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         sample->period = 1;
 181         sample->weight = 1;
 182
 183         he = hists__add_entry(hists, al, NULL, NULL, NULL, sample, true);
 184         if (he == NULL)
 185                 return -ENOMEM;
 186
 187         ret = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
 188         hists__inc_nr_samples(hists, true);
 189         return ret;
 190 }
 191
 192 static int process_sample_event(struct perf_tool *tool,
 193                                 union perf_event *event,
 194                                 struct perf_sample *sample,
 195                                 struct perf_evsel *evsel,
 196                                 struct machine *machine)
 197 {
 198         struct perf_annotate *ann = container_of(tool, struct perf_annotate, tool);
 199         struct addr_location al;
 200         int ret = 0;
 201
 202         if (machine__resolve(machine, &al, sample) < 0) {
 203                 pr_warning("problem processing %d event, skipping it.\n",
 204                            event->header.type);
 205                 return -1;
 206         }
 207
 208         if (ann->cpu_list && !test_bit(sample->cpu, ann->cpu_bitmap))
 209                 goto out_put;
 210
 211         if (!al.filtered && perf_evsel__add_sample(evsel, sample, &al, ann)) {
 212                 pr_warning("problem incrementing symbol count, "
 213                            "skipping event\n");
 214                 ret = -1;
 215         }
 216 out_put:
 217         addr_location__put(&al);
 218         return ret;
 219 }
 220
 221 static int hist_entry__tty_annotate(struct hist_entry *he,
 222                                     struct perf_evsel *evsel,
 223                                     struct perf_annotate *ann)
 224 {
 225         return symbol__tty_annotate(he->ms.sym, he->ms.map, evsel,
 226                                     ann->print_line, ann->full_paths, 0, 0);
 227 }
 228
 229 static void hists__find_annotations(struct hists *hists,
 230                                     struct perf_evsel *evsel,
 231                                     struct perf_annotate *ann)
 232 {
 233         struct rb_node *nd = rb_first(&hists->entries), *next;
 234         int key = K_RIGHT;
 235
 236         while (nd) {
 237                 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
 238                 struct annotation *notes;
 239
 240                 if (he->ms.sym == NULL || he->ms.map->dso->annotate_warned)
 241                         goto find_next;
 242
 243                 notes = symbol__annotation(he->ms.sym);
 244                 if (notes->src == NULL) {
 245 find_next:
 246                         if (key == K_LEFT)
 247                                 nd = rb_prev(nd);
 248                         else
 249                                 nd = rb_next(nd);
 250                         continue;
 251                 }
 252
 253                 if (use_browser == 2) {
 254                         int ret;
 255                         int (*annotate)(struct hist_entry *he,
 256                                         struct perf_evsel *evsel,
 257                                         struct hist_browser_timer *hbt);
 258
 259                         annotate = dlsym(perf_gtk_handle,
 260                                          "hist_entry__gtk_annotate");
 261                         if (annotate == NULL) {
 262                                 ui__error("GTK browser not found!\n");
 263                                 return;
 264                         }
 265
 266                         ret = annotate(he, evsel, NULL);
 267                         if (!ret || !ann->skip_missing)
 268                                 return;
 269
 270                         /* skip missing symbols */
 271                         nd = rb_next(nd);
 272                 } else if (use_browser == 1) {
 273                         key = hist_entry__tui_annotate(he, evsel, NULL);
 274                         switch (key) {
 275                         case -1:
 276                                 if (!ann->skip_missing)
 277                                         return;
 278                                 /* fall through */
 279                         case K_RIGHT:
 280                                 next = rb_next(nd);
 281                                 break;
 282                         case K_LEFT:
 283                                 next = rb_prev(nd);
 284                                 break;
 285                         default:
 286                                 return;
 287                         }
 288
 289                         if (next != NULL)
 290                                 nd = next;
 291                 } else {
 292                         hist_entry__tty_annotate(he, evsel, ann);
 293                         nd = rb_next(nd);
 294                         /*
 295                          * Since we have a hist_entry per IP for the same
 296                          * symbol, free he->ms.sym->src to signal we already
 297                          * processed this symbol.
 298                          */
 299                         zfree(&notes->src->cycles_hist);
 300                         zfree(&notes->src);
 301                 }
 302         }
 303 }
 304
 305 static int __cmd_annotate(struct perf_annotate *ann)
 306 {
 307         int ret;
 308         struct perf_session *session = ann->session;
 309         struct perf_evsel *pos;
 310         u64 total_nr_samples;
 311
 312         if (ann->cpu_list) {
 313                 ret = perf_session__cpu_bitmap(session, ann->cpu_list,
 314                                                ann->cpu_bitmap);
 315                 if (ret)
 316                         goto out;
 317         }
 318
 319         if (!objdump_path) {
 320                 ret = perf_env__lookup_objdump(&session->header.env);
 321                 if (ret)
 322                         goto out;
 323         }
 324
 325         ret = perf_session__process_events(session);
 326         if (ret)
 327                 goto out;
 328
 329         if (dump_trace) {
 330                 perf_session__fprintf_nr_events(session, stdout);
 331                 perf_evlist__fprintf_nr_events(session->evlist, stdout);
 332                 goto out;
 333         }
 334
 335         if (verbose > 3)
 336                 perf_session__fprintf(session, stdout);
 337
 338         if (verbose > 2)
 339                 perf_session__fprintf_dsos(session, stdout);
 340
 341         total_nr_samples = 0;
 342         evlist__for_each_entry(session->evlist, pos) {
 343                 struct hists *hists = evsel__hists(pos);
 344                 u32 nr_samples = hists->stats.nr_events[PERF_RECORD_SAMPLE];
 345
 346                 if (nr_samples > 0) {
 347                         total_nr_samples += nr_samples;
 348                         hists__collapse_resort(hists, NULL);
 349                         /* Don't sort callchain */
 350                         perf_evsel__reset_sample_bit(pos, CALLCHAIN);
 351                         perf_evsel__output_resort(pos, NULL);
 352
 353                         if (symbol_conf.event_group &&
 354                             !perf_evsel__is_group_leader(pos))
 355                                 continue;
 356
 357                         hists__find_annotations(hists, pos, ann);
 358                 }
 359         }
 360
 361         if (total_nr_samples == 0) {
 362                 ui__error("The %s file has no samples!\n", session->file->path);
 363                 goto out;
 364         }
 365
 366         if (use_browser == 2) {
 367                 void (*show_annotations)(void);
 368
 369                 show_annotations = dlsym(perf_gtk_handle,
 370                                          "perf_gtk__show_annotations");
 371                 if (show_annotations == NULL) {
 372                         ui__error("GTK browser not found!\n");
 373                         goto out;
 374                 }
 375                 show_annotations();
 376         }
 377
 378 out:
 379         return ret;
 380 }
 381
 382 static const char * const annotate_usage[] = {
 383         "perf annotate [<options>]",
 384         NULL
 385 };
 386
 387 int cmd_annotate(int argc, const char **argv)
 388 {
 389         struct perf_annotate annotate = {
 390                 .tool = {
 391                         .sample = process_sample_event,
 392                         .mmap   = perf_event__process_mmap,
 393                         .mmap2  = perf_event__process_mmap2,
 394                         .comm   = perf_event__process_comm,
 395                         .exit   = perf_event__process_exit,
 396                         .fork   = perf_event__process_fork,
 397                         .namespaces = perf_event__process_namespaces,
 398                         .attr   = perf_event__process_attr,
 399                         .build_id = perf_event__process_build_id,
 400                         .ordered_events = true,
 401                         .ordering_requires_timestamps = true,
 402                 },
 403         };
 404         struct perf_data_file file = {
 405                 .mode  = PERF_DATA_MODE_READ,
 406         };
 407         const struct option options[] = {
 408         OPT_STRING('i', "input", &input_name, "file",
 409                     "input file name"),
 410         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
 411                    "only consider symbols in these dsos"),
 412         OPT_STRING('s', "symbol", &annotate.sym_hist_filter, "symbol",
 413                     "symbol to annotate"),
 414         OPT_BOOLEAN('f', "force", &file.force, "don't complain, do it"),
 415         OPT_INCR('v', "verbose", &verbose,
 416                     "be more verbose (show symbol address, etc)"),
 417         OPT_BOOLEAN('q', "quiet", &quiet, "do now show any message"),
 418         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
 419                     "dump raw trace in ASCII"),
 420         OPT_BOOLEAN(0, "gtk", &annotate.use_gtk, "Use the GTK interface"),
 421         OPT_BOOLEAN(0, "tui", &annotate.use_tui, "Use the TUI interface"),
 422         OPT_BOOLEAN(0, "stdio", &annotate.use_stdio, "Use the stdio interface"),
 423         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 424                    "file", "vmlinux pathname"),
 425         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
 426                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
 427         OPT_BOOLEAN('l', "print-line", &annotate.print_line,
 428                     "print matching source lines (may be slow)"),
 429         OPT_BOOLEAN('P', "full-paths", &annotate.full_paths,
 430                     "Don't shorten the displayed pathnames"),
 431         OPT_BOOLEAN(0, "skip-missing", &annotate.skip_missing,
 432                     "Skip symbols that cannot be annotated"),
 433         OPT_STRING('C', "cpu", &annotate.cpu_list, "cpu", "list of cpus to profile"),
 434         OPT_CALLBACK(0, "symfs", NULL, "directory",
 435                      "Look for files with symbols relative to this directory",
 436                      symbol__config_symfs),
 437         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
 438                     "Interleave source code with assembly code (default)"),
 439         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
 440                     "Display raw encoding of assembly instructions (default)"),
 441         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
 442                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
 443         OPT_STRING(0, "objdump", &objdump_path, "path",
 444                    "objdump binary to use for disassembly and annotations"),
 445         OPT_BOOLEAN(0, "group", &symbol_conf.event_group,
 446                     "Show event group information together"),
 447         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
 448                     "Show a column with the sum of periods"),
 449         OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode",
 450                              "'always' (default), 'never' or 'auto' only applicable to --stdio mode",
 451                              stdio__config_color, "always"),
 452         OPT_END()
 453         };
 454         int ret = hists__init();
 455
 456         if (ret < 0)
 457                 return ret;
 458
 459         argc = parse_options(argc, argv, options, annotate_usage, 0);
 460         if (argc) {
 461                 /*
 462                  * Special case: if there's an argument left then assume that
 463                  * it's a symbol filter:
 464                  */
 465                 if (argc > 1)
 466                         usage_with_options(annotate_usage, options);
 467
 468                 annotate.sym_hist_filter = argv[0];
 469         }
 470
 471         if (quiet)
 472                 perf_quiet_option();
 473
 474         file.path  = input_name;
 475
 476         annotate.session = perf_session__new(&file, false, &annotate.tool);
 477         if (annotate.session == NULL)
 478                 return -1;
 479
 480         ret = symbol__annotation_init();
 481         if (ret < 0)
 482                 goto out_delete;
 483
 484         symbol_conf.try_vmlinux_path = true;
 485
 486         ret = symbol__init(&annotate.session->header.env);
 487         if (ret < 0)
 488                 goto out_delete;
 489
 490         if (setup_sorting(NULL) < 0)
 491                 usage_with_options(annotate_usage, options);
 492
 493         if (annotate.use_stdio)
 494                 use_browser = 0;
 495         else if (annotate.use_tui)
 496                 use_browser = 1;
 497         else if (annotate.use_gtk)
 498                 use_browser = 2;
 499
 500         setup_browser(true);
 501
 502         ret = __cmd_annotate(&annotate);
 503
 504 out_delete:
 505         /*
 506          * Speed up the exit process, for large files this can
 507          * take quite a while.
 508          *
 509          * XXX Enable this when using valgrind or if we ever
 510          * librarize this command.
 511          *
 512          * Also experiment with obstacks to see how much speed
 513          * up we'll get here.
 514          *
 515          * perf_session__delete(session);
 516          */
 517         return ret;
 518 }