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/util.h"
  12#include "util/color.h"
  13#include <linux/list.h>
  14#include "util/cache.h"
  15#include <linux/rbtree.h>
  16#include "util/symbol.h"
  17
  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/thread.h"
  28#include "util/sort.h"
  29#include "util/hist.h"
  30#include "util/session.h"
  31#include "util/tool.h"
  32#include "util/data.h"
  33#include "arch/common.h"
  34#include "util/block-range.h"
  35
  36#include <dlfcn.h>
  37#include <errno.h>
  38#include <linux/bitmap.h>
  39
  40struct perf_annotate {
  41        struct perf_tool tool;
  42        struct perf_session *session;
  43        bool       use_tui, use_stdio, use_stdio2, use_gtk;
  44        bool       full_paths;
  45        bool       print_line;
  46        bool       skip_missing;
  47        bool       has_br_stack;
  48        const char *sym_hist_filter;
  49        const char *cpu_list;
  50        DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  51};
  52
  53/*
  54 * Given one basic block:
  55 *
  56 *      from    to              branch_i
  57 *      * ----> *
  58 *              |
  59 *              | block
  60 *              v
  61 *              * ----> *
  62 *              from    to      branch_i+1
  63 *
  64 * where the horizontal are the branches and the vertical is the executed
  65 * block of instructions.
  66 *
  67 * We count, for each 'instruction', the number of blocks that covered it as
  68 * well as count the ratio each branch is taken.
  69 *
  70 * We can do this without knowing the actual instruction stream by keeping
  71 * track of the address ranges. We break down ranges such that there is no
  72 * overlap and iterate from the start until the end.
  73 *
  74 * @acme: once we parse the objdump output _before_ processing the samples,
  75 * we can easily fold the branch.cycles IPC bits in.
  76 */
  77static void process_basic_block(struct addr_map_symbol *start,
  78                                struct addr_map_symbol *end,
  79                                struct branch_flags *flags)
  80{
  81        struct symbol *sym = start->sym;
  82        struct annotation *notes = sym ? symbol__annotation(sym) : NULL;
  83        struct block_range_iter iter;
  84        struct block_range *entry;
  85
  86        /*
  87         * Sanity; NULL isn't executable and the CPU cannot execute backwards
  88         */
  89        if (!start->addr || start->addr > end->addr)
  90                return;
  91
  92        iter = block_range__create(start->addr, end->addr);
  93        if (!block_range_iter__valid(&iter))
  94                return;
  95
  96        /*
  97         * First block in range is a branch target.
  98         */
  99        entry = block_range_iter(&iter);
 100        assert(entry->is_target);
 101        entry->entry++;
 102
 103        do {
 104                entry = block_range_iter(&iter);
 105
 106                entry->coverage++;
 107                entry->sym = sym;
 108
 109                if (notes)
 110                        notes->max_coverage = max(notes->max_coverage, entry->coverage);
 111
 112        } while (block_range_iter__next(&iter));
 113
 114        /*
 115         * Last block in rage is a branch.
 116         */
 117        entry = block_range_iter(&iter);
 118        assert(entry->is_branch);
 119        entry->taken++;
 120        if (flags->predicted)
 121                entry->pred++;
 122}
 123
 124static void process_branch_stack(struct branch_stack *bs, struct addr_location *al,
 125                                 struct perf_sample *sample)
 126{
 127        struct addr_map_symbol *prev = NULL;
 128        struct branch_info *bi;
 129        int i;
 130
 131        if (!bs || !bs->nr)
 132                return;
 133
 134        bi = sample__resolve_bstack(sample, al);
 135        if (!bi)
 136                return;
 137
 138        for (i = bs->nr - 1; i >= 0; i--) {
 139                /*
 140                 * XXX filter against symbol
 141                 */
 142                if (prev)
 143                        process_basic_block(prev, &bi[i].from, &bi[i].flags);
 144                prev = &bi[i].to;
 145        }
 146
 147        free(bi);
 148}
 149
 150static int hist_iter__branch_callback(struct hist_entry_iter *iter,
 151                                      struct addr_location *al __maybe_unused,
 152                                      bool single __maybe_unused,
 153                                      void *arg __maybe_unused)
 154{
 155        struct hist_entry *he = iter->he;
 156        struct branch_info *bi;
 157        struct perf_sample *sample = iter->sample;
 158        struct perf_evsel *evsel = iter->evsel;
 159        int err;
 160
 161        hist__account_cycles(sample->branch_stack, al, sample, false);
 162
 163        bi = he->branch_info;
 164        err = addr_map_symbol__inc_samples(&bi->from, sample, evsel->idx);
 165
 166        if (err)
 167                goto out;
 168
 169        err = addr_map_symbol__inc_samples(&bi->to, sample, evsel->idx);
 170
 171out:
 172        return err;
 173}
 174
 175static int process_branch_callback(struct perf_evsel *evsel,
 176                                   struct perf_sample *sample,
 177                                   struct addr_location *al __maybe_unused,
 178                                   struct perf_annotate *ann,
 179                                   struct machine *machine)
 180{
 181        struct hist_entry_iter iter = {
 182                .evsel          = evsel,
 183                .sample         = sample,
 184                .add_entry_cb   = hist_iter__branch_callback,
 185                .hide_unresolved        = symbol_conf.hide_unresolved,
 186                .ops            = &hist_iter_branch,
 187        };
 188
 189        struct addr_location a;
 190        int ret;
 191
 192        if (machine__resolve(machine, &a, sample) < 0)
 193                return -1;
 194
 195        if (a.sym == NULL)
 196                return 0;
 197
 198        if (a.map != NULL)
 199                a.map->dso->hit = 1;
 200
 201        ret = hist_entry_iter__add(&iter, &a, PERF_MAX_STACK_DEPTH, ann);
 202        return ret;
 203}
 204
 205static bool has_annotation(struct perf_annotate *ann)
 206{
 207        return ui__has_annotation() || ann->use_stdio2;
 208}
 209
 210static int perf_evsel__add_sample(struct perf_evsel *evsel,
 211                                  struct perf_sample *sample,
 212                                  struct addr_location *al,
 213                                  struct perf_annotate *ann,
 214                                  struct machine *machine)
 215{
 216        struct hists *hists = evsel__hists(evsel);
 217        struct hist_entry *he;
 218        int ret;
 219
 220        if ((!ann->has_br_stack || !has_annotation(ann)) &&
 221            ann->sym_hist_filter != NULL &&
 222            (al->sym == NULL ||
 223             strcmp(ann->sym_hist_filter, al->sym->name) != 0)) {
 224                /* We're only interested in a symbol named sym_hist_filter */
 225                /*
 226                 * FIXME: why isn't this done in the symbol_filter when loading
 227                 * the DSO?
 228                 */
 229                if (al->sym != NULL) {
 230                        rb_erase(&al->sym->rb_node,
 231                                 &al->map->dso->symbols[al->map->type]);
 232                        symbol__delete(al->sym);
 233                        dso__reset_find_symbol_cache(al->map->dso);
 234                }
 235                return 0;
 236        }
 237
 238        /*
 239         * XXX filtered samples can still have branch entires pointing into our
 240         * symbol and are missed.
 241         */
 242        process_branch_stack(sample->branch_stack, al, sample);
 243
 244        if (ann->has_br_stack && has_annotation(ann))
 245                return process_branch_callback(evsel, sample, al, ann, machine);
 246
 247        he = hists__add_entry(hists, al, NULL, NULL, NULL, sample, true);
 248        if (he == NULL)
 249                return -ENOMEM;
 250
 251        ret = hist_entry__inc_addr_samples(he, sample, evsel->idx, al->addr);
 252        hists__inc_nr_samples(hists, true);
 253        return ret;
 254}
 255
 256static int process_sample_event(struct perf_tool *tool,
 257                                union perf_event *event,
 258                                struct perf_sample *sample,
 259                                struct perf_evsel *evsel,
 260                                struct machine *machine)
 261{
 262        struct perf_annotate *ann = container_of(tool, struct perf_annotate, tool);
 263        struct addr_location al;
 264        int ret = 0;
 265
 266        if (machine__resolve(machine, &al, sample) < 0) {
 267                pr_warning("problem processing %d event, skipping it.\n",
 268                           event->header.type);
 269                return -1;
 270        }
 271
 272        if (ann->cpu_list && !test_bit(sample->cpu, ann->cpu_bitmap))
 273                goto out_put;
 274
 275        if (!al.filtered &&
 276            perf_evsel__add_sample(evsel, sample, &al, ann, machine)) {
 277                pr_warning("problem incrementing symbol count, "
 278                           "skipping event\n");
 279                ret = -1;
 280        }
 281out_put:
 282        addr_location__put(&al);
 283        return ret;
 284}
 285
 286static int hist_entry__tty_annotate(struct hist_entry *he,
 287                                    struct perf_evsel *evsel,
 288                                    struct perf_annotate *ann)
 289{
 290        if (!ann->use_stdio2)
 291                return symbol__tty_annotate(he->ms.sym, he->ms.map, evsel,
 292                                            ann->print_line, ann->full_paths, 0, 0);
 293        return symbol__tty_annotate2(he->ms.sym, he->ms.map, evsel,
 294                                     ann->print_line, ann->full_paths);
 295}
 296
 297static void hists__find_annotations(struct hists *hists,
 298                                    struct perf_evsel *evsel,
 299                                    struct perf_annotate *ann)
 300{
 301        struct rb_node *nd = rb_first(&hists->entries), *next;
 302        int key = K_RIGHT;
 303
 304        while (nd) {
 305                struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
 306                struct annotation *notes;
 307
 308                if (he->ms.sym == NULL || he->ms.map->dso->annotate_warned)
 309                        goto find_next;
 310
 311                if (ann->sym_hist_filter &&
 312                    (strcmp(he->ms.sym->name, ann->sym_hist_filter) != 0))
 313                        goto find_next;
 314
 315                notes = symbol__annotation(he->ms.sym);
 316                if (notes->src == NULL) {
 317find_next:
 318                        if (key == K_LEFT)
 319                                nd = rb_prev(nd);
 320                        else
 321                                nd = rb_next(nd);
 322                        continue;
 323                }
 324
 325                if (use_browser == 2) {
 326                        int ret;
 327                        int (*annotate)(struct hist_entry *he,
 328                                        struct perf_evsel *evsel,
 329                                        struct hist_browser_timer *hbt);
 330
 331                        annotate = dlsym(perf_gtk_handle,
 332                                         "hist_entry__gtk_annotate");
 333                        if (annotate == NULL) {
 334                                ui__error("GTK browser not found!\n");
 335                                return;
 336                        }
 337
 338                        ret = annotate(he, evsel, NULL);
 339                        if (!ret || !ann->skip_missing)
 340                                return;
 341
 342                        /* skip missing symbols */
 343                        nd = rb_next(nd);
 344                } else if (use_browser == 1) {
 345                        key = hist_entry__tui_annotate(he, evsel, NULL);
 346
 347                        switch (key) {
 348                        case -1:
 349                                if (!ann->skip_missing)
 350                                        return;
 351                                /* fall through */
 352                        case K_RIGHT:
 353                                next = rb_next(nd);
 354                                break;
 355                        case K_LEFT:
 356                                next = rb_prev(nd);
 357                                break;
 358                        default:
 359                                return;
 360                        }
 361
 362                        if (next != NULL)
 363                                nd = next;
 364                } else {
 365                        hist_entry__tty_annotate(he, evsel, ann);
 366                        nd = rb_next(nd);
 367                        /*
 368                         * Since we have a hist_entry per IP for the same
 369                         * symbol, free he->ms.sym->src to signal we already
 370                         * processed this symbol.
 371                         */
 372                        zfree(&notes->src->cycles_hist);
 373                        zfree(&notes->src);
 374                }
 375        }
 376}
 377
 378static int __cmd_annotate(struct perf_annotate *ann)
 379{
 380        int ret;
 381        struct perf_session *session = ann->session;
 382        struct perf_evsel *pos;
 383        u64 total_nr_samples;
 384
 385        if (ann->cpu_list) {
 386                ret = perf_session__cpu_bitmap(session, ann->cpu_list,
 387                                               ann->cpu_bitmap);
 388                if (ret)
 389                        goto out;
 390        }
 391
 392        if (!objdump_path) {
 393                ret = perf_env__lookup_objdump(&session->header.env);
 394                if (ret)
 395                        goto out;
 396        }
 397
 398        ret = perf_session__process_events(session);
 399        if (ret)
 400                goto out;
 401
 402        if (dump_trace) {
 403                perf_session__fprintf_nr_events(session, stdout);
 404                perf_evlist__fprintf_nr_events(session->evlist, stdout);
 405                goto out;
 406        }
 407
 408        if (verbose > 3)
 409                perf_session__fprintf(session, stdout);
 410
 411        if (verbose > 2)
 412                perf_session__fprintf_dsos(session, stdout);
 413
 414        total_nr_samples = 0;
 415        evlist__for_each_entry(session->evlist, pos) {
 416                struct hists *hists = evsel__hists(pos);
 417                u32 nr_samples = hists->stats.nr_events[PERF_RECORD_SAMPLE];
 418
 419                if (nr_samples > 0) {
 420                        total_nr_samples += nr_samples;
 421                        hists__collapse_resort(hists, NULL);
 422                        /* Don't sort callchain */
 423                        perf_evsel__reset_sample_bit(pos, CALLCHAIN);
 424                        perf_evsel__output_resort(pos, NULL);
 425
 426                        if (symbol_conf.event_group &&
 427                            !perf_evsel__is_group_leader(pos))
 428                                continue;
 429
 430                        hists__find_annotations(hists, pos, ann);
 431                }
 432        }
 433
 434        if (total_nr_samples == 0) {
 435                ui__error("The %s file has no samples!\n", session->data->file.path);
 436                goto out;
 437        }
 438
 439        if (use_browser == 2) {
 440                void (*show_annotations)(void);
 441
 442                show_annotations = dlsym(perf_gtk_handle,
 443                                         "perf_gtk__show_annotations");
 444                if (show_annotations == NULL) {
 445                        ui__error("GTK browser not found!\n");
 446                        goto out;
 447                }
 448                show_annotations();
 449        }
 450
 451out:
 452        return ret;
 453}
 454
 455static const char * const annotate_usage[] = {
 456        "perf annotate [<options>]",
 457        NULL
 458};
 459
 460int cmd_annotate(int argc, const char **argv)
 461{
 462        struct perf_annotate annotate = {
 463                .tool = {
 464                        .sample = process_sample_event,
 465                        .mmap   = perf_event__process_mmap,
 466                        .mmap2  = perf_event__process_mmap2,
 467                        .comm   = perf_event__process_comm,
 468                        .exit   = perf_event__process_exit,
 469                        .fork   = perf_event__process_fork,
 470                        .namespaces = perf_event__process_namespaces,
 471                        .attr   = perf_event__process_attr,
 472                        .build_id = perf_event__process_build_id,
 473                        .tracing_data   = perf_event__process_tracing_data,
 474                        .feature        = perf_event__process_feature,
 475                        .ordered_events = true,
 476                        .ordering_requires_timestamps = true,
 477                },
 478        };
 479        struct perf_data data = {
 480                .mode  = PERF_DATA_MODE_READ,
 481        };
 482        struct option options[] = {
 483        OPT_STRING('i', "input", &input_name, "file",
 484                    "input file name"),
 485        OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
 486                   "only consider symbols in these dsos"),
 487        OPT_STRING('s', "symbol", &annotate.sym_hist_filter, "symbol",
 488                    "symbol to annotate"),
 489        OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"),
 490        OPT_INCR('v', "verbose", &verbose,
 491                    "be more verbose (show symbol address, etc)"),
 492        OPT_BOOLEAN('q', "quiet", &quiet, "do now show any message"),
 493        OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
 494                    "dump raw trace in ASCII"),
 495        OPT_BOOLEAN(0, "gtk", &annotate.use_gtk, "Use the GTK interface"),
 496        OPT_BOOLEAN(0, "tui", &annotate.use_tui, "Use the TUI interface"),
 497        OPT_BOOLEAN(0, "stdio", &annotate.use_stdio, "Use the stdio interface"),
 498        OPT_BOOLEAN(0, "stdio2", &annotate.use_stdio2, "Use the stdio interface"),
 499        OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
 500                    "don't load vmlinux even if found"),
 501        OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 502                   "file", "vmlinux pathname"),
 503        OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
 504                    "load module symbols - WARNING: use only with -k and LIVE kernel"),
 505        OPT_BOOLEAN('l', "print-line", &annotate.print_line,
 506                    "print matching source lines (may be slow)"),
 507        OPT_BOOLEAN('P', "full-paths", &annotate.full_paths,
 508                    "Don't shorten the displayed pathnames"),
 509        OPT_BOOLEAN(0, "skip-missing", &annotate.skip_missing,
 510                    "Skip symbols that cannot be annotated"),
 511        OPT_STRING('C', "cpu", &annotate.cpu_list, "cpu", "list of cpus to profile"),
 512        OPT_CALLBACK(0, "symfs", NULL, "directory",
 513                     "Look for files with symbols relative to this directory",
 514                     symbol__config_symfs),
 515        OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
 516                    "Interleave source code with assembly code (default)"),
 517        OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
 518                    "Display raw encoding of assembly instructions (default)"),
 519        OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
 520                   "Specify disassembler style (e.g. -M intel for intel syntax)"),
 521        OPT_STRING(0, "objdump", &objdump_path, "path",
 522                   "objdump binary to use for disassembly and annotations"),
 523        OPT_BOOLEAN(0, "group", &symbol_conf.event_group,
 524                    "Show event group information together"),
 525        OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
 526                    "Show a column with the sum of periods"),
 527        OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
 528                    "Show a column with the number of samples"),
 529        OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode",
 530                             "'always' (default), 'never' or 'auto' only applicable to --stdio mode",
 531                             stdio__config_color, "always"),
 532        OPT_END()
 533        };
 534        int ret;
 535
 536        set_option_flag(options, 0, "show-total-period", PARSE_OPT_EXCLUSIVE);
 537        set_option_flag(options, 0, "show-nr-samples", PARSE_OPT_EXCLUSIVE);
 538
 539
 540        ret = hists__init();
 541        if (ret < 0)
 542                return ret;
 543
 544        argc = parse_options(argc, argv, options, annotate_usage, 0);
 545        if (argc) {
 546                /*
 547                 * Special case: if there's an argument left then assume that
 548                 * it's a symbol filter:
 549                 */
 550                if (argc > 1)
 551                        usage_with_options(annotate_usage, options);
 552
 553                annotate.sym_hist_filter = argv[0];
 554        }
 555
 556        if (symbol_conf.show_nr_samples && annotate.use_gtk) {
 557                pr_err("--show-nr-samples is not available in --gtk mode at this time\n");
 558                return ret;
 559        }
 560
 561        if (quiet)
 562                perf_quiet_option();
 563
 564        data.file.path = input_name;
 565
 566        annotate.session = perf_session__new(&data, false, &annotate.tool);
 567        if (annotate.session == NULL)
 568                return -1;
 569
 570        annotate.has_br_stack = perf_header__has_feat(&annotate.session->header,
 571                                                      HEADER_BRANCH_STACK);
 572
 573        ret = symbol__annotation_init();
 574        if (ret < 0)
 575                goto out_delete;
 576
 577        annotation_config__init();
 578
 579        symbol_conf.try_vmlinux_path = true;
 580
 581        ret = symbol__init(&annotate.session->header.env);
 582        if (ret < 0)
 583                goto out_delete;
 584
 585        if (annotate.use_stdio || annotate.use_stdio2)
 586                use_browser = 0;
 587        else if (annotate.use_tui)
 588                use_browser = 1;
 589        else if (annotate.use_gtk)
 590                use_browser = 2;
 591
 592        setup_browser(true);
 593
 594        if ((use_browser == 1 || annotate.use_stdio2) && annotate.has_br_stack) {
 595                sort__mode = SORT_MODE__BRANCH;
 596                if (setup_sorting(annotate.session->evlist) < 0)
 597                        usage_with_options(annotate_usage, options);
 598        } else {
 599                if (setup_sorting(NULL) < 0)
 600                        usage_with_options(annotate_usage, options);
 601        }
 602
 603        ret = __cmd_annotate(&annotate);
 604
 605out_delete:
 606        /*
 607         * Speed up the exit process, for large files this can
 608         * take quite a while.
 609         *
 610         * XXX Enable this when using valgrind or if we ever
 611         * librarize this command.
 612         *
 613         * Also experiment with obstacks to see how much speed
 614         * up we'll get here.
 615         *
 616         * perf_session__delete(session);
 617         */
 618        return ret;
 619}
 620