tools/perf/builtin-annotate.c

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