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 #include <linux/bitmap.h>
  37
  38 static char             const *input_name = "perf.data";
  39
  40 static bool             force, use_tui, use_stdio;
  41 static bool             hide_unresolved;
  42 static bool             dont_use_callchains;
  43 static bool             show_full_info;
  44
  45 static bool             show_threads;
  46 static struct perf_read_values  show_threads_values;
  47
  48 static const char       default_pretty_printing_style[] = "normal";
  49 static const char       *pretty_printing_style = default_pretty_printing_style;
  50
  51 static char             callchain_default_opt[] = "fractal,0.5,callee";
  52 static bool             inverted_callchain;
  53 static symbol_filter_t  annotate_init;
  54
  55 static const char       *cpu_list;
  56 static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  57
  58 static int perf_session__add_hist_entry(struct perf_session *session,
  59                                         struct addr_location *al,
  60                                         struct perf_sample *sample,
  61                                         struct perf_evsel *evsel)
  62 {
  63         struct symbol *parent = NULL;
  64         int err = 0;
  65         struct hist_entry *he;
  66
  67         if ((sort__has_parent || symbol_conf.use_callchain) && sample->callchain) {
  68                 err = perf_session__resolve_callchain(session, al->thread,
  69                                                       sample->callchain, &parent);
  70                 if (err)
  71                         return err;
  72         }
  73
  74         he = __hists__add_entry(&evsel->hists, al, parent, sample->period);
  75         if (he == NULL)
  76                 return -ENOMEM;
  77
  78         if (symbol_conf.use_callchain) {
  79                 err = callchain_append(he->callchain, &session->callchain_cursor,
  80                                        sample->period);
  81                 if (err)
  82                         return err;
  83         }
  84         /*
  85          * Only in the newt browser we are doing integrated annotation,
  86          * so we don't allocated the extra space needed because the stdio
  87          * code will not use it.
  88          */
  89         if (al->sym != NULL && use_browser > 0) {
  90                 struct annotation *notes = symbol__annotation(he->ms.sym);
  91
  92                 assert(evsel != NULL);
  93
  94                 err = -ENOMEM;
  95                 if (notes->src == NULL && symbol__alloc_hist(he->ms.sym) < 0)
  96                         goto out;
  97
  98                 err = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
  99         }
 100
 101         evsel->hists.stats.total_period += sample->period;
 102         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 103 out:
 104         return err;
 105 }
 106
 107
 108 static int process_sample_event(union perf_event *event,
 109                                 struct perf_sample *sample,
 110                                 struct perf_evsel *evsel,
 111                                 struct perf_session *session)
 112 {
 113         struct addr_location al;
 114
 115         if (perf_event__preprocess_sample(event, session, &al, sample,
 116                                           annotate_init) < 0) {
 117                 fprintf(stderr, "problem processing %d event, skipping it.\n",
 118                         event->header.type);
 119                 return -1;
 120         }
 121
 122         if (al.filtered || (hide_unresolved && al.sym == NULL))
 123                 return 0;
 124
 125         if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
 126                 return 0;
 127
 128         if (al.map != NULL)
 129                 al.map->dso->hit = 1;
 130
 131         if (perf_session__add_hist_entry(session, &al, sample, evsel)) {
 132                 pr_debug("problem incrementing symbol period, skipping event\n");
 133                 return -1;
 134         }
 135
 136         return 0;
 137 }
 138
 139 static int process_read_event(union perf_event *event,
 140                               struct perf_sample *sample __used,
 141                               struct perf_session *session)
 142 {
 143         struct perf_evsel *evsel = perf_evlist__id2evsel(session->evlist,
 144                                                          event->read.id);
 145         if (show_threads) {
 146                 const char *name = evsel ? event_name(evsel) : "unknown";
 147                 perf_read_values_add_value(&show_threads_values,
 148                                            event->read.pid, event->read.tid,
 149                                            event->read.id,
 150                                            name,
 151                                            event->read.value);
 152         }
 153
 154         dump_printf(": %d %d %s %" PRIu64 "\n", event->read.pid, event->read.tid,
 155                     evsel ? event_name(evsel) : "FAIL",
 156                     event->read.value);
 157
 158         return 0;
 159 }
 160
 161 static int perf_session__setup_sample_type(struct perf_session *self)
 162 {
 163         if (!(self->sample_type & PERF_SAMPLE_CALLCHAIN)) {
 164                 if (sort__has_parent) {
 165                         ui__warning("Selected --sort parent, but no "
 166                                     "callchain data. Did you call "
 167                                     "'perf record' without -g?\n");
 168                         return -EINVAL;
 169                 }
 170                 if (symbol_conf.use_callchain) {
 171                         ui__warning("Selected -g but no callchain data. Did "
 172                                     "you call 'perf record' without -g?\n");
 173                         return -1;
 174                 }
 175         } else if (!dont_use_callchains && callchain_param.mode != CHAIN_NONE &&
 176                    !symbol_conf.use_callchain) {
 177                         symbol_conf.use_callchain = true;
 178                         if (callchain_register_param(&callchain_param) < 0) {
 179                                 ui__warning("Can't register callchain "
 180                                             "params.\n");
 181                                 return -EINVAL;
 182                         }
 183         }
 184
 185         return 0;
 186 }
 187
 188 static struct perf_event_ops event_ops = {
 189         .sample          = process_sample_event,
 190         .mmap            = perf_event__process_mmap,
 191         .comm            = perf_event__process_comm,
 192         .exit            = perf_event__process_task,
 193         .fork            = perf_event__process_task,
 194         .lost            = perf_event__process_lost,
 195         .read            = process_read_event,
 196         .attr            = perf_event__process_attr,
 197         .event_type      = perf_event__process_event_type,
 198         .tracing_data    = perf_event__process_tracing_data,
 199         .build_id        = perf_event__process_build_id,
 200         .ordered_samples = true,
 201         .ordering_requires_timestamps = true,
 202 };
 203
 204 extern volatile int session_done;
 205
 206 static void sig_handler(int sig __used)
 207 {
 208         session_done = 1;
 209 }
 210
 211 static size_t hists__fprintf_nr_sample_events(struct hists *self,
 212                                               const char *evname, FILE *fp)
 213 {
 214         size_t ret;
 215         char unit;
 216         unsigned long nr_events = self->stats.nr_events[PERF_RECORD_SAMPLE];
 217
 218         nr_events = convert_unit(nr_events, &unit);
 219         ret = fprintf(fp, "# Events: %lu%c", nr_events, unit);
 220         if (evname != NULL)
 221                 ret += fprintf(fp, " %s", evname);
 222         return ret + fprintf(fp, "\n#\n");
 223 }
 224
 225 static int perf_evlist__tty_browse_hists(struct perf_evlist *evlist,
 226                                          const char *help)
 227 {
 228         struct perf_evsel *pos;
 229
 230         list_for_each_entry(pos, &evlist->entries, node) {
 231                 struct hists *hists = &pos->hists;
 232                 const char *evname = event_name(pos);
 233
 234                 hists__fprintf_nr_sample_events(hists, evname, stdout);
 235                 hists__fprintf(hists, NULL, false, true, 0, 0, stdout);
 236                 fprintf(stdout, "\n\n");
 237         }
 238
 239         if (sort_order == default_sort_order &&
 240             parent_pattern == default_parent_pattern) {
 241                 fprintf(stdout, "#\n# (%s)\n#\n", help);
 242
 243                 if (show_threads) {
 244                         bool style = !strcmp(pretty_printing_style, "raw");
 245                         perf_read_values_display(stdout, &show_threads_values,
 246                                                  style);
 247                         perf_read_values_destroy(&show_threads_values);
 248                 }
 249         }
 250
 251         return 0;
 252 }
 253
 254 static int __cmd_report(void)
 255 {
 256         int ret = -EINVAL;
 257         u64 nr_samples;
 258         struct perf_session *session;
 259         struct perf_evsel *pos;
 260         struct map *kernel_map;
 261         struct kmap *kernel_kmap;
 262         const char *help = "For a higher level overview, try: perf report --sort comm,dso";
 263
 264         signal(SIGINT, sig_handler);
 265
 266         session = perf_session__new(input_name, O_RDONLY, force, false, &event_ops);
 267         if (session == NULL)
 268                 return -ENOMEM;
 269
 270         if (cpu_list) {
 271                 ret = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
 272                 if (ret)
 273                         goto out_delete;
 274         }
 275
 276         if (use_browser <= 0)
 277                 perf_session__fprintf_info(session, stdout, show_full_info);
 278
 279         if (show_threads)
 280                 perf_read_values_init(&show_threads_values);
 281
 282         ret = perf_session__setup_sample_type(session);
 283         if (ret)
 284                 goto out_delete;
 285
 286         ret = perf_session__process_events(session, &event_ops);
 287         if (ret)
 288                 goto out_delete;
 289
 290         kernel_map = session->host_machine.vmlinux_maps[MAP__FUNCTION];
 291         kernel_kmap = map__kmap(kernel_map);
 292         if (kernel_map == NULL ||
 293             (kernel_map->dso->hit &&
 294              (kernel_kmap->ref_reloc_sym == NULL ||
 295               kernel_kmap->ref_reloc_sym->addr == 0))) {
 296                 const struct dso *kdso = kernel_map->dso;
 297
 298                 ui__warning(
 299 "Kernel address maps (/proc/{kallsyms,modules}) were restricted.\n\n"
 300 "Check /proc/sys/kernel/kptr_restrict before running 'perf record'.\n\n%s\n\n"
 301 "Samples in kernel modules can't be resolved as well.\n\n",
 302                             RB_EMPTY_ROOT(&kdso->symbols[MAP__FUNCTION]) ?
 303 "As no suitable kallsyms nor vmlinux was found, kernel samples\n"
 304 "can't be resolved." :
 305 "If some relocation was applied (e.g. kexec) symbols may be misresolved.");
 306         }
 307
 308         if (dump_trace) {
 309                 perf_session__fprintf_nr_events(session, stdout);
 310                 goto out_delete;
 311         }
 312
 313         if (verbose > 3)
 314                 perf_session__fprintf(session, stdout);
 315
 316         if (verbose > 2)
 317                 perf_session__fprintf_dsos(session, stdout);
 318
 319         nr_samples = 0;
 320         list_for_each_entry(pos, &session->evlist->entries, node) {
 321                 struct hists *hists = &pos->hists;
 322
 323                 hists__collapse_resort(hists);
 324                 hists__output_resort(hists);
 325                 nr_samples += hists->stats.nr_events[PERF_RECORD_SAMPLE];
 326         }
 327
 328         if (nr_samples == 0) {
 329                 ui__warning("The %s file has no samples!\n", input_name);
 330                 goto out_delete;
 331         }
 332
 333         if (use_browser > 0) {
 334                 perf_evlist__tui_browse_hists(session->evlist, help,
 335                                               NULL, NULL, 0);
 336         } else
 337                 perf_evlist__tty_browse_hists(session->evlist, help);
 338
 339 out_delete:
 340         /*
 341          * Speed up the exit process, for large files this can
 342          * take quite a while.
 343          *
 344          * XXX Enable this when using valgrind or if we ever
 345          * librarize this command.
 346          *
 347          * Also experiment with obstacks to see how much speed
 348          * up we'll get here.
 349          *
 350          * perf_session__delete(session);
 351          */
 352         return ret;
 353 }
 354
 355 static int
 356 parse_callchain_opt(const struct option *opt __used, const char *arg,
 357                     int unset)
 358 {
 359         char *tok, *tok2;
 360         char *endptr;
 361
 362         /*
 363          * --no-call-graph
 364          */
 365         if (unset) {
 366                 dont_use_callchains = true;
 367                 return 0;
 368         }
 369
 370         symbol_conf.use_callchain = true;
 371
 372         if (!arg)
 373                 return 0;
 374
 375         tok = strtok((char *)arg, ",");
 376         if (!tok)
 377                 return -1;
 378
 379         /* get the output mode */
 380         if (!strncmp(tok, "graph", strlen(arg)))
 381                 callchain_param.mode = CHAIN_GRAPH_ABS;
 382
 383         else if (!strncmp(tok, "flat", strlen(arg)))
 384                 callchain_param.mode = CHAIN_FLAT;
 385
 386         else if (!strncmp(tok, "fractal", strlen(arg)))
 387                 callchain_param.mode = CHAIN_GRAPH_REL;
 388
 389         else if (!strncmp(tok, "none", strlen(arg))) {
 390                 callchain_param.mode = CHAIN_NONE;
 391                 symbol_conf.use_callchain = false;
 392
 393                 return 0;
 394         }
 395
 396         else
 397                 return -1;
 398
 399         /* get the min percentage */
 400         tok = strtok(NULL, ",");
 401         if (!tok)
 402                 goto setup;
 403
 404         callchain_param.min_percent = strtod(tok, &endptr);
 405         if (tok == endptr)
 406                 return -1;
 407
 408         /* get the print limit */
 409         tok2 = strtok(NULL, ",");
 410         if (!tok2)
 411                 goto setup;
 412
 413         if (tok2[0] != 'c') {
 414                 callchain_param.print_limit = strtod(tok2, &endptr);
 415                 tok2 = strtok(NULL, ",");
 416                 if (!tok2)
 417                         goto setup;
 418         }
 419
 420         /* get the call chain order */
 421         if (!strcmp(tok2, "caller"))
 422                 callchain_param.order = ORDER_CALLER;
 423         else if (!strcmp(tok2, "callee"))
 424                 callchain_param.order = ORDER_CALLEE;
 425         else
 426                 return -1;
 427 setup:
 428         if (callchain_register_param(&callchain_param) < 0) {
 429                 fprintf(stderr, "Can't register callchain params\n");
 430                 return -1;
 431         }
 432         return 0;
 433 }
 434
 435 static const char * const report_usage[] = {
 436         "perf report [<options>] <command>",
 437         NULL
 438 };
 439
 440 static const struct option options[] = {
 441         OPT_STRING('i', "input", &input_name, "file",
 442                     "input file name"),
 443         OPT_INCR('v', "verbose", &verbose,
 444                     "be more verbose (show symbol address, etc)"),
 445         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
 446                     "dump raw trace in ASCII"),
 447         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 448                    "file", "vmlinux pathname"),
 449         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
 450                    "file", "kallsyms pathname"),
 451         OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
 452         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
 453                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
 454         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
 455                     "Show a column with the number of samples"),
 456         OPT_BOOLEAN('T', "threads", &show_threads,
 457                     "Show per-thread event counters"),
 458         OPT_STRING(0, "pretty", &pretty_printing_style, "key",
 459                    "pretty printing style key: normal raw"),
 460         OPT_BOOLEAN(0, "tui", &use_tui, "Use the TUI interface"),
 461         OPT_BOOLEAN(0, "stdio", &use_stdio, "Use the stdio interface"),
 462         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
 463                    "sort by key(s): pid, comm, dso, symbol, parent"),
 464         OPT_BOOLEAN(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
 465                     "Show sample percentage for different cpu modes"),
 466         OPT_STRING('p', "parent", &parent_pattern, "regex",
 467                    "regex filter to identify parent, see: '--sort parent'"),
 468         OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
 469                     "Only display entries with parent-match"),
 470         OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent, call_order",
 471                      "Display callchains using output_type (graph, flat, fractal, or none) , min percent threshold and callchain order. "
 472                      "Default: fractal,0.5,callee", &parse_callchain_opt, callchain_default_opt),
 473         OPT_BOOLEAN('G', "inverted", &inverted_callchain, "alias for inverted call graph"),
 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_STRING('c', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
 491         OPT_BOOLEAN('I', "show-info", &show_full_info,
 492                     "Display extended information about perf.data file"),
 493         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
 494                     "Interleave source code with assembly code (default)"),
 495         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
 496                     "Display raw encoding of assembly instructions (default)"),
 497         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
 498                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
 499         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
 500                     "Show a column with the sum of periods"),
 501         OPT_END()
 502 };
 503
 504 int cmd_report(int argc, const char **argv, const char *prefix __used)
 505 {
 506         argc = parse_options(argc, argv, options, report_usage, 0);
 507
 508         if (use_stdio)
 509                 use_browser = 0;
 510         else if (use_tui)
 511                 use_browser = 1;
 512
 513         if (inverted_callchain)
 514                 callchain_param.order = ORDER_CALLER;
 515
 516         if (strcmp(input_name, "-") != 0)
 517                 setup_browser(true);
 518         else
 519                 use_browser = 0;
 520         /*
 521          * Only in the newt browser we are doing integrated annotation,
 522          * so don't allocate extra space that won't be used in the stdio
 523          * implementation.
 524          */
 525         if (use_browser > 0) {
 526                 symbol_conf.priv_size = sizeof(struct annotation);
 527                 annotate_init         = symbol__annotate_init;
 528                 /*
 529                  * For searching by name on the "Browse map details".
 530                  * providing it only in verbose mode not to bloat too
 531                  * much struct symbol.
 532                  */
 533                 if (verbose) {
 534                         /*
 535                          * XXX: Need to provide a less kludgy way to ask for
 536                          * more space per symbol, the u32 is for the index on
 537                          * the ui browser.
 538                          * See symbol__browser_index.
 539                          */
 540                         symbol_conf.priv_size += sizeof(u32);
 541                         symbol_conf.sort_by_name = true;
 542                 }
 543         }
 544
 545         if (symbol__init() < 0)
 546                 return -1;
 547
 548         setup_sorting(report_usage, options);
 549
 550         if (parent_pattern != default_parent_pattern) {
 551                 if (sort_dimension__add("parent") < 0)
 552                         return -1;
 553
 554                 /*
 555                  * Only show the parent fields if we explicitly
 556                  * sort that way. If we only use parent machinery
 557                  * for filtering, we don't want it.
 558                  */
 559                 if (!strstr(sort_order, "parent"))
 560                         sort_parent.elide = 1;
 561         } else
 562                 symbol_conf.exclude_other = false;
 563
 564         /*
 565          * Any (unrecognized) arguments left?
 566          */
 567         if (argc)
 568                 usage_with_options(report_usage, options);
 569
 570         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
 571         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
 572         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
 573
 574         return __cmd_report();
 575 }