tools/perf/builtin-top.c

   1 /*
   2  * builtin-top.c
   3  *
   4  * Builtin top command: Display a continuously updated profile of
   5  * any workload, CPU or specific PID.
   6  *
   7  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
   8  *               2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   9  *
  10  * Improvements and fixes by:
  11  *
  12  *   Arjan van de Ven <arjan@linux.intel.com>
  13  *   Yanmin Zhang <yanmin.zhang@intel.com>
  14  *   Wu Fengguang <fengguang.wu@intel.com>
  15  *   Mike Galbraith <efault@gmx.de>
  16  *   Paul Mackerras <paulus@samba.org>
  17  *
  18  * Released under the GPL v2. (and only v2, not any later version)
  19  */
  20 #include "builtin.h"
  21
  22 #include "perf.h"
  23
  24 #include "util/annotate.h"
  25 #include "util/config.h"
  26 #include "util/color.h"
  27 #include "util/drv_configs.h"
  28 #include "util/evlist.h"
  29 #include "util/evsel.h"
  30 #include "util/event.h"
  31 #include "util/machine.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/term.h"
  44 #include "util/intlist.h"
  45 #include "util/parse-branch-options.h"
  46 #include "arch/common.h"
  47
  48 #include "util/debug.h"
  49
  50 #include <assert.h>
  51 #include <elf.h>
  52 #include <fcntl.h>
  53
  54 #include <stdio.h>
  55 #include <termios.h>
  56 #include <unistd.h>
  57 #include <inttypes.h>
  58
  59 #include <errno.h>
  60 #include <time.h>
  61 #include <sched.h>
  62 #include <signal.h>
  63
  64 #include <sys/syscall.h>
  65 #include <sys/ioctl.h>
  66 #include <poll.h>
  67 #include <sys/prctl.h>
  68 #include <sys/wait.h>
  69 #include <sys/uio.h>
  70 #include <sys/utsname.h>
  71 #include <sys/mman.h>
  72
  73 #include <linux/stringify.h>
  74 #include <linux/time64.h>
  75 #include <linux/types.h>
  76
  77 #include "sane_ctype.h"
  78
  79 static volatile int done;
  80 static volatile int resize;
  81
  82 #define HEADER_LINE_NR  5
  83
  84 static void perf_top__update_print_entries(struct perf_top *top)
  85 {
  86         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
  87 }
  88
  89 static void winch_sig(int sig __maybe_unused)
  90 {
  91         resize = 1;
  92 }
  93
  94 static void perf_top__resize(struct perf_top *top)
  95 {
  96         get_term_dimensions(&top->winsize);
  97         perf_top__update_print_entries(top);
  98 }
  99
 100 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
 101 {
 102         struct perf_evsel *evsel = hists_to_evsel(he->hists);
 103         struct symbol *sym;
 104         struct annotation *notes;
 105         struct map *map;
 106         int err = -1;
 107
 108         if (!he || !he->ms.sym)
 109                 return -1;
 110
 111         sym = he->ms.sym;
 112         map = he->ms.map;
 113
 114         /*
 115          * We can't annotate with just /proc/kallsyms
 116          */
 117         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 118             !dso__is_kcore(map->dso)) {
 119                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 120                        "path\n", sym->name);
 121                 sleep(1);
 122                 return -1;
 123         }
 124
 125         notes = symbol__annotation(sym);
 126         pthread_mutex_lock(&notes->lock);
 127
 128         if (!symbol__hists(sym, top->evlist->nr_entries)) {
 129                 pthread_mutex_unlock(&notes->lock);
 130                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 131                        sym->name);
 132                 sleep(1);
 133                 return err;
 134         }
 135
 136         err = symbol__annotate(sym, map, evsel, 0, &top->annotation_opts, NULL);
 137         if (err == 0) {
 138                 top->sym_filter_entry = he;
 139         } else {
 140                 char msg[BUFSIZ];
 141                 symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
 142                 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
 143         }
 144
 145         pthread_mutex_unlock(&notes->lock);
 146         return err;
 147 }
 148
 149 static void __zero_source_counters(struct hist_entry *he)
 150 {
 151         struct symbol *sym = he->ms.sym;
 152         symbol__annotate_zero_histograms(sym);
 153 }
 154
 155 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 156 {
 157         struct utsname uts;
 158         int err = uname(&uts);
 159
 160         ui__warning("Out of bounds address found:\n\n"
 161                     "Addr:   %" PRIx64 "\n"
 162                     "DSO:    %s %c\n"
 163                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
 164                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 165                     "Arch:   %s\n"
 166                     "Kernel: %s\n"
 167                     "Tools:  %s\n\n"
 168                     "Not all samples will be on the annotation output.\n\n"
 169                     "Please report to linux-kernel@vger.kernel.org\n",
 170                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
 171                     map->start, map->end, sym->start, sym->end,
 172                     sym->binding == STB_GLOBAL ? 'g' :
 173                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 174                     err ? "[unknown]" : uts.machine,
 175                     err ? "[unknown]" : uts.release, perf_version_string);
 176         if (use_browser <= 0)
 177                 sleep(5);
 178
 179         map->erange_warned = true;
 180 }
 181
 182 static void perf_top__record_precise_ip(struct perf_top *top,
 183                                         struct hist_entry *he,
 184                                         struct perf_sample *sample,
 185                                         struct perf_evsel *evsel, u64 ip)
 186 {
 187         struct annotation *notes;
 188         struct symbol *sym = he->ms.sym;
 189         int err = 0;
 190
 191         if (sym == NULL || (use_browser == 0 &&
 192                             (top->sym_filter_entry == NULL ||
 193                              top->sym_filter_entry->ms.sym != sym)))
 194                 return;
 195
 196         notes = symbol__annotation(sym);
 197
 198         if (pthread_mutex_trylock(&notes->lock))
 199                 return;
 200
 201         err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
 202
 203         pthread_mutex_unlock(&notes->lock);
 204
 205         if (unlikely(err)) {
 206                 /*
 207                  * This function is now called with he->hists->lock held.
 208                  * Release it before going to sleep.
 209                  */
 210                 pthread_mutex_unlock(&he->hists->lock);
 211
 212                 if (err == -ERANGE && !he->ms.map->erange_warned)
 213                         ui__warn_map_erange(he->ms.map, sym, ip);
 214                 else if (err == -ENOMEM) {
 215                         pr_err("Not enough memory for annotating '%s' symbol!\n",
 216                                sym->name);
 217                         sleep(1);
 218                 }
 219
 220                 pthread_mutex_lock(&he->hists->lock);
 221         }
 222 }
 223
 224 static void perf_top__show_details(struct perf_top *top)
 225 {
 226         struct hist_entry *he = top->sym_filter_entry;
 227         struct perf_evsel *evsel = hists_to_evsel(he->hists);
 228         struct annotation *notes;
 229         struct symbol *symbol;
 230         int more;
 231
 232         if (!he)
 233                 return;
 234
 235         symbol = he->ms.sym;
 236         notes = symbol__annotation(symbol);
 237
 238         pthread_mutex_lock(&notes->lock);
 239
 240         symbol__calc_percent(symbol, evsel);
 241
 242         if (notes->src == NULL)
 243                 goto out_unlock;
 244
 245         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
 246         printf("  Events  Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
 247
 248         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel, &top->annotation_opts);
 249
 250         if (top->evlist->enabled) {
 251                 if (top->zero)
 252                         symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 253                 else
 254                         symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 255         }
 256         if (more != 0)
 257                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 258 out_unlock:
 259         pthread_mutex_unlock(&notes->lock);
 260 }
 261
 262 static void perf_top__print_sym_table(struct perf_top *top)
 263 {
 264         char bf[160];
 265         int printed = 0;
 266         const int win_width = top->winsize.ws_col - 1;
 267         struct perf_evsel *evsel = top->sym_evsel;
 268         struct hists *hists = evsel__hists(evsel);
 269
 270         puts(CONSOLE_CLEAR);
 271
 272         perf_top__header_snprintf(top, bf, sizeof(bf));
 273         printf("%s\n", bf);
 274
 275         perf_top__reset_sample_counters(top);
 276
 277         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 278
 279         if (!top->record_opts.overwrite &&
 280             (hists->stats.nr_lost_warned !=
 281             hists->stats.nr_events[PERF_RECORD_LOST])) {
 282                 hists->stats.nr_lost_warned =
 283                               hists->stats.nr_events[PERF_RECORD_LOST];
 284                 color_fprintf(stdout, PERF_COLOR_RED,
 285                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 286                               hists->stats.nr_lost_warned);
 287                 ++printed;
 288         }
 289
 290         if (top->sym_filter_entry) {
 291                 perf_top__show_details(top);
 292                 return;
 293         }
 294
 295         if (top->evlist->enabled) {
 296                 if (top->zero) {
 297                         hists__delete_entries(hists);
 298                 } else {
 299                         hists__decay_entries(hists, top->hide_user_symbols,
 300                                              top->hide_kernel_symbols);
 301                 }
 302         }
 303
 304         hists__collapse_resort(hists, NULL);
 305         perf_evsel__output_resort(evsel, NULL);
 306
 307         hists__output_recalc_col_len(hists, top->print_entries - printed);
 308         putchar('\n');
 309         hists__fprintf(hists, false, top->print_entries - printed, win_width,
 310                        top->min_percent, stdout, symbol_conf.use_callchain);
 311 }
 312
 313 static void prompt_integer(int *target, const char *msg)
 314 {
 315         char *buf = malloc(0), *p;
 316         size_t dummy = 0;
 317         int tmp;
 318
 319         fprintf(stdout, "\n%s: ", msg);
 320         if (getline(&buf, &dummy, stdin) < 0)
 321                 return;
 322
 323         p = strchr(buf, '\n');
 324         if (p)
 325                 *p = 0;
 326
 327         p = buf;
 328         while(*p) {
 329                 if (!isdigit(*p))
 330                         goto out_free;
 331                 p++;
 332         }
 333         tmp = strtoul(buf, NULL, 10);
 334         *target = tmp;
 335 out_free:
 336         free(buf);
 337 }
 338
 339 static void prompt_percent(int *target, const char *msg)
 340 {
 341         int tmp = 0;
 342
 343         prompt_integer(&tmp, msg);
 344         if (tmp >= 0 && tmp <= 100)
 345                 *target = tmp;
 346 }
 347
 348 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 349 {
 350         char *buf = malloc(0), *p;
 351         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 352         struct hists *hists = evsel__hists(top->sym_evsel);
 353         struct rb_node *next;
 354         size_t dummy = 0;
 355
 356         /* zero counters of active symbol */
 357         if (syme) {
 358                 __zero_source_counters(syme);
 359                 top->sym_filter_entry = NULL;
 360         }
 361
 362         fprintf(stdout, "\n%s: ", msg);
 363         if (getline(&buf, &dummy, stdin) < 0)
 364                 goto out_free;
 365
 366         p = strchr(buf, '\n');
 367         if (p)
 368                 *p = 0;
 369
 370         next = rb_first(&hists->entries);
 371         while (next) {
 372                 n = rb_entry(next, struct hist_entry, rb_node);
 373                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 374                         found = n;
 375                         break;
 376                 }
 377                 next = rb_next(&n->rb_node);
 378         }
 379
 380         if (!found) {
 381                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 382                 sleep(1);
 383         } else
 384                 perf_top__parse_source(top, found);
 385
 386 out_free:
 387         free(buf);
 388 }
 389
 390 static void perf_top__print_mapped_keys(struct perf_top *top)
 391 {
 392         char *name = NULL;
 393
 394         if (top->sym_filter_entry) {
 395                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 396                 name = sym->name;
 397         }
 398
 399         fprintf(stdout, "\nMapped keys:\n");
 400         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 401         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 402
 403         if (top->evlist->nr_entries > 1)
 404                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 405
 406         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 407
 408         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
 409         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 410         fprintf(stdout, "\t[S]     stop annotation.\n");
 411
 412         fprintf(stdout,
 413                 "\t[K]     hide kernel symbols.             \t(%s)\n",
 414                 top->hide_kernel_symbols ? "yes" : "no");
 415         fprintf(stdout,
 416                 "\t[U]     hide user symbols.               \t(%s)\n",
 417                 top->hide_user_symbols ? "yes" : "no");
 418         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 419         fprintf(stdout, "\t[qQ]    quit.\n");
 420 }
 421
 422 static int perf_top__key_mapped(struct perf_top *top, int c)
 423 {
 424         switch (c) {
 425                 case 'd':
 426                 case 'e':
 427                 case 'f':
 428                 case 'z':
 429                 case 'q':
 430                 case 'Q':
 431                 case 'K':
 432                 case 'U':
 433                 case 'F':
 434                 case 's':
 435                 case 'S':
 436                         return 1;
 437                 case 'E':
 438                         return top->evlist->nr_entries > 1 ? 1 : 0;
 439                 default:
 440                         break;
 441         }
 442
 443         return 0;
 444 }
 445
 446 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 447 {
 448         bool ret = true;
 449
 450         if (!perf_top__key_mapped(top, c)) {
 451                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 452                 struct termios save;
 453
 454                 perf_top__print_mapped_keys(top);
 455                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 456                 fflush(stdout);
 457
 458                 set_term_quiet_input(&save);
 459
 460                 poll(&stdin_poll, 1, -1);
 461                 c = getc(stdin);
 462
 463                 tcsetattr(0, TCSAFLUSH, &save);
 464                 if (!perf_top__key_mapped(top, c))
 465                         return ret;
 466         }
 467
 468         switch (c) {
 469                 case 'd':
 470                         prompt_integer(&top->delay_secs, "Enter display delay");
 471                         if (top->delay_secs < 1)
 472                                 top->delay_secs = 1;
 473                         break;
 474                 case 'e':
 475                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 476                         if (top->print_entries == 0) {
 477                                 perf_top__resize(top);
 478                                 signal(SIGWINCH, winch_sig);
 479                         } else {
 480                                 signal(SIGWINCH, SIG_DFL);
 481                         }
 482                         break;
 483                 case 'E':
 484                         if (top->evlist->nr_entries > 1) {
 485                                 /* Select 0 as the default event: */
 486                                 int counter = 0;
 487
 488                                 fprintf(stderr, "\nAvailable events:");
 489
 490                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
 491                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
 492
 493                                 prompt_integer(&counter, "Enter details event counter");
 494
 495                                 if (counter >= top->evlist->nr_entries) {
 496                                         top->sym_evsel = perf_evlist__first(top->evlist);
 497                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
 498                                         sleep(1);
 499                                         break;
 500                                 }
 501                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
 502                                         if (top->sym_evsel->idx == counter)
 503                                                 break;
 504                         } else
 505                                 top->sym_evsel = perf_evlist__first(top->evlist);
 506                         break;
 507                 case 'f':
 508                         prompt_integer(&top->count_filter, "Enter display event count filter");
 509                         break;
 510                 case 'F':
 511                         prompt_percent(&top->annotation_opts.min_pcnt,
 512                                        "Enter details display event filter (percent)");
 513                         break;
 514                 case 'K':
 515                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 516                         break;
 517                 case 'q':
 518                 case 'Q':
 519                         printf("exiting.\n");
 520                         if (top->dump_symtab)
 521                                 perf_session__fprintf_dsos(top->session, stderr);
 522                         ret = false;
 523                         break;
 524                 case 's':
 525                         perf_top__prompt_symbol(top, "Enter details symbol");
 526                         break;
 527                 case 'S':
 528                         if (!top->sym_filter_entry)
 529                                 break;
 530                         else {
 531                                 struct hist_entry *syme = top->sym_filter_entry;
 532
 533                                 top->sym_filter_entry = NULL;
 534                                 __zero_source_counters(syme);
 535                         }
 536                         break;
 537                 case 'U':
 538                         top->hide_user_symbols = !top->hide_user_symbols;
 539                         break;
 540                 case 'z':
 541                         top->zero = !top->zero;
 542                         break;
 543                 default:
 544                         break;
 545         }
 546
 547         return ret;
 548 }
 549
 550 static void perf_top__sort_new_samples(void *arg)
 551 {
 552         struct perf_top *t = arg;
 553         struct perf_evsel *evsel = t->sym_evsel;
 554         struct hists *hists;
 555
 556         perf_top__reset_sample_counters(t);
 557
 558         if (t->evlist->selected != NULL)
 559                 t->sym_evsel = t->evlist->selected;
 560
 561         hists = evsel__hists(evsel);
 562
 563         if (t->evlist->enabled) {
 564                 if (t->zero) {
 565                         hists__delete_entries(hists);
 566                 } else {
 567                         hists__decay_entries(hists, t->hide_user_symbols,
 568                                              t->hide_kernel_symbols);
 569                 }
 570         }
 571
 572         hists__collapse_resort(hists, NULL);
 573         perf_evsel__output_resort(evsel, NULL);
 574 }
 575
 576 static void *display_thread_tui(void *arg)
 577 {
 578         struct perf_evsel *pos;
 579         struct perf_top *top = arg;
 580         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 581         struct hist_browser_timer hbt = {
 582                 .timer          = perf_top__sort_new_samples,
 583                 .arg            = top,
 584                 .refresh        = top->delay_secs,
 585         };
 586
 587         /* In order to read symbols from other namespaces perf to  needs to call
 588          * setns(2).  This isn't permitted if the struct_fs has multiple users.
 589          * unshare(2) the fs so that we may continue to setns into namespaces
 590          * that we're observing.
 591          */
 592         unshare(CLONE_FS);
 593
 594         perf_top__sort_new_samples(top);
 595
 596         /*
 597          * Initialize the uid_filter_str, in the future the TUI will allow
 598          * Zooming in/out UIDs. For now juse use whatever the user passed
 599          * via --uid.
 600          */
 601         evlist__for_each_entry(top->evlist, pos) {
 602                 struct hists *hists = evsel__hists(pos);
 603                 hists->uid_filter_str = top->record_opts.target.uid_str;
 604         }
 605
 606         perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
 607                                       top->min_percent,
 608                                       &top->session->header.env,
 609                                       !top->record_opts.overwrite);
 610
 611         done = 1;
 612         return NULL;
 613 }
 614
 615 static void display_sig(int sig __maybe_unused)
 616 {
 617         done = 1;
 618 }
 619
 620 static void display_setup_sig(void)
 621 {
 622         signal(SIGSEGV, sighandler_dump_stack);
 623         signal(SIGFPE, sighandler_dump_stack);
 624         signal(SIGINT,  display_sig);
 625         signal(SIGQUIT, display_sig);
 626         signal(SIGTERM, display_sig);
 627 }
 628
 629 static void *display_thread(void *arg)
 630 {
 631         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 632         struct termios save;
 633         struct perf_top *top = arg;
 634         int delay_msecs, c;
 635
 636         /* In order to read symbols from other namespaces perf to  needs to call
 637          * setns(2).  This isn't permitted if the struct_fs has multiple users.
 638          * unshare(2) the fs so that we may continue to setns into namespaces
 639          * that we're observing.
 640          */
 641         unshare(CLONE_FS);
 642
 643         display_setup_sig();
 644         pthread__unblock_sigwinch();
 645 repeat:
 646         delay_msecs = top->delay_secs * MSEC_PER_SEC;
 647         set_term_quiet_input(&save);
 648         /* trash return*/
 649         getc(stdin);
 650
 651         while (!done) {
 652                 perf_top__print_sym_table(top);
 653                 /*
 654                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 655                  * refresh screen in both cases.
 656                  */
 657                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 658                 case 0:
 659                         continue;
 660                 case -1:
 661                         if (errno == EINTR)
 662                                 continue;
 663                         __fallthrough;
 664                 default:
 665                         c = getc(stdin);
 666                         tcsetattr(0, TCSAFLUSH, &save);
 667
 668                         if (perf_top__handle_keypress(top, c))
 669                                 goto repeat;
 670                         done = 1;
 671                 }
 672         }
 673
 674         tcsetattr(0, TCSAFLUSH, &save);
 675         return NULL;
 676 }
 677
 678 static int hist_iter__top_callback(struct hist_entry_iter *iter,
 679                                    struct addr_location *al, bool single,
 680                                    void *arg)
 681 {
 682         struct perf_top *top = arg;
 683         struct hist_entry *he = iter->he;
 684         struct perf_evsel *evsel = iter->evsel;
 685
 686         if (perf_hpp_list.sym && single)
 687                 perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
 688
 689         hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
 690                      !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
 691         return 0;
 692 }
 693
 694 static void perf_event__process_sample(struct perf_tool *tool,
 695                                        const union perf_event *event,
 696                                        struct perf_evsel *evsel,
 697                                        struct perf_sample *sample,
 698                                        struct machine *machine)
 699 {
 700         struct perf_top *top = container_of(tool, struct perf_top, tool);
 701         struct addr_location al;
 702         int err;
 703
 704         if (!machine && perf_guest) {
 705                 static struct intlist *seen;
 706
 707                 if (!seen)
 708                         seen = intlist__new(NULL);
 709
 710                 if (!intlist__has_entry(seen, sample->pid)) {
 711                         pr_err("Can't find guest [%d]'s kernel information\n",
 712                                 sample->pid);
 713                         intlist__add(seen, sample->pid);
 714                 }
 715                 return;
 716         }
 717
 718         if (!machine) {
 719                 pr_err("%u unprocessable samples recorded.\r",
 720                        top->session->evlist->stats.nr_unprocessable_samples++);
 721                 return;
 722         }
 723
 724         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 725                 top->exact_samples++;
 726
 727         if (machine__resolve(machine, &al, sample) < 0)
 728                 return;
 729
 730         if (!machine->kptr_restrict_warned &&
 731             symbol_conf.kptr_restrict &&
 732             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 733                 if (!perf_evlist__exclude_kernel(top->session->evlist)) {
 734                         ui__warning(
 735 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 736 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 737 "Kernel%s samples will not be resolved.\n",
 738                           al.map && map__has_symbols(al.map) ?
 739                           " modules" : "");
 740                         if (use_browser <= 0)
 741                                 sleep(5);
 742                 }
 743                 machine->kptr_restrict_warned = true;
 744         }
 745
 746         if (al.sym == NULL && al.map != NULL) {
 747                 const char *msg = "Kernel samples will not be resolved.\n";
 748                 /*
 749                  * As we do lazy loading of symtabs we only will know if the
 750                  * specified vmlinux file is invalid when we actually have a
 751                  * hit in kernel space and then try to load it. So if we get
 752                  * here and there are _no_ symbols in the DSO backing the
 753                  * kernel map, bail out.
 754                  *
 755                  * We may never get here, for instance, if we use -K/
 756                  * --hide-kernel-symbols, even if the user specifies an
 757                  * invalid --vmlinux ;-)
 758                  */
 759                 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
 760                     __map__is_kernel(al.map) && map__has_symbols(al.map)) {
 761                         if (symbol_conf.vmlinux_name) {
 762                                 char serr[256];
 763                                 dso__strerror_load(al.map->dso, serr, sizeof(serr));
 764                                 ui__warning("The %s file can't be used: %s\n%s",
 765                                             symbol_conf.vmlinux_name, serr, msg);
 766                         } else {
 767                                 ui__warning("A vmlinux file was not found.\n%s",
 768                                             msg);
 769                         }
 770
 771                         if (use_browser <= 0)
 772                                 sleep(5);
 773                         top->vmlinux_warned = true;
 774                 }
 775         }
 776
 777         if (al.sym == NULL || !al.sym->idle) {
 778                 struct hists *hists = evsel__hists(evsel);
 779                 struct hist_entry_iter iter = {
 780                         .evsel          = evsel,
 781                         .sample         = sample,
 782                         .add_entry_cb   = hist_iter__top_callback,
 783                 };
 784
 785                 if (symbol_conf.cumulate_callchain)
 786                         iter.ops = &hist_iter_cumulative;
 787                 else
 788                         iter.ops = &hist_iter_normal;
 789
 790                 pthread_mutex_lock(&hists->lock);
 791
 792                 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
 793                 if (err < 0)
 794                         pr_err("Problem incrementing symbol period, skipping event\n");
 795
 796                 pthread_mutex_unlock(&hists->lock);
 797         }
 798
 799         addr_location__put(&al);
 800 }
 801
 802 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 803 {
 804         struct record_opts *opts = &top->record_opts;
 805         struct perf_evlist *evlist = top->evlist;
 806         struct perf_sample sample;
 807         struct perf_evsel *evsel;
 808         struct perf_mmap *md;
 809         struct perf_session *session = top->session;
 810         union perf_event *event;
 811         struct machine *machine;
 812         int ret;
 813
 814         md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
 815         if (perf_mmap__read_init(md) < 0)
 816                 return;
 817
 818         while ((event = perf_mmap__read_event(md)) != NULL) {
 819                 ret = perf_evlist__parse_sample(evlist, event, &sample);
 820                 if (ret) {
 821                         pr_err("Can't parse sample, err = %d\n", ret);
 822                         goto next_event;
 823                 }
 824
 825                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 826                 assert(evsel != NULL);
 827
 828                 if (event->header.type == PERF_RECORD_SAMPLE)
 829                         ++top->samples;
 830
 831                 switch (sample.cpumode) {
 832                 case PERF_RECORD_MISC_USER:
 833                         ++top->us_samples;
 834                         if (top->hide_user_symbols)
 835                                 goto next_event;
 836                         machine = &session->machines.host;
 837                         break;
 838                 case PERF_RECORD_MISC_KERNEL:
 839                         ++top->kernel_samples;
 840                         if (top->hide_kernel_symbols)
 841                                 goto next_event;
 842                         machine = &session->machines.host;
 843                         break;
 844                 case PERF_RECORD_MISC_GUEST_KERNEL:
 845                         ++top->guest_kernel_samples;
 846                         machine = perf_session__find_machine(session,
 847                                                              sample.pid);
 848                         break;
 849                 case PERF_RECORD_MISC_GUEST_USER:
 850                         ++top->guest_us_samples;
 851                         /*
 852                          * TODO: we don't process guest user from host side
 853                          * except simple counting.
 854                          */
 855                         goto next_event;
 856                 default:
 857                         if (event->header.type == PERF_RECORD_SAMPLE)
 858                                 goto next_event;
 859                         machine = &session->machines.host;
 860                         break;
 861                 }
 862
 863
 864                 if (event->header.type == PERF_RECORD_SAMPLE) {
 865                         perf_event__process_sample(&top->tool, event, evsel,
 866                                                    &sample, machine);
 867                 } else if (event->header.type < PERF_RECORD_MAX) {
 868                         hists__inc_nr_events(evsel__hists(evsel), event->header.type);
 869                         machine__process_event(machine, event, &sample);
 870                 } else
 871                         ++session->evlist->stats.nr_unknown_events;
 872 next_event:
 873                 perf_mmap__consume(md);
 874         }
 875
 876         perf_mmap__read_done(md);
 877 }
 878
 879 static void perf_top__mmap_read(struct perf_top *top)
 880 {
 881         bool overwrite = top->record_opts.overwrite;
 882         struct perf_evlist *evlist = top->evlist;
 883         unsigned long long start, end;
 884         int i;
 885
 886         start = rdclock();
 887         if (overwrite)
 888                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
 889
 890         for (i = 0; i < top->evlist->nr_mmaps; i++)
 891                 perf_top__mmap_read_idx(top, i);
 892
 893         if (overwrite) {
 894                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
 895                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 896         }
 897         end = rdclock();
 898
 899         if ((end - start) > (unsigned long long)top->delay_secs * NSEC_PER_SEC)
 900                 ui__warning("Too slow to read ring buffer.\n"
 901                             "Please try increasing the period (-c) or\n"
 902                             "decreasing the freq (-F) or\n"
 903                             "limiting the number of CPUs (-C)\n");
 904 }
 905
 906 /*
 907  * Check per-event overwrite term.
 908  * perf top should support consistent term for all events.
 909  * - All events don't have per-event term
 910  *   E.g. "cpu/cpu-cycles/,cpu/instructions/"
 911  *   Nothing change, return 0.
 912  * - All events have same per-event term
 913  *   E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
 914  *   Using the per-event setting to replace the opts->overwrite if
 915  *   they are different, then return 0.
 916  * - Events have different per-event term
 917  *   E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
 918  *   Return -1
 919  * - Some of the event set per-event term, but some not.
 920  *   E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
 921  *   Return -1
 922  */
 923 static int perf_top__overwrite_check(struct perf_top *top)
 924 {
 925         struct record_opts *opts = &top->record_opts;
 926         struct perf_evlist *evlist = top->evlist;
 927         struct perf_evsel_config_term *term;
 928         struct list_head *config_terms;
 929         struct perf_evsel *evsel;
 930         int set, overwrite = -1;
 931
 932         evlist__for_each_entry(evlist, evsel) {
 933                 set = -1;
 934                 config_terms = &evsel->config_terms;
 935                 list_for_each_entry(term, config_terms, list) {
 936                         if (term->type == PERF_EVSEL__CONFIG_TERM_OVERWRITE)
 937                                 set = term->val.overwrite ? 1 : 0;
 938                 }
 939
 940                 /* no term for current and previous event (likely) */
 941                 if ((overwrite < 0) && (set < 0))
 942                         continue;
 943
 944                 /* has term for both current and previous event, compare */
 945                 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
 946                         return -1;
 947
 948                 /* no term for current event but has term for previous one */
 949                 if ((overwrite >= 0) && (set < 0))
 950                         return -1;
 951
 952                 /* has term for current event */
 953                 if ((overwrite < 0) && (set >= 0)) {
 954                         /* if it's first event, set overwrite */
 955                         if (evsel == perf_evlist__first(evlist))
 956                                 overwrite = set;
 957                         else
 958                                 return -1;
 959                 }
 960         }
 961
 962         if ((overwrite >= 0) && (opts->overwrite != overwrite))
 963                 opts->overwrite = overwrite;
 964
 965         return 0;
 966 }
 967
 968 static int perf_top_overwrite_fallback(struct perf_top *top,
 969                                        struct perf_evsel *evsel)
 970 {
 971         struct record_opts *opts = &top->record_opts;
 972         struct perf_evlist *evlist = top->evlist;
 973         struct perf_evsel *counter;
 974
 975         if (!opts->overwrite)
 976                 return 0;
 977
 978         /* only fall back when first event fails */
 979         if (evsel != perf_evlist__first(evlist))
 980                 return 0;
 981
 982         evlist__for_each_entry(evlist, counter)
 983                 counter->attr.write_backward = false;
 984         opts->overwrite = false;
 985         pr_debug2("fall back to non-overwrite mode\n");
 986         return 1;
 987 }
 988
 989 static int perf_top__start_counters(struct perf_top *top)
 990 {
 991         char msg[BUFSIZ];
 992         struct perf_evsel *counter;
 993         struct perf_evlist *evlist = top->evlist;
 994         struct record_opts *opts = &top->record_opts;
 995
 996         if (perf_top__overwrite_check(top)) {
 997                 ui__error("perf top only support consistent per-event "
 998                           "overwrite setting for all events\n");
 999                 goto out_err;
1000         }
1001
1002         perf_evlist__config(evlist, opts, &callchain_param);
1003
1004         evlist__for_each_entry(evlist, counter) {
1005 try_again:
1006                 if (perf_evsel__open(counter, top->evlist->cpus,
1007                                      top->evlist->threads) < 0) {
1008
1009                         /*
1010                          * Specially handle overwrite fall back.
1011                          * Because perf top is the only tool which has
1012                          * overwrite mode by default, support
1013                          * both overwrite and non-overwrite mode, and
1014                          * require consistent mode for all events.
1015                          *
1016                          * May move it to generic code with more tools
1017                          * have similar attribute.
1018                          */
1019                         if (perf_missing_features.write_backward &&
1020                             perf_top_overwrite_fallback(top, counter))
1021                                 goto try_again;
1022
1023                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
1024                                 if (verbose > 0)
1025                                         ui__warning("%s\n", msg);
1026                                 goto try_again;
1027                         }
1028
1029                         perf_evsel__open_strerror(counter, &opts->target,
1030                                                   errno, msg, sizeof(msg));
1031                         ui__error("%s\n", msg);
1032                         goto out_err;
1033                 }
1034         }
1035
1036         if (perf_evlist__mmap(evlist, opts->mmap_pages) < 0) {
1037                 ui__error("Failed to mmap with %d (%s)\n",
1038                             errno, str_error_r(errno, msg, sizeof(msg)));
1039                 goto out_err;
1040         }
1041
1042         return 0;
1043
1044 out_err:
1045         return -1;
1046 }
1047
1048 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1049 {
1050         if (!perf_hpp_list.sym) {
1051                 if (callchain->enabled) {
1052                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
1053                         return -EINVAL;
1054                 }
1055         } else if (callchain->mode != CHAIN_NONE) {
1056                 if (callchain_register_param(callchain) < 0) {
1057                         ui__error("Can't register callchain params.\n");
1058                         return -EINVAL;
1059                 }
1060         }
1061
1062         return 0;
1063 }
1064
1065 static int __cmd_top(struct perf_top *top)
1066 {
1067         char msg[512];
1068         struct perf_evsel *pos;
1069         struct perf_evsel_config_term *err_term;
1070         struct perf_evlist *evlist = top->evlist;
1071         struct record_opts *opts = &top->record_opts;
1072         pthread_t thread;
1073         int ret;
1074
1075         top->session = perf_session__new(NULL, false, NULL);
1076         if (top->session == NULL)
1077                 return -1;
1078
1079         if (!objdump_path) {
1080                 ret = perf_env__lookup_objdump(&top->session->header.env);
1081                 if (ret)
1082                         goto out_delete;
1083         }
1084
1085         ret = callchain_param__setup_sample_type(&callchain_param);
1086         if (ret)
1087                 goto out_delete;
1088
1089         if (perf_session__register_idle_thread(top->session) < 0)
1090                 goto out_delete;
1091
1092         if (top->nr_threads_synthesize > 1)
1093                 perf_set_multithreaded();
1094
1095         machine__synthesize_threads(&top->session->machines.host, &opts->target,
1096                                     top->evlist->threads, false,
1097                                     opts->proc_map_timeout,
1098                                     top->nr_threads_synthesize);
1099
1100         if (top->nr_threads_synthesize > 1)
1101                 perf_set_singlethreaded();
1102
1103         if (perf_hpp_list.socket) {
1104                 ret = perf_env__read_cpu_topology_map(&perf_env);
1105                 if (ret < 0)
1106                         goto out_err_cpu_topo;
1107         }
1108
1109         ret = perf_top__start_counters(top);
1110         if (ret)
1111                 goto out_delete;
1112
1113         ret = perf_evlist__apply_drv_configs(evlist, &pos, &err_term);
1114         if (ret) {
1115                 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
1116                         err_term->val.drv_cfg, perf_evsel__name(pos), errno,
1117                         str_error_r(errno, msg, sizeof(msg)));
1118                 goto out_delete;
1119         }
1120
1121         top->session->evlist = top->evlist;
1122         perf_session__set_id_hdr_size(top->session);
1123
1124         /*
1125          * When perf is starting the traced process, all the events (apart from
1126          * group members) have enable_on_exec=1 set, so don't spoil it by
1127          * prematurely enabling them.
1128          *
1129          * XXX 'top' still doesn't start workloads like record, trace, but should,
1130          * so leave the check here.
1131          */
1132         if (!target__none(&opts->target))
1133                 perf_evlist__enable(top->evlist);
1134
1135         /* Wait for a minimal set of events before starting the snapshot */
1136         perf_evlist__poll(top->evlist, 100);
1137
1138         perf_top__mmap_read(top);
1139
1140         ret = -1;
1141         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1142                                                             display_thread), top)) {
1143                 ui__error("Could not create display thread.\n");
1144                 goto out_delete;
1145         }
1146
1147         if (top->realtime_prio) {
1148                 struct sched_param param;
1149
1150                 param.sched_priority = top->realtime_prio;
1151                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1152                         ui__error("Could not set realtime priority.\n");
1153                         goto out_join;
1154                 }
1155         }
1156
1157         while (!done) {
1158                 u64 hits = top->samples;
1159
1160                 perf_top__mmap_read(top);
1161
1162                 if (opts->overwrite || (hits == top->samples))
1163                         ret = perf_evlist__poll(top->evlist, 100);
1164
1165                 if (resize) {
1166                         perf_top__resize(top);
1167                         resize = 0;
1168                 }
1169         }
1170
1171         ret = 0;
1172 out_join:
1173         pthread_join(thread, NULL);
1174 out_delete:
1175         perf_session__delete(top->session);
1176         top->session = NULL;
1177
1178         return ret;
1179
1180 out_err_cpu_topo: {
1181         char errbuf[BUFSIZ];
1182         const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1183
1184         ui__error("Could not read the CPU topology map: %s\n", err);
1185         goto out_delete;
1186 }
1187 }
1188
1189 static int
1190 callchain_opt(const struct option *opt, const char *arg, int unset)
1191 {
1192         symbol_conf.use_callchain = true;
1193         return record_callchain_opt(opt, arg, unset);
1194 }
1195
1196 static int
1197 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1198 {
1199         struct callchain_param *callchain = opt->value;
1200
1201         callchain->enabled = !unset;
1202         callchain->record_mode = CALLCHAIN_FP;
1203
1204         /*
1205          * --no-call-graph
1206          */
1207         if (unset) {
1208                 symbol_conf.use_callchain = false;
1209                 callchain->record_mode = CALLCHAIN_NONE;
1210                 return 0;
1211         }
1212
1213         return parse_callchain_top_opt(arg);
1214 }
1215
1216 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1217 {
1218         if (!strcmp(var, "top.call-graph")) {
1219                 var = "call-graph.record-mode";
1220                 return perf_default_config(var, value, cb);
1221         }
1222         if (!strcmp(var, "top.children")) {
1223                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1224                 return 0;
1225         }
1226
1227         return 0;
1228 }
1229
1230 static int
1231 parse_percent_limit(const struct option *opt, const char *arg,
1232                     int unset __maybe_unused)
1233 {
1234         struct perf_top *top = opt->value;
1235
1236         top->min_percent = strtof(arg, NULL);
1237         return 0;
1238 }
1239
1240 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1241         "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1242
1243 int cmd_top(int argc, const char **argv)
1244 {
1245         char errbuf[BUFSIZ];
1246         struct perf_top top = {
1247                 .count_filter        = 5,
1248                 .delay_secs          = 2,
1249                 .record_opts = {
1250                         .mmap_pages     = UINT_MAX,
1251                         .user_freq      = UINT_MAX,
1252                         .user_interval  = ULLONG_MAX,
1253                         .freq           = 4000, /* 4 KHz */
1254                         .target         = {
1255                                 .uses_mmap   = true,
1256                         },
1257                         .proc_map_timeout    = 500,
1258                         .overwrite      = 1,
1259                 },
1260                 .max_stack           = sysctl__max_stack(),
1261                 .annotation_opts     = annotation__default_options,
1262                 .nr_threads_synthesize = UINT_MAX,
1263         };
1264         struct record_opts *opts = &top.record_opts;
1265         struct target *target = &opts->target;
1266         const struct option options[] = {
1267         OPT_CALLBACK('e', "event", &top.evlist, "event",
1268                      "event selector. use 'perf list' to list available events",
1269                      parse_events_option),
1270         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1271         OPT_STRING('p', "pid", &target->pid, "pid",
1272                     "profile events on existing process id"),
1273         OPT_STRING('t', "tid", &target->tid, "tid",
1274                     "profile events on existing thread id"),
1275         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1276                             "system-wide collection from all CPUs"),
1277         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1278                     "list of cpus to monitor"),
1279         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1280                    "file", "vmlinux pathname"),
1281         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1282                     "don't load vmlinux even if found"),
1283         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1284                     "hide kernel symbols"),
1285         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1286                      "number of mmap data pages",
1287                      perf_evlist__parse_mmap_pages),
1288         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1289                     "collect data with this RT SCHED_FIFO priority"),
1290         OPT_INTEGER('d', "delay", &top.delay_secs,
1291                     "number of seconds to delay between refreshes"),
1292         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1293                             "dump the symbol table used for profiling"),
1294         OPT_INTEGER('f', "count-filter", &top.count_filter,
1295                     "only display functions with more events than this"),
1296         OPT_BOOLEAN(0, "group", &opts->group,
1297                             "put the counters into a counter group"),
1298         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1299                     "child tasks do not inherit counters"),
1300         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1301                     "symbol to annotate"),
1302         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1303         OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1304                      "profile at this frequency",
1305                       record__parse_freq),
1306         OPT_INTEGER('E', "entries", &top.print_entries,
1307                     "display this many functions"),
1308         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1309                     "hide user symbols"),
1310         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1311         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1312         OPT_INCR('v', "verbose", &verbose,
1313                     "be more verbose (show counter open errors, etc)"),
1314         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1315                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1316                    " Please refer the man page for the complete list."),
1317         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1318                    "output field(s): overhead, period, sample plus all of sort keys"),
1319         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1320                     "Show a column with the number of samples"),
1321         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1322                            NULL, "enables call-graph recording and display",
1323                            &callchain_opt),
1324         OPT_CALLBACK(0, "call-graph", &callchain_param,
1325                      "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1326                      top_callchain_help, &parse_callchain_opt),
1327         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1328                     "Accumulate callchains of children and show total overhead as well"),
1329         OPT_INTEGER(0, "max-stack", &top.max_stack,
1330                     "Set the maximum stack depth when parsing the callchain. "
1331                     "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1332         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1333                    "ignore callees of these functions in call graphs",
1334                    report_parse_ignore_callees_opt),
1335         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1336                     "Show a column with the sum of periods"),
1337         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1338                    "only consider symbols in these dsos"),
1339         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1340                    "only consider symbols in these comms"),
1341         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1342                    "only consider these symbols"),
1343         OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1344                     "Interleave source code with assembly code (default)"),
1345         OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1346                     "Display raw encoding of assembly instructions (default)"),
1347         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1348                     "Enable kernel symbol demangling"),
1349         OPT_STRING(0, "objdump", &objdump_path, "path",
1350                     "objdump binary to use for disassembly and annotations"),
1351         OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1352                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1353         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1354         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1355                      "Don't show entries under that percent", parse_percent_limit),
1356         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1357                      "How to display percentage of filtered entries", parse_filter_percentage),
1358         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1359                    "width[,width...]",
1360                    "don't try to adjust column width, use these fixed values"),
1361         OPT_UINTEGER(0, "proc-map-timeout", &opts->proc_map_timeout,
1362                         "per thread proc mmap processing timeout in ms"),
1363         OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1364                      "branch any", "sample any taken branches",
1365                      parse_branch_stack),
1366         OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1367                      "branch filter mask", "branch stack filter modes",
1368                      parse_branch_stack),
1369         OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1370                     "Show raw trace event output (do not use print fmt or plugins)"),
1371         OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1372                     "Show entries in a hierarchy"),
1373         OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1374         OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1375                         "number of thread to run event synthesize"),
1376         OPT_END()
1377         };
1378         const char * const top_usage[] = {
1379                 "perf top [<options>]",
1380                 NULL
1381         };
1382         int status = hists__init();
1383
1384         if (status < 0)
1385                 return status;
1386
1387         top.annotation_opts.min_pcnt = 5;
1388         top.annotation_opts.context  = 4;
1389
1390         top.evlist = perf_evlist__new();
1391         if (top.evlist == NULL)
1392                 return -ENOMEM;
1393
1394         status = perf_config(perf_top_config, &top);
1395         if (status)
1396                 return status;
1397
1398         argc = parse_options(argc, argv, options, top_usage, 0);
1399         if (argc)
1400                 usage_with_options(top_usage, options);
1401
1402         if (!top.evlist->nr_entries &&
1403             perf_evlist__add_default(top.evlist) < 0) {
1404                 pr_err("Not enough memory for event selector list\n");
1405                 goto out_delete_evlist;
1406         }
1407
1408         if (symbol_conf.report_hierarchy) {
1409                 /* disable incompatible options */
1410                 symbol_conf.event_group = false;
1411                 symbol_conf.cumulate_callchain = false;
1412
1413                 if (field_order) {
1414                         pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1415                         parse_options_usage(top_usage, options, "fields", 0);
1416                         parse_options_usage(NULL, options, "hierarchy", 0);
1417                         goto out_delete_evlist;
1418                 }
1419         }
1420
1421         sort__mode = SORT_MODE__TOP;
1422         /* display thread wants entries to be collapsed in a different tree */
1423         perf_hpp_list.need_collapse = 1;
1424
1425         if (top.use_stdio)
1426                 use_browser = 0;
1427         else if (top.use_tui)
1428                 use_browser = 1;
1429
1430         setup_browser(false);
1431
1432         if (setup_sorting(top.evlist) < 0) {
1433                 if (sort_order)
1434                         parse_options_usage(top_usage, options, "s", 1);
1435                 if (field_order)
1436                         parse_options_usage(sort_order ? NULL : top_usage,
1437                                             options, "fields", 0);
1438                 goto out_delete_evlist;
1439         }
1440
1441         status = target__validate(target);
1442         if (status) {
1443                 target__strerror(target, status, errbuf, BUFSIZ);
1444                 ui__warning("%s\n", errbuf);
1445         }
1446
1447         status = target__parse_uid(target);
1448         if (status) {
1449                 int saved_errno = errno;
1450
1451                 target__strerror(target, status, errbuf, BUFSIZ);
1452                 ui__error("%s\n", errbuf);
1453
1454                 status = -saved_errno;
1455                 goto out_delete_evlist;
1456         }
1457
1458         if (target__none(target))
1459                 target->system_wide = true;
1460
1461         if (perf_evlist__create_maps(top.evlist, target) < 0) {
1462                 ui__error("Couldn't create thread/CPU maps: %s\n",
1463                           errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1464                 goto out_delete_evlist;
1465         }
1466
1467         if (top.delay_secs < 1)
1468                 top.delay_secs = 1;
1469
1470         if (record_opts__config(opts)) {
1471                 status = -EINVAL;
1472                 goto out_delete_evlist;
1473         }
1474
1475         top.sym_evsel = perf_evlist__first(top.evlist);
1476
1477         if (!callchain_param.enabled) {
1478                 symbol_conf.cumulate_callchain = false;
1479                 perf_hpp__cancel_cumulate();
1480         }
1481
1482         if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1483                 callchain_param.order = ORDER_CALLER;
1484
1485         status = symbol__annotation_init();
1486         if (status < 0)
1487                 goto out_delete_evlist;
1488
1489         annotation_config__init();
1490
1491         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1492         if (symbol__init(NULL) < 0)
1493                 return -1;
1494
1495         sort__setup_elide(stdout);
1496
1497         get_term_dimensions(&top.winsize);
1498         if (top.print_entries == 0) {
1499                 perf_top__update_print_entries(&top);
1500                 signal(SIGWINCH, winch_sig);
1501         }
1502
1503         status = __cmd_top(&top);
1504
1505 out_delete_evlist:
1506         perf_evlist__delete(top.evlist);
1507
1508         return status;
1509 }