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/string.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 int              force;
  37 static bool             hide_unresolved;
  38 static bool             dont_use_callchains;
  39
  40 static int              show_threads;
  41 static struct perf_read_values  show_threads_values;
  42
  43 static char             default_pretty_printing_style[] = "normal";
  44 static char             *pretty_printing_style = default_pretty_printing_style;
  45
  46 static char             callchain_default_opt[] = "fractal,0.5";
  47
  48 static struct event_stat_id *get_stats(struct perf_session *self,
  49                                        u64 event_stream, u32 type, u64 config)
  50 {
  51         struct rb_node **p = &self->stats_by_id.rb_node;
  52         struct rb_node *parent = NULL;
  53         struct event_stat_id *iter, *new;
  54
  55         while (*p != NULL) {
  56                 parent = *p;
  57                 iter = rb_entry(parent, struct event_stat_id, 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 event_stat_id));
  69         if (new == NULL)
  70                 return NULL;
  71         memset(new, 0, sizeof(struct event_stat_id));
  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->stats_by_id);
  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 symbol **syms = NULL, *parent = NULL;
  85         bool hit;
  86         int err;
  87         struct hist_entry *he;
  88         struct event_stat_id *stats;
  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
  95         attr = perf_header__find_attr(data->id, &self->header);
  96         if (attr)
  97                 stats = get_stats(self, data->id, attr->type, attr->config);
  98         else
  99                 stats = get_stats(self, data->id, 0, 0);
 100         if (stats == NULL)
 101                 return -ENOMEM;
 102         he = __perf_session__add_hist_entry(&stats->hists, al, parent,
 103                                             data->period, &hit);
 104         if (he == NULL)
 105                 return -ENOMEM;
 106
 107         if (hit)
 108                 he->count += data->period;
 109
 110         if (symbol_conf.use_callchain) {
 111                 if (!hit)
 112                         callchain_init(&he->callchain);
 113                 err = append_chain(&he->callchain, data->callchain, syms);
 114                 free(syms);
 115
 116                 if (err)
 117                         return err;
 118         }
 119
 120         return 0;
 121 }
 122
 123 static int validate_chain(struct ip_callchain *chain, event_t *event)
 124 {
 125         unsigned int chain_size;
 126
 127         chain_size = event->header.size;
 128         chain_size -= (unsigned long)&event->ip.__more_data - (unsigned long)event;
 129
 130         if (chain->nr*sizeof(u64) > chain_size)
 131                 return -1;
 132
 133         return 0;
 134 }
 135
 136 static int add_event_total(struct perf_session *session,
 137                            struct sample_data *data,
 138                            struct perf_event_attr *attr)
 139 {
 140         struct event_stat_id *stats;
 141
 142         if (attr)
 143                 stats = get_stats(session, data->id, attr->type, attr->config);
 144         else
 145                 stats = get_stats(session, data->id, 0, 0);
 146
 147         if (!stats)
 148                 return -ENOMEM;
 149
 150         stats->stats.total += data->period;
 151         session->events_stats.total += data->period;
 152         return 0;
 153 }
 154
 155 static int process_sample_event(event_t *event, struct perf_session *session)
 156 {
 157         struct sample_data data = { .period = 1, };
 158         struct addr_location al;
 159         struct perf_event_attr *attr;
 160
 161         event__parse_sample(event, session->sample_type, &data);
 162
 163         dump_printf("(IP, %d): %d/%d: %#Lx period: %Ld\n", event->header.misc,
 164                     data.pid, data.tid, data.ip, data.period);
 165
 166         if (session->sample_type & PERF_SAMPLE_CALLCHAIN) {
 167                 unsigned int i;
 168
 169                 dump_printf("... chain: nr:%Lu\n", data.callchain->nr);
 170
 171                 if (validate_chain(data.callchain, event) < 0) {
 172                         pr_debug("call-chain problem with event, "
 173                                  "skipping it.\n");
 174                         return 0;
 175                 }
 176
 177                 if (dump_trace) {
 178                         for (i = 0; i < data.callchain->nr; i++)
 179                                 dump_printf("..... %2d: %016Lx\n",
 180                                             i, data.callchain->ips[i]);
 181                 }
 182         }
 183
 184         if (event__preprocess_sample(event, session, &al, NULL) < 0) {
 185                 fprintf(stderr, "problem processing %d event, skipping it.\n",
 186                         event->header.type);
 187                 return -1;
 188         }
 189
 190         if (al.filtered || (hide_unresolved && al.sym == NULL))
 191                 return 0;
 192
 193         if (perf_session__add_hist_entry(session, &al, &data)) {
 194                 pr_debug("problem incrementing symbol count, skipping event\n");
 195                 return -1;
 196         }
 197
 198         attr = perf_header__find_attr(data.id, &session->header);
 199
 200         if (add_event_total(session, &data, attr)) {
 201                 pr_debug("problem adding event count\n");
 202                 return -1;
 203         }
 204
 205         return 0;
 206 }
 207
 208 static int process_read_event(event_t *event, struct perf_session *session __used)
 209 {
 210         struct perf_event_attr *attr;
 211
 212         attr = perf_header__find_attr(event->read.id, &session->header);
 213
 214         if (show_threads) {
 215                 const char *name = attr ? __event_name(attr->type, attr->config)
 216                                    : "unknown";
 217                 perf_read_values_add_value(&show_threads_values,
 218                                            event->read.pid, event->read.tid,
 219                                            event->read.id,
 220                                            name,
 221                                            event->read.value);
 222         }
 223
 224         dump_printf(": %d %d %s %Lu\n", event->read.pid, event->read.tid,
 225                     attr ? __event_name(attr->type, attr->config) : "FAIL",
 226                     event->read.value);
 227
 228         return 0;
 229 }
 230
 231 static int perf_session__setup_sample_type(struct perf_session *self)
 232 {
 233         if (!(self->sample_type & PERF_SAMPLE_CALLCHAIN)) {
 234                 if (sort__has_parent) {
 235                         fprintf(stderr, "selected --sort parent, but no"
 236                                         " callchain data. Did you call"
 237                                         " perf record without -g?\n");
 238                         return -EINVAL;
 239                 }
 240                 if (symbol_conf.use_callchain) {
 241                         fprintf(stderr, "selected -g but no callchain data."
 242                                         " Did you call perf record without"
 243                                         " -g?\n");
 244                         return -1;
 245                 }
 246         } else if (!dont_use_callchains && callchain_param.mode != CHAIN_NONE &&
 247                    !symbol_conf.use_callchain) {
 248                         symbol_conf.use_callchain = true;
 249                         if (register_callchain_param(&callchain_param) < 0) {
 250                                 fprintf(stderr, "Can't register callchain"
 251                                                 " params\n");
 252                                 return -EINVAL;
 253                         }
 254         }
 255
 256         return 0;
 257 }
 258
 259 static struct perf_event_ops event_ops = {
 260         .sample = process_sample_event,
 261         .mmap   = event__process_mmap,
 262         .comm   = event__process_comm,
 263         .exit   = event__process_task,
 264         .fork   = event__process_task,
 265         .lost   = event__process_lost,
 266         .read   = process_read_event,
 267 };
 268
 269 static int __cmd_report(void)
 270 {
 271         int ret = -EINVAL;
 272         struct perf_session *session;
 273         struct rb_node *next;
 274         const char *help = "For a higher level overview, try: perf report --sort comm,dso";
 275
 276         session = perf_session__new(input_name, O_RDONLY, force);
 277         if (session == NULL)
 278                 return -ENOMEM;
 279
 280         if (show_threads)
 281                 perf_read_values_init(&show_threads_values);
 282
 283         ret = perf_session__setup_sample_type(session);
 284         if (ret)
 285                 goto out_delete;
 286
 287         ret = perf_session__process_events(session, &event_ops);
 288         if (ret)
 289                 goto out_delete;
 290
 291         if (dump_trace) {
 292                 event__print_totals();
 293                 goto out_delete;
 294         }
 295
 296         if (verbose > 3)
 297                 perf_session__fprintf(session, stdout);
 298
 299         if (verbose > 2)
 300                 dsos__fprintf(stdout);
 301
 302         next = rb_first(&session->stats_by_id);
 303         while (next) {
 304                 struct event_stat_id *stats;
 305
 306                 stats = rb_entry(next, struct event_stat_id, rb_node);
 307                 perf_session__collapse_resort(&stats->hists);
 308                 perf_session__output_resort(&stats->hists, stats->stats.total);
 309
 310                 if (use_browser)
 311                         perf_session__browse_hists(&stats->hists,
 312                                                    stats->stats.total, help);
 313                 else {
 314                         if (rb_first(&session->stats_by_id) ==
 315                             rb_last(&session->stats_by_id))
 316                                 fprintf(stdout, "# Samples: %Ld\n#\n",
 317                                         stats->stats.total);
 318                         else
 319                                 fprintf(stdout, "# Samples: %Ld %s\n#\n",
 320                                         stats->stats.total,
 321                                         __event_name(stats->type, stats->config));
 322
 323                         perf_session__fprintf_hists(&stats->hists, NULL, false, stdout,
 324                                             stats->stats.total);
 325                         fprintf(stdout, "\n\n");
 326                 }
 327
 328                 next = rb_next(&stats->rb_node);
 329         }
 330
 331         if (!use_browser && sort_order == default_sort_order &&
 332             parent_pattern == default_parent_pattern) {
 333                 fprintf(stdout, "#\n# (%s)\n#\n", help);
 334
 335                 if (show_threads) {
 336                         bool style = !strcmp(pretty_printing_style, "raw");
 337                         perf_read_values_display(stdout, &show_threads_values,
 338                                                  style);
 339                         perf_read_values_destroy(&show_threads_values);
 340                 }
 341         }
 342 out_delete:
 343         perf_session__delete(session);
 344         return ret;
 345 }
 346
 347 static int
 348 parse_callchain_opt(const struct option *opt __used, const char *arg,
 349                     int unset)
 350 {
 351         char *tok;
 352         char *endptr;
 353
 354         /*
 355          * --no-call-graph
 356          */
 357         if (unset) {
 358                 dont_use_callchains = true;
 359                 return 0;
 360         }
 361
 362         symbol_conf.use_callchain = true;
 363
 364         if (!arg)
 365                 return 0;
 366
 367         tok = strtok((char *)arg, ",");
 368         if (!tok)
 369                 return -1;
 370
 371         /* get the output mode */
 372         if (!strncmp(tok, "graph", strlen(arg)))
 373                 callchain_param.mode = CHAIN_GRAPH_ABS;
 374
 375         else if (!strncmp(tok, "flat", strlen(arg)))
 376                 callchain_param.mode = CHAIN_FLAT;
 377
 378         else if (!strncmp(tok, "fractal", strlen(arg)))
 379                 callchain_param.mode = CHAIN_GRAPH_REL;
 380
 381         else if (!strncmp(tok, "none", strlen(arg))) {
 382                 callchain_param.mode = CHAIN_NONE;
 383                 symbol_conf.use_callchain = false;
 384
 385                 return 0;
 386         }
 387
 388         else
 389                 return -1;
 390
 391         /* get the min percentage */
 392         tok = strtok(NULL, ",");
 393         if (!tok)
 394                 goto setup;
 395
 396         callchain_param.min_percent = strtod(tok, &endptr);
 397         if (tok == endptr)
 398                 return -1;
 399
 400 setup:
 401         if (register_callchain_param(&callchain_param) < 0) {
 402                 fprintf(stderr, "Can't register callchain params\n");
 403                 return -1;
 404         }
 405         return 0;
 406 }
 407
 408 static const char * const report_usage[] = {
 409         "perf report [<options>] <command>",
 410         NULL
 411 };
 412
 413 static const struct option options[] = {
 414         OPT_STRING('i', "input", &input_name, "file",
 415                     "input file name"),
 416         OPT_BOOLEAN('v', "verbose", &verbose,
 417                     "be more verbose (show symbol address, etc)"),
 418         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
 419                     "dump raw trace in ASCII"),
 420         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 421                    "file", "vmlinux pathname"),
 422         OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
 423         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
 424                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
 425         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
 426                     "Show a column with the number of samples"),
 427         OPT_BOOLEAN('T', "threads", &show_threads,
 428                     "Show per-thread event counters"),
 429         OPT_STRING(0, "pretty", &pretty_printing_style, "key",
 430                    "pretty printing style key: normal raw"),
 431         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
 432                    "sort by key(s): pid, comm, dso, symbol, parent"),
 433         OPT_BOOLEAN('P', "full-paths", &symbol_conf.full_paths,
 434                     "Don't shorten the pathnames taking into account the cwd"),
 435         OPT_STRING('p', "parent", &parent_pattern, "regex",
 436                    "regex filter to identify parent, see: '--sort parent'"),
 437         OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
 438                     "Only display entries with parent-match"),
 439         OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent",
 440                      "Display callchains using output_type and min percent threshold. "
 441                      "Default: fractal,0.5", &parse_callchain_opt, callchain_default_opt),
 442         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
 443                    "only consider symbols in these dsos"),
 444         OPT_STRING('C', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
 445                    "only consider symbols in these comms"),
 446         OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
 447                    "only consider these symbols"),
 448         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
 449                    "width[,width...]",
 450                    "don't try to adjust column width, use these fixed values"),
 451         OPT_STRING('t', "field-separator", &symbol_conf.field_sep, "separator",
 452                    "separator for columns, no spaces will be added between "
 453                    "columns '.' is reserved."),
 454         OPT_BOOLEAN('U', "hide-unresolved", &hide_unresolved,
 455                     "Only display entries resolved to a symbol"),
 456         OPT_END()
 457 };
 458
 459 int cmd_report(int argc, const char **argv, const char *prefix __used)
 460 {
 461         argc = parse_options(argc, argv, options, report_usage, 0);
 462
 463         setup_browser();
 464
 465         if (symbol__init() < 0)
 466                 return -1;
 467
 468         setup_sorting(report_usage, options);
 469
 470         if (parent_pattern != default_parent_pattern) {
 471                 sort_dimension__add("parent");
 472                 sort_parent.elide = 1;
 473         } else
 474                 symbol_conf.exclude_other = false;
 475
 476         /*
 477          * Any (unrecognized) arguments left?
 478          */
 479         if (argc)
 480                 usage_with_options(report_usage, options);
 481
 482         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
 483         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
 484         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
 485
 486         return __cmd_report();
 487 }