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
 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         if (top->zero) {
 280                 hists__delete_entries(&top->sym_evsel->hists);
 281         } else {
 282                 hists__decay_entries(&top->sym_evsel->hists,
 283                                      top->hide_user_symbols,
 284                                      top->hide_kernel_symbols);
 285         }
 286
 287         hists__collapse_resort(&top->sym_evsel->hists, NULL);
 288         hists__output_resort(&top->sym_evsel->hists);
 289
 290         hists__output_recalc_col_len(&top->sym_evsel->hists,
 291                                      top->print_entries - printed);
 292         putchar('\n');
 293         hists__fprintf(&top->sym_evsel->hists, false,
 294                        top->print_entries - printed, win_width,
 295                        top->min_percent, stdout);
 296 }
 297
 298 static void prompt_integer(int *target, const char *msg)
 299 {
 300         char *buf = malloc(0), *p;
 301         size_t dummy = 0;
 302         int tmp;
 303
 304         fprintf(stdout, "\n%s: ", msg);
 305         if (getline(&buf, &dummy, stdin) < 0)
 306                 return;
 307
 308         p = strchr(buf, '\n');
 309         if (p)
 310                 *p = 0;
 311
 312         p = buf;
 313         while(*p) {
 314                 if (!isdigit(*p))
 315                         goto out_free;
 316                 p++;
 317         }
 318         tmp = strtoul(buf, NULL, 10);
 319         *target = tmp;
 320 out_free:
 321         free(buf);
 322 }
 323
 324 static void prompt_percent(int *target, const char *msg)
 325 {
 326         int tmp = 0;
 327
 328         prompt_integer(&tmp, msg);
 329         if (tmp >= 0 && tmp <= 100)
 330                 *target = tmp;
 331 }
 332
 333 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 334 {
 335         char *buf = malloc(0), *p;
 336         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 337         struct rb_node *next;
 338         size_t dummy = 0;
 339
 340         /* zero counters of active symbol */
 341         if (syme) {
 342                 __zero_source_counters(syme);
 343                 top->sym_filter_entry = NULL;
 344         }
 345
 346         fprintf(stdout, "\n%s: ", msg);
 347         if (getline(&buf, &dummy, stdin) < 0)
 348                 goto out_free;
 349
 350         p = strchr(buf, '\n');
 351         if (p)
 352                 *p = 0;
 353
 354         next = rb_first(&top->sym_evsel->hists.entries);
 355         while (next) {
 356                 n = rb_entry(next, struct hist_entry, rb_node);
 357                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 358                         found = n;
 359                         break;
 360                 }
 361                 next = rb_next(&n->rb_node);
 362         }
 363
 364         if (!found) {
 365                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 366                 sleep(1);
 367         } else
 368                 perf_top__parse_source(top, found);
 369
 370 out_free:
 371         free(buf);
 372 }
 373
 374 static void perf_top__print_mapped_keys(struct perf_top *top)
 375 {
 376         char *name = NULL;
 377
 378         if (top->sym_filter_entry) {
 379                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 380                 name = sym->name;
 381         }
 382
 383         fprintf(stdout, "\nMapped keys:\n");
 384         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 385         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 386
 387         if (top->evlist->nr_entries > 1)
 388                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 389
 390         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 391
 392         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
 393         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 394         fprintf(stdout, "\t[S]     stop annotation.\n");
 395
 396         fprintf(stdout,
 397                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 398                 top->hide_kernel_symbols ? "yes" : "no");
 399         fprintf(stdout,
 400                 "\t[U]     hide user symbols.               \t(%s)\n",
 401                 top->hide_user_symbols ? "yes" : "no");
 402         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 403         fprintf(stdout, "\t[qQ]    quit.\n");
 404 }
 405
 406 static int perf_top__key_mapped(struct perf_top *top, int c)
 407 {
 408         switch (c) {
 409                 case 'd':
 410                 case 'e':
 411                 case 'f':
 412                 case 'z':
 413                 case 'q':
 414                 case 'Q':
 415                 case 'K':
 416                 case 'U':
 417                 case 'F':
 418                 case 's':
 419                 case 'S':
 420                         return 1;
 421                 case 'E':
 422                         return top->evlist->nr_entries > 1 ? 1 : 0;
 423                 default:
 424                         break;
 425         }
 426
 427         return 0;
 428 }
 429
 430 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 431 {
 432         bool ret = true;
 433
 434         if (!perf_top__key_mapped(top, c)) {
 435                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 436                 struct termios save;
 437
 438                 perf_top__print_mapped_keys(top);
 439                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 440                 fflush(stdout);
 441
 442                 set_term_quiet_input(&save);
 443
 444                 poll(&stdin_poll, 1, -1);
 445                 c = getc(stdin);
 446
 447                 tcsetattr(0, TCSAFLUSH, &save);
 448                 if (!perf_top__key_mapped(top, c))
 449                         return ret;
 450         }
 451
 452         switch (c) {
 453                 case 'd':
 454                         prompt_integer(&top->delay_secs, "Enter display delay");
 455                         if (top->delay_secs < 1)
 456                                 top->delay_secs = 1;
 457                         break;
 458                 case 'e':
 459                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 460                         if (top->print_entries == 0) {
 461                                 struct sigaction act = {
 462                                         .sa_sigaction = perf_top__sig_winch,
 463                                         .sa_flags     = SA_SIGINFO,
 464                                 };
 465                                 perf_top__sig_winch(SIGWINCH, NULL, top);
 466                                 sigaction(SIGWINCH, &act, NULL);
 467                         } else {
 468                                 signal(SIGWINCH, SIG_DFL);
 469                         }
 470                         break;
 471                 case 'E':
 472                         if (top->evlist->nr_entries > 1) {
 473                                 /* Select 0 as the default event: */
 474                                 int counter = 0;
 475
 476                                 fprintf(stderr, "\nAvailable events:");
 477
 478                                 evlist__for_each(top->evlist, top->sym_evsel)
 479                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
 480
 481                                 prompt_integer(&counter, "Enter details event counter");
 482
 483                                 if (counter >= top->evlist->nr_entries) {
 484                                         top->sym_evsel = perf_evlist__first(top->evlist);
 485                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
 486                                         sleep(1);
 487                                         break;
 488                                 }
 489                                 evlist__for_each(top->evlist, top->sym_evsel)
 490                                         if (top->sym_evsel->idx == counter)
 491                                                 break;
 492                         } else
 493                                 top->sym_evsel = perf_evlist__first(top->evlist);
 494                         break;
 495                 case 'f':
 496                         prompt_integer(&top->count_filter, "Enter display event count filter");
 497                         break;
 498                 case 'F':
 499                         prompt_percent(&top->sym_pcnt_filter,
 500                                        "Enter details display event filter (percent)");
 501                         break;
 502                 case 'K':
 503                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 504                         break;
 505                 case 'q':
 506                 case 'Q':
 507                         printf("exiting.\n");
 508                         if (top->dump_symtab)
 509                                 perf_session__fprintf_dsos(top->session, stderr);
 510                         ret = false;
 511                         break;
 512                 case 's':
 513                         perf_top__prompt_symbol(top, "Enter details symbol");
 514                         break;
 515                 case 'S':
 516                         if (!top->sym_filter_entry)
 517                                 break;
 518                         else {
 519                                 struct hist_entry *syme = top->sym_filter_entry;
 520
 521                                 top->sym_filter_entry = NULL;
 522                                 __zero_source_counters(syme);
 523                         }
 524                         break;
 525                 case 'U':
 526                         top->hide_user_symbols = !top->hide_user_symbols;
 527                         break;
 528                 case 'z':
 529                         top->zero = !top->zero;
 530                         break;
 531                 default:
 532                         break;
 533         }
 534
 535         return ret;
 536 }
 537
 538 static void perf_top__sort_new_samples(void *arg)
 539 {
 540         struct perf_top *t = arg;
 541         perf_top__reset_sample_counters(t);
 542
 543         if (t->evlist->selected != NULL)
 544                 t->sym_evsel = t->evlist->selected;
 545
 546         if (t->zero) {
 547                 hists__delete_entries(&t->sym_evsel->hists);
 548         } else {
 549                 hists__decay_entries(&t->sym_evsel->hists,
 550                                      t->hide_user_symbols,
 551                                      t->hide_kernel_symbols);
 552         }
 553
 554         hists__collapse_resort(&t->sym_evsel->hists, NULL);
 555         hists__output_resort(&t->sym_evsel->hists);
 556 }
 557
 558 static void *display_thread_tui(void *arg)
 559 {
 560         struct perf_evsel *pos;
 561         struct perf_top *top = arg;
 562         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 563         struct hist_browser_timer hbt = {
 564                 .timer          = perf_top__sort_new_samples,
 565                 .arg            = top,
 566                 .refresh        = top->delay_secs,
 567         };
 568
 569         perf_top__sort_new_samples(top);
 570
 571         /*
 572          * Initialize the uid_filter_str, in the future the TUI will allow
 573          * Zooming in/out UIDs. For now juse use whatever the user passed
 574          * via --uid.
 575          */
 576         evlist__for_each(top->evlist, pos)
 577                 pos->hists.uid_filter_str = top->record_opts.target.uid_str;
 578
 579         perf_evlist__tui_browse_hists(top->evlist, help, &hbt, top->min_percent,
 580                                       &top->session->header.env);
 581
 582         done = 1;
 583         return NULL;
 584 }
 585
 586 static void display_sig(int sig __maybe_unused)
 587 {
 588         done = 1;
 589 }
 590
 591 static void display_setup_sig(void)
 592 {
 593         signal(SIGSEGV, display_sig);
 594         signal(SIGFPE,  display_sig);
 595         signal(SIGINT,  display_sig);
 596         signal(SIGQUIT, display_sig);
 597         signal(SIGTERM, display_sig);
 598 }
 599
 600 static void *display_thread(void *arg)
 601 {
 602         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 603         struct termios save;
 604         struct perf_top *top = arg;
 605         int delay_msecs, c;
 606
 607         display_setup_sig();
 608         pthread__unblock_sigwinch();
 609 repeat:
 610         delay_msecs = top->delay_secs * 1000;
 611         set_term_quiet_input(&save);
 612         /* trash return*/
 613         getc(stdin);
 614
 615         while (!done) {
 616                 perf_top__print_sym_table(top);
 617                 /*
 618                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 619                  * refresh screen in both cases.
 620                  */
 621                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 622                 case 0:
 623                         continue;
 624                 case -1:
 625                         if (errno == EINTR)
 626                                 continue;
 627                         /* Fall trhu */
 628                 default:
 629                         c = getc(stdin);
 630                         tcsetattr(0, TCSAFLUSH, &save);
 631
 632                         if (perf_top__handle_keypress(top, c))
 633                                 goto repeat;
 634                         done = 1;
 635                 }
 636         }
 637
 638         tcsetattr(0, TCSAFLUSH, &save);
 639         return NULL;
 640 }
 641
 642 static int symbol_filter(struct map *map, struct symbol *sym)
 643 {
 644         const char *name = sym->name;
 645
 646         if (!map->dso->kernel)
 647                 return 0;
 648         /*
 649          * ppc64 uses function descriptors and appends a '.' to the
 650          * start of every instruction address. Remove it.
 651          */
 652         if (name[0] == '.')
 653                 name++;
 654
 655         if (!strcmp(name, "_text") ||
 656             !strcmp(name, "_etext") ||
 657             !strcmp(name, "_sinittext") ||
 658             !strncmp("init_module", name, 11) ||
 659             !strncmp("cleanup_module", name, 14) ||
 660             strstr(name, "_text_start") ||
 661             strstr(name, "_text_end"))
 662                 return 1;
 663
 664         if (symbol__is_idle(sym))
 665                 sym->ignore = true;
 666
 667         return 0;
 668 }
 669
 670 static int hist_iter__top_callback(struct hist_entry_iter *iter,
 671                                    struct addr_location *al, bool single,
 672                                    void *arg)
 673 {
 674         struct perf_top *top = arg;
 675         struct hist_entry *he = iter->he;
 676         struct perf_evsel *evsel = iter->evsel;
 677
 678         if (sort__has_sym && single) {
 679                 u64 ip = al->addr;
 680
 681                 if (al->map)
 682                         ip = al->map->unmap_ip(al->map, ip);
 683
 684                 perf_top__record_precise_ip(top, he, evsel->idx, ip);
 685         }
 686
 687         return 0;
 688 }
 689
 690 static void perf_event__process_sample(struct perf_tool *tool,
 691                                        const union perf_event *event,
 692                                        struct perf_evsel *evsel,
 693                                        struct perf_sample *sample,
 694                                        struct machine *machine)
 695 {
 696         struct perf_top *top = container_of(tool, struct perf_top, tool);
 697         struct addr_location al;
 698         int err;
 699
 700         if (!machine && perf_guest) {
 701                 static struct intlist *seen;
 702
 703                 if (!seen)
 704                         seen = intlist__new(NULL);
 705
 706                 if (!intlist__has_entry(seen, sample->pid)) {
 707                         pr_err("Can't find guest [%d]'s kernel information\n",
 708                                 sample->pid);
 709                         intlist__add(seen, sample->pid);
 710                 }
 711                 return;
 712         }
 713
 714         if (!machine) {
 715                 pr_err("%u unprocessable samples recorded.\r",
 716                        top->session->stats.nr_unprocessable_samples++);
 717                 return;
 718         }
 719
 720         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 721                 top->exact_samples++;
 722
 723         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0)
 724                 return;
 725
 726         if (!top->kptr_restrict_warned &&
 727             symbol_conf.kptr_restrict &&
 728             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 729                 ui__warning(
 730 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 731 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 732 "Kernel%s samples will not be resolved.\n",
 733                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 734                           " modules" : "");
 735                 if (use_browser <= 0)
 736                         sleep(5);
 737                 top->kptr_restrict_warned = true;
 738         }
 739
 740         if (al.sym == NULL) {
 741                 const char *msg = "Kernel samples will not be resolved.\n";
 742                 /*
 743                  * As we do lazy loading of symtabs we only will know if the
 744                  * specified vmlinux file is invalid when we actually have a
 745                  * hit in kernel space and then try to load it. So if we get
 746                  * here and there are _no_ symbols in the DSO backing the
 747                  * kernel map, bail out.
 748                  *
 749                  * We may never get here, for instance, if we use -K/
 750                  * --hide-kernel-symbols, even if the user specifies an
 751                  * invalid --vmlinux ;-)
 752                  */
 753                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
 754                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 755                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 756                         if (symbol_conf.vmlinux_name) {
 757                                 ui__warning("The %s file can't be used.\n%s",
 758                                             symbol_conf.vmlinux_name, msg);
 759                         } else {
 760                                 ui__warning("A vmlinux file was not found.\n%s",
 761                                             msg);
 762                         }
 763
 764                         if (use_browser <= 0)
 765                                 sleep(5);
 766                         top->vmlinux_warned = true;
 767                 }
 768         }
 769
 770         if (al.sym == NULL || !al.sym->ignore) {
 771                 struct hist_entry_iter iter = {
 772                         .add_entry_cb = hist_iter__top_callback,
 773                 };
 774
 775                 if (symbol_conf.cumulate_callchain)
 776                         iter.ops = &hist_iter_cumulative;
 777                 else
 778                         iter.ops = &hist_iter_normal;
 779
 780                 pthread_mutex_lock(&evsel->hists.lock);
 781
 782                 err = hist_entry_iter__add(&iter, &al, evsel, sample,
 783                                            top->max_stack, top);
 784                 if (err < 0)
 785                         pr_err("Problem incrementing symbol period, skipping event\n");
 786
 787                 pthread_mutex_unlock(&evsel->hists.lock);
 788         }
 789
 790         return;
 791 }
 792
 793 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 794 {
 795         struct perf_sample sample;
 796         struct perf_evsel *evsel;
 797         struct perf_session *session = top->session;
 798         union perf_event *event;
 799         struct machine *machine;
 800         u8 origin;
 801         int ret;
 802
 803         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
 804                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
 805                 if (ret) {
 806                         pr_err("Can't parse sample, err = %d\n", ret);
 807                         goto next_event;
 808                 }
 809
 810                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 811                 assert(evsel != NULL);
 812
 813                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 814
 815                 if (event->header.type == PERF_RECORD_SAMPLE)
 816                         ++top->samples;
 817
 818                 switch (origin) {
 819                 case PERF_RECORD_MISC_USER:
 820                         ++top->us_samples;
 821                         if (top->hide_user_symbols)
 822                                 goto next_event;
 823                         machine = &session->machines.host;
 824                         break;
 825                 case PERF_RECORD_MISC_KERNEL:
 826                         ++top->kernel_samples;
 827                         if (top->hide_kernel_symbols)
 828                                 goto next_event;
 829                         machine = &session->machines.host;
 830                         break;
 831                 case PERF_RECORD_MISC_GUEST_KERNEL:
 832                         ++top->guest_kernel_samples;
 833                         machine = perf_session__find_machine(session,
 834                                                              sample.pid);
 835                         break;
 836                 case PERF_RECORD_MISC_GUEST_USER:
 837                         ++top->guest_us_samples;
 838                         /*
 839                          * TODO: we don't process guest user from host side
 840                          * except simple counting.
 841                          */
 842                         /* Fall thru */
 843                 default:
 844                         goto next_event;
 845                 }
 846
 847
 848                 if (event->header.type == PERF_RECORD_SAMPLE) {
 849                         perf_event__process_sample(&top->tool, event, evsel,
 850                                                    &sample, machine);
 851                 } else if (event->header.type < PERF_RECORD_MAX) {
 852                         hists__inc_nr_events(&evsel->hists, event->header.type);
 853                         machine__process_event(machine, event, &sample);
 854                 } else
 855                         ++session->stats.nr_unknown_events;
 856 next_event:
 857                 perf_evlist__mmap_consume(top->evlist, idx);
 858         }
 859 }
 860
 861 static void perf_top__mmap_read(struct perf_top *top)
 862 {
 863         int i;
 864
 865         for (i = 0; i < top->evlist->nr_mmaps; i++)
 866                 perf_top__mmap_read_idx(top, i);
 867 }
 868
 869 static int perf_top__start_counters(struct perf_top *top)
 870 {
 871         char msg[512];
 872         struct perf_evsel *counter;
 873         struct perf_evlist *evlist = top->evlist;
 874         struct record_opts *opts = &top->record_opts;
 875
 876         perf_evlist__config(evlist, opts);
 877
 878         evlist__for_each(evlist, counter) {
 879 try_again:
 880                 if (perf_evsel__open(counter, top->evlist->cpus,
 881                                      top->evlist->threads) < 0) {
 882                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 883                                 if (verbose)
 884                                         ui__warning("%s\n", msg);
 885                                 goto try_again;
 886                         }
 887
 888                         perf_evsel__open_strerror(counter, &opts->target,
 889                                                   errno, msg, sizeof(msg));
 890                         ui__error("%s\n", msg);
 891                         goto out_err;
 892                 }
 893         }
 894
 895         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 896                 ui__error("Failed to mmap with %d (%s)\n",
 897                             errno, strerror_r(errno, msg, sizeof(msg)));
 898                 goto out_err;
 899         }
 900
 901         return 0;
 902
 903 out_err:
 904         return -1;
 905 }
 906
 907 static int perf_top__setup_sample_type(struct perf_top *top __maybe_unused)
 908 {
 909         if (!sort__has_sym) {
 910                 if (symbol_conf.use_callchain) {
 911                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
 912                         return -EINVAL;
 913                 }
 914         } else if (callchain_param.mode != CHAIN_NONE) {
 915                 if (callchain_register_param(&callchain_param) < 0) {
 916                         ui__error("Can't register callchain params.\n");
 917                         return -EINVAL;
 918                 }
 919         }
 920
 921         return 0;
 922 }
 923
 924 static int __cmd_top(struct perf_top *top)
 925 {
 926         struct record_opts *opts = &top->record_opts;
 927         pthread_t thread;
 928         int ret;
 929
 930         top->session = perf_session__new(NULL, false, NULL);
 931         if (top->session == NULL)
 932                 return -1;
 933
 934         machines__set_symbol_filter(&top->session->machines, symbol_filter);
 935
 936         if (!objdump_path) {
 937                 ret = perf_session_env__lookup_objdump(&top->session->header.env);
 938                 if (ret)
 939                         goto out_delete;
 940         }
 941
 942         ret = perf_top__setup_sample_type(top);
 943         if (ret)
 944                 goto out_delete;
 945
 946         machine__synthesize_threads(&top->session->machines.host, &opts->target,
 947                                     top->evlist->threads, false);
 948         ret = perf_top__start_counters(top);
 949         if (ret)
 950                 goto out_delete;
 951
 952         top->session->evlist = top->evlist;
 953         perf_session__set_id_hdr_size(top->session);
 954
 955         /*
 956          * When perf is starting the traced process, all the events (apart from
 957          * group members) have enable_on_exec=1 set, so don't spoil it by
 958          * prematurely enabling them.
 959          *
 960          * XXX 'top' still doesn't start workloads like record, trace, but should,
 961          * so leave the check here.
 962          */
 963         if (!target__none(&opts->target))
 964                 perf_evlist__enable(top->evlist);
 965
 966         /* Wait for a minimal set of events before starting the snapshot */
 967         perf_evlist__poll(top->evlist, 100);
 968
 969         perf_top__mmap_read(top);
 970
 971         ret = -1;
 972         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 973                                                             display_thread), top)) {
 974                 ui__error("Could not create display thread.\n");
 975                 goto out_delete;
 976         }
 977
 978         if (top->realtime_prio) {
 979                 struct sched_param param;
 980
 981                 param.sched_priority = top->realtime_prio;
 982                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 983                         ui__error("Could not set realtime priority.\n");
 984                         goto out_join;
 985                 }
 986         }
 987
 988         while (!done) {
 989                 u64 hits = top->samples;
 990
 991                 perf_top__mmap_read(top);
 992
 993                 if (hits == top->samples)
 994                         ret = perf_evlist__poll(top->evlist, 100);
 995         }
 996
 997         ret = 0;
 998 out_join:
 999         pthread_join(thread, NULL);
1000 out_delete:
1001         perf_session__delete(top->session);
1002         top->session = NULL;
1003
1004         return ret;
1005 }
1006
1007 static int
1008 callchain_opt(const struct option *opt, const char *arg, int unset)
1009 {
1010         symbol_conf.use_callchain = true;
1011         return record_callchain_opt(opt, arg, unset);
1012 }
1013
1014 static int
1015 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1016 {
1017         symbol_conf.use_callchain = true;
1018         return record_parse_callchain_opt(opt, arg, unset);
1019 }
1020
1021 static int perf_top_config(const char *var, const char *value, void *cb)
1022 {
1023         if (!strcmp(var, "top.call-graph"))
1024                 var = "call-graph.record-mode"; /* fall-through */
1025         if (!strcmp(var, "top.children")) {
1026                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1027                 return 0;
1028         }
1029
1030         return perf_default_config(var, value, cb);
1031 }
1032
1033 static int
1034 parse_percent_limit(const struct option *opt, const char *arg,
1035                     int unset __maybe_unused)
1036 {
1037         struct perf_top *top = opt->value;
1038
1039         top->min_percent = strtof(arg, NULL);
1040         return 0;
1041 }
1042
1043 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1044 {
1045         int status = -1;
1046         char errbuf[BUFSIZ];
1047         struct perf_top top = {
1048                 .count_filter        = 5,
1049                 .delay_secs          = 2,
1050                 .record_opts = {
1051                         .mmap_pages     = UINT_MAX,
1052                         .user_freq      = UINT_MAX,
1053                         .user_interval  = ULLONG_MAX,
1054                         .freq           = 4000, /* 4 KHz */
1055                         .target         = {
1056                                 .uses_mmap   = true,
1057                         },
1058                 },
1059                 .max_stack           = PERF_MAX_STACK_DEPTH,
1060                 .sym_pcnt_filter     = 5,
1061         };
1062         struct record_opts *opts = &top.record_opts;
1063         struct target *target = &opts->target;
1064         const struct option options[] = {
1065         OPT_CALLBACK('e', "event", &top.evlist, "event",
1066                      "event selector. use 'perf list' to list available events",
1067                      parse_events_option),
1068         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1069         OPT_STRING('p', "pid", &target->pid, "pid",
1070                     "profile events on existing process id"),
1071         OPT_STRING('t', "tid", &target->tid, "tid",
1072                     "profile events on existing thread id"),
1073         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1074                             "system-wide collection from all CPUs"),
1075         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1076                     "list of cpus to monitor"),
1077         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1078                    "file", "vmlinux pathname"),
1079         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1080                     "don't load vmlinux even if found"),
1081         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1082                     "hide kernel symbols"),
1083         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1084                      "number of mmap data pages",
1085                      perf_evlist__parse_mmap_pages),
1086         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1087                     "collect data with this RT SCHED_FIFO priority"),
1088         OPT_INTEGER('d', "delay", &top.delay_secs,
1089                     "number of seconds to delay between refreshes"),
1090         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1091                             "dump the symbol table used for profiling"),
1092         OPT_INTEGER('f', "count-filter", &top.count_filter,
1093                     "only display functions with more events than this"),
1094         OPT_BOOLEAN(0, "group", &opts->group,
1095                             "put the counters into a counter group"),
1096         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1097                     "child tasks do not inherit counters"),
1098         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1099                     "symbol to annotate"),
1100         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1101         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1102         OPT_INTEGER('E', "entries", &top.print_entries,
1103                     "display this many functions"),
1104         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1105                     "hide user symbols"),
1106         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1107         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1108         OPT_INCR('v', "verbose", &verbose,
1109                     "be more verbose (show counter open errors, etc)"),
1110         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1111                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1112                    " Please refer the man page for the complete list."),
1113         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1114                    "output field(s): overhead, period, sample plus all of sort keys"),
1115         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1116                     "Show a column with the number of samples"),
1117         OPT_CALLBACK_NOOPT('g', NULL, &top.record_opts,
1118                            NULL, "enables call-graph recording",
1119                            &callchain_opt),
1120         OPT_CALLBACK(0, "call-graph", &top.record_opts,
1121                      "mode[,dump_size]", record_callchain_help,
1122                      &parse_callchain_opt),
1123         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1124                     "Accumulate callchains of children and show total overhead as well"),
1125         OPT_INTEGER(0, "max-stack", &top.max_stack,
1126                     "Set the maximum stack depth when parsing the callchain. "
1127                     "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
1128         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1129                    "ignore callees of these functions in call graphs",
1130                    report_parse_ignore_callees_opt),
1131         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1132                     "Show a column with the sum of periods"),
1133         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1134                    "only consider symbols in these dsos"),
1135         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1136                    "only consider symbols in these comms"),
1137         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1138                    "only consider these symbols"),
1139         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1140                     "Interleave source code with assembly code (default)"),
1141         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1142                     "Display raw encoding of assembly instructions (default)"),
1143         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1144                     "Enable kernel symbol demangling"),
1145         OPT_STRING(0, "objdump", &objdump_path, "path",
1146                     "objdump binary to use for disassembly and annotations"),
1147         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1148                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1149         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1150         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1151                      "Don't show entries under that percent", parse_percent_limit),
1152         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1153                      "How to display percentage of filtered entries", parse_filter_percentage),
1154         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1155                    "width[,width...]",
1156                    "don't try to adjust column width, use these fixed values"),
1157         OPT_END()
1158         };
1159         const char * const top_usage[] = {
1160                 "perf top [<options>]",
1161                 NULL
1162         };
1163
1164         top.evlist = perf_evlist__new();
1165         if (top.evlist == NULL)
1166                 return -ENOMEM;
1167
1168         perf_config(perf_top_config, &top);
1169
1170         argc = parse_options(argc, argv, options, top_usage, 0);
1171         if (argc)
1172                 usage_with_options(top_usage, options);
1173
1174         sort__mode = SORT_MODE__TOP;
1175         /* display thread wants entries to be collapsed in a different tree */
1176         sort__need_collapse = 1;
1177
1178         if (setup_sorting() < 0) {
1179                 if (sort_order)
1180                         parse_options_usage(top_usage, options, "s", 1);
1181                 if (field_order)
1182                         parse_options_usage(sort_order ? NULL : top_usage,
1183                                             options, "fields", 0);
1184                 goto out_delete_evlist;
1185         }
1186
1187         if (top.use_stdio)
1188                 use_browser = 0;
1189         else if (top.use_tui)
1190                 use_browser = 1;
1191
1192         setup_browser(false);
1193
1194         status = target__validate(target);
1195         if (status) {
1196                 target__strerror(target, status, errbuf, BUFSIZ);
1197                 ui__warning("%s\n", errbuf);
1198         }
1199
1200         status = target__parse_uid(target);
1201         if (status) {
1202                 int saved_errno = errno;
1203
1204                 target__strerror(target, status, errbuf, BUFSIZ);
1205                 ui__error("%s\n", errbuf);
1206
1207                 status = -saved_errno;
1208                 goto out_delete_evlist;
1209         }
1210
1211         if (target__none(target))
1212                 target->system_wide = true;
1213
1214         if (perf_evlist__create_maps(top.evlist, target) < 0)
1215                 usage_with_options(top_usage, options);
1216
1217         if (!top.evlist->nr_entries &&
1218             perf_evlist__add_default(top.evlist) < 0) {
1219                 ui__error("Not enough memory for event selector list\n");
1220                 goto out_delete_evlist;
1221         }
1222
1223         symbol_conf.nr_events = top.evlist->nr_entries;
1224
1225         if (top.delay_secs < 1)
1226                 top.delay_secs = 1;
1227
1228         if (record_opts__config(opts)) {
1229                 status = -EINVAL;
1230                 goto out_delete_evlist;
1231         }
1232
1233         top.sym_evsel = perf_evlist__first(top.evlist);
1234
1235         if (!symbol_conf.use_callchain) {
1236                 symbol_conf.cumulate_callchain = false;
1237                 perf_hpp__cancel_cumulate();
1238         }
1239
1240         symbol_conf.priv_size = sizeof(struct annotation);
1241
1242         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1243         if (symbol__init(NULL) < 0)
1244                 return -1;
1245
1246         sort__setup_elide(stdout);
1247
1248         get_term_dimensions(&top.winsize);
1249         if (top.print_entries == 0) {
1250                 struct sigaction act = {
1251                         .sa_sigaction = perf_top__sig_winch,
1252                         .sa_flags     = SA_SIGINFO,
1253                 };
1254                 perf_top__update_print_entries(&top);
1255                 sigaction(SIGWINCH, &act, NULL);
1256         }
1257
1258         status = __cmd_top(&top);
1259
1260 out_delete_evlist:
1261         perf_evlist__delete(top.evlist);
1262
1263         return status;
1264 }