rb_root *dest_root, struct rb_root *src_root)
1659 {
1660         struct source_line *src_line;
1661         struct rb_node *node;
1662
1663         node = rb_first(src_root);
1664         while (node) {
1665                 struct rb_node *next;
1666
1667                 src_line = rb_entry(node, struct source_line, node);
1668                 next = rb_next(node);
1669                 rb_erase(node, src_root);
1670
1671                 __resort_source_line(dest_root, src_line);
1672                 node = next;
1673         }
1674 }
1675
1676 static void symbol__free_source_line(struct symbol *sym, int len)
1677 {
1678         struct annotation *notes = symbol__annotation(sym);
1679         struct source_line *src_line = notes->src->lines;
1680         size_t sizeof_src_line;
1681         int i;
1682
1683         sizeof_src_line = sizeof(*src_line) +
1684                           (sizeof(src_line->samples) * (src_line->nr_pcnt - 1));
1685
1686         for (i = 0; i < len; i++) {
1687                 free_srcline(src_line->path);
1688                 src_line = (void *)src_line + sizeof_src_line;
1689         }
1690
1691         zfree(&notes->src->lines);
1692 }
1693
1694 /* Get the filename:line for the colored entries */
1695 static int symbol__get_source_line(struct symbol *sym, struct map *map,
1696                                    struct perf_evsel *evsel,
1697                                    struct rb_root *root, int len)
1698 {
1699         u64 start;
1700         int i, k;
1701         int evidx = evsel->idx;
1702         struct source_line *src_line;
1703         struct annotation *notes = symbol__annotation(sym);
1704         struct sym_hist *h = annotation__histogram(notes, evidx);
1705         struct rb_root tmp_root = RB_ROOT;
1706         int nr_pcnt = 1;
1707         u64 nr_samples = h->nr_samples;
1708         size_t sizeof_src_line = sizeof(struct source_line);
1709
1710         if (perf_evsel__is_group_event(evsel)) {
1711                 for (i = 1; i < evsel->nr_members; i++) {
1712                         h = annotation__histogram(notes, evidx + i);
1713                         nr_samples += h->nr_samples;
1714                 }
1715                 nr_pcnt = evsel->nr_members;
1716                 sizeof_src_line += (nr_pcnt - 1) * sizeof(src_line->samples);
1717         }
1718
1719         if (!nr_samples)
1720                 return 0;
1721
1722         src_line = notes->src->lines = calloc(len, sizeof_src_line);
1723         if (!notes->src->lines)
1724                 return -1;
1725
1726         start = map__rip_2objdump(map, sym->start);
1727
1728         for (i = 0; i < len; i++) {
1729                 u64 offset;
1730                 double percent_max = 0.0;
1731
1732                 src_line->nr_pcnt = nr_pcnt;
1733
1734                 for (k = 0; k < nr_pcnt; k++) {
1735                         double percent = 0.0;
1736
1737                         h = annotation__histogram(notes, evidx + k);
1738                         nr_samples = h->addr[i].nr_samples;
1739                         if (h->nr_samples)
1740                                 percent = 100.0 * nr_samples / h->nr_samples;
1741
1742                         if (percent > percent_max)
1743                                 percent_max = percent;
1744                         src_line->samples[k].percent = percent;
1745                         src_line->samples[k].nr = nr_samples;
1746                 }
1747
1748                 if (percent_max <= 0.5)
1749                         goto next;
1750
1751                 offset = start + i;
1752                 src_line->path = get_srcline(map->dso, offset, NULL,
1753                                              false, true);
1754                 insert_source_line(&tmp_root, src_line);
1755
1756         next:
1757                 src_line = (void *)src_line + sizeof_src_line;
1758         }
1759
1760         resort_source_line(root, &tmp_root);
1761         return 0;
1762 }
1763
1764 static void print_summary(struct rb_root *root, const char *filename)
1765 {
1766         struct source_line *src_line;
1767         struct rb_node *node;
1768
1769         printf("\nSorted summary for file %s\n", filename);
1770         printf("----------------------------------------------\n\n");
1771
1772         if (RB_EMPTY_ROOT(root)) {
1773                 printf(" Nothing higher than %1.1f%%\n", MIN_GREEN);
1774                 return;
1775         }
1776
1777         node = rb_first(root);
1778         while (node) {
1779                 double percent, percent_max = 0.0;
1780                 const char *color;
1781                 char *path;
1782                 int i;
1783
1784                 src_line = rb_entry(node, struct source_line, node);
1785                 for (i = 0; i < src_line->nr_pcnt; i++) {
1786                         percent = src_line->samples[i].percent_sum;
1787                         color = get_percent_color(percent);
1788                         color_fprintf(stdout, color, " %7.2f", percent);
1789
1790                         if (percent > percent_max)
1791                                 percent_max = percent;
1792                 }
1793
1794                 path = src_line->path;
1795                 color = get_percent_color(percent_max);
1796                 color_fprintf(stdout, color, " %s\n", path);
1797
1798                 node = rb_next(node);
1799         }
1800 }
1801
1802 static void symbol__annotate_hits(struct symbol *sym, struct perf_evsel *evsel)
1803 {
1804         struct annotation *notes = symbol__annotation(sym);
1805         struct sym_hist *h = annotation__histogram(notes, evsel->idx);
1806         u64 len = symbol__size(sym), offset;
1807
1808         for (offset = 0; offset < len; ++offset)
1809                 if (h->addr[offset].nr_samples != 0)
1810                         printf("%*" PRIx64 ": %" PRIu64 "\n", BITS_PER_LONG / 2,
1811                                sym->start + offset, h->addr[offset].nr_samples);
1812         printf("%*s: %" PRIu64 "\n", BITS_PER_LONG / 2, "h->nr_samples", h->nr_samples);
1813 }
1814
1815 int symbol__annotate_printf(struct symbol *sym, struct map *map,
1816                             struct perf_evsel *evsel, bool full_paths,
1817                             int min_pcnt, int max_lines, int context)
1818 {
1819         struct dso *dso = map->dso;
1820         char *filename;
1821         const char *d_filename;
1822         const char *evsel_name = perf_evsel__name(evsel);
1823         struct annotation *notes = symbol__annotation(sym);
1824         struct sym_hist *h = annotation__histogram(notes, evsel->idx);
1825         struct disasm_line *pos, *queue = NULL;
1826         u64 start = map__rip_2objdump(map, sym->start);
1827         int printed = 2, queue_len = 0;
1828         int more = 0;
1829         u64 len;
1830         int width = symbol_conf.show_total_period ? 12 : 8;
1831         int graph_dotted_len;
1832
1833         filename = strdup(dso->long_name);
1834         if (!filename)
1835                 return -ENOMEM;
1836
1837         if (full_paths)
1838                 d_filename = filename;
1839         else
1840                 d_filename = basename(filename);
1841
1842         len = symbol__size(sym);
1843
1844         if (perf_evsel__is_group_event(evsel))
1845                 width *= evsel->nr_members;
1846
1847         graph_dotted_len = printf(" %-*.*s|     Source code & Disassembly of %s for %s (%" PRIu64 " samples)\n",
1848                                   width, width, symbol_conf.show_total_period ? "Period" :
1849                                   symbol_conf.show_nr_samples ? "Samples" : "Percent",
1850                                   d_filename, evsel_name, h->nr_samples);
1851
1852         printf("%-*.*s----\n",
1853                graph_dotted_len, graph_dotted_len, graph_dotted_line);
1854
1855         if (verbose > 0)
1856                 symbol__annotate_hits(sym, evsel);
1857
1858         list_for_each_entry(pos, &notes->src->source, node) {
1859                 if (context && queue == NULL) {
1860                         queue = pos;
1861                         queue_len = 0;
1862                 }
1863
1864                 switch (disasm_line__print(pos, sym, start, evsel, len,
1865                                             min_pcnt, printed, max_lines,
1866                                             queue)) {
1867                 case 0:
1868                         ++printed;
1869                         if (context) {
1870                                 printed += queue_len;
1871                                 queue = NULL;
1872                                 queue_len = 0;
1873                         }
1874                         break;
1875                 case 1:
1876                         /* filtered by max_lines */
1877                         ++more;
1878                         break;
1879                 case -1:
1880                 default:
1881                         /*
1882                          * Filtered by min_pcnt or non IP lines when
1883                          * context != 0
1884                          */
1885                         if (!context)
1886                                 break;
1887                         if (queue_len == context)
1888                                 queue = list_entry(queue->node.next, typeof(*queue), node);
1889                         else
1890                                 ++queue_len;
1891                         break;
1892                 }
1893         }
1894
1895         free(filename);
1896
1897         return more;
1898 }
1899
1900 void symbol__annotate_zero_histogram(struct symbol *sym, int evidx)
1901 {
1902         struct annotation *notes = symbol__annotation(sym);
1903         struct sym_hist *h = annotation__histogram(notes, evidx);
1904
1905         memset(h, 0, notes->src->sizeof_sym_hist);
1906 }
1907
1908 void symbol__annotate_decay_histogram(struct symbol *sym, int evidx)
1909 {
1910         struct annotation *notes = symbol__annotation(sym);
1911         struct sym_hist *h = annotation__histogram(notes, evidx);
1912         int len = symbol__size(sym), offset;
1913
1914         h->nr_samples = 0;
1915         for (offset = 0; offset < len; ++offset) {
1916                 h->addr[offset].nr_samples = h->addr[offset].nr_samples * 7 / 8;
1917                 h->nr_samples += h->addr[offset].nr_samples;
1918         }
1919 }
1920
1921 void disasm__purge(struct list_head *head)
1922 {
1923         struct disasm_line *pos, *n;
1924
1925         list_for_each_entry_safe(pos, n, head, node) {
1926                 list_del(&pos->node);
1927                 disasm_line__free(pos);
1928         }
1929 }
1930
1931 static size_t disasm_line__fprintf(struct disasm_line *dl, FILE *fp)
1932 {
1933         size_t printed;
1934
1935         if (dl->offset == -1)
1936                 return fprintf(fp, "%s\n", dl->line);
1937
1938         printed = fprintf(fp, "%#" PRIx64 " %s", dl->offset, dl->ins.name);
1939
1940         if (dl->ops.raw[0] != '\0') {
1941                 printed += fprintf(fp, "%.*s %s\n", 6 - (int)printed, " ",
1942                                    dl->ops.raw);
1943         }
1944
1945         return printed + fprintf(fp, "\n");
1946 }
1947
1948 size_t disasm__fprintf(struct list_head *head, FILE *fp)
1949 {
1950         struct disasm_line *pos;
1951         size_t printed = 0;
1952
1953         list_for_each_entry(pos, head, node)
1954                 printed += disasm_line__fprintf(pos, fp);
1955
1956         return printed;
1957 }
1958
1959 int symbol__tty_annotate(struct symbol *sym, struct map *map,
1960                          struct perf_evsel *evsel, bool print_lines,
1961                          bool full_paths, int min_pcnt, int max_lines)
1962 {
1963         struct dso *dso = map->dso;
1964         struct rb_root source_line = RB_ROOT;
1965         u64 len;
1966
1967         if (symbol__disassemble(sym, map, perf_evsel__env_arch(evsel),
1968                                 0, NULL, NULL) < 0)
1969                 return -1;
1970
1971         len = symbol__size(sym);
1972
1973         if (print_lines) {
1974                 srcline_full_filename = full_paths;
1975                 symbol__get_source_line(sym, map, evsel, &source_line, len);
1976                 print_summary(&source_line, dso->long_name);
1977         }
1978
1979         symbol__annotate_printf(sym, map, evsel, full_paths,
1980                                 min_pcnt, max_lines, 0);
1981         if (print_lines)
1982                 symbol__free_source_line(sym, len);
1983
1984         disasm__purge(&symbol__annotation(sym)->src->source);
1985
1986         return 0;
1987 }
1988
1989 bool ui__has_annotation(void)
1990 {
1991         return use_browser == 1 && perf_hpp_list.sym;
1992 }