tools/perf/builtin-top.c

   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/config.h"
  26 #include "util/color.h"
  27 #include "util/dso.h"
  28 #include "util/evlist.h"
  29 #include "util/evsel.h"
  30 #include "util/evsel_config.h"
  31 #include "util/event.h"
  32 #include "util/machine.h"
  33 #include "util/map.h"
  34 #include "util/mmap.h"
  35 #include "util/session.h"
  36 #include "util/symbol.h"
  37 #include "util/synthetic-events.h"
  38 #include "util/top.h"
  39 #include "util/util.h"
  40 #include <linux/rbtree.h>
  41 #include <subcmd/parse-options.h>
  42 #include "util/parse-events.h"
  43 #include "util/callchain.h"
  44 #include "util/cpumap.h"
  45 #include "util/sort.h"
  46 #include "util/string2.h"
  47 #include "util/term.h"
  48 #include "util/intlist.h"
  49 #include "util/parse-branch-options.h"
  50 #include "arch/common.h"
  51 #include "ui/ui.h"
  52
  53 #include "util/debug.h"
  54 #include "util/ordered-events.h"
  55
  56 #include <assert.h>
  57 #include <elf.h>
  58 #include <fcntl.h>
  59
  60 #include <stdio.h>
  61 #include <termios.h>
  62 #include <unistd.h>
  63 #include <inttypes.h>
  64
  65 #include <errno.h>
  66 #include <time.h>
  67 #include <sched.h>
  68 #include <signal.h>
  69
  70 #include <sys/syscall.h>
  71 #include <sys/ioctl.h>
  72 #include <poll.h>
  73 #include <sys/prctl.h>
  74 #include <sys/wait.h>
  75 #include <sys/uio.h>
  76 #include <sys/utsname.h>
  77 #include <sys/mman.h>
  78
  79 #include <linux/stringify.h>
  80 #include <linux/time64.h>
  81 #include <linux/types.h>
  82 #include <linux/err.h>
  83
  84 #include <linux/ctype.h>
  85
  86 static volatile int done;
  87 static volatile int resize;
  88
  89 #define HEADER_LINE_NR  5
  90
  91 static void perf_top__update_print_entries(struct perf_top *top)
  92 {
  93         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
  94 }
  95
  96 static void winch_sig(int sig __maybe_unused)
  97 {
  98         resize = 1;
  99 }
 100
 101 static void perf_top__resize(struct perf_top *top)
 102 {
 103         get_term_dimensions(&top->winsize);
 104         perf_top__update_print_entries(top);
 105 }
 106
 107 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
 108 {
 109         struct evsel *evsel;
 110         struct symbol *sym;
 111         struct annotation *notes;
 112         struct map *map;
 113         int err = -1;
 114
 115         if (!he || !he->ms.sym)
 116                 return -1;
 117
 118         evsel = hists_to_evsel(he->hists);
 119
 120         sym = he->ms.sym;
 121         map = he->ms.map;
 122
 123         /*
 124          * We can't annotate with just /proc/kallsyms
 125          */
 126         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 127             !dso__is_kcore(map->dso)) {
 128                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 129                        "path\n", sym->name);
 130                 sleep(1);
 131                 return -1;
 132         }
 133
 134         notes = symbol__annotation(sym);
 135         pthread_mutex_lock(&notes->lock);
 136
 137         if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
 138                 pthread_mutex_unlock(&notes->lock);
 139                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 140                        sym->name);
 141                 sleep(1);
 142                 return err;
 143         }
 144
 145         err = symbol__annotate(sym, map, evsel, 0, &top->annotation_opts, NULL);
 146         if (err == 0) {
 147                 top->sym_filter_entry = he;
 148         } else {
 149                 char msg[BUFSIZ];
 150                 symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
 151                 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
 152         }
 153
 154         pthread_mutex_unlock(&notes->lock);
 155         return err;
 156 }
 157
 158 static void __zero_source_counters(struct hist_entry *he)
 159 {
 160         struct symbol *sym = he->ms.sym;
 161         symbol__annotate_zero_histograms(sym);
 162 }
 163
 164 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 165 {
 166         struct utsname uts;
 167         int err = uname(&uts);
 168
 169         ui__warning("Out of bounds address found:\n\n"
 170                     "Addr:   %" PRIx64 "\n"
 171                     "DSO:    %s %c\n"
 172                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
 173                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 174                     "Arch:   %s\n"
 175                     "Kernel: %s\n"
 176                     "Tools:  %s\n\n"
 177                     "Not all samples will be on the annotation output.\n\n"
 178                     "Please report to linux-kernel@vger.kernel.org\n",
 179                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
 180                     map->start, map->end, sym->start, sym->end,
 181                     sym->binding == STB_GLOBAL ? 'g' :
 182                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 183                     err ? "[unknown]" : uts.machine,
 184                     err ? "[unknown]" : uts.release, perf_version_string);
 185         if (use_browser <= 0)
 186                 sleep(5);
 187
 188         map->erange_warned = true;
 189 }
 190
 191 static void perf_top__record_precise_ip(struct perf_top *top,
 192                                         struct hist_entry *he,
 193                                         struct perf_sample *sample,
 194                                         struct evsel *evsel, u64 ip)
 195 {
 196         struct annotation *notes;
 197         struct symbol *sym = he->ms.sym;
 198         int err = 0;
 199
 200         if (sym == NULL || (use_browser == 0 &&
 201                             (top->sym_filter_entry == NULL ||
 202                              top->sym_filter_entry->ms.sym != sym)))
 203                 return;
 204
 205         notes = symbol__annotation(sym);
 206
 207         if (pthread_mutex_trylock(&notes->lock))
 208                 return;
 209
 210         err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
 211
 212         pthread_mutex_unlock(&notes->lock);
 213
 214         if (unlikely(err)) {
 215                 /*
 216                  * This function is now called with he->hists->lock held.
 217                  * Release it before going to sleep.
 218                  */
 219                 pthread_mutex_unlock(&he->hists->lock);
 220
 221                 if (err == -ERANGE && !he->ms.map->erange_warned)
 222                         ui__warn_map_erange(he->ms.map, sym, ip);
 223                 else if (err == -ENOMEM) {
 224                         pr_err("Not enough memory for annotating '%s' symbol!\n",
 225                                sym->name);
 226                         sleep(1);
 227                 }
 228
 229                 pthread_mutex_lock(&he->hists->lock);
 230         }
 231 }
 232
 233 static void perf_top__show_details(struct perf_top *top)
 234 {
 235         struct hist_entry *he = top->sym_filter_entry;
 236         struct evsel *evsel;
 237         struct annotation *notes;
 238         struct symbol *symbol;
 239         int more;
 240
 241         if (!he)
 242                 return;
 243
 244         evsel = hists_to_evsel(he->hists);
 245
 246         symbol = he->ms.sym;
 247         notes = symbol__annotation(symbol);
 248
 249         pthread_mutex_lock(&notes->lock);
 250
 251         symbol__calc_percent(symbol, evsel);
 252
 253         if (notes->src == NULL)
 254                 goto out_unlock;
 255
 256         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
 257         printf("  Events  Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
 258
 259         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel, &top->annotation_opts);
 260
 261         if (top->evlist->enabled) {
 262                 if (top->zero)
 263                         symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 264                 else
 265                         symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 266         }
 267         if (more != 0)
 268                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 269 out_unlock:
 270         pthread_mutex_unlock(&notes->lock);
 271 }
 272
 273 static void perf_top__resort_hists(struct perf_top *t)
 274 {
 275         struct evlist *evlist = t->evlist;
 276         struct evsel *pos;
 277
 278         evlist__for_each_entry(evlist, pos) {
 279                 struct hists *hists = evsel__hists(pos);
 280
 281                 /*
 282                  * unlink existing entries so that they can be linked
 283                  * in a correct order in hists__match() below.
 284                  */
 285                 hists__unlink(hists);
 286
 287                 if (evlist->enabled) {
 288                         if (t->zero) {
 289                                 hists__delete_entries(hists);
 290                         } else {
 291                                 hists__decay_entries(hists, t->hide_user_symbols,
 292                                                      t->hide_kernel_symbols);
 293                         }
 294                 }
 295
 296                 hists__collapse_resort(hists, NULL);
 297
 298                 /* Non-group events are considered as leader */
 299                 if (symbol_conf.event_group &&
 300                     !perf_evsel__is_group_leader(pos)) {
 301                         struct hists *leader_hists = evsel__hists(pos->leader);
 302
 303                         hists__match(leader_hists, hists);
 304                         hists__link(leader_hists, hists);
 305                 }
 306         }
 307
 308         evlist__for_each_entry(evlist, pos) {
 309                 perf_evsel__output_resort(pos, NULL);
 310         }
 311 }
 312
 313 static void perf_top__print_sym_table(struct perf_top *top)
 314 {
 315         char bf[160];
 316         int printed = 0;
 317         const int win_width = top->winsize.ws_col - 1;
 318         struct evsel *evsel = top->sym_evsel;
 319         struct hists *hists = evsel__hists(evsel);
 320
 321         puts(CONSOLE_CLEAR);
 322
 323         perf_top__header_snprintf(top, bf, sizeof(bf));
 324         printf("%s\n", bf);
 325
 326         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 327
 328         if (!top->record_opts.overwrite &&
 329             (hists->stats.nr_lost_warned !=
 330             hists->stats.nr_events[PERF_RECORD_LOST])) {
 331                 hists->stats.nr_lost_warned =
 332                               hists->stats.nr_events[PERF_RECORD_LOST];
 333                 color_fprintf(stdout, PERF_COLOR_RED,
 334                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 335                               hists->stats.nr_lost_warned);
 336                 ++printed;
 337         }
 338
 339         if (top->sym_filter_entry) {
 340                 perf_top__show_details(top);
 341                 return;
 342         }
 343
 344         perf_top__resort_hists(top);
 345
 346         hists__output_recalc_col_len(hists, top->print_entries - printed);
 347         putchar('\n');
 348         hists__fprintf(hists, false, top->print_entries - printed, win_width,
 349                        top->min_percent, stdout, !symbol_conf.use_callchain);
 350 }
 351
 352 static void prompt_integer(int *target, const char *msg)
 353 {
 354         char *buf = malloc(0), *p;
 355         size_t dummy = 0;
 356         int tmp;
 357
 358         fprintf(stdout, "\n%s: ", msg);
 359         if (getline(&buf, &dummy, stdin) < 0)
 360                 return;
 361
 362         p = strchr(buf, '\n');
 363         if (p)
 364                 *p = 0;
 365
 366         p = buf;
 367         while(*p) {
 368                 if (!isdigit(*p))
 369                         goto out_free;
 370                 p++;
 371         }
 372         tmp = strtoul(buf, NULL, 10);
 373         *target = tmp;
 374 out_free:
 375         free(buf);
 376 }
 377
 378 static void prompt_percent(int *target, const char *msg)
 379 {
 380         int tmp = 0;
 381
 382         prompt_integer(&tmp, msg);
 383         if (tmp >= 0 && tmp <= 100)
 384                 *target = tmp;
 385 }
 386
 387 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 388 {
 389         char *buf = malloc(0), *p;
 390         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 391         struct hists *hists = evsel__hists(top->sym_evsel);
 392         struct rb_node *next;
 393         size_t dummy = 0;
 394
 395         /* zero counters of active symbol */
 396         if (syme) {
 397                 __zero_source_counters(syme);
 398                 top->sym_filter_entry = NULL;
 399         }
 400
 401         fprintf(stdout, "\n%s: ", msg);
 402         if (getline(&buf, &dummy, stdin) < 0)
 403                 goto out_free;
 404
 405         p = strchr(buf, '\n');
 406         if (p)
 407                 *p = 0;
 408
 409         next = rb_first_cached(&hists->entries);
 410         while (next) {
 411                 n = rb_entry(next, struct hist_entry, rb_node);
 412                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 413                         found = n;
 414                         break;
 415                 }
 416                 next = rb_next(&n->rb_node);
 417         }
 418
 419         if (!found) {
 420                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 421                 sleep(1);
 422         } else
 423                 perf_top__parse_source(top, found);
 424
 425 out_free:
 426         free(buf);
 427 }
 428
 429 static void perf_top__print_mapped_keys(struct perf_top *top)
 430 {
 431         char *name = NULL;
 432
 433         if (top->sym_filter_entry) {
 434                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 435                 name = sym->name;
 436         }
 437
 438         fprintf(stdout, "\nMapped keys:\n");
 439         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 440         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 441
 442         if (top->evlist->core.nr_entries > 1)
 443                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 444
 445         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 446
 447         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
 448         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 449         fprintf(stdout, "\t[S]     stop annotation.\n");
 450
 451         fprintf(stdout,
 452                 "\t[K]     hide kernel symbols.             \t(%s)\n",
 453                 top->hide_kernel_symbols ? "yes" : "no");
 454         fprintf(stdout,
 455                 "\t[U]     hide user symbols.               \t(%s)\n",
 456                 top->hide_user_symbols ? "yes" : "no");
 457         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 458         fprintf(stdout, "\t[qQ]    quit.\n");
 459 }
 460
 461 static int perf_top__key_mapped(struct perf_top *top, int c)
 462 {
 463         switch (c) {
 464                 case 'd':
 465                 case 'e':
 466                 case 'f':
 467                 case 'z':
 468                 case 'q':
 469                 case 'Q':
 470                 case 'K':
 471                 case 'U':
 472                 case 'F':
 473                 case 's':
 474                 case 'S':
 475                         return 1;
 476                 case 'E':
 477                         return top->evlist->core.nr_entries > 1 ? 1 : 0;
 478                 default:
 479                         break;
 480         }
 481
 482         return 0;
 483 }
 484
 485 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 486 {
 487         bool ret = true;
 488
 489         if (!perf_top__key_mapped(top, c)) {
 490                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 491                 struct termios save;
 492
 493                 perf_top__print_mapped_keys(top);
 494                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 495                 fflush(stdout);
 496
 497                 set_term_quiet_input(&save);
 498
 499                 poll(&stdin_poll, 1, -1);
 500                 c = getc(stdin);
 501
 502                 tcsetattr(0, TCSAFLUSH, &save);
 503                 if (!perf_top__key_mapped(top, c))
 504                         return ret;
 505         }
 506
 507         switch (c) {
 508                 case 'd':
 509                         prompt_integer(&top->delay_secs, "Enter display delay");
 510                         if (top->delay_secs < 1)
 511                                 top->delay_secs = 1;
 512                         break;
 513                 case 'e':
 514                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 515                         if (top->print_entries == 0) {
 516                                 perf_top__resize(top);
 517                                 signal(SIGWINCH, winch_sig);
 518                         } else {
 519                                 signal(SIGWINCH, SIG_DFL);
 520                         }
 521                         break;
 522                 case 'E':
 523                         if (top->evlist->core.nr_entries > 1) {
 524                                 /* Select 0 as the default event: */
 525                                 int counter = 0;
 526
 527                                 fprintf(stderr, "\nAvailable events:");
 528
 529                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
 530                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
 531
 532                                 prompt_integer(&counter, "Enter details event counter");
 533
 534                                 if (counter >= top->evlist->core.nr_entries) {
 535                                         top->sym_evsel = evlist__first(top->evlist);
 536                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
 537                                         sleep(1);
 538                                         break;
 539                                 }
 540                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
 541                                         if (top->sym_evsel->idx == counter)
 542                                                 break;
 543                         } else
 544                                 top->sym_evsel = evlist__first(top->evlist);
 545                         break;
 546                 case 'f':
 547                         prompt_integer(&top->count_filter, "Enter display event count filter");
 548                         break;
 549                 case 'F':
 550                         prompt_percent(&top->annotation_opts.min_pcnt,
 551                                        "Enter details display event filter (percent)");
 552                         break;
 553                 case 'K':
 554                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 555                         break;
 556                 case 'q':
 557                 case 'Q':
 558                         printf("exiting.\n");
 559                         if (top->dump_symtab)
 560                                 perf_session__fprintf_dsos(top->session, stderr);
 561                         ret = false;
 562                         break;
 563                 case 's':
 564                         perf_top__prompt_symbol(top, "Enter details symbol");
 565                         break;
 566                 case 'S':
 567                         if (!top->sym_filter_entry)
 568                                 break;
 569                         else {
 570                                 struct hist_entry *syme = top->sym_filter_entry;
 571
 572                                 top->sym_filter_entry = NULL;
 573                                 __zero_source_counters(syme);
 574                         }
 575                         break;
 576                 case 'U':
 577                         top->hide_user_symbols = !top->hide_user_symbols;
 578                         break;
 579                 case 'z':
 580                         top->zero = !top->zero;
 581                         break;
 582                 default:
 583                         break;
 584         }
 585
 586         return ret;
 587 }
 588
 589 static void perf_top__sort_new_samples(void *arg)
 590 {
 591         struct perf_top *t = arg;
 592
 593         if (t->evlist->selected != NULL)
 594                 t->sym_evsel = t->evlist->selected;
 595
 596         perf_top__resort_hists(t);
 597
 598         if (t->lost || t->drop)
 599                 pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
 600 }
 601
 602 static void stop_top(void)
 603 {
 604         session_done = 1;
 605         done = 1;
 606 }
 607
 608 static void *display_thread_tui(void *arg)
 609 {
 610         struct evsel *pos;
 611         struct perf_top *top = arg;
 612         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 613         struct hist_browser_timer hbt = {
 614                 .timer          = perf_top__sort_new_samples,
 615                 .arg            = top,
 616                 .refresh        = top->delay_secs,
 617         };
 618
 619         /* In order to read symbols from other namespaces perf to  needs to call
 620          * setns(2).  This isn't permitted if the struct_fs has multiple users.
 621          * unshare(2) the fs so that we may continue to setns into namespaces
 622          * that we're observing.
 623          */
 624         unshare(CLONE_FS);
 625
 626         prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
 627
 628         perf_top__sort_new_samples(top);
 629
 630         /*
 631          * Initialize the uid_filter_str, in the future the TUI will allow
 632          * Zooming in/out UIDs. For now just use whatever the user passed
 633          * via --uid.
 634          */
 635         evlist__for_each_entry(top->evlist, pos) {
 636                 struct hists *hists = evsel__hists(pos);
 637                 hists->uid_filter_str = top->record_opts.target.uid_str;
 638         }
 639
 640         perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
 641                                       top->min_percent,
 642                                       &top->session->header.env,
 643                                       !top->record_opts.overwrite,
 644                                       &top->annotation_opts);
 645
 646         stop_top();
 647         return NULL;
 648 }
 649
 650 static void display_sig(int sig __maybe_unused)
 651 {
 652         stop_top();
 653 }
 654
 655 static void display_setup_sig(void)
 656 {
 657         signal(SIGSEGV, sighandler_dump_stack);
 658         signal(SIGFPE, sighandler_dump_stack);
 659         signal(SIGINT,  display_sig);
 660         signal(SIGQUIT, display_sig);
 661         signal(SIGTERM, display_sig);
 662 }
 663
 664 static void *display_thread(void *arg)
 665 {
 666         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 667         struct termios save;
 668         struct perf_top *top = arg;
 669         int delay_msecs, c;
 670
 671         /* In order to read symbols from other namespaces perf to  needs to call
 672          * setns(2).  This isn't permitted if the struct_fs has multiple users.
 673          * unshare(2) the fs so that we may continue to setns into namespaces
 674          * that we're observing.
 675          */
 676         unshare(CLONE_FS);
 677
 678         prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
 679
 680         display_setup_sig();
 681         pthread__unblock_sigwinch();
 682 repeat:
 683         delay_msecs = top->delay_secs * MSEC_PER_SEC;
 684         set_term_quiet_input(&save);
 685         /* trash return*/
 686         clearerr(stdin);
 687         if (poll(&stdin_poll, 1, 0) > 0)
 688                 getc(stdin);
 689
 690         while (!done) {
 691                 perf_top__print_sym_table(top);
 692                 /*
 693                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 694                  * refresh screen in both cases.
 695                  */
 696                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 697                 case 0:
 698                         continue;
 699                 case -1:
 700                         if (errno == EINTR)
 701                                 continue;
 702                         __fallthrough;
 703                 default:
 704                         c = getc(stdin);
 705                         tcsetattr(0, TCSAFLUSH, &save);
 706
 707                         if (perf_top__handle_keypress(top, c))
 708                                 goto repeat;
 709                         stop_top();
 710                 }
 711         }
 712
 713         tcsetattr(0, TCSAFLUSH, &save);
 714         return NULL;
 715 }
 716
 717 static int hist_iter__top_callback(struct hist_entry_iter *iter,
 718                                    struct addr_location *al, bool single,
 719                                    void *arg)
 720 {
 721         struct perf_top *top = arg;
 722         struct hist_entry *he = iter->he;
 723         struct evsel *evsel = iter->evsel;
 724
 725         if (perf_hpp_list.sym && single)
 726                 perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
 727
 728         hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
 729                      !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
 730         return 0;
 731 }
 732
 733 static void perf_event__process_sample(struct perf_tool *tool,
 734                                        const union perf_event *event,
 735                                        struct evsel *evsel,
 736                                        struct perf_sample *sample,
 737                                        struct machine *machine)
 738 {
 739         struct perf_top *top = container_of(tool, struct perf_top, tool);
 740         struct addr_location al;
 741         int err;
 742
 743         if (!machine && perf_guest) {
 744                 static struct intlist *seen;
 745
 746                 if (!seen)
 747                         seen = intlist__new(NULL);
 748
 749                 if (!intlist__has_entry(seen, sample->pid)) {
 750                         pr_err("Can't find guest [%d]'s kernel information\n",
 751                                 sample->pid);
 752                         intlist__add(seen, sample->pid);
 753                 }
 754                 return;
 755         }
 756
 757         if (!machine) {
 758                 pr_err("%u unprocessable samples recorded.\r",
 759                        top->session->evlist->stats.nr_unprocessable_samples++);
 760                 return;
 761         }
 762
 763         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 764                 top->exact_samples++;
 765
 766         if (machine__resolve(machine, &al, sample) < 0)
 767                 return;
 768
 769         if (!machine->kptr_restrict_warned &&
 770             symbol_conf.kptr_restrict &&
 771             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 772                 if (!perf_evlist__exclude_kernel(top->session->evlist)) {
 773                         ui__warning(
 774 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 775 "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
 776 "Kernel%s samples will not be resolved.\n",
 777                           al.map && map__has_symbols(al.map) ?
 778                           " modules" : "");
 779                         if (use_browser <= 0)
 780                                 sleep(5);
 781                 }
 782                 machine->kptr_restrict_warned = true;
 783         }
 784
 785         if (al.sym == NULL && al.map != NULL) {
 786                 const char *msg = "Kernel samples will not be resolved.\n";
 787                 /*
 788                  * As we do lazy loading of symtabs we only will know if the
 789                  * specified vmlinux file is invalid when we actually have a
 790                  * hit in kernel space and then try to load it. So if we get
 791                  * here and there are _no_ symbols in the DSO backing the
 792                  * kernel map, bail out.
 793                  *
 794                  * We may never get here, for instance, if we use -K/
 795                  * --hide-kernel-symbols, even if the user specifies an
 796                  * invalid --vmlinux ;-)
 797                  */
 798                 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
 799                     __map__is_kernel(al.map) && map__has_symbols(al.map)) {
 800                         if (symbol_conf.vmlinux_name) {
 801                                 char serr[256];
 802                                 dso__strerror_load(al.map->dso, serr, sizeof(serr));
 803                                 ui__warning("The %s file can't be used: %s\n%s",
 804                                             symbol_conf.vmlinux_name, serr, msg);
 805                         } else {
 806                                 ui__warning("A vmlinux file was not found.\n%s",
 807                                             msg);
 808                         }
 809
 810                         if (use_browser <= 0)
 811                                 sleep(5);
 812                         top->vmlinux_warned = true;
 813                 }
 814         }
 815
 816         if (al.sym == NULL || !al.sym->idle) {
 817                 struct hists *hists = evsel__hists(evsel);
 818                 struct hist_entry_iter iter = {
 819                         .evsel          = evsel,
 820                         .sample         = sample,
 821                         .add_entry_cb   = hist_iter__top_callback,
 822                 };
 823
 824                 if (symbol_conf.cumulate_callchain)
 825                         iter.ops = &hist_iter_cumulative;
 826                 else
 827                         iter.ops = &hist_iter_normal;
 828
 829                 pthread_mutex_lock(&hists->lock);
 830
 831                 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
 832                 if (err < 0)
 833                         pr_err("Problem incrementing symbol period, skipping event\n");
 834
 835                 pthread_mutex_unlock(&hists->lock);
 836         }
 837
 838         addr_location__put(&al);
 839 }
 840
 841 static void
 842 perf_top__process_lost(struct perf_top *top, union perf_event *event,
 843                        struct evsel *evsel)
 844 {
 845         struct hists *hists = evsel__hists(evsel);
 846
 847         top->lost += event->lost.lost;
 848         top->lost_total += event->lost.lost;
 849         hists->stats.total_lost += event->lost.lost;
 850 }
 851
 852 static void
 853 perf_top__process_lost_samples(struct perf_top *top,
 854                                union perf_event *event,
 855                                struct evsel *evsel)
 856 {
 857         struct hists *hists = evsel__hists(evsel);
 858
 859         top->lost += event->lost_samples.lost;
 860         top->lost_total += event->lost_samples.lost;
 861         hists->stats.total_lost_samples += event->lost_samples.lost;
 862 }
 863
 864 static u64 last_timestamp;
 865
 866 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 867 {
 868         struct record_opts *opts = &top->record_opts;
 869         struct evlist *evlist = top->evlist;
 870         struct mmap *md;
 871         union perf_event *event;
 872
 873         md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
 874         if (perf_mmap__read_init(md) < 0)
 875                 return;
 876
 877         while ((event = perf_mmap__read_event(md)) != NULL) {
 878                 int ret;
 879
 880                 ret = perf_evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
 881                 if (ret && ret != -1)
 882                         break;
 883
 884                 ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0);
 885                 if (ret)
 886                         break;
 887
 888                 perf_mmap__consume(md);
 889
 890                 if (top->qe.rotate) {
 891                         pthread_mutex_lock(&top->qe.mutex);
 892                         top->qe.rotate = false;
 893                         pthread_cond_signal(&top->qe.cond);
 894                         pthread_mutex_unlock(&top->qe.mutex);
 895                 }
 896         }
 897
 898         perf_mmap__read_done(md);
 899 }
 900
 901 static void perf_top__mmap_read(struct perf_top *top)
 902 {
 903         bool overwrite = top->record_opts.overwrite;
 904         struct evlist *evlist = top->evlist;
 905         int i;
 906
 907         if (overwrite)
 908                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
 909
 910         for (i = 0; i < top->evlist->core.nr_mmaps; i++)
 911                 perf_top__mmap_read_idx(top, i);
 912
 913         if (overwrite) {
 914                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
 915                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 916         }
 917 }
 918
 919 /*
 920  * Check per-event overwrite term.
 921  * perf top should support consistent term for all events.
 922  * - All events don't have per-event term
 923  *   E.g. "cpu/cpu-cycles/,cpu/instructions/"
 924  *   Nothing change, return 0.
 925  * - All events have same per-event term
 926  *   E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
 927  *   Using the per-event setting to replace the opts->overwrite if
 928  *   they are different, then return 0.
 929  * - Events have different per-event term
 930  *   E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
 931  *   Return -1
 932  * - Some of the event set per-event term, but some not.
 933  *   E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
 934  *   Return -1
 935  */
 936 static int perf_top__overwrite_check(struct perf_top *top)
 937 {
 938         struct record_opts *opts = &top->record_opts;
 939         struct evlist *evlist = top->evlist;
 940         struct perf_evsel_config_term *term;
 941         struct list_head *config_terms;
 942         struct evsel *evsel;
 943         int set, overwrite = -1;
 944
 945         evlist__for_each_entry(evlist, evsel) {
 946                 set = -1;
 947                 config_terms = &evsel->config_terms;
 948                 list_for_each_entry(term, config_terms, list) {
 949                         if (term->type == PERF_EVSEL__CONFIG_TERM_OVERWRITE)
 950                                 set = term->val.overwrite ? 1 : 0;
 951                 }
 952
 953                 /* no term for current and previous event (likely) */
 954                 if ((overwrite < 0) && (set < 0))
 955                         continue;
 956
 957                 /* has term for both current and previous event, compare */
 958                 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
 959                         return -1;
 960
 961                 /* no term for current event but has term for previous one */
 962                 if ((overwrite >= 0) && (set < 0))
 963                         return -1;
 964
 965                 /* has term for current event */
 966                 if ((overwrite < 0) && (set >= 0)) {
 967                         /* if it's first event, set overwrite */
 968                         if (evsel == evlist__first(evlist))
 969                                 overwrite = set;
 970                         else
 971                                 return -1;
 972                 }
 973         }
 974
 975         if ((overwrite >= 0) && (opts->overwrite != overwrite))
 976                 opts->overwrite = overwrite;
 977
 978         return 0;
 979 }
 980
 981 static int perf_top_overwrite_fallback(struct perf_top *top,
 982                                        struct evsel *evsel)
 983 {
 984         struct record_opts *opts = &top->record_opts;
 985         struct evlist *evlist = top->evlist;
 986         struct evsel *counter;
 987
 988         if (!opts->overwrite)
 989                 return 0;
 990
 991         /* only fall back when first event fails */
 992         if (evsel != evlist__first(evlist))
 993                 return 0;
 994
 995         evlist__for_each_entry(evlist, counter)
 996                 counter->core.attr.write_backward = false;
 997         opts->overwrite = false;
 998         pr_debug2("fall back to non-overwrite mode\n");
 999         return 1;
1000 }
1001
1002 static int perf_top__start_counters(struct perf_top *top)
1003 {
1004         char msg[BUFSIZ];
1005         struct evsel *counter;
1006         struct evlist *evlist = top->evlist;
1007         struct record_opts *opts = &top->record_opts;
1008
1009         if (perf_top__overwrite_check(top)) {
1010                 ui__error("perf top only support consistent per-event "
1011                           "overwrite setting for all events\n");
1012                 goto out_err;
1013         }
1014
1015         perf_evlist__config(evlist, opts, &callchain_param);
1016
1017         evlist__for_each_entry(evlist, counter) {
1018 try_again:
1019                 if (evsel__open(counter, top->evlist->core.cpus,
1020                                      top->evlist->core.threads) < 0) {
1021
1022                         /*
1023                          * Specially handle overwrite fall back.
1024                          * Because perf top is the only tool which has
1025                          * overwrite mode by default, support
1026                          * both overwrite and non-overwrite mode, and
1027                          * require consistent mode for all events.
1028                          *
1029                          * May move it to generic code with more tools
1030                          * have similar attribute.
1031                          */
1032                         if (perf_missing_features.write_backward &&
1033                             perf_top_overwrite_fallback(top, counter))
1034                                 goto try_again;
1035
1036                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
1037                                 if (verbose > 0)
1038                                         ui__warning("%s\n", msg);
1039                                 goto try_again;
1040                         }
1041
1042                         perf_evsel__open_strerror(counter, &opts->target,
1043                                                   errno, msg, sizeof(msg));
1044                         ui__error("%s\n", msg);
1045                         goto out_err;
1046                 }
1047         }
1048
1049         if (evlist__mmap(evlist, opts->mmap_pages) < 0) {
1050                 ui__error("Failed to mmap with %d (%s)\n",
1051                             errno, str_error_r(errno, msg, sizeof(msg)));
1052                 goto out_err;
1053         }
1054
1055         return 0;
1056
1057 out_err:
1058         return -1;
1059 }
1060
1061 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1062 {
1063         if (callchain->mode != CHAIN_NONE) {
1064                 if (callchain_register_param(callchain) < 0) {
1065                         ui__error("Can't register callchain params.\n");
1066                         return -EINVAL;
1067                 }
1068         }
1069
1070         return 0;
1071 }
1072
1073 static struct ordered_events *rotate_queues(struct perf_top *top)
1074 {
1075         struct ordered_events *in = top->qe.in;
1076
1077         if (top->qe.in == &top->qe.data[1])
1078                 top->qe.in = &top->qe.data[0];
1079         else
1080                 top->qe.in = &top->qe.data[1];
1081
1082         return in;
1083 }
1084
1085 static void *process_thread(void *arg)
1086 {
1087         struct perf_top *top = arg;
1088
1089         while (!done) {
1090                 struct ordered_events *out, *in = top->qe.in;
1091
1092                 if (!in->nr_events) {
1093                         usleep(100);
1094                         continue;
1095                 }
1096
1097                 out = rotate_queues(top);
1098
1099                 pthread_mutex_lock(&top->qe.mutex);
1100                 top->qe.rotate = true;
1101                 pthread_cond_wait(&top->qe.cond, &top->qe.mutex);
1102                 pthread_mutex_unlock(&top->qe.mutex);
1103
1104                 if (ordered_events__flush(out, OE_FLUSH__TOP))
1105                         pr_err("failed to process events\n");
1106         }
1107
1108         return NULL;
1109 }
1110
1111 /*
1112  * Allow only 'top->delay_secs' seconds behind samples.
1113  */
1114 static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1115 {
1116         union perf_event *event = qevent->event;
1117         u64 delay_timestamp;
1118
1119         if (event->header.type != PERF_RECORD_SAMPLE)
1120                 return false;
1121
1122         delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1123         return delay_timestamp < last_timestamp;
1124 }
1125
1126 static int deliver_event(struct ordered_events *qe,
1127                          struct ordered_event *qevent)
1128 {
1129         struct perf_top *top = qe->data;
1130         struct evlist *evlist = top->evlist;
1131         struct perf_session *session = top->session;
1132         union perf_event *event = qevent->event;
1133         struct perf_sample sample;
1134         struct evsel *evsel;
1135         struct machine *machine;
1136         int ret = -1;
1137
1138         if (should_drop(qevent, top)) {
1139                 top->drop++;
1140                 top->drop_total++;
1141                 return 0;
1142         }
1143
1144         ret = perf_evlist__parse_sample(evlist, event, &sample);
1145         if (ret) {
1146                 pr_err("Can't parse sample, err = %d\n", ret);
1147                 goto next_event;
1148         }
1149
1150         evsel = perf_evlist__id2evsel(session->evlist, sample.id);
1151         assert(evsel != NULL);
1152
1153         if (event->header.type == PERF_RECORD_SAMPLE) {
1154                 if (evswitch__discard(&top->evswitch, evsel))
1155                         return 0;
1156                 ++top->samples;
1157         }
1158
1159         switch (sample.cpumode) {
1160         case PERF_RECORD_MISC_USER:
1161                 ++top->us_samples;
1162                 if (top->hide_user_symbols)
1163                         goto next_event;
1164                 machine = &session->machines.host;
1165                 break;
1166         case PERF_RECORD_MISC_KERNEL:
1167                 ++top->kernel_samples;
1168                 if (top->hide_kernel_symbols)
1169                         goto next_event;
1170                 machine = &session->machines.host;
1171                 break;
1172         case PERF_RECORD_MISC_GUEST_KERNEL:
1173                 ++top->guest_kernel_samples;
1174                 machine = perf_session__find_machine(session,
1175                                                      sample.pid);
1176                 break;
1177         case PERF_RECORD_MISC_GUEST_USER:
1178                 ++top->guest_us_samples;
1179                 /*
1180                  * TODO: we don't process guest user from host side
1181                  * except simple counting.
1182                  */
1183                 goto next_event;
1184         default:
1185                 if (event->header.type == PERF_RECORD_SAMPLE)
1186                         goto next_event;
1187                 machine = &session->machines.host;
1188                 break;
1189         }
1190
1191         if (event->header.type == PERF_RECORD_SAMPLE) {
1192                 perf_event__process_sample(&top->tool, event, evsel,
1193                                            &sample, machine);
1194         } else if (event->header.type == PERF_RECORD_LOST) {
1195                 perf_top__process_lost(top, event, evsel);
1196         } else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1197                 perf_top__process_lost_samples(top, event, evsel);
1198         } else if (event->header.type < PERF_RECORD_MAX) {
1199                 hists__inc_nr_events(evsel__hists(evsel), event->header.type);
1200                 machine__process_event(machine, event, &sample);
1201         } else
1202                 ++session->evlist->stats.nr_unknown_events;
1203
1204         ret = 0;
1205 next_event:
1206         return ret;
1207 }
1208
1209 static void init_process_thread(struct perf_top *top)
1210 {
1211         ordered_events__init(&top->qe.data[0], deliver_event, top);
1212         ordered_events__init(&top->qe.data[1], deliver_event, top);
1213         ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1214         ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1215         top->qe.in = &top->qe.data[0];
1216         pthread_mutex_init(&top->qe.mutex, NULL);
1217         pthread_cond_init(&top->qe.cond, NULL);
1218 }
1219
1220 static int __cmd_top(struct perf_top *top)
1221 {
1222         struct record_opts *opts = &top->record_opts;
1223         pthread_t thread, thread_process;
1224         int ret;
1225
1226         if (!top->annotation_opts.objdump_path) {
1227                 ret = perf_env__lookup_objdump(&top->session->header.env,
1228                                                &top->annotation_opts.objdump_path);
1229                 if (ret)
1230                         return ret;
1231         }
1232
1233         ret = callchain_param__setup_sample_type(&callchain_param);
1234         if (ret)
1235                 return ret;
1236
1237         if (perf_session__register_idle_thread(top->session) < 0)
1238                 return ret;
1239
1240         if (top->nr_threads_synthesize > 1)
1241                 perf_set_multithreaded();
1242
1243         init_process_thread(top);
1244
1245         if (opts->record_namespaces)
1246                 top->tool.namespace_events = true;
1247
1248         ret = perf_event__synthesize_bpf_events(top->session, perf_event__process,
1249                                                 &top->session->machines.host,
1250                                                 &top->record_opts);
1251         if (ret < 0)
1252                 pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n");
1253
1254         machine__synthesize_threads(&top->session->machines.host, &opts->target,
1255                                     top->evlist->core.threads, false,
1256                                     top->nr_threads_synthesize);
1257
1258         if (top->nr_threads_synthesize > 1)
1259                 perf_set_singlethreaded();
1260
1261         if (perf_hpp_list.socket) {
1262                 ret = perf_env__read_cpu_topology_map(&perf_env);
1263                 if (ret < 0) {
1264                         char errbuf[BUFSIZ];
1265                         const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1266
1267                         ui__error("Could not read the CPU topology map: %s\n", err);
1268                         return ret;
1269                 }
1270         }
1271
1272         ret = perf_top__start_counters(top);
1273         if (ret)
1274                 return ret;
1275
1276         top->session->evlist = top->evlist;
1277         perf_session__set_id_hdr_size(top->session);
1278
1279         /*
1280          * When perf is starting the traced process, all the events (apart from
1281          * group members) have enable_on_exec=1 set, so don't spoil it by
1282          * prematurely enabling them.
1283          *
1284          * XXX 'top' still doesn't start workloads like record, trace, but should,
1285          * so leave the check here.
1286          */
1287         if (!target__none(&opts->target))
1288                 evlist__enable(top->evlist);
1289
1290         ret = -1;
1291         if (pthread_create(&thread_process, NULL, process_thread, top)) {
1292                 ui__error("Could not create process thread.\n");
1293                 return ret;
1294         }
1295
1296         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1297                                                             display_thread), top)) {
1298                 ui__error("Could not create display thread.\n");
1299                 goto out_join_thread;
1300         }
1301
1302         if (top->realtime_prio) {
1303                 struct sched_param param;
1304
1305                 param.sched_priority = top->realtime_prio;
1306                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1307                         ui__error("Could not set realtime priority.\n");
1308                         goto out_join;
1309                 }
1310         }
1311
1312         /* Wait for a minimal set of events before starting the snapshot */
1313         evlist__poll(top->evlist, 100);
1314
1315         perf_top__mmap_read(top);
1316
1317         while (!done) {
1318                 u64 hits = top->samples;
1319
1320                 perf_top__mmap_read(top);
1321
1322                 if (opts->overwrite || (hits == top->samples))
1323                         ret = evlist__poll(top->evlist, 100);
1324
1325                 if (resize) {
1326                         perf_top__resize(top);
1327                         resize = 0;
1328                 }
1329         }
1330
1331         ret = 0;
1332 out_join:
1333         pthread_join(thread, NULL);
1334 out_join_thread:
1335         pthread_cond_signal(&top->qe.cond);
1336         pthread_join(thread_process, NULL);
1337         return ret;
1338 }
1339
1340 static int
1341 callchain_opt(const struct option *opt, const char *arg, int unset)
1342 {
1343         symbol_conf.use_callchain = true;
1344         return record_callchain_opt(opt, arg, unset);
1345 }
1346
1347 static int
1348 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1349 {
1350         struct callchain_param *callchain = opt->value;
1351
1352         callchain->enabled = !unset;
1353         callchain->record_mode = CALLCHAIN_FP;
1354
1355         /*
1356          * --no-call-graph
1357          */
1358         if (unset) {
1359                 symbol_conf.use_callchain = false;
1360                 callchain->record_mode = CALLCHAIN_NONE;
1361                 return 0;
1362         }
1363
1364         return parse_callchain_top_opt(arg);
1365 }
1366
1367 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1368 {
1369         if (!strcmp(var, "top.call-graph")) {
1370                 var = "call-graph.record-mode";
1371                 return perf_default_config(var, value, cb);
1372         }
1373         if (!strcmp(var, "top.children")) {
1374                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1375                 return 0;
1376         }
1377
1378         return 0;
1379 }
1380
1381 static int
1382 parse_percent_limit(const struct option *opt, const char *arg,
1383                     int unset __maybe_unused)
1384 {
1385         struct perf_top *top = opt->value;
1386
1387         top->min_percent = strtof(arg, NULL);
1388         return 0;
1389 }
1390
1391 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1392         "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1393
1394 int cmd_top(int argc, const char **argv)
1395 {
1396         char errbuf[BUFSIZ];
1397         struct perf_top top = {
1398                 .count_filter        = 5,
1399                 .delay_secs          = 2,
1400                 .record_opts = {
1401                         .mmap_pages     = UINT_MAX,
1402                         .user_freq      = UINT_MAX,
1403                         .user_interval  = ULLONG_MAX,
1404                         .freq           = 4000, /* 4 KHz */
1405                         .target         = {
1406                                 .uses_mmap   = true,
1407                         },
1408                         /*
1409                          * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1410                          * when we pause, fix that and reenable. Probably using a
1411                          * separate evlist with a dummy event, i.e. a non-overwrite
1412                          * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1413                          * stays in overwrite mode. -acme
1414                          * */
1415                         .overwrite      = 0,
1416                         .sample_time    = true,
1417                         .sample_time_set = true,
1418                 },
1419                 .max_stack           = sysctl__max_stack(),
1420                 .annotation_opts     = annotation__default_options,
1421                 .nr_threads_synthesize = UINT_MAX,
1422         };
1423         struct record_opts *opts = &top.record_opts;
1424         struct target *target = &opts->target;
1425         const struct option options[] = {
1426         OPT_CALLBACK('e', "event", &top.evlist, "event",
1427                      "event selector. use 'perf list' to list available events",
1428                      parse_events_option),
1429         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1430         OPT_STRING('p', "pid", &target->pid, "pid",
1431                     "profile events on existing process id"),
1432         OPT_STRING('t', "tid", &target->tid, "tid",
1433                     "profile events on existing thread id"),
1434         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1435                             "system-wide collection from all CPUs"),
1436         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1437                     "list of cpus to monitor"),
1438         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1439                    "file", "vmlinux pathname"),
1440         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1441                     "don't load vmlinux even if found"),
1442         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1443                    "file", "kallsyms pathname"),
1444         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1445                     "hide kernel symbols"),
1446         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1447                      "number of mmap data pages",
1448                      perf_evlist__parse_mmap_pages),
1449         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1450                     "collect data with this RT SCHED_FIFO priority"),
1451         OPT_INTEGER('d', "delay", &top.delay_secs,
1452                     "number of seconds to delay between refreshes"),
1453         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1454                             "dump the symbol table used for profiling"),
1455         OPT_INTEGER('f', "count-filter", &top.count_filter,
1456                     "only display functions with more events than this"),
1457         OPT_BOOLEAN(0, "group", &opts->group,
1458                             "put the counters into a counter group"),
1459         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1460                     "child tasks do not inherit counters"),
1461         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1462                     "symbol to annotate"),
1463         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1464         OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1465                      "profile at this frequency",
1466                       record__parse_freq),
1467         OPT_INTEGER('E', "entries", &top.print_entries,
1468                     "display this many functions"),
1469         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1470                     "hide user symbols"),
1471         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1472         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1473         OPT_INCR('v', "verbose", &verbose,
1474                     "be more verbose (show counter open errors, etc)"),
1475         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1476                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1477                    " Please refer the man page for the complete list."),
1478         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1479                    "output field(s): overhead, period, sample plus all of sort keys"),
1480         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1481                     "Show a column with the number of samples"),
1482         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1483                            NULL, "enables call-graph recording and display",
1484                            &callchain_opt),
1485         OPT_CALLBACK(0, "call-graph", &callchain_param,
1486                      "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1487                      top_callchain_help, &parse_callchain_opt),
1488         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1489                     "Accumulate callchains of children and show total overhead as well"),
1490         OPT_INTEGER(0, "max-stack", &top.max_stack,
1491                     "Set the maximum stack depth when parsing the callchain. "
1492                     "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1493         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1494                    "ignore callees of these functions in call graphs",
1495                    report_parse_ignore_callees_opt),
1496         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1497                     "Show a column with the sum of periods"),
1498         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1499                    "only consider symbols in these dsos"),
1500         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1501                    "only consider symbols in these comms"),
1502         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1503                    "only consider these symbols"),
1504         OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1505                     "Interleave source code with assembly code (default)"),
1506         OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1507                     "Display raw encoding of assembly instructions (default)"),
1508         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1509                     "Enable kernel symbol demangling"),
1510         OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"),
1511         OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path",
1512                     "objdump binary to use for disassembly and annotations"),
1513         OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1514                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1515         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1516         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1517                      "Don't show entries under that percent", parse_percent_limit),
1518         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1519                      "How to display percentage of filtered entries", parse_filter_percentage),
1520         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1521                    "width[,width...]",
1522                    "don't try to adjust column width, use these fixed values"),
1523         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1524                         "per thread proc mmap processing timeout in ms"),
1525         OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1526                      "branch any", "sample any taken branches",
1527                      parse_branch_stack),
1528         OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1529                      "branch filter mask", "branch stack filter modes",
1530                      parse_branch_stack),
1531         OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1532                     "Show raw trace event output (do not use print fmt or plugins)"),
1533         OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1534                     "Show entries in a hierarchy"),
1535         OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1536                     "Use a backward ring buffer, default: no"),
1537         OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1538         OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1539                         "number of thread to run event synthesize"),
1540         OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces,
1541                     "Record namespaces events"),
1542         OPTS_EVSWITCH(&top.evswitch),
1543         OPT_END()
1544         };
1545         struct evlist *sb_evlist = NULL;
1546         const char * const top_usage[] = {
1547                 "perf top [<options>]",
1548                 NULL
1549         };
1550         int status = hists__init();
1551
1552         if (status < 0)
1553                 return status;
1554
1555         top.annotation_opts.min_pcnt = 5;
1556         top.annotation_opts.context  = 4;
1557
1558         top.evlist = evlist__new();
1559         if (top.evlist == NULL)
1560                 return -ENOMEM;
1561
1562         status = perf_config(perf_top_config, &top);
1563         if (status)
1564                 return status;
1565
1566         argc = parse_options(argc, argv, options, top_usage, 0);
1567         if (argc)
1568                 usage_with_options(top_usage, options);
1569
1570         if (!top.evlist->core.nr_entries &&
1571             perf_evlist__add_default(top.evlist) < 0) {
1572                 pr_err("Not enough memory for event selector list\n");
1573                 goto out_delete_evlist;
1574         }
1575
1576         status = evswitch__init(&top.evswitch, top.evlist, stderr);
1577         if (status)
1578                 goto out_delete_evlist;
1579
1580         if (symbol_conf.report_hierarchy) {
1581                 /* disable incompatible options */
1582                 symbol_conf.event_group = false;
1583                 symbol_conf.cumulate_callchain = false;
1584
1585                 if (field_order) {
1586                         pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1587                         parse_options_usage(top_usage, options, "fields", 0);
1588                         parse_options_usage(NULL, options, "hierarchy", 0);
1589                         goto out_delete_evlist;
1590                 }
1591         }
1592
1593         if (opts->branch_stack && callchain_param.enabled)
1594                 symbol_conf.show_branchflag_count = true;
1595
1596         sort__mode = SORT_MODE__TOP;
1597         /* display thread wants entries to be collapsed in a different tree */
1598         perf_hpp_list.need_collapse = 1;
1599
1600         if (top.use_stdio)
1601                 use_browser = 0;
1602         else if (top.use_tui)
1603                 use_browser = 1;
1604
1605         setup_browser(false);
1606
1607         if (setup_sorting(top.evlist) < 0) {
1608                 if (sort_order)
1609                         parse_options_usage(top_usage, options, "s", 1);
1610                 if (field_order)
1611                         parse_options_usage(sort_order ? NULL : top_usage,
1612                                             options, "fields", 0);
1613                 goto out_delete_evlist;
1614         }
1615
1616         status = target__validate(target);
1617         if (status) {
1618                 target__strerror(target, status, errbuf, BUFSIZ);
1619                 ui__warning("%s\n", errbuf);
1620         }
1621
1622         status = target__parse_uid(target);
1623         if (status) {
1624                 int saved_errno = errno;
1625
1626                 target__strerror(target, status, errbuf, BUFSIZ);
1627                 ui__error("%s\n", errbuf);
1628
1629                 status = -saved_errno;
1630                 goto out_delete_evlist;
1631         }
1632
1633         if (target__none(target))
1634                 target->system_wide = true;
1635
1636         if (perf_evlist__create_maps(top.evlist, target) < 0) {
1637                 ui__error("Couldn't create thread/CPU maps: %s\n",
1638                           errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1639                 goto out_delete_evlist;
1640         }
1641
1642         if (top.delay_secs < 1)
1643                 top.delay_secs = 1;
1644
1645         if (record_opts__config(opts)) {
1646                 status = -EINVAL;
1647                 goto out_delete_evlist;
1648         }
1649
1650         top.sym_evsel = evlist__first(top.evlist);
1651
1652         if (!callchain_param.enabled) {
1653                 symbol_conf.cumulate_callchain = false;
1654                 perf_hpp__cancel_cumulate();
1655         }
1656
1657         if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1658                 callchain_param.order = ORDER_CALLER;
1659
1660         status = symbol__annotation_init();
1661         if (status < 0)
1662                 goto out_delete_evlist;
1663
1664         annotation_config__init();
1665
1666         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1667         status = symbol__init(NULL);
1668         if (status < 0)
1669                 goto out_delete_evlist;
1670
1671         sort__setup_elide(stdout);
1672
1673         get_term_dimensions(&top.winsize);
1674         if (top.print_entries == 0) {
1675                 perf_top__update_print_entries(&top);
1676                 signal(SIGWINCH, winch_sig);
1677         }
1678
1679         top.session = perf_session__new(NULL, false, NULL);
1680         if (IS_ERR(top.session)) {
1681                 status = PTR_ERR(top.session);
1682                 goto out_delete_evlist;
1683         }
1684
1685         if (!top.record_opts.no_bpf_event)
1686                 bpf_event__add_sb_event(&sb_evlist, &perf_env);
1687
1688         if (perf_evlist__start_sb_thread(sb_evlist, target)) {
1689                 pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1690                 opts->no_bpf_event = true;
1691         }
1692
1693         status = __cmd_top(&top);
1694
1695         if (!opts->no_bpf_event)
1696                 perf_evlist__stop_sb_thread(sb_evlist);
1697
1698 out_delete_evlist:
1699         evlist__delete(top.evlist);
1700         perf_session__delete(top.session);
1701
1702         return status;
1703 }