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