1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * builtin-annotate.c 4 * 5 * Builtin annotate command: Analyze the perf.data input file, 6 * look up and read DSOs and symbol information and display 7 * a histogram of results, along various sorting keys. 8 */ 9 #include "builtin.h" 10 11 #include "util/color.h" 12 #include <linux/list.h> 13 #include "util/cache.h" 14 #include <linux/rbtree.h> 15 #include <linux/zalloc.h> 16 #include "util/symbol.h" 17 18 #include "util/debug.h" 19 20 #include "util/evlist.h" 21 #include "util/evsel.h" 22 #include "util/annotate.h" 23 #include "util/event.h" 24 #include <subcmd/parse-options.h> 25 #include "util/parse-events.h" 26 #include "util/sort.h" 27 #include "util/hist.h" 28 #include "util/dso.h" 29 #include "util/machine.h" 30 #include "util/map.h" 31 #include "util/session.h" 32 #include "util/tool.h" 33 #include "util/data.h" 34 #include "arch/common.h" 35 #include "util/block-range.h" 36 #include "util/map_symbol.h" 37 #include "util/branch.h" 38 #include "util/util.h" 39 40 #include <dlfcn.h> 41 #include <errno.h> 42 #include <linux/bitmap.h> 43 #include <linux/err.h> 44 45 struct perf_annotate { 46 struct perf_tool tool; 47 struct perf_session *session; 48 struct annotation_options opts; 49 #ifdef HAVE_SLANG_SUPPORT 50 bool use_tui; 51 #endif 52 bool use_stdio, use_stdio2; 53 #ifdef HAVE_GTK2_SUPPORT 54 bool use_gtk; 55 #endif 56 bool skip_missing; 57 bool has_br_stack; 58 bool group_set; 59 float min_percent; 60 const char *sym_hist_filter; 61 const char *cpu_list; 62 DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS); 63 }; 64 65 /* 66 * Given one basic block: 67 * 68 * from to branch_i 69 * * ----> * 70 * | 71 * | block 72 * v 73 * * ----> * 74 * from to branch_i+1 75 * 76 * where the horizontal are the branches and the vertical is the executed 77 * block of instructions. 78 * 79 * We count, for each 'instruction', the number of blocks that covered it as 80 * well as count the ratio each branch is taken. 81 * 82 * We can do this without knowing the actual instruction stream by keeping 83 * track of the address ranges. We break down ranges such that there is no 84 * overlap and iterate from the start until the end. 85 * 86 * @acme: once we parse the objdump output _before_ processing the samples, 87 * we can easily fold the branch.cycles IPC bits in. 88 */ 89 static void process_basic_block(struct addr_map_symbol *start, 90 struct addr_map_symbol *end, 91 struct branch_flags *flags) 92 { 93 struct symbol *sym = start->ms.sym; 94 struct annotation *notes = sym ? symbol__annotation(sym) : NULL; 95 struct block_range_iter iter; 96 struct block_range *entry; 97 98 /* 99 * Sanity; NULL isn't executable and the CPU cannot execute backwards 100 */ 101 if (!start->addr || start->addr > end->addr) 102 return; 103 104 iter = block_range__create(start->addr, end->addr); 105 if (!block_range_iter__valid(&iter)) 106 return; 107 108 /* 109 * First block in range is a branch target. 110 */ 111 entry = block_range_iter(&iter); 112 assert(entry->is_target); 113 entry->entry++; 114 115 do { 116 entry = block_range_iter(&iter); 117 118 entry->coverage++; 119 entry->sym = sym; 120 121 if (notes) 122 notes->max_coverage = max(notes->max_coverage, entry->coverage); 123 124 } while (block_range_iter__next(&iter)); 125 126 /* 127 * Last block in rage is a branch. 128 */ 129 entry = block_range_iter(&iter); 130 assert(entry->is_branch); 131 entry->taken++; 132 if (flags->predicted) 133 entry->pred++; 134 } 135 136 static void process_branch_stack(struct branch_stack *bs, struct addr_location *al, 137 struct perf_sample *sample) 138 { 139 struct addr_map_symbol *prev = NULL; 140 struct branch_info *bi; 141 int i; 142 143 if (!bs || !bs->nr) 144 return; 145 146 bi = sample__resolve_bstack(sample, al); 147 if (!bi) 148 return; 149 150 for (i = bs->nr - 1; i >= 0; i--) { 151 /* 152 * XXX filter against symbol 153 */ 154 if (prev) 155 process_basic_block(prev, &bi[i].from, &bi[i].flags); 156 prev = &bi[i].to; 157 } 158 159 free(bi); 160 } 161 162 static int hist_iter__branch_callback(struct hist_entry_iter *iter, 163 struct addr_location *al __maybe_unused, 164 bool single __maybe_unused, 165 void *arg __maybe_unused) 166 { 167 struct hist_entry *he = iter->he; 168 struct branch_info *bi; 169 struct perf_sample *sample = iter->sample; 170 struct evsel *evsel = iter->evsel; 171 int err; 172 173 bi = he->branch_info; 174 err = addr_map_symbol__inc_samples(&bi->from, sample, evsel); 175 176 if (err) 177 goto out; 178 179 err = addr_map_symbol__inc_samples(&bi->to, sample, evsel); 180 181 out: 182 return err; 183 } 184 185 static int process_branch_callback(struct evsel *evsel, 186 struct perf_sample *sample, 187 struct addr_location *al __maybe_unused, 188 struct perf_annotate *ann, 189 struct machine *machine) 190 { 191 struct hist_entry_iter iter = { 192 .evsel = evsel, 193 .sample = sample, 194 .add_entry_cb = hist_iter__branch_callback, 195 .hide_unresolved = symbol_conf.hide_unresolved, 196 .ops = &hist_iter_branch, 197 }; 198 199 struct addr_location a; 200 201 if (machine__resolve(machine, &a, sample) < 0) 202 return -1; 203 204 if (a.sym == NULL) 205 return 0; 206 207 if (a.map != NULL) 208 map__dso(a.map)->hit = 1; 209 210 hist__account_cycles(sample->branch_stack, al, sample, false, NULL); 211 212 return hist_entry_iter__add(&iter, &a, PERF_MAX_STACK_DEPTH, ann); 213 } 214 215 static bool has_annotation(struct perf_annotate *ann) 216 { 217 return ui__has_annotation() || ann->use_stdio2; 218 } 219 220 static int evsel__add_sample(struct evsel *evsel, struct perf_sample *sample, 221 struct addr_location *al, struct perf_annotate *ann, 222 struct machine *machine) 223 { 224 struct hists *hists = evsel__hists(evsel); 225 struct hist_entry *he; 226 int ret; 227 228 if ((!ann->has_br_stack || !has_annotation(ann)) && 229 ann->sym_hist_filter != NULL && 230 (al->sym == NULL || 231 strcmp(ann->sym_hist_filter, al->sym->name) != 0)) { 232 /* We're only interested in a symbol named sym_hist_filter */ 233 /* 234 * FIXME: why isn't this done in the symbol_filter when loading 235 * the DSO? 236 */ 237 if (al->sym != NULL) { 238 struct dso *dso = map__dso(al->map); 239 240 rb_erase_cached(&al->sym->rb_node, &dso->symbols); 241 symbol__delete(al->sym); 242 dso__reset_find_symbol_cache(dso); 243 } 244 return 0; 245 } 246 247 /* 248 * XXX filtered samples can still have branch entries pointing into our 249 * symbol and are missed. 250 */ 251 process_branch_stack(sample->branch_stack, al, sample); 252 253 if (ann->has_br_stack && has_annotation(ann)) 254 return process_branch_callback(evsel, sample, al, ann, machine); 255 256 he = hists__add_entry(hists, al, NULL, NULL, NULL, NULL, sample, true); 257 if (he == NULL) 258 return -ENOMEM; 259 260 ret = hist_entry__inc_addr_samples(he, sample, evsel, al->addr); 261 hists__inc_nr_samples(hists, true); 262 return ret; 263 } 264 265 static int process_sample_event(struct perf_tool *tool, 266 union perf_event *event, 267 struct perf_sample *sample, 268 struct evsel *evsel, 269 struct machine *machine) 270 { 271 struct perf_annotate *ann = container_of(tool, struct perf_annotate, tool); 272 struct addr_location al; 273 int ret = 0; 274 275 if (machine__resolve(machine, &al, sample) < 0) { 276 pr_warning("problem processing %d event, skipping it.\n", 277 event->header.type); 278 return -1; 279 } 280 281 if (ann->cpu_list && !test_bit(sample->cpu, ann->cpu_bitmap)) 282 goto out_put; 283 284 if (!al.filtered && 285 evsel__add_sample(evsel, sample, &al, ann, machine)) { 286 pr_warning("problem incrementing symbol count, " 287 "skipping event\n"); 288 ret = -1; 289 } 290 out_put: 291 addr_location__put(&al); 292 return ret; 293 } 294 295 static int process_feature_event(struct perf_session *session, 296 union perf_event *event) 297 { 298 if (event->feat.feat_id < HEADER_LAST_FEATURE) 299 return perf_event__process_feature(session, event); 300 return 0; 301 } 302 303 static int hist_entry__tty_annotate(struct hist_entry *he, 304 struct evsel *evsel, 305 struct perf_annotate *ann) 306 { 307 if (!ann->use_stdio2) 308 return symbol__tty_annotate(&he->ms, evsel, &ann->opts); 309 310 return symbol__tty_annotate2(&he->ms, evsel, &ann->opts); 311 } 312 313 static void hists__find_annotations(struct hists *hists, 314 struct evsel *evsel, 315 struct perf_annotate *ann) 316 { 317 struct rb_node *nd = rb_first_cached(&hists->entries), *next; 318 int key = K_RIGHT; 319 320 while (nd) { 321 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node); 322 struct annotation *notes; 323 324 if (he->ms.sym == NULL || map__dso(he->ms.map)->annotate_warned) 325 goto find_next; 326 327 if (ann->sym_hist_filter && 328 (strcmp(he->ms.sym->name, ann->sym_hist_filter) != 0)) 329 goto find_next; 330 331 if (ann->min_percent) { 332 float percent = 0; 333 u64 total = hists__total_period(hists); 334 335 if (total) 336 percent = 100.0 * he->stat.period / total; 337 338 if (percent < ann->min_percent) 339 goto find_next; 340 } 341 342 notes = symbol__annotation(he->ms.sym); 343 if (notes->src == NULL) { 344 find_next: 345 if (key == K_LEFT || key == '<') 346 nd = rb_prev(nd); 347 else 348 nd = rb_next(nd); 349 continue; 350 } 351 352 if (use_browser == 2) { 353 int ret; 354 int (*annotate)(struct hist_entry *he, 355 struct evsel *evsel, 356 struct annotation_options *options, 357 struct hist_browser_timer *hbt); 358 359 annotate = dlsym(perf_gtk_handle, 360 "hist_entry__gtk_annotate"); 361 if (annotate == NULL) { 362 ui__error("GTK browser not found!\n"); 363 return; 364 } 365 366 ret = annotate(he, evsel, &ann->opts, NULL); 367 if (!ret || !ann->skip_missing) 368 return; 369 370 /* skip missing symbols */ 371 nd = rb_next(nd); 372 } else if (use_browser == 1) { 373 key = hist_entry__tui_annotate(he, evsel, NULL, &ann->opts); 374 375 switch (key) { 376 case -1: 377 if (!ann->skip_missing) 378 return; 379 /* fall through */ 380 case K_RIGHT: 381 case '>': 382 next = rb_next(nd); 383 break; 384 case K_LEFT: 385 case '<': 386 next = rb_prev(nd); 387 break; 388 default: 389 return; 390 } 391 392 if (next != NULL) 393 nd = next; 394 } else { 395 hist_entry__tty_annotate(he, evsel, ann); 396 nd = rb_next(nd); 397 } 398 } 399 } 400 401 static int __cmd_annotate(struct perf_annotate *ann) 402 { 403 int ret; 404 struct perf_session *session = ann->session; 405 struct evsel *pos; 406 u64 total_nr_samples; 407 408 if (ann->cpu_list) { 409 ret = perf_session__cpu_bitmap(session, ann->cpu_list, 410 ann->cpu_bitmap); 411 if (ret) 412 goto out; 413 } 414 415 if (!ann->opts.objdump_path) { 416 ret = perf_env__lookup_objdump(&session->header.env, 417 &ann->opts.objdump_path); 418 if (ret) 419 goto out; 420 } 421 422 ret = perf_session__process_events(session); 423 if (ret) 424 goto out; 425 426 if (dump_trace) { 427 perf_session__fprintf_nr_events(session, stdout, false); 428 evlist__fprintf_nr_events(session->evlist, stdout, false); 429 goto out; 430 } 431 432 if (verbose > 3) 433 perf_session__fprintf(session, stdout); 434 435 if (verbose > 2) 436 perf_session__fprintf_dsos(session, stdout); 437 438 total_nr_samples = 0; 439 evlist__for_each_entry(session->evlist, pos) { 440 struct hists *hists = evsel__hists(pos); 441 u32 nr_samples = hists->stats.nr_samples; 442 443 if (nr_samples > 0) { 444 total_nr_samples += nr_samples; 445 hists__collapse_resort(hists, NULL); 446 /* Don't sort callchain */ 447 evsel__reset_sample_bit(pos, CALLCHAIN); 448 evsel__output_resort(pos, NULL); 449 450 if (symbol_conf.event_group && !evsel__is_group_leader(pos)) 451 continue; 452 453 hists__find_annotations(hists, pos, ann); 454 } 455 } 456 457 if (total_nr_samples == 0) { 458 ui__error("The %s data has no samples!\n", session->data->path); 459 goto out; 460 } 461 462 if (use_browser == 2) { 463 void (*show_annotations)(void); 464 465 show_annotations = dlsym(perf_gtk_handle, 466 "perf_gtk__show_annotations"); 467 if (show_annotations == NULL) { 468 ui__error("GTK browser not found!\n"); 469 goto out; 470 } 471 show_annotations(); 472 } 473 474 out: 475 return ret; 476 } 477 478 static int parse_percent_limit(const struct option *opt, const char *str, 479 int unset __maybe_unused) 480 { 481 struct perf_annotate *ann = opt->value; 482 double pcnt = strtof(str, NULL); 483 484 ann->min_percent = pcnt; 485 return 0; 486 } 487 488 static const char * const annotate_usage[] = { 489 "perf annotate [<options>]", 490 NULL 491 }; 492 493 int cmd_annotate(int argc, const char **argv) 494 { 495 struct perf_annotate annotate = { 496 .tool = { 497 .sample = process_sample_event, 498 .mmap = perf_event__process_mmap, 499 .mmap2 = perf_event__process_mmap2, 500 .comm = perf_event__process_comm, 501 .exit = perf_event__process_exit, 502 .fork = perf_event__process_fork, 503 .namespaces = perf_event__process_namespaces, 504 .attr = perf_event__process_attr, 505 .build_id = perf_event__process_build_id, 506 #ifdef HAVE_LIBTRACEEVENT 507 .tracing_data = perf_event__process_tracing_data, 508 #endif 509 .id_index = perf_event__process_id_index, 510 .auxtrace_info = perf_event__process_auxtrace_info, 511 .auxtrace = perf_event__process_auxtrace, 512 .feature = process_feature_event, 513 .ordered_events = true, 514 .ordering_requires_timestamps = true, 515 }, 516 }; 517 struct perf_data data = { 518 .mode = PERF_DATA_MODE_READ, 519 }; 520 struct itrace_synth_opts itrace_synth_opts = { 521 .set = 0, 522 }; 523 const char *disassembler_style = NULL, *objdump_path = NULL, *addr2line_path = NULL; 524 struct option options[] = { 525 OPT_STRING('i', "input", &input_name, "file", 526 "input file name"), 527 OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]", 528 "only consider symbols in these dsos"), 529 OPT_STRING('s', "symbol", &annotate.sym_hist_filter, "symbol", 530 "symbol to annotate"), 531 OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"), 532 OPT_INCR('v', "verbose", &verbose, 533 "be more verbose (show symbol address, etc)"), 534 OPT_BOOLEAN('q', "quiet", &quiet, "do now show any warnings or messages"), 535 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, 536 "dump raw trace in ASCII"), 537 #ifdef HAVE_GTK2_SUPPORT 538 OPT_BOOLEAN(0, "gtk", &annotate.use_gtk, "Use the GTK interface"), 539 #endif 540 #ifdef HAVE_SLANG_SUPPORT 541 OPT_BOOLEAN(0, "tui", &annotate.use_tui, "Use the TUI interface"), 542 #endif 543 OPT_BOOLEAN(0, "stdio", &annotate.use_stdio, "Use the stdio interface"), 544 OPT_BOOLEAN(0, "stdio2", &annotate.use_stdio2, "Use the stdio interface"), 545 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux, 546 "don't load vmlinux even if found"), 547 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name, 548 "file", "vmlinux pathname"), 549 OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules, 550 "load module symbols - WARNING: use only with -k and LIVE kernel"), 551 OPT_BOOLEAN('l', "print-line", &annotate.opts.print_lines, 552 "print matching source lines (may be slow)"), 553 OPT_BOOLEAN('P', "full-paths", &annotate.opts.full_path, 554 "Don't shorten the displayed pathnames"), 555 OPT_BOOLEAN(0, "skip-missing", &annotate.skip_missing, 556 "Skip symbols that cannot be annotated"), 557 OPT_BOOLEAN_SET(0, "group", &symbol_conf.event_group, 558 &annotate.group_set, 559 "Show event group information together"), 560 OPT_STRING('C', "cpu", &annotate.cpu_list, "cpu", "list of cpus to profile"), 561 OPT_CALLBACK(0, "symfs", NULL, "directory", 562 "Look for files with symbols relative to this directory", 563 symbol__config_symfs), 564 OPT_BOOLEAN(0, "source", &annotate.opts.annotate_src, 565 "Interleave source code with assembly code (default)"), 566 OPT_BOOLEAN(0, "asm-raw", &annotate.opts.show_asm_raw, 567 "Display raw encoding of assembly instructions (default)"), 568 OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style", 569 "Specify disassembler style (e.g. -M intel for intel syntax)"), 570 OPT_STRING(0, "prefix", &annotate.opts.prefix, "prefix", 571 "Add prefix to source file path names in programs (with --prefix-strip)"), 572 OPT_STRING(0, "prefix-strip", &annotate.opts.prefix_strip, "N", 573 "Strip first N entries of source file path name in programs (with --prefix)"), 574 OPT_STRING(0, "objdump", &objdump_path, "path", 575 "objdump binary to use for disassembly and annotations"), 576 OPT_STRING(0, "addr2line", &addr2line_path, "path", 577 "addr2line binary to use for line numbers"), 578 OPT_BOOLEAN(0, "demangle", &symbol_conf.demangle, 579 "Enable symbol demangling"), 580 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel, 581 "Enable kernel symbol demangling"), 582 OPT_BOOLEAN(0, "group", &symbol_conf.event_group, 583 "Show event group information together"), 584 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period, 585 "Show a column with the sum of periods"), 586 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples, 587 "Show a column with the number of samples"), 588 OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode", 589 "'always' (default), 'never' or 'auto' only applicable to --stdio mode", 590 stdio__config_color, "always"), 591 OPT_CALLBACK(0, "percent-type", &annotate.opts, "local-period", 592 "Set percent type local/global-period/hits", 593 annotate_parse_percent_type), 594 OPT_CALLBACK(0, "percent-limit", &annotate, "percent", 595 "Don't show entries under that percent", parse_percent_limit), 596 OPT_CALLBACK_OPTARG(0, "itrace", &itrace_synth_opts, NULL, "opts", 597 "Instruction Tracing options\n" ITRACE_HELP, 598 itrace_parse_synth_opts), 599 600 OPT_END() 601 }; 602 int ret; 603 604 set_option_flag(options, 0, "show-total-period", PARSE_OPT_EXCLUSIVE); 605 set_option_flag(options, 0, "show-nr-samples", PARSE_OPT_EXCLUSIVE); 606 607 annotation_options__init(&annotate.opts); 608 609 ret = hists__init(); 610 if (ret < 0) 611 return ret; 612 613 annotation_config__init(&annotate.opts); 614 615 argc = parse_options(argc, argv, options, annotate_usage, 0); 616 if (argc) { 617 /* 618 * Special case: if there's an argument left then assume that 619 * it's a symbol filter: 620 */ 621 if (argc > 1) 622 usage_with_options(annotate_usage, options); 623 624 annotate.sym_hist_filter = argv[0]; 625 } 626 627 if (disassembler_style) { 628 annotate.opts.disassembler_style = strdup(disassembler_style); 629 if (!annotate.opts.disassembler_style) 630 return -ENOMEM; 631 } 632 if (objdump_path) { 633 annotate.opts.objdump_path = strdup(objdump_path); 634 if (!annotate.opts.objdump_path) 635 return -ENOMEM; 636 } 637 if (addr2line_path) { 638 symbol_conf.addr2line_path = strdup(addr2line_path); 639 if (!symbol_conf.addr2line_path) 640 return -ENOMEM; 641 } 642 643 if (annotate_check_args(&annotate.opts) < 0) 644 return -EINVAL; 645 646 #ifdef HAVE_GTK2_SUPPORT 647 if (symbol_conf.show_nr_samples && annotate.use_gtk) { 648 pr_err("--show-nr-samples is not available in --gtk mode at this time\n"); 649 return ret; 650 } 651 #endif 652 653 ret = symbol__validate_sym_arguments(); 654 if (ret) 655 return ret; 656 657 if (quiet) 658 perf_quiet_option(); 659 660 data.path = input_name; 661 662 annotate.session = perf_session__new(&data, &annotate.tool); 663 if (IS_ERR(annotate.session)) 664 return PTR_ERR(annotate.session); 665 666 annotate.session->itrace_synth_opts = &itrace_synth_opts; 667 668 annotate.has_br_stack = perf_header__has_feat(&annotate.session->header, 669 HEADER_BRANCH_STACK); 670 671 if (annotate.group_set) 672 evlist__force_leader(annotate.session->evlist); 673 674 ret = symbol__annotation_init(); 675 if (ret < 0) 676 goto out_delete; 677 678 symbol_conf.try_vmlinux_path = true; 679 680 ret = symbol__init(&annotate.session->header.env); 681 if (ret < 0) 682 goto out_delete; 683 684 if (annotate.use_stdio || annotate.use_stdio2) 685 use_browser = 0; 686 #ifdef HAVE_SLANG_SUPPORT 687 else if (annotate.use_tui) 688 use_browser = 1; 689 #endif 690 #ifdef HAVE_GTK2_SUPPORT 691 else if (annotate.use_gtk) 692 use_browser = 2; 693 #endif 694 695 setup_browser(true); 696 697 /* 698 * Events of different processes may correspond to the same 699 * symbol, we do not care about the processes in annotate, 700 * set sort order to avoid repeated output. 701 */ 702 sort_order = "dso,symbol"; 703 704 /* 705 * Set SORT_MODE__BRANCH so that annotate display IPC/Cycle 706 * if branch info is in perf data in TUI mode. 707 */ 708 if ((use_browser == 1 || annotate.use_stdio2) && annotate.has_br_stack) 709 sort__mode = SORT_MODE__BRANCH; 710 711 if (setup_sorting(NULL) < 0) 712 usage_with_options(annotate_usage, options); 713 714 ret = __cmd_annotate(&annotate); 715 716 out_delete: 717 /* 718 * Speed up the exit process by only deleting for debug builds. For 719 * large files this can save time. 720 */ 721 #ifndef NDEBUG 722 perf_session__delete(annotate.session); 723 #endif 724 annotation_options__exit(&annotate.opts); 725 726 return ret; 727 } 728