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 #include "util/cache.h"
  12
  13 #include "util/annotate.h"
  14 #include "util/color.h"
  15 #include <linux/list.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 #include "util/tool.h"
  29
  30 #include "util/parse-options.h"
  31 #include "util/parse-events.h"
  32
  33 #include "util/thread.h"
  34 #include "util/sort.h"
  35 #include "util/hist.h"
  36 #include "util/data.h"
  37 #include "arch/common.h"
  38
  39 #include <dlfcn.h>
  40 #include <linux/bitmap.h>
  41
  42 struct report {
  43         struct perf_tool        tool;
  44         struct perf_session     *session;
  45         bool                    force, use_tui, use_gtk, use_stdio;
  46         bool                    hide_unresolved;
  47         bool                    dont_use_callchains;
  48         bool                    show_full_info;
  49         bool                    show_threads;
  50         bool                    inverted_callchain;
  51         bool                    mem_mode;
  52         bool                    header;
  53         bool                    header_only;
  54         int                     max_stack;
  55         struct perf_read_values show_threads_values;
  56         const char              *pretty_printing_style;
  57         const char              *cpu_list;
  58         const char              *symbol_filter_str;
  59         float                   min_percent;
  60         DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  61 };
  62
  63 static int report__config(const char *var, const char *value, void *cb)
  64 {
  65         if (!strcmp(var, "report.group")) {
  66                 symbol_conf.event_group = perf_config_bool(var, value);
  67                 return 0;
  68         }
  69         if (!strcmp(var, "report.percent-limit")) {
  70                 struct report *rep = cb;
  71                 rep->min_percent = strtof(value, NULL);
  72                 return 0;
  73         }
  74
  75         return perf_default_config(var, value, cb);
  76 }
  77
  78 static int report__resolve_callchain(struct report *rep, struct symbol **parent,
  79                                      struct perf_evsel *evsel, struct addr_location *al,
  80                                      struct perf_sample *sample)
  81 {
  82         if ((sort__has_parent || symbol_conf.use_callchain) && sample->callchain) {
  83                 return machine__resolve_callchain(al->machine, evsel, al->thread, sample,
  84                                                   parent, al, rep->max_stack);
  85         }
  86         return 0;
  87 }
  88
  89 static int hist_entry__append_callchain(struct hist_entry *he, struct perf_sample *sample)
  90 {
  91         if (!symbol_conf.use_callchain)
  92                 return 0;
  93         return callchain_append(he->callchain, &callchain_cursor, sample->period);
  94 }
  95
  96 static int report__add_mem_hist_entry(struct perf_tool *tool, struct addr_location *al,
  97                                       struct perf_sample *sample, struct perf_evsel *evsel,
  98                                       union perf_event *event)
  99 {
 100         struct report *rep = container_of(tool, struct report, tool);
 101         struct symbol *parent = NULL;
 102         u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 103         struct hist_entry *he;
 104         struct mem_info *mi, *mx;
 105         uint64_t cost;
 106         int err = report__resolve_callchain(rep, &parent, evsel, al, sample);
 107
 108         if (err)
 109                 return err;
 110
 111         mi = machine__resolve_mem(al->machine, al->thread, sample, cpumode);
 112         if (!mi)
 113                 return -ENOMEM;
 114
 115         if (rep->hide_unresolved && !al->sym)
 116                 return 0;
 117
 118         cost = sample->weight;
 119         if (!cost)
 120                 cost = 1;
 121
 122         /*
 123          * must pass period=weight in order to get the correct
 124          * sorting from hists__collapse_resort() which is solely
 125          * based on periods. We want sorting be done on nr_events * weight
 126          * and this is indirectly achieved by passing period=weight here
 127          * and the he_stat__add_period() function.
 128          */
 129         he = __hists__add_entry(&evsel->hists, al, parent, NULL, mi,
 130                                 cost, cost, 0);
 131         if (!he)
 132                 return -ENOMEM;
 133
 134         err = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
 135         if (err)
 136                 goto out;
 137
 138         mx = he->mem_info;
 139         err = addr_map_symbol__inc_samples(&mx->daddr, evsel->idx);
 140         if (err)
 141                 goto out;
 142
 143         evsel->hists.stats.total_period += cost;
 144         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 145         err = hist_entry__append_callchain(he, sample);
 146 out:
 147         return err;
 148 }
 149
 150 static int report__add_branch_hist_entry(struct perf_tool *tool, struct addr_location *al,
 151                                          struct perf_sample *sample, struct perf_evsel *evsel)
 152 {
 153         struct report *rep = container_of(tool, struct report, tool);
 154         struct symbol *parent = NULL;
 155         unsigned i;
 156         struct hist_entry *he;
 157         struct branch_info *bi, *bx;
 158         int err = report__resolve_callchain(rep, &parent, evsel, al, sample);
 159
 160         if (err)
 161                 return err;
 162
 163         bi = machine__resolve_bstack(al->machine, al->thread,
 164                                      sample->branch_stack);
 165         if (!bi)
 166                 return -ENOMEM;
 167
 168         for (i = 0; i < sample->branch_stack->nr; i++) {
 169                 if (rep->hide_unresolved && !(bi[i].from.sym && bi[i].to.sym))
 170                         continue;
 171
 172                 err = -ENOMEM;
 173
 174                 /* overwrite the 'al' to branch-to info */
 175                 al->map = bi[i].to.map;
 176                 al->sym = bi[i].to.sym;
 177                 al->addr = bi[i].to.addr;
 178                 /*
 179                  * The report shows the percentage of total branches captured
 180                  * and not events sampled. Thus we use a pseudo period of 1.
 181                  */
 182                 he = __hists__add_entry(&evsel->hists, al, parent, &bi[i], NULL,
 183                                         1, 1, 0);
 184                 if (he) {
 185                         bx = he->branch_info;
 186                         err = addr_map_symbol__inc_samples(&bx->from, evsel->idx);
 187                         if (err)
 188                                 goto out;
 189
 190                         err = addr_map_symbol__inc_samples(&bx->to, evsel->idx);
 191                         if (err)
 192                                 goto out;
 193
 194                         evsel->hists.stats.total_period += 1;
 195                         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 196                 } else
 197                         goto out;
 198         }
 199         err = 0;
 200 out:
 201         free(bi);
 202         return err;
 203 }
 204
 205 static int report__add_hist_entry(struct perf_tool *tool, struct perf_evsel *evsel,
 206                                   struct addr_location *al, struct perf_sample *sample)
 207 {
 208         struct report *rep = container_of(tool, struct report, tool);
 209         struct symbol *parent = NULL;
 210         struct hist_entry *he;
 211         int err = report__resolve_callchain(rep, &parent, evsel, al, sample);
 212
 213         if (err)
 214                 return err;
 215
 216         he = __hists__add_entry(&evsel->hists, al, parent, NULL, NULL,
 217                                 sample->period, sample->weight,
 218                                 sample->transaction);
 219         if (he == NULL)
 220                 return -ENOMEM;
 221
 222         err = hist_entry__append_callchain(he, sample);
 223         if (err)
 224                 goto out;
 225
 226         err = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
 227         evsel->hists.stats.total_period += sample->period;
 228         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 229 out:
 230         return err;
 231 }
 232
 233
 234 static int process_sample_event(struct perf_tool *tool,
 235                                 union perf_event *event,
 236                                 struct perf_sample *sample,
 237                                 struct perf_evsel *evsel,
 238                                 struct machine *machine)
 239 {
 240         struct report *rep = container_of(tool, struct report, tool);
 241         struct addr_location al;
 242         int ret;
 243
 244         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0) {
 245                 pr_debug("problem processing %d event, skipping it.\n",
 246                          event->header.type);
 247                 return -1;
 248         }
 249
 250         if (al.filtered || (rep->hide_unresolved && al.sym == NULL))
 251                 return 0;
 252
 253         if (rep->cpu_list && !test_bit(sample->cpu, rep->cpu_bitmap))
 254                 return 0;
 255
 256         if (sort__mode == SORT_MODE__BRANCH) {
 257                 ret = report__add_branch_hist_entry(tool, &al, sample, evsel);
 258                 if (ret < 0)
 259                         pr_debug("problem adding lbr entry, skipping event\n");
 260         } else if (rep->mem_mode == 1) {
 261                 ret = report__add_mem_hist_entry(tool, &al, sample, evsel, event);
 262                 if (ret < 0)
 263                         pr_debug("problem adding mem entry, skipping event\n");
 264         } else {
 265                 if (al.map != NULL)
 266                         al.map->dso->hit = 1;
 267
 268                 ret = report__add_hist_entry(tool, evsel, &al, sample);
 269                 if (ret < 0)
 270                         pr_debug("problem incrementing symbol period, skipping event\n");
 271         }
 272         return ret;
 273 }
 274
 275 static int process_read_event(struct perf_tool *tool,
 276                               union perf_event *event,
 277                               struct perf_sample *sample __maybe_unused,
 278                               struct perf_evsel *evsel,
 279                               struct machine *machine __maybe_unused)
 280 {
 281         struct report *rep = container_of(tool, struct report, tool);
 282
 283         if (rep->show_threads) {
 284                 const char *name = evsel ? perf_evsel__name(evsel) : "unknown";
 285                 perf_read_values_add_value(&rep->show_threads_values,
 286                                            event->read.pid, event->read.tid,
 287                                            event->read.id,
 288                                            name,
 289                                            event->read.value);
 290         }
 291
 292         dump_printf(": %d %d %s %" PRIu64 "\n", event->read.pid, event->read.tid,
 293                     evsel ? perf_evsel__name(evsel) : "FAIL",
 294                     event->read.value);
 295
 296         return 0;
 297 }
 298
 299 /* For pipe mode, sample_type is not currently set */
 300 static int report__setup_sample_type(struct report *rep)
 301 {
 302         struct perf_session *session = rep->session;
 303         u64 sample_type = perf_evlist__combined_sample_type(session->evlist);
 304         bool is_pipe = perf_data_file__is_pipe(session->file);
 305
 306         if (!is_pipe && !(sample_type & PERF_SAMPLE_CALLCHAIN)) {
 307                 if (sort__has_parent) {
 308                         ui__error("Selected --sort parent, but no "
 309                                     "callchain data. Did you call "
 310                                     "'perf record' without -g?\n");
 311                         return -EINVAL;
 312                 }
 313                 if (symbol_conf.use_callchain) {
 314                         ui__error("Selected -g but no callchain data. Did "
 315                                     "you call 'perf record' without -g?\n");
 316                         return -1;
 317                 }
 318         } else if (!rep->dont_use_callchains &&
 319                    callchain_param.mode != CHAIN_NONE &&
 320                    !symbol_conf.use_callchain) {
 321                         symbol_conf.use_callchain = true;
 322                         if (callchain_register_param(&callchain_param) < 0) {
 323                                 ui__error("Can't register callchain params.\n");
 324                                 return -EINVAL;
 325                         }
 326         }
 327
 328         if (sort__mode == SORT_MODE__BRANCH) {
 329                 if (!is_pipe &&
 330                     !(sample_type & PERF_SAMPLE_BRANCH_STACK)) {
 331                         ui__error("Selected -b but no branch data. "
 332                                   "Did you call perf record without -b?\n");
 333                         return -1;
 334                 }
 335         }
 336
 337         return 0;
 338 }
 339
 340 static void sig_handler(int sig __maybe_unused)
 341 {
 342         session_done = 1;
 343 }
 344
 345 static size_t hists__fprintf_nr_sample_events(struct hists *hists, struct report *rep,
 346                                               const char *evname, FILE *fp)
 347 {
 348         size_t ret;
 349         char unit;
 350         unsigned long nr_samples = hists->stats.nr_events[PERF_RECORD_SAMPLE];
 351         u64 nr_events = hists->stats.total_period;
 352         struct perf_evsel *evsel = hists_to_evsel(hists);
 353         char buf[512];
 354         size_t size = sizeof(buf);
 355
 356         if (perf_evsel__is_group_event(evsel)) {
 357                 struct perf_evsel *pos;
 358
 359                 perf_evsel__group_desc(evsel, buf, size);
 360                 evname = buf;
 361
 362                 for_each_group_member(pos, evsel) {
 363                         nr_samples += pos->hists.stats.nr_events[PERF_RECORD_SAMPLE];
 364                         nr_events += pos->hists.stats.total_period;
 365                 }
 366         }
 367
 368         nr_samples = convert_unit(nr_samples, &unit);
 369         ret = fprintf(fp, "# Samples: %lu%c", nr_samples, unit);
 370         if (evname != NULL)
 371                 ret += fprintf(fp, " of event '%s'", evname);
 372
 373         if (rep->mem_mode) {
 374                 ret += fprintf(fp, "\n# Total weight : %" PRIu64, nr_events);
 375                 ret += fprintf(fp, "\n# Sort order   : %s", sort_order);
 376         } else
 377                 ret += fprintf(fp, "\n# Event count (approx.): %" PRIu64, nr_events);
 378         return ret + fprintf(fp, "\n#\n");
 379 }
 380
 381 static int perf_evlist__tty_browse_hists(struct perf_evlist *evlist,
 382                                          struct report *rep,
 383                                          const char *help)
 384 {
 385         struct perf_evsel *pos;
 386
 387         list_for_each_entry(pos, &evlist->entries, node) {
 388                 struct hists *hists = &pos->hists;
 389                 const char *evname = perf_evsel__name(pos);
 390
 391                 if (symbol_conf.event_group &&
 392                     !perf_evsel__is_group_leader(pos))
 393                         continue;
 394
 395                 hists__fprintf_nr_sample_events(hists, rep, evname, stdout);
 396                 hists__fprintf(hists, true, 0, 0, rep->min_percent, stdout);
 397                 fprintf(stdout, "\n\n");
 398         }
 399
 400         if (sort_order == default_sort_order &&
 401             parent_pattern == default_parent_pattern) {
 402                 fprintf(stdout, "#\n# (%s)\n#\n", help);
 403
 404                 if (rep->show_threads) {
 405                         bool style = !strcmp(rep->pretty_printing_style, "raw");
 406                         perf_read_values_display(stdout, &rep->show_threads_values,
 407                                                  style);
 408                         perf_read_values_destroy(&rep->show_threads_values);
 409                 }
 410         }
 411
 412         return 0;
 413 }
 414
 415 static void report__warn_kptr_restrict(const struct report *rep)
 416 {
 417         struct map *kernel_map = rep->session->machines.host.vmlinux_maps[MAP__FUNCTION];
 418         struct kmap *kernel_kmap = map__kmap(kernel_map);
 419
 420         if (kernel_map == NULL ||
 421             (kernel_map->dso->hit &&
 422              (kernel_kmap->ref_reloc_sym == NULL ||
 423               kernel_kmap->ref_reloc_sym->addr == 0))) {
 424                 const char *desc =
 425                     "As no suitable kallsyms nor vmlinux was found, kernel samples\n"
 426                     "can't be resolved.";
 427
 428                 if (kernel_map) {
 429                         const struct dso *kdso = kernel_map->dso;
 430                         if (!RB_EMPTY_ROOT(&kdso->symbols[MAP__FUNCTION])) {
 431                                 desc = "If some relocation was applied (e.g. "
 432                                        "kexec) symbols may be misresolved.";
 433                         }
 434                 }
 435
 436                 ui__warning(
 437 "Kernel address maps (/proc/{kallsyms,modules}) were restricted.\n\n"
 438 "Check /proc/sys/kernel/kptr_restrict before running 'perf record'.\n\n%s\n\n"
 439 "Samples in kernel modules can't be resolved as well.\n\n",
 440                 desc);
 441         }
 442 }
 443
 444 static int report__gtk_browse_hists(struct report *rep, const char *help)
 445 {
 446         int (*hist_browser)(struct perf_evlist *evlist, const char *help,
 447                             struct hist_browser_timer *timer, float min_pcnt);
 448
 449         hist_browser = dlsym(perf_gtk_handle, "perf_evlist__gtk_browse_hists");
 450
 451         if (hist_browser == NULL) {
 452                 ui__error("GTK browser not found!\n");
 453                 return -1;
 454         }
 455
 456         return hist_browser(rep->session->evlist, help, NULL, rep->min_percent);
 457 }
 458
 459 static int report__browse_hists(struct report *rep)
 460 {
 461         int ret;
 462         struct perf_session *session = rep->session;
 463         struct perf_evlist *evlist = session->evlist;
 464         const char *help = "For a higher level overview, try: perf report --sort comm,dso";
 465
 466         switch (use_browser) {
 467         case 1:
 468                 ret = perf_evlist__tui_browse_hists(evlist, help, NULL,
 469                                                     rep->min_percent,
 470                                                     &session->header.env);
 471                 /*
 472                  * Usually "ret" is the last pressed key, and we only
 473                  * care if the key notifies us to switch data file.
 474                  */
 475                 if (ret != K_SWITCH_INPUT_DATA)
 476                         ret = 0;
 477                 break;
 478         case 2:
 479                 ret = report__gtk_browse_hists(rep, help);
 480                 break;
 481         default:
 482                 ret = perf_evlist__tty_browse_hists(evlist, rep, help);
 483                 break;
 484         }
 485
 486         return ret;
 487 }
 488
 489 static u64 report__collapse_hists(struct report *rep)
 490 {
 491         struct ui_progress prog;
 492         struct perf_evsel *pos;
 493         u64 nr_samples = 0;
 494         /*
 495          * Count number of histogram entries to use when showing progress,
 496          * reusing nr_samples variable.
 497          */
 498         list_for_each_entry(pos, &rep->session->evlist->entries, node)
 499                 nr_samples += pos->hists.nr_entries;
 500
 501         ui_progress__init(&prog, nr_samples, "Merging related events...");
 502         /*
 503          * Count total number of samples, will be used to check if this
 504          * session had any.
 505          */
 506         nr_samples = 0;
 507
 508         list_for_each_entry(pos, &rep->session->evlist->entries, node) {
 509                 struct hists *hists = &pos->hists;
 510
 511                 if (pos->idx == 0)
 512                         hists->symbol_filter_str = rep->symbol_filter_str;
 513
 514                 hists__collapse_resort(hists, &prog);
 515                 nr_samples += hists->stats.nr_events[PERF_RECORD_SAMPLE];
 516
 517                 /* Non-group events are considered as leader */
 518                 if (symbol_conf.event_group &&
 519                     !perf_evsel__is_group_leader(pos)) {
 520                         struct hists *leader_hists = &pos->leader->hists;
 521
 522                         hists__match(leader_hists, hists);
 523                         hists__link(leader_hists, hists);
 524                 }
 525         }
 526
 527         ui_progress__finish();
 528
 529         return nr_samples;
 530 }
 531
 532 static int __cmd_report(struct report *rep)
 533 {
 534         int ret;
 535         u64 nr_samples;
 536         struct perf_session *session = rep->session;
 537         struct perf_evsel *pos;
 538         struct perf_data_file *file = session->file;
 539
 540         signal(SIGINT, sig_handler);
 541
 542         if (rep->cpu_list) {
 543                 ret = perf_session__cpu_bitmap(session, rep->cpu_list,
 544                                                rep->cpu_bitmap);
 545                 if (ret)
 546                         return ret;
 547         }
 548
 549         if (rep->show_threads)
 550                 perf_read_values_init(&rep->show_threads_values);
 551
 552         ret = report__setup_sample_type(rep);
 553         if (ret)
 554                 return ret;
 555
 556         ret = perf_session__process_events(session, &rep->tool);
 557         if (ret)
 558                 return ret;
 559
 560         report__warn_kptr_restrict(rep);
 561
 562         if (use_browser == 0) {
 563                 if (verbose > 3)
 564                         perf_session__fprintf(session, stdout);
 565
 566                 if (verbose > 2)
 567                         perf_session__fprintf_dsos(session, stdout);
 568
 569                 if (dump_trace) {
 570                         perf_session__fprintf_nr_events(session, stdout);
 571                         return 0;
 572                 }
 573         }
 574
 575         nr_samples = report__collapse_hists(rep);
 576
 577         if (session_done())
 578                 return 0;
 579
 580         if (nr_samples == 0) {
 581                 ui__error("The %s file has no samples!\n", file->path);
 582                 return 0;
 583         }
 584
 585         list_for_each_entry(pos, &session->evlist->entries, node)
 586                 hists__output_resort(&pos->hists);
 587
 588         return report__browse_hists(rep);
 589 }
 590
 591 static int
 592 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
 593 {
 594         struct report *rep = (struct report *)opt->value;
 595         char *tok, *tok2;
 596         char *endptr;
 597
 598         /*
 599          * --no-call-graph
 600          */
 601         if (unset) {
 602                 rep->dont_use_callchains = true;
 603                 return 0;
 604         }
 605
 606         symbol_conf.use_callchain = true;
 607
 608         if (!arg)
 609                 return 0;
 610
 611         tok = strtok((char *)arg, ",");
 612         if (!tok)
 613                 return -1;
 614
 615         /* get the output mode */
 616         if (!strncmp(tok, "graph", strlen(arg)))
 617                 callchain_param.mode = CHAIN_GRAPH_ABS;
 618
 619         else if (!strncmp(tok, "flat", strlen(arg)))
 620                 callchain_param.mode = CHAIN_FLAT;
 621
 622         else if (!strncmp(tok, "fractal", strlen(arg)))
 623                 callchain_param.mode = CHAIN_GRAPH_REL;
 624
 625         else if (!strncmp(tok, "none", strlen(arg))) {
 626                 callchain_param.mode = CHAIN_NONE;
 627                 symbol_conf.use_callchain = false;
 628
 629                 return 0;
 630         }
 631
 632         else
 633                 return -1;
 634
 635         /* get the min percentage */
 636         tok = strtok(NULL, ",");
 637         if (!tok)
 638                 goto setup;
 639
 640         callchain_param.min_percent = strtod(tok, &endptr);
 641         if (tok == endptr)
 642                 return -1;
 643
 644         /* get the print limit */
 645         tok2 = strtok(NULL, ",");
 646         if (!tok2)
 647                 goto setup;
 648
 649         if (tok2[0] != 'c') {
 650                 callchain_param.print_limit = strtoul(tok2, &endptr, 0);
 651                 tok2 = strtok(NULL, ",");
 652                 if (!tok2)
 653                         goto setup;
 654         }
 655
 656         /* get the call chain order */
 657         if (!strncmp(tok2, "caller", strlen("caller")))
 658                 callchain_param.order = ORDER_CALLER;
 659         else if (!strncmp(tok2, "callee", strlen("callee")))
 660                 callchain_param.order = ORDER_CALLEE;
 661         else
 662                 return -1;
 663
 664         /* Get the sort key */
 665         tok2 = strtok(NULL, ",");
 666         if (!tok2)
 667                 goto setup;
 668         if (!strncmp(tok2, "function", strlen("function")))
 669                 callchain_param.key = CCKEY_FUNCTION;
 670         else if (!strncmp(tok2, "address", strlen("address")))
 671                 callchain_param.key = CCKEY_ADDRESS;
 672         else
 673                 return -1;
 674 setup:
 675         if (callchain_register_param(&callchain_param) < 0) {
 676                 pr_err("Can't register callchain params\n");
 677                 return -1;
 678         }
 679         return 0;
 680 }
 681
 682 int
 683 report_parse_ignore_callees_opt(const struct option *opt __maybe_unused,
 684                                 const char *arg, int unset __maybe_unused)
 685 {
 686         if (arg) {
 687                 int err = regcomp(&ignore_callees_regex, arg, REG_EXTENDED);
 688                 if (err) {
 689                         char buf[BUFSIZ];
 690                         regerror(err, &ignore_callees_regex, buf, sizeof(buf));
 691                         pr_err("Invalid --ignore-callees regex: %s\n%s", arg, buf);
 692                         return -1;
 693                 }
 694                 have_ignore_callees = 1;
 695         }
 696
 697         return 0;
 698 }
 699
 700 static int
 701 parse_branch_mode(const struct option *opt __maybe_unused,
 702                   const char *str __maybe_unused, int unset)
 703 {
 704         int *branch_mode = opt->value;
 705
 706         *branch_mode = !unset;
 707         return 0;
 708 }
 709
 710 static int
 711 parse_percent_limit(const struct option *opt, const char *str,
 712                     int unset __maybe_unused)
 713 {
 714         struct report *rep = opt->value;
 715
 716         rep->min_percent = strtof(str, NULL);
 717         return 0;
 718 }
 719
 720 int cmd_report(int argc, const char **argv, const char *prefix __maybe_unused)
 721 {
 722         struct perf_session *session;
 723         struct stat st;
 724         bool has_br_stack = false;
 725         int branch_mode = -1;
 726         int ret = -1;
 727         char callchain_default_opt[] = "fractal,0.5,callee";
 728         const char * const report_usage[] = {
 729                 "perf report [<options>]",
 730                 NULL
 731         };
 732         struct report report = {
 733                 .tool = {
 734                         .sample          = process_sample_event,
 735                         .mmap            = perf_event__process_mmap,
 736                         .mmap2           = perf_event__process_mmap2,
 737                         .comm            = perf_event__process_comm,
 738                         .exit            = perf_event__process_exit,
 739                         .fork            = perf_event__process_fork,
 740                         .lost            = perf_event__process_lost,
 741                         .read            = process_read_event,
 742                         .attr            = perf_event__process_attr,
 743                         .tracing_data    = perf_event__process_tracing_data,
 744                         .build_id        = perf_event__process_build_id,
 745                         .ordered_samples = true,
 746                         .ordering_requires_timestamps = true,
 747                 },
 748                 .max_stack               = PERF_MAX_STACK_DEPTH,
 749                 .pretty_printing_style   = "normal",
 750         };
 751         const struct option options[] = {
 752         OPT_STRING('i', "input", &input_name, "file",
 753                     "input file name"),
 754         OPT_INCR('v', "verbose", &verbose,
 755                     "be more verbose (show symbol address, etc)"),
 756         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
 757                     "dump raw trace in ASCII"),
 758         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 759                    "file", "vmlinux pathname"),
 760         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
 761                    "file", "kallsyms pathname"),
 762         OPT_BOOLEAN('f', "force", &report.force, "don't complain, do it"),
 763         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
 764                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
 765         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
 766                     "Show a column with the number of samples"),
 767         OPT_BOOLEAN('T', "threads", &report.show_threads,
 768                     "Show per-thread event counters"),
 769         OPT_STRING(0, "pretty", &report.pretty_printing_style, "key",
 770                    "pretty printing style key: normal raw"),
 771         OPT_BOOLEAN(0, "tui", &report.use_tui, "Use the TUI interface"),
 772         OPT_BOOLEAN(0, "gtk", &report.use_gtk, "Use the GTK2 interface"),
 773         OPT_BOOLEAN(0, "stdio", &report.use_stdio,
 774                     "Use the stdio interface"),
 775         OPT_BOOLEAN(0, "header", &report.header, "Show data header."),
 776         OPT_BOOLEAN(0, "header-only", &report.header_only,
 777                     "Show only data header."),
 778         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
 779                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline,"
 780                    " dso_to, dso_from, symbol_to, symbol_from, mispredict,"
 781                    " weight, local_weight, mem, symbol_daddr, dso_daddr, tlb, "
 782                    "snoop, locked, abort, in_tx, transaction"),
 783         OPT_BOOLEAN(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
 784                     "Show sample percentage for different cpu modes"),
 785         OPT_STRING('p', "parent", &parent_pattern, "regex",
 786                    "regex filter to identify parent, see: '--sort parent'"),
 787         OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
 788                     "Only display entries with parent-match"),
 789         OPT_CALLBACK_DEFAULT('g', "call-graph", &report, "output_type,min_percent[,print_limit],call_order",
 790                      "Display callchains using output_type (graph, flat, fractal, or none) , min percent threshold, optional print limit, callchain order, key (function or address). "
 791                      "Default: fractal,0.5,callee,function", &parse_callchain_opt, callchain_default_opt),
 792         OPT_INTEGER(0, "max-stack", &report.max_stack,
 793                     "Set the maximum stack depth when parsing the callchain, "
 794                     "anything beyond the specified depth will be ignored. "
 795                     "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
 796         OPT_BOOLEAN('G', "inverted", &report.inverted_callchain,
 797                     "alias for inverted call graph"),
 798         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
 799                    "ignore callees of these functions in call graphs",
 800                    report_parse_ignore_callees_opt),
 801         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
 802                    "only consider symbols in these dsos"),
 803         OPT_STRING('c', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
 804                    "only consider symbols in these comms"),
 805         OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
 806                    "only consider these symbols"),
 807         OPT_STRING(0, "symbol-filter", &report.symbol_filter_str, "filter",
 808                    "only show symbols that (partially) match with this filter"),
 809         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
 810                    "width[,width...]",
 811                    "don't try to adjust column width, use these fixed values"),
 812         OPT_STRING('t', "field-separator", &symbol_conf.field_sep, "separator",
 813                    "separator for columns, no spaces will be added between "
 814                    "columns '.' is reserved."),
 815         OPT_BOOLEAN('U', "hide-unresolved", &report.hide_unresolved,
 816                     "Only display entries resolved to a symbol"),
 817         OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
 818                     "Look for files with symbols relative to this directory"),
 819         OPT_STRING('C', "cpu", &report.cpu_list, "cpu",
 820                    "list of cpus to profile"),
 821         OPT_BOOLEAN('I', "show-info", &report.show_full_info,
 822                     "Display extended information about perf.data file"),
 823         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
 824                     "Interleave source code with assembly code (default)"),
 825         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
 826                     "Display raw encoding of assembly instructions (default)"),
 827         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
 828                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
 829         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
 830                     "Show a column with the sum of periods"),
 831         OPT_BOOLEAN(0, "group", &symbol_conf.event_group,
 832                     "Show event group information together"),
 833         OPT_CALLBACK_NOOPT('b', "branch-stack", &branch_mode, "",
 834                     "use branch records for histogram filling", parse_branch_mode),
 835         OPT_STRING(0, "objdump", &objdump_path, "path",
 836                    "objdump binary to use for disassembly and annotations"),
 837         OPT_BOOLEAN(0, "demangle", &symbol_conf.demangle,
 838                     "Disable symbol demangling"),
 839         OPT_BOOLEAN(0, "mem-mode", &report.mem_mode, "mem access profile"),
 840         OPT_CALLBACK(0, "percent-limit", &report, "percent",
 841                      "Don't show entries under that percent", parse_percent_limit),
 842         OPT_END()
 843         };
 844         struct perf_data_file file = {
 845                 .mode  = PERF_DATA_MODE_READ,
 846         };
 847
 848         perf_config(report__config, &report);
 849
 850         argc = parse_options(argc, argv, options, report_usage, 0);
 851
 852         if (report.use_stdio)
 853                 use_browser = 0;
 854         else if (report.use_tui)
 855                 use_browser = 1;
 856         else if (report.use_gtk)
 857                 use_browser = 2;
 858
 859         if (report.inverted_callchain)
 860                 callchain_param.order = ORDER_CALLER;
 861
 862         if (!input_name || !strlen(input_name)) {
 863                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
 864                         input_name = "-";
 865                 else
 866                         input_name = "perf.data";
 867         }
 868
 869         file.path  = input_name;
 870         file.force = report.force;
 871
 872 repeat:
 873         session = perf_session__new(&file, false, &report.tool);
 874         if (session == NULL)
 875                 return -ENOMEM;
 876
 877         report.session = session;
 878
 879         has_br_stack = perf_header__has_feat(&session->header,
 880                                              HEADER_BRANCH_STACK);
 881
 882         if (branch_mode == -1 && has_br_stack)
 883                 sort__mode = SORT_MODE__BRANCH;
 884
 885         /* sort__mode could be NORMAL if --no-branch-stack */
 886         if (sort__mode == SORT_MODE__BRANCH) {
 887                 /*
 888                  * if no sort_order is provided, then specify
 889                  * branch-mode specific order
 890                  */
 891                 if (sort_order == default_sort_order)
 892                         sort_order = "comm,dso_from,symbol_from,"
 893                                      "dso_to,symbol_to";
 894
 895         }
 896         if (report.mem_mode) {
 897                 if (sort__mode == SORT_MODE__BRANCH) {
 898                         pr_err("branch and mem mode incompatible\n");
 899                         goto error;
 900                 }
 901                 sort__mode = SORT_MODE__MEMORY;
 902
 903                 /*
 904                  * if no sort_order is provided, then specify
 905                  * branch-mode specific order
 906                  */
 907                 if (sort_order == default_sort_order)
 908                         sort_order = "local_weight,mem,sym,dso,symbol_daddr,dso_daddr,snoop,tlb,locked";
 909         }
 910
 911         if (setup_sorting() < 0) {
 912                 parse_options_usage(report_usage, options, "s", 1);
 913                 goto error;
 914         }
 915
 916         if (parent_pattern != default_parent_pattern) {
 917                 if (sort_dimension__add("parent") < 0)
 918                         goto error;
 919         }
 920
 921         /* Force tty output for header output. */
 922         if (report.header || report.header_only)
 923                 use_browser = 0;
 924
 925         if (strcmp(input_name, "-") != 0)
 926                 setup_browser(true);
 927         else {
 928                 use_browser = 0;
 929                 perf_hpp__init();
 930         }
 931
 932         if (report.header || report.header_only) {
 933                 perf_session__fprintf_info(session, stdout,
 934                                            report.show_full_info);
 935                 if (report.header_only)
 936                         return 0;
 937         } else if (use_browser == 0) {
 938                 fputs("# To display the perf.data header info, please use --header/--header-only options.\n#\n",
 939                       stdout);
 940         }
 941
 942         /*
 943          * Only in the TUI browser we are doing integrated annotation,
 944          * so don't allocate extra space that won't be used in the stdio
 945          * implementation.
 946          */
 947         if (use_browser == 1 && sort__has_sym) {
 948                 symbol_conf.priv_size = sizeof(struct annotation);
 949                 machines__set_symbol_filter(&session->machines,
 950                                             symbol__annotate_init);
 951                 /*
 952                  * For searching by name on the "Browse map details".
 953                  * providing it only in verbose mode not to bloat too
 954                  * much struct symbol.
 955                  */
 956                 if (verbose) {
 957                         /*
 958                          * XXX: Need to provide a less kludgy way to ask for
 959                          * more space per symbol, the u32 is for the index on
 960                          * the ui browser.
 961                          * See symbol__browser_index.
 962                          */
 963                         symbol_conf.priv_size += sizeof(u32);
 964                         symbol_conf.sort_by_name = true;
 965                 }
 966         }
 967
 968         if (symbol__init() < 0)
 969                 goto error;
 970
 971         if (argc) {
 972                 /*
 973                  * Special case: if there's an argument left then assume that
 974                  * it's a symbol filter:
 975                  */
 976                 if (argc > 1)
 977                         usage_with_options(report_usage, options);
 978
 979                 report.symbol_filter_str = argv[0];
 980         }
 981
 982         sort__setup_elide(stdout);
 983
 984         ret = __cmd_report(&report);
 985         if (ret == K_SWITCH_INPUT_DATA) {
 986                 perf_session__delete(session);
 987                 goto repeat;
 988         } else
 989                 ret = 0;
 990
 991 error:
 992         perf_session__delete(session);
 993         return ret;
 994 }