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,callee";
  49 static bool             inverted_callchain;
  50 static symbol_filter_t  annotate_init;
  51
  52 static int perf_session__add_hist_entry(struct perf_session *session,
  53                                         struct addr_location *al,
  54                                         struct perf_sample *sample,
  55                                         struct perf_evsel *evsel)
  56 {
  57         struct symbol *parent = NULL;
  58         int err = 0;
  59         struct hist_entry *he;
  60
  61         if ((sort__has_parent || symbol_conf.use_callchain) && sample->callchain) {
  62                 err = perf_session__resolve_callchain(session, al->thread,
  63                                                       sample->callchain, &parent);
  64                 if (err)
  65                         return err;
  66         }
  67
  68         he = __hists__add_entry(&evsel->hists, al, parent, sample->period);
  69         if (he == NULL)
  70                 return -ENOMEM;
  71
  72         if (symbol_conf.use_callchain) {
  73                 err = callchain_append(he->callchain, &session->callchain_cursor,
  74                                        sample->period);
  75                 if (err)
  76                         return err;
  77         }
  78         /*
  79          * Only in the newt browser we are doing integrated annotation,
  80          * so we don't allocated the extra space needed because the stdio
  81          * code will not use it.
  82          */
  83         if (al->sym != NULL && use_browser > 0) {
  84                 struct annotation *notes = symbol__annotation(he->ms.sym);
  85
  86                 assert(evsel != NULL);
  87
  88                 err = -ENOMEM;
  89                 if (notes->src == NULL &&
  90                     symbol__alloc_hist(he->ms.sym, session->evlist->nr_entries) < 0)
  91                         goto out;
  92
  93                 err = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
  94         }
  95
  96         evsel->hists.stats.total_period += sample->period;
  97         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
  98 out:
  99         return err;
 100 }
 101
 102
 103 static int process_sample_event(union perf_event *event,
 104                                 struct perf_sample *sample,
 105                                 struct perf_evsel *evsel,
 106                                 struct perf_session *session)
 107 {
 108         struct addr_location al;
 109
 110         if (perf_event__preprocess_sample(event, session, &al, sample,
 111                                           annotate_init) < 0) {
 112                 fprintf(stderr, "problem processing %d event, skipping it.\n",
 113                         event->header.type);
 114                 return -1;
 115         }
 116
 117         if (al.filtered || (hide_unresolved && al.sym == NULL))
 118                 return 0;
 119
 120         if (al.map != NULL)
 121                 al.map->dso->hit = 1;
 122
 123         if (perf_session__add_hist_entry(session, &al, sample, evsel)) {
 124                 pr_debug("problem incrementing symbol period, skipping event\n");
 125                 return -1;
 126         }
 127
 128         return 0;
 129 }
 130
 131 static int process_read_event(union perf_event *event,
 132                               struct perf_sample *sample __used,
 133                               struct perf_session *session)
 134 {
 135         struct perf_evsel *evsel = perf_evlist__id2evsel(session->evlist,
 136                                                          event->read.id);
 137         if (show_threads) {
 138                 const char *name = evsel ? event_name(evsel) : "unknown";
 139                 perf_read_values_add_value(&show_threads_values,
 140                                            event->read.pid, event->read.tid,
 141                                            event->read.id,
 142                                            name,
 143                                            event->read.value);
 144         }
 145
 146         dump_printf(": %d %d %s %" PRIu64 "\n", event->read.pid, event->read.tid,
 147                     evsel ? event_name(evsel) : "FAIL",
 148                     event->read.value);
 149
 150         return 0;
 151 }
 152
 153 static int perf_session__setup_sample_type(struct perf_session *self)
 154 {
 155         if (!(self->sample_type & PERF_SAMPLE_CALLCHAIN)) {
 156                 if (sort__has_parent) {
 157                         fprintf(stderr, "selected --sort parent, but no"
 158                                         " callchain data. Did you call"
 159                                         " perf record without -g?\n");
 160                         return -EINVAL;
 161                 }
 162                 if (symbol_conf.use_callchain) {
 163                         fprintf(stderr, "selected -g but no callchain data."
 164                                         " Did you call perf record without"
 165                                         " -g?\n");
 166                         return -1;
 167                 }
 168         } else if (!dont_use_callchains && callchain_param.mode != CHAIN_NONE &&
 169                    !symbol_conf.use_callchain) {
 170                         symbol_conf.use_callchain = true;
 171                         if (callchain_register_param(&callchain_param) < 0) {
 172                                 fprintf(stderr, "Can't register callchain"
 173                                                 " params\n");
 174                                 return -EINVAL;
 175                         }
 176         }
 177
 178         return 0;
 179 }
 180
 181 static struct perf_event_ops event_ops = {
 182         .sample          = process_sample_event,
 183         .mmap            = perf_event__process_mmap,
 184         .comm            = perf_event__process_comm,
 185         .exit            = perf_event__process_task,
 186         .fork            = perf_event__process_task,
 187         .lost            = perf_event__process_lost,
 188         .read            = process_read_event,
 189         .attr            = perf_event__process_attr,
 190         .event_type      = perf_event__process_event_type,
 191         .tracing_data    = perf_event__process_tracing_data,
 192         .build_id        = perf_event__process_build_id,
 193         .ordered_samples = true,
 194         .ordering_requires_timestamps = true,
 195 };
 196
 197 extern volatile int session_done;
 198
 199 static void sig_handler(int sig __used)
 200 {
 201         session_done = 1;
 202 }
 203
 204 static size_t hists__fprintf_nr_sample_events(struct hists *self,
 205                                               const char *evname, FILE *fp)
 206 {
 207         size_t ret;
 208         char unit;
 209         unsigned long nr_events = self->stats.nr_events[PERF_RECORD_SAMPLE];
 210
 211         nr_events = convert_unit(nr_events, &unit);
 212         ret = fprintf(fp, "# Events: %lu%c", nr_events, unit);
 213         if (evname != NULL)
 214                 ret += fprintf(fp, " %s", evname);
 215         return ret + fprintf(fp, "\n#\n");
 216 }
 217
 218 static int perf_evlist__tty_browse_hists(struct perf_evlist *evlist,
 219                                          const char *help)
 220 {
 221         struct perf_evsel *pos;
 222
 223         list_for_each_entry(pos, &evlist->entries, node) {
 224                 struct hists *hists = &pos->hists;
 225                 const char *evname = NULL;
 226
 227                 if (rb_first(&hists->entries) != rb_last(&hists->entries))
 228                         evname = event_name(pos);
 229
 230                 hists__fprintf_nr_sample_events(hists, evname, stdout);
 231                 hists__fprintf(hists, NULL, false, stdout);
 232                 fprintf(stdout, "\n\n");
 233         }
 234
 235         if (sort_order == default_sort_order &&
 236             parent_pattern == default_parent_pattern) {
 237                 fprintf(stdout, "#\n# (%s)\n#\n", help);
 238
 239                 if (show_threads) {
 240                         bool style = !strcmp(pretty_printing_style, "raw");
 241                         perf_read_values_display(stdout, &show_threads_values,
 242                                                  style);
 243                         perf_read_values_destroy(&show_threads_values);
 244                 }
 245         }
 246
 247         return 0;
 248 }
 249
 250 static int __cmd_report(void)
 251 {
 252         int ret = -EINVAL;
 253         u64 nr_samples;
 254         struct perf_session *session;
 255         struct perf_evsel *pos;
 256         struct map *kernel_map;
 257         struct kmap *kernel_kmap;
 258         const char *help = "For a higher level overview, try: perf report --sort comm,dso";
 259
 260         signal(SIGINT, sig_handler);
 261
 262         session = perf_session__new(input_name, O_RDONLY, force, false, &event_ops);
 263         if (session == NULL)
 264                 return -ENOMEM;
 265
 266         if (show_threads)
 267                 perf_read_values_init(&show_threads_values);
 268
 269         ret = perf_session__setup_sample_type(session);
 270         if (ret)
 271                 goto out_delete;
 272
 273         ret = perf_session__process_events(session, &event_ops);
 274         if (ret)
 275                 goto out_delete;
 276
 277         kernel_map = session->host_machine.vmlinux_maps[MAP__FUNCTION];
 278         kernel_kmap = map__kmap(kernel_map);
 279         if (kernel_map == NULL ||
 280             (kernel_map->dso->hit &&
 281              (kernel_kmap->ref_reloc_sym == NULL ||
 282               kernel_kmap->ref_reloc_sym->addr == 0))) {
 283                 const struct dso *kdso = kernel_map->dso;
 284
 285                 ui__warning(
 286 "Kernel address maps (/proc/{kallsyms,modules}) were restricted.\n\n"
 287 "Check /proc/sys/kernel/kptr_restrict before running 'perf record'.\n\n%s\n\n"
 288 "Samples in kernel modules can't be resolved as well.\n\n",
 289                             RB_EMPTY_ROOT(&kdso->symbols[MAP__FUNCTION]) ?
 290 "As no suitable kallsyms nor vmlinux was found, kernel samples\n"
 291 "can't be resolved." :
 292 "If some relocation was applied (e.g. kexec) symbols may be misresolved.");
 293         }
 294
 295         if (dump_trace) {
 296                 perf_session__fprintf_nr_events(session, stdout);
 297                 goto out_delete;
 298         }
 299
 300         if (verbose > 3)
 301                 perf_session__fprintf(session, stdout);
 302
 303         if (verbose > 2)
 304                 perf_session__fprintf_dsos(session, stdout);
 305
 306         nr_samples = 0;
 307         list_for_each_entry(pos, &session->evlist->entries, node) {
 308                 struct hists *hists = &pos->hists;
 309
 310                 hists__collapse_resort(hists);
 311                 hists__output_resort(hists);
 312                 nr_samples += hists->stats.nr_events[PERF_RECORD_SAMPLE];
 313         }
 314
 315         if (nr_samples == 0) {
 316                 ui__warning("The %s file has no samples!\n", input_name);
 317                 goto out_delete;
 318         }
 319
 320         if (use_browser > 0)
 321                 perf_evlist__tui_browse_hists(session->evlist, help);
 322         else
 323                 perf_evlist__tty_browse_hists(session->evlist, help);
 324
 325 out_delete:
 326         /*
 327          * Speed up the exit process, for large files this can
 328          * take quite a while.
 329          *
 330          * XXX Enable this when using valgrind or if we ever
 331          * librarize this command.
 332          *
 333          * Also experiment with obstacks to see how much speed
 334          * up we'll get here.
 335          *
 336          * perf_session__delete(session);
 337          */
 338         return ret;
 339 }
 340
 341 static int
 342 parse_callchain_opt(const struct option *opt __used, const char *arg,
 343                     int unset)
 344 {
 345         char *tok, *tok2;
 346         char *endptr;
 347
 348         /*
 349          * --no-call-graph
 350          */
 351         if (unset) {
 352                 dont_use_callchains = true;
 353                 return 0;
 354         }
 355
 356         symbol_conf.use_callchain = true;
 357
 358         if (!arg)
 359                 return 0;
 360
 361         tok = strtok((char *)arg, ",");
 362         if (!tok)
 363                 return -1;
 364
 365         /* get the output mode */
 366         if (!strncmp(tok, "graph", strlen(arg)))
 367                 callchain_param.mode = CHAIN_GRAPH_ABS;
 368
 369         else if (!strncmp(tok, "flat", strlen(arg)))
 370                 callchain_param.mode = CHAIN_FLAT;
 371
 372         else if (!strncmp(tok, "fractal", strlen(arg)))
 373                 callchain_param.mode = CHAIN_GRAPH_REL;
 374
 375         else if (!strncmp(tok, "none", strlen(arg))) {
 376                 callchain_param.mode = CHAIN_NONE;
 377                 symbol_conf.use_callchain = false;
 378
 379                 return 0;
 380         }
 381
 382         else
 383                 return -1;
 384
 385         /* get the min percentage */
 386         tok = strtok(NULL, ",");
 387         if (!tok)
 388                 goto setup;
 389
 390         callchain_param.min_percent = strtod(tok, &endptr);
 391         if (tok == endptr)
 392                 return -1;
 393
 394         /* get the print limit */
 395         tok2 = strtok(NULL, ",");
 396         if (!tok2)
 397                 goto setup;
 398
 399         if (tok2[0] != 'c') {
 400                 callchain_param.print_limit = strtod(tok2, &endptr);
 401                 tok2 = strtok(NULL, ",");
 402                 if (!tok2)
 403                         goto setup;
 404         }
 405
 406         /* get the call chain order */
 407         if (!strcmp(tok2, "caller"))
 408                 callchain_param.order = ORDER_CALLER;
 409         else if (!strcmp(tok2, "callee"))
 410                 callchain_param.order = ORDER_CALLEE;
 411         else
 412                 return -1;
 413 setup:
 414         if (callchain_register_param(&callchain_param) < 0) {
 415                 fprintf(stderr, "Can't register callchain params\n");
 416                 return -1;
 417         }
 418         return 0;
 419 }
 420
 421 static const char * const report_usage[] = {
 422         "perf report [<options>] <command>",
 423         NULL
 424 };
 425
 426 static const struct option options[] = {
 427         OPT_STRING('i', "input", &input_name, "file",
 428                     "input file name"),
 429         OPT_INCR('v', "verbose", &verbose,
 430                     "be more verbose (show symbol address, etc)"),
 431         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
 432                     "dump raw trace in ASCII"),
 433         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 434                    "file", "vmlinux pathname"),
 435         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
 436                    "file", "kallsyms pathname"),
 437         OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
 438         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
 439                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
 440         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
 441                     "Show a column with the number of samples"),
 442         OPT_BOOLEAN('T', "threads", &show_threads,
 443                     "Show per-thread event counters"),
 444         OPT_STRING(0, "pretty", &pretty_printing_style, "key",
 445                    "pretty printing style key: normal raw"),
 446         OPT_BOOLEAN(0, "tui", &use_tui, "Use the TUI interface"),
 447         OPT_BOOLEAN(0, "stdio", &use_stdio, "Use the stdio interface"),
 448         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
 449                    "sort by key(s): pid, comm, dso, symbol, parent"),
 450         OPT_BOOLEAN(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
 451                     "Show sample percentage for different cpu modes"),
 452         OPT_STRING('p', "parent", &parent_pattern, "regex",
 453                    "regex filter to identify parent, see: '--sort parent'"),
 454         OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
 455                     "Only display entries with parent-match"),
 456         OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent, call_order",
 457                      "Display callchains using output_type (graph, flat, fractal, or none) , min percent threshold and callchain order. "
 458                      "Default: fractal,0.5,callee", &parse_callchain_opt, callchain_default_opt),
 459         OPT_BOOLEAN('G', "inverted", &inverted_callchain, "alias for inverted call graph"),
 460         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
 461                    "only consider symbols in these dsos"),
 462         OPT_STRING('C', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
 463                    "only consider symbols in these comms"),
 464         OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
 465                    "only consider these symbols"),
 466         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
 467                    "width[,width...]",
 468                    "don't try to adjust column width, use these fixed values"),
 469         OPT_STRING('t', "field-separator", &symbol_conf.field_sep, "separator",
 470                    "separator for columns, no spaces will be added between "
 471                    "columns '.' is reserved."),
 472         OPT_BOOLEAN('U', "hide-unresolved", &hide_unresolved,
 473                     "Only display entries resolved to a symbol"),
 474         OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
 475                     "Look for files with symbols relative to this directory"),
 476         OPT_END()
 477 };
 478
 479 int cmd_report(int argc, const char **argv, const char *prefix __used)
 480 {
 481         argc = parse_options(argc, argv, options, report_usage, 0);
 482
 483         if (use_stdio)
 484                 use_browser = 0;
 485         else if (use_tui)
 486                 use_browser = 1;
 487
 488         if (inverted_callchain)
 489                 callchain_param.order = ORDER_CALLER;
 490
 491         if (strcmp(input_name, "-") != 0)
 492                 setup_browser(true);
 493         else
 494                 use_browser = 0;
 495         /*
 496          * Only in the newt browser we are doing integrated annotation,
 497          * so don't allocate extra space that won't be used in the stdio
 498          * implementation.
 499          */
 500         if (use_browser > 0) {
 501                 symbol_conf.priv_size = sizeof(struct annotation);
 502                 annotate_init         = symbol__annotate_init;
 503                 /*
 504                  * For searching by name on the "Browse map details".
 505                  * providing it only in verbose mode not to bloat too
 506                  * much struct symbol.
 507                  */
 508                 if (verbose) {
 509                         /*
 510                          * XXX: Need to provide a less kludgy way to ask for
 511                          * more space per symbol, the u32 is for the index on
 512                          * the ui browser.
 513                          * See symbol__browser_index.
 514                          */
 515                         symbol_conf.priv_size += sizeof(u32);
 516                         symbol_conf.sort_by_name = true;
 517                 }
 518         }
 519
 520         if (symbol__init() < 0)
 521                 return -1;
 522
 523         setup_sorting(report_usage, options);
 524
 525         if (parent_pattern != default_parent_pattern) {
 526                 if (sort_dimension__add("parent") < 0)
 527                         return -1;
 528
 529                 /*
 530                  * Only show the parent fields if we explicitly
 531                  * sort that way. If we only use parent machinery
 532                  * for filtering, we don't want it.
 533                  */
 534                 if (!strstr(sort_order, "parent"))
 535                         sort_parent.elide = 1;
 536         } else
 537                 symbol_conf.exclude_other = false;
 538
 539         /*
 540          * Any (unrecognized) arguments left?
 541          */
 542         if (argc)
 543                 usage_with_options(report_usage, options);
 544
 545         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
 546         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
 547         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
 548
 549         return __cmd_report();
 550 }