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