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 <sys/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
 255         puts(CONSOLE_CLEAR);
 256
 257         perf_top__header_snprintf(top, bf, sizeof(bf));
 258         printf("%s\n", bf);
 259
 260         perf_top__reset_sample_counters(top);
 261
 262         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 263
 264         if (top->sym_evsel->hists.stats.nr_lost_warned !=
 265             top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
 266                 top->sym_evsel->hists.stats.nr_lost_warned =
 267                         top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
 268                 color_fprintf(stdout, PERF_COLOR_RED,
 269                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 270                               top->sym_evsel->hists.stats.nr_lost_warned);
 271                 ++printed;
 272         }
 273
 274         if (top->sym_filter_entry) {
 275                 perf_top__show_details(top);
 276                 return;
 277         }
 278
 279         hists__collapse_resort(&top->sym_evsel->hists, NULL);
 280         hists__output_resort(&top->sym_evsel->hists);
 281         hists__decay_entries(&top->sym_evsel->hists,
 282                              top->hide_user_symbols,
 283                              top->hide_kernel_symbols);
 284         hists__output_recalc_col_len(&top->sym_evsel->hists,
 285                                      top->print_entries - printed);
 286         putchar('\n');
 287         hists__fprintf(&top->sym_evsel->hists, false,
 288                        top->print_entries - printed, win_width,
 289                        top->min_percent, stdout);
 290 }
 291
 292 static void prompt_integer(int *target, const char *msg)
 293 {
 294         char *buf = malloc(0), *p;
 295         size_t dummy = 0;
 296         int tmp;
 297
 298         fprintf(stdout, "\n%s: ", msg);
 299         if (getline(&buf, &dummy, stdin) < 0)
 300                 return;
 301
 302         p = strchr(buf, '\n');
 303         if (p)
 304                 *p = 0;
 305
 306         p = buf;
 307         while(*p) {
 308                 if (!isdigit(*p))
 309                         goto out_free;
 310                 p++;
 311         }
 312         tmp = strtoul(buf, NULL, 10);
 313         *target = tmp;
 314 out_free:
 315         free(buf);
 316 }
 317
 318 static void prompt_percent(int *target, const char *msg)
 319 {
 320         int tmp = 0;
 321
 322         prompt_integer(&tmp, msg);
 323         if (tmp >= 0 && tmp <= 100)
 324                 *target = tmp;
 325 }
 326
 327 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 328 {
 329         char *buf = malloc(0), *p;
 330         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 331         struct rb_node *next;
 332         size_t dummy = 0;
 333
 334         /* zero counters of active symbol */
 335         if (syme) {
 336                 __zero_source_counters(syme);
 337                 top->sym_filter_entry = NULL;
 338         }
 339
 340         fprintf(stdout, "\n%s: ", msg);
 341         if (getline(&buf, &dummy, stdin) < 0)
 342                 goto out_free;
 343
 344         p = strchr(buf, '\n');
 345         if (p)
 346                 *p = 0;
 347
 348         next = rb_first(&top->sym_evsel->hists.entries);
 349         while (next) {
 350                 n = rb_entry(next, struct hist_entry, rb_node);
 351                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 352                         found = n;
 353                         break;
 354                 }
 355                 next = rb_next(&n->rb_node);
 356         }
 357
 358         if (!found) {
 359                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 360                 sleep(1);
 361         } else
 362                 perf_top__parse_source(top, found);
 363
 364 out_free:
 365         free(buf);
 366 }
 367
 368 static void perf_top__print_mapped_keys(struct perf_top *top)
 369 {
 370         char *name = NULL;
 371
 372         if (top->sym_filter_entry) {
 373                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 374                 name = sym->name;
 375         }
 376
 377         fprintf(stdout, "\nMapped keys:\n");
 378         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 379         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 380
 381         if (top->evlist->nr_entries > 1)
 382                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 383
 384         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 385
 386         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
 387         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 388         fprintf(stdout, "\t[S]     stop annotation.\n");
 389
 390         fprintf(stdout,
 391                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 392                 top->hide_kernel_symbols ? "yes" : "no");
 393         fprintf(stdout,
 394                 "\t[U]     hide user symbols.               \t(%s)\n",
 395                 top->hide_user_symbols ? "yes" : "no");
 396         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 397         fprintf(stdout, "\t[qQ]    quit.\n");
 398 }
 399
 400 static int perf_top__key_mapped(struct perf_top *top, int c)
 401 {
 402         switch (c) {
 403                 case 'd':
 404                 case 'e':
 405                 case 'f':
 406                 case 'z':
 407                 case 'q':
 408                 case 'Q':
 409                 case 'K':
 410                 case 'U':
 411                 case 'F':
 412                 case 's':
 413                 case 'S':
 414                         return 1;
 415                 case 'E':
 416                         return top->evlist->nr_entries > 1 ? 1 : 0;
 417                 default:
 418                         break;
 419         }
 420
 421         return 0;
 422 }
 423
 424 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 425 {
 426         bool ret = true;
 427
 428         if (!perf_top__key_mapped(top, c)) {
 429                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 430                 struct termios tc, save;
 431
 432                 perf_top__print_mapped_keys(top);
 433                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 434                 fflush(stdout);
 435
 436                 tcgetattr(0, &save);
 437                 tc = save;
 438                 tc.c_lflag &= ~(ICANON | ECHO);
 439                 tc.c_cc[VMIN] = 0;
 440                 tc.c_cc[VTIME] = 0;
 441                 tcsetattr(0, TCSANOW, &tc);
 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         perf_top__reset_sample_counters(t);
 541
 542         if (t->evlist->selected != NULL)
 543                 t->sym_evsel = t->evlist->selected;
 544
 545         hists__collapse_resort(&t->sym_evsel->hists, NULL);
 546         hists__output_resort(&t->sym_evsel->hists);
 547         hists__decay_entries(&t->sym_evsel->hists,
 548                              t->hide_user_symbols,
 549                              t->hide_kernel_symbols);
 550 }
 551
 552 static void *display_thread_tui(void *arg)
 553 {
 554         struct perf_evsel *pos;
 555         struct perf_top *top = arg;
 556         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 557         struct hist_browser_timer hbt = {
 558                 .timer          = perf_top__sort_new_samples,
 559                 .arg            = top,
 560                 .refresh        = top->delay_secs,
 561         };
 562
 563         perf_top__sort_new_samples(top);
 564
 565         /*
 566          * Initialize the uid_filter_str, in the future the TUI will allow
 567          * Zooming in/out UIDs. For now juse use whatever the user passed
 568          * via --uid.
 569          */
 570         evlist__for_each(top->evlist, pos)
 571                 pos->hists.uid_filter_str = top->record_opts.target.uid_str;
 572
 573         perf_evlist__tui_browse_hists(top->evlist, help, &hbt, top->min_percent,
 574                                       &top->session->header.env);
 575
 576         done = 1;
 577         return NULL;
 578 }
 579
 580 static void display_sig(int sig __maybe_unused)
 581 {
 582         done = 1;
 583 }
 584
 585 static void display_setup_sig(void)
 586 {
 587         signal(SIGSEGV, display_sig);
 588         signal(SIGFPE,  display_sig);
 589         signal(SIGINT,  display_sig);
 590         signal(SIGQUIT, display_sig);
 591         signal(SIGTERM, display_sig);
 592 }
 593
 594 static void *display_thread(void *arg)
 595 {
 596         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 597         struct termios save;
 598         struct perf_top *top = arg;
 599         int delay_msecs, c;
 600
 601         display_setup_sig();
 602         pthread__unblock_sigwinch();
 603 repeat:
 604         delay_msecs = top->delay_secs * 1000;
 605         set_term_quiet_input(&save);
 606         /* trash return*/
 607         getc(stdin);
 608
 609         while (!done) {
 610                 perf_top__print_sym_table(top);
 611                 /*
 612                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 613                  * refresh screen in both cases.
 614                  */
 615                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 616                 case 0:
 617                         continue;
 618                 case -1:
 619                         if (errno == EINTR)
 620                                 continue;
 621                         /* Fall trhu */
 622                 default:
 623                         c = getc(stdin);
 624                         tcsetattr(0, TCSAFLUSH, &save);
 625
 626                         if (perf_top__handle_keypress(top, c))
 627                                 goto repeat;
 628                         done = 1;
 629                 }
 630         }
 631
 632         tcsetattr(0, TCSAFLUSH, &save);
 633         return NULL;
 634 }
 635
 636 static int symbol_filter(struct map *map, struct symbol *sym)
 637 {
 638         const char *name = sym->name;
 639
 640         if (!map->dso->kernel)
 641                 return 0;
 642         /*
 643          * ppc64 uses function descriptors and appends a '.' to the
 644          * start of every instruction address. Remove it.
 645          */
 646         if (name[0] == '.')
 647                 name++;
 648
 649         if (!strcmp(name, "_text") ||
 650             !strcmp(name, "_etext") ||
 651             !strcmp(name, "_sinittext") ||
 652             !strncmp("init_module", name, 11) ||
 653             !strncmp("cleanup_module", name, 14) ||
 654             strstr(name, "_text_start") ||
 655             strstr(name, "_text_end"))
 656                 return 1;
 657
 658         if (symbol__is_idle(sym))
 659                 sym->ignore = true;
 660
 661         return 0;
 662 }
 663
 664 static int hist_iter__top_callback(struct hist_entry_iter *iter,
 665                                    struct addr_location *al, bool single,
 666                                    void *arg)
 667 {
 668         struct perf_top *top = arg;
 669         struct hist_entry *he = iter->he;
 670         struct perf_evsel *evsel = iter->evsel;
 671
 672         if (sort__has_sym && single) {
 673                 u64 ip = al->addr;
 674
 675                 if (al->map)
 676                         ip = al->map->unmap_ip(al->map, ip);
 677
 678                 perf_top__record_precise_ip(top, he, evsel->idx, ip);
 679         }
 680
 681         return 0;
 682 }
 683
 684 static void perf_event__process_sample(struct perf_tool *tool,
 685                                        const union perf_event *event,
 686                                        struct perf_evsel *evsel,
 687                                        struct perf_sample *sample,
 688                                        struct machine *machine)
 689 {
 690         struct perf_top *top = container_of(tool, struct perf_top, tool);
 691         struct addr_location al;
 692         int err;
 693
 694         if (!machine && perf_guest) {
 695                 static struct intlist *seen;
 696
 697                 if (!seen)
 698                         seen = intlist__new(NULL);
 699
 700                 if (!intlist__has_entry(seen, sample->pid)) {
 701                         pr_err("Can't find guest [%d]'s kernel information\n",
 702                                 sample->pid);
 703                         intlist__add(seen, sample->pid);
 704                 }
 705                 return;
 706         }
 707
 708         if (!machine) {
 709                 pr_err("%u unprocessable samples recorded.\r",
 710                        top->session->stats.nr_unprocessable_samples++);
 711                 return;
 712         }
 713
 714         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 715                 top->exact_samples++;
 716
 717         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0)
 718                 return;
 719
 720         if (!top->kptr_restrict_warned &&
 721             symbol_conf.kptr_restrict &&
 722             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 723                 ui__warning(
 724 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 725 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 726 "Kernel%s samples will not be resolved.\n",
 727                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 728                           " modules" : "");
 729                 if (use_browser <= 0)
 730                         sleep(5);
 731                 top->kptr_restrict_warned = true;
 732         }
 733
 734         if (al.sym == NULL) {
 735                 const char *msg = "Kernel samples will not be resolved.\n";
 736                 /*
 737                  * As we do lazy loading of symtabs we only will know if the
 738                  * specified vmlinux file is invalid when we actually have a
 739                  * hit in kernel space and then try to load it. So if we get
 740                  * here and there are _no_ symbols in the DSO backing the
 741                  * kernel map, bail out.
 742                  *
 743                  * We may never get here, for instance, if we use -K/
 744                  * --hide-kernel-symbols, even if the user specifies an
 745                  * invalid --vmlinux ;-)
 746                  */
 747                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
 748                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 749                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 750                         if (symbol_conf.vmlinux_name) {
 751                                 ui__warning("The %s file can't be used.\n%s",
 752                                             symbol_conf.vmlinux_name, msg);
 753                         } else {
 754                                 ui__warning("A vmlinux file was not found.\n%s",
 755                                             msg);
 756                         }
 757
 758                         if (use_browser <= 0)
 759                                 sleep(5);
 760                         top->vmlinux_warned = true;
 761                 }
 762         }
 763
 764         if (al.sym == NULL || !al.sym->ignore) {
 765                 struct hist_entry_iter iter = {
 766                         .add_entry_cb = hist_iter__top_callback,
 767                 };
 768
 769                 if (symbol_conf.cumulate_callchain)
 770                         iter.ops = &hist_iter_cumulative;
 771                 else
 772                         iter.ops = &hist_iter_normal;
 773
 774                 pthread_mutex_lock(&evsel->hists.lock);
 775
 776                 err = hist_entry_iter__add(&iter, &al, evsel, sample,
 777                                            top->max_stack, top);
 778                 if (err < 0)
 779                         pr_err("Problem incrementing symbol period, skipping event\n");
 780
 781                 pthread_mutex_unlock(&evsel->hists.lock);
 782         }
 783
 784         return;
 785 }
 786
 787 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 788 {
 789         struct perf_sample sample;
 790         struct perf_evsel *evsel;
 791         struct perf_session *session = top->session;
 792         union perf_event *event;
 793         struct machine *machine;
 794         u8 origin;
 795         int ret;
 796
 797         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
 798                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
 799                 if (ret) {
 800                         pr_err("Can't parse sample, err = %d\n", ret);
 801                         goto next_event;
 802                 }
 803
 804                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 805                 assert(evsel != NULL);
 806
 807                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 808
 809                 if (event->header.type == PERF_RECORD_SAMPLE)
 810                         ++top->samples;
 811
 812                 switch (origin) {
 813                 case PERF_RECORD_MISC_USER:
 814                         ++top->us_samples;
 815                         if (top->hide_user_symbols)
 816                                 goto next_event;
 817                         machine = &session->machines.host;
 818                         break;
 819                 case PERF_RECORD_MISC_KERNEL:
 820                         ++top->kernel_samples;
 821                         if (top->hide_kernel_symbols)
 822                                 goto next_event;
 823                         machine = &session->machines.host;
 824                         break;
 825                 case PERF_RECORD_MISC_GUEST_KERNEL:
 826                         ++top->guest_kernel_samples;
 827                         machine = perf_session__find_machine(session,
 828                                                              sample.pid);
 829                         break;
 830                 case PERF_RECORD_MISC_GUEST_USER:
 831                         ++top->guest_us_samples;
 832                         /*
 833                          * TODO: we don't process guest user from host side
 834                          * except simple counting.
 835                          */
 836                         /* Fall thru */
 837                 default:
 838                         goto next_event;
 839                 }
 840
 841
 842                 if (event->header.type == PERF_RECORD_SAMPLE) {
 843                         perf_event__process_sample(&top->tool, event, evsel,
 844                                                    &sample, machine);
 845                 } else if (event->header.type < PERF_RECORD_MAX) {
 846                         hists__inc_nr_events(&evsel->hists, event->header.type);
 847                         machine__process_event(machine, event, &sample);
 848                 } else
 849                         ++session->stats.nr_unknown_events;
 850 next_event:
 851                 perf_evlist__mmap_consume(top->evlist, idx);
 852         }
 853 }
 854
 855 static void perf_top__mmap_read(struct perf_top *top)
 856 {
 857         int i;
 858
 859         for (i = 0; i < top->evlist->nr_mmaps; i++)
 860                 perf_top__mmap_read_idx(top, i);
 861 }
 862
 863 static int perf_top__start_counters(struct perf_top *top)
 864 {
 865         char msg[512];
 866         struct perf_evsel *counter;
 867         struct perf_evlist *evlist = top->evlist;
 868         struct record_opts *opts = &top->record_opts;
 869
 870         perf_evlist__config(evlist, opts);
 871
 872         evlist__for_each(evlist, counter) {
 873 try_again:
 874                 if (perf_evsel__open(counter, top->evlist->cpus,
 875                                      top->evlist->threads) < 0) {
 876                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 877                                 if (verbose)
 878                                         ui__warning("%s\n", msg);
 879                                 goto try_again;
 880                         }
 881
 882                         perf_evsel__open_strerror(counter, &opts->target,
 883                                                   errno, msg, sizeof(msg));
 884                         ui__error("%s\n", msg);
 885                         goto out_err;
 886                 }
 887         }
 888
 889         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 890                 ui__error("Failed to mmap with %d (%s)\n",
 891                             errno, strerror(errno));
 892                 goto out_err;
 893         }
 894
 895         return 0;
 896
 897 out_err:
 898         return -1;
 899 }
 900
 901 static int perf_top__setup_sample_type(struct perf_top *top __maybe_unused)
 902 {
 903         if (!sort__has_sym) {
 904                 if (symbol_conf.use_callchain) {
 905                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
 906                         return -EINVAL;
 907                 }
 908         } else if (callchain_param.mode != CHAIN_NONE) {
 909                 if (callchain_register_param(&callchain_param) < 0) {
 910                         ui__error("Can't register callchain params.\n");
 911                         return -EINVAL;
 912                 }
 913         }
 914
 915         return 0;
 916 }
 917
 918 static int __cmd_top(struct perf_top *top)
 919 {
 920         struct record_opts *opts = &top->record_opts;
 921         pthread_t thread;
 922         int ret;
 923
 924         top->session = perf_session__new(NULL, false, NULL);
 925         if (top->session == NULL)
 926                 return -ENOMEM;
 927
 928         machines__set_symbol_filter(&top->session->machines, symbol_filter);
 929
 930         if (!objdump_path) {
 931                 ret = perf_session_env__lookup_objdump(&top->session->header.env);
 932                 if (ret)
 933                         goto out_delete;
 934         }
 935
 936         ret = perf_top__setup_sample_type(top);
 937         if (ret)
 938                 goto out_delete;
 939
 940         machine__synthesize_threads(&top->session->machines.host, &opts->target,
 941                                     top->evlist->threads, false);
 942         ret = perf_top__start_counters(top);
 943         if (ret)
 944                 goto out_delete;
 945
 946         top->session->evlist = top->evlist;
 947         perf_session__set_id_hdr_size(top->session);
 948
 949         /*
 950          * When perf is starting the traced process, all the events (apart from
 951          * group members) have enable_on_exec=1 set, so don't spoil it by
 952          * prematurely enabling them.
 953          *
 954          * XXX 'top' still doesn't start workloads like record, trace, but should,
 955          * so leave the check here.
 956          */
 957         if (!target__none(&opts->target))
 958                 perf_evlist__enable(top->evlist);
 959
 960         /* Wait for a minimal set of events before starting the snapshot */
 961         poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
 962
 963         perf_top__mmap_read(top);
 964
 965         ret = -1;
 966         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 967                                                             display_thread), top)) {
 968                 ui__error("Could not create display thread.\n");
 969                 goto out_delete;
 970         }
 971
 972         if (top->realtime_prio) {
 973                 struct sched_param param;
 974
 975                 param.sched_priority = top->realtime_prio;
 976                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 977                         ui__error("Could not set realtime priority.\n");
 978                         goto out_join;
 979                 }
 980         }
 981
 982         while (!done) {
 983                 u64 hits = top->samples;
 984
 985                 perf_top__mmap_read(top);
 986
 987                 if (hits == top->samples)
 988                         ret = poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
 989         }
 990
 991         ret = 0;
 992 out_join:
 993         pthread_join(thread, NULL);
 994 out_delete:
 995         perf_session__delete(top->session);
 996         top->session = NULL;
 997
 998         return ret;
 999 }
1000
1001 static int
1002 callchain_opt(const struct option *opt, const char *arg, int unset)
1003 {
1004         symbol_conf.use_callchain = true;
1005         return record_callchain_opt(opt, arg, unset);
1006 }
1007
1008 static int
1009 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1010 {
1011         symbol_conf.use_callchain = true;
1012         return record_parse_callchain_opt(opt, arg, unset);
1013 }
1014
1015 static int perf_top_config(const char *var, const char *value, void *cb)
1016 {
1017         struct perf_top *top = cb;
1018
1019         if (!strcmp(var, "top.call-graph"))
1020                 return record_parse_callchain(value, &top->record_opts);
1021         if (!strcmp(var, "top.children")) {
1022                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1023                 return 0;
1024         }
1025
1026         return perf_default_config(var, value, cb);
1027 }
1028
1029 static int
1030 parse_percent_limit(const struct option *opt, const char *arg,
1031                     int unset __maybe_unused)
1032 {
1033         struct perf_top *top = opt->value;
1034
1035         top->min_percent = strtof(arg, NULL);
1036         return 0;
1037 }
1038
1039 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1040 {
1041         int status = -1;
1042         char errbuf[BUFSIZ];
1043         struct perf_top top = {
1044                 .count_filter        = 5,
1045                 .delay_secs          = 2,
1046                 .record_opts = {
1047                         .mmap_pages     = UINT_MAX,
1048                         .user_freq      = UINT_MAX,
1049                         .user_interval  = ULLONG_MAX,
1050                         .freq           = 4000, /* 4 KHz */
1051                         .target         = {
1052                                 .uses_mmap   = true,
1053                         },
1054                 },
1055                 .max_stack           = PERF_MAX_STACK_DEPTH,
1056                 .sym_pcnt_filter     = 5,
1057         };
1058         struct record_opts *opts = &top.record_opts;
1059         struct target *target = &opts->target;
1060         const struct option options[] = {
1061         OPT_CALLBACK('e', "event", &top.evlist, "event",
1062                      "event selector. use 'perf list' to list available events",
1063                      parse_events_option),
1064         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1065         OPT_STRING('p', "pid", &target->pid, "pid",
1066                     "profile events on existing process id"),
1067         OPT_STRING('t', "tid", &target->tid, "tid",
1068                     "profile events on existing thread id"),
1069         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1070                             "system-wide collection from all CPUs"),
1071         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1072                     "list of cpus to monitor"),
1073         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1074                    "file", "vmlinux pathname"),
1075         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1076                     "don't load vmlinux even if found"),
1077         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1078                     "hide kernel symbols"),
1079         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1080                      "number of mmap data pages",
1081                      perf_evlist__parse_mmap_pages),
1082         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1083                     "collect data with this RT SCHED_FIFO priority"),
1084         OPT_INTEGER('d', "delay", &top.delay_secs,
1085                     "number of seconds to delay between refreshes"),
1086         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1087                             "dump the symbol table used for profiling"),
1088         OPT_INTEGER('f', "count-filter", &top.count_filter,
1089                     "only display functions with more events than this"),
1090         OPT_BOOLEAN(0, "group", &opts->group,
1091                             "put the counters into a counter group"),
1092         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1093                     "child tasks do not inherit counters"),
1094         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1095                     "symbol to annotate"),
1096         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1097         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1098         OPT_INTEGER('E', "entries", &top.print_entries,
1099                     "display this many functions"),
1100         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1101                     "hide user symbols"),
1102         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1103         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1104         OPT_INCR('v', "verbose", &verbose,
1105                     "be more verbose (show counter open errors, etc)"),
1106         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1107                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1108                    " Please refer the man page for the complete list."),
1109         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1110                    "output field(s): overhead, period, sample plus all of sort keys"),
1111         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1112                     "Show a column with the number of samples"),
1113         OPT_CALLBACK_NOOPT('g', NULL, &top.record_opts,
1114                            NULL, "enables call-graph recording",
1115                            &callchain_opt),
1116         OPT_CALLBACK(0, "call-graph", &top.record_opts,
1117                      "mode[,dump_size]", record_callchain_help,
1118                      &parse_callchain_opt),
1119         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1120                     "Accumulate callchains of children and show total overhead as well"),
1121         OPT_INTEGER(0, "max-stack", &top.max_stack,
1122                     "Set the maximum stack depth when parsing the callchain. "
1123                     "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
1124         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1125                    "ignore callees of these functions in call graphs",
1126                    report_parse_ignore_callees_opt),
1127         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1128                     "Show a column with the sum of periods"),
1129         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1130                    "only consider symbols in these dsos"),
1131         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1132                    "only consider symbols in these comms"),
1133         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1134                    "only consider these symbols"),
1135         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1136                     "Interleave source code with assembly code (default)"),
1137         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1138                     "Display raw encoding of assembly instructions (default)"),
1139         OPT_STRING(0, "objdump", &objdump_path, "path",
1140                     "objdump binary to use for disassembly and annotations"),
1141         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1142                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1143         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1144         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1145                      "Don't show entries under that percent", parse_percent_limit),
1146         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1147                      "How to display percentage of filtered entries", parse_filter_percentage),
1148         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1149                    "width[,width...]",
1150                    "don't try to adjust column width, use these fixed values"),
1151         OPT_END()
1152         };
1153         const char * const top_usage[] = {
1154                 "perf top [<options>]",
1155                 NULL
1156         };
1157
1158         top.evlist = perf_evlist__new();
1159         if (top.evlist == NULL)
1160                 return -ENOMEM;
1161
1162         perf_config(perf_top_config, &top);
1163
1164         argc = parse_options(argc, argv, options, top_usage, 0);
1165         if (argc)
1166                 usage_with_options(top_usage, options);
1167
1168         sort__mode = SORT_MODE__TOP;
1169         /* display thread wants entries to be collapsed in a different tree */
1170         sort__need_collapse = 1;
1171
1172         if (setup_sorting() < 0) {
1173                 if (sort_order)
1174                         parse_options_usage(top_usage, options, "s", 1);
1175                 if (field_order)
1176                         parse_options_usage(sort_order ? NULL : top_usage,
1177                                             options, "fields", 0);
1178                 goto out_delete_evlist;
1179         }
1180
1181         if (top.use_stdio)
1182                 use_browser = 0;
1183         else if (top.use_tui)
1184                 use_browser = 1;
1185
1186         setup_browser(false);
1187
1188         status = target__validate(target);
1189         if (status) {
1190                 target__strerror(target, status, errbuf, BUFSIZ);
1191                 ui__warning("%s\n", errbuf);
1192         }
1193
1194         status = target__parse_uid(target);
1195         if (status) {
1196                 int saved_errno = errno;
1197
1198                 target__strerror(target, status, errbuf, BUFSIZ);
1199                 ui__error("%s\n", errbuf);
1200
1201                 status = -saved_errno;
1202                 goto out_delete_evlist;
1203         }
1204
1205         if (target__none(target))
1206                 target->system_wide = true;
1207
1208         if (perf_evlist__create_maps(top.evlist, target) < 0)
1209                 usage_with_options(top_usage, options);
1210
1211         if (!top.evlist->nr_entries &&
1212             perf_evlist__add_default(top.evlist) < 0) {
1213                 ui__error("Not enough memory for event selector list\n");
1214                 goto out_delete_evlist;
1215         }
1216
1217         symbol_conf.nr_events = top.evlist->nr_entries;
1218
1219         if (top.delay_secs < 1)
1220                 top.delay_secs = 1;
1221
1222         if (record_opts__config(opts)) {
1223                 status = -EINVAL;
1224                 goto out_delete_evlist;
1225         }
1226
1227         top.sym_evsel = perf_evlist__first(top.evlist);
1228
1229         if (!symbol_conf.use_callchain) {
1230                 symbol_conf.cumulate_callchain = false;
1231                 perf_hpp__cancel_cumulate();
1232         }
1233
1234         symbol_conf.priv_size = sizeof(struct annotation);
1235
1236         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1237         if (symbol__init() < 0)
1238                 return -1;
1239
1240         sort__setup_elide(stdout);
1241
1242         get_term_dimensions(&top.winsize);
1243         if (top.print_entries == 0) {
1244                 struct sigaction act = {
1245                         .sa_sigaction = perf_top__sig_winch,
1246                         .sa_flags     = SA_SIGINFO,
1247                 };
1248                 perf_top__update_print_entries(&top);
1249                 sigaction(SIGWINCH, &act, NULL);
1250         }
1251
1252         status = __cmd_top(&top);
1253
1254 out_delete_evlist:
1255         perf_evlist__delete(top.evlist);
1256
1257         return status;
1258 }