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/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
  40 struct perf_annotate {
  41         struct perf_tool tool;
  42         struct perf_session *session;
  43         struct annotation_options opts;
  44         bool       use_tui, use_stdio, use_stdio2, use_gtk;
  45         bool       skip_missing;
  46         bool       has_br_stack;
  47         bool       group_set;
  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  */
  77 static 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
 124 static 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
 150 static 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);
 165
 166         if (err)
 167                 goto out;
 168
 169         err = addr_map_symbol__inc_samples(&bi->to, sample, evsel);
 170
 171 out:
 172         return err;
 173 }
 174
 175 static 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
 205 static bool has_annotation(struct perf_annotate *ann)
 206 {
 207         return ui__has_annotation() || ann->use_stdio2;
 208 }
 209
 210 static 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);
 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, al->addr);
 252         hists__inc_nr_samples(hists, true);
 253         return ret;
 254 }
 255
 256 static 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         }
 281 out_put:
 282         addr_location__put(&al);
 283         return ret;
 284 }
 285
 286 static int process_feature_event(struct perf_session *session,
 287                                  union perf_event *event)
 288 {
 289         if (event->feat.feat_id < HEADER_LAST_FEATURE)
 290                 return perf_event__process_feature(session, event);
 291         return 0;
 292 }
 293
 294 static int hist_entry__tty_annotate(struct hist_entry *he,
 295                                     struct perf_evsel *evsel,
 296                                     struct perf_annotate *ann)
 297 {
 298         if (!ann->use_stdio2)
 299                 return symbol__tty_annotate(he->ms.sym, he->ms.map, evsel, &ann->opts);
 300
 301         return symbol__tty_annotate2(he->ms.sym, he->ms.map, evsel, &ann->opts);
 302 }
 303
 304 static void hists__find_annotations(struct hists *hists,
 305                                     struct perf_evsel *evsel,
 306                                     struct perf_annotate *ann)
 307 {
 308         struct rb_node *nd = rb_first(&hists->entries), *next;
 309         int key = K_RIGHT;
 310
 311         while (nd) {
 312                 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
 313                 struct annotation *notes;
 314
 315                 if (he->ms.sym == NULL || he->ms.map->dso->annotate_warned)
 316                         goto find_next;
 317
 318                 if (ann->sym_hist_filter &&
 319                     (strcmp(he->ms.sym->name, ann->sym_hist_filter) != 0))
 320                         goto find_next;
 321
 322                 notes = symbol__annotation(he->ms.sym);
 323                 if (notes->src == NULL) {
 324 find_next:
 325                         if (key == K_LEFT)
 326                                 nd = rb_prev(nd);
 327                         else
 328                                 nd = rb_next(nd);
 329                         continue;
 330                 }
 331
 332                 if (use_browser == 2) {
 333                         int ret;
 334                         int (*annotate)(struct hist_entry *he,
 335                                         struct perf_evsel *evsel,
 336                                         struct hist_browser_timer *hbt);
 337
 338                         annotate = dlsym(perf_gtk_handle,
 339                                          "hist_entry__gtk_annotate");
 340                         if (annotate == NULL) {
 341                                 ui__error("GTK browser not found!\n");
 342                                 return;
 343                         }
 344
 345                         ret = annotate(he, evsel, NULL);
 346                         if (!ret || !ann->skip_missing)
 347                                 return;
 348
 349                         /* skip missing symbols */
 350                         nd = rb_next(nd);
 351                 } else if (use_browser == 1) {
 352                         key = hist_entry__tui_annotate(he, evsel, NULL, &ann->opts);
 353
 354                         switch (key) {
 355                         case -1:
 356                                 if (!ann->skip_missing)
 357                                         return;
 358                                 /* fall through */
 359                         case K_RIGHT:
 360                                 next = rb_next(nd);
 361                                 break;
 362                         case K_LEFT:
 363                                 next = rb_prev(nd);
 364                                 break;
 365                         default:
 366                                 return;
 367                         }
 368
 369                         if (next != NULL)
 370                                 nd = next;
 371                 } else {
 372                         hist_entry__tty_annotate(he, evsel, ann);
 373                         nd = rb_next(nd);
 374                         /*
 375                          * Since we have a hist_entry per IP for the same
 376                          * symbol, free he->ms.sym->src to signal we already
 377                          * processed this symbol.
 378                          */
 379                         zfree(&notes->src->cycles_hist);
 380                         zfree(&notes->src);
 381                 }
 382         }
 383 }
 384
 385 static int __cmd_annotate(struct perf_annotate *ann)
 386 {
 387         int ret;
 388         struct perf_session *session = ann->session;
 389         struct perf_evsel *pos;
 390         u64 total_nr_samples;
 391
 392         if (ann->cpu_list) {
 393                 ret = perf_session__cpu_bitmap(session, ann->cpu_list,
 394                                                ann->cpu_bitmap);
 395                 if (ret)
 396                         goto out;
 397         }
 398
 399         if (!ann->opts.objdump_path) {
 400                 ret = perf_env__lookup_objdump(&session->header.env,
 401                                                &ann->opts.objdump_path);
 402                 if (ret)
 403                         goto out;
 404         }
 405
 406         ret = perf_session__process_events(session);
 407         if (ret)
 408                 goto out;
 409
 410         if (dump_trace) {
 411                 perf_session__fprintf_nr_events(session, stdout);
 412                 perf_evlist__fprintf_nr_events(session->evlist, stdout);
 413                 goto out;
 414         }
 415
 416         if (verbose > 3)
 417                 perf_session__fprintf(session, stdout);
 418
 419         if (verbose > 2)
 420                 perf_session__fprintf_dsos(session, stdout);
 421
 422         total_nr_samples = 0;
 423         evlist__for_each_entry(session->evlist, pos) {
 424                 struct hists *hists = evsel__hists(pos);
 425                 u32 nr_samples = hists->stats.nr_events[PERF_RECORD_SAMPLE];
 426
 427                 if (nr_samples > 0) {
 428                         total_nr_samples += nr_samples;
 429                         hists__collapse_resort(hists, NULL);
 430                         /* Don't sort callchain */
 431                         perf_evsel__reset_sample_bit(pos, CALLCHAIN);
 432                         perf_evsel__output_resort(pos, NULL);
 433
 434                         if (symbol_conf.event_group &&
 435                             !perf_evsel__is_group_leader(pos))
 436                                 continue;
 437
 438                         hists__find_annotations(hists, pos, ann);
 439                 }
 440         }
 441
 442         if (total_nr_samples == 0) {
 443                 ui__error("The %s file has no samples!\n", session->data->file.path);
 444                 goto out;
 445         }
 446
 447         if (use_browser == 2) {
 448                 void (*show_annotations)(void);
 449
 450                 show_annotations = dlsym(perf_gtk_handle,
 451                                          "perf_gtk__show_annotations");
 452                 if (show_annotations == NULL) {
 453                         ui__error("GTK browser not found!\n");
 454                         goto out;
 455                 }
 456                 show_annotations();
 457         }
 458
 459 out:
 460         return ret;
 461 }
 462
 463 static const char * const annotate_usage[] = {
 464         "perf annotate [<options>]",
 465         NULL
 466 };
 467
 468 int cmd_annotate(int argc, const char **argv)
 469 {
 470         struct perf_annotate annotate = {
 471                 .tool = {
 472                         .sample = process_sample_event,
 473                         .mmap   = perf_event__process_mmap,
 474                         .mmap2  = perf_event__process_mmap2,
 475                         .comm   = perf_event__process_comm,
 476                         .exit   = perf_event__process_exit,
 477                         .fork   = perf_event__process_fork,
 478                         .namespaces = perf_event__process_namespaces,
 479                         .attr   = perf_event__process_attr,
 480                         .build_id = perf_event__process_build_id,
 481                         .tracing_data   = perf_event__process_tracing_data,
 482                         .feature        = process_feature_event,
 483                         .ordered_events = true,
 484                         .ordering_requires_timestamps = true,
 485                 },
 486                 .opts = annotation__default_options,
 487         };
 488         struct perf_data data = {
 489                 .mode  = PERF_DATA_MODE_READ,
 490         };
 491         struct option options[] = {
 492         OPT_STRING('i', "input", &input_name, "file",
 493                     "input file name"),
 494         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
 495                    "only consider symbols in these dsos"),
 496         OPT_STRING('s', "symbol", &annotate.sym_hist_filter, "symbol",
 497                     "symbol to annotate"),
 498         OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"),
 499         OPT_INCR('v', "verbose", &verbose,
 500                     "be more verbose (show symbol address, etc)"),
 501         OPT_BOOLEAN('q', "quiet", &quiet, "do now show any message"),
 502         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
 503                     "dump raw trace in ASCII"),
 504         OPT_BOOLEAN(0, "gtk", &annotate.use_gtk, "Use the GTK interface"),
 505         OPT_BOOLEAN(0, "tui", &annotate.use_tui, "Use the TUI interface"),
 506         OPT_BOOLEAN(0, "stdio", &annotate.use_stdio, "Use the stdio interface"),
 507         OPT_BOOLEAN(0, "stdio2", &annotate.use_stdio2, "Use the stdio interface"),
 508         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
 509                     "don't load vmlinux even if found"),
 510         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 511                    "file", "vmlinux pathname"),
 512         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
 513                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
 514         OPT_BOOLEAN('l', "print-line", &annotate.opts.print_lines,
 515                     "print matching source lines (may be slow)"),
 516         OPT_BOOLEAN('P', "full-paths", &annotate.opts.full_path,
 517                     "Don't shorten the displayed pathnames"),
 518         OPT_BOOLEAN(0, "skip-missing", &annotate.skip_missing,
 519                     "Skip symbols that cannot be annotated"),
 520         OPT_BOOLEAN_SET(0, "group", &symbol_conf.event_group,
 521                         &annotate.group_set,
 522                         "Show event group information together"),
 523         OPT_STRING('C', "cpu", &annotate.cpu_list, "cpu", "list of cpus to profile"),
 524         OPT_CALLBACK(0, "symfs", NULL, "directory",
 525                      "Look for files with symbols relative to this directory",
 526                      symbol__config_symfs),
 527         OPT_BOOLEAN(0, "source", &annotate.opts.annotate_src,
 528                     "Interleave source code with assembly code (default)"),
 529         OPT_BOOLEAN(0, "asm-raw", &annotate.opts.show_asm_raw,
 530                     "Display raw encoding of assembly instructions (default)"),
 531         OPT_STRING('M', "disassembler-style", &annotate.opts.disassembler_style, "disassembler style",
 532                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
 533         OPT_STRING(0, "objdump", &annotate.opts.objdump_path, "path",
 534                    "objdump binary to use for disassembly and annotations"),
 535         OPT_BOOLEAN(0, "group", &symbol_conf.event_group,
 536                     "Show event group information together"),
 537         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
 538                     "Show a column with the sum of periods"),
 539         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
 540                     "Show a column with the number of samples"),
 541         OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode",
 542                              "'always' (default), 'never' or 'auto' only applicable to --stdio mode",
 543                              stdio__config_color, "always"),
 544         OPT_CALLBACK(0, "percent-type", &annotate.opts, "local-period",
 545                      "Set percent type local/global-period/hits",
 546                      annotate_parse_percent_type),
 547
 548         OPT_END()
 549         };
 550         int ret;
 551
 552         set_option_flag(options, 0, "show-total-period", PARSE_OPT_EXCLUSIVE);
 553         set_option_flag(options, 0, "show-nr-samples", PARSE_OPT_EXCLUSIVE);
 554
 555
 556         ret = hists__init();
 557         if (ret < 0)
 558                 return ret;
 559
 560         argc = parse_options(argc, argv, options, annotate_usage, 0);
 561         if (argc) {
 562                 /*
 563                  * Special case: if there's an argument left then assume that
 564                  * it's a symbol filter:
 565                  */
 566                 if (argc > 1)
 567                         usage_with_options(annotate_usage, options);
 568
 569                 annotate.sym_hist_filter = argv[0];
 570         }
 571
 572         if (symbol_conf.show_nr_samples && annotate.use_gtk) {
 573                 pr_err("--show-nr-samples is not available in --gtk mode at this time\n");
 574                 return ret;
 575         }
 576
 577         if (quiet)
 578                 perf_quiet_option();
 579
 580         data.file.path = input_name;
 581
 582         annotate.session = perf_session__new(&data, false, &annotate.tool);
 583         if (annotate.session == NULL)
 584                 return -1;
 585
 586         annotate.has_br_stack = perf_header__has_feat(&annotate.session->header,
 587                                                       HEADER_BRANCH_STACK);
 588
 589         if (annotate.group_set)
 590                 perf_evlist__force_leader(annotate.session->evlist);
 591
 592         ret = symbol__annotation_init();
 593         if (ret < 0)
 594                 goto out_delete;
 595
 596         annotation_config__init();
 597
 598         symbol_conf.try_vmlinux_path = true;
 599
 600         ret = symbol__init(&annotate.session->header.env);
 601         if (ret < 0)
 602                 goto out_delete;
 603
 604         if (annotate.use_stdio || annotate.use_stdio2)
 605                 use_browser = 0;
 606         else if (annotate.use_tui)
 607                 use_browser = 1;
 608         else if (annotate.use_gtk)
 609                 use_browser = 2;
 610
 611         setup_browser(true);
 612
 613         if ((use_browser == 1 || annotate.use_stdio2) && annotate.has_br_stack) {
 614                 sort__mode = SORT_MODE__BRANCH;
 615                 if (setup_sorting(annotate.session->evlist) < 0)
 616                         usage_with_options(annotate_usage, options);
 617         } else {
 618                 if (setup_sorting(NULL) < 0)
 619                         usage_with_options(annotate_usage, options);
 620         }
 621
 622         ret = __cmd_annotate(&annotate);
 623
 624 out_delete:
 625         /*
 626          * Speed up the exit process, for large files this can
 627          * take quite a while.
 628          *
 629          * XXX Enable this when using valgrind or if we ever
 630          * librarize this command.
 631          *
 632          * Also experiment with obstacks to see how much speed
 633          * up we'll get here.
 634          *
 635          * perf_session__delete(session);
 636          */
 637         return ret;
 638 }