1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * builtin-top.c 4 * 5 * Builtin top command: Display a continuously updated profile of 6 * any workload, CPU or specific PID. 7 * 8 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com> 9 * 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com> 10 * 11 * Improvements and fixes by: 12 * 13 * Arjan van de Ven <arjan@linux.intel.com> 14 * Yanmin Zhang <yanmin.zhang@intel.com> 15 * Wu Fengguang <fengguang.wu@intel.com> 16 * Mike Galbraith <efault@gmx.de> 17 * Paul Mackerras <paulus@samba.org> 18 */ 19 #include "builtin.h" 20 21 #include "perf.h" 22 23 #include "util/annotate.h" 24 #include "util/bpf-event.h" 25 #include "util/cgroup.h" 26 #include "util/config.h" 27 #include "util/color.h" 28 #include "util/dso.h" 29 #include "util/evlist.h" 30 #include "util/evsel.h" 31 #include "util/evsel_config.h" 32 #include "util/event.h" 33 #include "util/machine.h" 34 #include "util/map.h" 35 #include "util/mmap.h" 36 #include "util/session.h" 37 #include "util/thread.h" 38 #include "util/symbol.h" 39 #include "util/synthetic-events.h" 40 #include "util/top.h" 41 #include "util/util.h" 42 #include <linux/rbtree.h> 43 #include <subcmd/parse-options.h> 44 #include "util/parse-events.h" 45 #include "util/callchain.h" 46 #include "util/cpumap.h" 47 #include "util/sort.h" 48 #include "util/string2.h" 49 #include "util/term.h" 50 #include "util/intlist.h" 51 #include "util/parse-branch-options.h" 52 #include "arch/common.h" 53 #include "ui/ui.h" 54 55 #include "util/debug.h" 56 #include "util/ordered-events.h" 57 #include "util/pfm.h" 58 59 #include <assert.h> 60 #include <elf.h> 61 #include <fcntl.h> 62 63 #include <stdio.h> 64 #include <termios.h> 65 #include <unistd.h> 66 #include <inttypes.h> 67 68 #include <errno.h> 69 #include <time.h> 70 #include <sched.h> 71 #include <signal.h> 72 73 #include <sys/syscall.h> 74 #include <sys/ioctl.h> 75 #include <poll.h> 76 #include <sys/prctl.h> 77 #include <sys/wait.h> 78 #include <sys/uio.h> 79 #include <sys/utsname.h> 80 #include <sys/mman.h> 81 82 #include <linux/stringify.h> 83 #include <linux/time64.h> 84 #include <linux/types.h> 85 #include <linux/err.h> 86 87 #include <linux/ctype.h> 88 #include <perf/mmap.h> 89 90 static volatile int done; 91 static volatile int resize; 92 93 #define HEADER_LINE_NR 5 94 95 static void perf_top__update_print_entries(struct perf_top *top) 96 { 97 top->print_entries = top->winsize.ws_row - HEADER_LINE_NR; 98 } 99 100 static void winch_sig(int sig __maybe_unused) 101 { 102 resize = 1; 103 } 104 105 static void perf_top__resize(struct perf_top *top) 106 { 107 get_term_dimensions(&top->winsize); 108 perf_top__update_print_entries(top); 109 } 110 111 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he) 112 { 113 struct evsel *evsel; 114 struct symbol *sym; 115 struct annotation *notes; 116 struct map *map; 117 int err = -1; 118 119 if (!he || !he->ms.sym) 120 return -1; 121 122 evsel = hists_to_evsel(he->hists); 123 124 sym = he->ms.sym; 125 map = he->ms.map; 126 127 /* 128 * We can't annotate with just /proc/kallsyms 129 */ 130 if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS && 131 !dso__is_kcore(map->dso)) { 132 pr_err("Can't annotate %s: No vmlinux file was found in the " 133 "path\n", sym->name); 134 sleep(1); 135 return -1; 136 } 137 138 notes = symbol__annotation(sym); 139 pthread_mutex_lock(¬es->lock); 140 141 if (!symbol__hists(sym, top->evlist->core.nr_entries)) { 142 pthread_mutex_unlock(¬es->lock); 143 pr_err("Not enough memory for annotating '%s' symbol!\n", 144 sym->name); 145 sleep(1); 146 return err; 147 } 148 149 err = symbol__annotate(&he->ms, evsel, &top->annotation_opts, NULL); 150 if (err == 0) { 151 top->sym_filter_entry = he; 152 } else { 153 char msg[BUFSIZ]; 154 symbol__strerror_disassemble(&he->ms, err, msg, sizeof(msg)); 155 pr_err("Couldn't annotate %s: %s\n", sym->name, msg); 156 } 157 158 pthread_mutex_unlock(¬es->lock); 159 return err; 160 } 161 162 static void __zero_source_counters(struct hist_entry *he) 163 { 164 struct symbol *sym = he->ms.sym; 165 symbol__annotate_zero_histograms(sym); 166 } 167 168 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip) 169 { 170 struct utsname uts; 171 int err = uname(&uts); 172 173 ui__warning("Out of bounds address found:\n\n" 174 "Addr: %" PRIx64 "\n" 175 "DSO: %s %c\n" 176 "Map: %" PRIx64 "-%" PRIx64 "\n" 177 "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n" 178 "Arch: %s\n" 179 "Kernel: %s\n" 180 "Tools: %s\n\n" 181 "Not all samples will be on the annotation output.\n\n" 182 "Please report to linux-kernel@vger.kernel.org\n", 183 ip, map->dso->long_name, dso__symtab_origin(map->dso), 184 map->start, map->end, sym->start, sym->end, 185 sym->binding == STB_GLOBAL ? 'g' : 186 sym->binding == STB_LOCAL ? 'l' : 'w', sym->name, 187 err ? "[unknown]" : uts.machine, 188 err ? "[unknown]" : uts.release, perf_version_string); 189 if (use_browser <= 0) 190 sleep(5); 191 192 map->erange_warned = true; 193 } 194 195 static void perf_top__record_precise_ip(struct perf_top *top, 196 struct hist_entry *he, 197 struct perf_sample *sample, 198 struct evsel *evsel, u64 ip) 199 { 200 struct annotation *notes; 201 struct symbol *sym = he->ms.sym; 202 int err = 0; 203 204 if (sym == NULL || (use_browser == 0 && 205 (top->sym_filter_entry == NULL || 206 top->sym_filter_entry->ms.sym != sym))) 207 return; 208 209 notes = symbol__annotation(sym); 210 211 if (pthread_mutex_trylock(¬es->lock)) 212 return; 213 214 err = hist_entry__inc_addr_samples(he, sample, evsel, ip); 215 216 pthread_mutex_unlock(¬es->lock); 217 218 if (unlikely(err)) { 219 /* 220 * This function is now called with he->hists->lock held. 221 * Release it before going to sleep. 222 */ 223 pthread_mutex_unlock(&he->hists->lock); 224 225 if (err == -ERANGE && !he->ms.map->erange_warned) 226 ui__warn_map_erange(he->ms.map, sym, ip); 227 else if (err == -ENOMEM) { 228 pr_err("Not enough memory for annotating '%s' symbol!\n", 229 sym->name); 230 sleep(1); 231 } 232 233 pthread_mutex_lock(&he->hists->lock); 234 } 235 } 236 237 static void perf_top__show_details(struct perf_top *top) 238 { 239 struct hist_entry *he = top->sym_filter_entry; 240 struct evsel *evsel; 241 struct annotation *notes; 242 struct symbol *symbol; 243 int more; 244 245 if (!he) 246 return; 247 248 evsel = hists_to_evsel(he->hists); 249 250 symbol = he->ms.sym; 251 notes = symbol__annotation(symbol); 252 253 pthread_mutex_lock(¬es->lock); 254 255 symbol__calc_percent(symbol, evsel); 256 257 if (notes->src == NULL) 258 goto out_unlock; 259 260 printf("Showing %s for %s\n", evsel__name(top->sym_evsel), symbol->name); 261 printf(" Events Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt); 262 263 more = symbol__annotate_printf(&he->ms, top->sym_evsel, &top->annotation_opts); 264 265 if (top->evlist->enabled) { 266 if (top->zero) 267 symbol__annotate_zero_histogram(symbol, top->sym_evsel->core.idx); 268 else 269 symbol__annotate_decay_histogram(symbol, top->sym_evsel->core.idx); 270 } 271 if (more != 0) 272 printf("%d lines not displayed, maybe increase display entries [e]\n", more); 273 out_unlock: 274 pthread_mutex_unlock(¬es->lock); 275 } 276 277 static void perf_top__resort_hists(struct perf_top *t) 278 { 279 struct evlist *evlist = t->evlist; 280 struct evsel *pos; 281 282 evlist__for_each_entry(evlist, pos) { 283 struct hists *hists = evsel__hists(pos); 284 285 /* 286 * unlink existing entries so that they can be linked 287 * in a correct order in hists__match() below. 288 */ 289 hists__unlink(hists); 290 291 if (evlist->enabled) { 292 if (t->zero) { 293 hists__delete_entries(hists); 294 } else { 295 hists__decay_entries(hists, t->hide_user_symbols, 296 t->hide_kernel_symbols); 297 } 298 } 299 300 hists__collapse_resort(hists, NULL); 301 302 /* Non-group events are considered as leader */ 303 if (symbol_conf.event_group && !evsel__is_group_leader(pos)) { 304 struct hists *leader_hists = evsel__hists(evsel__leader(pos)); 305 306 hists__match(leader_hists, hists); 307 hists__link(leader_hists, hists); 308 } 309 } 310 311 evlist__for_each_entry(evlist, pos) { 312 evsel__output_resort(pos, NULL); 313 } 314 } 315 316 static void perf_top__print_sym_table(struct perf_top *top) 317 { 318 char bf[160]; 319 int printed = 0; 320 const int win_width = top->winsize.ws_col - 1; 321 struct evsel *evsel = top->sym_evsel; 322 struct hists *hists = evsel__hists(evsel); 323 324 puts(CONSOLE_CLEAR); 325 326 perf_top__header_snprintf(top, bf, sizeof(bf)); 327 printf("%s\n", bf); 328 329 printf("%-*.*s\n", win_width, win_width, graph_dotted_line); 330 331 if (!top->record_opts.overwrite && 332 (top->evlist->stats.nr_lost_warned != 333 top->evlist->stats.nr_events[PERF_RECORD_LOST])) { 334 top->evlist->stats.nr_lost_warned = 335 top->evlist->stats.nr_events[PERF_RECORD_LOST]; 336 color_fprintf(stdout, PERF_COLOR_RED, 337 "WARNING: LOST %d chunks, Check IO/CPU overload", 338 top->evlist->stats.nr_lost_warned); 339 ++printed; 340 } 341 342 if (top->sym_filter_entry) { 343 perf_top__show_details(top); 344 return; 345 } 346 347 perf_top__resort_hists(top); 348 349 hists__output_recalc_col_len(hists, top->print_entries - printed); 350 putchar('\n'); 351 hists__fprintf(hists, false, top->print_entries - printed, win_width, 352 top->min_percent, stdout, !symbol_conf.use_callchain); 353 } 354 355 static void prompt_integer(int *target, const char *msg) 356 { 357 char *buf = malloc(0), *p; 358 size_t dummy = 0; 359 int tmp; 360 361 fprintf(stdout, "\n%s: ", msg); 362 if (getline(&buf, &dummy, stdin) < 0) 363 return; 364 365 p = strchr(buf, '\n'); 366 if (p) 367 *p = 0; 368 369 p = buf; 370 while(*p) { 371 if (!isdigit(*p)) 372 goto out_free; 373 p++; 374 } 375 tmp = strtoul(buf, NULL, 10); 376 *target = tmp; 377 out_free: 378 free(buf); 379 } 380 381 static void prompt_percent(int *target, const char *msg) 382 { 383 int tmp = 0; 384 385 prompt_integer(&tmp, msg); 386 if (tmp >= 0 && tmp <= 100) 387 *target = tmp; 388 } 389 390 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg) 391 { 392 char *buf = malloc(0), *p; 393 struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL; 394 struct hists *hists = evsel__hists(top->sym_evsel); 395 struct rb_node *next; 396 size_t dummy = 0; 397 398 /* zero counters of active symbol */ 399 if (syme) { 400 __zero_source_counters(syme); 401 top->sym_filter_entry = NULL; 402 } 403 404 fprintf(stdout, "\n%s: ", msg); 405 if (getline(&buf, &dummy, stdin) < 0) 406 goto out_free; 407 408 p = strchr(buf, '\n'); 409 if (p) 410 *p = 0; 411 412 next = rb_first_cached(&hists->entries); 413 while (next) { 414 n = rb_entry(next, struct hist_entry, rb_node); 415 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) { 416 found = n; 417 break; 418 } 419 next = rb_next(&n->rb_node); 420 } 421 422 if (!found) { 423 fprintf(stderr, "Sorry, %s is not active.\n", buf); 424 sleep(1); 425 } else 426 perf_top__parse_source(top, found); 427 428 out_free: 429 free(buf); 430 } 431 432 static void perf_top__print_mapped_keys(struct perf_top *top) 433 { 434 char *name = NULL; 435 436 if (top->sym_filter_entry) { 437 struct symbol *sym = top->sym_filter_entry->ms.sym; 438 name = sym->name; 439 } 440 441 fprintf(stdout, "\nMapped keys:\n"); 442 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", top->delay_secs); 443 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", top->print_entries); 444 445 if (top->evlist->core.nr_entries > 1) 446 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", evsel__name(top->sym_evsel)); 447 448 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", top->count_filter); 449 450 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt); 451 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL"); 452 fprintf(stdout, "\t[S] stop annotation.\n"); 453 454 fprintf(stdout, 455 "\t[K] hide kernel symbols. \t(%s)\n", 456 top->hide_kernel_symbols ? "yes" : "no"); 457 fprintf(stdout, 458 "\t[U] hide user symbols. \t(%s)\n", 459 top->hide_user_symbols ? "yes" : "no"); 460 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", top->zero ? 1 : 0); 461 fprintf(stdout, "\t[qQ] quit.\n"); 462 } 463 464 static int perf_top__key_mapped(struct perf_top *top, int c) 465 { 466 switch (c) { 467 case 'd': 468 case 'e': 469 case 'f': 470 case 'z': 471 case 'q': 472 case 'Q': 473 case 'K': 474 case 'U': 475 case 'F': 476 case 's': 477 case 'S': 478 return 1; 479 case 'E': 480 return top->evlist->core.nr_entries > 1 ? 1 : 0; 481 default: 482 break; 483 } 484 485 return 0; 486 } 487 488 static bool perf_top__handle_keypress(struct perf_top *top, int c) 489 { 490 bool ret = true; 491 492 if (!perf_top__key_mapped(top, c)) { 493 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN }; 494 struct termios save; 495 496 perf_top__print_mapped_keys(top); 497 fprintf(stdout, "\nEnter selection, or unmapped key to continue: "); 498 fflush(stdout); 499 500 set_term_quiet_input(&save); 501 502 poll(&stdin_poll, 1, -1); 503 c = getc(stdin); 504 505 tcsetattr(0, TCSAFLUSH, &save); 506 if (!perf_top__key_mapped(top, c)) 507 return ret; 508 } 509 510 switch (c) { 511 case 'd': 512 prompt_integer(&top->delay_secs, "Enter display delay"); 513 if (top->delay_secs < 1) 514 top->delay_secs = 1; 515 break; 516 case 'e': 517 prompt_integer(&top->print_entries, "Enter display entries (lines)"); 518 if (top->print_entries == 0) { 519 perf_top__resize(top); 520 signal(SIGWINCH, winch_sig); 521 } else { 522 signal(SIGWINCH, SIG_DFL); 523 } 524 break; 525 case 'E': 526 if (top->evlist->core.nr_entries > 1) { 527 /* Select 0 as the default event: */ 528 int counter = 0; 529 530 fprintf(stderr, "\nAvailable events:"); 531 532 evlist__for_each_entry(top->evlist, top->sym_evsel) 533 fprintf(stderr, "\n\t%d %s", top->sym_evsel->core.idx, evsel__name(top->sym_evsel)); 534 535 prompt_integer(&counter, "Enter details event counter"); 536 537 if (counter >= top->evlist->core.nr_entries) { 538 top->sym_evsel = evlist__first(top->evlist); 539 fprintf(stderr, "Sorry, no such event, using %s.\n", evsel__name(top->sym_evsel)); 540 sleep(1); 541 break; 542 } 543 evlist__for_each_entry(top->evlist, top->sym_evsel) 544 if (top->sym_evsel->core.idx == counter) 545 break; 546 } else 547 top->sym_evsel = evlist__first(top->evlist); 548 break; 549 case 'f': 550 prompt_integer(&top->count_filter, "Enter display event count filter"); 551 break; 552 case 'F': 553 prompt_percent(&top->annotation_opts.min_pcnt, 554 "Enter details display event filter (percent)"); 555 break; 556 case 'K': 557 top->hide_kernel_symbols = !top->hide_kernel_symbols; 558 break; 559 case 'q': 560 case 'Q': 561 printf("exiting.\n"); 562 if (top->dump_symtab) 563 perf_session__fprintf_dsos(top->session, stderr); 564 ret = false; 565 break; 566 case 's': 567 perf_top__prompt_symbol(top, "Enter details symbol"); 568 break; 569 case 'S': 570 if (!top->sym_filter_entry) 571 break; 572 else { 573 struct hist_entry *syme = top->sym_filter_entry; 574 575 top->sym_filter_entry = NULL; 576 __zero_source_counters(syme); 577 } 578 break; 579 case 'U': 580 top->hide_user_symbols = !top->hide_user_symbols; 581 break; 582 case 'z': 583 top->zero = !top->zero; 584 break; 585 default: 586 break; 587 } 588 589 return ret; 590 } 591 592 static void perf_top__sort_new_samples(void *arg) 593 { 594 struct perf_top *t = arg; 595 596 if (t->evlist->selected != NULL) 597 t->sym_evsel = t->evlist->selected; 598 599 perf_top__resort_hists(t); 600 601 if (t->lost || t->drop) 602 pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n"); 603 } 604 605 static void stop_top(void) 606 { 607 session_done = 1; 608 done = 1; 609 } 610 611 static void *display_thread_tui(void *arg) 612 { 613 struct evsel *pos; 614 struct perf_top *top = arg; 615 const char *help = "For a higher level overview, try: perf top --sort comm,dso"; 616 struct hist_browser_timer hbt = { 617 .timer = perf_top__sort_new_samples, 618 .arg = top, 619 .refresh = top->delay_secs, 620 }; 621 int ret; 622 623 /* In order to read symbols from other namespaces perf to needs to call 624 * setns(2). This isn't permitted if the struct_fs has multiple users. 625 * unshare(2) the fs so that we may continue to setns into namespaces 626 * that we're observing. 627 */ 628 unshare(CLONE_FS); 629 630 prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0); 631 632 repeat: 633 perf_top__sort_new_samples(top); 634 635 /* 636 * Initialize the uid_filter_str, in the future the TUI will allow 637 * Zooming in/out UIDs. For now just use whatever the user passed 638 * via --uid. 639 */ 640 evlist__for_each_entry(top->evlist, pos) { 641 struct hists *hists = evsel__hists(pos); 642 hists->uid_filter_str = top->record_opts.target.uid_str; 643 } 644 645 ret = evlist__tui_browse_hists(top->evlist, help, &hbt, top->min_percent, 646 &top->session->header.env, !top->record_opts.overwrite, 647 &top->annotation_opts); 648 if (ret == K_RELOAD) { 649 top->zero = true; 650 goto repeat; 651 } else 652 stop_top(); 653 654 return NULL; 655 } 656 657 static void display_sig(int sig __maybe_unused) 658 { 659 stop_top(); 660 } 661 662 static void display_setup_sig(void) 663 { 664 signal(SIGSEGV, sighandler_dump_stack); 665 signal(SIGFPE, sighandler_dump_stack); 666 signal(SIGINT, display_sig); 667 signal(SIGQUIT, display_sig); 668 signal(SIGTERM, display_sig); 669 } 670 671 static void *display_thread(void *arg) 672 { 673 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN }; 674 struct termios save; 675 struct perf_top *top = arg; 676 int delay_msecs, c; 677 678 /* In order to read symbols from other namespaces perf to needs to call 679 * setns(2). This isn't permitted if the struct_fs has multiple users. 680 * unshare(2) the fs so that we may continue to setns into namespaces 681 * that we're observing. 682 */ 683 unshare(CLONE_FS); 684 685 prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0); 686 687 display_setup_sig(); 688 pthread__unblock_sigwinch(); 689 repeat: 690 delay_msecs = top->delay_secs * MSEC_PER_SEC; 691 set_term_quiet_input(&save); 692 /* trash return*/ 693 clearerr(stdin); 694 if (poll(&stdin_poll, 1, 0) > 0) 695 getc(stdin); 696 697 while (!done) { 698 perf_top__print_sym_table(top); 699 /* 700 * Either timeout expired or we got an EINTR due to SIGWINCH, 701 * refresh screen in both cases. 702 */ 703 switch (poll(&stdin_poll, 1, delay_msecs)) { 704 case 0: 705 continue; 706 case -1: 707 if (errno == EINTR) 708 continue; 709 __fallthrough; 710 default: 711 c = getc(stdin); 712 tcsetattr(0, TCSAFLUSH, &save); 713 714 if (perf_top__handle_keypress(top, c)) 715 goto repeat; 716 stop_top(); 717 } 718 } 719 720 tcsetattr(0, TCSAFLUSH, &save); 721 return NULL; 722 } 723 724 static int hist_iter__top_callback(struct hist_entry_iter *iter, 725 struct addr_location *al, bool single, 726 void *arg) 727 { 728 struct perf_top *top = arg; 729 struct hist_entry *he = iter->he; 730 struct evsel *evsel = iter->evsel; 731 732 if (perf_hpp_list.sym && single) 733 perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr); 734 735 hist__account_cycles(iter->sample->branch_stack, al, iter->sample, 736 !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY), 737 NULL); 738 return 0; 739 } 740 741 static void perf_event__process_sample(struct perf_tool *tool, 742 const union perf_event *event, 743 struct evsel *evsel, 744 struct perf_sample *sample, 745 struct machine *machine) 746 { 747 struct perf_top *top = container_of(tool, struct perf_top, tool); 748 struct addr_location al; 749 750 if (!machine && perf_guest) { 751 static struct intlist *seen; 752 753 if (!seen) 754 seen = intlist__new(NULL); 755 756 if (!intlist__has_entry(seen, sample->pid)) { 757 pr_err("Can't find guest [%d]'s kernel information\n", 758 sample->pid); 759 intlist__add(seen, sample->pid); 760 } 761 return; 762 } 763 764 if (!machine) { 765 pr_err("%u unprocessable samples recorded.\r", 766 top->session->evlist->stats.nr_unprocessable_samples++); 767 return; 768 } 769 770 if (event->header.misc & PERF_RECORD_MISC_EXACT_IP) 771 top->exact_samples++; 772 773 if (machine__resolve(machine, &al, sample) < 0) 774 return; 775 776 if (top->stitch_lbr) 777 al.thread->lbr_stitch_enable = true; 778 779 if (!machine->kptr_restrict_warned && 780 symbol_conf.kptr_restrict && 781 al.cpumode == PERF_RECORD_MISC_KERNEL) { 782 if (!evlist__exclude_kernel(top->session->evlist)) { 783 ui__warning( 784 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n" 785 "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n" 786 "Kernel%s samples will not be resolved.\n", 787 al.map && map__has_symbols(al.map) ? 788 " modules" : ""); 789 if (use_browser <= 0) 790 sleep(5); 791 } 792 machine->kptr_restrict_warned = true; 793 } 794 795 if (al.sym == NULL && al.map != NULL) { 796 const char *msg = "Kernel samples will not be resolved.\n"; 797 /* 798 * As we do lazy loading of symtabs we only will know if the 799 * specified vmlinux file is invalid when we actually have a 800 * hit in kernel space and then try to load it. So if we get 801 * here and there are _no_ symbols in the DSO backing the 802 * kernel map, bail out. 803 * 804 * We may never get here, for instance, if we use -K/ 805 * --hide-kernel-symbols, even if the user specifies an 806 * invalid --vmlinux ;-) 807 */ 808 if (!machine->kptr_restrict_warned && !top->vmlinux_warned && 809 __map__is_kernel(al.map) && map__has_symbols(al.map)) { 810 if (symbol_conf.vmlinux_name) { 811 char serr[256]; 812 dso__strerror_load(al.map->dso, serr, sizeof(serr)); 813 ui__warning("The %s file can't be used: %s\n%s", 814 symbol_conf.vmlinux_name, serr, msg); 815 } else { 816 ui__warning("A vmlinux file was not found.\n%s", 817 msg); 818 } 819 820 if (use_browser <= 0) 821 sleep(5); 822 top->vmlinux_warned = true; 823 } 824 } 825 826 if (al.sym == NULL || !al.sym->idle) { 827 struct hists *hists = evsel__hists(evsel); 828 struct hist_entry_iter iter = { 829 .evsel = evsel, 830 .sample = sample, 831 .add_entry_cb = hist_iter__top_callback, 832 }; 833 834 if (symbol_conf.cumulate_callchain) 835 iter.ops = &hist_iter_cumulative; 836 else 837 iter.ops = &hist_iter_normal; 838 839 pthread_mutex_lock(&hists->lock); 840 841 if (hist_entry_iter__add(&iter, &al, top->max_stack, top) < 0) 842 pr_err("Problem incrementing symbol period, skipping event\n"); 843 844 pthread_mutex_unlock(&hists->lock); 845 } 846 847 addr_location__put(&al); 848 } 849 850 static void 851 perf_top__process_lost(struct perf_top *top, union perf_event *event, 852 struct evsel *evsel) 853 { 854 top->lost += event->lost.lost; 855 top->lost_total += event->lost.lost; 856 evsel->evlist->stats.total_lost += event->lost.lost; 857 } 858 859 static void 860 perf_top__process_lost_samples(struct perf_top *top, 861 union perf_event *event, 862 struct evsel *evsel) 863 { 864 top->lost += event->lost_samples.lost; 865 top->lost_total += event->lost_samples.lost; 866 evsel->evlist->stats.total_lost_samples += event->lost_samples.lost; 867 } 868 869 static u64 last_timestamp; 870 871 static void perf_top__mmap_read_idx(struct perf_top *top, int idx) 872 { 873 struct record_opts *opts = &top->record_opts; 874 struct evlist *evlist = top->evlist; 875 struct mmap *md; 876 union perf_event *event; 877 878 md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx]; 879 if (perf_mmap__read_init(&md->core) < 0) 880 return; 881 882 while ((event = perf_mmap__read_event(&md->core)) != NULL) { 883 int ret; 884 885 ret = evlist__parse_sample_timestamp(evlist, event, &last_timestamp); 886 if (ret && ret != -1) 887 break; 888 889 ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0, NULL); 890 if (ret) 891 break; 892 893 perf_mmap__consume(&md->core); 894 895 if (top->qe.rotate) { 896 pthread_mutex_lock(&top->qe.mutex); 897 top->qe.rotate = false; 898 pthread_cond_signal(&top->qe.cond); 899 pthread_mutex_unlock(&top->qe.mutex); 900 } 901 } 902 903 perf_mmap__read_done(&md->core); 904 } 905 906 static void perf_top__mmap_read(struct perf_top *top) 907 { 908 bool overwrite = top->record_opts.overwrite; 909 struct evlist *evlist = top->evlist; 910 int i; 911 912 if (overwrite) 913 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING); 914 915 for (i = 0; i < top->evlist->core.nr_mmaps; i++) 916 perf_top__mmap_read_idx(top, i); 917 918 if (overwrite) { 919 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY); 920 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING); 921 } 922 } 923 924 /* 925 * Check per-event overwrite term. 926 * perf top should support consistent term for all events. 927 * - All events don't have per-event term 928 * E.g. "cpu/cpu-cycles/,cpu/instructions/" 929 * Nothing change, return 0. 930 * - All events have same per-event term 931 * E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/ 932 * Using the per-event setting to replace the opts->overwrite if 933 * they are different, then return 0. 934 * - Events have different per-event term 935 * E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/" 936 * Return -1 937 * - Some of the event set per-event term, but some not. 938 * E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/" 939 * Return -1 940 */ 941 static int perf_top__overwrite_check(struct perf_top *top) 942 { 943 struct record_opts *opts = &top->record_opts; 944 struct evlist *evlist = top->evlist; 945 struct evsel_config_term *term; 946 struct list_head *config_terms; 947 struct evsel *evsel; 948 int set, overwrite = -1; 949 950 evlist__for_each_entry(evlist, evsel) { 951 set = -1; 952 config_terms = &evsel->config_terms; 953 list_for_each_entry(term, config_terms, list) { 954 if (term->type == EVSEL__CONFIG_TERM_OVERWRITE) 955 set = term->val.overwrite ? 1 : 0; 956 } 957 958 /* no term for current and previous event (likely) */ 959 if ((overwrite < 0) && (set < 0)) 960 continue; 961 962 /* has term for both current and previous event, compare */ 963 if ((overwrite >= 0) && (set >= 0) && (overwrite != set)) 964 return -1; 965 966 /* no term for current event but has term for previous one */ 967 if ((overwrite >= 0) && (set < 0)) 968 return -1; 969 970 /* has term for current event */ 971 if ((overwrite < 0) && (set >= 0)) { 972 /* if it's first event, set overwrite */ 973 if (evsel == evlist__first(evlist)) 974 overwrite = set; 975 else 976 return -1; 977 } 978 } 979 980 if ((overwrite >= 0) && (opts->overwrite != overwrite)) 981 opts->overwrite = overwrite; 982 983 return 0; 984 } 985 986 static int perf_top_overwrite_fallback(struct perf_top *top, 987 struct evsel *evsel) 988 { 989 struct record_opts *opts = &top->record_opts; 990 struct evlist *evlist = top->evlist; 991 struct evsel *counter; 992 993 if (!opts->overwrite) 994 return 0; 995 996 /* only fall back when first event fails */ 997 if (evsel != evlist__first(evlist)) 998 return 0; 999 1000 evlist__for_each_entry(evlist, counter) 1001 counter->core.attr.write_backward = false; 1002 opts->overwrite = false; 1003 pr_debug2("fall back to non-overwrite mode\n"); 1004 return 1; 1005 } 1006 1007 static int perf_top__start_counters(struct perf_top *top) 1008 { 1009 char msg[BUFSIZ]; 1010 struct evsel *counter; 1011 struct evlist *evlist = top->evlist; 1012 struct record_opts *opts = &top->record_opts; 1013 1014 if (perf_top__overwrite_check(top)) { 1015 ui__error("perf top only support consistent per-event " 1016 "overwrite setting for all events\n"); 1017 goto out_err; 1018 } 1019 1020 evlist__config(evlist, opts, &callchain_param); 1021 1022 evlist__for_each_entry(evlist, counter) { 1023 try_again: 1024 if (evsel__open(counter, top->evlist->core.user_requested_cpus, 1025 top->evlist->core.threads) < 0) { 1026 1027 /* 1028 * Specially handle overwrite fall back. 1029 * Because perf top is the only tool which has 1030 * overwrite mode by default, support 1031 * both overwrite and non-overwrite mode, and 1032 * require consistent mode for all events. 1033 * 1034 * May move it to generic code with more tools 1035 * have similar attribute. 1036 */ 1037 if (perf_missing_features.write_backward && 1038 perf_top_overwrite_fallback(top, counter)) 1039 goto try_again; 1040 1041 if (evsel__fallback(counter, errno, msg, sizeof(msg))) { 1042 if (verbose > 0) 1043 ui__warning("%s\n", msg); 1044 goto try_again; 1045 } 1046 1047 evsel__open_strerror(counter, &opts->target, errno, msg, sizeof(msg)); 1048 ui__error("%s\n", msg); 1049 goto out_err; 1050 } 1051 } 1052 1053 if (evlist__mmap(evlist, opts->mmap_pages) < 0) { 1054 ui__error("Failed to mmap with %d (%s)\n", 1055 errno, str_error_r(errno, msg, sizeof(msg))); 1056 goto out_err; 1057 } 1058 1059 return 0; 1060 1061 out_err: 1062 return -1; 1063 } 1064 1065 static int callchain_param__setup_sample_type(struct callchain_param *callchain) 1066 { 1067 if (callchain->mode != CHAIN_NONE) { 1068 if (callchain_register_param(callchain) < 0) { 1069 ui__error("Can't register callchain params.\n"); 1070 return -EINVAL; 1071 } 1072 } 1073 1074 return 0; 1075 } 1076 1077 static struct ordered_events *rotate_queues(struct perf_top *top) 1078 { 1079 struct ordered_events *in = top->qe.in; 1080 1081 if (top->qe.in == &top->qe.data[1]) 1082 top->qe.in = &top->qe.data[0]; 1083 else 1084 top->qe.in = &top->qe.data[1]; 1085 1086 return in; 1087 } 1088 1089 static void *process_thread(void *arg) 1090 { 1091 struct perf_top *top = arg; 1092 1093 while (!done) { 1094 struct ordered_events *out, *in = top->qe.in; 1095 1096 if (!in->nr_events) { 1097 usleep(100); 1098 continue; 1099 } 1100 1101 out = rotate_queues(top); 1102 1103 pthread_mutex_lock(&top->qe.mutex); 1104 top->qe.rotate = true; 1105 pthread_cond_wait(&top->qe.cond, &top->qe.mutex); 1106 pthread_mutex_unlock(&top->qe.mutex); 1107 1108 if (ordered_events__flush(out, OE_FLUSH__TOP)) 1109 pr_err("failed to process events\n"); 1110 } 1111 1112 return NULL; 1113 } 1114 1115 /* 1116 * Allow only 'top->delay_secs' seconds behind samples. 1117 */ 1118 static int should_drop(struct ordered_event *qevent, struct perf_top *top) 1119 { 1120 union perf_event *event = qevent->event; 1121 u64 delay_timestamp; 1122 1123 if (event->header.type != PERF_RECORD_SAMPLE) 1124 return false; 1125 1126 delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC; 1127 return delay_timestamp < last_timestamp; 1128 } 1129 1130 static int deliver_event(struct ordered_events *qe, 1131 struct ordered_event *qevent) 1132 { 1133 struct perf_top *top = qe->data; 1134 struct evlist *evlist = top->evlist; 1135 struct perf_session *session = top->session; 1136 union perf_event *event = qevent->event; 1137 struct perf_sample sample; 1138 struct evsel *evsel; 1139 struct machine *machine; 1140 int ret = -1; 1141 1142 if (should_drop(qevent, top)) { 1143 top->drop++; 1144 top->drop_total++; 1145 return 0; 1146 } 1147 1148 ret = evlist__parse_sample(evlist, event, &sample); 1149 if (ret) { 1150 pr_err("Can't parse sample, err = %d\n", ret); 1151 goto next_event; 1152 } 1153 1154 evsel = evlist__id2evsel(session->evlist, sample.id); 1155 assert(evsel != NULL); 1156 1157 if (event->header.type == PERF_RECORD_SAMPLE) { 1158 if (evswitch__discard(&top->evswitch, evsel)) 1159 return 0; 1160 ++top->samples; 1161 } 1162 1163 switch (sample.cpumode) { 1164 case PERF_RECORD_MISC_USER: 1165 ++top->us_samples; 1166 if (top->hide_user_symbols) 1167 goto next_event; 1168 machine = &session->machines.host; 1169 break; 1170 case PERF_RECORD_MISC_KERNEL: 1171 ++top->kernel_samples; 1172 if (top->hide_kernel_symbols) 1173 goto next_event; 1174 machine = &session->machines.host; 1175 break; 1176 case PERF_RECORD_MISC_GUEST_KERNEL: 1177 ++top->guest_kernel_samples; 1178 machine = perf_session__find_machine(session, 1179 sample.pid); 1180 break; 1181 case PERF_RECORD_MISC_GUEST_USER: 1182 ++top->guest_us_samples; 1183 /* 1184 * TODO: we don't process guest user from host side 1185 * except simple counting. 1186 */ 1187 goto next_event; 1188 default: 1189 if (event->header.type == PERF_RECORD_SAMPLE) 1190 goto next_event; 1191 machine = &session->machines.host; 1192 break; 1193 } 1194 1195 if (event->header.type == PERF_RECORD_SAMPLE) { 1196 perf_event__process_sample(&top->tool, event, evsel, 1197 &sample, machine); 1198 } else if (event->header.type == PERF_RECORD_LOST) { 1199 perf_top__process_lost(top, event, evsel); 1200 } else if (event->header.type == PERF_RECORD_LOST_SAMPLES) { 1201 perf_top__process_lost_samples(top, event, evsel); 1202 } else if (event->header.type < PERF_RECORD_MAX) { 1203 events_stats__inc(&session->evlist->stats, event->header.type); 1204 machine__process_event(machine, event, &sample); 1205 } else 1206 ++session->evlist->stats.nr_unknown_events; 1207 1208 ret = 0; 1209 next_event: 1210 return ret; 1211 } 1212 1213 static void init_process_thread(struct perf_top *top) 1214 { 1215 ordered_events__init(&top->qe.data[0], deliver_event, top); 1216 ordered_events__init(&top->qe.data[1], deliver_event, top); 1217 ordered_events__set_copy_on_queue(&top->qe.data[0], true); 1218 ordered_events__set_copy_on_queue(&top->qe.data[1], true); 1219 top->qe.in = &top->qe.data[0]; 1220 pthread_mutex_init(&top->qe.mutex, NULL); 1221 pthread_cond_init(&top->qe.cond, NULL); 1222 } 1223 1224 static int __cmd_top(struct perf_top *top) 1225 { 1226 struct record_opts *opts = &top->record_opts; 1227 pthread_t thread, thread_process; 1228 int ret; 1229 1230 if (!top->annotation_opts.objdump_path) { 1231 ret = perf_env__lookup_objdump(&top->session->header.env, 1232 &top->annotation_opts.objdump_path); 1233 if (ret) 1234 return ret; 1235 } 1236 1237 ret = callchain_param__setup_sample_type(&callchain_param); 1238 if (ret) 1239 return ret; 1240 1241 if (perf_session__register_idle_thread(top->session) < 0) 1242 return ret; 1243 1244 if (top->nr_threads_synthesize > 1) 1245 perf_set_multithreaded(); 1246 1247 init_process_thread(top); 1248 1249 if (opts->record_namespaces) 1250 top->tool.namespace_events = true; 1251 if (opts->record_cgroup) { 1252 #ifdef HAVE_FILE_HANDLE 1253 top->tool.cgroup_events = true; 1254 #else 1255 pr_err("cgroup tracking is not supported.\n"); 1256 return -1; 1257 #endif 1258 } 1259 1260 ret = perf_event__synthesize_bpf_events(top->session, perf_event__process, 1261 &top->session->machines.host, 1262 &top->record_opts); 1263 if (ret < 0) 1264 pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n"); 1265 1266 ret = perf_event__synthesize_cgroups(&top->tool, perf_event__process, 1267 &top->session->machines.host); 1268 if (ret < 0) 1269 pr_debug("Couldn't synthesize cgroup events.\n"); 1270 1271 machine__synthesize_threads(&top->session->machines.host, &opts->target, 1272 top->evlist->core.threads, true, false, 1273 top->nr_threads_synthesize); 1274 1275 if (top->nr_threads_synthesize > 1) 1276 perf_set_singlethreaded(); 1277 1278 if (perf_hpp_list.socket) { 1279 ret = perf_env__read_cpu_topology_map(&perf_env); 1280 if (ret < 0) { 1281 char errbuf[BUFSIZ]; 1282 const char *err = str_error_r(-ret, errbuf, sizeof(errbuf)); 1283 1284 ui__error("Could not read the CPU topology map: %s\n", err); 1285 return ret; 1286 } 1287 } 1288 1289 ret = perf_top__start_counters(top); 1290 if (ret) 1291 return ret; 1292 1293 top->session->evlist = top->evlist; 1294 perf_session__set_id_hdr_size(top->session); 1295 1296 /* 1297 * When perf is starting the traced process, all the events (apart from 1298 * group members) have enable_on_exec=1 set, so don't spoil it by 1299 * prematurely enabling them. 1300 * 1301 * XXX 'top' still doesn't start workloads like record, trace, but should, 1302 * so leave the check here. 1303 */ 1304 if (!target__none(&opts->target)) 1305 evlist__enable(top->evlist); 1306 1307 ret = -1; 1308 if (pthread_create(&thread_process, NULL, process_thread, top)) { 1309 ui__error("Could not create process thread.\n"); 1310 return ret; 1311 } 1312 1313 if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui : 1314 display_thread), top)) { 1315 ui__error("Could not create display thread.\n"); 1316 goto out_join_thread; 1317 } 1318 1319 if (top->realtime_prio) { 1320 struct sched_param param; 1321 1322 param.sched_priority = top->realtime_prio; 1323 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) { 1324 ui__error("Could not set realtime priority.\n"); 1325 goto out_join; 1326 } 1327 } 1328 1329 /* Wait for a minimal set of events before starting the snapshot */ 1330 evlist__poll(top->evlist, 100); 1331 1332 perf_top__mmap_read(top); 1333 1334 while (!done) { 1335 u64 hits = top->samples; 1336 1337 perf_top__mmap_read(top); 1338 1339 if (opts->overwrite || (hits == top->samples)) 1340 ret = evlist__poll(top->evlist, 100); 1341 1342 if (resize) { 1343 perf_top__resize(top); 1344 resize = 0; 1345 } 1346 } 1347 1348 ret = 0; 1349 out_join: 1350 pthread_join(thread, NULL); 1351 out_join_thread: 1352 pthread_cond_signal(&top->qe.cond); 1353 pthread_join(thread_process, NULL); 1354 return ret; 1355 } 1356 1357 static int 1358 callchain_opt(const struct option *opt, const char *arg, int unset) 1359 { 1360 symbol_conf.use_callchain = true; 1361 return record_callchain_opt(opt, arg, unset); 1362 } 1363 1364 static int 1365 parse_callchain_opt(const struct option *opt, const char *arg, int unset) 1366 { 1367 struct callchain_param *callchain = opt->value; 1368 1369 callchain->enabled = !unset; 1370 callchain->record_mode = CALLCHAIN_FP; 1371 1372 /* 1373 * --no-call-graph 1374 */ 1375 if (unset) { 1376 symbol_conf.use_callchain = false; 1377 callchain->record_mode = CALLCHAIN_NONE; 1378 return 0; 1379 } 1380 1381 return parse_callchain_top_opt(arg); 1382 } 1383 1384 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused) 1385 { 1386 if (!strcmp(var, "top.call-graph")) { 1387 var = "call-graph.record-mode"; 1388 return perf_default_config(var, value, cb); 1389 } 1390 if (!strcmp(var, "top.children")) { 1391 symbol_conf.cumulate_callchain = perf_config_bool(var, value); 1392 return 0; 1393 } 1394 1395 return 0; 1396 } 1397 1398 static int 1399 parse_percent_limit(const struct option *opt, const char *arg, 1400 int unset __maybe_unused) 1401 { 1402 struct perf_top *top = opt->value; 1403 1404 top->min_percent = strtof(arg, NULL); 1405 return 0; 1406 } 1407 1408 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP 1409 "\n\t\t\t\tDefault: fp,graph,0.5,caller,function"; 1410 1411 int cmd_top(int argc, const char **argv) 1412 { 1413 char errbuf[BUFSIZ]; 1414 struct perf_top top = { 1415 .count_filter = 5, 1416 .delay_secs = 2, 1417 .record_opts = { 1418 .mmap_pages = UINT_MAX, 1419 .user_freq = UINT_MAX, 1420 .user_interval = ULLONG_MAX, 1421 .freq = 4000, /* 4 KHz */ 1422 .target = { 1423 .uses_mmap = true, 1424 }, 1425 /* 1426 * FIXME: This will lose PERF_RECORD_MMAP and other metadata 1427 * when we pause, fix that and reenable. Probably using a 1428 * separate evlist with a dummy event, i.e. a non-overwrite 1429 * ring buffer just for metadata events, while PERF_RECORD_SAMPLE 1430 * stays in overwrite mode. -acme 1431 * */ 1432 .overwrite = 0, 1433 .sample_time = true, 1434 .sample_time_set = true, 1435 }, 1436 .max_stack = sysctl__max_stack(), 1437 .annotation_opts = annotation__default_options, 1438 .nr_threads_synthesize = UINT_MAX, 1439 }; 1440 struct record_opts *opts = &top.record_opts; 1441 struct target *target = &opts->target; 1442 const struct option options[] = { 1443 OPT_CALLBACK('e', "event", &top.evlist, "event", 1444 "event selector. use 'perf list' to list available events", 1445 parse_events_option), 1446 OPT_U64('c', "count", &opts->user_interval, "event period to sample"), 1447 OPT_STRING('p', "pid", &target->pid, "pid", 1448 "profile events on existing process id"), 1449 OPT_STRING('t', "tid", &target->tid, "tid", 1450 "profile events on existing thread id"), 1451 OPT_BOOLEAN('a', "all-cpus", &target->system_wide, 1452 "system-wide collection from all CPUs"), 1453 OPT_STRING('C', "cpu", &target->cpu_list, "cpu", 1454 "list of cpus to monitor"), 1455 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name, 1456 "file", "vmlinux pathname"), 1457 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux, 1458 "don't load vmlinux even if found"), 1459 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name, 1460 "file", "kallsyms pathname"), 1461 OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols, 1462 "hide kernel symbols"), 1463 OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages", 1464 "number of mmap data pages", evlist__parse_mmap_pages), 1465 OPT_INTEGER('r', "realtime", &top.realtime_prio, 1466 "collect data with this RT SCHED_FIFO priority"), 1467 OPT_INTEGER('d', "delay", &top.delay_secs, 1468 "number of seconds to delay between refreshes"), 1469 OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab, 1470 "dump the symbol table used for profiling"), 1471 OPT_INTEGER('f', "count-filter", &top.count_filter, 1472 "only display functions with more events than this"), 1473 OPT_BOOLEAN(0, "group", &opts->group, 1474 "put the counters into a counter group"), 1475 OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit, 1476 "child tasks do not inherit counters"), 1477 OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name", 1478 "symbol to annotate"), 1479 OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"), 1480 OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'", 1481 "profile at this frequency", 1482 record__parse_freq), 1483 OPT_INTEGER('E', "entries", &top.print_entries, 1484 "display this many functions"), 1485 OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols, 1486 "hide user symbols"), 1487 #ifdef HAVE_SLANG_SUPPORT 1488 OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"), 1489 #endif 1490 OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"), 1491 OPT_INCR('v', "verbose", &verbose, 1492 "be more verbose (show counter open errors, etc)"), 1493 OPT_STRING('s', "sort", &sort_order, "key[,key2...]", 1494 "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..." 1495 " Please refer the man page for the complete list."), 1496 OPT_STRING(0, "fields", &field_order, "key[,keys...]", 1497 "output field(s): overhead, period, sample plus all of sort keys"), 1498 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples, 1499 "Show a column with the number of samples"), 1500 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param, 1501 NULL, "enables call-graph recording and display", 1502 &callchain_opt), 1503 OPT_CALLBACK(0, "call-graph", &callchain_param, 1504 "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]", 1505 top_callchain_help, &parse_callchain_opt), 1506 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain, 1507 "Accumulate callchains of children and show total overhead as well"), 1508 OPT_INTEGER(0, "max-stack", &top.max_stack, 1509 "Set the maximum stack depth when parsing the callchain. " 1510 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)), 1511 OPT_CALLBACK(0, "ignore-callees", NULL, "regex", 1512 "ignore callees of these functions in call graphs", 1513 report_parse_ignore_callees_opt), 1514 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period, 1515 "Show a column with the sum of periods"), 1516 OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]", 1517 "only consider symbols in these dsos"), 1518 OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]", 1519 "only consider symbols in these comms"), 1520 OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]", 1521 "only consider these symbols"), 1522 OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src, 1523 "Interleave source code with assembly code (default)"), 1524 OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw, 1525 "Display raw encoding of assembly instructions (default)"), 1526 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel, 1527 "Enable kernel symbol demangling"), 1528 OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"), 1529 OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path", 1530 "objdump binary to use for disassembly and annotations"), 1531 OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style", 1532 "Specify disassembler style (e.g. -M intel for intel syntax)"), 1533 OPT_STRING(0, "prefix", &top.annotation_opts.prefix, "prefix", 1534 "Add prefix to source file path names in programs (with --prefix-strip)"), 1535 OPT_STRING(0, "prefix-strip", &top.annotation_opts.prefix_strip, "N", 1536 "Strip first N entries of source file path name in programs (with --prefix)"), 1537 OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"), 1538 OPT_CALLBACK(0, "percent-limit", &top, "percent", 1539 "Don't show entries under that percent", parse_percent_limit), 1540 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute", 1541 "How to display percentage of filtered entries", parse_filter_percentage), 1542 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str, 1543 "width[,width...]", 1544 "don't try to adjust column width, use these fixed values"), 1545 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout, 1546 "per thread proc mmap processing timeout in ms"), 1547 OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack, 1548 "branch any", "sample any taken branches", 1549 parse_branch_stack), 1550 OPT_CALLBACK('j', "branch-filter", &opts->branch_stack, 1551 "branch filter mask", "branch stack filter modes", 1552 parse_branch_stack), 1553 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace, 1554 "Show raw trace event output (do not use print fmt or plugins)"), 1555 OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy, 1556 "Show entries in a hierarchy"), 1557 OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite, 1558 "Use a backward ring buffer, default: no"), 1559 OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"), 1560 OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize, 1561 "number of thread to run event synthesize"), 1562 OPT_CALLBACK('G', "cgroup", &top.evlist, "name", 1563 "monitor event in cgroup name only", parse_cgroups), 1564 OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces, 1565 "Record namespaces events"), 1566 OPT_BOOLEAN(0, "all-cgroups", &opts->record_cgroup, 1567 "Record cgroup events"), 1568 OPT_INTEGER(0, "group-sort-idx", &symbol_conf.group_sort_idx, 1569 "Sort the output by the event at the index n in group. " 1570 "If n is invalid, sort by the first event. " 1571 "WARNING: should be used on grouped events."), 1572 OPT_BOOLEAN(0, "stitch-lbr", &top.stitch_lbr, 1573 "Enable LBR callgraph stitching approach"), 1574 #ifdef HAVE_LIBPFM 1575 OPT_CALLBACK(0, "pfm-events", &top.evlist, "event", 1576 "libpfm4 event selector. use 'perf list' to list available events", 1577 parse_libpfm_events_option), 1578 #endif 1579 OPTS_EVSWITCH(&top.evswitch), 1580 OPT_END() 1581 }; 1582 const char * const top_usage[] = { 1583 "perf top [<options>]", 1584 NULL 1585 }; 1586 int status = hists__init(); 1587 1588 if (status < 0) 1589 return status; 1590 1591 top.annotation_opts.min_pcnt = 5; 1592 top.annotation_opts.context = 4; 1593 1594 top.evlist = evlist__new(); 1595 if (top.evlist == NULL) 1596 return -ENOMEM; 1597 1598 status = perf_config(perf_top_config, &top); 1599 if (status) 1600 return status; 1601 /* 1602 * Since the per arch annotation init routine may need the cpuid, read 1603 * it here, since we are not getting this from the perf.data header. 1604 */ 1605 status = perf_env__read_cpuid(&perf_env); 1606 if (status) { 1607 /* 1608 * Some arches do not provide a get_cpuid(), so just use pr_debug, otherwise 1609 * warn the user explicitly. 1610 */ 1611 eprintf(status == ENOSYS ? 1 : 0, verbose, 1612 "Couldn't read the cpuid for this machine: %s\n", 1613 str_error_r(errno, errbuf, sizeof(errbuf))); 1614 } 1615 top.evlist->env = &perf_env; 1616 1617 argc = parse_options(argc, argv, options, top_usage, 0); 1618 if (argc) 1619 usage_with_options(top_usage, options); 1620 1621 status = symbol__validate_sym_arguments(); 1622 if (status) 1623 goto out_delete_evlist; 1624 1625 if (annotate_check_args(&top.annotation_opts) < 0) 1626 goto out_delete_evlist; 1627 1628 if (!top.evlist->core.nr_entries && 1629 evlist__add_default(top.evlist) < 0) { 1630 pr_err("Not enough memory for event selector list\n"); 1631 goto out_delete_evlist; 1632 } 1633 1634 status = evswitch__init(&top.evswitch, top.evlist, stderr); 1635 if (status) 1636 goto out_delete_evlist; 1637 1638 if (symbol_conf.report_hierarchy) { 1639 /* disable incompatible options */ 1640 symbol_conf.event_group = false; 1641 symbol_conf.cumulate_callchain = false; 1642 1643 if (field_order) { 1644 pr_err("Error: --hierarchy and --fields options cannot be used together\n"); 1645 parse_options_usage(top_usage, options, "fields", 0); 1646 parse_options_usage(NULL, options, "hierarchy", 0); 1647 goto out_delete_evlist; 1648 } 1649 } 1650 1651 if (top.stitch_lbr && !(callchain_param.record_mode == CALLCHAIN_LBR)) { 1652 pr_err("Error: --stitch-lbr must be used with --call-graph lbr\n"); 1653 goto out_delete_evlist; 1654 } 1655 1656 if (nr_cgroups > 0 && opts->record_cgroup) { 1657 pr_err("--cgroup and --all-cgroups cannot be used together\n"); 1658 goto out_delete_evlist; 1659 } 1660 1661 if (opts->branch_stack && callchain_param.enabled) 1662 symbol_conf.show_branchflag_count = true; 1663 1664 sort__mode = SORT_MODE__TOP; 1665 /* display thread wants entries to be collapsed in a different tree */ 1666 perf_hpp_list.need_collapse = 1; 1667 1668 if (top.use_stdio) 1669 use_browser = 0; 1670 #ifdef HAVE_SLANG_SUPPORT 1671 else if (top.use_tui) 1672 use_browser = 1; 1673 #endif 1674 1675 setup_browser(false); 1676 1677 if (setup_sorting(top.evlist) < 0) { 1678 if (sort_order) 1679 parse_options_usage(top_usage, options, "s", 1); 1680 if (field_order) 1681 parse_options_usage(sort_order ? NULL : top_usage, 1682 options, "fields", 0); 1683 goto out_delete_evlist; 1684 } 1685 1686 status = target__validate(target); 1687 if (status) { 1688 target__strerror(target, status, errbuf, BUFSIZ); 1689 ui__warning("%s\n", errbuf); 1690 } 1691 1692 status = target__parse_uid(target); 1693 if (status) { 1694 int saved_errno = errno; 1695 1696 target__strerror(target, status, errbuf, BUFSIZ); 1697 ui__error("%s\n", errbuf); 1698 1699 status = -saved_errno; 1700 goto out_delete_evlist; 1701 } 1702 1703 if (target__none(target)) 1704 target->system_wide = true; 1705 1706 if (evlist__create_maps(top.evlist, target) < 0) { 1707 ui__error("Couldn't create thread/CPU maps: %s\n", 1708 errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf))); 1709 goto out_delete_evlist; 1710 } 1711 1712 if (top.delay_secs < 1) 1713 top.delay_secs = 1; 1714 1715 if (record_opts__config(opts)) { 1716 status = -EINVAL; 1717 goto out_delete_evlist; 1718 } 1719 1720 top.sym_evsel = evlist__first(top.evlist); 1721 1722 if (!callchain_param.enabled) { 1723 symbol_conf.cumulate_callchain = false; 1724 perf_hpp__cancel_cumulate(); 1725 } 1726 1727 if (symbol_conf.cumulate_callchain && !callchain_param.order_set) 1728 callchain_param.order = ORDER_CALLER; 1729 1730 status = symbol__annotation_init(); 1731 if (status < 0) 1732 goto out_delete_evlist; 1733 1734 annotation_config__init(&top.annotation_opts); 1735 1736 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL); 1737 status = symbol__init(NULL); 1738 if (status < 0) 1739 goto out_delete_evlist; 1740 1741 sort__setup_elide(stdout); 1742 1743 get_term_dimensions(&top.winsize); 1744 if (top.print_entries == 0) { 1745 perf_top__update_print_entries(&top); 1746 signal(SIGWINCH, winch_sig); 1747 } 1748 1749 top.session = perf_session__new(NULL, NULL); 1750 if (IS_ERR(top.session)) { 1751 status = PTR_ERR(top.session); 1752 goto out_delete_evlist; 1753 } 1754 1755 #ifdef HAVE_LIBBPF_SUPPORT 1756 if (!top.record_opts.no_bpf_event) { 1757 top.sb_evlist = evlist__new(); 1758 1759 if (top.sb_evlist == NULL) { 1760 pr_err("Couldn't create side band evlist.\n."); 1761 goto out_delete_evlist; 1762 } 1763 1764 if (evlist__add_bpf_sb_event(top.sb_evlist, &perf_env)) { 1765 pr_err("Couldn't ask for PERF_RECORD_BPF_EVENT side band events.\n."); 1766 goto out_delete_evlist; 1767 } 1768 } 1769 #endif 1770 1771 if (evlist__start_sb_thread(top.sb_evlist, target)) { 1772 pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n"); 1773 opts->no_bpf_event = true; 1774 } 1775 1776 status = __cmd_top(&top); 1777 1778 if (!opts->no_bpf_event) 1779 evlist__stop_sb_thread(top.sb_evlist); 1780 1781 out_delete_evlist: 1782 evlist__delete(top.evlist); 1783 perf_session__delete(top.session); 1784 1785 return status; 1786 } 1787