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/color.h"
  13 #include <linux/list.h>
  14 #include "util/cache.h"
  15 #include <linux/rbtree.h>
  16 #include "util/symbol.h"
  17 #include "util/callchain.h"
  18 #include "util/strlist.h"
  19 #include "util/values.h"
  20
  21 #include "perf.h"
  22 #include "util/debug.h"
  23 #include "util/header.h"
  24 #include "util/session.h"
  25
  26 #include "util/parse-options.h"
  27 #include "util/parse-events.h"
  28
  29 #include "util/thread.h"
  30 #include "util/sort.h"
  31 #include "util/hist.h"
  32
  33 static char             const *input_name = "perf.data";
  34
  35 static bool             force, use_tui, use_stdio;
  36 static bool             hide_unresolved;
  37 static bool             dont_use_callchains;
  38
  39 static bool             show_threads;
  40 static struct perf_read_values  show_threads_values;
  41
  42 static const char       default_pretty_printing_style[] = "normal";
  43 static const char       *pretty_printing_style = default_pretty_printing_style;
  44
  45 static char             callchain_default_opt[] = "fractal,0.5";
  46
  47 static struct hists *perf_session__hists_findnew(struct perf_session *self,
  48                                                  u64 event_stream, u32 type,
  49                                                  u64 config)
  50 {
  51         struct rb_node **p = &self->hists_tree.rb_node;
  52         struct rb_node *parent = NULL;
  53         struct hists *iter, *new;
  54
  55         while (*p != NULL) {
  56                 parent = *p;
  57                 iter = rb_entry(parent, struct hists, rb_node);
  58                 if (iter->config == config)
  59                         return iter;
  60
  61
  62                 if (config > iter->config)
  63                         p = &(*p)->rb_right;
  64                 else
  65                         p = &(*p)->rb_left;
  66         }
  67
  68         new = malloc(sizeof(struct hists));
  69         if (new == NULL)
  70                 return NULL;
  71         memset(new, 0, sizeof(struct hists));
  72         new->event_stream = event_stream;
  73         new->config = config;
  74         new->type = type;
  75         rb_link_node(&new->rb_node, parent, p);
  76         rb_insert_color(&new->rb_node, &self->hists_tree);
  77         return new;
  78 }
  79
  80 static int perf_session__add_hist_entry(struct perf_session *self,
  81                                         struct addr_location *al,
  82                                         struct sample_data *data)
  83 {
  84         struct map_symbol *syms = NULL;
  85         struct symbol *parent = NULL;
  86         int err = -ENOMEM;
  87         struct hist_entry *he;
  88         struct hists *hists;
  89         struct perf_event_attr *attr;
  90
  91         if ((sort__has_parent || symbol_conf.use_callchain) && data->callchain) {
  92                 syms = perf_session__resolve_callchain(self, al->thread,
  93                                                        data->callchain, &parent);
  94                 if (syms == NULL)
  95                         return -ENOMEM;
  96         }
  97
  98         attr = perf_header__find_attr(data->id, &self->header);
  99         if (attr)
 100                 hists = perf_session__hists_findnew(self, data->id, attr->type, attr->config);
 101         else
 102                 hists = perf_session__hists_findnew(self, data->id, 0, 0);
 103         if (hists == NULL)
 104                 goto out_free_syms;
 105         he = __hists__add_entry(hists, al, parent, data->period);
 106         if (he == NULL)
 107                 goto out_free_syms;
 108         err = 0;
 109         if (symbol_conf.use_callchain) {
 110                 err = callchain_append(he->callchain, data->callchain, syms,
 111                                        data->period);
 112                 if (err)
 113                         goto out_free_syms;
 114         }
 115         /*
 116          * Only in the newt browser we are doing integrated annotation,
 117          * so we don't allocated the extra space needed because the stdio
 118          * code will not use it.
 119          */
 120         if (use_browser > 0)
 121                 err = hist_entry__inc_addr_samples(he, al->addr);
 122 out_free_syms:
 123         free(syms);
 124         return err;
 125 }
 126
 127 static int add_event_total(struct perf_session *session,
 128                            struct sample_data *data,
 129                            struct perf_event_attr *attr)
 130 {
 131         struct hists *hists;
 132
 133         if (attr)
 134                 hists = perf_session__hists_findnew(session, data->id,
 135                                                     attr->type, attr->config);
 136         else
 137                 hists = perf_session__hists_findnew(session, data->id, 0, 0);
 138
 139         if (!hists)
 140                 return -ENOMEM;
 141
 142         hists->stats.total_period += data->period;
 143         /*
 144          * FIXME: add_event_total should be moved from here to
 145          * perf_session__process_event so that the proper hist is passed to
 146          * the event_op methods.
 147          */
 148         hists__inc_nr_events(hists, PERF_RECORD_SAMPLE);
 149         session->hists.stats.total_period += data->period;
 150         return 0;
 151 }
 152
 153 static int process_sample_event(event_t *event, struct sample_data *sample,
 154                                 struct perf_session *session)
 155 {
 156         struct addr_location al;
 157         struct perf_event_attr *attr;
 158
 159         if (event__preprocess_sample(event, session, &al, sample, NULL) < 0) {
 160                 fprintf(stderr, "problem processing %d event, skipping it.\n",
 161                         event->header.type);
 162                 return -1;
 163         }
 164
 165         if (al.filtered || (hide_unresolved && al.sym == NULL))
 166                 return 0;
 167
 168         if (perf_session__add_hist_entry(session, &al, sample)) {
 169                 pr_debug("problem incrementing symbol period, skipping event\n");
 170                 return -1;
 171         }
 172
 173         attr = perf_header__find_attr(sample->id, &session->header);
 174
 175         if (add_event_total(session, sample, attr)) {
 176                 pr_debug("problem adding event period\n");
 177                 return -1;
 178         }
 179
 180         return 0;
 181 }
 182
 183 static int process_read_event(event_t *event, struct sample_data *sample __used,
 184                               struct perf_session *session __used)
 185 {
 186         struct perf_event_attr *attr;
 187
 188         attr = perf_header__find_attr(event->read.id, &session->header);
 189
 190         if (show_threads) {
 191                 const char *name = attr ? __event_name(attr->type, attr->config)
 192                                    : "unknown";
 193                 perf_read_values_add_value(&show_threads_values,
 194                                            event->read.pid, event->read.tid,
 195                                            event->read.id,
 196                                            name,
 197                                            event->read.value);
 198         }
 199
 200         dump_printf(": %d %d %s %Lu\n", event->read.pid, event->read.tid,
 201                     attr ? __event_name(attr->type, attr->config) : "FAIL",
 202                     event->read.value);
 203
 204         return 0;
 205 }
 206
 207 static int perf_session__setup_sample_type(struct perf_session *self)
 208 {
 209         if (!(self->sample_type & PERF_SAMPLE_CALLCHAIN)) {
 210                 if (sort__has_parent) {
 211                         fprintf(stderr, "selected --sort parent, but no"
 212                                         " callchain data. Did you call"
 213                                         " perf record without -g?\n");
 214                         return -EINVAL;
 215                 }
 216                 if (symbol_conf.use_callchain) {
 217                         fprintf(stderr, "selected -g but no callchain data."
 218                                         " Did you call perf record without"
 219                                         " -g?\n");
 220                         return -1;
 221                 }
 222         } else if (!dont_use_callchains && callchain_param.mode != CHAIN_NONE &&
 223                    !symbol_conf.use_callchain) {
 224                         symbol_conf.use_callchain = true;
 225                         if (register_callchain_param(&callchain_param) < 0) {
 226                                 fprintf(stderr, "Can't register callchain"
 227                                                 " params\n");
 228                                 return -EINVAL;
 229                         }
 230         }
 231
 232         return 0;
 233 }
 234
 235 static struct perf_event_ops event_ops = {
 236         .sample = process_sample_event,
 237         .mmap   = event__process_mmap,
 238         .comm   = event__process_comm,
 239         .exit   = event__process_task,
 240         .fork   = event__process_task,
 241         .lost   = event__process_lost,
 242         .read   = process_read_event,
 243         .attr   = event__process_attr,
 244         .event_type = event__process_event_type,
 245         .tracing_data = event__process_tracing_data,
 246         .build_id = event__process_build_id,
 247 };
 248
 249 extern volatile int session_done;
 250
 251 static void sig_handler(int sig __used)
 252 {
 253         session_done = 1;
 254 }
 255
 256 static size_t hists__fprintf_nr_sample_events(struct hists *self,
 257                                               const char *evname, FILE *fp)
 258 {
 259         size_t ret;
 260         char unit;
 261         unsigned long nr_events = self->stats.nr_events[PERF_RECORD_SAMPLE];
 262
 263         nr_events = convert_unit(nr_events, &unit);
 264         ret = fprintf(fp, "# Events: %lu%c", nr_events, unit);
 265         if (evname != NULL)
 266                 ret += fprintf(fp, " %s", evname);
 267         return ret + fprintf(fp, "\n#\n");
 268 }
 269
 270 static int hists__tty_browse_tree(struct rb_root *tree, const char *help)
 271 {
 272         struct rb_node *next = rb_first(tree);
 273
 274         while (next) {
 275                 struct hists *hists = rb_entry(next, struct hists, rb_node);
 276                 const char *evname = NULL;
 277
 278                 if (rb_first(&hists->entries) != rb_last(&hists->entries))
 279                         evname = __event_name(hists->type, hists->config);
 280
 281                 hists__fprintf_nr_sample_events(hists, evname, stdout);
 282                 hists__fprintf(hists, NULL, false, stdout);
 283                 fprintf(stdout, "\n\n");
 284                 next = rb_next(&hists->rb_node);
 285         }
 286
 287         if (sort_order == default_sort_order &&
 288             parent_pattern == default_parent_pattern) {
 289                 fprintf(stdout, "#\n# (%s)\n#\n", help);
 290
 291                 if (show_threads) {
 292                         bool style = !strcmp(pretty_printing_style, "raw");
 293                         perf_read_values_display(stdout, &show_threads_values,
 294                                                  style);
 295                         perf_read_values_destroy(&show_threads_values);
 296                 }
 297         }
 298
 299         return 0;
 300 }
 301
 302 static int __cmd_report(void)
 303 {
 304         int ret = -EINVAL;
 305         struct perf_session *session;
 306         struct rb_node *next;
 307         const char *help = "For a higher level overview, try: perf report --sort comm,dso";
 308
 309         signal(SIGINT, sig_handler);
 310
 311         session = perf_session__new(input_name, O_RDONLY, force, false, &event_ops);
 312         if (session == NULL)
 313                 return -ENOMEM;
 314
 315         if (show_threads)
 316                 perf_read_values_init(&show_threads_values);
 317
 318         ret = perf_session__setup_sample_type(session);
 319         if (ret)
 320                 goto out_delete;
 321
 322         ret = perf_session__process_events(session, &event_ops);
 323         if (ret)
 324                 goto out_delete;
 325
 326         if (dump_trace) {
 327                 perf_session__fprintf_nr_events(session, stdout);
 328                 goto out_delete;
 329         }
 330
 331         if (verbose > 3)
 332                 perf_session__fprintf(session, stdout);
 333
 334         if (verbose > 2)
 335                 perf_session__fprintf_dsos(session, stdout);
 336
 337         next = rb_first(&session->hists_tree);
 338         while (next) {
 339                 struct hists *hists;
 340
 341                 hists = rb_entry(next, struct hists, rb_node);
 342                 hists__collapse_resort(hists);
 343                 hists__output_resort(hists);
 344                 next = rb_next(&hists->rb_node);
 345         }
 346
 347         if (use_browser > 0)
 348                 hists__tui_browse_tree(&session->hists_tree, help);
 349         else
 350                 hists__tty_browse_tree(&session->hists_tree, help);
 351
 352 out_delete:
 353         /*
 354          * Speed up the exit process, for large files this can
 355          * take quite a while.
 356          *
 357          * XXX Enable this when using valgrind or if we ever
 358          * librarize this command.
 359          *
 360          * Also experiment with obstacks to see how much speed
 361          * up we'll get here.
 362          *
 363          * perf_session__delete(session);
 364          */
 365         return ret;
 366 }
 367
 368 static int
 369 parse_callchain_opt(const struct option *opt __used, const char *arg,
 370                     int unset)
 371 {
 372         char *tok, *tok2;
 373         char *endptr;
 374
 375         /*
 376          * --no-call-graph
 377          */
 378         if (unset) {
 379                 dont_use_callchains = true;
 380                 return 0;
 381         }
 382
 383         symbol_conf.use_callchain = true;
 384
 385         if (!arg)
 386                 return 0;
 387
 388         tok = strtok((char *)arg, ",");
 389         if (!tok)
 390                 return -1;
 391
 392         /* get the output mode */
 393         if (!strncmp(tok, "graph", strlen(arg)))
 394                 callchain_param.mode = CHAIN_GRAPH_ABS;
 395
 396         else if (!strncmp(tok, "flat", strlen(arg)))
 397                 callchain_param.mode = CHAIN_FLAT;
 398
 399         else if (!strncmp(tok, "fractal", strlen(arg)))
 400                 callchain_param.mode = CHAIN_GRAPH_REL;
 401
 402         else if (!strncmp(tok, "none", strlen(arg))) {
 403                 callchain_param.mode = CHAIN_NONE;
 404                 symbol_conf.use_callchain = false;
 405
 406                 return 0;
 407         }
 408
 409         else
 410                 return -1;
 411
 412         /* get the min percentage */
 413         tok = strtok(NULL, ",");
 414         if (!tok)
 415                 goto setup;
 416
 417         tok2 = strtok(NULL, ",");
 418         callchain_param.min_percent = strtod(tok, &endptr);
 419         if (tok == endptr)
 420                 return -1;
 421
 422         if (tok2)
 423                 callchain_param.print_limit = strtod(tok2, &endptr);
 424 setup:
 425         if (register_callchain_param(&callchain_param) < 0) {
 426                 fprintf(stderr, "Can't register callchain params\n");
 427                 return -1;
 428         }
 429         return 0;
 430 }
 431
 432 static const char * const report_usage[] = {
 433         "perf report [<options>] <command>",
 434         NULL
 435 };
 436
 437 static const struct option options[] = {
 438         OPT_STRING('i', "input", &input_name, "file",
 439                     "input file name"),
 440         OPT_INCR('v', "verbose", &verbose,
 441                     "be more verbose (show symbol address, etc)"),
 442         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
 443                     "dump raw trace in ASCII"),
 444         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 445                    "file", "vmlinux pathname"),
 446         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
 447                    "file", "kallsyms pathname"),
 448         OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
 449         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
 450                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
 451         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
 452                     "Show a column with the number of samples"),
 453         OPT_BOOLEAN('T', "threads", &show_threads,
 454                     "Show per-thread event counters"),
 455         OPT_STRING(0, "pretty", &pretty_printing_style, "key",
 456                    "pretty printing style key: normal raw"),
 457         OPT_BOOLEAN(0, "tui", &use_tui, "Use the TUI interface"),
 458         OPT_BOOLEAN(0, "stdio", &use_stdio, "Use the stdio interface"),
 459         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
 460                    "sort by key(s): pid, comm, dso, symbol, parent"),
 461         OPT_BOOLEAN(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
 462                     "Show sample percentage for different cpu modes"),
 463         OPT_STRING('p', "parent", &parent_pattern, "regex",
 464                    "regex filter to identify parent, see: '--sort parent'"),
 465         OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
 466                     "Only display entries with parent-match"),
 467         OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent",
 468                      "Display callchains using output_type (graph, flat, fractal, or none) and min percent threshold. "
 469                      "Default: fractal,0.5", &parse_callchain_opt, callchain_default_opt),
 470         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
 471                    "only consider symbols in these dsos"),
 472         OPT_STRING('C', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
 473                    "only consider symbols in these comms"),
 474         OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
 475                    "only consider these symbols"),
 476         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
 477                    "width[,width...]",
 478                    "don't try to adjust column width, use these fixed values"),
 479         OPT_STRING('t', "field-separator", &symbol_conf.field_sep, "separator",
 480                    "separator for columns, no spaces will be added between "
 481                    "columns '.' is reserved."),
 482         OPT_BOOLEAN('U', "hide-unresolved", &hide_unresolved,
 483                     "Only display entries resolved to a symbol"),
 484         OPT_END()
 485 };
 486
 487 int cmd_report(int argc, const char **argv, const char *prefix __used)
 488 {
 489         argc = parse_options(argc, argv, options, report_usage, 0);
 490
 491         if (use_stdio)
 492                 use_browser = 0;
 493         else if (use_tui)
 494                 use_browser = 1;
 495
 496         if (strcmp(input_name, "-") != 0)
 497                 setup_browser();
 498         else
 499                 use_browser = 0;
 500         /*
 501          * Only in the newt browser we are doing integrated annotation,
 502          * so don't allocate extra space that won't be used in the stdio
 503          * implementation.
 504          */
 505         if (use_browser > 0) {
 506                 symbol_conf.priv_size = sizeof(struct sym_priv);
 507                 /*
 508                  * For searching by name on the "Browse map details".
 509                  * providing it only in verbose mode not to bloat too
 510                  * much struct symbol.
 511                  */
 512                 if (verbose) {
 513                         /*
 514                          * XXX: Need to provide a less kludgy way to ask for
 515                          * more space per symbol, the u32 is for the index on
 516                          * the ui browser.
 517                          * See symbol__browser_index.
 518                          */
 519                         symbol_conf.priv_size += sizeof(u32);
 520                         symbol_conf.sort_by_name = true;
 521                 }
 522         }
 523
 524         if (symbol__init() < 0)
 525                 return -1;
 526
 527         setup_sorting(report_usage, options);
 528
 529         if (parent_pattern != default_parent_pattern) {
 530                 if (sort_dimension__add("parent") < 0)
 531                         return -1;
 532                 sort_parent.elide = 1;
 533         } else
 534                 symbol_conf.exclude_other = false;
 535
 536         /*
 537          * Any (unrecognized) arguments left?
 538          */
 539         if (argc)
 540                 usage_with_options(report_usage, options);
 541
 542         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
 543         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
 544         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
 545
 546         return __cmd_report();
 547 }