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
  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,
  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         sample->period = 1;
  76         sample->weight = 1;
  77
  78         he = __hists__add_entry(hists, al, NULL, NULL, NULL, sample, true);
  79         if (he == NULL)
  80                 return -ENOMEM;
  81
  82         ret = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
  83         hists__inc_nr_samples(hists, true);
  84         return ret;
  85 }
  86
  87 static int process_sample_event(struct perf_tool *tool,
  88                                 union perf_event *event,
  89                                 struct perf_sample *sample,
  90                                 struct perf_evsel *evsel,
  91                                 struct machine *machine)
  92 {
  93         struct perf_annotate *ann = container_of(tool, struct perf_annotate, tool);
  94         struct addr_location al;
  95         int ret = 0;
  96
  97         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0) {
  98                 pr_warning("problem processing %d event, skipping it.\n",
  99                            event->header.type);
 100                 return -1;
 101         }
 102
 103         if (ann->cpu_list && !test_bit(sample->cpu, ann->cpu_bitmap))
 104                 goto out_put;
 105
 106         if (!al.filtered && perf_evsel__add_sample(evsel, sample, &al, ann)) {
 107                 pr_warning("problem incrementing symbol count, "
 108                            "skipping event\n");
 109                 ret = -1;
 110         }
 111 out_put:
 112         addr_location__put(&al);
 113         return ret;
 114 }
 115
 116 static int hist_entry__tty_annotate(struct hist_entry *he,
 117                                     struct perf_evsel *evsel,
 118                                     struct perf_annotate *ann)
 119 {
 120         return symbol__tty_annotate(he->ms.sym, he->ms.map, evsel,
 121                                     ann->print_line, ann->full_paths, 0, 0);
 122 }
 123
 124 static void hists__find_annotations(struct hists *hists,
 125                                     struct perf_evsel *evsel,
 126                                     struct perf_annotate *ann)
 127 {
 128         struct rb_node *nd = rb_first(&hists->entries), *next;
 129         int key = K_RIGHT;
 130
 131         while (nd) {
 132                 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
 133                 struct annotation *notes;
 134
 135                 if (he->ms.sym == NULL || he->ms.map->dso->annotate_warned)
 136                         goto find_next;
 137
 138                 notes = symbol__annotation(he->ms.sym);
 139                 if (notes->src == NULL) {
 140 find_next:
 141                         if (key == K_LEFT)
 142                                 nd = rb_prev(nd);
 143                         else
 144                                 nd = rb_next(nd);
 145                         continue;
 146                 }
 147
 148                 if (use_browser == 2) {
 149                         int ret;
 150                         int (*annotate)(struct hist_entry *he,
 151                                         struct perf_evsel *evsel,
 152                                         struct hist_browser_timer *hbt);
 153
 154                         annotate = dlsym(perf_gtk_handle,
 155                                          "hist_entry__gtk_annotate");
 156                         if (annotate == NULL) {
 157                                 ui__error("GTK browser not found!\n");
 158                                 return;
 159                         }
 160
 161                         ret = annotate(he, evsel, NULL);
 162                         if (!ret || !ann->skip_missing)
 163                                 return;
 164
 165                         /* skip missing symbols */
 166                         nd = rb_next(nd);
 167                 } else if (use_browser == 1) {
 168                         key = hist_entry__tui_annotate(he, evsel, NULL);
 169                         switch (key) {
 170                         case -1:
 171                                 if (!ann->skip_missing)
 172                                         return;
 173                                 /* fall through */
 174                         case K_RIGHT:
 175                                 next = rb_next(nd);
 176                                 break;
 177                         case K_LEFT:
 178                                 next = rb_prev(nd);
 179                                 break;
 180                         default:
 181                                 return;
 182                         }
 183
 184                         if (next != NULL)
 185                                 nd = next;
 186                 } else {
 187                         hist_entry__tty_annotate(he, evsel, ann);
 188                         nd = rb_next(nd);
 189                         /*
 190                          * Since we have a hist_entry per IP for the same
 191                          * symbol, free he->ms.sym->src to signal we already
 192                          * processed this symbol.
 193                          */
 194                         zfree(&notes->src->cycles_hist);
 195                         zfree(&notes->src);
 196                 }
 197         }
 198 }
 199
 200 static int __cmd_annotate(struct perf_annotate *ann)
 201 {
 202         int ret;
 203         struct perf_session *session = ann->session;
 204         struct perf_evsel *pos;
 205         u64 total_nr_samples;
 206
 207         machines__set_symbol_filter(&session->machines, symbol__annotate_init);
 208
 209         if (ann->cpu_list) {
 210                 ret = perf_session__cpu_bitmap(session, ann->cpu_list,
 211                                                ann->cpu_bitmap);
 212                 if (ret)
 213                         goto out;
 214         }
 215
 216         if (!objdump_path) {
 217                 ret = perf_env__lookup_objdump(&session->header.env);
 218                 if (ret)
 219                         goto out;
 220         }
 221
 222         ret = perf_session__process_events(session);
 223         if (ret)
 224                 goto out;
 225
 226         if (dump_trace) {
 227                 perf_session__fprintf_nr_events(session, stdout);
 228                 perf_evlist__fprintf_nr_events(session->evlist, stdout);
 229                 goto out;
 230         }
 231
 232         if (verbose > 3)
 233                 perf_session__fprintf(session, stdout);
 234
 235         if (verbose > 2)
 236                 perf_session__fprintf_dsos(session, stdout);
 237
 238         total_nr_samples = 0;
 239         evlist__for_each(session->evlist, pos) {
 240                 struct hists *hists = evsel__hists(pos);
 241                 u32 nr_samples = hists->stats.nr_events[PERF_RECORD_SAMPLE];
 242
 243                 if (nr_samples > 0) {
 244                         total_nr_samples += nr_samples;
 245                         hists__collapse_resort(hists, NULL);
 246                         /* Don't sort callchain */
 247                         perf_evsel__reset_sample_bit(pos, CALLCHAIN);
 248                         hists__output_resort(hists, NULL);
 249
 250                         if (symbol_conf.event_group &&
 251                             !perf_evsel__is_group_leader(pos))
 252                                 continue;
 253
 254                         hists__find_annotations(hists, pos, ann);
 255                 }
 256         }
 257
 258         if (total_nr_samples == 0) {
 259                 ui__error("The %s file has no samples!\n", session->file->path);
 260                 goto out;
 261         }
 262
 263         if (use_browser == 2) {
 264                 void (*show_annotations)(void);
 265
 266                 show_annotations = dlsym(perf_gtk_handle,
 267                                          "perf_gtk__show_annotations");
 268                 if (show_annotations == NULL) {
 269                         ui__error("GTK browser not found!\n");
 270                         goto out;
 271                 }
 272                 show_annotations();
 273         }
 274
 275 out:
 276         return ret;
 277 }
 278
 279 static const char * const annotate_usage[] = {
 280         "perf annotate [<options>]",
 281         NULL
 282 };
 283
 284 int cmd_annotate(int argc, const char **argv, const char *prefix __maybe_unused)
 285 {
 286         struct perf_annotate annotate = {
 287                 .tool = {
 288                         .sample = process_sample_event,
 289                         .mmap   = perf_event__process_mmap,
 290                         .mmap2  = perf_event__process_mmap2,
 291                         .comm   = perf_event__process_comm,
 292                         .exit   = perf_event__process_exit,
 293                         .fork   = perf_event__process_fork,
 294                         .ordered_events = true,
 295                         .ordering_requires_timestamps = true,
 296                 },
 297         };
 298         struct perf_data_file file = {
 299                 .mode  = PERF_DATA_MODE_READ,
 300         };
 301         const struct option options[] = {
 302         OPT_STRING('i', "input", &input_name, "file",
 303                     "input file name"),
 304         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
 305                    "only consider symbols in these dsos"),
 306         OPT_STRING('s', "symbol", &annotate.sym_hist_filter, "symbol",
 307                     "symbol to annotate"),
 308         OPT_BOOLEAN('f', "force", &file.force, "don't complain, do it"),
 309         OPT_INCR('v', "verbose", &verbose,
 310                     "be more verbose (show symbol address, etc)"),
 311         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
 312                     "dump raw trace in ASCII"),
 313         OPT_BOOLEAN(0, "gtk", &annotate.use_gtk, "Use the GTK interface"),
 314         OPT_BOOLEAN(0, "tui", &annotate.use_tui, "Use the TUI interface"),
 315         OPT_BOOLEAN(0, "stdio", &annotate.use_stdio, "Use the stdio interface"),
 316         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 317                    "file", "vmlinux pathname"),
 318         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
 319                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
 320         OPT_BOOLEAN('l', "print-line", &annotate.print_line,
 321                     "print matching source lines (may be slow)"),
 322         OPT_BOOLEAN('P', "full-paths", &annotate.full_paths,
 323                     "Don't shorten the displayed pathnames"),
 324         OPT_BOOLEAN(0, "skip-missing", &annotate.skip_missing,
 325                     "Skip symbols that cannot be annotated"),
 326         OPT_STRING('C', "cpu", &annotate.cpu_list, "cpu", "list of cpus to profile"),
 327         OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
 328                    "Look for files with symbols relative to this directory"),
 329         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
 330                     "Interleave source code with assembly code (default)"),
 331         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
 332                     "Display raw encoding of assembly instructions (default)"),
 333         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
 334                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
 335         OPT_STRING(0, "objdump", &objdump_path, "path",
 336                    "objdump binary to use for disassembly and annotations"),
 337         OPT_BOOLEAN(0, "group", &symbol_conf.event_group,
 338                     "Show event group information together"),
 339         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
 340                     "Show a column with the sum of periods"),
 341         OPT_END()
 342         };
 343         int ret = hists__init();
 344
 345         if (ret < 0)
 346                 return ret;
 347
 348         argc = parse_options(argc, argv, options, annotate_usage, 0);
 349         if (argc) {
 350                 /*
 351                  * Special case: if there's an argument left then assume that
 352                  * it's a symbol filter:
 353                  */
 354                 if (argc > 1)
 355                         usage_with_options(annotate_usage, options);
 356
 357                 annotate.sym_hist_filter = argv[0];
 358         }
 359
 360         file.path  = input_name;
 361
 362         annotate.session = perf_session__new(&file, false, &annotate.tool);
 363         if (annotate.session == NULL)
 364                 return -1;
 365
 366         symbol_conf.priv_size = sizeof(struct annotation);
 367         symbol_conf.try_vmlinux_path = true;
 368
 369         ret = symbol__init(&annotate.session->header.env);
 370         if (ret < 0)
 371                 goto out_delete;
 372
 373         if (setup_sorting(NULL) < 0)
 374                 usage_with_options(annotate_usage, options);
 375
 376         if (annotate.use_stdio)
 377                 use_browser = 0;
 378         else if (annotate.use_tui)
 379                 use_browser = 1;
 380         else if (annotate.use_gtk)
 381                 use_browser = 2;
 382
 383         setup_browser(true);
 384
 385         ret = __cmd_annotate(&annotate);
 386
 387 out_delete:
 388         /*
 389          * Speed up the exit process, for large files this can
 390          * take quite a while.
 391          *
 392          * XXX Enable this when using valgrind or if we ever
 393          * librarize this command.
 394          *
 395          * Also experiment with obstacks to see how much speed
 396          * up we'll get here.
 397          *
 398          * perf_session__delete(session);
 399          */
 400         return ret;
 401 }