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