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/thread.h"
  37 #include "util/symbol.h"
  38 #include "util/synthetic-events.h"
  39 #include "util/top.h"
  40 #include "util/util.h"
  41 #include <linux/rbtree.h>
  42 #include <subcmd/parse-options.h>
  43 #include "util/parse-events.h"
  44 #include "util/callchain.h"
  45 #include "util/cpumap.h"
  46 #include "util/sort.h"
  47 #include "util/string2.h"
  48 #include "util/term.h"
  49 #include "util/intlist.h"
  50 #include "util/parse-branch-options.h"
  51 #include "arch/common.h"
  52 #include "ui/ui.h"
  53
  54 #include "util/debug.h"
  55 #include "util/ordered-events.h"
  56 #include "util/pfm.h"
  57
  58 #include <assert.h>
  59 #include <elf.h>
  60 #include <fcntl.h>
  61
  62 #include <stdio.h>
  63 #include <termios.h>
  64 #include <unistd.h>
  65 #include <inttypes.h>
  66
  67 #include <errno.h>
  68 #include <time.h>
  69 #include <sched.h>
  70 #include <signal.h>
  71
  72 #include <sys/syscall.h>
  73 #include <sys/ioctl.h>
  74 #include <poll.h>
  75 #include <sys/prctl.h>
  76 #include <sys/wait.h>
  77 #include <sys/uio.h>
  78 #include <sys/utsname.h>
  79 #include <sys/mman.h>
  80
  81 #include <linux/stringify.h>
  82 #include <linux/time64.h>
  83 #include <linux/types.h>
  84 #include <linux/err.h>
  85
  86 #include <linux/ctype.h>
  87 #include <perf/mmap.h>
  88
  89 static volatile int done;
  90 static volatile int resize;
  91
  92 #define HEADER_LINE_NR  5
  93
  94 static void perf_top__update_print_entries(struct perf_top *top)
  95 {
  96         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
  97 }
  98
  99 static void winch_sig(int sig __maybe_unused)
 100 {
 101         resize = 1;
 102 }
 103
 104 static void perf_top__resize(struct perf_top *top)
 105 {
 106         get_term_dimensions(&top->winsize);
 107         perf_top__update_print_entries(top);
 108 }
 109
 110 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
 111 {
 112         struct evsel *evsel;
 113         struct symbol *sym;
 114         struct annotation *notes;
 115         struct map *map;
 116         int err = -1;
 117
 118         if (!he || !he->ms.sym)
 119                 return -1;
 120
 121         evsel = hists_to_evsel(he->hists);
 122
 123         sym = he->ms.sym;
 124         map = he->ms.map;
 125
 126         /*
 127          * We can't annotate with just /proc/kallsyms
 128          */
 129         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 130             !dso__is_kcore(map->dso)) {
 131                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 132                        "path\n", sym->name);
 133                 sleep(1);
 134                 return -1;
 135         }
 136
 137         notes = symbol__annotation(sym);
 138         pthread_mutex_lock(&notes->lock);
 139
 140         if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
 141                 pthread_mutex_unlock(&notes->lock);
 142                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 143                        sym->name);
 144                 sleep(1);
 145                 return err;
 146         }
 147
 148         err = symbol__annotate(&he->ms, evsel, &top->annotation_opts, NULL);
 149         if (err == 0) {
 150                 top->sym_filter_entry = he;
 151         } else {
 152                 char msg[BUFSIZ];
 153                 symbol__strerror_disassemble(&he->ms, err, msg, sizeof(msg));
 154                 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
 155         }
 156
 157         pthread_mutex_unlock(&notes->lock);
 158         return err;
 159 }
 160
 161 static void __zero_source_counters(struct hist_entry *he)
 162 {
 163         struct symbol *sym = he->ms.sym;
 164         symbol__annotate_zero_histograms(sym);
 165 }
 166
 167 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 168 {
 169         struct utsname uts;
 170         int err = uname(&uts);
 171
 172         ui__warning("Out of bounds address found:\n\n"
 173                     "Addr:   %" PRIx64 "\n"
 174                     "DSO:    %s %c\n"
 175                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
 176                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 177                     "Arch:   %s\n"
 178                     "Kernel: %s\n"
 179                     "Tools:  %s\n\n"
 180                     "Not all samples will be on the annotation output.\n\n"
 181                     "Please report to linux-kernel@vger.kernel.org\n",
 182                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
 183                     map->start, map->end, sym->start, sym->end,
 184                     sym->binding == STB_GLOBAL ? 'g' :
 185                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 186                     err ? "[unknown]" : uts.machine,
 187                     err ? "[unknown]" : uts.release, perf_version_string);
 188         if (use_browser <= 0)
 189                 sleep(5);
 190
 191         map->erange_warned = true;
 192 }
 193
 194 static void perf_top__record_precise_ip(struct perf_top *top,
 195                                         struct hist_entry *he,
 196                                         struct perf_sample *sample,
 197                                         struct evsel *evsel, u64 ip)
 198 {
 199         struct annotation *notes;
 200         struct symbol *sym = he->ms.sym;
 201         int err = 0;
 202
 203         if (sym == NULL || (use_browser == 0 &&
 204                             (top->sym_filter_entry == NULL ||
 205                              top->sym_filter_entry->ms.sym != sym)))
 206                 return;
 207
 208         notes = symbol__annotation(sym);
 209
 210         if (pthread_mutex_trylock(&notes->lock))
 211                 return;
 212
 213         err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
 214
 215         pthread_mutex_unlock(&notes->lock);
 216
 217         if (unlikely(err)) {
 218                 /*
 219                  * This function is now called with he->hists->lock held.
 220                  * Release it before going to sleep.
 221                  */
 222                 pthread_mutex_unlock(&he->hists->lock);
 223
 224                 if (err == -ERANGE && !he->ms.map->erange_warned)
 225                         ui__warn_map_erange(he->ms.map, sym, ip);
 226                 else if (err == -ENOMEM) {
 227                         pr_err("Not enough memory for annotating '%s' symbol!\n",
 228                                sym->name);
 229                         sleep(1);
 230                 }
 231
 232                 pthread_mutex_lock(&he->hists->lock);
 233         }
 234 }
 235
 236 static void perf_top__show_details(struct perf_top *top)
 237 {
 238         struct hist_entry *he = top->sym_filter_entry;
 239         struct evsel *evsel;
 240         struct annotation *notes;
 241         struct symbol *symbol;
 242         int more;
 243
 244         if (!he)
 245                 return;
 246
 247         evsel = hists_to_evsel(he->hists);
 248
 249         symbol = he->ms.sym;
 250         notes = symbol__annotation(symbol);
 251
 252         pthread_mutex_lock(&notes->lock);
 253
 254         symbol__calc_percent(symbol, evsel);
 255
 256         if (notes->src == NULL)
 257                 goto out_unlock;
 258
 259         printf("Showing %s for %s\n", evsel__name(top->sym_evsel), symbol->name);
 260         printf("  Events  Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
 261
 262         more = symbol__annotate_printf(&he->ms, top->sym_evsel, &top->annotation_opts);
 263
 264         if (top->evlist->enabled) {
 265                 if (top->zero)
 266                         symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 267                 else
 268                         symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 269         }
 270         if (more != 0)
 271                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 272 out_unlock:
 273         pthread_mutex_unlock(&notes->lock);
 274 }
 275
 276 static void perf_top__resort_hists(struct perf_top *t)
 277 {
 278         struct evlist *evlist = t->evlist;
 279         struct evsel *pos;
 280
 281         evlist__for_each_entry(evlist, pos) {
 282                 struct hists *hists = evsel__hists(pos);
 283
 284                 /*
 285                  * unlink existing entries so that they can be linked
 286                  * in a correct order in hists__match() below.
 287                  */
 288                 hists__unlink(hists);
 289
 290                 if (evlist->enabled) {
 291                         if (t->zero) {
 292                                 hists__delete_entries(hists);
 293                         } else {
 294                                 hists__decay_entries(hists, t->hide_user_symbols,
 295                                                      t->hide_kernel_symbols);
 296                         }
 297                 }
 298
 299                 hists__collapse_resort(hists, NULL);
 300
 301                 /* Non-group events are considered as leader */
 302                 if (symbol_conf.event_group && !evsel__is_group_leader(pos)) {
 303                         struct hists *leader_hists = evsel__hists(pos->leader);
 304
 305                         hists__match(leader_hists, hists);
 306                         hists__link(leader_hists, hists);
 307                 }
 308         }
 309
 310         evlist__for_each_entry(evlist, pos) {
 311                 evsel__output_resort(pos, NULL);
 312         }
 313 }
 314
 315 static void perf_top__print_sym_table(struct perf_top *top)
 316 {
 317         char bf[160];
 318         int printed = 0;
 319         const int win_width = top->winsize.ws_col - 1;
 320         struct evsel *evsel = top->sym_evsel;
 321         struct hists *hists = evsel__hists(evsel);
 322
 323         puts(CONSOLE_CLEAR);
 324
 325         perf_top__header_snprintf(top, bf, sizeof(bf));
 326         printf("%s\n", bf);
 327
 328         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 329
 330         if (!top->record_opts.overwrite &&
 331             (hists->stats.nr_lost_warned !=
 332             hists->stats.nr_events[PERF_RECORD_LOST])) {
 333                 hists->stats.nr_lost_warned =
 334                               hists->stats.nr_events[PERF_RECORD_LOST];
 335                 color_fprintf(stdout, PERF_COLOR_RED,
 336                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 337                               hists->stats.nr_lost_warned);
 338                 ++printed;
 339         }
 340
 341         if (top->sym_filter_entry) {
 342                 perf_top__show_details(top);
 343                 return;
 344         }
 345
 346         perf_top__resort_hists(top);
 347
 348         hists__output_recalc_col_len(hists, top->print_entries - printed);
 349         putchar('\n');
 350         hists__fprintf(hists, false, top->print_entries - printed, win_width,
 351                        top->min_percent, stdout, !symbol_conf.use_callchain);
 352 }
 353
 354 static void prompt_integer(int *target, const char *msg)
 355 {
 356         char *buf = malloc(0), *p;
 357         size_t dummy = 0;
 358         int tmp;
 359
 360         fprintf(stdout, "\n%s: ", msg);
 361         if (getline(&buf, &dummy, stdin) < 0)
 362                 return;
 363
 364         p = strchr(buf, '\n');
 365         if (p)
 366                 *p = 0;
 367
 368         p = buf;
 369         while(*p) {
 370                 if (!isdigit(*p))
 371                         goto out_free;
 372                 p++;
 373         }
 374         tmp = strtoul(buf, NULL, 10);
 375         *target = tmp;
 376 out_free:
 377         free(buf);
 378 }
 379
 380 static void prompt_percent(int *target, const char *msg)
 381 {
 382         int tmp = 0;
 383
 384         prompt_integer(&tmp, msg);
 385         if (tmp >= 0 && tmp <= 100)
 386                 *target = tmp;
 387 }
 388
 389 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 390 {
 391         char *buf = malloc(0), *p;
 392         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 393         struct hists *hists = evsel__hists(top->sym_evsel);
 394         struct rb_node *next;
 395         size_t dummy = 0;
 396
 397         /* zero counters of active symbol */
 398         if (syme) {
 399                 __zero_source_counters(syme);
 400                 top->sym_filter_entry = NULL;
 401         }
 402
 403         fprintf(stdout, "\n%s: ", msg);
 404         if (getline(&buf, &dummy, stdin) < 0)
 405                 goto out_free;
 406
 407         p = strchr(buf, '\n');
 408         if (p)
 409                 *p = 0;
 410
 411         next = rb_first_cached(&hists->entries);
 412         while (next) {
 413                 n = rb_entry(next, struct hist_entry, rb_node);
 414                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 415                         found = n;
 416                         break;
 417                 }
 418                 next = rb_next(&n->rb_node);
 419         }
 420
 421         if (!found) {
 422                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 423                 sleep(1);
 424         } else
 425                 perf_top__parse_source(top, found);
 426
 427 out_free:
 428         free(buf);
 429 }
 430
 431 static void perf_top__print_mapped_keys(struct perf_top *top)
 432 {
 433         char *name = NULL;
 434
 435         if (top->sym_filter_entry) {
 436                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 437                 name = sym->name;
 438         }
 439
 440         fprintf(stdout, "\nMapped keys:\n");
 441         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 442         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 443
 444         if (top->evlist->core.nr_entries > 1)
 445                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", evsel__name(top->sym_evsel));
 446
 447         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 448
 449         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
 450         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 451         fprintf(stdout, "\t[S]     stop annotation.\n");
 452
 453         fprintf(stdout,
 454                 "\t[K]     hide kernel symbols.             \t(%s)\n",
 455                 top->hide_kernel_symbols ? "yes" : "no");
 456         fprintf(stdout,
 457                 "\t[U]     hide user symbols.               \t(%s)\n",
 458                 top->hide_user_symbols ? "yes" : "no");
 459         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 460         fprintf(stdout, "\t[qQ]    quit.\n");
 461 }
 462
 463 static int perf_top__key_mapped(struct perf_top *top, int c)
 464 {
 465         switch (c) {
 466                 case 'd':
 467                 case 'e':
 468                 case 'f':
 469                 case 'z':
 470                 case 'q':
 471                 case 'Q':
 472                 case 'K':
 473                 case 'U':
 474                 case 'F':
 475                 case 's':
 476                 case 'S':
 477                         return 1;
 478                 case 'E':
 479                         return top->evlist->core.nr_entries > 1 ? 1 : 0;
 480                 default:
 481                         break;
 482         }
 483
 484         return 0;
 485 }
 486
 487 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 488 {
 489         bool ret = true;
 490
 491         if (!perf_top__key_mapped(top, c)) {
 492                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 493                 struct termios save;
 494
 495                 perf_top__print_mapped_keys(top);
 496                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 497                 fflush(stdout);
 498
 499                 set_term_quiet_input(&save);
 500
 501                 poll(&stdin_poll, 1, -1);
 502                 c = getc(stdin);
 503
 504                 tcsetattr(0, TCSAFLUSH, &save);
 505                 if (!perf_top__key_mapped(top, c))
 506                         return ret;
 507         }
 508
 509         switch (c) {
 510                 case 'd':
 511                         prompt_integer(&top->delay_secs, "Enter display delay");
 512                         if (top->delay_secs < 1)
 513                                 top->delay_secs = 1;
 514                         break;
 515                 case 'e':
 516                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 517                         if (top->print_entries == 0) {
 518                                 perf_top__resize(top);
 519                                 signal(SIGWINCH, winch_sig);
 520                         } else {
 521                                 signal(SIGWINCH, SIG_DFL);
 522                         }
 523                         break;
 524                 case 'E':
 525                         if (top->evlist->core.nr_entries > 1) {
 526                                 /* Select 0 as the default event: */
 527                                 int counter = 0;
 528
 529                                 fprintf(stderr, "\nAvailable events:");
 530
 531                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
 532                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, evsel__name(top->sym_evsel));
 533
 534                                 prompt_integer(&counter, "Enter details event counter");
 535
 536                                 if (counter >= top->evlist->core.nr_entries) {
 537                                         top->sym_evsel = evlist__first(top->evlist);
 538                                         fprintf(stderr, "Sorry, no such event, using %s.\n", evsel__name(top->sym_evsel));
 539                                         sleep(1);
 540                                         break;
 541                                 }
 542                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
 543                                         if (top->sym_evsel->idx == counter)
 544                                                 break;
 545                         } else
 546                                 top->sym_evsel = evlist__first(top->evlist);
 547                         break;
 548                 case 'f':
 549                         prompt_integer(&top->count_filter, "Enter display event count filter");
 550                         break;
 551                 case 'F':
 552                         prompt_percent(&top->annotation_opts.min_pcnt,
 553                                        "Enter details display event filter (percent)");
 554                         break;
 555                 case 'K':
 556                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 557                         break;
 558                 case 'q':
 559                 case 'Q':
 560                         printf("exiting.\n");
 561                         if (top->dump_symtab)
 562                                 perf_session__fprintf_dsos(top->session, stderr);
 563                         ret = false;
 564                         break;
 565                 case 's':
 566                         perf_top__prompt_symbol(top, "Enter details symbol");
 567                         break;
 568                 case 'S':
 569                         if (!top->sym_filter_entry)
 570                                 break;
 571                         else {
 572                                 struct hist_entry *syme = top->sym_filter_entry;
 573
 574                                 top->sym_filter_entry = NULL;
 575                                 __zero_source_counters(syme);
 576                         }
 577                         break;
 578                 case 'U':
 579                         top->hide_user_symbols = !top->hide_user_symbols;
 580                         break;
 581                 case 'z':
 582                         top->zero = !top->zero;
 583                         break;
 584                 default:
 585                         break;
 586         }
 587
 588         return ret;
 589 }
 590
 591 static void perf_top__sort_new_samples(void *arg)
 592 {
 593         struct perf_top *t = arg;
 594
 595         if (t->evlist->selected != NULL)
 596                 t->sym_evsel = t->evlist->selected;
 597
 598         perf_top__resort_hists(t);
 599
 600         if (t->lost || t->drop)
 601                 pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
 602 }
 603
 604 static void stop_top(void)
 605 {
 606         session_done = 1;
 607         done = 1;
 608 }
 609
 610 static void *display_thread_tui(void *arg)
 611 {
 612         struct evsel *pos;
 613         struct perf_top *top = arg;
 614         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 615         struct hist_browser_timer hbt = {
 616                 .timer          = perf_top__sort_new_samples,
 617                 .arg            = top,
 618                 .refresh        = top->delay_secs,
 619         };
 620         int ret;
 621
 622         /* In order to read symbols from other namespaces perf to  needs to call
 623          * setns(2).  This isn't permitted if the struct_fs has multiple users.
 624          * unshare(2) the fs so that we may continue to setns into namespaces
 625          * that we're observing.
 626          */
 627         unshare(CLONE_FS);
 628
 629         prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
 630
 631 repeat:
 632         perf_top__sort_new_samples(top);
 633
 634         /*
 635          * Initialize the uid_filter_str, in the future the TUI will allow
 636          * Zooming in/out UIDs. For now just use whatever the user passed
 637          * via --uid.
 638          */
 639         evlist__for_each_entry(top->evlist, pos) {
 640                 struct hists *hists = evsel__hists(pos);
 641                 hists->uid_filter_str = top->record_opts.target.uid_str;
 642         }
 643
 644         ret = evlist__tui_browse_hists(top->evlist, help, &hbt, top->min_percent,
 645                                        &top->session->header.env, !top->record_opts.overwrite,
 646                                        &top->annotation_opts);
 647         if (ret == K_RELOAD) {
 648                 top->zero = true;
 649                 goto repeat;
 650         } else
 651                 stop_top();
 652
 653         return NULL;
 654 }
 655
 656 static void display_sig(int sig __maybe_unused)
 657 {
 658         stop_top();
 659 }
 660
 661 static void display_setup_sig(void)
 662 {
 663         signal(SIGSEGV, sighandler_dump_stack);
 664         signal(SIGFPE, sighandler_dump_stack);
 665         signal(SIGINT,  display_sig);
 666         signal(SIGQUIT, display_sig);
 667         signal(SIGTERM, display_sig);
 668 }
 669
 670 static void *display_thread(void *arg)
 671 {
 672         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 673         struct termios save;
 674         struct perf_top *top = arg;
 675         int delay_msecs, c;
 676
 677         /* In order to read symbols from other namespaces perf to  needs to call
 678          * setns(2).  This isn't permitted if the struct_fs has multiple users.
 679          * unshare(2) the fs so that we may continue to setns into namespaces
 680          * that we're observing.
 681          */
 682         unshare(CLONE_FS);
 683
 684         prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
 685
 686         display_setup_sig();
 687         pthread__unblock_sigwinch();
 688 repeat:
 689         delay_msecs = top->delay_secs * MSEC_PER_SEC;
 690         set_term_quiet_input(&save);
 691         /* trash return*/
 692         clearerr(stdin);
 693         if (poll(&stdin_poll, 1, 0) > 0)
 694                 getc(stdin);
 695
 696         while (!done) {
 697                 perf_top__print_sym_table(top);
 698                 /*
 699                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 700                  * refresh screen in both cases.
 701                  */
 702                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 703                 case 0:
 704                         continue;
 705                 case -1:
 706                         if (errno == EINTR)
 707                                 continue;
 708                         __fallthrough;
 709                 default:
 710                         c = getc(stdin);
 711                         tcsetattr(0, TCSAFLUSH, &save);
 712
 713                         if (perf_top__handle_keypress(top, c))
 714                                 goto repeat;
 715                         stop_top();
 716                 }
 717         }
 718
 719         tcsetattr(0, TCSAFLUSH, &save);
 720         return NULL;
 721 }
 722
 723 static int hist_iter__top_callback(struct hist_entry_iter *iter,
 724                                    struct addr_location *al, bool single,
 725                                    void *arg)
 726 {
 727         struct perf_top *top = arg;
 728         struct hist_entry *he = iter->he;
 729         struct evsel *evsel = iter->evsel;
 730
 731         if (perf_hpp_list.sym && single)
 732                 perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
 733
 734         hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
 735                      !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY),
 736                      NULL);
 737         return 0;
 738 }
 739
 740 static void perf_event__process_sample(struct perf_tool *tool,
 741                                        const union perf_event *event,
 742                                        struct evsel *evsel,
 743                                        struct perf_sample *sample,
 744                                        struct machine *machine)
 745 {
 746         struct perf_top *top = container_of(tool, struct perf_top, tool);
 747         struct addr_location al;
 748         int err;
 749
 750         if (!machine && perf_guest) {
 751                 static struct intlist *seen;
 752
 753                 if (!seen)
 754                         seen = intlist__new(NULL);
 755
 756                 if (!intlist__has_entry(seen, sample->pid)) {
 757                         pr_err("Can't find guest [%d]'s kernel information\n",
 758                                 sample->pid);
 759                         intlist__add(seen, sample->pid);
 760                 }
 761                 return;
 762         }
 763
 764         if (!machine) {
 765                 pr_err("%u unprocessable samples recorded.\r",
 766                        top->session->evlist->stats.nr_unprocessable_samples++);
 767                 return;
 768         }
 769
 770         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 771                 top->exact_samples++;
 772
 773         if (machine__resolve(machine, &al, sample) < 0)
 774                 return;
 775
 776         if (top->stitch_lbr)
 777                 al.thread->lbr_stitch_enable = true;
 778
 779         if (!machine->kptr_restrict_warned &&
 780             symbol_conf.kptr_restrict &&
 781             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 782                 if (!perf_evlist__exclude_kernel(top->session->evlist)) {
 783                         ui__warning(
 784 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 785 "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
 786 "Kernel%s samples will not be resolved.\n",
 787                           al.map && map__has_symbols(al.map) ?
 788                           " modules" : "");
 789                         if (use_browser <= 0)
 790                                 sleep(5);
 791                 }
 792                 machine->kptr_restrict_warned = true;
 793         }
 794
 795         if (al.sym == NULL && al.map != NULL) {
 796                 const char *msg = "Kernel samples will not be resolved.\n";
 797                 /*
 798                  * As we do lazy loading of symtabs we only will know if the
 799                  * specified vmlinux file is invalid when we actually have a
 800                  * hit in kernel space and then try to load it. So if we get
 801                  * here and there are _no_ symbols in the DSO backing the
 802                  * kernel map, bail out.
 803                  *
 804                  * We may never get here, for instance, if we use -K/
 805                  * --hide-kernel-symbols, even if the user specifies an
 806                  * invalid --vmlinux ;-)
 807                  */
 808                 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
 809                     __map__is_kernel(al.map) && map__has_symbols(al.map)) {
 810                         if (symbol_conf.vmlinux_name) {
 811                                 char serr[256];
 812                                 dso__strerror_load(al.map->dso, serr, sizeof(serr));
 813                                 ui__warning("The %s file can't be used: %s\n%s",
 814                                             symbol_conf.vmlinux_name, serr, msg);
 815                         } else {
 816                                 ui__warning("A vmlinux file was not found.\n%s",
 817                                             msg);
 818                         }
 819
 820                         if (use_browser <= 0)
 821                                 sleep(5);
 822                         top->vmlinux_warned = true;
 823                 }
 824         }
 825
 826         if (al.sym == NULL || !al.sym->idle) {
 827                 struct hists *hists = evsel__hists(evsel);
 828                 struct hist_entry_iter iter = {
 829                         .evsel          = evsel,
 830                         .sample         = sample,
 831                         .add_entry_cb   = hist_iter__top_callback,
 832                 };
 833
 834                 if (symbol_conf.cumulate_callchain)
 835                         iter.ops = &hist_iter_cumulative;
 836                 else
 837                         iter.ops = &hist_iter_normal;
 838
 839                 pthread_mutex_lock(&hists->lock);
 840
 841                 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
 842                 if (err < 0)
 843                         pr_err("Problem incrementing symbol period, skipping event\n");
 844
 845                 pthread_mutex_unlock(&hists->lock);
 846         }
 847
 848         addr_location__put(&al);
 849 }
 850
 851 static void
 852 perf_top__process_lost(struct perf_top *top, union perf_event *event,
 853                        struct evsel *evsel)
 854 {
 855         struct hists *hists = evsel__hists(evsel);
 856
 857         top->lost += event->lost.lost;
 858         top->lost_total += event->lost.lost;
 859         hists->stats.total_lost += event->lost.lost;
 860 }
 861
 862 static void
 863 perf_top__process_lost_samples(struct perf_top *top,
 864                                union perf_event *event,
 865                                struct evsel *evsel)
 866 {
 867         struct hists *hists = evsel__hists(evsel);
 868
 869         top->lost += event->lost_samples.lost;
 870         top->lost_total += event->lost_samples.lost;
 871         hists->stats.total_lost_samples += event->lost_samples.lost;
 872 }
 873
 874 static u64 last_timestamp;
 875
 876 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 877 {
 878         struct record_opts *opts = &top->record_opts;
 879         struct evlist *evlist = top->evlist;
 880         struct mmap *md;
 881         union perf_event *event;
 882
 883         md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
 884         if (perf_mmap__read_init(&md->core) < 0)
 885                 return;
 886
 887         while ((event = perf_mmap__read_event(&md->core)) != NULL) {
 888                 int ret;
 889
 890                 ret = evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
 891                 if (ret && ret != -1)
 892                         break;
 893
 894                 ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0);
 895                 if (ret)
 896                         break;
 897
 898                 perf_mmap__consume(&md->core);
 899
 900                 if (top->qe.rotate) {
 901                         pthread_mutex_lock(&top->qe.mutex);
 902                         top->qe.rotate = false;
 903                         pthread_cond_signal(&top->qe.cond);
 904                         pthread_mutex_unlock(&top->qe.mutex);
 905                 }
 906         }
 907
 908         perf_mmap__read_done(&md->core);
 909 }
 910
 911 static void perf_top__mmap_read(struct perf_top *top)
 912 {
 913         bool overwrite = top->record_opts.overwrite;
 914         struct evlist *evlist = top->evlist;
 915         int i;
 916
 917         if (overwrite)
 918                 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
 919
 920         for (i = 0; i < top->evlist->core.nr_mmaps; i++)
 921                 perf_top__mmap_read_idx(top, i);
 922
 923         if (overwrite) {
 924                 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
 925                 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 926         }
 927 }
 928
 929 /*
 930  * Check per-event overwrite term.
 931  * perf top should support consistent term for all events.
 932  * - All events don't have per-event term
 933  *   E.g. "cpu/cpu-cycles/,cpu/instructions/"
 934  *   Nothing change, return 0.
 935  * - All events have same per-event term
 936  *   E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
 937  *   Using the per-event setting to replace the opts->overwrite if
 938  *   they are different, then return 0.
 939  * - Events have different per-event term
 940  *   E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
 941  *   Return -1
 942  * - Some of the event set per-event term, but some not.
 943  *   E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
 944  *   Return -1
 945  */
 946 static int perf_top__overwrite_check(struct perf_top *top)
 947 {
 948         struct record_opts *opts = &top->record_opts;
 949         struct evlist *evlist = top->evlist;
 950         struct evsel_config_term *term;
 951         struct list_head *config_terms;
 952         struct evsel *evsel;
 953         int set, overwrite = -1;
 954
 955         evlist__for_each_entry(evlist, evsel) {
 956                 set = -1;
 957                 config_terms = &evsel->config_terms;
 958                 list_for_each_entry(term, config_terms, list) {
 959                         if (term->type == EVSEL__CONFIG_TERM_OVERWRITE)
 960                                 set = term->val.overwrite ? 1 : 0;
 961                 }
 962
 963                 /* no term for current and previous event (likely) */
 964                 if ((overwrite < 0) && (set < 0))
 965                         continue;
 966
 967                 /* has term for both current and previous event, compare */
 968                 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
 969                         return -1;
 970
 971                 /* no term for current event but has term for previous one */
 972                 if ((overwrite >= 0) && (set < 0))
 973                         return -1;
 974
 975                 /* has term for current event */
 976                 if ((overwrite < 0) && (set >= 0)) {
 977                         /* if it's first event, set overwrite */
 978                         if (evsel == evlist__first(evlist))
 979                                 overwrite = set;
 980                         else
 981                                 return -1;
 982                 }
 983         }
 984
 985         if ((overwrite >= 0) && (opts->overwrite != overwrite))
 986                 opts->overwrite = overwrite;
 987
 988         return 0;
 989 }
 990
 991 static int perf_top_overwrite_fallback(struct perf_top *top,
 992                                        struct evsel *evsel)
 993 {
 994         struct record_opts *opts = &top->record_opts;
 995         struct evlist *evlist = top->evlist;
 996         struct evsel *counter;
 997
 998         if (!opts->overwrite)
 999                 return 0;
1000
1001         /* only fall back when first event fails */
1002         if (evsel != evlist__first(evlist))
1003                 return 0;
1004
1005         evlist__for_each_entry(evlist, counter)
1006                 counter->core.attr.write_backward = false;
1007         opts->overwrite = false;
1008         pr_debug2("fall back to non-overwrite mode\n");
1009         return 1;
1010 }
1011
1012 static int perf_top__start_counters(struct perf_top *top)
1013 {
1014         char msg[BUFSIZ];
1015         struct evsel *counter;
1016         struct evlist *evlist = top->evlist;
1017         struct record_opts *opts = &top->record_opts;
1018
1019         if (perf_top__overwrite_check(top)) {
1020                 ui__error("perf top only support consistent per-event "
1021                           "overwrite setting for all events\n");
1022                 goto out_err;
1023         }
1024
1025         perf_evlist__config(evlist, opts, &callchain_param);
1026
1027         evlist__for_each_entry(evlist, counter) {
1028 try_again:
1029                 if (evsel__open(counter, top->evlist->core.cpus,
1030                                      top->evlist->core.threads) < 0) {
1031
1032                         /*
1033                          * Specially handle overwrite fall back.
1034                          * Because perf top is the only tool which has
1035                          * overwrite mode by default, support
1036                          * both overwrite and non-overwrite mode, and
1037                          * require consistent mode for all events.
1038                          *
1039                          * May move it to generic code with more tools
1040                          * have similar attribute.
1041                          */
1042                         if (perf_missing_features.write_backward &&
1043                             perf_top_overwrite_fallback(top, counter))
1044                                 goto try_again;
1045
1046                         if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
1047                                 if (verbose > 0)
1048                                         ui__warning("%s\n", msg);
1049                                 goto try_again;
1050                         }
1051
1052                         evsel__open_strerror(counter, &opts->target, 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 = 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 = 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         OPT_BOOLEAN(0, "stitch-lbr", &top.stitch_lbr,
1575                     "Enable LBR callgraph stitching approach"),
1576 #ifdef HAVE_LIBPFM
1577         OPT_CALLBACK(0, "pfm-events", &top.evlist, "event",
1578                 "libpfm4 event selector. use 'perf list' to list available events",
1579                 parse_libpfm_events_option),
1580 #endif
1581         OPTS_EVSWITCH(&top.evswitch),
1582         OPT_END()
1583         };
1584         const char * const top_usage[] = {
1585                 "perf top [<options>]",
1586                 NULL
1587         };
1588         int status = hists__init();
1589
1590         if (status < 0)
1591                 return status;
1592
1593         top.annotation_opts.min_pcnt = 5;
1594         top.annotation_opts.context  = 4;
1595
1596         top.evlist = evlist__new();
1597         if (top.evlist == NULL)
1598                 return -ENOMEM;
1599
1600         status = perf_config(perf_top_config, &top);
1601         if (status)
1602                 return status;
1603         /*
1604          * Since the per arch annotation init routine may need the cpuid, read
1605          * it here, since we are not getting this from the perf.data header.
1606          */
1607         status = perf_env__read_cpuid(&perf_env);
1608         if (status) {
1609                 /*
1610                  * Some arches do not provide a get_cpuid(), so just use pr_debug, otherwise
1611                  * warn the user explicitely.
1612                  */
1613                 eprintf(status == ENOSYS ? 1 : 0, verbose,
1614                         "Couldn't read the cpuid for this machine: %s\n",
1615                         str_error_r(errno, errbuf, sizeof(errbuf)));
1616         }
1617         top.evlist->env = &perf_env;
1618
1619         argc = parse_options(argc, argv, options, top_usage, 0);
1620         if (argc)
1621                 usage_with_options(top_usage, options);
1622
1623         if (annotate_check_args(&top.annotation_opts) < 0)
1624                 goto out_delete_evlist;
1625
1626         if (!top.evlist->core.nr_entries &&
1627             evlist__add_default(top.evlist) < 0) {
1628                 pr_err("Not enough memory for event selector list\n");
1629                 goto out_delete_evlist;
1630         }
1631
1632         status = evswitch__init(&top.evswitch, top.evlist, stderr);
1633         if (status)
1634                 goto out_delete_evlist;
1635
1636         if (symbol_conf.report_hierarchy) {
1637                 /* disable incompatible options */
1638                 symbol_conf.event_group = false;
1639                 symbol_conf.cumulate_callchain = false;
1640
1641                 if (field_order) {
1642                         pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1643                         parse_options_usage(top_usage, options, "fields", 0);
1644                         parse_options_usage(NULL, options, "hierarchy", 0);
1645                         goto out_delete_evlist;
1646                 }
1647         }
1648
1649         if (top.stitch_lbr && !(callchain_param.record_mode == CALLCHAIN_LBR)) {
1650                 pr_err("Error: --stitch-lbr must be used with --call-graph lbr\n");
1651                 goto out_delete_evlist;
1652         }
1653
1654         if (opts->branch_stack && callchain_param.enabled)
1655                 symbol_conf.show_branchflag_count = true;
1656
1657         sort__mode = SORT_MODE__TOP;
1658         /* display thread wants entries to be collapsed in a different tree */
1659         perf_hpp_list.need_collapse = 1;
1660
1661         if (top.use_stdio)
1662                 use_browser = 0;
1663         else if (top.use_tui)
1664                 use_browser = 1;
1665
1666         setup_browser(false);
1667
1668         if (setup_sorting(top.evlist) < 0) {
1669                 if (sort_order)
1670                         parse_options_usage(top_usage, options, "s", 1);
1671                 if (field_order)
1672                         parse_options_usage(sort_order ? NULL : top_usage,
1673                                             options, "fields", 0);
1674                 goto out_delete_evlist;
1675         }
1676
1677         status = target__validate(target);
1678         if (status) {
1679                 target__strerror(target, status, errbuf, BUFSIZ);
1680                 ui__warning("%s\n", errbuf);
1681         }
1682
1683         status = target__parse_uid(target);
1684         if (status) {
1685                 int saved_errno = errno;
1686
1687                 target__strerror(target, status, errbuf, BUFSIZ);
1688                 ui__error("%s\n", errbuf);
1689
1690                 status = -saved_errno;
1691                 goto out_delete_evlist;
1692         }
1693
1694         if (target__none(target))
1695                 target->system_wide = true;
1696
1697         if (perf_evlist__create_maps(top.evlist, target) < 0) {
1698                 ui__error("Couldn't create thread/CPU maps: %s\n",
1699                           errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1700                 goto out_delete_evlist;
1701         }
1702
1703         if (top.delay_secs < 1)
1704                 top.delay_secs = 1;
1705
1706         if (record_opts__config(opts)) {
1707                 status = -EINVAL;
1708                 goto out_delete_evlist;
1709         }
1710
1711         top.sym_evsel = evlist__first(top.evlist);
1712
1713         if (!callchain_param.enabled) {
1714                 symbol_conf.cumulate_callchain = false;
1715                 perf_hpp__cancel_cumulate();
1716         }
1717
1718         if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1719                 callchain_param.order = ORDER_CALLER;
1720
1721         status = symbol__annotation_init();
1722         if (status < 0)
1723                 goto out_delete_evlist;
1724
1725         annotation_config__init(&top.annotation_opts);
1726
1727         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1728         status = symbol__init(NULL);
1729         if (status < 0)
1730                 goto out_delete_evlist;
1731
1732         sort__setup_elide(stdout);
1733
1734         get_term_dimensions(&top.winsize);
1735         if (top.print_entries == 0) {
1736                 perf_top__update_print_entries(&top);
1737                 signal(SIGWINCH, winch_sig);
1738         }
1739
1740         top.session = perf_session__new(NULL, false, NULL);
1741         if (IS_ERR(top.session)) {
1742                 status = PTR_ERR(top.session);
1743                 goto out_delete_evlist;
1744         }
1745
1746 #ifdef HAVE_LIBBPF_SUPPORT
1747         if (!top.record_opts.no_bpf_event) {
1748                 top.sb_evlist = evlist__new();
1749
1750                 if (top.sb_evlist == NULL) {
1751                         pr_err("Couldn't create side band evlist.\n.");
1752                         goto out_delete_evlist;
1753                 }
1754
1755                 if (evlist__add_bpf_sb_event(top.sb_evlist, &perf_env)) {
1756                         pr_err("Couldn't ask for PERF_RECORD_BPF_EVENT side band events.\n.");
1757                         goto out_delete_evlist;
1758                 }
1759         }
1760 #endif
1761
1762         if (evlist__start_sb_thread(top.sb_evlist, target)) {
1763                 pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1764                 opts->no_bpf_event = true;
1765         }
1766
1767         status = __cmd_top(&top);
1768
1769         if (!opts->no_bpf_event)
1770                 evlist__stop_sb_thread(top.sb_evlist);
1771
1772 out_delete_evlist:
1773         evlist__delete(top.evlist);
1774         perf_session__delete(top.session);
1775
1776         return status;
1777 }