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/header.h"
  25 #include "util/session.h"
  26
  27 #include "util/parse-options.h"
  28 #include "util/parse-events.h"
  29
  30 #include "util/thread.h"
  31 #include "util/sort.h"
  32 #include "util/hist.h"
  33
  34 static char             const *input_name = "perf.data";
  35
  36 static bool             force, use_tui, use_stdio;
  37 static bool             hide_unresolved;
  38 static bool             dont_use_callchains;
  39
  40 static bool             show_threads;
  41 static struct perf_read_values  show_threads_values;
  42
  43 static const char       default_pretty_printing_style[] = "normal";
  44 static const char       *pretty_printing_style = default_pretty_printing_style;
  45
  46 static char             callchain_default_opt[] = "fractal,0.5";
  47 static symbol_filter_t  annotate_init;
  48
  49 static struct hists *perf_session__hists_findnew(struct perf_session *self,
  50                                                  u64 event_stream, u32 type,
  51                                                  u64 config)
  52 {
  53         struct rb_node **p = &self->hists_tree.rb_node;
  54         struct rb_node *parent = NULL;
  55         struct hists *iter, *new;
  56
  57         while (*p != NULL) {
  58                 parent = *p;
  59                 iter = rb_entry(parent, struct hists, rb_node);
  60                 if (iter->config == config)
  61                         return iter;
  62
  63
  64                 if (config > iter->config)
  65                         p = &(*p)->rb_right;
  66                 else
  67                         p = &(*p)->rb_left;
  68         }
  69
  70         new = malloc(sizeof(struct hists));
  71         if (new == NULL)
  72                 return NULL;
  73         memset(new, 0, sizeof(struct hists));
  74         new->event_stream = event_stream;
  75         new->config = config;
  76         new->type = type;
  77         rb_link_node(&new->rb_node, parent, p);
  78         rb_insert_color(&new->rb_node, &self->hists_tree);
  79         return new;
  80 }
  81
  82 static int perf_session__add_hist_entry(struct perf_session *session,
  83                                         struct addr_location *al,
  84                                         struct perf_sample *sample)
  85 {
  86         struct symbol *parent = NULL;
  87         int err = 0;
  88         struct hist_entry *he;
  89         struct hists *hists;
  90         struct perf_event_attr *attr;
  91
  92         if ((sort__has_parent || symbol_conf.use_callchain) && sample->callchain) {
  93                 err = perf_session__resolve_callchain(session, al->thread,
  94                                                       sample->callchain, &parent);
  95                 if (err)
  96                         return err;
  97         }
  98
  99         attr = perf_header__find_attr(sample->id, &session->header);
 100         if (attr)
 101                 hists = perf_session__hists_findnew(session, sample->id, attr->type, attr->config);
 102         else
 103                 hists = perf_session__hists_findnew(session, sample->id, 0, 0);
 104         if (hists == NULL)
 105                 return -ENOMEM;
 106
 107         he = __hists__add_entry(hists, al, parent, sample->period);
 108         if (he == NULL)
 109                 return -ENOMEM;
 110
 111         if (symbol_conf.use_callchain) {
 112                 err = callchain_append(he->callchain, &session->callchain_cursor,
 113                                        sample->period);
 114                 if (err)
 115                         return err;
 116         }
 117         /*
 118          * Only in the newt browser we are doing integrated annotation,
 119          * so we don't allocated the extra space needed because the stdio
 120          * code will not use it.
 121          */
 122         if (al->sym != NULL && use_browser > 0) {
 123                 /*
 124                  * All aggregated on the first sym_hist.
 125                  */
 126                 struct annotation *notes = symbol__annotation(he->ms.sym);
 127                 if (notes->src == NULL &&
 128                     symbol__alloc_hist(he->ms.sym, 1) < 0)
 129                         err = -ENOMEM;
 130                 else
 131                         err = hist_entry__inc_addr_samples(he, 0, al->addr);
 132         }
 133
 134         return err;
 135 }
 136
 137 static int add_event_total(struct perf_session *session,
 138                            struct perf_sample *sample,
 139                            struct perf_event_attr *attr)
 140 {
 141         struct hists *hists;
 142
 143         if (attr)
 144                 hists = perf_session__hists_findnew(session, sample->id,
 145                                                     attr->type, attr->config);
 146         else
 147                 hists = perf_session__hists_findnew(session, sample->id, 0, 0);
 148
 149         if (!hists)
 150                 return -ENOMEM;
 151
 152         hists->stats.total_period += sample->period;
 153         /*
 154          * FIXME: add_event_total should be moved from here to
 155          * perf_session__process_event so that the proper hist is passed to
 156          * the event_op methods.
 157          */
 158         hists__inc_nr_events(hists, PERF_RECORD_SAMPLE);
 159         session->hists.stats.total_period += sample->period;
 160         return 0;
 161 }
 162
 163 static int process_sample_event(union perf_event *event,
 164                                 struct perf_sample *sample,
 165                                 struct perf_session *session)
 166 {
 167         struct addr_location al;
 168         struct perf_event_attr *attr;
 169
 170         if (perf_event__preprocess_sample(event, session, &al, sample,
 171                                           annotate_init) < 0) {
 172                 fprintf(stderr, "problem processing %d event, skipping it.\n",
 173                         event->header.type);
 174                 return -1;
 175         }
 176
 177         if (al.filtered || (hide_unresolved && al.sym == NULL))
 178                 return 0;
 179
 180         if (perf_session__add_hist_entry(session, &al, sample)) {
 181                 pr_debug("problem incrementing symbol period, skipping event\n");
 182                 return -1;
 183         }
 184
 185         attr = perf_header__find_attr(sample->id, &session->header);
 186
 187         if (add_event_total(session, sample, attr)) {
 188                 pr_debug("problem adding event period\n");
 189                 return -1;
 190         }
 191
 192         return 0;
 193 }
 194
 195 static int process_read_event(union perf_event *event,
 196                               struct perf_sample *sample __used,
 197                               struct perf_session *session __used)
 198 {
 199         struct perf_event_attr *attr;
 200
 201         attr = perf_header__find_attr(event->read.id, &session->header);
 202
 203         if (show_threads) {
 204                 const char *name = attr ? __event_name(attr->type, attr->config)
 205                                    : "unknown";
 206                 perf_read_values_add_value(&show_threads_values,
 207                                            event->read.pid, event->read.tid,
 208                                            event->read.id,
 209                                            name,
 210                                            event->read.value);
 211         }
 212
 213         dump_printf(": %d %d %s %" PRIu64 "\n", event->read.pid, event->read.tid,
 214                     attr ? __event_name(attr->type, attr->config) : "FAIL",
 215                     event->read.value);
 216
 217         return 0;
 218 }
 219
 220 static int perf_session__setup_sample_type(struct perf_session *self)
 221 {
 222         if (!(self->sample_type & PERF_SAMPLE_CALLCHAIN)) {
 223                 if (sort__has_parent) {
 224                         fprintf(stderr, "selected --sort parent, but no"
 225                                         " callchain data. Did you call"
 226                                         " perf record without -g?\n");
 227                         return -EINVAL;
 228                 }
 229                 if (symbol_conf.use_callchain) {
 230                         fprintf(stderr, "selected -g but no callchain data."
 231                                         " Did you call perf record without"
 232                                         " -g?\n");
 233                         return -1;
 234                 }
 235         } else if (!dont_use_callchains && callchain_param.mode != CHAIN_NONE &&
 236                    !symbol_conf.use_callchain) {
 237                         symbol_conf.use_callchain = true;
 238                         if (callchain_register_param(&callchain_param) < 0) {
 239                                 fprintf(stderr, "Can't register callchain"
 240                                                 " params\n");
 241                                 return -EINVAL;
 242                         }
 243         }
 244
 245         return 0;
 246 }
 247
 248 static struct perf_event_ops event_ops = {
 249         .sample          = process_sample_event,
 250         .mmap            = perf_event__process_mmap,
 251         .comm            = perf_event__process_comm,
 252         .exit            = perf_event__process_task,
 253         .fork            = perf_event__process_task,
 254         .lost            = perf_event__process_lost,
 255         .read            = process_read_event,
 256         .attr            = perf_event__process_attr,
 257         .event_type      = perf_event__process_event_type,
 258         .tracing_data    = perf_event__process_tracing_data,
 259         .build_id        = perf_event__process_build_id,
 260         .ordered_samples = true,
 261         .ordering_requires_timestamps = true,
 262 };
 263
 264 extern volatile int session_done;
 265
 266 static void sig_handler(int sig __used)
 267 {
 268         session_done = 1;
 269 }
 270
 271 static size_t hists__fprintf_nr_sample_events(struct hists *self,
 272                                               const char *evname, FILE *fp)
 273 {
 274         size_t ret;
 275         char unit;
 276         unsigned long nr_events = self->stats.nr_events[PERF_RECORD_SAMPLE];
 277
 278         nr_events = convert_unit(nr_events, &unit);
 279         ret = fprintf(fp, "# Events: %lu%c", nr_events, unit);
 280         if (evname != NULL)
 281                 ret += fprintf(fp, " %s", evname);
 282         return ret + fprintf(fp, "\n#\n");
 283 }
 284
 285 static int hists__tty_browse_tree(struct rb_root *tree, const char *help)
 286 {
 287         struct rb_node *next = rb_first(tree);
 288
 289         while (next) {
 290                 struct hists *hists = rb_entry(next, struct hists, rb_node);
 291                 const char *evname = NULL;
 292
 293                 if (rb_first(&hists->entries) != rb_last(&hists->entries))
 294                         evname = __event_name(hists->type, hists->config);
 295
 296                 hists__fprintf_nr_sample_events(hists, evname, stdout);
 297                 hists__fprintf(hists, NULL, false, stdout);
 298                 fprintf(stdout, "\n\n");
 299                 next = rb_next(&hists->rb_node);
 300         }
 301
 302         if (sort_order == default_sort_order &&
 303             parent_pattern == default_parent_pattern) {
 304                 fprintf(stdout, "#\n# (%s)\n#\n", help);
 305
 306                 if (show_threads) {
 307                         bool style = !strcmp(pretty_printing_style, "raw");
 308                         perf_read_values_display(stdout, &show_threads_values,
 309                                                  style);
 310                         perf_read_values_destroy(&show_threads_values);
 311                 }
 312         }
 313
 314         return 0;
 315 }
 316
 317 static int __cmd_report(void)
 318 {
 319         int ret = -EINVAL;
 320         struct perf_session *session;
 321         struct rb_node *next;
 322         const char *help = "For a higher level overview, try: perf report --sort comm,dso";
 323
 324         signal(SIGINT, sig_handler);
 325
 326         session = perf_session__new(input_name, O_RDONLY, force, false, &event_ops);
 327         if (session == NULL)
 328                 return -ENOMEM;
 329
 330         if (show_threads)
 331                 perf_read_values_init(&show_threads_values);
 332
 333         ret = perf_session__setup_sample_type(session);
 334         if (ret)
 335                 goto out_delete;
 336
 337         ret = perf_session__process_events(session, &event_ops);
 338         if (ret)
 339                 goto out_delete;
 340
 341         if (dump_trace) {
 342                 perf_session__fprintf_nr_events(session, stdout);
 343                 goto out_delete;
 344         }
 345
 346         if (verbose > 3)
 347                 perf_session__fprintf(session, stdout);
 348
 349         if (verbose > 2)
 350                 perf_session__fprintf_dsos(session, stdout);
 351
 352         next = rb_first(&session->hists_tree);
 353
 354         if (next == NULL) {
 355                 ui__warning("The %s file has no samples!\n", input_name);
 356                 goto out_delete;
 357         }
 358
 359         while (next) {
 360                 struct hists *hists;
 361
 362                 hists = rb_entry(next, struct hists, rb_node);
 363                 hists__collapse_resort(hists);
 364                 hists__output_resort(hists);
 365                 next = rb_next(&hists->rb_node);
 366         }
 367
 368         if (use_browser > 0)
 369                 hists__tui_browse_tree(&session->hists_tree, help, 0);
 370         else
 371                 hists__tty_browse_tree(&session->hists_tree, help);
 372
 373 out_delete:
 374         /*
 375          * Speed up the exit process, for large files this can
 376          * take quite a while.
 377          *
 378          * XXX Enable this when using valgrind or if we ever
 379          * librarize this command.
 380          *
 381          * Also experiment with obstacks to see how much speed
 382          * up we'll get here.
 383          *
 384          * perf_session__delete(session);
 385          */
 386         return ret;
 387 }
 388
 389 static int
 390 parse_callchain_opt(const struct option *opt __used, const char *arg,
 391                     int unset)
 392 {
 393         char *tok, *tok2;
 394         char *endptr;
 395
 396         /*
 397          * --no-call-graph
 398          */
 399         if (unset) {
 400                 dont_use_callchains = true;
 401                 return 0;
 402         }
 403
 404         symbol_conf.use_callchain = true;
 405
 406         if (!arg)
 407                 return 0;
 408
 409         tok = strtok((char *)arg, ",");
 410         if (!tok)
 411                 return -1;
 412
 413         /* get the output mode */
 414         if (!strncmp(tok, "graph", strlen(arg)))
 415                 callchain_param.mode = CHAIN_GRAPH_ABS;
 416
 417         else if (!strncmp(tok, "flat", strlen(arg)))
 418                 callchain_param.mode = CHAIN_FLAT;
 419
 420         else if (!strncmp(tok, "fractal", strlen(arg)))
 421                 callchain_param.mode = CHAIN_GRAPH_REL;
 422
 423         else if (!strncmp(tok, "none", strlen(arg))) {
 424                 callchain_param.mode = CHAIN_NONE;
 425                 symbol_conf.use_callchain = false;
 426
 427                 return 0;
 428         }
 429
 430         else
 431                 return -1;
 432
 433         /* get the min percentage */
 434         tok = strtok(NULL, ",");
 435         if (!tok)
 436                 goto setup;
 437
 438         tok2 = strtok(NULL, ",");
 439         callchain_param.min_percent = strtod(tok, &endptr);
 440         if (tok == endptr)
 441                 return -1;
 442
 443         if (tok2)
 444                 callchain_param.print_limit = strtod(tok2, &endptr);
 445 setup:
 446         if (callchain_register_param(&callchain_param) < 0) {
 447                 fprintf(stderr, "Can't register callchain params\n");
 448                 return -1;
 449         }
 450         return 0;
 451 }
 452
 453 static const char * const report_usage[] = {
 454         "perf report [<options>] <command>",
 455         NULL
 456 };
 457
 458 static const struct option options[] = {
 459         OPT_STRING('i', "input", &input_name, "file",
 460                     "input file name"),
 461         OPT_INCR('v', "verbose", &verbose,
 462                     "be more verbose (show symbol address, etc)"),
 463         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
 464                     "dump raw trace in ASCII"),
 465         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 466                    "file", "vmlinux pathname"),
 467         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
 468                    "file", "kallsyms pathname"),
 469         OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
 470         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
 471                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
 472         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
 473                     "Show a column with the number of samples"),
 474         OPT_BOOLEAN('T', "threads", &show_threads,
 475                     "Show per-thread event counters"),
 476         OPT_STRING(0, "pretty", &pretty_printing_style, "key",
 477                    "pretty printing style key: normal raw"),
 478         OPT_BOOLEAN(0, "tui", &use_tui, "Use the TUI interface"),
 479         OPT_BOOLEAN(0, "stdio", &use_stdio, "Use the stdio interface"),
 480         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
 481                    "sort by key(s): pid, comm, dso, symbol, parent"),
 482         OPT_BOOLEAN(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
 483                     "Show sample percentage for different cpu modes"),
 484         OPT_STRING('p', "parent", &parent_pattern, "regex",
 485                    "regex filter to identify parent, see: '--sort parent'"),
 486         OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
 487                     "Only display entries with parent-match"),
 488         OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent",
 489                      "Display callchains using output_type (graph, flat, fractal, or none) and min percent threshold. "
 490                      "Default: fractal,0.5", &parse_callchain_opt, callchain_default_opt),
 491         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
 492                    "only consider symbols in these dsos"),
 493         OPT_STRING('C', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
 494                    "only consider symbols in these comms"),
 495         OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
 496                    "only consider these symbols"),
 497         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
 498                    "width[,width...]",
 499                    "don't try to adjust column width, use these fixed values"),
 500         OPT_STRING('t', "field-separator", &symbol_conf.field_sep, "separator",
 501                    "separator for columns, no spaces will be added between "
 502                    "columns '.' is reserved."),
 503         OPT_BOOLEAN('U', "hide-unresolved", &hide_unresolved,
 504                     "Only display entries resolved to a symbol"),
 505         OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
 506                     "Look for files with symbols relative to this directory"),
 507         OPT_END()
 508 };
 509
 510 int cmd_report(int argc, const char **argv, const char *prefix __used)
 511 {
 512         argc = parse_options(argc, argv, options, report_usage, 0);
 513
 514         if (use_stdio)
 515                 use_browser = 0;
 516         else if (use_tui)
 517                 use_browser = 1;
 518
 519         if (strcmp(input_name, "-") != 0)
 520                 setup_browser(true);
 521         else
 522                 use_browser = 0;
 523         /*
 524          * Only in the newt browser we are doing integrated annotation,
 525          * so don't allocate extra space that won't be used in the stdio
 526          * implementation.
 527          */
 528         if (use_browser > 0) {
 529                 symbol_conf.priv_size = sizeof(struct annotation);
 530                 annotate_init         = symbol__annotate_init;
 531                 /*
 532                  * For searching by name on the "Browse map details".
 533                  * providing it only in verbose mode not to bloat too
 534                  * much struct symbol.
 535                  */
 536                 if (verbose) {
 537                         /*
 538                          * XXX: Need to provide a less kludgy way to ask for
 539                          * more space per symbol, the u32 is for the index on
 540                          * the ui browser.
 541                          * See symbol__browser_index.
 542                          */
 543                         symbol_conf.priv_size += sizeof(u32);
 544                         symbol_conf.sort_by_name = true;
 545                 }
 546         }
 547
 548         if (symbol__init() < 0)
 549                 return -1;
 550
 551         setup_sorting(report_usage, options);
 552
 553         if (parent_pattern != default_parent_pattern) {
 554                 if (sort_dimension__add("parent") < 0)
 555                         return -1;
 556                 sort_parent.elide = 1;
 557         } else
 558                 symbol_conf.exclude_other = false;
 559
 560         /*
 561          * Any (unrecognized) arguments left?
 562          */
 563         if (argc)
 564                 usage_with_options(report_usage, options);
 565
 566         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
 567         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
 568         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
 569
 570         return __cmd_report();
 571 }