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 #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 37 #include <linux/bitmap.h> 38 39 struct perf_report { 40 struct perf_tool tool; 41 struct perf_session *session; 42 char const *input_name; 43 bool force, use_tui, use_stdio; 44 bool hide_unresolved; 45 bool dont_use_callchains; 46 bool show_full_info; 47 bool show_threads; 48 bool inverted_callchain; 49 struct perf_read_values show_threads_values; 50 const char *pretty_printing_style; 51 symbol_filter_t annotate_init; 52 const char *cpu_list; 53 DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS); 54 }; 55 56 static int perf_evsel__add_hist_entry(struct perf_evsel *evsel, 57 struct addr_location *al, 58 struct perf_sample *sample, 59 struct machine *machine) 60 { 61 struct symbol *parent = NULL; 62 int err = 0; 63 struct hist_entry *he; 64 65 if ((sort__has_parent || symbol_conf.use_callchain) && sample->callchain) { 66 err = machine__resolve_callchain(machine, evsel, al->thread, 67 sample->callchain, &parent); 68 if (err) 69 return err; 70 } 71 72 he = __hists__add_entry(&evsel->hists, al, parent, sample->period); 73 if (he == NULL) 74 return -ENOMEM; 75 76 if (symbol_conf.use_callchain) { 77 err = callchain_append(he->callchain, 78 &evsel->hists.callchain_cursor, 79 sample->period); 80 if (err) 81 return err; 82 } 83 /* 84 * Only in the newt browser we are doing integrated annotation, 85 * so we don't allocated the extra space needed because the stdio 86 * code will not use it. 87 */ 88 if (al->sym != NULL && use_browser > 0) { 89 struct annotation *notes = symbol__annotation(he->ms.sym); 90 91 assert(evsel != NULL); 92 93 err = -ENOMEM; 94 if (notes->src == NULL && symbol__alloc_hist(he->ms.sym) < 0) 95 goto out; 96 97 err = hist_entry__inc_addr_samples(he, evsel->idx, al->addr); 98 } 99 100 evsel->hists.stats.total_period += sample->period; 101 hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE); 102 out: 103 return err; 104 } 105 106 107 static int process_sample_event(struct perf_tool *tool, 108 union perf_event *event, 109 struct perf_sample *sample, 110 struct perf_evsel *evsel, 111 struct machine *machine) 112 { 113 struct perf_report *rep = container_of(tool, struct perf_report, tool); 114 struct addr_location al; 115 116 if (perf_event__preprocess_sample(event, machine, &al, sample, 117 rep->annotate_init) < 0) { 118 fprintf(stderr, "problem processing %d event, skipping it.\n", 119 event->header.type); 120 return -1; 121 } 122 123 if (al.filtered || (rep->hide_unresolved && al.sym == NULL)) 124 return 0; 125 126 if (rep->cpu_list && !test_bit(sample->cpu, rep->cpu_bitmap)) 127 return 0; 128 129 if (al.map != NULL) 130 al.map->dso->hit = 1; 131 132 if (perf_evsel__add_hist_entry(evsel, &al, sample, machine)) { 133 pr_debug("problem incrementing symbol period, skipping event\n"); 134 return -1; 135 } 136 137 return 0; 138 } 139 140 static int process_read_event(struct perf_tool *tool, 141 union perf_event *event, 142 struct perf_sample *sample __used, 143 struct perf_evsel *evsel, 144 struct machine *machine __used) 145 { 146 struct perf_report *rep = container_of(tool, struct perf_report, tool); 147 148 if (rep->show_threads) { 149 const char *name = evsel ? event_name(evsel) : "unknown"; 150 perf_read_values_add_value(&rep->show_threads_values, 151 event->read.pid, event->read.tid, 152 event->read.id, 153 name, 154 event->read.value); 155 } 156 157 dump_printf(": %d %d %s %" PRIu64 "\n", event->read.pid, event->read.tid, 158 evsel ? event_name(evsel) : "FAIL", 159 event->read.value); 160 161 return 0; 162 } 163 164 static int perf_report__setup_sample_type(struct perf_report *rep) 165 { 166 struct perf_session *self = rep->session; 167 168 if (!(self->sample_type & PERF_SAMPLE_CALLCHAIN)) { 169 if (sort__has_parent) { 170 ui__warning("Selected --sort parent, but no " 171 "callchain data. Did you call " 172 "'perf record' without -g?\n"); 173 return -EINVAL; 174 } 175 if (symbol_conf.use_callchain) { 176 ui__warning("Selected -g but no callchain data. Did " 177 "you call 'perf record' without -g?\n"); 178 return -1; 179 } 180 } else if (!rep->dont_use_callchains && 181 callchain_param.mode != CHAIN_NONE && 182 !symbol_conf.use_callchain) { 183 symbol_conf.use_callchain = true; 184 if (callchain_register_param(&callchain_param) < 0) { 185 ui__warning("Can't register callchain " 186 "params.\n"); 187 return -EINVAL; 188 } 189 } 190 191 return 0; 192 } 193 194 extern volatile int session_done; 195 196 static void sig_handler(int sig __used) 197 { 198 session_done = 1; 199 } 200 201 static size_t hists__fprintf_nr_sample_events(struct hists *self, 202 const char *evname, FILE *fp) 203 { 204 size_t ret; 205 char unit; 206 unsigned long nr_events = self->stats.nr_events[PERF_RECORD_SAMPLE]; 207 208 nr_events = convert_unit(nr_events, &unit); 209 ret = fprintf(fp, "# Events: %lu%c", nr_events, unit); 210 if (evname != NULL) 211 ret += fprintf(fp, " %s", evname); 212 return ret + fprintf(fp, "\n#\n"); 213 } 214 215 static int perf_evlist__tty_browse_hists(struct perf_evlist *evlist, 216 struct perf_report *rep, 217 const char *help) 218 { 219 struct perf_evsel *pos; 220 221 list_for_each_entry(pos, &evlist->entries, node) { 222 struct hists *hists = &pos->hists; 223 const char *evname = event_name(pos); 224 225 hists__fprintf_nr_sample_events(hists, evname, stdout); 226 hists__fprintf(hists, NULL, false, true, 0, 0, stdout); 227 fprintf(stdout, "\n\n"); 228 } 229 230 if (sort_order == default_sort_order && 231 parent_pattern == default_parent_pattern) { 232 fprintf(stdout, "#\n# (%s)\n#\n", help); 233 234 if (rep->show_threads) { 235 bool style = !strcmp(rep->pretty_printing_style, "raw"); 236 perf_read_values_display(stdout, &rep->show_threads_values, 237 style); 238 perf_read_values_destroy(&rep->show_threads_values); 239 } 240 } 241 242 return 0; 243 } 244 245 static int __cmd_report(struct perf_report *rep) 246 { 247 int ret = -EINVAL; 248 u64 nr_samples; 249 struct perf_session *session; 250 struct perf_evsel *pos; 251 struct map *kernel_map; 252 struct kmap *kernel_kmap; 253 const char *help = "For a higher level overview, try: perf report --sort comm,dso"; 254 255 signal(SIGINT, sig_handler); 256 257 session = perf_session__new(rep->input_name, O_RDONLY, 258 rep->force, false, &rep->tool); 259 if (session == NULL) 260 return -ENOMEM; 261 262 rep->session = session; 263 264 if (rep->cpu_list) { 265 ret = perf_session__cpu_bitmap(session, rep->cpu_list, 266 rep->cpu_bitmap); 267 if (ret) 268 goto out_delete; 269 } 270 271 if (use_browser <= 0) 272 perf_session__fprintf_info(session, stdout, rep->show_full_info); 273 274 if (rep->show_threads) 275 perf_read_values_init(&rep->show_threads_values); 276 277 ret = perf_report__setup_sample_type(rep); 278 if (ret) 279 goto out_delete; 280 281 ret = perf_session__process_events(session, &rep->tool); 282 if (ret) 283 goto out_delete; 284 285 kernel_map = session->host_machine.vmlinux_maps[MAP__FUNCTION]; 286 kernel_kmap = map__kmap(kernel_map); 287 if (kernel_map == NULL || 288 (kernel_map->dso->hit && 289 (kernel_kmap->ref_reloc_sym == NULL || 290 kernel_kmap->ref_reloc_sym->addr == 0))) { 291 const struct dso *kdso = kernel_map->dso; 292 293 ui__warning( 294 "Kernel address maps (/proc/{kallsyms,modules}) were restricted.\n\n" 295 "Check /proc/sys/kernel/kptr_restrict before running 'perf record'.\n\n%s\n\n" 296 "Samples in kernel modules can't be resolved as well.\n\n", 297 RB_EMPTY_ROOT(&kdso->symbols[MAP__FUNCTION]) ? 298 "As no suitable kallsyms nor vmlinux was found, kernel samples\n" 299 "can't be resolved." : 300 "If some relocation was applied (e.g. kexec) symbols may be misresolved."); 301 } 302 303 if (dump_trace) { 304 perf_session__fprintf_nr_events(session, stdout); 305 goto out_delete; 306 } 307 308 if (verbose > 3) 309 perf_session__fprintf(session, stdout); 310 311 if (verbose > 2) 312 perf_session__fprintf_dsos(session, stdout); 313 314 nr_samples = 0; 315 list_for_each_entry(pos, &session->evlist->entries, node) { 316 struct hists *hists = &pos->hists; 317 318 hists__collapse_resort(hists); 319 hists__output_resort(hists); 320 nr_samples += hists->stats.nr_events[PERF_RECORD_SAMPLE]; 321 } 322 323 if (nr_samples == 0) { 324 ui__warning("The %s file has no samples!\n", 325 rep->input_name); 326 goto out_delete; 327 } 328 329 if (use_browser > 0) { 330 perf_evlist__tui_browse_hists(session->evlist, help, 331 NULL, NULL, 0); 332 } else 333 perf_evlist__tty_browse_hists(session->evlist, rep, help); 334 335 out_delete: 336 /* 337 * Speed up the exit process, for large files this can 338 * take quite a while. 339 * 340 * XXX Enable this when using valgrind or if we ever 341 * librarize this command. 342 * 343 * Also experiment with obstacks to see how much speed 344 * up we'll get here. 345 * 346 * perf_session__delete(session); 347 */ 348 return ret; 349 } 350 351 static int 352 parse_callchain_opt(const struct option *opt, const char *arg, int unset) 353 { 354 struct perf_report *rep = (struct perf_report *)opt->value; 355 char *tok, *tok2; 356 char *endptr; 357 358 /* 359 * --no-call-graph 360 */ 361 if (unset) { 362 rep->dont_use_callchains = true; 363 return 0; 364 } 365 366 symbol_conf.use_callchain = true; 367 368 if (!arg) 369 return 0; 370 371 tok = strtok((char *)arg, ","); 372 if (!tok) 373 return -1; 374 375 /* get the output mode */ 376 if (!strncmp(tok, "graph", strlen(arg))) 377 callchain_param.mode = CHAIN_GRAPH_ABS; 378 379 else if (!strncmp(tok, "flat", strlen(arg))) 380 callchain_param.mode = CHAIN_FLAT; 381 382 else if (!strncmp(tok, "fractal", strlen(arg))) 383 callchain_param.mode = CHAIN_GRAPH_REL; 384 385 else if (!strncmp(tok, "none", strlen(arg))) { 386 callchain_param.mode = CHAIN_NONE; 387 symbol_conf.use_callchain = false; 388 389 return 0; 390 } 391 392 else 393 return -1; 394 395 /* get the min percentage */ 396 tok = strtok(NULL, ","); 397 if (!tok) 398 goto setup; 399 400 callchain_param.min_percent = strtod(tok, &endptr); 401 if (tok == endptr) 402 return -1; 403 404 /* get the print limit */ 405 tok2 = strtok(NULL, ","); 406 if (!tok2) 407 goto setup; 408 409 if (tok2[0] != 'c') { 410 callchain_param.print_limit = strtoul(tok2, &endptr, 0); 411 tok2 = strtok(NULL, ","); 412 if (!tok2) 413 goto setup; 414 } 415 416 /* get the call chain order */ 417 if (!strcmp(tok2, "caller")) 418 callchain_param.order = ORDER_CALLER; 419 else if (!strcmp(tok2, "callee")) 420 callchain_param.order = ORDER_CALLEE; 421 else 422 return -1; 423 setup: 424 if (callchain_register_param(&callchain_param) < 0) { 425 fprintf(stderr, "Can't register callchain params\n"); 426 return -1; 427 } 428 return 0; 429 } 430 431 int cmd_report(int argc, const char **argv, const char *prefix __used) 432 { 433 char callchain_default_opt[] = "fractal,0.5,callee"; 434 const char * const report_usage[] = { 435 "perf report [<options>]", 436 NULL 437 }; 438 struct perf_report report = { 439 .tool = { 440 .sample = process_sample_event, 441 .mmap = perf_event__process_mmap, 442 .comm = perf_event__process_comm, 443 .exit = perf_event__process_task, 444 .fork = perf_event__process_task, 445 .lost = perf_event__process_lost, 446 .read = process_read_event, 447 .attr = perf_event__process_attr, 448 .event_type = perf_event__process_event_type, 449 .tracing_data = perf_event__process_tracing_data, 450 .build_id = perf_event__process_build_id, 451 .ordered_samples = true, 452 .ordering_requires_timestamps = true, 453 }, 454 .input_name = "perf.data", 455 .pretty_printing_style = "normal", 456 }; 457 const struct option options[] = { 458 OPT_STRING('i', "input", &report.input_name, "file", 459 "input file name"), 460 OPT_INCR('v', "verbose", &verbose, 461 "be more verbose (show symbol address, etc)"), 462 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, 463 "dump raw trace in ASCII"), 464 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name, 465 "file", "vmlinux pathname"), 466 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name, 467 "file", "kallsyms pathname"), 468 OPT_BOOLEAN('f', "force", &report.force, "don't complain, do it"), 469 OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules, 470 "load module symbols - WARNING: use only with -k and LIVE kernel"), 471 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples, 472 "Show a column with the number of samples"), 473 OPT_BOOLEAN('T', "threads", &report.show_threads, 474 "Show per-thread event counters"), 475 OPT_STRING(0, "pretty", &report.pretty_printing_style, "key", 476 "pretty printing style key: normal raw"), 477 OPT_BOOLEAN(0, "tui", &report.use_tui, "Use the TUI interface"), 478 OPT_BOOLEAN(0, "stdio", &report.use_stdio, 479 "Use the stdio interface"), 480 OPT_STRING('s', "sort", &sort_order, "key[,key2...]", 481 "sort by key(s): pid, comm, dso, symbol, parent"), 482 OPT_BOOLEAN(0, "showcpuutilization", &symbol_conf.show_cpu_utilization, 483 "Show sample percentage for different cpu modes"), 484 OPT_STRING('p', "parent", &parent_pattern, "regex", 485 "regex filter to identify parent, see: '--sort parent'"), 486 OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other, 487 "Only display entries with parent-match"), 488 OPT_CALLBACK_DEFAULT('g', "call-graph", &report, "output_type,min_percent[,print_limit],call_order", 489 "Display callchains using output_type (graph, flat, fractal, or none) , min percent threshold, optional print limit and callchain order. " 490 "Default: fractal,0.5,callee", &parse_callchain_opt, callchain_default_opt), 491 OPT_BOOLEAN('G', "inverted", &report.inverted_callchain, 492 "alias for inverted call graph"), 493 OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]", 494 "only consider symbols in these dsos"), 495 OPT_STRING('c', "comms", &symbol_conf.comm_list_str, "comm[,comm...]", 496 "only consider symbols in these comms"), 497 OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]", 498 "only consider these symbols"), 499 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str, 500 "width[,width...]", 501 "don't try to adjust column width, use these fixed values"), 502 OPT_STRING('t', "field-separator", &symbol_conf.field_sep, "separator", 503 "separator for columns, no spaces will be added between " 504 "columns '.' is reserved."), 505 OPT_BOOLEAN('U', "hide-unresolved", &report.hide_unresolved, 506 "Only display entries resolved to a symbol"), 507 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory", 508 "Look for files with symbols relative to this directory"), 509 OPT_STRING('C', "cpu", &report.cpu_list, "cpu", 510 "list of cpus to profile"), 511 OPT_BOOLEAN('I', "show-info", &report.show_full_info, 512 "Display extended information about perf.data file"), 513 OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src, 514 "Interleave source code with assembly code (default)"), 515 OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw, 516 "Display raw encoding of assembly instructions (default)"), 517 OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style", 518 "Specify disassembler style (e.g. -M intel for intel syntax)"), 519 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period, 520 "Show a column with the sum of periods"), 521 OPT_END() 522 }; 523 524 argc = parse_options(argc, argv, options, report_usage, 0); 525 526 if (report.use_stdio) 527 use_browser = 0; 528 else if (report.use_tui) 529 use_browser = 1; 530 531 if (report.inverted_callchain) 532 callchain_param.order = ORDER_CALLER; 533 534 if (strcmp(report.input_name, "-") != 0) 535 setup_browser(true); 536 else 537 use_browser = 0; 538 /* 539 * Only in the newt browser we are doing integrated annotation, 540 * so don't allocate extra space that won't be used in the stdio 541 * implementation. 542 */ 543 if (use_browser > 0) { 544 symbol_conf.priv_size = sizeof(struct annotation); 545 report.annotate_init = symbol__annotate_init; 546 /* 547 * For searching by name on the "Browse map details". 548 * providing it only in verbose mode not to bloat too 549 * much struct symbol. 550 */ 551 if (verbose) { 552 /* 553 * XXX: Need to provide a less kludgy way to ask for 554 * more space per symbol, the u32 is for the index on 555 * the ui browser. 556 * See symbol__browser_index. 557 */ 558 symbol_conf.priv_size += sizeof(u32); 559 symbol_conf.sort_by_name = true; 560 } 561 } 562 563 if (symbol__init() < 0) 564 return -1; 565 566 setup_sorting(report_usage, options); 567 568 if (parent_pattern != default_parent_pattern) { 569 if (sort_dimension__add("parent") < 0) 570 return -1; 571 572 /* 573 * Only show the parent fields if we explicitly 574 * sort that way. If we only use parent machinery 575 * for filtering, we don't want it. 576 */ 577 if (!strstr(sort_order, "parent")) 578 sort_parent.elide = 1; 579 } else 580 symbol_conf.exclude_other = false; 581 582 /* 583 * Any (unrecognized) arguments left? 584 */ 585 if (argc) 586 usage_with_options(report_usage, options); 587 588 sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout); 589 sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout); 590 sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout); 591 592 return __cmd_report(&report); 593 } 594