1 /* 2 * builtin-top.c 3 * 4 * Builtin top command: Display a continuously updated profile of 5 * any workload, CPU or specific PID. 6 * 7 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com> 8 * 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com> 9 * 10 * Improvements and fixes by: 11 * 12 * Arjan van de Ven <arjan@linux.intel.com> 13 * Yanmin Zhang <yanmin.zhang@intel.com> 14 * Wu Fengguang <fengguang.wu@intel.com> 15 * Mike Galbraith <efault@gmx.de> 16 * Paul Mackerras <paulus@samba.org> 17 * 18 * Released under the GPL v2. (and only v2, not any later version) 19 */ 20 #include "builtin.h" 21 22 #include "perf.h" 23 24 #include "util/annotate.h" 25 #include "util/config.h" 26 #include "util/color.h" 27 #include "util/drv_configs.h" 28 #include "util/evlist.h" 29 #include "util/evsel.h" 30 #include "util/event.h" 31 #include "util/machine.h" 32 #include "util/session.h" 33 #include "util/symbol.h" 34 #include "util/thread.h" 35 #include "util/thread_map.h" 36 #include "util/top.h" 37 #include <linux/rbtree.h> 38 #include <subcmd/parse-options.h> 39 #include "util/parse-events.h" 40 #include "util/cpumap.h" 41 #include "util/xyarray.h" 42 #include "util/sort.h" 43 #include "util/term.h" 44 #include "util/intlist.h" 45 #include "util/parse-branch-options.h" 46 #include "arch/common.h" 47 48 #include "util/debug.h" 49 50 #include <assert.h> 51 #include <elf.h> 52 #include <fcntl.h> 53 54 #include <stdio.h> 55 #include <termios.h> 56 #include <unistd.h> 57 #include <inttypes.h> 58 59 #include <errno.h> 60 #include <time.h> 61 #include <sched.h> 62 #include <signal.h> 63 64 #include <sys/syscall.h> 65 #include <sys/ioctl.h> 66 #include <poll.h> 67 #include <sys/prctl.h> 68 #include <sys/wait.h> 69 #include <sys/uio.h> 70 #include <sys/utsname.h> 71 #include <sys/mman.h> 72 73 #include <linux/stringify.h> 74 #include <linux/time64.h> 75 #include <linux/types.h> 76 77 #include "sane_ctype.h" 78 79 static volatile int done; 80 static volatile int resize; 81 82 #define HEADER_LINE_NR 5 83 84 static void perf_top__update_print_entries(struct perf_top *top) 85 { 86 top->print_entries = top->winsize.ws_row - HEADER_LINE_NR; 87 } 88 89 static void winch_sig(int sig __maybe_unused) 90 { 91 resize = 1; 92 } 93 94 static void perf_top__resize(struct perf_top *top) 95 { 96 get_term_dimensions(&top->winsize); 97 perf_top__update_print_entries(top); 98 } 99 100 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he) 101 { 102 struct perf_evsel *evsel = hists_to_evsel(he->hists); 103 struct symbol *sym; 104 struct annotation *notes; 105 struct map *map; 106 int err = -1; 107 108 if (!he || !he->ms.sym) 109 return -1; 110 111 sym = he->ms.sym; 112 map = he->ms.map; 113 114 /* 115 * We can't annotate with just /proc/kallsyms 116 */ 117 if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS && 118 !dso__is_kcore(map->dso)) { 119 pr_err("Can't annotate %s: No vmlinux file was found in the " 120 "path\n", sym->name); 121 sleep(1); 122 return -1; 123 } 124 125 notes = symbol__annotation(sym); 126 if (notes->src != NULL) { 127 pthread_mutex_lock(¬es->lock); 128 goto out_assign; 129 } 130 131 pthread_mutex_lock(¬es->lock); 132 133 if (symbol__alloc_hist(sym) < 0) { 134 pthread_mutex_unlock(¬es->lock); 135 pr_err("Not enough memory for annotating '%s' symbol!\n", 136 sym->name); 137 sleep(1); 138 return err; 139 } 140 141 err = symbol__annotate(sym, map, evsel, 0, NULL); 142 if (err == 0) { 143 out_assign: 144 top->sym_filter_entry = he; 145 } else { 146 char msg[BUFSIZ]; 147 symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg)); 148 pr_err("Couldn't annotate %s: %s\n", sym->name, msg); 149 } 150 151 pthread_mutex_unlock(¬es->lock); 152 return err; 153 } 154 155 static void __zero_source_counters(struct hist_entry *he) 156 { 157 struct symbol *sym = he->ms.sym; 158 symbol__annotate_zero_histograms(sym); 159 } 160 161 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip) 162 { 163 struct utsname uts; 164 int err = uname(&uts); 165 166 ui__warning("Out of bounds address found:\n\n" 167 "Addr: %" PRIx64 "\n" 168 "DSO: %s %c\n" 169 "Map: %" PRIx64 "-%" PRIx64 "\n" 170 "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n" 171 "Arch: %s\n" 172 "Kernel: %s\n" 173 "Tools: %s\n\n" 174 "Not all samples will be on the annotation output.\n\n" 175 "Please report to linux-kernel@vger.kernel.org\n", 176 ip, map->dso->long_name, dso__symtab_origin(map->dso), 177 map->start, map->end, sym->start, sym->end, 178 sym->binding == STB_GLOBAL ? 'g' : 179 sym->binding == STB_LOCAL ? 'l' : 'w', sym->name, 180 err ? "[unknown]" : uts.machine, 181 err ? "[unknown]" : uts.release, perf_version_string); 182 if (use_browser <= 0) 183 sleep(5); 184 185 map->erange_warned = true; 186 } 187 188 static void perf_top__record_precise_ip(struct perf_top *top, 189 struct hist_entry *he, 190 struct perf_sample *sample, 191 int counter, u64 ip) 192 { 193 struct annotation *notes; 194 struct symbol *sym = he->ms.sym; 195 int err = 0; 196 197 if (sym == NULL || (use_browser == 0 && 198 (top->sym_filter_entry == NULL || 199 top->sym_filter_entry->ms.sym != sym))) 200 return; 201 202 notes = symbol__annotation(sym); 203 204 if (pthread_mutex_trylock(¬es->lock)) 205 return; 206 207 err = hist_entry__inc_addr_samples(he, sample, counter, ip); 208 209 pthread_mutex_unlock(¬es->lock); 210 211 if (unlikely(err)) { 212 /* 213 * This function is now called with he->hists->lock held. 214 * Release it before going to sleep. 215 */ 216 pthread_mutex_unlock(&he->hists->lock); 217 218 if (err == -ERANGE && !he->ms.map->erange_warned) 219 ui__warn_map_erange(he->ms.map, sym, ip); 220 else if (err == -ENOMEM) { 221 pr_err("Not enough memory for annotating '%s' symbol!\n", 222 sym->name); 223 sleep(1); 224 } 225 226 pthread_mutex_lock(&he->hists->lock); 227 } 228 } 229 230 static void perf_top__show_details(struct perf_top *top) 231 { 232 struct hist_entry *he = top->sym_filter_entry; 233 struct perf_evsel *evsel = hists_to_evsel(he->hists); 234 struct annotation *notes; 235 struct symbol *symbol; 236 int more; 237 238 if (!he) 239 return; 240 241 symbol = he->ms.sym; 242 notes = symbol__annotation(symbol); 243 244 pthread_mutex_lock(¬es->lock); 245 246 symbol__calc_percent(symbol, evsel); 247 248 if (notes->src == NULL) 249 goto out_unlock; 250 251 printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name); 252 printf(" Events Pcnt (>=%d%%)\n", top->sym_pcnt_filter); 253 254 more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel, 255 0, top->sym_pcnt_filter, top->print_entries, 4); 256 257 if (top->evlist->enabled) { 258 if (top->zero) 259 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx); 260 else 261 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx); 262 } 263 if (more != 0) 264 printf("%d lines not displayed, maybe increase display entries [e]\n", more); 265 out_unlock: 266 pthread_mutex_unlock(¬es->lock); 267 } 268 269 static void perf_top__print_sym_table(struct perf_top *top) 270 { 271 char bf[160]; 272 int printed = 0; 273 const int win_width = top->winsize.ws_col - 1; 274 struct perf_evsel *evsel = top->sym_evsel; 275 struct hists *hists = evsel__hists(evsel); 276 277 puts(CONSOLE_CLEAR); 278 279 perf_top__header_snprintf(top, bf, sizeof(bf)); 280 printf("%s\n", bf); 281 282 perf_top__reset_sample_counters(top); 283 284 printf("%-*.*s\n", win_width, win_width, graph_dotted_line); 285 286 if (hists->stats.nr_lost_warned != 287 hists->stats.nr_events[PERF_RECORD_LOST]) { 288 hists->stats.nr_lost_warned = 289 hists->stats.nr_events[PERF_RECORD_LOST]; 290 color_fprintf(stdout, PERF_COLOR_RED, 291 "WARNING: LOST %d chunks, Check IO/CPU overload", 292 hists->stats.nr_lost_warned); 293 ++printed; 294 } 295 296 if (top->sym_filter_entry) { 297 perf_top__show_details(top); 298 return; 299 } 300 301 if (top->evlist->enabled) { 302 if (top->zero) { 303 hists__delete_entries(hists); 304 } else { 305 hists__decay_entries(hists, top->hide_user_symbols, 306 top->hide_kernel_symbols); 307 } 308 } 309 310 hists__collapse_resort(hists, NULL); 311 perf_evsel__output_resort(evsel, NULL); 312 313 hists__output_recalc_col_len(hists, top->print_entries - printed); 314 putchar('\n'); 315 hists__fprintf(hists, false, top->print_entries - printed, win_width, 316 top->min_percent, stdout, symbol_conf.use_callchain); 317 } 318 319 static void prompt_integer(int *target, const char *msg) 320 { 321 char *buf = malloc(0), *p; 322 size_t dummy = 0; 323 int tmp; 324 325 fprintf(stdout, "\n%s: ", msg); 326 if (getline(&buf, &dummy, stdin) < 0) 327 return; 328 329 p = strchr(buf, '\n'); 330 if (p) 331 *p = 0; 332 333 p = buf; 334 while(*p) { 335 if (!isdigit(*p)) 336 goto out_free; 337 p++; 338 } 339 tmp = strtoul(buf, NULL, 10); 340 *target = tmp; 341 out_free: 342 free(buf); 343 } 344 345 static void prompt_percent(int *target, const char *msg) 346 { 347 int tmp = 0; 348 349 prompt_integer(&tmp, msg); 350 if (tmp >= 0 && tmp <= 100) 351 *target = tmp; 352 } 353 354 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg) 355 { 356 char *buf = malloc(0), *p; 357 struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL; 358 struct hists *hists = evsel__hists(top->sym_evsel); 359 struct rb_node *next; 360 size_t dummy = 0; 361 362 /* zero counters of active symbol */ 363 if (syme) { 364 __zero_source_counters(syme); 365 top->sym_filter_entry = NULL; 366 } 367 368 fprintf(stdout, "\n%s: ", msg); 369 if (getline(&buf, &dummy, stdin) < 0) 370 goto out_free; 371 372 p = strchr(buf, '\n'); 373 if (p) 374 *p = 0; 375 376 next = rb_first(&hists->entries); 377 while (next) { 378 n = rb_entry(next, struct hist_entry, rb_node); 379 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) { 380 found = n; 381 break; 382 } 383 next = rb_next(&n->rb_node); 384 } 385 386 if (!found) { 387 fprintf(stderr, "Sorry, %s is not active.\n", buf); 388 sleep(1); 389 } else 390 perf_top__parse_source(top, found); 391 392 out_free: 393 free(buf); 394 } 395 396 static void perf_top__print_mapped_keys(struct perf_top *top) 397 { 398 char *name = NULL; 399 400 if (top->sym_filter_entry) { 401 struct symbol *sym = top->sym_filter_entry->ms.sym; 402 name = sym->name; 403 } 404 405 fprintf(stdout, "\nMapped keys:\n"); 406 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", top->delay_secs); 407 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", top->print_entries); 408 409 if (top->evlist->nr_entries > 1) 410 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", perf_evsel__name(top->sym_evsel)); 411 412 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", top->count_filter); 413 414 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter); 415 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL"); 416 fprintf(stdout, "\t[S] stop annotation.\n"); 417 418 fprintf(stdout, 419 "\t[K] hide kernel symbols. \t(%s)\n", 420 top->hide_kernel_symbols ? "yes" : "no"); 421 fprintf(stdout, 422 "\t[U] hide user symbols. \t(%s)\n", 423 top->hide_user_symbols ? "yes" : "no"); 424 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", top->zero ? 1 : 0); 425 fprintf(stdout, "\t[qQ] quit.\n"); 426 } 427 428 static int perf_top__key_mapped(struct perf_top *top, int c) 429 { 430 switch (c) { 431 case 'd': 432 case 'e': 433 case 'f': 434 case 'z': 435 case 'q': 436 case 'Q': 437 case 'K': 438 case 'U': 439 case 'F': 440 case 's': 441 case 'S': 442 return 1; 443 case 'E': 444 return top->evlist->nr_entries > 1 ? 1 : 0; 445 default: 446 break; 447 } 448 449 return 0; 450 } 451 452 static bool perf_top__handle_keypress(struct perf_top *top, int c) 453 { 454 bool ret = true; 455 456 if (!perf_top__key_mapped(top, c)) { 457 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN }; 458 struct termios save; 459 460 perf_top__print_mapped_keys(top); 461 fprintf(stdout, "\nEnter selection, or unmapped key to continue: "); 462 fflush(stdout); 463 464 set_term_quiet_input(&save); 465 466 poll(&stdin_poll, 1, -1); 467 c = getc(stdin); 468 469 tcsetattr(0, TCSAFLUSH, &save); 470 if (!perf_top__key_mapped(top, c)) 471 return ret; 472 } 473 474 switch (c) { 475 case 'd': 476 prompt_integer(&top->delay_secs, "Enter display delay"); 477 if (top->delay_secs < 1) 478 top->delay_secs = 1; 479 break; 480 case 'e': 481 prompt_integer(&top->print_entries, "Enter display entries (lines)"); 482 if (top->print_entries == 0) { 483 perf_top__resize(top); 484 signal(SIGWINCH, winch_sig); 485 } else { 486 signal(SIGWINCH, SIG_DFL); 487 } 488 break; 489 case 'E': 490 if (top->evlist->nr_entries > 1) { 491 /* Select 0 as the default event: */ 492 int counter = 0; 493 494 fprintf(stderr, "\nAvailable events:"); 495 496 evlist__for_each_entry(top->evlist, top->sym_evsel) 497 fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel)); 498 499 prompt_integer(&counter, "Enter details event counter"); 500 501 if (counter >= top->evlist->nr_entries) { 502 top->sym_evsel = perf_evlist__first(top->evlist); 503 fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel)); 504 sleep(1); 505 break; 506 } 507 evlist__for_each_entry(top->evlist, top->sym_evsel) 508 if (top->sym_evsel->idx == counter) 509 break; 510 } else 511 top->sym_evsel = perf_evlist__first(top->evlist); 512 break; 513 case 'f': 514 prompt_integer(&top->count_filter, "Enter display event count filter"); 515 break; 516 case 'F': 517 prompt_percent(&top->sym_pcnt_filter, 518 "Enter details display event filter (percent)"); 519 break; 520 case 'K': 521 top->hide_kernel_symbols = !top->hide_kernel_symbols; 522 break; 523 case 'q': 524 case 'Q': 525 printf("exiting.\n"); 526 if (top->dump_symtab) 527 perf_session__fprintf_dsos(top->session, stderr); 528 ret = false; 529 break; 530 case 's': 531 perf_top__prompt_symbol(top, "Enter details symbol"); 532 break; 533 case 'S': 534 if (!top->sym_filter_entry) 535 break; 536 else { 537 struct hist_entry *syme = top->sym_filter_entry; 538 539 top->sym_filter_entry = NULL; 540 __zero_source_counters(syme); 541 } 542 break; 543 case 'U': 544 top->hide_user_symbols = !top->hide_user_symbols; 545 break; 546 case 'z': 547 top->zero = !top->zero; 548 break; 549 default: 550 break; 551 } 552 553 return ret; 554 } 555 556 static void perf_top__sort_new_samples(void *arg) 557 { 558 struct perf_top *t = arg; 559 struct perf_evsel *evsel = t->sym_evsel; 560 struct hists *hists; 561 562 perf_top__reset_sample_counters(t); 563 564 if (t->evlist->selected != NULL) 565 t->sym_evsel = t->evlist->selected; 566 567 hists = evsel__hists(evsel); 568 569 if (t->evlist->enabled) { 570 if (t->zero) { 571 hists__delete_entries(hists); 572 } else { 573 hists__decay_entries(hists, t->hide_user_symbols, 574 t->hide_kernel_symbols); 575 } 576 } 577 578 hists__collapse_resort(hists, NULL); 579 perf_evsel__output_resort(evsel, NULL); 580 } 581 582 static void *display_thread_tui(void *arg) 583 { 584 struct perf_evsel *pos; 585 struct perf_top *top = arg; 586 const char *help = "For a higher level overview, try: perf top --sort comm,dso"; 587 struct hist_browser_timer hbt = { 588 .timer = perf_top__sort_new_samples, 589 .arg = top, 590 .refresh = top->delay_secs, 591 }; 592 593 /* In order to read symbols from other namespaces perf to needs to call 594 * setns(2). This isn't permitted if the struct_fs has multiple users. 595 * unshare(2) the fs so that we may continue to setns into namespaces 596 * that we're observing. 597 */ 598 unshare(CLONE_FS); 599 600 perf_top__sort_new_samples(top); 601 602 /* 603 * Initialize the uid_filter_str, in the future the TUI will allow 604 * Zooming in/out UIDs. For now juse use whatever the user passed 605 * via --uid. 606 */ 607 evlist__for_each_entry(top->evlist, pos) { 608 struct hists *hists = evsel__hists(pos); 609 hists->uid_filter_str = top->record_opts.target.uid_str; 610 } 611 612 perf_evlist__tui_browse_hists(top->evlist, help, &hbt, 613 top->min_percent, 614 &top->session->header.env); 615 616 done = 1; 617 return NULL; 618 } 619 620 static void display_sig(int sig __maybe_unused) 621 { 622 done = 1; 623 } 624 625 static void display_setup_sig(void) 626 { 627 signal(SIGSEGV, sighandler_dump_stack); 628 signal(SIGFPE, sighandler_dump_stack); 629 signal(SIGINT, display_sig); 630 signal(SIGQUIT, display_sig); 631 signal(SIGTERM, display_sig); 632 } 633 634 static void *display_thread(void *arg) 635 { 636 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN }; 637 struct termios save; 638 struct perf_top *top = arg; 639 int delay_msecs, c; 640 641 /* In order to read symbols from other namespaces perf to needs to call 642 * setns(2). This isn't permitted if the struct_fs has multiple users. 643 * unshare(2) the fs so that we may continue to setns into namespaces 644 * that we're observing. 645 */ 646 unshare(CLONE_FS); 647 648 display_setup_sig(); 649 pthread__unblock_sigwinch(); 650 repeat: 651 delay_msecs = top->delay_secs * MSEC_PER_SEC; 652 set_term_quiet_input(&save); 653 /* trash return*/ 654 getc(stdin); 655 656 while (!done) { 657 perf_top__print_sym_table(top); 658 /* 659 * Either timeout expired or we got an EINTR due to SIGWINCH, 660 * refresh screen in both cases. 661 */ 662 switch (poll(&stdin_poll, 1, delay_msecs)) { 663 case 0: 664 continue; 665 case -1: 666 if (errno == EINTR) 667 continue; 668 __fallthrough; 669 default: 670 c = getc(stdin); 671 tcsetattr(0, TCSAFLUSH, &save); 672 673 if (perf_top__handle_keypress(top, c)) 674 goto repeat; 675 done = 1; 676 } 677 } 678 679 tcsetattr(0, TCSAFLUSH, &save); 680 return NULL; 681 } 682 683 static int hist_iter__top_callback(struct hist_entry_iter *iter, 684 struct addr_location *al, bool single, 685 void *arg) 686 { 687 struct perf_top *top = arg; 688 struct hist_entry *he = iter->he; 689 struct perf_evsel *evsel = iter->evsel; 690 691 if (perf_hpp_list.sym && single) 692 perf_top__record_precise_ip(top, he, iter->sample, evsel->idx, al->addr); 693 694 hist__account_cycles(iter->sample->branch_stack, al, iter->sample, 695 !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY)); 696 return 0; 697 } 698 699 static void perf_event__process_sample(struct perf_tool *tool, 700 const union perf_event *event, 701 struct perf_evsel *evsel, 702 struct perf_sample *sample, 703 struct machine *machine) 704 { 705 struct perf_top *top = container_of(tool, struct perf_top, tool); 706 struct addr_location al; 707 int err; 708 709 if (!machine && perf_guest) { 710 static struct intlist *seen; 711 712 if (!seen) 713 seen = intlist__new(NULL); 714 715 if (!intlist__has_entry(seen, sample->pid)) { 716 pr_err("Can't find guest [%d]'s kernel information\n", 717 sample->pid); 718 intlist__add(seen, sample->pid); 719 } 720 return; 721 } 722 723 if (!machine) { 724 pr_err("%u unprocessable samples recorded.\r", 725 top->session->evlist->stats.nr_unprocessable_samples++); 726 return; 727 } 728 729 if (event->header.misc & PERF_RECORD_MISC_EXACT_IP) 730 top->exact_samples++; 731 732 if (machine__resolve(machine, &al, sample) < 0) 733 return; 734 735 if (!machine->kptr_restrict_warned && 736 symbol_conf.kptr_restrict && 737 al.cpumode == PERF_RECORD_MISC_KERNEL) { 738 if (!perf_evlist__exclude_kernel(top->session->evlist)) { 739 ui__warning( 740 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n" 741 "Check /proc/sys/kernel/kptr_restrict.\n\n" 742 "Kernel%s samples will not be resolved.\n", 743 al.map && !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ? 744 " modules" : ""); 745 if (use_browser <= 0) 746 sleep(5); 747 } 748 machine->kptr_restrict_warned = true; 749 } 750 751 if (al.sym == NULL) { 752 const char *msg = "Kernel samples will not be resolved.\n"; 753 /* 754 * As we do lazy loading of symtabs we only will know if the 755 * specified vmlinux file is invalid when we actually have a 756 * hit in kernel space and then try to load it. So if we get 757 * here and there are _no_ symbols in the DSO backing the 758 * kernel map, bail out. 759 * 760 * We may never get here, for instance, if we use -K/ 761 * --hide-kernel-symbols, even if the user specifies an 762 * invalid --vmlinux ;-) 763 */ 764 if (!machine->kptr_restrict_warned && !top->vmlinux_warned && 765 al.map == machine->vmlinux_maps[MAP__FUNCTION] && 766 RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) { 767 if (symbol_conf.vmlinux_name) { 768 char serr[256]; 769 dso__strerror_load(al.map->dso, serr, sizeof(serr)); 770 ui__warning("The %s file can't be used: %s\n%s", 771 symbol_conf.vmlinux_name, serr, msg); 772 } else { 773 ui__warning("A vmlinux file was not found.\n%s", 774 msg); 775 } 776 777 if (use_browser <= 0) 778 sleep(5); 779 top->vmlinux_warned = true; 780 } 781 } 782 783 if (al.sym == NULL || !al.sym->idle) { 784 struct hists *hists = evsel__hists(evsel); 785 struct hist_entry_iter iter = { 786 .evsel = evsel, 787 .sample = sample, 788 .add_entry_cb = hist_iter__top_callback, 789 }; 790 791 if (symbol_conf.cumulate_callchain) 792 iter.ops = &hist_iter_cumulative; 793 else 794 iter.ops = &hist_iter_normal; 795 796 pthread_mutex_lock(&hists->lock); 797 798 err = hist_entry_iter__add(&iter, &al, top->max_stack, top); 799 if (err < 0) 800 pr_err("Problem incrementing symbol period, skipping event\n"); 801 802 pthread_mutex_unlock(&hists->lock); 803 } 804 805 addr_location__put(&al); 806 } 807 808 static void perf_top__mmap_read_idx(struct perf_top *top, int idx) 809 { 810 struct perf_sample sample; 811 struct perf_evsel *evsel; 812 struct perf_session *session = top->session; 813 union perf_event *event; 814 struct machine *machine; 815 int ret; 816 817 while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) { 818 ret = perf_evlist__parse_sample(top->evlist, event, &sample); 819 if (ret) { 820 pr_err("Can't parse sample, err = %d\n", ret); 821 goto next_event; 822 } 823 824 evsel = perf_evlist__id2evsel(session->evlist, sample.id); 825 assert(evsel != NULL); 826 827 if (event->header.type == PERF_RECORD_SAMPLE) 828 ++top->samples; 829 830 switch (sample.cpumode) { 831 case PERF_RECORD_MISC_USER: 832 ++top->us_samples; 833 if (top->hide_user_symbols) 834 goto next_event; 835 machine = &session->machines.host; 836 break; 837 case PERF_RECORD_MISC_KERNEL: 838 ++top->kernel_samples; 839 if (top->hide_kernel_symbols) 840 goto next_event; 841 machine = &session->machines.host; 842 break; 843 case PERF_RECORD_MISC_GUEST_KERNEL: 844 ++top->guest_kernel_samples; 845 machine = perf_session__find_machine(session, 846 sample.pid); 847 break; 848 case PERF_RECORD_MISC_GUEST_USER: 849 ++top->guest_us_samples; 850 /* 851 * TODO: we don't process guest user from host side 852 * except simple counting. 853 */ 854 goto next_event; 855 default: 856 if (event->header.type == PERF_RECORD_SAMPLE) 857 goto next_event; 858 machine = &session->machines.host; 859 break; 860 } 861 862 863 if (event->header.type == PERF_RECORD_SAMPLE) { 864 perf_event__process_sample(&top->tool, event, evsel, 865 &sample, machine); 866 } else if (event->header.type < PERF_RECORD_MAX) { 867 hists__inc_nr_events(evsel__hists(evsel), event->header.type); 868 machine__process_event(machine, event, &sample); 869 } else 870 ++session->evlist->stats.nr_unknown_events; 871 next_event: 872 perf_evlist__mmap_consume(top->evlist, idx); 873 } 874 } 875 876 static void perf_top__mmap_read(struct perf_top *top) 877 { 878 int i; 879 880 for (i = 0; i < top->evlist->nr_mmaps; i++) 881 perf_top__mmap_read_idx(top, i); 882 } 883 884 static int perf_top__start_counters(struct perf_top *top) 885 { 886 char msg[BUFSIZ]; 887 struct perf_evsel *counter; 888 struct perf_evlist *evlist = top->evlist; 889 struct record_opts *opts = &top->record_opts; 890 891 perf_evlist__config(evlist, opts, &callchain_param); 892 893 evlist__for_each_entry(evlist, counter) { 894 try_again: 895 if (perf_evsel__open(counter, top->evlist->cpus, 896 top->evlist->threads) < 0) { 897 if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) { 898 if (verbose > 0) 899 ui__warning("%s\n", msg); 900 goto try_again; 901 } 902 903 perf_evsel__open_strerror(counter, &opts->target, 904 errno, msg, sizeof(msg)); 905 ui__error("%s\n", msg); 906 goto out_err; 907 } 908 } 909 910 if (perf_evlist__mmap(evlist, opts->mmap_pages) < 0) { 911 ui__error("Failed to mmap with %d (%s)\n", 912 errno, str_error_r(errno, msg, sizeof(msg))); 913 goto out_err; 914 } 915 916 return 0; 917 918 out_err: 919 return -1; 920 } 921 922 static int callchain_param__setup_sample_type(struct callchain_param *callchain) 923 { 924 if (!perf_hpp_list.sym) { 925 if (callchain->enabled) { 926 ui__error("Selected -g but \"sym\" not present in --sort/-s."); 927 return -EINVAL; 928 } 929 } else if (callchain->mode != CHAIN_NONE) { 930 if (callchain_register_param(callchain) < 0) { 931 ui__error("Can't register callchain params.\n"); 932 return -EINVAL; 933 } 934 } 935 936 return 0; 937 } 938 939 static int __cmd_top(struct perf_top *top) 940 { 941 char msg[512]; 942 struct perf_evsel *pos; 943 struct perf_evsel_config_term *err_term; 944 struct perf_evlist *evlist = top->evlist; 945 struct record_opts *opts = &top->record_opts; 946 pthread_t thread; 947 int ret; 948 949 top->session = perf_session__new(NULL, false, NULL); 950 if (top->session == NULL) 951 return -1; 952 953 if (!objdump_path) { 954 ret = perf_env__lookup_objdump(&top->session->header.env); 955 if (ret) 956 goto out_delete; 957 } 958 959 ret = callchain_param__setup_sample_type(&callchain_param); 960 if (ret) 961 goto out_delete; 962 963 if (perf_session__register_idle_thread(top->session) < 0) 964 goto out_delete; 965 966 if (top->nr_threads_synthesize > 1) 967 perf_set_multithreaded(); 968 969 machine__synthesize_threads(&top->session->machines.host, &opts->target, 970 top->evlist->threads, false, 971 opts->proc_map_timeout, 972 top->nr_threads_synthesize); 973 974 if (top->nr_threads_synthesize > 1) 975 perf_set_singlethreaded(); 976 977 if (perf_hpp_list.socket) { 978 ret = perf_env__read_cpu_topology_map(&perf_env); 979 if (ret < 0) 980 goto out_err_cpu_topo; 981 } 982 983 ret = perf_top__start_counters(top); 984 if (ret) 985 goto out_delete; 986 987 ret = perf_evlist__apply_drv_configs(evlist, &pos, &err_term); 988 if (ret) { 989 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n", 990 err_term->val.drv_cfg, perf_evsel__name(pos), errno, 991 str_error_r(errno, msg, sizeof(msg))); 992 goto out_delete; 993 } 994 995 top->session->evlist = top->evlist; 996 perf_session__set_id_hdr_size(top->session); 997 998 /* 999 * When perf is starting the traced process, all the events (apart from 1000 * group members) have enable_on_exec=1 set, so don't spoil it by 1001 * prematurely enabling them. 1002 * 1003 * XXX 'top' still doesn't start workloads like record, trace, but should, 1004 * so leave the check here. 1005 */ 1006 if (!target__none(&opts->target)) 1007 perf_evlist__enable(top->evlist); 1008 1009 /* Wait for a minimal set of events before starting the snapshot */ 1010 perf_evlist__poll(top->evlist, 100); 1011 1012 perf_top__mmap_read(top); 1013 1014 ret = -1; 1015 if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui : 1016 display_thread), top)) { 1017 ui__error("Could not create display thread.\n"); 1018 goto out_delete; 1019 } 1020 1021 if (top->realtime_prio) { 1022 struct sched_param param; 1023 1024 param.sched_priority = top->realtime_prio; 1025 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) { 1026 ui__error("Could not set realtime priority.\n"); 1027 goto out_join; 1028 } 1029 } 1030 1031 while (!done) { 1032 u64 hits = top->samples; 1033 1034 perf_top__mmap_read(top); 1035 1036 if (hits == top->samples) 1037 ret = perf_evlist__poll(top->evlist, 100); 1038 1039 if (resize) { 1040 perf_top__resize(top); 1041 resize = 0; 1042 } 1043 } 1044 1045 ret = 0; 1046 out_join: 1047 pthread_join(thread, NULL); 1048 out_delete: 1049 perf_session__delete(top->session); 1050 top->session = NULL; 1051 1052 return ret; 1053 1054 out_err_cpu_topo: { 1055 char errbuf[BUFSIZ]; 1056 const char *err = str_error_r(-ret, errbuf, sizeof(errbuf)); 1057 1058 ui__error("Could not read the CPU topology map: %s\n", err); 1059 goto out_delete; 1060 } 1061 } 1062 1063 static int 1064 callchain_opt(const struct option *opt, const char *arg, int unset) 1065 { 1066 symbol_conf.use_callchain = true; 1067 return record_callchain_opt(opt, arg, unset); 1068 } 1069 1070 static int 1071 parse_callchain_opt(const struct option *opt, const char *arg, int unset) 1072 { 1073 struct callchain_param *callchain = opt->value; 1074 1075 callchain->enabled = !unset; 1076 callchain->record_mode = CALLCHAIN_FP; 1077 1078 /* 1079 * --no-call-graph 1080 */ 1081 if (unset) { 1082 symbol_conf.use_callchain = false; 1083 callchain->record_mode = CALLCHAIN_NONE; 1084 return 0; 1085 } 1086 1087 return parse_callchain_top_opt(arg); 1088 } 1089 1090 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused) 1091 { 1092 if (!strcmp(var, "top.call-graph")) 1093 var = "call-graph.record-mode"; /* fall-through */ 1094 if (!strcmp(var, "top.children")) { 1095 symbol_conf.cumulate_callchain = perf_config_bool(var, value); 1096 return 0; 1097 } 1098 1099 return 0; 1100 } 1101 1102 static int 1103 parse_percent_limit(const struct option *opt, const char *arg, 1104 int unset __maybe_unused) 1105 { 1106 struct perf_top *top = opt->value; 1107 1108 top->min_percent = strtof(arg, NULL); 1109 return 0; 1110 } 1111 1112 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP 1113 "\n\t\t\t\tDefault: fp,graph,0.5,caller,function"; 1114 1115 int cmd_top(int argc, const char **argv) 1116 { 1117 char errbuf[BUFSIZ]; 1118 struct perf_top top = { 1119 .count_filter = 5, 1120 .delay_secs = 2, 1121 .record_opts = { 1122 .mmap_pages = UINT_MAX, 1123 .user_freq = UINT_MAX, 1124 .user_interval = ULLONG_MAX, 1125 .freq = 4000, /* 4 KHz */ 1126 .target = { 1127 .uses_mmap = true, 1128 }, 1129 .proc_map_timeout = 500, 1130 }, 1131 .max_stack = sysctl_perf_event_max_stack, 1132 .sym_pcnt_filter = 5, 1133 .nr_threads_synthesize = UINT_MAX, 1134 }; 1135 struct record_opts *opts = &top.record_opts; 1136 struct target *target = &opts->target; 1137 const struct option options[] = { 1138 OPT_CALLBACK('e', "event", &top.evlist, "event", 1139 "event selector. use 'perf list' to list available events", 1140 parse_events_option), 1141 OPT_U64('c', "count", &opts->user_interval, "event period to sample"), 1142 OPT_STRING('p', "pid", &target->pid, "pid", 1143 "profile events on existing process id"), 1144 OPT_STRING('t', "tid", &target->tid, "tid", 1145 "profile events on existing thread id"), 1146 OPT_BOOLEAN('a', "all-cpus", &target->system_wide, 1147 "system-wide collection from all CPUs"), 1148 OPT_STRING('C', "cpu", &target->cpu_list, "cpu", 1149 "list of cpus to monitor"), 1150 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name, 1151 "file", "vmlinux pathname"), 1152 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux, 1153 "don't load vmlinux even if found"), 1154 OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols, 1155 "hide kernel symbols"), 1156 OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages", 1157 "number of mmap data pages", 1158 perf_evlist__parse_mmap_pages), 1159 OPT_INTEGER('r', "realtime", &top.realtime_prio, 1160 "collect data with this RT SCHED_FIFO priority"), 1161 OPT_INTEGER('d', "delay", &top.delay_secs, 1162 "number of seconds to delay between refreshes"), 1163 OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab, 1164 "dump the symbol table used for profiling"), 1165 OPT_INTEGER('f', "count-filter", &top.count_filter, 1166 "only display functions with more events than this"), 1167 OPT_BOOLEAN(0, "group", &opts->group, 1168 "put the counters into a counter group"), 1169 OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit, 1170 "child tasks do not inherit counters"), 1171 OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name", 1172 "symbol to annotate"), 1173 OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"), 1174 OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"), 1175 OPT_INTEGER('E', "entries", &top.print_entries, 1176 "display this many functions"), 1177 OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols, 1178 "hide user symbols"), 1179 OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"), 1180 OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"), 1181 OPT_INCR('v', "verbose", &verbose, 1182 "be more verbose (show counter open errors, etc)"), 1183 OPT_STRING('s', "sort", &sort_order, "key[,key2...]", 1184 "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..." 1185 " Please refer the man page for the complete list."), 1186 OPT_STRING(0, "fields", &field_order, "key[,keys...]", 1187 "output field(s): overhead, period, sample plus all of sort keys"), 1188 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples, 1189 "Show a column with the number of samples"), 1190 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param, 1191 NULL, "enables call-graph recording and display", 1192 &callchain_opt), 1193 OPT_CALLBACK(0, "call-graph", &callchain_param, 1194 "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]", 1195 top_callchain_help, &parse_callchain_opt), 1196 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain, 1197 "Accumulate callchains of children and show total overhead as well"), 1198 OPT_INTEGER(0, "max-stack", &top.max_stack, 1199 "Set the maximum stack depth when parsing the callchain. " 1200 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)), 1201 OPT_CALLBACK(0, "ignore-callees", NULL, "regex", 1202 "ignore callees of these functions in call graphs", 1203 report_parse_ignore_callees_opt), 1204 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period, 1205 "Show a column with the sum of periods"), 1206 OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]", 1207 "only consider symbols in these dsos"), 1208 OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]", 1209 "only consider symbols in these comms"), 1210 OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]", 1211 "only consider these symbols"), 1212 OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src, 1213 "Interleave source code with assembly code (default)"), 1214 OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw, 1215 "Display raw encoding of assembly instructions (default)"), 1216 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel, 1217 "Enable kernel symbol demangling"), 1218 OPT_STRING(0, "objdump", &objdump_path, "path", 1219 "objdump binary to use for disassembly and annotations"), 1220 OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style", 1221 "Specify disassembler style (e.g. -M intel for intel syntax)"), 1222 OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"), 1223 OPT_CALLBACK(0, "percent-limit", &top, "percent", 1224 "Don't show entries under that percent", parse_percent_limit), 1225 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute", 1226 "How to display percentage of filtered entries", parse_filter_percentage), 1227 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str, 1228 "width[,width...]", 1229 "don't try to adjust column width, use these fixed values"), 1230 OPT_UINTEGER(0, "proc-map-timeout", &opts->proc_map_timeout, 1231 "per thread proc mmap processing timeout in ms"), 1232 OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack, 1233 "branch any", "sample any taken branches", 1234 parse_branch_stack), 1235 OPT_CALLBACK('j', "branch-filter", &opts->branch_stack, 1236 "branch filter mask", "branch stack filter modes", 1237 parse_branch_stack), 1238 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace, 1239 "Show raw trace event output (do not use print fmt or plugins)"), 1240 OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy, 1241 "Show entries in a hierarchy"), 1242 OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"), 1243 OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize, 1244 "number of thread to run event synthesize"), 1245 OPT_END() 1246 }; 1247 const char * const top_usage[] = { 1248 "perf top [<options>]", 1249 NULL 1250 }; 1251 int status = hists__init(); 1252 1253 if (status < 0) 1254 return status; 1255 1256 top.evlist = perf_evlist__new(); 1257 if (top.evlist == NULL) 1258 return -ENOMEM; 1259 1260 status = perf_config(perf_top_config, &top); 1261 if (status) 1262 return status; 1263 1264 argc = parse_options(argc, argv, options, top_usage, 0); 1265 if (argc) 1266 usage_with_options(top_usage, options); 1267 1268 if (!top.evlist->nr_entries && 1269 perf_evlist__add_default(top.evlist) < 0) { 1270 pr_err("Not enough memory for event selector list\n"); 1271 goto out_delete_evlist; 1272 } 1273 1274 if (symbol_conf.report_hierarchy) { 1275 /* disable incompatible options */ 1276 symbol_conf.event_group = false; 1277 symbol_conf.cumulate_callchain = false; 1278 1279 if (field_order) { 1280 pr_err("Error: --hierarchy and --fields options cannot be used together\n"); 1281 parse_options_usage(top_usage, options, "fields", 0); 1282 parse_options_usage(NULL, options, "hierarchy", 0); 1283 goto out_delete_evlist; 1284 } 1285 } 1286 1287 sort__mode = SORT_MODE__TOP; 1288 /* display thread wants entries to be collapsed in a different tree */ 1289 perf_hpp_list.need_collapse = 1; 1290 1291 if (top.use_stdio) 1292 use_browser = 0; 1293 else if (top.use_tui) 1294 use_browser = 1; 1295 1296 setup_browser(false); 1297 1298 if (setup_sorting(top.evlist) < 0) { 1299 if (sort_order) 1300 parse_options_usage(top_usage, options, "s", 1); 1301 if (field_order) 1302 parse_options_usage(sort_order ? NULL : top_usage, 1303 options, "fields", 0); 1304 goto out_delete_evlist; 1305 } 1306 1307 status = target__validate(target); 1308 if (status) { 1309 target__strerror(target, status, errbuf, BUFSIZ); 1310 ui__warning("%s\n", errbuf); 1311 } 1312 1313 status = target__parse_uid(target); 1314 if (status) { 1315 int saved_errno = errno; 1316 1317 target__strerror(target, status, errbuf, BUFSIZ); 1318 ui__error("%s\n", errbuf); 1319 1320 status = -saved_errno; 1321 goto out_delete_evlist; 1322 } 1323 1324 if (target__none(target)) 1325 target->system_wide = true; 1326 1327 if (perf_evlist__create_maps(top.evlist, target) < 0) { 1328 ui__error("Couldn't create thread/CPU maps: %s\n", 1329 errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf))); 1330 goto out_delete_evlist; 1331 } 1332 1333 symbol_conf.nr_events = top.evlist->nr_entries; 1334 1335 if (top.delay_secs < 1) 1336 top.delay_secs = 1; 1337 1338 if (record_opts__config(opts)) { 1339 status = -EINVAL; 1340 goto out_delete_evlist; 1341 } 1342 1343 top.sym_evsel = perf_evlist__first(top.evlist); 1344 1345 if (!callchain_param.enabled) { 1346 symbol_conf.cumulate_callchain = false; 1347 perf_hpp__cancel_cumulate(); 1348 } 1349 1350 if (symbol_conf.cumulate_callchain && !callchain_param.order_set) 1351 callchain_param.order = ORDER_CALLER; 1352 1353 status = symbol__annotation_init(); 1354 if (status < 0) 1355 goto out_delete_evlist; 1356 1357 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL); 1358 if (symbol__init(NULL) < 0) 1359 return -1; 1360 1361 sort__setup_elide(stdout); 1362 1363 get_term_dimensions(&top.winsize); 1364 if (top.print_entries == 0) { 1365 perf_top__update_print_entries(&top); 1366 signal(SIGWINCH, winch_sig); 1367 } 1368 1369 status = __cmd_top(&top); 1370 1371 out_delete_evlist: 1372 perf_evlist__delete(top.evlist); 1373 1374 return status; 1375 } 1376