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/cache.h"
  26 #include "util/color.h"
  27 #include "util/evlist.h"
  28 #include "util/evsel.h"
  29 #include "util/machine.h"
  30 #include "util/session.h"
  31 #include "util/symbol.h"
  32 #include "util/thread.h"
  33 #include "util/thread_map.h"
  34 #include "util/top.h"
  35 #include "util/util.h"
  36 #include <linux/rbtree.h>
  37 #include "util/parse-options.h"
  38 #include "util/parse-events.h"
  39 #include "util/cpumap.h"
  40 #include "util/xyarray.h"
  41 #include "util/sort.h"
  42 #include "util/intlist.h"
  43 #include "arch/common.h"
  44
  45 #include "util/debug.h"
  46
  47 #include <assert.h>
  48 #include <elf.h>
  49 #include <fcntl.h>
  50
  51 #include <stdio.h>
  52 #include <termios.h>
  53 #include <unistd.h>
  54 #include <inttypes.h>
  55
  56 #include <errno.h>
  57 #include <time.h>
  58 #include <sched.h>
  59
  60 #include <sys/syscall.h>
  61 #include <sys/ioctl.h>
  62 #include <poll.h>
  63 #include <sys/prctl.h>
  64 #include <sys/wait.h>
  65 #include <sys/uio.h>
  66 #include <sys/utsname.h>
  67 #include <sys/mman.h>
  68
  69 #include <linux/unistd.h>
  70 #include <linux/types.h>
  71
  72 static volatile int done;
  73
  74 #define HEADER_LINE_NR  5
  75
  76 static void perf_top__update_print_entries(struct perf_top *top)
  77 {
  78         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
  79 }
  80
  81 static void perf_top__sig_winch(int sig __maybe_unused,
  82                                 siginfo_t *info __maybe_unused, void *arg)
  83 {
  84         struct perf_top *top = arg;
  85
  86         get_term_dimensions(&top->winsize);
  87         perf_top__update_print_entries(top);
  88 }
  89
  90 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
  91 {
  92         struct symbol *sym;
  93         struct annotation *notes;
  94         struct map *map;
  95         int err = -1;
  96
  97         if (!he || !he->ms.sym)
  98                 return -1;
  99
 100         sym = he->ms.sym;
 101         map = he->ms.map;
 102
 103         /*
 104          * We can't annotate with just /proc/kallsyms
 105          */
 106         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 107             !dso__is_kcore(map->dso)) {
 108                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 109                        "path\n", sym->name);
 110                 sleep(1);
 111                 return -1;
 112         }
 113
 114         notes = symbol__annotation(sym);
 115         if (notes->src != NULL) {
 116                 pthread_mutex_lock(&notes->lock);
 117                 goto out_assign;
 118         }
 119
 120         pthread_mutex_lock(&notes->lock);
 121
 122         if (symbol__alloc_hist(sym) < 0) {
 123                 pthread_mutex_unlock(&notes->lock);
 124                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 125                        sym->name);
 126                 sleep(1);
 127                 return err;
 128         }
 129
 130         err = symbol__annotate(sym, map, 0);
 131         if (err == 0) {
 132 out_assign:
 133                 top->sym_filter_entry = he;
 134         }
 135
 136         pthread_mutex_unlock(&notes->lock);
 137         return err;
 138 }
 139
 140 static void __zero_source_counters(struct hist_entry *he)
 141 {
 142         struct symbol *sym = he->ms.sym;
 143         symbol__annotate_zero_histograms(sym);
 144 }
 145
 146 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 147 {
 148         struct utsname uts;
 149         int err = uname(&uts);
 150
 151         ui__warning("Out of bounds address found:\n\n"
 152                     "Addr:   %" PRIx64 "\n"
 153                     "DSO:    %s %c\n"
 154                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
 155                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 156                     "Arch:   %s\n"
 157                     "Kernel: %s\n"
 158                     "Tools:  %s\n\n"
 159                     "Not all samples will be on the annotation output.\n\n"
 160                     "Please report to linux-kernel@vger.kernel.org\n",
 161                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
 162                     map->start, map->end, sym->start, sym->end,
 163                     sym->binding == STB_GLOBAL ? 'g' :
 164                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 165                     err ? "[unknown]" : uts.machine,
 166                     err ? "[unknown]" : uts.release, perf_version_string);
 167         if (use_browser <= 0)
 168                 sleep(5);
 169
 170         map->erange_warned = true;
 171 }
 172
 173 static void perf_top__record_precise_ip(struct perf_top *top,
 174                                         struct hist_entry *he,
 175                                         int counter, u64 ip)
 176 {
 177         struct annotation *notes;
 178         struct symbol *sym;
 179         int err = 0;
 180
 181         if (he == NULL || he->ms.sym == NULL ||
 182             ((top->sym_filter_entry == NULL ||
 183               top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
 184                 return;
 185
 186         sym = he->ms.sym;
 187         notes = symbol__annotation(sym);
 188
 189         if (pthread_mutex_trylock(&notes->lock))
 190                 return;
 191
 192         ip = he->ms.map->map_ip(he->ms.map, ip);
 193
 194         if (ui__has_annotation())
 195                 err = hist_entry__inc_addr_samples(he, counter, ip);
 196
 197         pthread_mutex_unlock(&notes->lock);
 198
 199         /*
 200          * This function is now called with he->hists->lock held.
 201          * Release it before going to sleep.
 202          */
 203         pthread_mutex_unlock(&he->hists->lock);
 204
 205         if (err == -ERANGE && !he->ms.map->erange_warned)
 206                 ui__warn_map_erange(he->ms.map, sym, ip);
 207         else if (err == -ENOMEM) {
 208                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 209                        sym->name);
 210                 sleep(1);
 211         }
 212
 213         pthread_mutex_lock(&he->hists->lock);
 214 }
 215
 216 static void perf_top__show_details(struct perf_top *top)
 217 {
 218         struct hist_entry *he = top->sym_filter_entry;
 219         struct annotation *notes;
 220         struct symbol *symbol;
 221         int more;
 222
 223         if (!he)
 224                 return;
 225
 226         symbol = he->ms.sym;
 227         notes = symbol__annotation(symbol);
 228
 229         pthread_mutex_lock(&notes->lock);
 230
 231         if (notes->src == NULL)
 232                 goto out_unlock;
 233
 234         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
 235         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
 236
 237         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel,
 238                                        0, top->sym_pcnt_filter, top->print_entries, 4);
 239         if (top->zero)
 240                 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 241         else
 242                 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 243         if (more != 0)
 244                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 245 out_unlock:
 246         pthread_mutex_unlock(&notes->lock);
 247 }
 248
 249 static void perf_top__print_sym_table(struct perf_top *top)
 250 {
 251         char bf[160];
 252         int printed = 0;
 253         const int win_width = top->winsize.ws_col - 1;
 254         struct hists *hists = evsel__hists(top->sym_evsel);
 255
 256         puts(CONSOLE_CLEAR);
 257
 258         perf_top__header_snprintf(top, bf, sizeof(bf));
 259         printf("%s\n", bf);
 260
 261         perf_top__reset_sample_counters(top);
 262
 263         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 264
 265         if (hists->stats.nr_lost_warned !=
 266             hists->stats.nr_events[PERF_RECORD_LOST]) {
 267                 hists->stats.nr_lost_warned =
 268                               hists->stats.nr_events[PERF_RECORD_LOST];
 269                 color_fprintf(stdout, PERF_COLOR_RED,
 270                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 271                               hists->stats.nr_lost_warned);
 272                 ++printed;
 273         }
 274
 275         if (top->sym_filter_entry) {
 276                 perf_top__show_details(top);
 277                 return;
 278         }
 279
 280         if (top->zero) {
 281                 hists__delete_entries(hists);
 282         } else {
 283                 hists__decay_entries(hists, top->hide_user_symbols,
 284                                      top->hide_kernel_symbols);
 285         }
 286
 287         hists__collapse_resort(hists, NULL);
 288         hists__output_resort(hists);
 289
 290         hists__output_recalc_col_len(hists, top->print_entries - printed);
 291         putchar('\n');
 292         hists__fprintf(hists, false, top->print_entries - printed, win_width,
 293                        top->min_percent, stdout);
 294 }
 295
 296 static void prompt_integer(int *target, const char *msg)
 297 {
 298         char *buf = malloc(0), *p;
 299         size_t dummy = 0;
 300         int tmp;
 301
 302         fprintf(stdout, "\n%s: ", msg);
 303         if (getline(&buf, &dummy, stdin) < 0)
 304                 return;
 305
 306         p = strchr(buf, '\n');
 307         if (p)
 308                 *p = 0;
 309
 310         p = buf;
 311         while(*p) {
 312                 if (!isdigit(*p))
 313                         goto out_free;
 314                 p++;
 315         }
 316         tmp = strtoul(buf, NULL, 10);
 317         *target = tmp;
 318 out_free:
 319         free(buf);
 320 }
 321
 322 static void prompt_percent(int *target, const char *msg)
 323 {
 324         int tmp = 0;
 325
 326         prompt_integer(&tmp, msg);
 327         if (tmp >= 0 && tmp <= 100)
 328                 *target = tmp;
 329 }
 330
 331 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 332 {
 333         char *buf = malloc(0), *p;
 334         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 335         struct hists *hists = evsel__hists(top->sym_evsel);
 336         struct rb_node *next;
 337         size_t dummy = 0;
 338
 339         /* zero counters of active symbol */
 340         if (syme) {
 341                 __zero_source_counters(syme);
 342                 top->sym_filter_entry = NULL;
 343         }
 344
 345         fprintf(stdout, "\n%s: ", msg);
 346         if (getline(&buf, &dummy, stdin) < 0)
 347                 goto out_free;
 348
 349         p = strchr(buf, '\n');
 350         if (p)
 351                 *p = 0;
 352
 353         next = rb_first(&hists->entries);
 354         while (next) {
 355                 n = rb_entry(next, struct hist_entry, rb_node);
 356                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 357                         found = n;
 358                         break;
 359                 }
 360                 next = rb_next(&n->rb_node);
 361         }
 362
 363         if (!found) {
 364                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 365                 sleep(1);
 366         } else
 367                 perf_top__parse_source(top, found);
 368
 369 out_free:
 370         free(buf);
 371 }
 372
 373 static void perf_top__print_mapped_keys(struct perf_top *top)
 374 {
 375         char *name = NULL;
 376
 377         if (top->sym_filter_entry) {
 378                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 379                 name = sym->name;
 380         }
 381
 382         fprintf(stdout, "\nMapped keys:\n");
 383         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 384         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 385
 386         if (top->evlist->nr_entries > 1)
 387                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 388
 389         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 390
 391         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
 392         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 393         fprintf(stdout, "\t[S]     stop annotation.\n");
 394
 395         fprintf(stdout,
 396                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 397                 top->hide_kernel_symbols ? "yes" : "no");
 398         fprintf(stdout,
 399                 "\t[U]     hide user symbols.               \t(%s)\n",
 400                 top->hide_user_symbols ? "yes" : "no");
 401         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 402         fprintf(stdout, "\t[qQ]    quit.\n");
 403 }
 404
 405 static int perf_top__key_mapped(struct perf_top *top, int c)
 406 {
 407         switch (c) {
 408                 case 'd':
 409                 case 'e':
 410                 case 'f':
 411                 case 'z':
 412                 case 'q':
 413                 case 'Q':
 414                 case 'K':
 415                 case 'U':
 416                 case 'F':
 417                 case 's':
 418                 case 'S':
 419                         return 1;
 420                 case 'E':
 421                         return top->evlist->nr_entries > 1 ? 1 : 0;
 422                 default:
 423                         break;
 424         }
 425
 426         return 0;
 427 }
 428
 429 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 430 {
 431         bool ret = true;
 432
 433         if (!perf_top__key_mapped(top, c)) {
 434                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 435                 struct termios save;
 436
 437                 perf_top__print_mapped_keys(top);
 438                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 439                 fflush(stdout);
 440
 441                 set_term_quiet_input(&save);
 442
 443                 poll(&stdin_poll, 1, -1);
 444                 c = getc(stdin);
 445
 446                 tcsetattr(0, TCSAFLUSH, &save);
 447                 if (!perf_top__key_mapped(top, c))
 448                         return ret;
 449         }
 450
 451         switch (c) {
 452                 case 'd':
 453                         prompt_integer(&top->delay_secs, "Enter display delay");
 454                         if (top->delay_secs < 1)
 455                                 top->delay_secs = 1;
 456                         break;
 457                 case 'e':
 458                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 459                         if (top->print_entries == 0) {
 460                                 struct sigaction act = {
 461                                         .sa_sigaction = perf_top__sig_winch,
 462                                         .sa_flags     = SA_SIGINFO,
 463                                 };
 464                                 perf_top__sig_winch(SIGWINCH, NULL, top);
 465                                 sigaction(SIGWINCH, &act, NULL);
 466                         } else {
 467                                 signal(SIGWINCH, SIG_DFL);
 468                         }
 469                         break;
 470                 case 'E':
 471                         if (top->evlist->nr_entries > 1) {
 472                                 /* Select 0 as the default event: */
 473                                 int counter = 0;
 474
 475                                 fprintf(stderr, "\nAvailable events:");
 476
 477                                 evlist__for_each(top->evlist, top->sym_evsel)
 478                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
 479
 480                                 prompt_integer(&counter, "Enter details event counter");
 481
 482                                 if (counter >= top->evlist->nr_entries) {
 483                                         top->sym_evsel = perf_evlist__first(top->evlist);
 484                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
 485                                         sleep(1);
 486                                         break;
 487                                 }
 488                                 evlist__for_each(top->evlist, top->sym_evsel)
 489                                         if (top->sym_evsel->idx == counter)
 490                                                 break;
 491                         } else
 492                                 top->sym_evsel = perf_evlist__first(top->evlist);
 493                         break;
 494                 case 'f':
 495                         prompt_integer(&top->count_filter, "Enter display event count filter");
 496                         break;
 497                 case 'F':
 498                         prompt_percent(&top->sym_pcnt_filter,
 499                                        "Enter details display event filter (percent)");
 500                         break;
 501                 case 'K':
 502                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 503                         break;
 504                 case 'q':
 505                 case 'Q':
 506                         printf("exiting.\n");
 507                         if (top->dump_symtab)
 508                                 perf_session__fprintf_dsos(top->session, stderr);
 509                         ret = false;
 510                         break;
 511                 case 's':
 512                         perf_top__prompt_symbol(top, "Enter details symbol");
 513                         break;
 514                 case 'S':
 515                         if (!top->sym_filter_entry)
 516                                 break;
 517                         else {
 518                                 struct hist_entry *syme = top->sym_filter_entry;
 519
 520                                 top->sym_filter_entry = NULL;
 521                                 __zero_source_counters(syme);
 522                         }
 523                         break;
 524                 case 'U':
 525                         top->hide_user_symbols = !top->hide_user_symbols;
 526                         break;
 527                 case 'z':
 528                         top->zero = !top->zero;
 529                         break;
 530                 default:
 531                         break;
 532         }
 533
 534         return ret;
 535 }
 536
 537 static void perf_top__sort_new_samples(void *arg)
 538 {
 539         struct perf_top *t = arg;
 540         struct hists *hists;
 541
 542         perf_top__reset_sample_counters(t);
 543
 544         if (t->evlist->selected != NULL)
 545                 t->sym_evsel = t->evlist->selected;
 546
 547         hists = evsel__hists(t->sym_evsel);
 548
 549         if (t->zero) {
 550                 hists__delete_entries(hists);
 551         } else {
 552                 hists__decay_entries(hists, t->hide_user_symbols,
 553                                      t->hide_kernel_symbols);
 554         }
 555
 556         hists__collapse_resort(hists, NULL);
 557         hists__output_resort(hists);
 558 }
 559
 560 static void *display_thread_tui(void *arg)
 561 {
 562         struct perf_evsel *pos;
 563         struct perf_top *top = arg;
 564         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 565         struct hist_browser_timer hbt = {
 566                 .timer          = perf_top__sort_new_samples,
 567                 .arg            = top,
 568                 .refresh        = top->delay_secs,
 569         };
 570
 571         perf_top__sort_new_samples(top);
 572
 573         /*
 574          * Initialize the uid_filter_str, in the future the TUI will allow
 575          * Zooming in/out UIDs. For now juse use whatever the user passed
 576          * via --uid.
 577          */
 578         evlist__for_each(top->evlist, pos) {
 579                 struct hists *hists = evsel__hists(pos);
 580                 hists->uid_filter_str = top->record_opts.target.uid_str;
 581         }
 582
 583         perf_evlist__tui_browse_hists(top->evlist, help, &hbt, top->min_percent,
 584                                       &top->session->header.env);
 585
 586         done = 1;
 587         return NULL;
 588 }
 589
 590 static void display_sig(int sig __maybe_unused)
 591 {
 592         done = 1;
 593 }
 594
 595 static void display_setup_sig(void)
 596 {
 597         signal(SIGSEGV, display_sig);
 598         signal(SIGFPE,  display_sig);
 599         signal(SIGINT,  display_sig);
 600         signal(SIGQUIT, display_sig);
 601         signal(SIGTERM, display_sig);
 602 }
 603
 604 static void *display_thread(void *arg)
 605 {
 606         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 607         struct termios save;
 608         struct perf_top *top = arg;
 609         int delay_msecs, c;
 610
 611         display_setup_sig();
 612         pthread__unblock_sigwinch();
 613 repeat:
 614         delay_msecs = top->delay_secs * 1000;
 615         set_term_quiet_input(&save);
 616         /* trash return*/
 617         getc(stdin);
 618
 619         while (!done) {
 620                 perf_top__print_sym_table(top);
 621                 /*
 622                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 623                  * refresh screen in both cases.
 624                  */
 625                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 626                 case 0:
 627                         continue;
 628                 case -1:
 629                         if (errno == EINTR)
 630                                 continue;
 631                         /* Fall trhu */
 632                 default:
 633                         c = getc(stdin);
 634                         tcsetattr(0, TCSAFLUSH, &save);
 635
 636                         if (perf_top__handle_keypress(top, c))
 637                                 goto repeat;
 638                         done = 1;
 639                 }
 640         }
 641
 642         tcsetattr(0, TCSAFLUSH, &save);
 643         return NULL;
 644 }
 645
 646 static int symbol_filter(struct map *map, struct symbol *sym)
 647 {
 648         const char *name = sym->name;
 649
 650         if (!map->dso->kernel)
 651                 return 0;
 652         /*
 653          * ppc64 uses function descriptors and appends a '.' to the
 654          * start of every instruction address. Remove it.
 655          */
 656         if (name[0] == '.')
 657                 name++;
 658
 659         if (!strcmp(name, "_text") ||
 660             !strcmp(name, "_etext") ||
 661             !strcmp(name, "_sinittext") ||
 662             !strncmp("init_module", name, 11) ||
 663             !strncmp("cleanup_module", name, 14) ||
 664             strstr(name, "_text_start") ||
 665             strstr(name, "_text_end"))
 666                 return 1;
 667
 668         if (symbol__is_idle(sym))
 669                 sym->ignore = true;
 670
 671         return 0;
 672 }
 673
 674 static int hist_iter__top_callback(struct hist_entry_iter *iter,
 675                                    struct addr_location *al, bool single,
 676                                    void *arg)
 677 {
 678         struct perf_top *top = arg;
 679         struct hist_entry *he = iter->he;
 680         struct perf_evsel *evsel = iter->evsel;
 681
 682         if (sort__has_sym && single) {
 683                 u64 ip = al->addr;
 684
 685                 if (al->map)
 686                         ip = al->map->unmap_ip(al->map, ip);
 687
 688                 perf_top__record_precise_ip(top, he, evsel->idx, ip);
 689         }
 690
 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->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 (perf_event__preprocess_sample(event, machine, &al, sample) < 0)
 728                 return;
 729
 730         if (!top->kptr_restrict_warned &&
 731             symbol_conf.kptr_restrict &&
 732             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 733                 ui__warning(
 734 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 735 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 736 "Kernel%s samples will not be resolved.\n",
 737                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 738                           " modules" : "");
 739                 if (use_browser <= 0)
 740                         sleep(5);
 741                 top->kptr_restrict_warned = true;
 742         }
 743
 744         if (al.sym == NULL) {
 745                 const char *msg = "Kernel samples will not be resolved.\n";
 746                 /*
 747                  * As we do lazy loading of symtabs we only will know if the
 748                  * specified vmlinux file is invalid when we actually have a
 749                  * hit in kernel space and then try to load it. So if we get
 750                  * here and there are _no_ symbols in the DSO backing the
 751                  * kernel map, bail out.
 752                  *
 753                  * We may never get here, for instance, if we use -K/
 754                  * --hide-kernel-symbols, even if the user specifies an
 755                  * invalid --vmlinux ;-)
 756                  */
 757                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
 758                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 759                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 760                         if (symbol_conf.vmlinux_name) {
 761                                 ui__warning("The %s file can't be used.\n%s",
 762                                             symbol_conf.vmlinux_name, msg);
 763                         } else {
 764                                 ui__warning("A vmlinux file was not found.\n%s",
 765                                             msg);
 766                         }
 767
 768                         if (use_browser <= 0)
 769                                 sleep(5);
 770                         top->vmlinux_warned = true;
 771                 }
 772         }
 773
 774         if (al.sym == NULL || !al.sym->ignore) {
 775                 struct hists *hists = evsel__hists(evsel);
 776                 struct hist_entry_iter iter = {
 777                         .add_entry_cb = hist_iter__top_callback,
 778                 };
 779
 780                 if (symbol_conf.cumulate_callchain)
 781                         iter.ops = &hist_iter_cumulative;
 782                 else
 783                         iter.ops = &hist_iter_normal;
 784
 785                 pthread_mutex_lock(&hists->lock);
 786
 787                 err = hist_entry_iter__add(&iter, &al, evsel, sample,
 788                                            top->max_stack, top);
 789                 if (err < 0)
 790                         pr_err("Problem incrementing symbol period, skipping event\n");
 791
 792                 pthread_mutex_unlock(&hists->lock);
 793         }
 794
 795         return;
 796 }
 797
 798 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 799 {
 800         struct perf_sample sample;
 801         struct perf_evsel *evsel;
 802         struct perf_session *session = top->session;
 803         union perf_event *event;
 804         struct machine *machine;
 805         u8 origin;
 806         int ret;
 807
 808         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
 809                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
 810                 if (ret) {
 811                         pr_err("Can't parse sample, err = %d\n", ret);
 812                         goto next_event;
 813                 }
 814
 815                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 816                 assert(evsel != NULL);
 817
 818                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 819
 820                 if (event->header.type == PERF_RECORD_SAMPLE)
 821                         ++top->samples;
 822
 823                 switch (origin) {
 824                 case PERF_RECORD_MISC_USER:
 825                         ++top->us_samples;
 826                         if (top->hide_user_symbols)
 827                                 goto next_event;
 828                         machine = &session->machines.host;
 829                         break;
 830                 case PERF_RECORD_MISC_KERNEL:
 831                         ++top->kernel_samples;
 832                         if (top->hide_kernel_symbols)
 833                                 goto next_event;
 834                         machine = &session->machines.host;
 835                         break;
 836                 case PERF_RECORD_MISC_GUEST_KERNEL:
 837                         ++top->guest_kernel_samples;
 838                         machine = perf_session__find_machine(session,
 839                                                              sample.pid);
 840                         break;
 841                 case PERF_RECORD_MISC_GUEST_USER:
 842                         ++top->guest_us_samples;
 843                         /*
 844                          * TODO: we don't process guest user from host side
 845                          * except simple counting.
 846                          */
 847                         /* Fall thru */
 848                 default:
 849                         goto next_event;
 850                 }
 851
 852
 853                 if (event->header.type == PERF_RECORD_SAMPLE) {
 854                         perf_event__process_sample(&top->tool, event, evsel,
 855                                                    &sample, machine);
 856                 } else if (event->header.type < PERF_RECORD_MAX) {
 857                         hists__inc_nr_events(evsel__hists(evsel), event->header.type);
 858                         machine__process_event(machine, event, &sample);
 859                 } else
 860                         ++session->stats.nr_unknown_events;
 861 next_event:
 862                 perf_evlist__mmap_consume(top->evlist, idx);
 863         }
 864 }
 865
 866 static void perf_top__mmap_read(struct perf_top *top)
 867 {
 868         int i;
 869
 870         for (i = 0; i < top->evlist->nr_mmaps; i++)
 871                 perf_top__mmap_read_idx(top, i);
 872 }
 873
 874 static int perf_top__start_counters(struct perf_top *top)
 875 {
 876         char msg[512];
 877         struct perf_evsel *counter;
 878         struct perf_evlist *evlist = top->evlist;
 879         struct record_opts *opts = &top->record_opts;
 880
 881         perf_evlist__config(evlist, opts);
 882
 883         evlist__for_each(evlist, counter) {
 884 try_again:
 885                 if (perf_evsel__open(counter, top->evlist->cpus,
 886                                      top->evlist->threads) < 0) {
 887                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 888                                 if (verbose)
 889                                         ui__warning("%s\n", msg);
 890                                 goto try_again;
 891                         }
 892
 893                         perf_evsel__open_strerror(counter, &opts->target,
 894                                                   errno, msg, sizeof(msg));
 895                         ui__error("%s\n", msg);
 896                         goto out_err;
 897                 }
 898         }
 899
 900         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 901                 ui__error("Failed to mmap with %d (%s)\n",
 902                             errno, strerror_r(errno, msg, sizeof(msg)));
 903                 goto out_err;
 904         }
 905
 906         return 0;
 907
 908 out_err:
 909         return -1;
 910 }
 911
 912 static int perf_top__setup_sample_type(struct perf_top *top __maybe_unused)
 913 {
 914         if (!sort__has_sym) {
 915                 if (symbol_conf.use_callchain) {
 916                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
 917                         return -EINVAL;
 918                 }
 919         } else if (callchain_param.mode != CHAIN_NONE) {
 920                 if (callchain_register_param(&callchain_param) < 0) {
 921                         ui__error("Can't register callchain params.\n");
 922                         return -EINVAL;
 923                 }
 924         }
 925
 926         return 0;
 927 }
 928
 929 static int __cmd_top(struct perf_top *top)
 930 {
 931         struct record_opts *opts = &top->record_opts;
 932         pthread_t thread;
 933         int ret;
 934
 935         top->session = perf_session__new(NULL, false, NULL);
 936         if (top->session == NULL)
 937                 return -1;
 938
 939         machines__set_symbol_filter(&top->session->machines, symbol_filter);
 940
 941         if (!objdump_path) {
 942                 ret = perf_session_env__lookup_objdump(&top->session->header.env);
 943                 if (ret)
 944                         goto out_delete;
 945         }
 946
 947         ret = perf_top__setup_sample_type(top);
 948         if (ret)
 949                 goto out_delete;
 950
 951         machine__synthesize_threads(&top->session->machines.host, &opts->target,
 952                                     top->evlist->threads, false);
 953         ret = perf_top__start_counters(top);
 954         if (ret)
 955                 goto out_delete;
 956
 957         top->session->evlist = top->evlist;
 958         perf_session__set_id_hdr_size(top->session);
 959
 960         /*
 961          * When perf is starting the traced process, all the events (apart from
 962          * group members) have enable_on_exec=1 set, so don't spoil it by
 963          * prematurely enabling them.
 964          *
 965          * XXX 'top' still doesn't start workloads like record, trace, but should,
 966          * so leave the check here.
 967          */
 968         if (!target__none(&opts->target))
 969                 perf_evlist__enable(top->evlist);
 970
 971         /* Wait for a minimal set of events before starting the snapshot */
 972         perf_evlist__poll(top->evlist, 100);
 973
 974         perf_top__mmap_read(top);
 975
 976         ret = -1;
 977         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 978                                                             display_thread), top)) {
 979                 ui__error("Could not create display thread.\n");
 980                 goto out_delete;
 981         }
 982
 983         if (top->realtime_prio) {
 984                 struct sched_param param;
 985
 986                 param.sched_priority = top->realtime_prio;
 987                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 988                         ui__error("Could not set realtime priority.\n");
 989                         goto out_join;
 990                 }
 991         }
 992
 993         while (!done) {
 994                 u64 hits = top->samples;
 995
 996                 perf_top__mmap_read(top);
 997
 998                 if (hits == top->samples)
 999                         ret = perf_evlist__poll(top->evlist, 100);
1000         }
1001
1002         ret = 0;
1003 out_join:
1004         pthread_join(thread, NULL);
1005 out_delete:
1006         perf_session__delete(top->session);
1007         top->session = NULL;
1008
1009         return ret;
1010 }
1011
1012 static int
1013 callchain_opt(const struct option *opt, const char *arg, int unset)
1014 {
1015         symbol_conf.use_callchain = true;
1016         return record_callchain_opt(opt, arg, unset);
1017 }
1018
1019 static int
1020 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1021 {
1022         symbol_conf.use_callchain = true;
1023         return record_parse_callchain_opt(opt, arg, unset);
1024 }
1025
1026 static int perf_top_config(const char *var, const char *value, void *cb)
1027 {
1028         if (!strcmp(var, "top.call-graph"))
1029                 var = "call-graph.record-mode"; /* fall-through */
1030         if (!strcmp(var, "top.children")) {
1031                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1032                 return 0;
1033         }
1034
1035         return perf_default_config(var, value, cb);
1036 }
1037
1038 static int
1039 parse_percent_limit(const struct option *opt, const char *arg,
1040                     int unset __maybe_unused)
1041 {
1042         struct perf_top *top = opt->value;
1043
1044         top->min_percent = strtof(arg, NULL);
1045         return 0;
1046 }
1047
1048 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1049 {
1050         char errbuf[BUFSIZ];
1051         struct perf_top top = {
1052                 .count_filter        = 5,
1053                 .delay_secs          = 2,
1054                 .record_opts = {
1055                         .mmap_pages     = UINT_MAX,
1056                         .user_freq      = UINT_MAX,
1057                         .user_interval  = ULLONG_MAX,
1058                         .freq           = 4000, /* 4 KHz */
1059                         .target         = {
1060                                 .uses_mmap   = true,
1061                         },
1062                 },
1063                 .max_stack           = PERF_MAX_STACK_DEPTH,
1064                 .sym_pcnt_filter     = 5,
1065         };
1066         struct record_opts *opts = &top.record_opts;
1067         struct target *target = &opts->target;
1068         const struct option options[] = {
1069         OPT_CALLBACK('e', "event", &top.evlist, "event",
1070                      "event selector. use 'perf list' to list available events",
1071                      parse_events_option),
1072         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1073         OPT_STRING('p', "pid", &target->pid, "pid",
1074                     "profile events on existing process id"),
1075         OPT_STRING('t', "tid", &target->tid, "tid",
1076                     "profile events on existing thread id"),
1077         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1078                             "system-wide collection from all CPUs"),
1079         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1080                     "list of cpus to monitor"),
1081         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1082                    "file", "vmlinux pathname"),
1083         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1084                     "don't load vmlinux even if found"),
1085         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1086                     "hide kernel symbols"),
1087         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1088                      "number of mmap data pages",
1089                      perf_evlist__parse_mmap_pages),
1090         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1091                     "collect data with this RT SCHED_FIFO priority"),
1092         OPT_INTEGER('d', "delay", &top.delay_secs,
1093                     "number of seconds to delay between refreshes"),
1094         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1095                             "dump the symbol table used for profiling"),
1096         OPT_INTEGER('f', "count-filter", &top.count_filter,
1097                     "only display functions with more events than this"),
1098         OPT_BOOLEAN(0, "group", &opts->group,
1099                             "put the counters into a counter group"),
1100         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1101                     "child tasks do not inherit counters"),
1102         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1103                     "symbol to annotate"),
1104         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1105         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1106         OPT_INTEGER('E', "entries", &top.print_entries,
1107                     "display this many functions"),
1108         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1109                     "hide user symbols"),
1110         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1111         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1112         OPT_INCR('v', "verbose", &verbose,
1113                     "be more verbose (show counter open errors, etc)"),
1114         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1115                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1116                    " Please refer the man page for the complete list."),
1117         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1118                    "output field(s): overhead, period, sample plus all of sort keys"),
1119         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1120                     "Show a column with the number of samples"),
1121         OPT_CALLBACK_NOOPT('g', NULL, &top.record_opts,
1122                            NULL, "enables call-graph recording",
1123                            &callchain_opt),
1124         OPT_CALLBACK(0, "call-graph", &top.record_opts,
1125                      "mode[,dump_size]", record_callchain_help,
1126                      &parse_callchain_opt),
1127         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1128                     "Accumulate callchains of children and show total overhead as well"),
1129         OPT_INTEGER(0, "max-stack", &top.max_stack,
1130                     "Set the maximum stack depth when parsing the callchain. "
1131                     "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
1132         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1133                    "ignore callees of these functions in call graphs",
1134                    report_parse_ignore_callees_opt),
1135         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1136                     "Show a column with the sum of periods"),
1137         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1138                    "only consider symbols in these dsos"),
1139         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1140                    "only consider symbols in these comms"),
1141         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1142                    "only consider these symbols"),
1143         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1144                     "Interleave source code with assembly code (default)"),
1145         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1146                     "Display raw encoding of assembly instructions (default)"),
1147         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1148                     "Enable kernel symbol demangling"),
1149         OPT_STRING(0, "objdump", &objdump_path, "path",
1150                     "objdump binary to use for disassembly and annotations"),
1151         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1152                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1153         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1154         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1155                      "Don't show entries under that percent", parse_percent_limit),
1156         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1157                      "How to display percentage of filtered entries", parse_filter_percentage),
1158         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1159                    "width[,width...]",
1160                    "don't try to adjust column width, use these fixed values"),
1161         OPT_END()
1162         };
1163         const char * const top_usage[] = {
1164                 "perf top [<options>]",
1165                 NULL
1166         };
1167         int status = hists__init();
1168
1169         if (status < 0)
1170                 return status;
1171
1172         top.evlist = perf_evlist__new();
1173         if (top.evlist == NULL)
1174                 return -ENOMEM;
1175
1176         perf_config(perf_top_config, &top);
1177
1178         argc = parse_options(argc, argv, options, top_usage, 0);
1179         if (argc)
1180                 usage_with_options(top_usage, options);
1181
1182         sort__mode = SORT_MODE__TOP;
1183         /* display thread wants entries to be collapsed in a different tree */
1184         sort__need_collapse = 1;
1185
1186         if (setup_sorting() < 0) {
1187                 if (sort_order)
1188                         parse_options_usage(top_usage, options, "s", 1);
1189                 if (field_order)
1190                         parse_options_usage(sort_order ? NULL : top_usage,
1191                                             options, "fields", 0);
1192                 goto out_delete_evlist;
1193         }
1194
1195         if (top.use_stdio)
1196                 use_browser = 0;
1197         else if (top.use_tui)
1198                 use_browser = 1;
1199
1200         setup_browser(false);
1201
1202         status = target__validate(target);
1203         if (status) {
1204                 target__strerror(target, status, errbuf, BUFSIZ);
1205                 ui__warning("%s\n", errbuf);
1206         }
1207
1208         status = target__parse_uid(target);
1209         if (status) {
1210                 int saved_errno = errno;
1211
1212                 target__strerror(target, status, errbuf, BUFSIZ);
1213                 ui__error("%s\n", errbuf);
1214
1215                 status = -saved_errno;
1216                 goto out_delete_evlist;
1217         }
1218
1219         if (target__none(target))
1220                 target->system_wide = true;
1221
1222         if (perf_evlist__create_maps(top.evlist, target) < 0)
1223                 usage_with_options(top_usage, options);
1224
1225         if (!top.evlist->nr_entries &&
1226             perf_evlist__add_default(top.evlist) < 0) {
1227                 ui__error("Not enough memory for event selector list\n");
1228                 goto out_delete_evlist;
1229         }
1230
1231         symbol_conf.nr_events = top.evlist->nr_entries;
1232
1233         if (top.delay_secs < 1)
1234                 top.delay_secs = 1;
1235
1236         if (record_opts__config(opts)) {
1237                 status = -EINVAL;
1238                 goto out_delete_evlist;
1239         }
1240
1241         top.sym_evsel = perf_evlist__first(top.evlist);
1242
1243         if (!symbol_conf.use_callchain) {
1244                 symbol_conf.cumulate_callchain = false;
1245                 perf_hpp__cancel_cumulate();
1246         }
1247
1248         symbol_conf.priv_size = sizeof(struct annotation);
1249
1250         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1251         if (symbol__init(NULL) < 0)
1252                 return -1;
1253
1254         sort__setup_elide(stdout);
1255
1256         get_term_dimensions(&top.winsize);
1257         if (top.print_entries == 0) {
1258                 struct sigaction act = {
1259                         .sa_sigaction = perf_top__sig_winch,
1260                         .sa_flags     = SA_SIGINFO,
1261                 };
1262                 perf_top__update_print_entries(&top);
1263                 sigaction(SIGWINCH, &act, NULL);
1264         }
1265
1266         status = __cmd_top(&top);
1267
1268 out_delete_evlist:
1269         perf_evlist__delete(top.evlist);
1270
1271         return status;
1272 }