tools/perf/builtin-report.c

   1 /*
   2  * builtin-report.c
   3  *
   4  * Builtin report 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
  12 #include "util/annotate.h"
  13 #include "util/color.h"
  14 #include <linux/list.h>
  15 #include "util/cache.h"
  16 #include <linux/rbtree.h>
  17 #include "util/symbol.h"
  18 #include "util/callchain.h"
  19 #include "util/strlist.h"
  20 #include "util/values.h"
  21
  22 #include "perf.h"
  23 #include "util/debug.h"
  24 #include "util/evlist.h"
  25 #include "util/evsel.h"
  26 #include "util/header.h"
  27 #include "util/session.h"
  28
  29 #include "util/parse-options.h"
  30 #include "util/parse-events.h"
  31
  32 #include "util/thread.h"
  33 #include "util/sort.h"
  34 #include "util/hist.h"
  35
  36 static char             const *input_name = "perf.data";
  37
  38 static bool             force, use_tui, use_stdio;
  39 static bool             hide_unresolved;
  40 static bool             dont_use_callchains;
  41
  42 static bool             show_threads;
  43 static struct perf_read_values  show_threads_values;
  44
  45 static const char       default_pretty_printing_style[] = "normal";
  46 static const char       *pretty_printing_style = default_pretty_printing_style;
  47
  48 static char             callchain_default_opt[] = "fractal,0.5";
  49 static symbol_filter_t  annotate_init;
  50
  51 static int perf_session__add_hist_entry(struct perf_session *session,
  52                                         struct addr_location *al,
  53                                         struct perf_sample *sample,
  54                                         struct perf_evsel *evsel)
  55 {
  56         struct symbol *parent = NULL;
  57         int err = 0;
  58         struct hist_entry *he;
  59
  60         if ((sort__has_parent || symbol_conf.use_callchain) && sample->callchain) {
  61                 err = perf_session__resolve_callchain(session, al->thread,
  62                                                       sample->callchain, &parent);
  63                 if (err)
  64                         return err;
  65         }
  66
  67         he = __hists__add_entry(&evsel->hists, al, parent, sample->period);
  68         if (he == NULL)
  69                 return -ENOMEM;
  70
  71         if (symbol_conf.use_callchain) {
  72                 err = callchain_append(he->callchain, &session->callchain_cursor,
  73                                        sample->period);
  74                 if (err)
  75                         return err;
  76         }
  77         /*
  78          * Only in the newt browser we are doing integrated annotation,
  79          * so we don't allocated the extra space needed because the stdio
  80          * code will not use it.
  81          */
  82         if (al->sym != NULL && use_browser > 0) {
  83                 struct annotation *notes = symbol__annotation(he->ms.sym);
  84
  85                 assert(evsel != NULL);
  86
  87                 err = -ENOMEM;
  88                 if (notes->src == NULL &&
  89                     symbol__alloc_hist(he->ms.sym, session->evlist->nr_entries) < 0)
  90                         goto out;
  91
  92                 err = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
  93         }
  94
  95         evsel->hists.stats.total_period += sample->period;
  96         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
  97 out:
  98         return err;
  99 }
 100
 101
 102 static int process_sample_event(union perf_event *event,
 103                                 struct perf_sample *sample,
 104                                 struct perf_evsel *evsel,
 105                                 struct perf_session *session)
 106 {
 107         struct addr_location al;
 108
 109         if (perf_event__preprocess_sample(event, session, &al, sample,
 110                                           annotate_init) < 0) {
 111                 fprintf(stderr, "problem processing %d event, skipping it.\n",
 112                         event->header.type);
 113                 return -1;
 114         }
 115
 116         if (al.filtered || (hide_unresolved && al.sym == NULL))
 117                 return 0;
 118
 119         if (al.map != NULL)
 120                 al.map->dso->hit = 1;
 121
 122         if (perf_session__add_hist_entry(session, &al, sample, evsel)) {
 123                 pr_debug("problem incrementing symbol period, skipping event\n");
 124                 return -1;
 125         }
 126
 127         return 0;
 128 }
 129
 130 static int process_read_event(union perf_event *event,
 131                               struct perf_sample *sample __used,
 132                               struct perf_session *session)
 133 {
 134         struct perf_evsel *evsel = perf_evlist__id2evsel(session->evlist,
 135                                                          event->read.id);
 136         if (show_threads) {
 137                 const char *name = evsel ? event_name(evsel) : "unknown";
 138                 perf_read_values_add_value(&show_threads_values,
 139                                            event->read.pid, event->read.tid,
 140                                            event->read.id,
 141                                            name,
 142                                            event->read.value);
 143         }
 144
 145         dump_printf(": %d %d %s %" PRIu64 "\n", event->read.pid, event->read.tid,
 146                     evsel ? event_name(evsel) : "FAIL",
 147                     event->read.value);
 148
 149         return 0;
 150 }
 151
 152 static int perf_session__setup_sample_type(struct perf_session *self)
 153 {
 154         if (!(self->sample_type & PERF_SAMPLE_CALLCHAIN)) {
 155                 if (sort__has_parent) {
 156                         fprintf(stderr, "selected --sort parent, but no"
 157                                         " callchain data. Did you call"
 158                                         " perf record without -g?\n");
 159                         return -EINVAL;
 160                 }
 161                 if (symbol_conf.use_callchain) {
 162                         fprintf(stderr, "selected -g but no callchain data."
 163                                         " Did you call perf record without"
 164                                         " -g?\n");
 165                         return -1;
 166                 }
 167         } else if (!dont_use_callchains && callchain_param.mode != CHAIN_NONE &&
 168                    !symbol_conf.use_callchain) {
 169                         symbol_conf.use_callchain = true;
 170                         if (callchain_register_param(&callchain_param) < 0) {
 171                                 fprintf(stderr, "Can't register callchain"
 172                                                 " params\n");
 173                                 return -EINVAL;
 174                         }
 175         }
 176
 177         return 0;
 178 }
 179
 180 static struct perf_event_ops event_ops = {
 181         .sample          = process_sample_event,
 182         .mmap            = perf_event__process_mmap,
 183         .comm            = perf_event__process_comm,
 184         .exit            = perf_event__process_task,
 185         .fork            = perf_event__process_task,
 186         .lost            = perf_event__process_lost,
 187         .read            = process_read_event,
 188         .attr            = perf_event__process_attr,
 189         .event_type      = perf_event__process_event_type,
 190         .tracing_data    = perf_event__process_tracing_data,
 191         .build_id        = perf_event__process_build_id,
 192         .ordered_samples = true,
 193         .ordering_requires_timestamps = true,
 194 };
 195
 196 extern volatile int session_done;
 197
 198 static void sig_handler(int sig __used)
 199 {
 200         session_done = 1;
 201 }
 202
 203 static size_t hists__fprintf_nr_sample_events(struct hists *self,
 204                                               const char *evname, FILE *fp)
 205 {
 206         size_t ret;
 207         char unit;
 208         unsigned long nr_events = self->stats.nr_events[PERF_RECORD_SAMPLE];
 209
 210         nr_events = convert_unit(nr_events, &unit);
 211         ret = fprintf(fp, "# Events: %lu%c", nr_events, unit);
 212         if (evname != NULL)
 213                 ret += fprintf(fp, " %s", evname);
 214         return ret + fprintf(fp, "\n#\n");
 215 }
 216
 217 static int perf_evlist__tty_browse_hists(struct perf_evlist *evlist,
 218                                          const char *help)
 219 {
 220         struct perf_evsel *pos;
 221
 222         list_for_each_entry(pos, &evlist->entries, node) {
 223                 struct hists *hists = &pos->hists;
 224                 const char *evname = NULL;
 225
 226                 if (rb_first(&hists->entries) != rb_last(&hists->entries))
 227                         evname = event_name(pos);
 228
 229                 hists__fprintf_nr_sample_events(hists, evname, stdout);
 230                 hists__fprintf(hists, NULL, false, stdout);
 231                 fprintf(stdout, "\n\n");
 232         }
 233
 234         if (sort_order == default_sort_order &&
 235             parent_pattern == default_parent_pattern) {
 236                 fprintf(stdout, "#\n# (%s)\n#\n", help);
 237
 238                 if (show_threads) {
 239                         bool style = !strcmp(pretty_printing_style, "raw");
 240                         perf_read_values_display(stdout, &show_threads_values,
 241                                                  style);
 242                         perf_read_values_destroy(&show_threads_values);
 243                 }
 244         }
 245
 246         return 0;
 247 }
 248
 249 static int __cmd_report(void)
 250 {
 251         int ret = -EINVAL;
 252         u64 nr_samples;
 253         struct perf_session *session;
 254         struct perf_evsel *pos;
 255         struct map *kernel_map;
 256         struct kmap *kernel_kmap;
 257         const char *help = "For a higher level overview, try: perf report --sort comm,dso";
 258
 259         signal(SIGINT, sig_handler);
 260
 261         session = perf_session__new(input_name, O_RDONLY, force, false, &event_ops);
 262         if (session == NULL)
 263                 return -ENOMEM;
 264
 265         if (show_threads)
 266                 perf_read_values_init(&show_threads_values);
 267
 268         ret = perf_session__setup_sample_type(session);
 269         if (ret)
 270                 goto out_delete;
 271
 272         ret = perf_session__process_events(session, &event_ops);
 273         if (ret)
 274                 goto out_delete;
 275
 276         kernel_map = session->host_machine.vmlinux_maps[MAP__FUNCTION];
 277         kernel_kmap = map__kmap(kernel_map);
 278         if (kernel_map == NULL ||
 279             (kernel_map->dso->hit &&
 280              (kernel_kmap->ref_reloc_sym == NULL ||
 281               kernel_kmap->ref_reloc_sym->addr == 0))) {
 282                 const struct dso *kdso = kernel_map->dso;
 283
 284                 ui__warning(
 285 "Kernel address maps (/proc/{kallsyms,modules}) were restricted.\n\n"
 286 "Check /proc/sys/kernel/kptr_restrict before running 'perf record'.\n\n%s\n\n"
 287 "Samples in kernel modules can't be resolved as well.\n\n",
 288                             RB_EMPTY_ROOT(&kdso->symbols[MAP__FUNCTION]) ?
 289 "As no suitable kallsyms nor vmlinux was found, kernel samples\n"
 290 "can't be resolved." :
 291 "If some relocation was applied (e.g. kexec) symbols may be misresolved.");
 292         }
 293
 294         if (dump_trace) {
 295                 perf_session__fprintf_nr_events(session, stdout);
 296                 goto out_delete;
 297         }
 298
 299         if (verbose > 3)
 300                 perf_session__fprintf(session, stdout);
 301
 302         if (verbose > 2)
 303                 perf_session__fprintf_dsos(session, stdout);
 304
 305         nr_samples = 0;
 306         list_for_each_entry(pos, &session->evlist->entries, node) {
 307                 struct hists *hists = &pos->hists;
 308
 309                 hists__collapse_resort(hists);
 310                 hists__output_resort(hists);
 311                 nr_samples += hists->stats.nr_events[PERF_RECORD_SAMPLE];
 312         }
 313
 314         if (nr_samples == 0) {
 315                 ui__warning("The %s file has no samples!\n", input_name);
 316                 goto out_delete;
 317         }
 318
 319         if (use_browser > 0)
 320                 perf_evlist__tui_browse_hists(session->evlist, help);
 321         else
 322                 perf_evlist__tty_browse_hists(session->evlist, help);
 323
 324 out_delete:
 325         /*
 326          * Speed up the exit process, for large files this can
 327          * take quite a while.
 328          *
 329          * XXX Enable this when using valgrind or if we ever
 330          * librarize this command.
 331          *
 332          * Also experiment with obstacks to see how much speed
 333          * up we'll get here.
 334          *
 335          * perf_session__delete(session);
 336          */
 337         return ret;
 338 }
 339
 340 static int
 341 parse_callchain_opt(const struct option *opt __used, const char *arg,
 342                     int unset)
 343 {
 344         char *tok, *tok2;
 345         char *endptr;
 346
 347         /*
 348          * --no-call-graph
 349          */
 350         if (unset) {
 351                 dont_use_callchains = true;
 352                 return 0;
 353         }
 354
 355         symbol_conf.use_callchain = true;
 356
 357         if (!arg)
 358                 return 0;
 359
 360         tok = strtok((char *)arg, ",");
 361         if (!tok)
 362                 return -1;
 363
 364         /* get the output mode */
 365         if (!strncmp(tok, "graph", strlen(arg)))
 366                 callchain_param.mode = CHAIN_GRAPH_ABS;
 367
 368         else if (!strncmp(tok, "flat", strlen(arg)))
 369                 callchain_param.mode = CHAIN_FLAT;
 370
 371         else if (!strncmp(tok, "fractal", strlen(arg)))
 372                 callchain_param.mode = CHAIN_GRAPH_REL;
 373
 374         else if (!strncmp(tok, "none", strlen(arg))) {
 375                 callchain_param.mode = CHAIN_NONE;
 376                 symbol_conf.use_callchain = false;
 377
 378                 return 0;
 379         }
 380
 381         else
 382                 return -1;
 383
 384         /* get the min percentage */
 385         tok = strtok(NULL, ",");
 386         if (!tok)
 387                 goto setup;
 388
 389         tok2 = strtok(NULL, ",");
 390         callchain_param.min_percent = strtod(tok, &endptr);
 391         if (tok == endptr)
 392                 return -1;
 393
 394         if (tok2)
 395                 callchain_param.print_limit = strtod(tok2, &endptr);
 396 setup:
 397         if (callchain_register_param(&callchain_param) < 0) {
 398                 fprintf(stderr, "Can't register callchain params\n");
 399                 return -1;
 400         }
 401         return 0;
 402 }
 403
 404 static const char * const report_usage[] = {
 405         "perf report [<options>] <command>",
 406         NULL
 407 };
 408
 409 static const struct option options[] = {
 410         OPT_STRING('i', "input", &input_name, "file",
 411                     "input file name"),
 412         OPT_INCR('v', "verbose", &verbose,
 413                     "be more verbose (show symbol address, etc)"),
 414         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
 415                     "dump raw trace in ASCII"),
 416         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 417                    "file", "vmlinux pathname"),
 418         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
 419                    "file", "kallsyms pathname"),
 420         OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
 421         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
 422                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
 423         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
 424                     "Show a column with the number of samples"),
 425         OPT_BOOLEAN('T', "threads", &show_threads,
 426                     "Show per-thread event counters"),
 427         OPT_STRING(0, "pretty", &pretty_printing_style, "key",
 428                    "pretty printing style key: normal raw"),
 429         OPT_BOOLEAN(0, "tui", &use_tui, "Use the TUI interface"),
 430         OPT_BOOLEAN(0, "stdio", &use_stdio, "Use the stdio interface"),
 431         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
 432                    "sort by key(s): pid, comm, dso, symbol, parent"),
 433         OPT_BOOLEAN(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
 434                     "Show sample percentage for different cpu modes"),
 435         OPT_STRING('p', "parent", &parent_pattern, "regex",
 436                    "regex filter to identify parent, see: '--sort parent'"),
 437         OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
 438                     "Only display entries with parent-match"),
 439         OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent",
 440                      "Display callchains using output_type (graph, flat, fractal, or none) and min percent threshold. "
 441                      "Default: fractal,0.5", &parse_callchain_opt, callchain_default_opt),
 442         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
 443                    "only consider symbols in these dsos"),
 444         OPT_STRING('C', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
 445                    "only consider symbols in these comms"),
 446         OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
 447                    "only consider these symbols"),
 448         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
 449                    "width[,width...]",
 450                    "don't try to adjust column width, use these fixed values"),
 451         OPT_STRING('t', "field-separator", &symbol_conf.field_sep, "separator",
 452                    "separator for columns, no spaces will be added between "
 453                    "columns '.' is reserved."),
 454         OPT_BOOLEAN('U', "hide-unresolved", &hide_unresolved,
 455                     "Only display entries resolved to a symbol"),
 456         OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
 457                     "Look for files with symbols relative to this directory"),
 458         OPT_END()
 459 };
 460
 461 int cmd_report(int argc, const char **argv, const char *prefix __used)
 462 {
 463         argc = parse_options(argc, argv, options, report_usage, 0);
 464
 465         if (use_stdio)
 466                 use_browser = 0;
 467         else if (use_tui)
 468                 use_browser = 1;
 469
 470         if (strcmp(input_name, "-") != 0)
 471                 setup_browser(true);
 472         else
 473                 use_browser = 0;
 474         /*
 475          * Only in the newt browser we are doing integrated annotation,
 476          * so don't allocate extra space that won't be used in the stdio
 477          * implementation.
 478          */
 479         if (use_browser > 0) {
 480                 symbol_conf.priv_size = sizeof(struct annotation);
 481                 annotate_init         = symbol__annotate_init;
 482                 /*
 483                  * For searching by name on the "Browse map details".
 484                  * providing it only in verbose mode not to bloat too
 485                  * much struct symbol.
 486                  */
 487                 if (verbose) {
 488                         /*
 489                          * XXX: Need to provide a less kludgy way to ask for
 490                          * more space per symbol, the u32 is for the index on
 491                          * the ui browser.
 492                          * See symbol__browser_index.
 493                          */
 494                         symbol_conf.priv_size += sizeof(u32);
 495                         symbol_conf.sort_by_name = true;
 496                 }
 497         }
 498
 499         if (symbol__init() < 0)
 500                 return -1;
 501
 502         setup_sorting(report_usage, options);
 503
 504         if (parent_pattern != default_parent_pattern) {
 505                 if (sort_dimension__add("parent") < 0)
 506                         return -1;
 507                 sort_parent.elide = 1;
 508         } else
 509                 symbol_conf.exclude_other = false;
 510
 511         /*
 512          * Any (unrecognized) arguments left?
 513          */
 514         if (argc)
 515                 usage_with_options(report_usage, options);
 516
 517         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
 518         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
 519         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
 520
 521         return __cmd_report();
 522 }