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