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 #include <perf/mmap.h>
  86
  87 static volatile int done;
  88 static volatile int resize;
  89
  90 #define HEADER_LINE_NR  5
  91
  92 static void perf_top__update_print_entries(struct perf_top *top)
  93 {
  94         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
  95 }
  96
  97 static void winch_sig(int sig __maybe_unused)
  98 {
  99         resize = 1;
 100 }
 101
 102 static void perf_top__resize(struct perf_top *top)
 103 {
 104         get_term_dimensions(&top->winsize);
 105         perf_top__update_print_entries(top);
 106 }
 107
 108 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
 109 {
 110         struct evsel *evsel;
 111         struct symbol *sym;
 112         struct annotation *notes;
 113         struct map *map;
 114         int err = -1;
 115
 116         if (!he || !he->ms.sym)
 117                 return -1;
 118
 119         evsel = hists_to_evsel(he->hists);
 120
 121         sym = he->ms.sym;
 122         map = he->ms.map;
 123
 124         /*
 125          * We can't annotate with just /proc/kallsyms
 126          */
 127         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 128             !dso__is_kcore(map->dso)) {
 129                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 130                        "path\n", sym->name);
 131                 sleep(1);
 132                 return -1;
 133         }
 134
 135         notes = symbol__annotation(sym);
 136         pthread_mutex_lock(&notes->lock);
 137
 138         if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
 139                 pthread_mutex_unlock(&notes->lock);
 140                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 141                        sym->name);
 142                 sleep(1);
 143                 return err;
 144         }
 145
 146         err = symbol__annotate(&he->ms, evsel, &top->annotation_opts, NULL);
 147         if (err == 0) {
 148                 top->sym_filter_entry = he;
 149         } else {
 150                 char msg[BUFSIZ];
 151                 symbol__strerror_disassemble(&he->ms, err, msg, sizeof(msg));
 152                 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
 153         }
 154
 155         pthread_mutex_unlock(&notes->lock);
 156         return err;
 157 }
 158
 159 static void __zero_source_counters(struct hist_entry *he)
 160 {
 161         struct symbol *sym = he->ms.sym;
 162         symbol__annotate_zero_histograms(sym);
 163 }
 164
 165 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 166 {
 167         struct utsname uts;
 168         int err = uname(&uts);
 169
 170         ui__warning("Out of bounds address found:\n\n"
 171                     "Addr:   %" PRIx64 "\n"
 172                     "DSO:    %s %c\n"
 173                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
 174                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 175                     "Arch:   %s\n"
 176                     "Kernel: %s\n"
 177                     "Tools:  %s\n\n"
 178                     "Not all samples will be on the annotation output.\n\n"
 179                     "Please report to linux-kernel@vger.kernel.org\n",
 180                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
 181                     map->start, map->end, sym->start, sym->end,
 182                     sym->binding == STB_GLOBAL ? 'g' :
 183                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 184                     err ? "[unknown]" : uts.machine,
 185                     err ? "[unknown]" : uts.release, perf_version_string);
 186         if (use_browser <= 0)
 187                 sleep(5);
 188
 189         map->erange_warned = true;
 190 }
 191
 192 static void perf_top__record_precise_ip(struct perf_top *top,
 193                                         struct hist_entry *he,
 194                                         struct perf_sample *sample,
 195                                         struct evsel *evsel, u64 ip)
 196 {
 197         struct annotation *notes;
 198         struct symbol *sym = he->ms.sym;
 199         int err = 0;
 200
 201         if (sym == NULL || (use_browser == 0 &&
 202                             (top->sym_filter_entry == NULL ||
 203                              top->sym_filter_entry->ms.sym != sym)))
 204                 return;
 205
 206         notes = symbol__annotation(sym);
 207
 208         if (pthread_mutex_trylock(&notes->lock))
 209                 return;
 210
 211         err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
 212
 213         pthread_mutex_unlock(&notes->lock);
 214
 215         if (unlikely(err)) {
 216                 /*
 217                  * This function is now called with he->hists->lock held.
 218                  * Release it before going to sleep.
 219                  */
 220                 pthread_mutex_unlock(&he->hists->lock);
 221
 222                 if (err == -ERANGE && !he->ms.map->erange_warned)
 223                         ui__warn_map_erange(he->ms.map, sym, ip);
 224                 else if (err == -ENOMEM) {
 225                         pr_err("Not enough memory for annotating '%s' symbol!\n",
 226                                sym->name);
 227                         sleep(1);
 228                 }
 229
 230                 pthread_mutex_lock(&he->hists->lock);
 231         }
 232 }
 233
 234 static void perf_top__show_details(struct perf_top *top)
 235 {
 236         struct hist_entry *he = top->sym_filter_entry;
 237         struct evsel *evsel;
 238         struct annotation *notes;
 239         struct symbol *symbol;
 240         int more;
 241
 242         if (!he)
 243                 return;
 244
 245         evsel = hists_to_evsel(he->hists);
 246
 247         symbol = he->ms.sym;
 248         notes = symbol__annotation(symbol);
 249
 250         pthread_mutex_lock(&notes->lock);
 251
 252         symbol__calc_percent(symbol, evsel);
 253
 254         if (notes->src == NULL)
 255                 goto out_unlock;
 256
 257         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
 258         printf("  Events  Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
 259
 260         more = symbol__annotate_printf(&he->ms, top->sym_evsel, &top->annotation_opts);
 261
 262         if (top->evlist->enabled) {
 263                 if (top->zero)
 264                         symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 265                 else
 266                         symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 267         }
 268         if (more != 0)
 269                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 270 out_unlock:
 271         pthread_mutex_unlock(&notes->lock);
 272 }
 273
 274 static void perf_top__resort_hists(struct perf_top *t)
 275 {
 276         struct evlist *evlist = t->evlist;
 277         struct evsel *pos;
 278
 279         evlist__for_each_entry(evlist, pos) {
 280                 struct hists *hists = evsel__hists(pos);
 281
 282                 /*
 283                  * unlink existing entries so that they can be linked
 284                  * in a correct order in hists__match() below.
 285                  */
 286                 hists__unlink(hists);
 287
 288                 if (evlist->enabled) {
 289                         if (t->zero) {
 290                                 hists__delete_entries(hists);
 291                         } else {
 292                                 hists__decay_entries(hists, t->hide_user_symbols,
 293                                                      t->hide_kernel_symbols);
 294                         }
 295                 }
 296
 297                 hists__collapse_resort(hists, NULL);
 298
 299                 /* Non-group events are considered as leader */
 300                 if (symbol_conf.event_group &&
 301                     !perf_evsel__is_group_leader(pos)) {
 302                         struct hists *leader_hists = evsel__hists(pos->leader);
 303
 304                         hists__match(leader_hists, hists);
 305                         hists__link(leader_hists, hists);
 306                 }
 307         }
 308
 309         evlist__for_each_entry(evlist, pos) {
 310                 perf_evsel__output_resort(pos, NULL);
 311         }
 312 }
 313
 314 static void perf_top__print_sym_table(struct perf_top *top)
 315 {
 316         char bf[160];
 317         int printed = 0;
 318         const int win_width = top->winsize.ws_col - 1;
 319         struct evsel *evsel = top->sym_evsel;
 320         struct hists *hists = evsel__hists(evsel);
 321
 322         puts(CONSOLE_CLEAR);
 323
 324         perf_top__header_snprintf(top, bf, sizeof(bf));
 325         printf("%s\n", bf);
 326
 327         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 328
 329         if (!top->record_opts.overwrite &&
 330             (hists->stats.nr_lost_warned !=
 331             hists->stats.nr_events[PERF_RECORD_LOST])) {
 332                 hists->stats.nr_lost_warned =
 333                               hists->stats.nr_events[PERF_RECORD_LOST];
 334                 color_fprintf(stdout, PERF_COLOR_RED,
 335                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 336                               hists->stats.nr_lost_warned);
 337                 ++printed;
 338         }
 339
 340         if (top->sym_filter_entry) {
 341                 perf_top__show_details(top);
 342                 return;
 343         }
 344
 345         perf_top__resort_hists(top);
 346
 347         hists__output_recalc_col_len(hists, top->print_entries - printed);
 348         putchar('\n');
 349         hists__fprintf(hists, false, top->print_entries - printed, win_width,
 350                        top->min_percent, stdout, !symbol_conf.use_callchain);
 351 }
 352
 353 static void prompt_integer(int *target, const char *msg)
 354 {
 355         char *buf = malloc(0), *p;
 356         size_t dummy = 0;
 357         int tmp;
 358
 359         fprintf(stdout, "\n%s: ", msg);
 360         if (getline(&buf, &dummy, stdin) < 0)
 361                 return;
 362
 363         p = strchr(buf, '\n');
 364         if (p)
 365                 *p = 0;
 366
 367         p = buf;
 368         while(*p) {
 369                 if (!isdigit(*p))
 370                         goto out_free;
 371                 p++;
 372         }
 373         tmp = strtoul(buf, NULL, 10);
 374         *target = tmp;
 375 out_free:
 376         free(buf);
 377 }
 378
 379 static void prompt_percent(int *target, const char *msg)
 380 {
 381         int tmp = 0;
 382
 383         prompt_integer(&tmp, msg);
 384         if (tmp >= 0 && tmp <= 100)
 385                 *target = tmp;
 386 }
 387
 388 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 389 {
 390         char *buf = malloc(0), *p;
 391         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 392         struct hists *hists = evsel__hists(top->sym_evsel);
 393         struct rb_node *next;
 394         size_t dummy = 0;
 395
 396         /* zero counters of active symbol */
 397         if (syme) {
 398                 __zero_source_counters(syme);
 399                 top->sym_filter_entry = NULL;
 400         }
 401
 402         fprintf(stdout, "\n%s: ", msg);
 403         if (getline(&buf, &dummy, stdin) < 0)
 404                 goto out_free;
 405
 406         p = strchr(buf, '\n');
 407         if (p)
 408                 *p = 0;
 409
 410         next = rb_first_cached(&hists->entries);
 411         while (next) {
 412                 n = rb_entry(next, struct hist_entry, rb_node);
 413                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 414                         found = n;
 415                         break;
 416                 }
 417                 next = rb_next(&n->rb_node);
 418         }
 419
 420         if (!found) {
 421                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 422                 sleep(1);
 423         } else
 424                 perf_top__parse_source(top, found);
 425
 426 out_free:
 427         free(buf);
 428 }
 429
 430 static void perf_top__print_mapped_keys(struct perf_top *top)
 431 {
 432         char *name = NULL;
 433
 434         if (top->sym_filter_entry) {
 435                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 436                 name = sym->name;
 437         }
 438
 439         fprintf(stdout, "\nMapped keys:\n");
 440         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 441         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 442
 443         if (top->evlist->core.nr_entries > 1)
 444                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 445
 446         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 447
 448         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
 449         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 450         fprintf(stdout, "\t[S]     stop annotation.\n");
 451
 452         fprintf(stdout,
 453                 "\t[K]     hide kernel symbols.             \t(%s)\n",
 454                 top->hide_kernel_symbols ? "yes" : "no");
 455         fprintf(stdout,
 456                 "\t[U]     hide user symbols.               \t(%s)\n",
 457                 top->hide_user_symbols ? "yes" : "no");
 458         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 459         fprintf(stdout, "\t[qQ]    quit.\n");
 460 }
 461
 462 static int perf_top__key_mapped(struct perf_top *top, int c)
 463 {
 464         switch (c) {
 465                 case 'd':
 466                 case 'e':
 467                 case 'f':
 468                 case 'z':
 469                 case 'q':
 470                 case 'Q':
 471                 case 'K':
 472                 case 'U':
 473                 case 'F':
 474                 case 's':
 475                 case 'S':
 476                         return 1;
 477                 case 'E':
 478                         return top->evlist->core.nr_entries > 1 ? 1 : 0;
 479                 default:
 480                         break;
 481         }
 482
 483         return 0;
 484 }
 485
 486 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 487 {
 488         bool ret = true;
 489
 490         if (!perf_top__key_mapped(top, c)) {
 491                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 492                 struct termios save;
 493
 494                 perf_top__print_mapped_keys(top);
 495                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 496                 fflush(stdout);
 497
 498                 set_term_quiet_input(&save);
 499
 500                 poll(&stdin_poll, 1, -1);
 501                 c = getc(stdin);
 502
 503                 tcsetattr(0, TCSAFLUSH, &save);
 504                 if (!perf_top__key_mapped(top, c))
 505                         return ret;
 506         }
 507
 508         switch (c) {
 509                 case 'd':
 510                         prompt_integer(&top->delay_secs, "Enter display delay");
 511                         if (top->delay_secs < 1)
 512                                 top->delay_secs = 1;
 513                         break;
 514                 case 'e':
 515                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 516                         if (top->print_entries == 0) {
 517                                 perf_top__resize(top);
 518                                 signal(SIGWINCH, winch_sig);
 519                         } else {
 520                                 signal(SIGWINCH, SIG_DFL);
 521                         }
 522                         break;
 523                 case 'E':
 524                         if (top->evlist->core.nr_entries > 1) {
 525                                 /* Select 0 as the default event: */
 526                                 int counter = 0;
 527
 528                                 fprintf(stderr, "\nAvailable events:");
 529
 530                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
 531                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
 532
 533                                 prompt_integer(&counter, "Enter details event counter");
 534
 535                                 if (counter >= top->evlist->core.nr_entries) {
 536                                         top->sym_evsel = evlist__first(top->evlist);
 537                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
 538                                         sleep(1);
 539                                         break;
 540                                 }
 541                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
 542                                         if (top->sym_evsel->idx == counter)
 543                                                 break;
 544                         } else
 545                                 top->sym_evsel = evlist__first(top->evlist);
 546                         break;
 547                 case 'f':
 548                         prompt_integer(&top->count_filter, "Enter display event count filter");
 549                         break;
 550                 case 'F':
 551                         prompt_percent(&top->annotation_opts.min_pcnt,
 552                                        "Enter details display event filter (percent)");
 553                         break;
 554                 case 'K':
 555                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 556                         break;
 557                 case 'q':
 558                 case 'Q':
 559                         printf("exiting.\n");
 560                         if (top->dump_symtab)
 561                                 perf_session__fprintf_dsos(top->session, stderr);
 562                         ret = false;
 563                         break;
 564                 case 's':
 565                         perf_top__prompt_symbol(top, "Enter details symbol");
 566                         break;
 567                 case 'S':
 568                         if (!top->sym_filter_entry)
 569                                 break;
 570                         else {
 571                                 struct hist_entry *syme = top->sym_filter_entry;
 572
 573                                 top->sym_filter_entry = NULL;
 574                                 __zero_source_counters(syme);
 575                         }
 576                         break;
 577                 case 'U':
 578                         top->hide_user_symbols = !top->hide_user_symbols;
 579                         break;
 580                 case 'z':
 581                         top->zero = !top->zero;
 582                         break;
 583                 default:
 584                         break;
 585         }
 586
 587         return ret;
 588 }
 589
 590 static void perf_top__sort_new_samples(void *arg)
 591 {
 592         struct perf_top *t = arg;
 593
 594         if (t->evlist->selected != NULL)
 595                 t->sym_evsel = t->evlist->selected;
 596
 597         perf_top__resort_hists(t);
 598
 599         if (t->lost || t->drop)
 600                 pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
 601 }
 602
 603 static void stop_top(void)
 604 {
 605         session_done = 1;
 606         done = 1;
 607 }
 608
 609 static void *display_thread_tui(void *arg)
 610 {
 611         struct evsel *pos;
 612         struct perf_top *top = arg;
 613         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 614         struct hist_browser_timer hbt = {
 615                 .timer          = perf_top__sort_new_samples,
 616                 .arg            = top,
 617                 .refresh        = top->delay_secs,
 618         };
 619         int ret;
 620
 621         /* In order to read symbols from other namespaces perf to  needs to call
 622          * setns(2).  This isn't permitted if the struct_fs has multiple users.
 623          * unshare(2) the fs so that we may continue to setns into namespaces
 624          * that we're observing.
 625          */
 626         unshare(CLONE_FS);
 627
 628         prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
 629
 630 repeat:
 631         perf_top__sort_new_samples(top);
 632
 633         /*
 634          * Initialize the uid_filter_str, in the future the TUI will allow
 635          * Zooming in/out UIDs. For now just use whatever the user passed
 636          * via --uid.
 637          */
 638         evlist__for_each_entry(top->evlist, pos) {
 639                 struct hists *hists = evsel__hists(pos);
 640                 hists->uid_filter_str = top->record_opts.target.uid_str;
 641         }
 642
 643         ret = perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
 644                                       top->min_percent,
 645                                       &top->session->header.env,
 646                                       !top->record_opts.overwrite,
 647                                       &top->annotation_opts);
 648
 649         if (ret == K_RELOAD) {
 650                 top->zero = true;
 651                 goto repeat;
 652         } else
 653                 stop_top();
 654
 655         return NULL;
 656 }
 657
 658 static void display_sig(int sig __maybe_unused)
 659 {
 660         stop_top();
 661 }
 662
 663 static void display_setup_sig(void)
 664 {
 665         signal(SIGSEGV, sighandler_dump_stack);
 666         signal(SIGFPE, sighandler_dump_stack);
 667         signal(SIGINT,  display_sig);
 668         signal(SIGQUIT, display_sig);
 669         signal(SIGTERM, display_sig);
 670 }
 671
 672 static void *display_thread(void *arg)
 673 {
 674         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 675         struct termios save;
 676         struct perf_top *top = arg;
 677         int delay_msecs, c;
 678
 679         /* In order to read symbols from other namespaces perf to  needs to call
 680          * setns(2).  This isn't permitted if the struct_fs has multiple users.
 681          * unshare(2) the fs so that we may continue to setns into namespaces
 682          * that we're observing.
 683          */
 684         unshare(CLONE_FS);
 685
 686         prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
 687
 688         display_setup_sig();
 689         pthread__unblock_sigwinch();
 690 repeat:
 691         delay_msecs = top->delay_secs * MSEC_PER_SEC;
 692         set_term_quiet_input(&save);
 693         /* trash return*/
 694         clearerr(stdin);
 695         if (poll(&stdin_poll, 1, 0) > 0)
 696                 getc(stdin);
 697
 698         while (!done) {
 699                 perf_top__print_sym_table(top);
 700                 /*
 701                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 702                  * refresh screen in both cases.
 703                  */
 704                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 705                 case 0:
 706                         continue;
 707                 case -1:
 708                         if (errno == EINTR)
 709                                 continue;
 710                         __fallthrough;
 711                 default:
 712                         c = getc(stdin);
 713                         tcsetattr(0, TCSAFLUSH, &save);
 714
 715                         if (perf_top__handle_keypress(top, c))
 716                                 goto repeat;
 717                         stop_top();
 718                 }
 719         }
 720
 721         tcsetattr(0, TCSAFLUSH, &save);
 722         return NULL;
 723 }
 724
 725 static int hist_iter__top_callback(struct hist_entry_iter *iter,
 726                                    struct addr_location *al, bool single,
 727                                    void *arg)
 728 {
 729         struct perf_top *top = arg;
 730         struct hist_entry *he = iter->he;
 731         struct evsel *evsel = iter->evsel;
 732
 733         if (perf_hpp_list.sym && single)
 734                 perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
 735
 736         hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
 737                      !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY),
 738                      NULL);
 739         return 0;
 740 }
 741
 742 static void perf_event__process_sample(struct perf_tool *tool,
 743                                        const union perf_event *event,
 744                                        struct evsel *evsel,
 745                                        struct perf_sample *sample,
 746                                        struct machine *machine)
 747 {
 748         struct perf_top *top = container_of(tool, struct perf_top, tool);
 749         struct addr_location al;
 750         int err;
 751
 752         if (!machine && perf_guest) {
 753                 static struct intlist *seen;
 754
 755                 if (!seen)
 756                         seen = intlist__new(NULL);
 757
 758                 if (!intlist__has_entry(seen, sample->pid)) {
 759                         pr_err("Can't find guest [%d]'s kernel information\n",
 760                                 sample->pid);
 761                         intlist__add(seen, sample->pid);
 762                 }
 763                 return;
 764         }
 765
 766         if (!machine) {
 767                 pr_err("%u unprocessable samples recorded.\r",
 768                        top->session->evlist->stats.nr_unprocessable_samples++);
 769                 return;
 770         }
 771
 772         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 773                 top->exact_samples++;
 774
 775         if (machine__resolve(machine, &al, sample) < 0)
 776                 return;
 777
 778         if (!machine->kptr_restrict_warned &&
 779             symbol_conf.kptr_restrict &&
 780             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 781                 if (!perf_evlist__exclude_kernel(top->session->evlist)) {
 782                         ui__warning(
 783 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 784 "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
 785 "Kernel%s samples will not be resolved.\n",
 786                           al.map && map__has_symbols(al.map) ?
 787                           " modules" : "");
 788                         if (use_browser <= 0)
 789                                 sleep(5);
 790                 }
 791                 machine->kptr_restrict_warned = true;
 792         }
 793
 794         if (al.sym == NULL && al.map != NULL) {
 795                 const char *msg = "Kernel samples will not be resolved.\n";
 796                 /*
 797                  * As we do lazy loading of symtabs we only will know if the
 798                  * specified vmlinux file is invalid when we actually have a
 799                  * hit in kernel space and then try to load it. So if we get
 800                  * here and there are _no_ symbols in the DSO backing the
 801                  * kernel map, bail out.
 802                  *
 803                  * We may never get here, for instance, if we use -K/
 804                  * --hide-kernel-symbols, even if the user specifies an
 805                  * invalid --vmlinux ;-)
 806                  */
 807                 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
 808                     __map__is_kernel(al.map) && map__has_symbols(al.map)) {
 809                         if (symbol_conf.vmlinux_name) {
 810                                 char serr[256];
 811                                 dso__strerror_load(al.map->dso, serr, sizeof(serr));
 812                                 ui__warning("The %s file can't be used: %s\n%s",
 813                                             symbol_conf.vmlinux_name, serr, msg);
 814                         } else {
 815                                 ui__warning("A vmlinux file was not found.\n%s",
 816                                             msg);
 817                         }
 818
 819                         if (use_browser <= 0)
 820                                 sleep(5);
 821                         top->vmlinux_warned = true;
 822                 }
 823         }
 824
 825         if (al.sym == NULL || !al.sym->idle) {
 826                 struct hists *hists = evsel__hists(evsel);
 827                 struct hist_entry_iter iter = {
 828                         .evsel          = evsel,
 829                         .sample         = sample,
 830                         .add_entry_cb   = hist_iter__top_callback,
 831                 };
 832
 833                 if (symbol_conf.cumulate_callchain)
 834                         iter.ops = &hist_iter_cumulative;
 835                 else
 836                         iter.ops = &hist_iter_normal;
 837
 838                 pthread_mutex_lock(&hists->lock);
 839
 840                 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
 841                 if (err < 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         struct hists *hists = evsel__hists(evsel);
 855
 856         top->lost += event->lost.lost;
 857         top->lost_total += event->lost.lost;
 858         hists->stats.total_lost += event->lost.lost;
 859 }
 860
 861 static void
 862 perf_top__process_lost_samples(struct perf_top *top,
 863                                union perf_event *event,
 864                                struct evsel *evsel)
 865 {
 866         struct hists *hists = evsel__hists(evsel);
 867
 868         top->lost += event->lost_samples.lost;
 869         top->lost_total += event->lost_samples.lost;
 870         hists->stats.total_lost_samples += event->lost_samples.lost;
 871 }
 872
 873 static u64 last_timestamp;
 874
 875 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 876 {
 877         struct record_opts *opts = &top->record_opts;
 878         struct evlist *evlist = top->evlist;
 879         struct mmap *md;
 880         union perf_event *event;
 881
 882         md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
 883         if (perf_mmap__read_init(&md->core) < 0)
 884                 return;
 885
 886         while ((event = perf_mmap__read_event(&md->core)) != NULL) {
 887                 int ret;
 888
 889                 ret = perf_evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
 890                 if (ret && ret != -1)
 891                         break;
 892
 893                 ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0);
 894                 if (ret)
 895                         break;
 896
 897                 perf_mmap__consume(&md->core);
 898
 899                 if (top->qe.rotate) {
 900                         pthread_mutex_lock(&top->qe.mutex);
 901                         top->qe.rotate = false;
 902                         pthread_cond_signal(&top->qe.cond);
 903                         pthread_mutex_unlock(&top->qe.mutex);
 904                 }
 905         }
 906
 907         perf_mmap__read_done(&md->core);
 908 }
 909
 910 static void perf_top__mmap_read(struct perf_top *top)
 911 {
 912         bool overwrite = top->record_opts.overwrite;
 913         struct evlist *evlist = top->evlist;
 914         int i;
 915
 916         if (overwrite)
 917                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
 918
 919         for (i = 0; i < top->evlist->core.nr_mmaps; i++)
 920                 perf_top__mmap_read_idx(top, i);
 921
 922         if (overwrite) {
 923                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
 924                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 925         }
 926 }
 927
 928 /*
 929  * Check per-event overwrite term.
 930  * perf top should support consistent term for all events.
 931  * - All events don't have per-event term
 932  *   E.g. "cpu/cpu-cycles/,cpu/instructions/"
 933  *   Nothing change, return 0.
 934  * - All events have same per-event term
 935  *   E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
 936  *   Using the per-event setting to replace the opts->overwrite if
 937  *   they are different, then return 0.
 938  * - Events have different per-event term
 939  *   E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
 940  *   Return -1
 941  * - Some of the event set per-event term, but some not.
 942  *   E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
 943  *   Return -1
 944  */
 945 static int perf_top__overwrite_check(struct perf_top *top)
 946 {
 947         struct record_opts *opts = &top->record_opts;
 948         struct evlist *evlist = top->evlist;
 949         struct perf_evsel_config_term *term;
 950         struct list_head *config_terms;
 951         struct evsel *evsel;
 952         int set, overwrite = -1;
 953
 954         evlist__for_each_entry(evlist, evsel) {
 955                 set = -1;
 956                 config_terms = &evsel->config_terms;
 957                 list_for_each_entry(term, config_terms, list) {
 958                         if (term->type == PERF_EVSEL__CONFIG_TERM_OVERWRITE)
 959                                 set = term->val.overwrite ? 1 : 0;
 960                 }
 961
 962                 /* no term for current and previous event (likely) */
 963                 if ((overwrite < 0) && (set < 0))
 964                         continue;
 965
 966                 /* has term for both current and previous event, compare */
 967                 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
 968                         return -1;
 969
 970                 /* no term for current event but has term for previous one */
 971                 if ((overwrite >= 0) && (set < 0))
 972                         return -1;
 973
 974                 /* has term for current event */
 975                 if ((overwrite < 0) && (set >= 0)) {
 976                         /* if it's first event, set overwrite */
 977                         if (evsel == evlist__first(evlist))
 978                                 overwrite = set;
 979                         else
 980                                 return -1;
 981                 }
 982         }
 983
 984         if ((overwrite >= 0) && (opts->overwrite != overwrite))
 985                 opts->overwrite = overwrite;
 986
 987         return 0;
 988 }
 989
 990 static int perf_top_overwrite_fallback(struct perf_top *top,
 991                                        struct evsel *evsel)
 992 {
 993         struct record_opts *opts = &top->record_opts;
 994         struct evlist *evlist = top->evlist;
 995         struct evsel *counter;
 996
 997         if (!opts->overwrite)
 998                 return 0;
 999
1000         /* only fall back when first event fails */
1001         if (evsel != evlist__first(evlist))
1002                 return 0;
1003
1004         evlist__for_each_entry(evlist, counter)
1005                 counter->core.attr.write_backward = false;
1006         opts->overwrite = false;
1007         pr_debug2("fall back to non-overwrite mode\n");
1008         return 1;
1009 }
1010
1011 static int perf_top__start_counters(struct perf_top *top)
1012 {
1013         char msg[BUFSIZ];
1014         struct evsel *counter;
1015         struct evlist *evlist = top->evlist;
1016         struct record_opts *opts = &top->record_opts;
1017
1018         if (perf_top__overwrite_check(top)) {
1019                 ui__error("perf top only support consistent per-event "
1020                           "overwrite setting for all events\n");
1021                 goto out_err;
1022         }
1023
1024         perf_evlist__config(evlist, opts, &callchain_param);
1025
1026         evlist__for_each_entry(evlist, counter) {
1027 try_again:
1028                 if (evsel__open(counter, top->evlist->core.cpus,
1029                                      top->evlist->core.threads) < 0) {
1030
1031                         /*
1032                          * Specially handle overwrite fall back.
1033                          * Because perf top is the only tool which has
1034                          * overwrite mode by default, support
1035                          * both overwrite and non-overwrite mode, and
1036                          * require consistent mode for all events.
1037                          *
1038                          * May move it to generic code with more tools
1039                          * have similar attribute.
1040                          */
1041                         if (perf_missing_features.write_backward &&
1042                             perf_top_overwrite_fallback(top, counter))
1043                                 goto try_again;
1044
1045                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
1046                                 if (verbose > 0)
1047                                         ui__warning("%s\n", msg);
1048                                 goto try_again;
1049                         }
1050
1051                         perf_evsel__open_strerror(counter, &opts->target,
1052                                                   errno, msg, sizeof(msg));
1053                         ui__error("%s\n", msg);
1054                         goto out_err;
1055                 }
1056         }
1057
1058         if (evlist__mmap(evlist, opts->mmap_pages) < 0) {
1059                 ui__error("Failed to mmap with %d (%s)\n",
1060                             errno, str_error_r(errno, msg, sizeof(msg)));
1061                 goto out_err;
1062         }
1063
1064         return 0;
1065
1066 out_err:
1067         return -1;
1068 }
1069
1070 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1071 {
1072         if (callchain->mode != CHAIN_NONE) {
1073                 if (callchain_register_param(callchain) < 0) {
1074                         ui__error("Can't register callchain params.\n");
1075                         return -EINVAL;
1076                 }
1077         }
1078
1079         return 0;
1080 }
1081
1082 static struct ordered_events *rotate_queues(struct perf_top *top)
1083 {
1084         struct ordered_events *in = top->qe.in;
1085
1086         if (top->qe.in == &top->qe.data[1])
1087                 top->qe.in = &top->qe.data[0];
1088         else
1089                 top->qe.in = &top->qe.data[1];
1090
1091         return in;
1092 }
1093
1094 static void *process_thread(void *arg)
1095 {
1096         struct perf_top *top = arg;
1097
1098         while (!done) {
1099                 struct ordered_events *out, *in = top->qe.in;
1100
1101                 if (!in->nr_events) {
1102                         usleep(100);
1103                         continue;
1104                 }
1105
1106                 out = rotate_queues(top);
1107
1108                 pthread_mutex_lock(&top->qe.mutex);
1109                 top->qe.rotate = true;
1110                 pthread_cond_wait(&top->qe.cond, &top->qe.mutex);
1111                 pthread_mutex_unlock(&top->qe.mutex);
1112
1113                 if (ordered_events__flush(out, OE_FLUSH__TOP))
1114                         pr_err("failed to process events\n");
1115         }
1116
1117         return NULL;
1118 }
1119
1120 /*
1121  * Allow only 'top->delay_secs' seconds behind samples.
1122  */
1123 static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1124 {
1125         union perf_event *event = qevent->event;
1126         u64 delay_timestamp;
1127
1128         if (event->header.type != PERF_RECORD_SAMPLE)
1129                 return false;
1130
1131         delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1132         return delay_timestamp < last_timestamp;
1133 }
1134
1135 static int deliver_event(struct ordered_events *qe,
1136                          struct ordered_event *qevent)
1137 {
1138         struct perf_top *top = qe->data;
1139         struct evlist *evlist = top->evlist;
1140         struct perf_session *session = top->session;
1141         union perf_event *event = qevent->event;
1142         struct perf_sample sample;
1143         struct evsel *evsel;
1144         struct machine *machine;
1145         int ret = -1;
1146
1147         if (should_drop(qevent, top)) {
1148                 top->drop++;
1149                 top->drop_total++;
1150                 return 0;
1151         }
1152
1153         ret = perf_evlist__parse_sample(evlist, event, &sample);
1154         if (ret) {
1155                 pr_err("Can't parse sample, err = %d\n", ret);
1156                 goto next_event;
1157         }
1158
1159         evsel = perf_evlist__id2evsel(session->evlist, sample.id);
1160         assert(evsel != NULL);
1161
1162         if (event->header.type == PERF_RECORD_SAMPLE) {
1163                 if (evswitch__discard(&top->evswitch, evsel))
1164                         return 0;
1165                 ++top->samples;
1166         }
1167
1168         switch (sample.cpumode) {
1169         case PERF_RECORD_MISC_USER:
1170                 ++top->us_samples;
1171                 if (top->hide_user_symbols)
1172                         goto next_event;
1173                 machine = &session->machines.host;
1174                 break;
1175         case PERF_RECORD_MISC_KERNEL:
1176                 ++top->kernel_samples;
1177                 if (top->hide_kernel_symbols)
1178                         goto next_event;
1179                 machine = &session->machines.host;
1180                 break;
1181         case PERF_RECORD_MISC_GUEST_KERNEL:
1182                 ++top->guest_kernel_samples;
1183                 machine = perf_session__find_machine(session,
1184                                                      sample.pid);
1185                 break;
1186         case PERF_RECORD_MISC_GUEST_USER:
1187                 ++top->guest_us_samples;
1188                 /*
1189                  * TODO: we don't process guest user from host side
1190                  * except simple counting.
1191                  */
1192                 goto next_event;
1193         default:
1194                 if (event->header.type == PERF_RECORD_SAMPLE)
1195                         goto next_event;
1196                 machine = &session->machines.host;
1197                 break;
1198         }
1199
1200         if (event->header.type == PERF_RECORD_SAMPLE) {
1201                 perf_event__process_sample(&top->tool, event, evsel,
1202                                            &sample, machine);
1203         } else if (event->header.type == PERF_RECORD_LOST) {
1204                 perf_top__process_lost(top, event, evsel);
1205         } else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1206                 perf_top__process_lost_samples(top, event, evsel);
1207         } else if (event->header.type < PERF_RECORD_MAX) {
1208                 hists__inc_nr_events(evsel__hists(evsel), event->header.type);
1209                 machine__process_event(machine, event, &sample);
1210         } else
1211                 ++session->evlist->stats.nr_unknown_events;
1212
1213         ret = 0;
1214 next_event:
1215         return ret;
1216 }
1217
1218 static void init_process_thread(struct perf_top *top)
1219 {
1220         ordered_events__init(&top->qe.data[0], deliver_event, top);
1221         ordered_events__init(&top->qe.data[1], deliver_event, top);
1222         ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1223         ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1224         top->qe.in = &top->qe.data[0];
1225         pthread_mutex_init(&top->qe.mutex, NULL);
1226         pthread_cond_init(&top->qe.cond, NULL);
1227 }
1228
1229 static int __cmd_top(struct perf_top *top)
1230 {
1231         struct record_opts *opts = &top->record_opts;
1232         pthread_t thread, thread_process;
1233         int ret;
1234
1235         if (!top->annotation_opts.objdump_path) {
1236                 ret = perf_env__lookup_objdump(&top->session->header.env,
1237                                                &top->annotation_opts.objdump_path);
1238                 if (ret)
1239                         return ret;
1240         }
1241
1242         ret = callchain_param__setup_sample_type(&callchain_param);
1243         if (ret)
1244                 return ret;
1245
1246         if (perf_session__register_idle_thread(top->session) < 0)
1247                 return ret;
1248
1249         if (top->nr_threads_synthesize > 1)
1250                 perf_set_multithreaded();
1251
1252         init_process_thread(top);
1253
1254         if (opts->record_namespaces)
1255                 top->tool.namespace_events = true;
1256         if (opts->record_cgroup) {
1257 #ifdef HAVE_FILE_HANDLE
1258                 top->tool.cgroup_events = true;
1259 #else
1260                 pr_err("cgroup tracking is not supported.\n");
1261                 return -1;
1262 #endif
1263         }
1264
1265         ret = perf_event__synthesize_bpf_events(top->session, perf_event__process,
1266                                                 &top->session->machines.host,
1267                                                 &top->record_opts);
1268         if (ret < 0)
1269                 pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n");
1270
1271         ret = perf_event__synthesize_cgroups(&top->tool, perf_event__process,
1272                                              &top->session->machines.host);
1273         if (ret < 0)
1274                 pr_debug("Couldn't synthesize cgroup events.\n");
1275
1276         machine__synthesize_threads(&top->session->machines.host, &opts->target,
1277                                     top->evlist->core.threads, false,
1278                                     top->nr_threads_synthesize);
1279
1280         if (top->nr_threads_synthesize > 1)
1281                 perf_set_singlethreaded();
1282
1283         if (perf_hpp_list.socket) {
1284                 ret = perf_env__read_cpu_topology_map(&perf_env);
1285                 if (ret < 0) {
1286                         char errbuf[BUFSIZ];
1287                         const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1288
1289                         ui__error("Could not read the CPU topology map: %s\n", err);
1290                         return ret;
1291                 }
1292         }
1293
1294         ret = perf_top__start_counters(top);
1295         if (ret)
1296                 return ret;
1297
1298         top->session->evlist = top->evlist;
1299         perf_session__set_id_hdr_size(top->session);
1300
1301         /*
1302          * When perf is starting the traced process, all the events (apart from
1303          * group members) have enable_on_exec=1 set, so don't spoil it by
1304          * prematurely enabling them.
1305          *
1306          * XXX 'top' still doesn't start workloads like record, trace, but should,
1307          * so leave the check here.
1308          */
1309         if (!target__none(&opts->target))
1310                 evlist__enable(top->evlist);
1311
1312         ret = -1;
1313         if (pthread_create(&thread_process, NULL, process_thread, top)) {
1314                 ui__error("Could not create process thread.\n");
1315                 return ret;
1316         }
1317
1318         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1319                                                             display_thread), top)) {
1320                 ui__error("Could not create display thread.\n");
1321                 goto out_join_thread;
1322         }
1323
1324         if (top->realtime_prio) {
1325                 struct sched_param param;
1326
1327                 param.sched_priority = top->realtime_prio;
1328                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1329                         ui__error("Could not set realtime priority.\n");
1330                         goto out_join;
1331                 }
1332         }
1333
1334         /* Wait for a minimal set of events before starting the snapshot */
1335         evlist__poll(top->evlist, 100);
1336
1337         perf_top__mmap_read(top);
1338
1339         while (!done) {
1340                 u64 hits = top->samples;
1341
1342                 perf_top__mmap_read(top);
1343
1344                 if (opts->overwrite || (hits == top->samples))
1345                         ret = evlist__poll(top->evlist, 100);
1346
1347                 if (resize) {
1348                         perf_top__resize(top);
1349                         resize = 0;
1350                 }
1351         }
1352
1353         ret = 0;
1354 out_join:
1355         pthread_join(thread, NULL);
1356 out_join_thread:
1357         pthread_cond_signal(&top->qe.cond);
1358         pthread_join(thread_process, NULL);
1359         return ret;
1360 }
1361
1362 static int
1363 callchain_opt(const struct option *opt, const char *arg, int unset)
1364 {
1365         symbol_conf.use_callchain = true;
1366         return record_callchain_opt(opt, arg, unset);
1367 }
1368
1369 static int
1370 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1371 {
1372         struct callchain_param *callchain = opt->value;
1373
1374         callchain->enabled = !unset;
1375         callchain->record_mode = CALLCHAIN_FP;
1376
1377         /*
1378          * --no-call-graph
1379          */
1380         if (unset) {
1381                 symbol_conf.use_callchain = false;
1382                 callchain->record_mode = CALLCHAIN_NONE;
1383                 return 0;
1384         }
1385
1386         return parse_callchain_top_opt(arg);
1387 }
1388
1389 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1390 {
1391         if (!strcmp(var, "top.call-graph")) {
1392                 var = "call-graph.record-mode";
1393                 return perf_default_config(var, value, cb);
1394         }
1395         if (!strcmp(var, "top.children")) {
1396                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1397                 return 0;
1398         }
1399
1400         return 0;
1401 }
1402
1403 static int
1404 parse_percent_limit(const struct option *opt, const char *arg,
1405                     int unset __maybe_unused)
1406 {
1407         struct perf_top *top = opt->value;
1408
1409         top->min_percent = strtof(arg, NULL);
1410         return 0;
1411 }
1412
1413 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1414         "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1415
1416 int cmd_top(int argc, const char **argv)
1417 {
1418         char errbuf[BUFSIZ];
1419         struct perf_top top = {
1420                 .count_filter        = 5,
1421                 .delay_secs          = 2,
1422                 .record_opts = {
1423                         .mmap_pages     = UINT_MAX,
1424                         .user_freq      = UINT_MAX,
1425                         .user_interval  = ULLONG_MAX,
1426                         .freq           = 4000, /* 4 KHz */
1427                         .target         = {
1428                                 .uses_mmap   = true,
1429                         },
1430                         /*
1431                          * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1432                          * when we pause, fix that and reenable. Probably using a
1433                          * separate evlist with a dummy event, i.e. a non-overwrite
1434                          * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1435                          * stays in overwrite mode. -acme
1436                          * */
1437                         .overwrite      = 0,
1438                         .sample_time    = true,
1439                         .sample_time_set = true,
1440                 },
1441                 .max_stack           = sysctl__max_stack(),
1442                 .annotation_opts     = annotation__default_options,
1443                 .nr_threads_synthesize = UINT_MAX,
1444         };
1445         struct record_opts *opts = &top.record_opts;
1446         struct target *target = &opts->target;
1447         const struct option options[] = {
1448         OPT_CALLBACK('e', "event", &top.evlist, "event",
1449                      "event selector. use 'perf list' to list available events",
1450                      parse_events_option),
1451         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1452         OPT_STRING('p', "pid", &target->pid, "pid",
1453                     "profile events on existing process id"),
1454         OPT_STRING('t', "tid", &target->tid, "tid",
1455                     "profile events on existing thread id"),
1456         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1457                             "system-wide collection from all CPUs"),
1458         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1459                     "list of cpus to monitor"),
1460         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1461                    "file", "vmlinux pathname"),
1462         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1463                     "don't load vmlinux even if found"),
1464         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1465                    "file", "kallsyms pathname"),
1466         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1467                     "hide kernel symbols"),
1468         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1469                      "number of mmap data pages",
1470                      perf_evlist__parse_mmap_pages),
1471         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1472                     "collect data with this RT SCHED_FIFO priority"),
1473         OPT_INTEGER('d', "delay", &top.delay_secs,
1474                     "number of seconds to delay between refreshes"),
1475         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1476                             "dump the symbol table used for profiling"),
1477         OPT_INTEGER('f', "count-filter", &top.count_filter,
1478                     "only display functions with more events than this"),
1479         OPT_BOOLEAN(0, "group", &opts->group,
1480                             "put the counters into a counter group"),
1481         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1482                     "child tasks do not inherit counters"),
1483         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1484                     "symbol to annotate"),
1485         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1486         OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1487                      "profile at this frequency",
1488                       record__parse_freq),
1489         OPT_INTEGER('E', "entries", &top.print_entries,
1490                     "display this many functions"),
1491         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1492                     "hide user symbols"),
1493         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1494         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1495         OPT_INCR('v', "verbose", &verbose,
1496                     "be more verbose (show counter open errors, etc)"),
1497         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1498                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1499                    " Please refer the man page for the complete list."),
1500         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1501                    "output field(s): overhead, period, sample plus all of sort keys"),
1502         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1503                     "Show a column with the number of samples"),
1504         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1505                            NULL, "enables call-graph recording and display",
1506                            &callchain_opt),
1507         OPT_CALLBACK(0, "call-graph", &callchain_param,
1508                      "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1509                      top_callchain_help, &parse_callchain_opt),
1510         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1511                     "Accumulate callchains of children and show total overhead as well"),
1512         OPT_INTEGER(0, "max-stack", &top.max_stack,
1513                     "Set the maximum stack depth when parsing the callchain. "
1514                     "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1515         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1516                    "ignore callees of these functions in call graphs",
1517                    report_parse_ignore_callees_opt),
1518         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1519                     "Show a column with the sum of periods"),
1520         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1521                    "only consider symbols in these dsos"),
1522         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1523                    "only consider symbols in these comms"),
1524         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1525                    "only consider these symbols"),
1526         OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1527                     "Interleave source code with assembly code (default)"),
1528         OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1529                     "Display raw encoding of assembly instructions (default)"),
1530         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1531                     "Enable kernel symbol demangling"),
1532         OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"),
1533         OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path",
1534                     "objdump binary to use for disassembly and annotations"),
1535         OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1536                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1537         OPT_STRING(0, "prefix", &top.annotation_opts.prefix, "prefix",
1538                     "Add prefix to source file path names in programs (with --prefix-strip)"),
1539         OPT_STRING(0, "prefix-strip", &top.annotation_opts.prefix_strip, "N",
1540                     "Strip first N entries of source file path name in programs (with --prefix)"),
1541         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1542         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1543                      "Don't show entries under that percent", parse_percent_limit),
1544         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1545                      "How to display percentage of filtered entries", parse_filter_percentage),
1546         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1547                    "width[,width...]",
1548                    "don't try to adjust column width, use these fixed values"),
1549         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1550                         "per thread proc mmap processing timeout in ms"),
1551         OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1552                      "branch any", "sample any taken branches",
1553                      parse_branch_stack),
1554         OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1555                      "branch filter mask", "branch stack filter modes",
1556                      parse_branch_stack),
1557         OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1558                     "Show raw trace event output (do not use print fmt or plugins)"),
1559         OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1560                     "Show entries in a hierarchy"),
1561         OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1562                     "Use a backward ring buffer, default: no"),
1563         OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1564         OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1565                         "number of thread to run event synthesize"),
1566         OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces,
1567                     "Record namespaces events"),
1568         OPT_BOOLEAN(0, "all-cgroups", &opts->record_cgroup,
1569                     "Record cgroup events"),
1570         OPT_INTEGER(0, "group-sort-idx", &symbol_conf.group_sort_idx,
1571                     "Sort the output by the event at the index n in group. "
1572                     "If n is invalid, sort by the first event. "
1573                     "WARNING: should be used on grouped events."),
1574         OPTS_EVSWITCH(&top.evswitch),
1575         OPT_END()
1576         };
1577         struct evlist *sb_evlist = NULL;
1578         const char * const top_usage[] = {
1579                 "perf top [<options>]",
1580                 NULL
1581         };
1582         int status = hists__init();
1583
1584         if (status < 0)
1585                 return status;
1586
1587         top.annotation_opts.min_pcnt = 5;
1588         top.annotation_opts.context  = 4;
1589
1590         top.evlist = evlist__new();
1591         if (top.evlist == NULL)
1592                 return -ENOMEM;
1593
1594         status = perf_config(perf_top_config, &top);
1595         if (status)
1596                 return status;
1597         /*
1598          * Since the per arch annotation init routine may need the cpuid, read
1599          * it here, since we are not getting this from the perf.data header.
1600          */
1601         status = perf_env__read_cpuid(&perf_env);
1602         if (status) {
1603                 /*
1604                  * Some arches do not provide a get_cpuid(), so just use pr_debug, otherwise
1605                  * warn the user explicitely.
1606                  */
1607                 eprintf(status == ENOSYS ? 1 : 0, verbose,
1608                         "Couldn't read the cpuid for this machine: %s\n",
1609                         str_error_r(errno, errbuf, sizeof(errbuf)));
1610         }
1611         top.evlist->env = &perf_env;
1612
1613         argc = parse_options(argc, argv, options, top_usage, 0);
1614         if (argc)
1615                 usage_with_options(top_usage, options);
1616
1617         if (annotate_check_args(&top.annotation_opts) < 0)
1618                 goto out_delete_evlist;
1619
1620         if (!top.evlist->core.nr_entries &&
1621             perf_evlist__add_default(top.evlist) < 0) {
1622                 pr_err("Not enough memory for event selector list\n");
1623                 goto out_delete_evlist;
1624         }
1625
1626         status = evswitch__init(&top.evswitch, top.evlist, stderr);
1627         if (status)
1628                 goto out_delete_evlist;
1629
1630         if (symbol_conf.report_hierarchy) {
1631                 /* disable incompatible options */
1632                 symbol_conf.event_group = false;
1633                 symbol_conf.cumulate_callchain = false;
1634
1635                 if (field_order) {
1636                         pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1637                         parse_options_usage(top_usage, options, "fields", 0);
1638                         parse_options_usage(NULL, options, "hierarchy", 0);
1639                         goto out_delete_evlist;
1640                 }
1641         }
1642
1643         if (opts->branch_stack && callchain_param.enabled)
1644                 symbol_conf.show_branchflag_count = true;
1645
1646         sort__mode = SORT_MODE__TOP;
1647         /* display thread wants entries to be collapsed in a different tree */
1648         perf_hpp_list.need_collapse = 1;
1649
1650         if (top.use_stdio)
1651                 use_browser = 0;
1652         else if (top.use_tui)
1653                 use_browser = 1;
1654
1655         setup_browser(false);
1656
1657         if (setup_sorting(top.evlist) < 0) {
1658                 if (sort_order)
1659                         parse_options_usage(top_usage, options, "s", 1);
1660                 if (field_order)
1661                         parse_options_usage(sort_order ? NULL : top_usage,
1662                                             options, "fields", 0);
1663                 goto out_delete_evlist;
1664         }
1665
1666         status = target__validate(target);
1667         if (status) {
1668                 target__strerror(target, status, errbuf, BUFSIZ);
1669                 ui__warning("%s\n", errbuf);
1670         }
1671
1672         status = target__parse_uid(target);
1673         if (status) {
1674                 int saved_errno = errno;
1675
1676                 target__strerror(target, status, errbuf, BUFSIZ);
1677                 ui__error("%s\n", errbuf);
1678
1679                 status = -saved_errno;
1680                 goto out_delete_evlist;
1681         }
1682
1683         if (target__none(target))
1684                 target->system_wide = true;
1685
1686         if (perf_evlist__create_maps(top.evlist, target) < 0) {
1687                 ui__error("Couldn't create thread/CPU maps: %s\n",
1688                           errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1689                 goto out_delete_evlist;
1690         }
1691
1692         if (top.delay_secs < 1)
1693                 top.delay_secs = 1;
1694
1695         if (record_opts__config(opts)) {
1696                 status = -EINVAL;
1697                 goto out_delete_evlist;
1698         }
1699
1700         top.sym_evsel = evlist__first(top.evlist);
1701
1702         if (!callchain_param.enabled) {
1703                 symbol_conf.cumulate_callchain = false;
1704                 perf_hpp__cancel_cumulate();
1705         }
1706
1707         if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1708                 callchain_param.order = ORDER_CALLER;
1709
1710         status = symbol__annotation_init();
1711         if (status < 0)
1712                 goto out_delete_evlist;
1713
1714         annotation_config__init(&top.annotation_opts);
1715
1716         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1717         status = symbol__init(NULL);
1718         if (status < 0)
1719                 goto out_delete_evlist;
1720
1721         sort__setup_elide(stdout);
1722
1723         get_term_dimensions(&top.winsize);
1724         if (top.print_entries == 0) {
1725                 perf_top__update_print_entries(&top);
1726                 signal(SIGWINCH, winch_sig);
1727         }
1728
1729         top.session = perf_session__new(NULL, false, NULL);
1730         if (IS_ERR(top.session)) {
1731                 status = PTR_ERR(top.session);
1732                 goto out_delete_evlist;
1733         }
1734
1735         if (!top.record_opts.no_bpf_event)
1736                 bpf_event__add_sb_event(&sb_evlist, &perf_env);
1737
1738         if (perf_evlist__start_sb_thread(sb_evlist, target)) {
1739                 pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1740                 opts->no_bpf_event = true;
1741         }
1742
1743         status = __cmd_top(&top);
1744
1745         if (!opts->no_bpf_event)
1746                 perf_evlist__stop_sb_thread(sb_evlist);
1747
1748 out_delete_evlist:
1749         evlist__delete(top.evlist);
1750         perf_session__delete(top.session);
1751
1752         return status;
1753 }