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;
 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         if (notes->src == NULL && symbol__alloc_hist(sym) < 0) {
 193                 pthread_mutex_unlock(&notes->lock);
 194                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 195                        sym->name);
 196                 sleep(1);
 197                 return;
 198         }
 199
 200         ip = he->ms.map->map_ip(he->ms.map, ip);
 201         err = symbol__inc_addr_samples(sym, he->ms.map, counter, ip);
 202
 203         pthread_mutex_unlock(&notes->lock);
 204
 205         if (err == -ERANGE && !he->ms.map->erange_warned)
 206                 ui__warn_map_erange(he->ms.map, sym, ip);
 207 }
 208
 209 static void perf_top__show_details(struct perf_top *top)
 210 {
 211         struct hist_entry *he = top->sym_filter_entry;
 212         struct annotation *notes;
 213         struct symbol *symbol;
 214         int more;
 215
 216         if (!he)
 217                 return;
 218
 219         symbol = he->ms.sym;
 220         notes = symbol__annotation(symbol);
 221
 222         pthread_mutex_lock(&notes->lock);
 223
 224         if (notes->src == NULL)
 225                 goto out_unlock;
 226
 227         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
 228         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
 229
 230         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel,
 231                                        0, top->sym_pcnt_filter, top->print_entries, 4);
 232         if (top->zero)
 233                 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 234         else
 235                 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 236         if (more != 0)
 237                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 238 out_unlock:
 239         pthread_mutex_unlock(&notes->lock);
 240 }
 241
 242 static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
 243                                                      struct addr_location *al,
 244                                                      struct perf_sample *sample)
 245 {
 246         struct hist_entry *he;
 247
 248         pthread_mutex_lock(&evsel->hists.lock);
 249         he = __hists__add_entry(&evsel->hists, al, NULL, sample->period,
 250                                 sample->weight);
 251         pthread_mutex_unlock(&evsel->hists.lock);
 252
 253         if (he == NULL)
 254                 return NULL;
 255
 256         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 257         return he;
 258 }
 259
 260 static void perf_top__print_sym_table(struct perf_top *top)
 261 {
 262         char bf[160];
 263         int printed = 0;
 264         const int win_width = top->winsize.ws_col - 1;
 265
 266         puts(CONSOLE_CLEAR);
 267
 268         perf_top__header_snprintf(top, bf, sizeof(bf));
 269         printf("%s\n", bf);
 270
 271         perf_top__reset_sample_counters(top);
 272
 273         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 274
 275         if (top->sym_evsel->hists.stats.nr_lost_warned !=
 276             top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
 277                 top->sym_evsel->hists.stats.nr_lost_warned =
 278                         top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
 279                 color_fprintf(stdout, PERF_COLOR_RED,
 280                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 281                               top->sym_evsel->hists.stats.nr_lost_warned);
 282                 ++printed;
 283         }
 284
 285         if (top->sym_filter_entry) {
 286                 perf_top__show_details(top);
 287                 return;
 288         }
 289
 290         hists__collapse_resort(&top->sym_evsel->hists);
 291         hists__output_resort(&top->sym_evsel->hists);
 292         hists__decay_entries(&top->sym_evsel->hists,
 293                              top->hide_user_symbols,
 294                              top->hide_kernel_symbols);
 295         hists__output_recalc_col_len(&top->sym_evsel->hists,
 296                                      top->print_entries - printed);
 297         putchar('\n');
 298         hists__fprintf(&top->sym_evsel->hists, false,
 299                        top->print_entries - printed, win_width,
 300                        top->min_percent, stdout);
 301 }
 302
 303 static void prompt_integer(int *target, const char *msg)
 304 {
 305         char *buf = malloc(0), *p;
 306         size_t dummy = 0;
 307         int tmp;
 308
 309         fprintf(stdout, "\n%s: ", msg);
 310         if (getline(&buf, &dummy, stdin) < 0)
 311                 return;
 312
 313         p = strchr(buf, '\n');
 314         if (p)
 315                 *p = 0;
 316
 317         p = buf;
 318         while(*p) {
 319                 if (!isdigit(*p))
 320                         goto out_free;
 321                 p++;
 322         }
 323         tmp = strtoul(buf, NULL, 10);
 324         *target = tmp;
 325 out_free:
 326         free(buf);
 327 }
 328
 329 static void prompt_percent(int *target, const char *msg)
 330 {
 331         int tmp = 0;
 332
 333         prompt_integer(&tmp, msg);
 334         if (tmp >= 0 && tmp <= 100)
 335                 *target = tmp;
 336 }
 337
 338 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 339 {
 340         char *buf = malloc(0), *p;
 341         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 342         struct rb_node *next;
 343         size_t dummy = 0;
 344
 345         /* zero counters of active symbol */
 346         if (syme) {
 347                 __zero_source_counters(syme);
 348                 top->sym_filter_entry = NULL;
 349         }
 350
 351         fprintf(stdout, "\n%s: ", msg);
 352         if (getline(&buf, &dummy, stdin) < 0)
 353                 goto out_free;
 354
 355         p = strchr(buf, '\n');
 356         if (p)
 357                 *p = 0;
 358
 359         next = rb_first(&top->sym_evsel->hists.entries);
 360         while (next) {
 361                 n = rb_entry(next, struct hist_entry, rb_node);
 362                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 363                         found = n;
 364                         break;
 365                 }
 366                 next = rb_next(&n->rb_node);
 367         }
 368
 369         if (!found) {
 370                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 371                 sleep(1);
 372         } else
 373                 perf_top__parse_source(top, found);
 374
 375 out_free:
 376         free(buf);
 377 }
 378
 379 static void perf_top__print_mapped_keys(struct perf_top *top)
 380 {
 381         char *name = NULL;
 382
 383         if (top->sym_filter_entry) {
 384                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 385                 name = sym->name;
 386         }
 387
 388         fprintf(stdout, "\nMapped keys:\n");
 389         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 390         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 391
 392         if (top->evlist->nr_entries > 1)
 393                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 394
 395         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 396
 397         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
 398         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 399         fprintf(stdout, "\t[S]     stop annotation.\n");
 400
 401         fprintf(stdout,
 402                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 403                 top->hide_kernel_symbols ? "yes" : "no");
 404         fprintf(stdout,
 405                 "\t[U]     hide user symbols.               \t(%s)\n",
 406                 top->hide_user_symbols ? "yes" : "no");
 407         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 408         fprintf(stdout, "\t[qQ]    quit.\n");
 409 }
 410
 411 static int perf_top__key_mapped(struct perf_top *top, int c)
 412 {
 413         switch (c) {
 414                 case 'd':
 415                 case 'e':
 416                 case 'f':
 417                 case 'z':
 418                 case 'q':
 419                 case 'Q':
 420                 case 'K':
 421                 case 'U':
 422                 case 'F':
 423                 case 's':
 424                 case 'S':
 425                         return 1;
 426                 case 'E':
 427                         return top->evlist->nr_entries > 1 ? 1 : 0;
 428                 default:
 429                         break;
 430         }
 431
 432         return 0;
 433 }
 434
 435 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 436 {
 437         bool ret = true;
 438
 439         if (!perf_top__key_mapped(top, c)) {
 440                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 441                 struct termios tc, save;
 442
 443                 perf_top__print_mapped_keys(top);
 444                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 445                 fflush(stdout);
 446
 447                 tcgetattr(0, &save);
 448                 tc = save;
 449                 tc.c_lflag &= ~(ICANON | ECHO);
 450                 tc.c_cc[VMIN] = 0;
 451                 tc.c_cc[VTIME] = 0;
 452                 tcsetattr(0, TCSANOW, &tc);
 453
 454                 poll(&stdin_poll, 1, -1);
 455                 c = getc(stdin);
 456
 457                 tcsetattr(0, TCSAFLUSH, &save);
 458                 if (!perf_top__key_mapped(top, c))
 459                         return ret;
 460         }
 461
 462         switch (c) {
 463                 case 'd':
 464                         prompt_integer(&top->delay_secs, "Enter display delay");
 465                         if (top->delay_secs < 1)
 466                                 top->delay_secs = 1;
 467                         break;
 468                 case 'e':
 469                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 470                         if (top->print_entries == 0) {
 471                                 struct sigaction act = {
 472                                         .sa_sigaction = perf_top__sig_winch,
 473                                         .sa_flags     = SA_SIGINFO,
 474                                 };
 475                                 perf_top__sig_winch(SIGWINCH, NULL, top);
 476                                 sigaction(SIGWINCH, &act, NULL);
 477                         } else {
 478                                 signal(SIGWINCH, SIG_DFL);
 479                         }
 480                         break;
 481                 case 'E':
 482                         if (top->evlist->nr_entries > 1) {
 483                                 /* Select 0 as the default event: */
 484                                 int counter = 0;
 485
 486                                 fprintf(stderr, "\nAvailable events:");
 487
 488                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
 489                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
 490
 491                                 prompt_integer(&counter, "Enter details event counter");
 492
 493                                 if (counter >= top->evlist->nr_entries) {
 494                                         top->sym_evsel = perf_evlist__first(top->evlist);
 495                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
 496                                         sleep(1);
 497                                         break;
 498                                 }
 499                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
 500                                         if (top->sym_evsel->idx == counter)
 501                                                 break;
 502                         } else
 503                                 top->sym_evsel = perf_evlist__first(top->evlist);
 504                         break;
 505                 case 'f':
 506                         prompt_integer(&top->count_filter, "Enter display event count filter");
 507                         break;
 508                 case 'F':
 509                         prompt_percent(&top->sym_pcnt_filter,
 510                                        "Enter details display event filter (percent)");
 511                         break;
 512                 case 'K':
 513                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 514                         break;
 515                 case 'q':
 516                 case 'Q':
 517                         printf("exiting.\n");
 518                         if (top->dump_symtab)
 519                                 perf_session__fprintf_dsos(top->session, stderr);
 520                         ret = false;
 521                         break;
 522                 case 's':
 523                         perf_top__prompt_symbol(top, "Enter details symbol");
 524                         break;
 525                 case 'S':
 526                         if (!top->sym_filter_entry)
 527                                 break;
 528                         else {
 529                                 struct hist_entry *syme = top->sym_filter_entry;
 530
 531                                 top->sym_filter_entry = NULL;
 532                                 __zero_source_counters(syme);
 533                         }
 534                         break;
 535                 case 'U':
 536                         top->hide_user_symbols = !top->hide_user_symbols;
 537                         break;
 538                 case 'z':
 539                         top->zero = !top->zero;
 540                         break;
 541                 default:
 542                         break;
 543         }
 544
 545         return ret;
 546 }
 547
 548 static void perf_top__sort_new_samples(void *arg)
 549 {
 550         struct perf_top *t = arg;
 551         perf_top__reset_sample_counters(t);
 552
 553         if (t->evlist->selected != NULL)
 554                 t->sym_evsel = t->evlist->selected;
 555
 556         hists__collapse_resort(&t->sym_evsel->hists);
 557         hists__output_resort(&t->sym_evsel->hists);
 558         hists__decay_entries(&t->sym_evsel->hists,
 559                              t->hide_user_symbols,
 560                              t->hide_kernel_symbols);
 561 }
 562
 563 static void *display_thread_tui(void *arg)
 564 {
 565         struct perf_evsel *pos;
 566         struct perf_top *top = arg;
 567         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 568         struct hist_browser_timer hbt = {
 569                 .timer          = perf_top__sort_new_samples,
 570                 .arg            = top,
 571                 .refresh        = top->delay_secs,
 572         };
 573
 574         perf_top__sort_new_samples(top);
 575
 576         /*
 577          * Initialize the uid_filter_str, in the future the TUI will allow
 578          * Zooming in/out UIDs. For now juse use whatever the user passed
 579          * via --uid.
 580          */
 581         list_for_each_entry(pos, &top->evlist->entries, node)
 582                 pos->hists.uid_filter_str = top->record_opts.target.uid_str;
 583
 584         perf_evlist__tui_browse_hists(top->evlist, help, &hbt, top->min_percent,
 585                                       &top->session->header.env);
 586
 587         done = 1;
 588         return NULL;
 589 }
 590
 591 static void *display_thread(void *arg)
 592 {
 593         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 594         struct termios tc, save;
 595         struct perf_top *top = arg;
 596         int delay_msecs, c;
 597
 598         tcgetattr(0, &save);
 599         tc = save;
 600         tc.c_lflag &= ~(ICANON | ECHO);
 601         tc.c_cc[VMIN] = 0;
 602         tc.c_cc[VTIME] = 0;
 603
 604         pthread__unblock_sigwinch();
 605 repeat:
 606         delay_msecs = top->delay_secs * 1000;
 607         tcsetattr(0, TCSANOW, &tc);
 608         /* trash return*/
 609         getc(stdin);
 610
 611         while (!done) {
 612                 perf_top__print_sym_table(top);
 613                 /*
 614                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 615                  * refresh screen in both cases.
 616                  */
 617                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 618                 case 0:
 619                         continue;
 620                 case -1:
 621                         if (errno == EINTR)
 622                                 continue;
 623                         /* Fall trhu */
 624                 default:
 625                         c = getc(stdin);
 626                         tcsetattr(0, TCSAFLUSH, &save);
 627
 628                         if (perf_top__handle_keypress(top, c))
 629                                 goto repeat;
 630                         done = 1;
 631                 }
 632         }
 633
 634         return NULL;
 635 }
 636
 637 /* Tag samples to be skipped. */
 638 static const char *skip_symbols[] = {
 639         "intel_idle",
 640         "default_idle",
 641         "native_safe_halt",
 642         "cpu_idle",
 643         "enter_idle",
 644         "exit_idle",
 645         "mwait_idle",
 646         "mwait_idle_with_hints",
 647         "poll_idle",
 648         "ppc64_runlatch_off",
 649         "pseries_dedicated_idle_sleep",
 650         NULL
 651 };
 652
 653 static int symbol_filter(struct map *map __maybe_unused, struct symbol *sym)
 654 {
 655         const char *name = sym->name;
 656         int i;
 657
 658         /*
 659          * ppc64 uses function descriptors and appends a '.' to the
 660          * start of every instruction address. Remove it.
 661          */
 662         if (name[0] == '.')
 663                 name++;
 664
 665         if (!strcmp(name, "_text") ||
 666             !strcmp(name, "_etext") ||
 667             !strcmp(name, "_sinittext") ||
 668             !strncmp("init_module", name, 11) ||
 669             !strncmp("cleanup_module", name, 14) ||
 670             strstr(name, "_text_start") ||
 671             strstr(name, "_text_end"))
 672                 return 1;
 673
 674         for (i = 0; skip_symbols[i]; i++) {
 675                 if (!strcmp(skip_symbols[i], name)) {
 676                         sym->ignore = true;
 677                         break;
 678                 }
 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 symbol *parent = NULL;
 692         u64 ip = sample->ip;
 693         struct addr_location al;
 694         int err;
 695
 696         if (!machine && perf_guest) {
 697                 static struct intlist *seen;
 698
 699                 if (!seen)
 700                         seen = intlist__new(NULL);
 701
 702                 if (!intlist__has_entry(seen, sample->pid)) {
 703                         pr_err("Can't find guest [%d]'s kernel information\n",
 704                                 sample->pid);
 705                         intlist__add(seen, sample->pid);
 706                 }
 707                 return;
 708         }
 709
 710         if (!machine) {
 711                 pr_err("%u unprocessable samples recorded.\r",
 712                        top->session->stats.nr_unprocessable_samples++);
 713                 return;
 714         }
 715
 716         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 717                 top->exact_samples++;
 718
 719         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0 ||
 720             al.filtered)
 721                 return;
 722
 723         if (!top->kptr_restrict_warned &&
 724             symbol_conf.kptr_restrict &&
 725             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 726                 ui__warning(
 727 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 728 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 729 "Kernel%s samples will not be resolved.\n",
 730                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 731                           " modules" : "");
 732                 if (use_browser <= 0)
 733                         sleep(5);
 734                 top->kptr_restrict_warned = true;
 735         }
 736
 737         if (al.sym == NULL) {
 738                 const char *msg = "Kernel samples will not be resolved.\n";
 739                 /*
 740                  * As we do lazy loading of symtabs we only will know if the
 741                  * specified vmlinux file is invalid when we actually have a
 742                  * hit in kernel space and then try to load it. So if we get
 743                  * here and there are _no_ symbols in the DSO backing the
 744                  * kernel map, bail out.
 745                  *
 746                  * We may never get here, for instance, if we use -K/
 747                  * --hide-kernel-symbols, even if the user specifies an
 748                  * invalid --vmlinux ;-)
 749                  */
 750                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
 751                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 752                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 753                         if (symbol_conf.vmlinux_name) {
 754                                 ui__warning("The %s file can't be used.\n%s",
 755                                             symbol_conf.vmlinux_name, msg);
 756                         } else {
 757                                 ui__warning("A vmlinux file was not found.\n%s",
 758                                             msg);
 759                         }
 760
 761                         if (use_browser <= 0)
 762                                 sleep(5);
 763                         top->vmlinux_warned = true;
 764                 }
 765         }
 766
 767         if (al.sym == NULL || !al.sym->ignore) {
 768                 struct hist_entry *he;
 769
 770                 if ((sort__has_parent || symbol_conf.use_callchain) &&
 771                     sample->callchain) {
 772                         err = machine__resolve_callchain(machine, evsel,
 773                                                          al.thread, sample,
 774                                                          &parent, &al);
 775                         if (err)
 776                                 return;
 777                 }
 778
 779                 he = perf_evsel__add_hist_entry(evsel, &al, sample);
 780                 if (he == NULL) {
 781                         pr_err("Problem incrementing symbol period, skipping event\n");
 782                         return;
 783                 }
 784
 785                 if (symbol_conf.use_callchain) {
 786                         err = callchain_append(he->callchain, &callchain_cursor,
 787                                                sample->period);
 788                         if (err)
 789                                 return;
 790                 }
 791
 792                 if (sort__has_sym)
 793                         perf_top__record_precise_ip(top, he, evsel->idx, ip);
 794         }
 795
 796         return;
 797 }
 798
 799 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 800 {
 801         struct perf_sample sample;
 802         struct perf_evsel *evsel;
 803         struct perf_session *session = top->session;
 804         union perf_event *event;
 805         struct machine *machine;
 806         u8 origin;
 807         int ret;
 808
 809         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
 810                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
 811                 if (ret) {
 812                         pr_err("Can't parse sample, err = %d\n", ret);
 813                         goto next_event;
 814                 }
 815
 816                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 817                 assert(evsel != NULL);
 818
 819                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 820
 821                 if (event->header.type == PERF_RECORD_SAMPLE)
 822                         ++top->samples;
 823
 824                 switch (origin) {
 825                 case PERF_RECORD_MISC_USER:
 826                         ++top->us_samples;
 827                         if (top->hide_user_symbols)
 828                                 goto next_event;
 829                         machine = &session->machines.host;
 830                         break;
 831                 case PERF_RECORD_MISC_KERNEL:
 832                         ++top->kernel_samples;
 833                         if (top->hide_kernel_symbols)
 834                                 goto next_event;
 835                         machine = &session->machines.host;
 836                         break;
 837                 case PERF_RECORD_MISC_GUEST_KERNEL:
 838                         ++top->guest_kernel_samples;
 839                         machine = perf_session__find_machine(session,
 840                                                              sample.pid);
 841                         break;
 842                 case PERF_RECORD_MISC_GUEST_USER:
 843                         ++top->guest_us_samples;
 844                         /*
 845                          * TODO: we don't process guest user from host side
 846                          * except simple counting.
 847                          */
 848                         /* Fall thru */
 849                 default:
 850                         goto next_event;
 851                 }
 852
 853
 854                 if (event->header.type == PERF_RECORD_SAMPLE) {
 855                         perf_event__process_sample(&top->tool, event, evsel,
 856                                                    &sample, machine);
 857                 } else if (event->header.type < PERF_RECORD_MAX) {
 858                         hists__inc_nr_events(&evsel->hists, event->header.type);
 859                         machine__process_event(machine, event);
 860                 } else
 861                         ++session->stats.nr_unknown_events;
 862 next_event:
 863                 perf_evlist__mmap_consume(top->evlist, idx);
 864         }
 865 }
 866
 867 static void perf_top__mmap_read(struct perf_top *top)
 868 {
 869         int i;
 870
 871         for (i = 0; i < top->evlist->nr_mmaps; i++)
 872                 perf_top__mmap_read_idx(top, i);
 873 }
 874
 875 static int perf_top__start_counters(struct perf_top *top)
 876 {
 877         char msg[512];
 878         struct perf_evsel *counter;
 879         struct perf_evlist *evlist = top->evlist;
 880         struct perf_record_opts *opts = &top->record_opts;
 881
 882         perf_evlist__config(evlist, opts);
 883
 884         list_for_each_entry(counter, &evlist->entries, node) {
 885 try_again:
 886                 if (perf_evsel__open(counter, top->evlist->cpus,
 887                                      top->evlist->threads) < 0) {
 888                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 889                                 if (verbose)
 890                                         ui__warning("%s\n", msg);
 891                                 goto try_again;
 892                         }
 893
 894                         perf_evsel__open_strerror(counter, &opts->target,
 895                                                   errno, msg, sizeof(msg));
 896                         ui__error("%s\n", msg);
 897                         goto out_err;
 898                 }
 899         }
 900
 901         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 902                 ui__error("Failed to mmap with %d (%s)\n",
 903                             errno, strerror(errno));
 904                 goto out_err;
 905         }
 906
 907         return 0;
 908
 909 out_err:
 910         return -1;
 911 }
 912
 913 static int perf_top__setup_sample_type(struct perf_top *top __maybe_unused)
 914 {
 915         if (!sort__has_sym) {
 916                 if (symbol_conf.use_callchain) {
 917                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
 918                         return -EINVAL;
 919                 }
 920         } else if (callchain_param.mode != CHAIN_NONE) {
 921                 if (callchain_register_param(&callchain_param) < 0) {
 922                         ui__error("Can't register callchain params.\n");
 923                         return -EINVAL;
 924                 }
 925         }
 926
 927         return 0;
 928 }
 929
 930 static int __cmd_top(struct perf_top *top)
 931 {
 932         struct perf_record_opts *opts = &top->record_opts;
 933         pthread_t thread;
 934         int ret;
 935         /*
 936          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
 937          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
 938          */
 939         top->session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
 940         if (top->session == NULL)
 941                 return -ENOMEM;
 942
 943         machines__set_symbol_filter(&top->session->machines, symbol_filter);
 944
 945         if (!objdump_path) {
 946                 ret = perf_session_env__lookup_objdump(&top->session->header.env);
 947                 if (ret)
 948                         goto out_delete;
 949         }
 950
 951         ret = perf_top__setup_sample_type(top);
 952         if (ret)
 953                 goto out_delete;
 954
 955         if (perf_target__has_task(&opts->target))
 956                 perf_event__synthesize_thread_map(&top->tool, top->evlist->threads,
 957                                                   perf_event__process,
 958                                                   &top->session->machines.host);
 959         else
 960                 perf_event__synthesize_threads(&top->tool, perf_event__process,
 961                                                &top->session->machines.host);
 962
 963         ret = perf_top__start_counters(top);
 964         if (ret)
 965                 goto out_delete;
 966
 967         top->session->evlist = top->evlist;
 968         perf_session__set_id_hdr_size(top->session);
 969
 970         /*
 971          * When perf is starting the traced process, all the events (apart from
 972          * group members) have enable_on_exec=1 set, so don't spoil it by
 973          * prematurely enabling them.
 974          *
 975          * XXX 'top' still doesn't start workloads like record, trace, but should,
 976          * so leave the check here.
 977          */
 978         if (!perf_target__none(&opts->target))
 979                 perf_evlist__enable(top->evlist);
 980
 981         /* Wait for a minimal set of events before starting the snapshot */
 982         poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
 983
 984         perf_top__mmap_read(top);
 985
 986         ret = -1;
 987         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 988                                                             display_thread), top)) {
 989                 ui__error("Could not create display thread.\n");
 990                 goto out_delete;
 991         }
 992
 993         if (top->realtime_prio) {
 994                 struct sched_param param;
 995
 996                 param.sched_priority = top->realtime_prio;
 997                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 998                         ui__error("Could not set realtime priority.\n");
 999                         goto out_delete;
1000                 }
1001         }
1002
1003         while (!done) {
1004                 u64 hits = top->samples;
1005
1006                 perf_top__mmap_read(top);
1007
1008                 if (hits == top->samples)
1009                         ret = poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
1010         }
1011
1012         ret = 0;
1013 out_delete:
1014         perf_session__delete(top->session);
1015         top->session = NULL;
1016
1017         return ret;
1018 }
1019
1020 static int
1021 callchain_opt(const struct option *opt, const char *arg, int unset)
1022 {
1023         symbol_conf.use_callchain = true;
1024         return record_callchain_opt(opt, arg, unset);
1025 }
1026
1027 static int
1028 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1029 {
1030         symbol_conf.use_callchain = true;
1031         return record_parse_callchain_opt(opt, arg, unset);
1032 }
1033
1034 static int
1035 parse_percent_limit(const struct option *opt, const char *arg,
1036                     int unset __maybe_unused)
1037 {
1038         struct perf_top *top = opt->value;
1039
1040         top->min_percent = strtof(arg, NULL);
1041         return 0;
1042 }
1043
1044 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1045 {
1046         int status;
1047         char errbuf[BUFSIZ];
1048         struct perf_top top = {
1049                 .count_filter        = 5,
1050                 .delay_secs          = 2,
1051                 .record_opts = {
1052                         .mmap_pages     = UINT_MAX,
1053                         .user_freq      = UINT_MAX,
1054                         .user_interval  = ULLONG_MAX,
1055                         .freq           = 4000, /* 4 KHz */
1056                         .target              = {
1057                                 .uses_mmap   = true,
1058                         },
1059                 },
1060                 .sym_pcnt_filter     = 5,
1061         };
1062         struct perf_record_opts *opts = &top.record_opts;
1063         struct perf_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('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1080                     "hide kernel symbols"),
1081         OPT_UINTEGER('m', "mmap-pages", &opts->mmap_pages,
1082                      "number of mmap data pages"),
1083         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1084                     "collect data with this RT SCHED_FIFO priority"),
1085         OPT_INTEGER('d', "delay", &top.delay_secs,
1086                     "number of seconds to delay between refreshes"),
1087         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1088                             "dump the symbol table used for profiling"),
1089         OPT_INTEGER('f', "count-filter", &top.count_filter,
1090                     "only display functions with more events than this"),
1091         OPT_BOOLEAN('g', "group", &opts->group,
1092                             "put the counters into a counter group"),
1093         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1094                     "child tasks do not inherit counters"),
1095         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1096                     "symbol to annotate"),
1097         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1098         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1099         OPT_INTEGER('E', "entries", &top.print_entries,
1100                     "display this many functions"),
1101         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1102                     "hide user symbols"),
1103         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1104         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1105         OPT_INCR('v', "verbose", &verbose,
1106                     "be more verbose (show counter open errors, etc)"),
1107         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1108                    "sort by key(s): pid, comm, dso, symbol, parent, weight, local_weight"),
1109         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1110                     "Show a column with the number of samples"),
1111         OPT_CALLBACK_NOOPT('G', NULL, &top.record_opts,
1112                            NULL, "enables call-graph recording",
1113                            &callchain_opt),
1114         OPT_CALLBACK(0, "call-graph", &top.record_opts,
1115                      "mode[,dump_size]", record_callchain_help,
1116                      &parse_callchain_opt),
1117         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1118                    "ignore callees of these functions in call graphs",
1119                    report_parse_ignore_callees_opt),
1120         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1121                     "Show a column with the sum of periods"),
1122         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1123                    "only consider symbols in these dsos"),
1124         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1125                    "only consider symbols in these comms"),
1126         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1127                    "only consider these symbols"),
1128         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1129                     "Interleave source code with assembly code (default)"),
1130         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1131                     "Display raw encoding of assembly instructions (default)"),
1132         OPT_STRING(0, "objdump", &objdump_path, "path",
1133                     "objdump binary to use for disassembly and annotations"),
1134         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1135                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1136         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1137         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1138                      "Don't show entries under that percent", parse_percent_limit),
1139         OPT_END()
1140         };
1141         const char * const top_usage[] = {
1142                 "perf top [<options>]",
1143                 NULL
1144         };
1145
1146         top.evlist = perf_evlist__new();
1147         if (top.evlist == NULL)
1148                 return -ENOMEM;
1149
1150         argc = parse_options(argc, argv, options, top_usage, 0);
1151         if (argc)
1152                 usage_with_options(top_usage, options);
1153
1154         if (sort_order == default_sort_order)
1155                 sort_order = "dso,symbol";
1156
1157         if (setup_sorting() < 0)
1158                 usage_with_options(top_usage, options);
1159
1160         /* display thread wants entries to be collapsed in a different tree */
1161         sort__need_collapse = 1;
1162
1163         if (top.use_stdio)
1164                 use_browser = 0;
1165         else if (top.use_tui)
1166                 use_browser = 1;
1167
1168         setup_browser(false);
1169
1170         status = perf_target__validate(target);
1171         if (status) {
1172                 perf_target__strerror(target, status, errbuf, BUFSIZ);
1173                 ui__warning("%s", errbuf);
1174         }
1175
1176         status = perf_target__parse_uid(target);
1177         if (status) {
1178                 int saved_errno = errno;
1179
1180                 perf_target__strerror(target, status, errbuf, BUFSIZ);
1181                 ui__error("%s", errbuf);
1182
1183                 status = -saved_errno;
1184                 goto out_delete_evlist;
1185         }
1186
1187         if (perf_target__none(target))
1188                 target->system_wide = true;
1189
1190         if (perf_evlist__create_maps(top.evlist, target) < 0)
1191                 usage_with_options(top_usage, options);
1192
1193         if (!top.evlist->nr_entries &&
1194             perf_evlist__add_default(top.evlist) < 0) {
1195                 ui__error("Not enough memory for event selector list\n");
1196                 goto out_delete_maps;
1197         }
1198
1199         symbol_conf.nr_events = top.evlist->nr_entries;
1200
1201         if (top.delay_secs < 1)
1202                 top.delay_secs = 1;
1203
1204         if (opts->user_interval != ULLONG_MAX)
1205                 opts->default_interval = opts->user_interval;
1206         if (opts->user_freq != UINT_MAX)
1207                 opts->freq = opts->user_freq;
1208
1209         /*
1210          * User specified count overrides default frequency.
1211          */
1212         if (opts->default_interval)
1213                 opts->freq = 0;
1214         else if (opts->freq) {
1215                 opts->default_interval = opts->freq;
1216         } else {
1217                 ui__error("frequency and count are zero, aborting\n");
1218                 status = -EINVAL;
1219                 goto out_delete_maps;
1220         }
1221
1222         top.sym_evsel = perf_evlist__first(top.evlist);
1223
1224         symbol_conf.priv_size = sizeof(struct annotation);
1225
1226         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1227         if (symbol__init() < 0)
1228                 return -1;
1229
1230         sort__setup_elide(stdout);
1231
1232         get_term_dimensions(&top.winsize);
1233         if (top.print_entries == 0) {
1234                 struct sigaction act = {
1235                         .sa_sigaction = perf_top__sig_winch,
1236                         .sa_flags     = SA_SIGINFO,
1237                 };
1238                 perf_top__update_print_entries(&top);
1239                 sigaction(SIGWINCH, &act, NULL);
1240         }
1241
1242         status = __cmd_top(&top);
1243
1244 out_delete_maps:
1245         perf_evlist__delete_maps(top.evlist);
1246 out_delete_evlist:
1247         perf_evlist__delete(top.evlist);
1248
1249         return status;
1250 }