tools/perf/builtin-annotate.c

   1 /*
   2  * builtin-annotate.c
   3  *
   4  * Builtin annotate command: Analyze the perf.data input file,
   5  * look up and read DSOs and symbol information and display
   6  * a histogram of results, along various sorting keys.
   7  */
   8 #include "builtin.h"
   9
  10 #include "util/util.h"
  11 #include "util/color.h"
  12 #include <linux/list.h>
  13 #include "util/cache.h"
  14 #include <linux/rbtree.h>
  15 #include "util/symbol.h"
  16
  17 #include "perf.h"
  18 #include "util/debug.h"
  19
  20 #include "util/evlist.h"
  21 #include "util/evsel.h"
  22 #include "util/annotate.h"
  23 #include "util/event.h"
  24 #include <subcmd/parse-options.h>
  25 #include "util/parse-events.h"
  26 #include "util/thread.h"
  27 #include "util/sort.h"
  28 #include "util/hist.h"
  29 #include "util/session.h"
  30 #include "util/tool.h"
  31 #include "util/data.h"
  32 #include "arch/common.h"
  33 #include "util/block-range.h"
  34
  35 #include <dlfcn.h>
  36 #include <errno.h>
  37 #include <linux/bitmap.h>
  38
  39 struct perf_annotate {
  40         struct perf_tool tool;
  41         struct perf_session *session;
  42         bool       use_tui, use_stdio, use_gtk;
  43         bool       full_paths;
  44         bool       print_line;
  45         bool       skip_missing;
  46         const char *sym_hist_filter;
  47         const char *cpu_list;
  48         DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  49 };
  50
  51 /*
  52  * Given one basic block:
  53  *
  54  *      from    to              branch_i
  55  *      * ----> *
  56  *              |
  57  *              | block
  58  *              v
  59  *              * ----> *
  60  *              from    to      branch_i+1
  61  *
  62  * where the horizontal are the branches and the vertical is the executed
  63  * block of instructions.
  64  *
  65  * We count, for each 'instruction', the number of blocks that covered it as
  66  * well as count the ratio each branch is taken.
  67  *
  68  * We can do this without knowing the actual instruction stream by keeping
  69  * track of the address ranges. We break down ranges such that there is no
  70  * overlap and iterate from the start until the end.
  71  *
  72  * @acme: once we parse the objdump output _before_ processing the samples,
  73  * we can easily fold the branch.cycles IPC bits in.
  74  */
  75 static void process_basic_block(struct addr_map_symbol *start,
  76                                 struct addr_map_symbol *end,
  77                                 struct branch_flags *flags)
  78 {
  79         struct symbol *sym = start->sym;
  80         struct annotation *notes = sym ? symbol__annotation(sym) : NULL;
  81         struct block_range_iter iter;
  82         struct block_range *entry;
  83
  84         /*
  85          * Sanity; NULL isn't executable and the CPU cannot execute backwards
  86          */
  87         if (!start->addr || start->addr > end->addr)
  88                 return;
  89
  90         iter = block_range__create(start->addr, end->addr);
  91         if (!block_range_iter__valid(&iter))
  92                 return;
  93
  94         /*
  95          * First block in range is a branch target.
  96          */
  97         entry = block_range_iter(&iter);
  98         assert(entry->is_target);
  99         entry->entry++;
 100
 101         do {
 102                 entry = block_range_iter(&iter);
 103
 104                 entry->coverage++;
 105                 entry->sym = sym;
 106
 107                 if (notes)
 108                         notes->max_coverage = max(notes->max_coverage, entry->coverage);
 109
 110         } while (block_range_iter__next(&iter));
 111
 112         /*
 113          * Last block in rage is a branch.
 114          */
 115         entry = block_range_iter(&iter);
 116         assert(entry->is_branch);
 117         entry->taken++;
 118         if (flags->predicted)
 119                 entry->pred++;
 120 }
 121
 122 static void process_branch_stack(struct branch_stack *bs, struct addr_location *al,
 123                                  struct perf_sample *sample)
 124 {
 125         struct addr_map_symbol *prev = NULL;
 126         struct branch_info *bi;
 127         int i;
 128
 129         if (!bs || !bs->nr)
 130                 return;
 131
 132         bi = sample__resolve_bstack(sample, al);
 133         if (!bi)
 134                 return;
 135
 136         for (i = bs->nr - 1; i >= 0; i--) {
 137                 /*
 138                  * XXX filter against symbol
 139                  */
 140                 if (prev)
 141                         process_basic_block(prev, &bi[i].from, &bi[i].flags);
 142                 prev = &bi[i].to;
 143         }
 144
 145         free(bi);
 146 }
 147
 148 static int perf_evsel__add_sample(struct perf_evsel *evsel,
 149                                   struct perf_sample *sample,
 150                                   struct addr_location *al,
 151                                   struct perf_annotate *ann)
 152 {
 153         struct hists *hists = evsel__hists(evsel);
 154         struct hist_entry *he;
 155         int ret;
 156
 157         if (ann->sym_hist_filter != NULL &&
 158             (al->sym == NULL ||
 159              strcmp(ann->sym_hist_filter, al->sym->name) != 0)) {
 160                 /* We're only interested in a symbol named sym_hist_filter */
 161                 /*
 162                  * FIXME: why isn't this done in the symbol_filter when loading
 163                  * the DSO?
 164                  */
 165                 if (al->sym != NULL) {
 166                         rb_erase(&al->sym->rb_node,
 167                                  &al->map->dso->symbols[al->map->type]);
 168                         symbol__delete(al->sym);
 169                         dso__reset_find_symbol_cache(al->map->dso);
 170                 }
 171                 return 0;
 172         }
 173
 174         /*
 175          * XXX filtered samples can still have branch entires pointing into our
 176          * symbol and are missed.
 177          */
 178         process_branch_stack(sample->branch_stack, al, sample);
 179
 180         he = hists__add_entry(hists, al, NULL, NULL, NULL, sample, true);
 181         if (he == NULL)
 182                 return -ENOMEM;
 183
 184         ret = hist_entry__inc_addr_samples(he, sample, evsel->idx, al->addr);
 185         hists__inc_nr_samples(hists, true);
 186         return ret;
 187 }
 188
 189 static int process_sample_event(struct perf_tool *tool,
 190                                 union perf_event *event,
 191                                 struct perf_sample *sample,
 192                                 struct perf_evsel *evsel,
 193                                 struct machine *machine)
 194 {
 195         struct perf_annotate *ann = container_of(tool, struct perf_annotate, tool);
 196         struct addr_location al;
 197         int ret = 0;
 198
 199         if (machine__resolve(machine, &al, sample) < 0) {
 200                 pr_warning("problem processing %d event, skipping it.\n",
 201                            event->header.type);
 202                 return -1;
 203         }
 204
 205         if (ann->cpu_list && !test_bit(sample->cpu, ann->cpu_bitmap))
 206                 goto out_put;
 207
 208         if (!al.filtered && perf_evsel__add_sample(evsel, sample, &al, ann)) {
 209                 pr_warning("problem incrementing symbol count, "
 210                            "skipping event\n");
 211                 ret = -1;
 212         }
 213 out_put:
 214         addr_location__put(&al);
 215         return ret;
 216 }
 217
 218 static int hist_entry__tty_annotate(struct hist_entry *he,
 219                                     struct perf_evsel *evsel,
 220                                     struct perf_annotate *ann)
 221 {
 222         return symbol__tty_annotate(he->ms.sym, he->ms.map, evsel,
 223                                     ann->print_line, ann->full_paths, 0, 0);
 224 }
 225
 226 static void hists__find_annotations(struct hists *hists,
 227                                     struct perf_evsel *evsel,
 228                                     struct perf_annotate *ann)
 229 {
 230         struct rb_node *nd = rb_first(&hists->entries), *next;
 231         int key = K_RIGHT;
 232
 233         while (nd) {
 234                 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
 235                 struct annotation *notes;
 236
 237                 if (he->ms.sym == NULL || he->ms.map->dso->annotate_warned)
 238                         goto find_next;
 239
 240                 notes = symbol__annotation(he->ms.sym);
 241                 if (notes->src == NULL) {
 242 find_next:
 243                         if (key == K_LEFT)
 244                                 nd = rb_prev(nd);
 245                         else
 246                                 nd = rb_next(nd);
 247                         continue;
 248                 }
 249
 250                 if (use_browser == 2) {
 251                         int ret;
 252                         int (*annotate)(struct hist_entry *he,
 253                                         struct perf_evsel *evsel,
 254                                         struct hist_browser_timer *hbt);
 255
 256                         annotate = dlsym(perf_gtk_handle,
 257                                          "hist_entry__gtk_annotate");
 258                         if (annotate == NULL) {
 259                                 ui__error("GTK browser not found!\n");
 260                                 return;
 261                         }
 262
 263                         ret = annotate(he, evsel, NULL);
 264                         if (!ret || !ann->skip_missing)
 265                                 return;
 266
 267                         /* skip missing symbols */
 268                         nd = rb_next(nd);
 269                 } else if (use_browser == 1) {
 270                         key = hist_entry__tui_annotate(he, evsel, NULL);
 271                         switch (key) {
 272                         case -1:
 273                                 if (!ann->skip_missing)
 274                                         return;
 275                                 /* fall through */
 276                         case K_RIGHT:
 277                                 next = rb_next(nd);
 278                                 break;
 279                         case K_LEFT:
 280                                 next = rb_prev(nd);
 281                                 break;
 282                         default:
 283                                 return;
 284                         }
 285
 286                         if (next != NULL)
 287                                 nd = next;
 288                 } else {
 289                         hist_entry__tty_annotate(he, evsel, ann);
 290                         nd = rb_next(nd);
 291                         /*
 292                          * Since we have a hist_entry per IP for the same
 293                          * symbol, free he->ms.sym->src to signal we already
 294                          * processed this symbol.
 295                          */
 296                         zfree(&notes->src->cycles_hist);
 297                         zfree(&notes->src);
 298                 }
 299         }
 300 }
 301
 302 static int __cmd_annotate(struct perf_annotate *ann)
 303 {
 304         int ret;
 305         struct perf_session *session = ann->session;
 306         struct perf_evsel *pos;
 307         u64 total_nr_samples;
 308
 309         if (ann->cpu_list) {
 310                 ret = perf_session__cpu_bitmap(session, ann->cpu_list,
 311                                                ann->cpu_bitmap);
 312                 if (ret)
 313                         goto out;
 314         }
 315
 316         if (!objdump_path) {
 317                 ret = perf_env__lookup_objdump(&session->header.env);
 318                 if (ret)
 319                         goto out;
 320         }
 321
 322         ret = perf_session__process_events(session);
 323         if (ret)
 324                 goto out;
 325
 326         if (dump_trace) {
 327                 perf_session__fprintf_nr_events(session, stdout);
 328                 perf_evlist__fprintf_nr_events(session->evlist, stdout);
 329                 goto out;
 330         }
 331
 332         if (verbose > 3)
 333                 perf_session__fprintf(session, stdout);
 334
 335         if (verbose > 2)
 336                 perf_session__fprintf_dsos(session, stdout);
 337
 338         total_nr_samples = 0;
 339         evlist__for_each_entry(session->evlist, pos) {
 340                 struct hists *hists = evsel__hists(pos);
 341                 u32 nr_samples = hists->stats.nr_events[PERF_RECORD_SAMPLE];
 342
 343                 if (nr_samples > 0) {
 344                         total_nr_samples += nr_samples;
 345                         hists__collapse_resort(hists, NULL);
 346                         /* Don't sort callchain */
 347                         perf_evsel__reset_sample_bit(pos, CALLCHAIN);
 348                         perf_evsel__output_resort(pos, NULL);
 349
 350                         if (symbol_conf.event_group &&
 351                             !perf_evsel__is_group_leader(pos))
 352                                 continue;
 353
 354                         hists__find_annotations(hists, pos, ann);
 355                 }
 356         }
 357
 358         if (total_nr_samples == 0) {
 359                 ui__error("The %s file has no samples!\n", session->file->path);
 360                 goto out;
 361         }
 362
 363         if (use_browser == 2) {
 364                 void (*show_annotations)(void);
 365
 366                 show_annotations = dlsym(perf_gtk_handle,
 367                                          "perf_gtk__show_annotations");
 368                 if (show_annotations == NULL) {
 369                         ui__error("GTK browser not found!\n");
 370                         goto out;
 371                 }
 372                 show_annotations();
 373         }
 374
 375 out:
 376         return ret;
 377 }
 378
 379 static const char * const annotate_usage[] = {
 380         "perf annotate [<options>]",
 381         NULL
 382 };
 383
 384 int cmd_annotate(int argc, const char **argv)
 385 {
 386         struct perf_annotate annotate = {
 387                 .tool = {
 388                         .sample = process_sample_event,
 389                         .mmap   = perf_event__process_mmap,
 390                         .mmap2  = perf_event__process_mmap2,
 391                         .comm   = perf_event__process_comm,
 392                         .exit   = perf_event__process_exit,
 393                         .fork   = perf_event__process_fork,
 394                         .namespaces = perf_event__process_namespaces,
 395                         .attr   = perf_event__process_attr,
 396                         .build_id = perf_event__process_build_id,
 397                         .tracing_data   = perf_event__process_tracing_data,
 398                         .feature        = perf_event__process_feature,
 399                         .ordered_events = true,
 400                         .ordering_requires_timestamps = true,
 401                 },
 402         };
 403         struct perf_data_file file = {
 404                 .mode  = PERF_DATA_MODE_READ,
 405         };
 406         struct option options[] = {
 407         OPT_STRING('i', "input", &input_name, "file",
 408                     "input file name"),
 409         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
 410                    "only consider symbols in these dsos"),
 411         OPT_STRING('s', "symbol", &annotate.sym_hist_filter, "symbol",
 412                     "symbol to annotate"),
 413         OPT_BOOLEAN('f', "force", &file.force, "don't complain, do it"),
 414         OPT_INCR('v', "verbose", &verbose,
 415                     "be more verbose (show symbol address, etc)"),
 416         OPT_BOOLEAN('q', "quiet", &quiet, "do now show any message"),
 417         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
 418                     "dump raw trace in ASCII"),
 419         OPT_BOOLEAN(0, "gtk", &annotate.use_gtk, "Use the GTK interface"),
 420         OPT_BOOLEAN(0, "tui", &annotate.use_tui, "Use the TUI interface"),
 421         OPT_BOOLEAN(0, "stdio", &annotate.use_stdio, "Use the stdio interface"),
 422         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 423                    "file", "vmlinux pathname"),
 424         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
 425                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
 426         OPT_BOOLEAN('l', "print-line", &annotate.print_line,
 427                     "print matching source lines (may be slow)"),
 428         OPT_BOOLEAN('P', "full-paths", &annotate.full_paths,
 429                     "Don't shorten the displayed pathnames"),
 430         OPT_BOOLEAN(0, "skip-missing", &annotate.skip_missing,
 431                     "Skip symbols that cannot be annotated"),
 432         OPT_STRING('C', "cpu", &annotate.cpu_list, "cpu", "list of cpus to profile"),
 433         OPT_CALLBACK(0, "symfs", NULL, "directory",
 434                      "Look for files with symbols relative to this directory",
 435                      symbol__config_symfs),
 436         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
 437                     "Interleave source code with assembly code (default)"),
 438         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
 439                     "Display raw encoding of assembly instructions (default)"),
 440         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
 441                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
 442         OPT_STRING(0, "objdump", &objdump_path, "path",
 443                    "objdump binary to use for disassembly and annotations"),
 444         OPT_BOOLEAN(0, "group", &symbol_conf.event_group,
 445                     "Show event group information together"),
 446         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
 447                     "Show a column with the sum of periods"),
 448         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
 449                     "Show a column with the number of samples"),
 450         OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode",
 451                              "'always' (default), 'never' or 'auto' only applicable to --stdio mode",
 452                              stdio__config_color, "always"),
 453         OPT_END()
 454         };
 455         int ret;
 456
 457         set_option_flag(options, 0, "show-total-period", PARSE_OPT_EXCLUSIVE);
 458         set_option_flag(options, 0, "show-nr-samples", PARSE_OPT_EXCLUSIVE);
 459
 460
 461         ret = hists__init();
 462         if (ret < 0)
 463                 return ret;
 464
 465         argc = parse_options(argc, argv, options, annotate_usage, 0);
 466         if (argc) {
 467                 /*
 468                  * Special case: if there's an argument left then assume that
 469                  * it's a symbol filter:
 470                  */
 471                 if (argc > 1)
 472                         usage_with_options(annotate_usage, options);
 473
 474                 annotate.sym_hist_filter = argv[0];
 475         }
 476
 477         if (symbol_conf.show_nr_samples && annotate.use_gtk) {
 478                 pr_err("--show-nr-samples is not available in --gtk mode at this time\n");
 479                 return ret;
 480         }
 481
 482         if (quiet)
 483                 perf_quiet_option();
 484
 485         file.path  = input_name;
 486
 487         annotate.session = perf_session__new(&file, false, &annotate.tool);
 488         if (annotate.session == NULL)
 489                 return -1;
 490
 491         ret = symbol__annotation_init();
 492         if (ret < 0)
 493                 goto out_delete;
 494
 495         symbol_conf.try_vmlinux_path = true;
 496
 497         ret = symbol__init(&annotate.session->header.env);
 498         if (ret < 0)
 499                 goto out_delete;
 500
 501         if (setup_sorting(NULL) < 0)
 502                 usage_with_options(annotate_usage, options);
 503
 504         if (annotate.use_stdio)
 505                 use_browser = 0;
 506         else if (annotate.use_tui)
 507                 use_browser = 1;
 508         else if (annotate.use_gtk)
 509                 use_browser = 2;
 510
 511         setup_browser(true);
 512
 513         ret = __cmd_annotate(&annotate);
 514
 515 out_delete:
 516         /*
 517          * Speed up the exit process, for large files this can
 518          * take quite a while.
 519          *
 520          * XXX Enable this when using valgrind or if we ever
 521          * librarize this command.
 522          *
 523          * Also experiment with obstacks to see how much speed
 524          * up we'll get here.
 525          *
 526          * perf_session__delete(session);
 527          */
 528         return ret;
 529 }