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