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 "util/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
  34 #include <dlfcn.h>
  35 #include <linux/bitmap.h>
  36
  37 struct perf_annotate {
  38         struct perf_tool tool;
  39         struct perf_session *session;
  40         bool       use_tui, use_stdio, use_gtk;
  41         bool       full_paths;
  42         bool       print_line;
  43         bool       skip_missing;
  44         const char *sym_hist_filter;
  45         const char *cpu_list;
  46         DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  47 };
  48
  49 static int perf_evsel__add_sample(struct perf_evsel *evsel,
  50                                   struct perf_sample *sample __maybe_unused,
  51                                   struct addr_location *al,
  52                                   struct perf_annotate *ann)
  53 {
  54         struct hists *hists = evsel__hists(evsel);
  55         struct hist_entry *he;
  56         int ret;
  57
  58         if (ann->sym_hist_filter != NULL &&
  59             (al->sym == NULL ||
  60              strcmp(ann->sym_hist_filter, al->sym->name) != 0)) {
  61                 /* We're only interested in a symbol named sym_hist_filter */
  62                 /*
  63                  * FIXME: why isn't this done in the symbol_filter when loading
  64                  * the DSO?
  65                  */
  66                 if (al->sym != NULL) {
  67                         rb_erase(&al->sym->rb_node,
  68                                  &al->map->dso->symbols[al->map->type]);
  69                         symbol__delete(al->sym);
  70                         dso__reset_find_symbol_cache(al->map->dso);
  71                 }
  72                 return 0;
  73         }
  74
  75         he = __hists__add_entry(hists, al, NULL, NULL, NULL, 1, 1, 0, true);
  76         if (he == NULL)
  77                 return -ENOMEM;
  78
  79         ret = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
  80         hists__inc_nr_samples(hists, true);
  81         return ret;
  82 }
  83
  84 static int process_sample_event(struct perf_tool *tool,
  85                                 union perf_event *event,
  86                                 struct perf_sample *sample,
  87                                 struct perf_evsel *evsel,
  88                                 struct machine *machine)
  89 {
  90         struct perf_annotate *ann = container_of(tool, struct perf_annotate, tool);
  91         struct addr_location al;
  92         int ret = 0;
  93
  94         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0) {
  95                 pr_warning("problem processing %d event, skipping it.\n",
  96                            event->header.type);
  97                 return -1;
  98         }
  99
 100         if (ann->cpu_list && !test_bit(sample->cpu, ann->cpu_bitmap))
 101                 goto out_put;
 102
 103         if (!al.filtered && perf_evsel__add_sample(evsel, sample, &al, ann)) {
 104                 pr_warning("problem incrementing symbol count, "
 105                            "skipping event\n");
 106                 ret = -1;
 107         }
 108 out_put:
 109         addr_location__put(&al);
 110         return ret;
 111 }
 112
 113 static int hist_entry__tty_annotate(struct hist_entry *he,
 114                                     struct perf_evsel *evsel,
 115                                     struct perf_annotate *ann)
 116 {
 117         return symbol__tty_annotate(he->ms.sym, he->ms.map, evsel,
 118                                     ann->print_line, ann->full_paths, 0, 0);
 119 }
 120
 121 static void hists__find_annotations(struct hists *hists,
 122                                     struct perf_evsel *evsel,
 123                                     struct perf_annotate *ann)
 124 {
 125         struct rb_node *nd = rb_first(&hists->entries), *next;
 126         int key = K_RIGHT;
 127
 128         while (nd) {
 129                 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
 130                 struct annotation *notes;
 131
 132                 if (he->ms.sym == NULL || he->ms.map->dso->annotate_warned)
 133                         goto find_next;
 134
 135                 notes = symbol__annotation(he->ms.sym);
 136                 if (notes->src == NULL) {
 137 find_next:
 138                         if (key == K_LEFT)
 139                                 nd = rb_prev(nd);
 140                         else
 141                                 nd = rb_next(nd);
 142                         continue;
 143                 }
 144
 145                 if (use_browser == 2) {
 146                         int ret;
 147                         int (*annotate)(struct hist_entry *he,
 148                                         struct perf_evsel *evsel,
 149                                         struct hist_browser_timer *hbt);
 150
 151                         annotate = dlsym(perf_gtk_handle,
 152                                          "hist_entry__gtk_annotate");
 153                         if (annotate == NULL) {
 154                                 ui__error("GTK browser not found!\n");
 155                                 return;
 156                         }
 157
 158                         ret = annotate(he, evsel, NULL);
 159                         if (!ret || !ann->skip_missing)
 160                                 return;
 161
 162                         /* skip missing symbols */
 163                         nd = rb_next(nd);
 164                 } else if (use_browser == 1) {
 165                         key = hist_entry__tui_annotate(he, evsel, NULL);
 166                         switch (key) {
 167                         case -1:
 168                                 if (!ann->skip_missing)
 169                                         return;
 170                                 /* fall through */
 171                         case K_RIGHT:
 172                                 next = rb_next(nd);
 173                                 break;
 174                         case K_LEFT:
 175                                 next = rb_prev(nd);
 176                                 break;
 177                         default:
 178                                 return;
 179                         }
 180
 181                         if (next != NULL)
 182                                 nd = next;
 183                 } else {
 184                         hist_entry__tty_annotate(he, evsel, ann);
 185                         nd = rb_next(nd);
 186                         /*
 187                          * Since we have a hist_entry per IP for the same
 188                          * symbol, free he->ms.sym->src to signal we already
 189                          * processed this symbol.
 190                          */
 191                         zfree(&notes->src->cycles_hist);
 192                         zfree(&notes->src);
 193                 }
 194         }
 195 }
 196
 197 static int __cmd_annotate(struct perf_annotate *ann)
 198 {
 199         int ret;
 200         struct perf_session *session = ann->session;
 201         struct perf_evsel *pos;
 202         u64 total_nr_samples;
 203
 204         machines__set_symbol_filter(&session->machines, symbol__annotate_init);
 205
 206         if (ann->cpu_list) {
 207                 ret = perf_session__cpu_bitmap(session, ann->cpu_list,
 208                                                ann->cpu_bitmap);
 209                 if (ret)
 210                         goto out;
 211         }
 212
 213         if (!objdump_path) {
 214                 ret = perf_env__lookup_objdump(&session->header.env);
 215                 if (ret)
 216                         goto out;
 217         }
 218
 219         ret = perf_session__process_events(session);
 220         if (ret)
 221                 goto out;
 222
 223         if (dump_trace) {
 224                 perf_session__fprintf_nr_events(session, stdout);
 225                 perf_evlist__fprintf_nr_events(session->evlist, stdout);
 226                 goto out;
 227         }
 228
 229         if (verbose > 3)
 230                 perf_session__fprintf(session, stdout);
 231
 232         if (verbose > 2)
 233                 perf_session__fprintf_dsos(session, stdout);
 234
 235         total_nr_samples = 0;
 236         evlist__for_each(session->evlist, pos) {
 237                 struct hists *hists = evsel__hists(pos);
 238                 u32 nr_samples = hists->stats.nr_events[PERF_RECORD_SAMPLE];
 239
 240                 if (nr_samples > 0) {
 241                         total_nr_samples += nr_samples;
 242                         hists__collapse_resort(hists, NULL);
 243                         /* Don't sort callchain */
 244                         perf_evsel__reset_sample_bit(pos, CALLCHAIN);
 245                         hists__output_resort(hists, NULL);
 246
 247                         if (symbol_conf.event_group &&
 248                             !perf_evsel__is_group_leader(pos))
 249                                 continue;
 250
 251                         hists__find_annotations(hists, pos, ann);
 252                 }
 253         }
 254
 255         if (total_nr_samples == 0) {
 256                 ui__error("The %s file has no samples!\n", session->file->path);
 257                 goto out;
 258         }
 259
 260         if (use_browser == 2) {
 261                 void (*show_annotations)(void);
 262
 263                 show_annotations = dlsym(perf_gtk_handle,
 264                                          "perf_gtk__show_annotations");
 265                 if (show_annotations == NULL) {
 266                         ui__error("GTK browser not found!\n");
 267                         goto out;
 268                 }
 269                 show_annotations();
 270         }
 271
 272 out:
 273         return ret;
 274 }
 275
 276 static const char * const annotate_usage[] = {
 277         "perf annotate [<options>]",
 278         NULL
 279 };
 280
 281 int cmd_annotate(int argc, const char **argv, const char *prefix __maybe_unused)
 282 {
 283         struct perf_annotate annotate = {
 284                 .tool = {
 285                         .sample = process_sample_event,
 286                         .mmap   = perf_event__process_mmap,
 287                         .mmap2  = perf_event__process_mmap2,
 288                         .comm   = perf_event__process_comm,
 289                         .exit   = perf_event__process_exit,
 290                         .fork   = perf_event__process_fork,
 291                         .ordered_events = true,
 292                         .ordering_requires_timestamps = true,
 293                 },
 294         };
 295         struct perf_data_file file = {
 296                 .mode  = PERF_DATA_MODE_READ,
 297         };
 298         const struct option options[] = {
 299         OPT_STRING('i', "input", &input_name, "file",
 300                     "input file name"),
 301         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
 302                    "only consider symbols in these dsos"),
 303         OPT_STRING('s', "symbol", &annotate.sym_hist_filter, "symbol",
 304                     "symbol to annotate"),
 305         OPT_BOOLEAN('f', "force", &file.force, "don't complain, do it"),
 306         OPT_INCR('v', "verbose", &verbose,
 307                     "be more verbose (show symbol address, etc)"),
 308         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
 309                     "dump raw trace in ASCII"),
 310         OPT_BOOLEAN(0, "gtk", &annotate.use_gtk, "Use the GTK interface"),
 311         OPT_BOOLEAN(0, "tui", &annotate.use_tui, "Use the TUI interface"),
 312         OPT_BOOLEAN(0, "stdio", &annotate.use_stdio, "Use the stdio interface"),
 313         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 314                    "file", "vmlinux pathname"),
 315         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
 316                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
 317         OPT_BOOLEAN('l', "print-line", &annotate.print_line,
 318                     "print matching source lines (may be slow)"),
 319         OPT_BOOLEAN('P', "full-paths", &annotate.full_paths,
 320                     "Don't shorten the displayed pathnames"),
 321         OPT_BOOLEAN(0, "skip-missing", &annotate.skip_missing,
 322                     "Skip symbols that cannot be annotated"),
 323         OPT_STRING('C', "cpu", &annotate.cpu_list, "cpu", "list of cpus to profile"),
 324         OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
 325                    "Look for files with symbols relative to this directory"),
 326         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
 327                     "Interleave source code with assembly code (default)"),
 328         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
 329                     "Display raw encoding of assembly instructions (default)"),
 330         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
 331                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
 332         OPT_STRING(0, "objdump", &objdump_path, "path",
 333                    "objdump binary to use for disassembly and annotations"),
 334         OPT_BOOLEAN(0, "group", &symbol_conf.event_group,
 335                     "Show event group information together"),
 336         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
 337                     "Show a column with the sum of periods"),
 338         OPT_END()
 339         };
 340         int ret = hists__init();
 341
 342         if (ret < 0)
 343                 return ret;
 344
 345         argc = parse_options(argc, argv, options, annotate_usage, 0);
 346
 347         if (annotate.use_stdio)
 348                 use_browser = 0;
 349         else if (annotate.use_tui)
 350                 use_browser = 1;
 351         else if (annotate.use_gtk)
 352                 use_browser = 2;
 353
 354         file.path  = input_name;
 355
 356         setup_browser(true);
 357
 358         annotate.session = perf_session__new(&file, false, &annotate.tool);
 359         if (annotate.session == NULL)
 360                 return -1;
 361
 362         symbol_conf.priv_size = sizeof(struct annotation);
 363         symbol_conf.try_vmlinux_path = true;
 364
 365         ret = symbol__init(&annotate.session->header.env);
 366         if (ret < 0)
 367                 goto out_delete;
 368
 369         if (setup_sorting() < 0)
 370                 usage_with_options(annotate_usage, options);
 371
 372         if (argc) {
 373                 /*
 374                  * Special case: if there's an argument left then assume that
 375                  * it's a symbol filter:
 376                  */
 377                 if (argc > 1)
 378                         usage_with_options(annotate_usage, options);
 379
 380                 annotate.sym_hist_filter = argv[0];
 381         }
 382
 383         ret = __cmd_annotate(&annotate);
 384
 385 out_delete:
 386         /*
 387          * Speed up the exit process, for large files this can
 388          * take quite a while.
 389          *
 390          * XXX Enable this when using valgrind or if we ever
 391          * librarize this command.
 392          *
 393          * Also experiment with obstacks to see how much speed
 394          * up we'll get here.
 395          *
 396          * perf_session__delete(session);
 397          */
 398         return ret;
 399 }