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