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