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