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;
  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 = append_chain(he->callchain, data->callchain, syms, data->period);
 111                 if (err)
 112                         goto out_free_syms;
 113         }
 114         /*
 115          * Only in the newt browser we are doing integrated annotation,
 116          * so we don't allocated the extra space needed because the stdio
 117          * code will not use it.
 118          */
 119         if (use_browser > 0)
 120                 err = hist_entry__inc_addr_samples(he, al->addr);
 121 out_free_syms:
 122         free(syms);
 123         return err;
 124 }
 125
 126 static int add_event_total(struct perf_session *session,
 127                            struct sample_data *data,
 128                            struct perf_event_attr *attr)
 129 {
 130         struct hists *hists;
 131
 132         if (attr)
 133                 hists = perf_session__hists_findnew(session, data->id,
 134                                                     attr->type, attr->config);
 135         else
 136                 hists = perf_session__hists_findnew(session, data->id, 0, 0);
 137
 138         if (!hists)
 139                 return -ENOMEM;
 140
 141         hists->stats.total_period += data->period;
 142         /*
 143          * FIXME: add_event_total should be moved from here to
 144          * perf_session__process_event so that the proper hist is passed to
 145          * the event_op methods.
 146          */
 147         hists__inc_nr_events(hists, PERF_RECORD_SAMPLE);
 148         session->hists.stats.total_period += data->period;
 149         return 0;
 150 }
 151
 152 static int process_sample_event(event_t *event, struct perf_session *session)
 153 {
 154         struct sample_data data = { .period = 1, };
 155         struct addr_location al;
 156         struct perf_event_attr *attr;
 157
 158         event__parse_sample(event, session->sample_type, &data);
 159
 160         dump_printf("(IP, %d): %d/%d: %#Lx period: %Ld\n", event->header.misc,
 161                     data.pid, data.tid, data.ip, data.period);
 162
 163         if (session->sample_type & PERF_SAMPLE_CALLCHAIN) {
 164                 unsigned int i;
 165
 166                 dump_printf("... chain: nr:%Lu\n", data.callchain->nr);
 167
 168                 if (!ip_callchain__valid(data.callchain, event)) {
 169                         pr_debug("call-chain problem with event, "
 170                                  "skipping it.\n");
 171                         return 0;
 172                 }
 173
 174                 if (dump_trace) {
 175                         for (i = 0; i < data.callchain->nr; i++)
 176                                 dump_printf("..... %2d: %016Lx\n",
 177                                             i, data.callchain->ips[i]);
 178                 }
 179         }
 180
 181         if (event__preprocess_sample(event, session, &al, NULL) < 0) {
 182                 fprintf(stderr, "problem processing %d event, skipping it.\n",
 183                         event->header.type);
 184                 return -1;
 185         }
 186
 187         if (al.filtered || (hide_unresolved && al.sym == NULL))
 188                 return 0;
 189
 190         if (perf_session__add_hist_entry(session, &al, &data)) {
 191                 pr_debug("problem incrementing symbol period, skipping event\n");
 192                 return -1;
 193         }
 194
 195         attr = perf_header__find_attr(data.id, &session->header);
 196
 197         if (add_event_total(session, &data, attr)) {
 198                 pr_debug("problem adding event period\n");
 199                 return -1;
 200         }
 201
 202         return 0;
 203 }
 204
 205 static int process_read_event(event_t *event, struct perf_session *session __used)
 206 {
 207         struct perf_event_attr *attr;
 208
 209         attr = perf_header__find_attr(event->read.id, &session->header);
 210
 211         if (show_threads) {
 212                 const char *name = attr ? __event_name(attr->type, attr->config)
 213                                    : "unknown";
 214                 perf_read_values_add_value(&show_threads_values,
 215                                            event->read.pid, event->read.tid,
 216                                            event->read.id,
 217                                            name,
 218                                            event->read.value);
 219         }
 220
 221         dump_printf(": %d %d %s %Lu\n", event->read.pid, event->read.tid,
 222                     attr ? __event_name(attr->type, attr->config) : "FAIL",
 223                     event->read.value);
 224
 225         return 0;
 226 }
 227
 228 static int perf_session__setup_sample_type(struct perf_session *self)
 229 {
 230         if (!(self->sample_type & PERF_SAMPLE_CALLCHAIN)) {
 231                 if (sort__has_parent) {
 232                         fprintf(stderr, "selected --sort parent, but no"
 233                                         " callchain data. Did you call"
 234                                         " perf record without -g?\n");
 235                         return -EINVAL;
 236                 }
 237                 if (symbol_conf.use_callchain) {
 238                         fprintf(stderr, "selected -g but no callchain data."
 239                                         " Did you call perf record without"
 240                                         " -g?\n");
 241                         return -1;
 242                 }
 243         } else if (!dont_use_callchains && callchain_param.mode != CHAIN_NONE &&
 244                    !symbol_conf.use_callchain) {
 245                         symbol_conf.use_callchain = true;
 246                         if (register_callchain_param(&callchain_param) < 0) {
 247                                 fprintf(stderr, "Can't register callchain"
 248                                                 " params\n");
 249                                 return -EINVAL;
 250                         }
 251         }
 252
 253         return 0;
 254 }
 255
 256 static struct perf_event_ops event_ops = {
 257         .sample = process_sample_event,
 258         .mmap   = event__process_mmap,
 259         .comm   = event__process_comm,
 260         .exit   = event__process_task,
 261         .fork   = event__process_task,
 262         .lost   = event__process_lost,
 263         .read   = process_read_event,
 264         .attr   = event__process_attr,
 265         .event_type = event__process_event_type,
 266         .tracing_data = event__process_tracing_data,
 267         .build_id = event__process_build_id,
 268 };
 269
 270 extern volatile int session_done;
 271
 272 static void sig_handler(int sig __used)
 273 {
 274         session_done = 1;
 275 }
 276
 277 static size_t hists__fprintf_nr_sample_events(struct hists *self,
 278                                               const char *evname, FILE *fp)
 279 {
 280         size_t ret;
 281         char unit;
 282         unsigned long nr_events = self->stats.nr_events[PERF_RECORD_SAMPLE];
 283
 284         nr_events = convert_unit(nr_events, &unit);
 285         ret = fprintf(fp, "# Events: %lu%c", nr_events, unit);
 286         if (evname != NULL)
 287                 ret += fprintf(fp, " %s", evname);
 288         return ret + fprintf(fp, "\n#\n");
 289 }
 290
 291 static int hists__tty_browse_tree(struct rb_root *tree, const char *help)
 292 {
 293         struct rb_node *next = rb_first(tree);
 294
 295         while (next) {
 296                 struct hists *hists = rb_entry(next, struct hists, rb_node);
 297                 const char *evname = NULL;
 298
 299                 if (rb_first(&hists->entries) != rb_last(&hists->entries))
 300                         evname = __event_name(hists->type, hists->config);
 301
 302                 hists__fprintf_nr_sample_events(hists, evname, stdout);
 303                 hists__fprintf(hists, NULL, false, stdout);
 304                 fprintf(stdout, "\n\n");
 305                 next = rb_next(&hists->rb_node);
 306         }
 307
 308         if (sort_order == default_sort_order &&
 309             parent_pattern == default_parent_pattern) {
 310                 fprintf(stdout, "#\n# (%s)\n#\n", help);
 311
 312                 if (show_threads) {
 313                         bool style = !strcmp(pretty_printing_style, "raw");
 314                         perf_read_values_display(stdout, &show_threads_values,
 315                                                  style);
 316                         perf_read_values_destroy(&show_threads_values);
 317                 }
 318         }
 319
 320         return 0;
 321 }
 322
 323 static int __cmd_report(void)
 324 {
 325         int ret = -EINVAL;
 326         struct perf_session *session;
 327         struct rb_node *next;
 328         const char *help = "For a higher level overview, try: perf report --sort comm,dso";
 329
 330         signal(SIGINT, sig_handler);
 331
 332         session = perf_session__new(input_name, O_RDONLY, force, false);
 333         if (session == NULL)
 334                 return -ENOMEM;
 335
 336         if (show_threads)
 337                 perf_read_values_init(&show_threads_values);
 338
 339         ret = perf_session__setup_sample_type(session);
 340         if (ret)
 341                 goto out_delete;
 342
 343         ret = perf_session__process_events(session, &event_ops);
 344         if (ret)
 345                 goto out_delete;
 346
 347         if (dump_trace) {
 348                 perf_session__fprintf_nr_events(session, stdout);
 349                 goto out_delete;
 350         }
 351
 352         if (verbose > 3)
 353                 perf_session__fprintf(session, stdout);
 354
 355         if (verbose > 2)
 356                 perf_session__fprintf_dsos(session, stdout);
 357
 358         next = rb_first(&session->hists_tree);
 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);
 370         else
 371                 hists__tty_browse_tree(&session->hists_tree, help);
 372
 373 out_delete:
 374         perf_session__delete(session);
 375         return ret;
 376 }
 377
 378 static int
 379 parse_callchain_opt(const struct option *opt __used, const char *arg,
 380                     int unset)
 381 {
 382         char *tok, *tok2;
 383         char *endptr;
 384
 385         /*
 386          * --no-call-graph
 387          */
 388         if (unset) {
 389                 dont_use_callchains = true;
 390                 return 0;
 391         }
 392
 393         symbol_conf.use_callchain = true;
 394
 395         if (!arg)
 396                 return 0;
 397
 398         tok = strtok((char *)arg, ",");
 399         if (!tok)
 400                 return -1;
 401
 402         /* get the output mode */
 403         if (!strncmp(tok, "graph", strlen(arg)))
 404                 callchain_param.mode = CHAIN_GRAPH_ABS;
 405
 406         else if (!strncmp(tok, "flat", strlen(arg)))
 407                 callchain_param.mode = CHAIN_FLAT;
 408
 409         else if (!strncmp(tok, "fractal", strlen(arg)))
 410                 callchain_param.mode = CHAIN_GRAPH_REL;
 411
 412         else if (!strncmp(tok, "none", strlen(arg))) {
 413                 callchain_param.mode = CHAIN_NONE;
 414                 symbol_conf.use_callchain = false;
 415
 416                 return 0;
 417         }
 418
 419         else
 420                 return -1;
 421
 422         /* get the min percentage */
 423         tok = strtok(NULL, ",");
 424         if (!tok)
 425                 goto setup;
 426
 427         tok2 = strtok(NULL, ",");
 428         callchain_param.min_percent = strtod(tok, &endptr);
 429         if (tok == endptr)
 430                 return -1;
 431
 432         if (tok2)
 433                 callchain_param.print_limit = strtod(tok2, &endptr);
 434 setup:
 435         if (register_callchain_param(&callchain_param) < 0) {
 436                 fprintf(stderr, "Can't register callchain params\n");
 437                 return -1;
 438         }
 439         return 0;
 440 }
 441
 442 static const char * const report_usage[] = {
 443         "perf report [<options>] <command>",
 444         NULL
 445 };
 446
 447 static const struct option options[] = {
 448         OPT_STRING('i', "input", &input_name, "file",
 449                     "input file name"),
 450         OPT_INCR('v', "verbose", &verbose,
 451                     "be more verbose (show symbol address, etc)"),
 452         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
 453                     "dump raw trace in ASCII"),
 454         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 455                    "file", "vmlinux pathname"),
 456         OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
 457         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
 458                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
 459         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
 460                     "Show a column with the number of samples"),
 461         OPT_BOOLEAN('T', "threads", &show_threads,
 462                     "Show per-thread event counters"),
 463         OPT_STRING(0, "pretty", &pretty_printing_style, "key",
 464                    "pretty printing style key: normal raw"),
 465         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
 466                    "sort by key(s): pid, comm, dso, symbol, parent"),
 467         OPT_BOOLEAN('P', "full-paths", &symbol_conf.full_paths,
 468                     "Don't shorten the pathnames taking into account the cwd"),
 469         OPT_BOOLEAN(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
 470                     "Show sample percentage for different cpu modes"),
 471         OPT_STRING('p', "parent", &parent_pattern, "regex",
 472                    "regex filter to identify parent, see: '--sort parent'"),
 473         OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
 474                     "Only display entries with parent-match"),
 475         OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent",
 476                      "Display callchains using output_type (graph, flat, fractal, or none) and min percent threshold. "
 477                      "Default: fractal,0.5", &parse_callchain_opt, callchain_default_opt),
 478         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
 479                    "only consider symbols in these dsos"),
 480         OPT_STRING('C', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
 481                    "only consider symbols in these comms"),
 482         OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
 483                    "only consider these symbols"),
 484         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
 485                    "width[,width...]",
 486                    "don't try to adjust column width, use these fixed values"),
 487         OPT_STRING('t', "field-separator", &symbol_conf.field_sep, "separator",
 488                    "separator for columns, no spaces will be added between "
 489                    "columns '.' is reserved."),
 490         OPT_BOOLEAN('U', "hide-unresolved", &hide_unresolved,
 491                     "Only display entries resolved to a symbol"),
 492         OPT_END()
 493 };
 494
 495 int cmd_report(int argc, const char **argv, const char *prefix __used)
 496 {
 497         argc = parse_options(argc, argv, options, report_usage, 0);
 498
 499         if (strcmp(input_name, "-") != 0)
 500                 setup_browser();
 501         /*
 502          * Only in the newt browser we are doing integrated annotation,
 503          * so don't allocate extra space that won't be used in the stdio
 504          * implementation.
 505          */
 506         if (use_browser > 0)
 507                 symbol_conf.priv_size = sizeof(struct sym_priv);
 508
 509         if (symbol__init() < 0)
 510                 return -1;
 511
 512         setup_sorting(report_usage, options);
 513
 514         if (parent_pattern != default_parent_pattern) {
 515                 if (sort_dimension__add("parent") < 0)
 516                         return -1;
 517                 sort_parent.elide = 1;
 518         } else
 519                 symbol_conf.exclude_other = false;
 520
 521         /*
 522          * Any (unrecognized) arguments left?
 523          */
 524         if (argc)
 525                 usage_with_options(report_usage, options);
 526
 527         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
 528         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
 529         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
 530
 531         return __cmd_report();
 532 }