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