tools/perf/builtin-top.c

   1 /*
   2  * builtin-top.c
   3  *
   4  * Builtin top command: Display a continuously updated profile of
   5  * any workload, CPU or specific PID.
   6  *
   7  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
   8  *               2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   9  *
  10  * Improvements and fixes by:
  11  *
  12  *   Arjan van de Ven <arjan@linux.intel.com>
  13  *   Yanmin Zhang <yanmin.zhang@intel.com>
  14  *   Wu Fengguang <fengguang.wu@intel.com>
  15  *   Mike Galbraith <efault@gmx.de>
  16  *   Paul Mackerras <paulus@samba.org>
  17  *
  18  * Released under the GPL v2. (and only v2, not any later version)
  19  */
  20 #include "builtin.h"
  21
  22 #include "perf.h"
  23
  24 #include "util/annotate.h"
  25 #include "util/cache.h"
  26 #include "util/color.h"
  27 #include "util/evlist.h"
  28 #include "util/evsel.h"
  29 #include "util/machine.h"
  30 #include "util/session.h"
  31 #include "util/symbol.h"
  32 #include "util/thread.h"
  33 #include "util/thread_map.h"
  34 #include "util/top.h"
  35 #include "util/util.h"
  36 #include <linux/rbtree.h>
  37 #include "util/parse-options.h"
  38 #include "util/parse-events.h"
  39 #include "util/cpumap.h"
  40 #include "util/xyarray.h"
  41 #include "util/sort.h"
  42 #include "util/intlist.h"
  43 #include "arch/common.h"
  44
  45 #include "util/debug.h"
  46
  47 #include <assert.h>
  48 #include <elf.h>
  49 #include <fcntl.h>
  50
  51 #include <stdio.h>
  52 #include <termios.h>
  53 #include <unistd.h>
  54 #include <inttypes.h>
  55
  56 #include <errno.h>
  57 #include <time.h>
  58 #include <sched.h>
  59
  60 #include <sys/syscall.h>
  61 #include <sys/ioctl.h>
  62 #include <sys/poll.h>
  63 #include <sys/prctl.h>
  64 #include <sys/wait.h>
  65 #include <sys/uio.h>
  66 #include <sys/utsname.h>
  67 #include <sys/mman.h>
  68
  69 #include <linux/unistd.h>
  70 #include <linux/types.h>
  71
  72 static volatile int done;
  73
  74 #define HEADER_LINE_NR  5
  75
  76 static void perf_top__update_print_entries(struct perf_top *top)
  77 {
  78         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
  79 }
  80
  81 static void perf_top__sig_winch(int sig __maybe_unused,
  82                                 siginfo_t *info __maybe_unused, void *arg)
  83 {
  84         struct perf_top *top = arg;
  85
  86         get_term_dimensions(&top->winsize);
  87         perf_top__update_print_entries(top);
  88 }
  89
  90 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
  91 {
  92         struct symbol *sym;
  93         struct annotation *notes;
  94         struct map *map;
  95         int err = -1;
  96
  97         if (!he || !he->ms.sym)
  98                 return -1;
  99
 100         sym = he->ms.sym;
 101         map = he->ms.map;
 102
 103         /*
 104          * We can't annotate with just /proc/kallsyms
 105          */
 106         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 107             !dso__is_kcore(map->dso)) {
 108                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 109                        "path\n", sym->name);
 110                 sleep(1);
 111                 return -1;
 112         }
 113
 114         notes = symbol__annotation(sym);
 115         if (notes->src != NULL) {
 116                 pthread_mutex_lock(&notes->lock);
 117                 goto out_assign;
 118         }
 119
 120         pthread_mutex_lock(&notes->lock);
 121
 122         if (symbol__alloc_hist(sym) < 0) {
 123                 pthread_mutex_unlock(&notes->lock);
 124                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 125                        sym->name);
 126                 sleep(1);
 127                 return err;
 128         }
 129
 130         err = symbol__annotate(sym, map, 0);
 131         if (err == 0) {
 132 out_assign:
 133                 top->sym_filter_entry = he;
 134         }
 135
 136         pthread_mutex_unlock(&notes->lock);
 137         return err;
 138 }
 139
 140 static void __zero_source_counters(struct hist_entry *he)
 141 {
 142         struct symbol *sym = he->ms.sym;
 143         symbol__annotate_zero_histograms(sym);
 144 }
 145
 146 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 147 {
 148         struct utsname uts;
 149         int err = uname(&uts);
 150
 151         ui__warning("Out of bounds address found:\n\n"
 152                     "Addr:   %" PRIx64 "\n"
 153                     "DSO:    %s %c\n"
 154                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
 155                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 156                     "Arch:   %s\n"
 157                     "Kernel: %s\n"
 158                     "Tools:  %s\n\n"
 159                     "Not all samples will be on the annotation output.\n\n"
 160                     "Please report to linux-kernel@vger.kernel.org\n",
 161                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
 162                     map->start, map->end, sym->start, sym->end,
 163                     sym->binding == STB_GLOBAL ? 'g' :
 164                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 165                     err ? "[unknown]" : uts.machine,
 166                     err ? "[unknown]" : uts.release, perf_version_string);
 167         if (use_browser <= 0)
 168                 sleep(5);
 169
 170         map->erange_warned = true;
 171 }
 172
 173 static void perf_top__record_precise_ip(struct perf_top *top,
 174                                         struct hist_entry *he,
 175                                         int counter, u64 ip)
 176 {
 177         struct annotation *notes;
 178         struct symbol *sym;
 179         int err = 0;
 180
 181         if (he == NULL || he->ms.sym == NULL ||
 182             ((top->sym_filter_entry == NULL ||
 183               top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
 184                 return;
 185
 186         sym = he->ms.sym;
 187         notes = symbol__annotation(sym);
 188
 189         if (pthread_mutex_trylock(&notes->lock))
 190                 return;
 191
 192         ip = he->ms.map->map_ip(he->ms.map, ip);
 193
 194         if (ui__has_annotation())
 195                 err = hist_entry__inc_addr_samples(he, counter, ip);
 196
 197         pthread_mutex_unlock(&notes->lock);
 198
 199         if (err == -ERANGE && !he->ms.map->erange_warned)
 200                 ui__warn_map_erange(he->ms.map, sym, ip);
 201         else if (err == -ENOMEM) {
 202                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 203                        sym->name);
 204                 sleep(1);
 205         }
 206 }
 207
 208 static void perf_top__show_details(struct perf_top *top)
 209 {
 210         struct hist_entry *he = top->sym_filter_entry;
 211         struct annotation *notes;
 212         struct symbol *symbol;
 213         int more;
 214
 215         if (!he)
 216                 return;
 217
 218         symbol = he->ms.sym;
 219         notes = symbol__annotation(symbol);
 220
 221         pthread_mutex_lock(&notes->lock);
 222
 223         if (notes->src == NULL)
 224                 goto out_unlock;
 225
 226         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
 227         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
 228
 229         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel,
 230                                        0, top->sym_pcnt_filter, top->print_entries, 4);
 231         if (top->zero)
 232                 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 233         else
 234                 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 235         if (more != 0)
 236                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 237 out_unlock:
 238         pthread_mutex_unlock(&notes->lock);
 239 }
 240
 241 static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
 242                                                      struct addr_location *al,
 243                                                      struct perf_sample *sample)
 244 {
 245         struct hist_entry *he;
 246
 247         pthread_mutex_lock(&evsel->hists.lock);
 248         he = __hists__add_entry(&evsel->hists, al, NULL, NULL, NULL,
 249                                 sample->period, sample->weight,
 250                                 sample->transaction);
 251         pthread_mutex_unlock(&evsel->hists.lock);
 252         if (he == NULL)
 253                 return NULL;
 254
 255         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 256         return he;
 257 }
 258
 259 static void perf_top__print_sym_table(struct perf_top *top)
 260 {
 261         char bf[160];
 262         int printed = 0;
 263         const int win_width = top->winsize.ws_col - 1;
 264
 265         puts(CONSOLE_CLEAR);
 266
 267         perf_top__header_snprintf(top, bf, sizeof(bf));
 268         printf("%s\n", bf);
 269
 270         perf_top__reset_sample_counters(top);
 271
 272         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 273
 274         if (top->sym_evsel->hists.stats.nr_lost_warned !=
 275             top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
 276                 top->sym_evsel->hists.stats.nr_lost_warned =
 277                         top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
 278                 color_fprintf(stdout, PERF_COLOR_RED,
 279                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 280                               top->sym_evsel->hists.stats.nr_lost_warned);
 281                 ++printed;
 282         }
 283
 284         if (top->sym_filter_entry) {
 285                 perf_top__show_details(top);
 286                 return;
 287         }
 288
 289         hists__collapse_resort(&top->sym_evsel->hists, NULL);
 290         hists__output_resort(&top->sym_evsel->hists);
 291         hists__decay_entries(&top->sym_evsel->hists,
 292                              top->hide_user_symbols,
 293                              top->hide_kernel_symbols);
 294         hists__output_recalc_col_len(&top->sym_evsel->hists,
 295                                      top->print_entries - printed);
 296         putchar('\n');
 297         hists__fprintf(&top->sym_evsel->hists, false,
 298                        top->print_entries - printed, win_width,
 299                        top->min_percent, stdout);
 300 }
 301
 302 static void prompt_integer(int *target, const char *msg)
 303 {
 304         char *buf = malloc(0), *p;
 305         size_t dummy = 0;
 306         int tmp;
 307
 308         fprintf(stdout, "\n%s: ", msg);
 309         if (getline(&buf, &dummy, stdin) < 0)
 310                 return;
 311
 312         p = strchr(buf, '\n');
 313         if (p)
 314                 *p = 0;
 315
 316         p = buf;
 317         while(*p) {
 318                 if (!isdigit(*p))
 319                         goto out_free;
 320                 p++;
 321         }
 322         tmp = strtoul(buf, NULL, 10);
 323         *target = tmp;
 324 out_free:
 325         free(buf);
 326 }
 327
 328 static void prompt_percent(int *target, const char *msg)
 329 {
 330         int tmp = 0;
 331
 332         prompt_integer(&tmp, msg);
 333         if (tmp >= 0 && tmp <= 100)
 334                 *target = tmp;
 335 }
 336
 337 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 338 {
 339         char *buf = malloc(0), *p;
 340         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 341         struct rb_node *next;
 342         size_t dummy = 0;
 343
 344         /* zero counters of active symbol */
 345         if (syme) {
 346                 __zero_source_counters(syme);
 347                 top->sym_filter_entry = NULL;
 348         }
 349
 350         fprintf(stdout, "\n%s: ", msg);
 351         if (getline(&buf, &dummy, stdin) < 0)
 352                 goto out_free;
 353
 354         p = strchr(buf, '\n');
 355         if (p)
 356                 *p = 0;
 357
 358         next = rb_first(&top->sym_evsel->hists.entries);
 359         while (next) {
 360                 n = rb_entry(next, struct hist_entry, rb_node);
 361                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 362                         found = n;
 363                         break;
 364                 }
 365                 next = rb_next(&n->rb_node);
 366         }
 367
 368         if (!found) {
 369                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 370                 sleep(1);
 371         } else
 372                 perf_top__parse_source(top, found);
 373
 374 out_free:
 375         free(buf);
 376 }
 377
 378 static void perf_top__print_mapped_keys(struct perf_top *top)
 379 {
 380         char *name = NULL;
 381
 382         if (top->sym_filter_entry) {
 383                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 384                 name = sym->name;
 385         }
 386
 387         fprintf(stdout, "\nMapped keys:\n");
 388         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 389         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 390
 391         if (top->evlist->nr_entries > 1)
 392                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 393
 394         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 395
 396         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
 397         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 398         fprintf(stdout, "\t[S]     stop annotation.\n");
 399
 400         fprintf(stdout,
 401                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 402                 top->hide_kernel_symbols ? "yes" : "no");
 403         fprintf(stdout,
 404                 "\t[U]     hide user symbols.               \t(%s)\n",
 405                 top->hide_user_symbols ? "yes" : "no");
 406         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 407         fprintf(stdout, "\t[qQ]    quit.\n");
 408 }
 409
 410 static int perf_top__key_mapped(struct perf_top *top, int c)
 411 {
 412         switch (c) {
 413                 case 'd':
 414                 case 'e':
 415                 case 'f':
 416                 case 'z':
 417                 case 'q':
 418                 case 'Q':
 419                 case 'K':
 420                 case 'U':
 421                 case 'F':
 422                 case 's':
 423                 case 'S':
 424                         return 1;
 425                 case 'E':
 426                         return top->evlist->nr_entries > 1 ? 1 : 0;
 427                 default:
 428                         break;
 429         }
 430
 431         return 0;
 432 }
 433
 434 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 435 {
 436         bool ret = true;
 437
 438         if (!perf_top__key_mapped(top, c)) {
 439                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 440                 struct termios tc, save;
 441
 442                 perf_top__print_mapped_keys(top);
 443                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 444                 fflush(stdout);
 445
 446                 tcgetattr(0, &save);
 447                 tc = save;
 448                 tc.c_lflag &= ~(ICANON | ECHO);
 449                 tc.c_cc[VMIN] = 0;
 450                 tc.c_cc[VTIME] = 0;
 451                 tcsetattr(0, TCSANOW, &tc);
 452
 453                 poll(&stdin_poll, 1, -1);
 454                 c = getc(stdin);
 455
 456                 tcsetattr(0, TCSAFLUSH, &save);
 457                 if (!perf_top__key_mapped(top, c))
 458                         return ret;
 459         }
 460
 461         switch (c) {
 462                 case 'd':
 463                         prompt_integer(&top->delay_secs, "Enter display delay");
 464                         if (top->delay_secs < 1)
 465                                 top->delay_secs = 1;
 466                         break;
 467                 case 'e':
 468                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 469                         if (top->print_entries == 0) {
 470                                 struct sigaction act = {
 471                                         .sa_sigaction = perf_top__sig_winch,
 472                                         .sa_flags     = SA_SIGINFO,
 473                                 };
 474                                 perf_top__sig_winch(SIGWINCH, NULL, top);
 475                                 sigaction(SIGWINCH, &act, NULL);
 476                         } else {
 477                                 signal(SIGWINCH, SIG_DFL);
 478                         }
 479                         break;
 480                 case 'E':
 481                         if (top->evlist->nr_entries > 1) {
 482                                 /* Select 0 as the default event: */
 483                                 int counter = 0;
 484
 485                                 fprintf(stderr, "\nAvailable events:");
 486
 487                                 evlist__for_each(top->evlist, top->sym_evsel)
 488                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
 489
 490                                 prompt_integer(&counter, "Enter details event counter");
 491
 492                                 if (counter >= top->evlist->nr_entries) {
 493                                         top->sym_evsel = perf_evlist__first(top->evlist);
 494                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
 495                                         sleep(1);
 496                                         break;
 497                                 }
 498                                 evlist__for_each(top->evlist, top->sym_evsel)
 499                                         if (top->sym_evsel->idx == counter)
 500                                                 break;
 501                         } else
 502                                 top->sym_evsel = perf_evlist__first(top->evlist);
 503                         break;
 504                 case 'f':
 505                         prompt_integer(&top->count_filter, "Enter display event count filter");
 506                         break;
 507                 case 'F':
 508                         prompt_percent(&top->sym_pcnt_filter,
 509                                        "Enter details display event filter (percent)");
 510                         break;
 511                 case 'K':
 512                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 513                         break;
 514                 case 'q':
 515                 case 'Q':
 516                         printf("exiting.\n");
 517                         if (top->dump_symtab)
 518                                 perf_session__fprintf_dsos(top->session, stderr);
 519                         ret = false;
 520                         break;
 521                 case 's':
 522                         perf_top__prompt_symbol(top, "Enter details symbol");
 523                         break;
 524                 case 'S':
 525                         if (!top->sym_filter_entry)
 526                                 break;
 527                         else {
 528                                 struct hist_entry *syme = top->sym_filter_entry;
 529
 530                                 top->sym_filter_entry = NULL;
 531                                 __zero_source_counters(syme);
 532                         }
 533                         break;
 534                 case 'U':
 535                         top->hide_user_symbols = !top->hide_user_symbols;
 536                         break;
 537                 case 'z':
 538                         top->zero = !top->zero;
 539                         break;
 540                 default:
 541                         break;
 542         }
 543
 544         return ret;
 545 }
 546
 547 static void perf_top__sort_new_samples(void *arg)
 548 {
 549         struct perf_top *t = arg;
 550         perf_top__reset_sample_counters(t);
 551
 552         if (t->evlist->selected != NULL)
 553                 t->sym_evsel = t->evlist->selected;
 554
 555         hists__collapse_resort(&t->sym_evsel->hists, NULL);
 556         hists__output_resort(&t->sym_evsel->hists);
 557         hists__decay_entries(&t->sym_evsel->hists,
 558                              t->hide_user_symbols,
 559                              t->hide_kernel_symbols);
 560 }
 561
 562 static void *display_thread_tui(void *arg)
 563 {
 564         struct perf_evsel *pos;
 565         struct perf_top *top = arg;
 566         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 567         struct hist_browser_timer hbt = {
 568                 .timer          = perf_top__sort_new_samples,
 569                 .arg            = top,
 570                 .refresh        = top->delay_secs,
 571         };
 572
 573         perf_top__sort_new_samples(top);
 574
 575         /*
 576          * Initialize the uid_filter_str, in the future the TUI will allow
 577          * Zooming in/out UIDs. For now juse use whatever the user passed
 578          * via --uid.
 579          */
 580         evlist__for_each(top->evlist, pos)
 581                 pos->hists.uid_filter_str = top->record_opts.target.uid_str;
 582
 583         perf_evlist__tui_browse_hists(top->evlist, help, &hbt, top->min_percent,
 584                                       &top->session->header.env);
 585
 586         done = 1;
 587         return NULL;
 588 }
 589
 590 static void *display_thread(void *arg)
 591 {
 592         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 593         struct termios tc, save;
 594         struct perf_top *top = arg;
 595         int delay_msecs, c;
 596
 597         tcgetattr(0, &save);
 598         tc = save;
 599         tc.c_lflag &= ~(ICANON | ECHO);
 600         tc.c_cc[VMIN] = 0;
 601         tc.c_cc[VTIME] = 0;
 602
 603         pthread__unblock_sigwinch();
 604 repeat:
 605         delay_msecs = top->delay_secs * 1000;
 606         tcsetattr(0, TCSANOW, &tc);
 607         /* trash return*/
 608         getc(stdin);
 609
 610         while (!done) {
 611                 perf_top__print_sym_table(top);
 612                 /*
 613                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 614                  * refresh screen in both cases.
 615                  */
 616                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 617                 case 0:
 618                         continue;
 619                 case -1:
 620                         if (errno == EINTR)
 621                                 continue;
 622                         /* Fall trhu */
 623                 default:
 624                         c = getc(stdin);
 625                         tcsetattr(0, TCSAFLUSH, &save);
 626
 627                         if (perf_top__handle_keypress(top, c))
 628                                 goto repeat;
 629                         done = 1;
 630                 }
 631         }
 632
 633         return NULL;
 634 }
 635
 636 static int symbol_filter(struct map *map __maybe_unused, struct symbol *sym)
 637 {
 638         const char *name = sym->name;
 639
 640         /*
 641          * ppc64 uses function descriptors and appends a '.' to the
 642          * start of every instruction address. Remove it.
 643          */
 644         if (name[0] == '.')
 645                 name++;
 646
 647         if (!strcmp(name, "_text") ||
 648             !strcmp(name, "_etext") ||
 649             !strcmp(name, "_sinittext") ||
 650             !strncmp("init_module", name, 11) ||
 651             !strncmp("cleanup_module", name, 14) ||
 652             strstr(name, "_text_start") ||
 653             strstr(name, "_text_end"))
 654                 return 1;
 655
 656         if (symbol__is_idle(sym))
 657                 sym->ignore = true;
 658
 659         return 0;
 660 }
 661
 662 static void perf_event__process_sample(struct perf_tool *tool,
 663                                        const union perf_event *event,
 664                                        struct perf_evsel *evsel,
 665                                        struct perf_sample *sample,
 666                                        struct machine *machine)
 667 {
 668         struct perf_top *top = container_of(tool, struct perf_top, tool);
 669         struct symbol *parent = NULL;
 670         u64 ip = sample->ip;
 671         struct addr_location al;
 672         int err;
 673
 674         if (!machine && perf_guest) {
 675                 static struct intlist *seen;
 676
 677                 if (!seen)
 678                         seen = intlist__new(NULL);
 679
 680                 if (!intlist__has_entry(seen, sample->pid)) {
 681                         pr_err("Can't find guest [%d]'s kernel information\n",
 682                                 sample->pid);
 683                         intlist__add(seen, sample->pid);
 684                 }
 685                 return;
 686         }
 687
 688         if (!machine) {
 689                 pr_err("%u unprocessable samples recorded.\r",
 690                        top->session->stats.nr_unprocessable_samples++);
 691                 return;
 692         }
 693
 694         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 695                 top->exact_samples++;
 696
 697         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0 ||
 698             al.filtered)
 699                 return;
 700
 701         if (!top->kptr_restrict_warned &&
 702             symbol_conf.kptr_restrict &&
 703             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 704                 ui__warning(
 705 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 706 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 707 "Kernel%s samples will not be resolved.\n",
 708                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 709                           " modules" : "");
 710                 if (use_browser <= 0)
 711                         sleep(5);
 712                 top->kptr_restrict_warned = true;
 713         }
 714
 715         if (al.sym == NULL) {
 716                 const char *msg = "Kernel samples will not be resolved.\n";
 717                 /*
 718                  * As we do lazy loading of symtabs we only will know if the
 719                  * specified vmlinux file is invalid when we actually have a
 720                  * hit in kernel space and then try to load it. So if we get
 721                  * here and there are _no_ symbols in the DSO backing the
 722                  * kernel map, bail out.
 723                  *
 724                  * We may never get here, for instance, if we use -K/
 725                  * --hide-kernel-symbols, even if the user specifies an
 726                  * invalid --vmlinux ;-)
 727                  */
 728                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
 729                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 730                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 731                         if (symbol_conf.vmlinux_name) {
 732                                 ui__warning("The %s file can't be used.\n%s",
 733                                             symbol_conf.vmlinux_name, msg);
 734                         } else {
 735                                 ui__warning("A vmlinux file was not found.\n%s",
 736                                             msg);
 737                         }
 738
 739                         if (use_browser <= 0)
 740                                 sleep(5);
 741                         top->vmlinux_warned = true;
 742                 }
 743         }
 744
 745         if (al.sym == NULL || !al.sym->ignore) {
 746                 struct hist_entry *he;
 747
 748                 err = sample__resolve_callchain(sample, &parent, evsel, &al,
 749                                                 top->max_stack);
 750                 if (err)
 751                         return;
 752
 753                 he = perf_evsel__add_hist_entry(evsel, &al, sample);
 754                 if (he == NULL) {
 755                         pr_err("Problem incrementing symbol period, skipping event\n");
 756                         return;
 757                 }
 758
 759                 err = hist_entry__append_callchain(he, sample);
 760                 if (err)
 761                         return;
 762
 763                 if (sort__has_sym)
 764                         perf_top__record_precise_ip(top, he, evsel->idx, ip);
 765         }
 766
 767         return;
 768 }
 769
 770 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 771 {
 772         struct perf_sample sample;
 773         struct perf_evsel *evsel;
 774         struct perf_session *session = top->session;
 775         union perf_event *event;
 776         struct machine *machine;
 777         u8 origin;
 778         int ret;
 779
 780         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
 781                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
 782                 if (ret) {
 783                         pr_err("Can't parse sample, err = %d\n", ret);
 784                         goto next_event;
 785                 }
 786
 787                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 788                 assert(evsel != NULL);
 789
 790                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 791
 792                 if (event->header.type == PERF_RECORD_SAMPLE)
 793                         ++top->samples;
 794
 795                 switch (origin) {
 796                 case PERF_RECORD_MISC_USER:
 797                         ++top->us_samples;
 798                         if (top->hide_user_symbols)
 799                                 goto next_event;
 800                         machine = &session->machines.host;
 801                         break;
 802                 case PERF_RECORD_MISC_KERNEL:
 803                         ++top->kernel_samples;
 804                         if (top->hide_kernel_symbols)
 805                                 goto next_event;
 806                         machine = &session->machines.host;
 807                         break;
 808                 case PERF_RECORD_MISC_GUEST_KERNEL:
 809                         ++top->guest_kernel_samples;
 810                         machine = perf_session__find_machine(session,
 811                                                              sample.pid);
 812                         break;
 813                 case PERF_RECORD_MISC_GUEST_USER:
 814                         ++top->guest_us_samples;
 815                         /*
 816                          * TODO: we don't process guest user from host side
 817                          * except simple counting.
 818                          */
 819                         /* Fall thru */
 820                 default:
 821                         goto next_event;
 822                 }
 823
 824
 825                 if (event->header.type == PERF_RECORD_SAMPLE) {
 826                         perf_event__process_sample(&top->tool, event, evsel,
 827                                                    &sample, machine);
 828                 } else if (event->header.type < PERF_RECORD_MAX) {
 829                         hists__inc_nr_events(&evsel->hists, event->header.type);
 830                         machine__process_event(machine, event, &sample);
 831                 } else
 832                         ++session->stats.nr_unknown_events;
 833 next_event:
 834                 perf_evlist__mmap_consume(top->evlist, idx);
 835         }
 836 }
 837
 838 static void perf_top__mmap_read(struct perf_top *top)
 839 {
 840         int i;
 841
 842         for (i = 0; i < top->evlist->nr_mmaps; i++)
 843                 perf_top__mmap_read_idx(top, i);
 844 }
 845
 846 static int perf_top__start_counters(struct perf_top *top)
 847 {
 848         char msg[512];
 849         struct perf_evsel *counter;
 850         struct perf_evlist *evlist = top->evlist;
 851         struct record_opts *opts = &top->record_opts;
 852
 853         perf_evlist__config(evlist, opts);
 854
 855         evlist__for_each(evlist, counter) {
 856 try_again:
 857                 if (perf_evsel__open(counter, top->evlist->cpus,
 858                                      top->evlist->threads) < 0) {
 859                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 860                                 if (verbose)
 861                                         ui__warning("%s\n", msg);
 862                                 goto try_again;
 863                         }
 864
 865                         perf_evsel__open_strerror(counter, &opts->target,
 866                                                   errno, msg, sizeof(msg));
 867                         ui__error("%s\n", msg);
 868                         goto out_err;
 869                 }
 870         }
 871
 872         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 873                 ui__error("Failed to mmap with %d (%s)\n",
 874                             errno, strerror(errno));
 875                 goto out_err;
 876         }
 877
 878         return 0;
 879
 880 out_err:
 881         return -1;
 882 }
 883
 884 static int perf_top__setup_sample_type(struct perf_top *top __maybe_unused)
 885 {
 886         if (!sort__has_sym) {
 887                 if (symbol_conf.use_callchain) {
 888                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
 889                         return -EINVAL;
 890                 }
 891         } else if (callchain_param.mode != CHAIN_NONE) {
 892                 if (callchain_register_param(&callchain_param) < 0) {
 893                         ui__error("Can't register callchain params.\n");
 894                         return -EINVAL;
 895                 }
 896         }
 897
 898         return 0;
 899 }
 900
 901 static int __cmd_top(struct perf_top *top)
 902 {
 903         struct record_opts *opts = &top->record_opts;
 904         pthread_t thread;
 905         int ret;
 906
 907         top->session = perf_session__new(NULL, false, NULL);
 908         if (top->session == NULL)
 909                 return -ENOMEM;
 910
 911         machines__set_symbol_filter(&top->session->machines, symbol_filter);
 912
 913         if (!objdump_path) {
 914                 ret = perf_session_env__lookup_objdump(&top->session->header.env);
 915                 if (ret)
 916                         goto out_delete;
 917         }
 918
 919         ret = perf_top__setup_sample_type(top);
 920         if (ret)
 921                 goto out_delete;
 922
 923         machine__synthesize_threads(&top->session->machines.host, &opts->target,
 924                                     top->evlist->threads, false);
 925         ret = perf_top__start_counters(top);
 926         if (ret)
 927                 goto out_delete;
 928
 929         top->session->evlist = top->evlist;
 930         perf_session__set_id_hdr_size(top->session);
 931
 932         /*
 933          * When perf is starting the traced process, all the events (apart from
 934          * group members) have enable_on_exec=1 set, so don't spoil it by
 935          * prematurely enabling them.
 936          *
 937          * XXX 'top' still doesn't start workloads like record, trace, but should,
 938          * so leave the check here.
 939          */
 940         if (!target__none(&opts->target))
 941                 perf_evlist__enable(top->evlist);
 942
 943         /* Wait for a minimal set of events before starting the snapshot */
 944         poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
 945
 946         perf_top__mmap_read(top);
 947
 948         ret = -1;
 949         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 950                                                             display_thread), top)) {
 951                 ui__error("Could not create display thread.\n");
 952                 goto out_delete;
 953         }
 954
 955         if (top->realtime_prio) {
 956                 struct sched_param param;
 957
 958                 param.sched_priority = top->realtime_prio;
 959                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 960                         ui__error("Could not set realtime priority.\n");
 961                         goto out_delete;
 962                 }
 963         }
 964
 965         while (!done) {
 966                 u64 hits = top->samples;
 967
 968                 perf_top__mmap_read(top);
 969
 970                 if (hits == top->samples)
 971                         ret = poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
 972         }
 973
 974         ret = 0;
 975 out_delete:
 976         perf_session__delete(top->session);
 977         top->session = NULL;
 978
 979         return ret;
 980 }
 981
 982 static int
 983 callchain_opt(const struct option *opt, const char *arg, int unset)
 984 {
 985         symbol_conf.use_callchain = true;
 986         return record_callchain_opt(opt, arg, unset);
 987 }
 988
 989 static int
 990 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
 991 {
 992         symbol_conf.use_callchain = true;
 993         return record_parse_callchain_opt(opt, arg, unset);
 994 }
 995
 996 static int perf_top_config(const char *var, const char *value, void *cb)
 997 {
 998         struct perf_top *top = cb;
 999
1000         if (!strcmp(var, "top.call-graph"))
1001                 return record_parse_callchain(value, &top->record_opts);
1002
1003         return perf_default_config(var, value, cb);
1004 }
1005
1006 static int
1007 parse_percent_limit(const struct option *opt, const char *arg,
1008                     int unset __maybe_unused)
1009 {
1010         struct perf_top *top = opt->value;
1011
1012         top->min_percent = strtof(arg, NULL);
1013         return 0;
1014 }
1015
1016 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1017 {
1018         int status = -1;
1019         char errbuf[BUFSIZ];
1020         struct perf_top top = {
1021                 .count_filter        = 5,
1022                 .delay_secs          = 2,
1023                 .record_opts = {
1024                         .mmap_pages     = UINT_MAX,
1025                         .user_freq      = UINT_MAX,
1026                         .user_interval  = ULLONG_MAX,
1027                         .freq           = 4000, /* 4 KHz */
1028                         .target         = {
1029                                 .uses_mmap   = true,
1030                         },
1031                 },
1032                 .max_stack           = PERF_MAX_STACK_DEPTH,
1033                 .sym_pcnt_filter     = 5,
1034         };
1035         struct record_opts *opts = &top.record_opts;
1036         struct target *target = &opts->target;
1037         const struct option options[] = {
1038         OPT_CALLBACK('e', "event", &top.evlist, "event",
1039                      "event selector. use 'perf list' to list available events",
1040                      parse_events_option),
1041         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1042         OPT_STRING('p', "pid", &target->pid, "pid",
1043                     "profile events on existing process id"),
1044         OPT_STRING('t', "tid", &target->tid, "tid",
1045                     "profile events on existing thread id"),
1046         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1047                             "system-wide collection from all CPUs"),
1048         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1049                     "list of cpus to monitor"),
1050         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1051                    "file", "vmlinux pathname"),
1052         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1053                     "don't load vmlinux even if found"),
1054         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1055                     "hide kernel symbols"),
1056         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1057                      "number of mmap data pages",
1058                      perf_evlist__parse_mmap_pages),
1059         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1060                     "collect data with this RT SCHED_FIFO priority"),
1061         OPT_INTEGER('d', "delay", &top.delay_secs,
1062                     "number of seconds to delay between refreshes"),
1063         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1064                             "dump the symbol table used for profiling"),
1065         OPT_INTEGER('f', "count-filter", &top.count_filter,
1066                     "only display functions with more events than this"),
1067         OPT_BOOLEAN(0, "group", &opts->group,
1068                             "put the counters into a counter group"),
1069         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1070                     "child tasks do not inherit counters"),
1071         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1072                     "symbol to annotate"),
1073         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1074         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1075         OPT_INTEGER('E', "entries", &top.print_entries,
1076                     "display this many functions"),
1077         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1078                     "hide user symbols"),
1079         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1080         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1081         OPT_INCR('v', "verbose", &verbose,
1082                     "be more verbose (show counter open errors, etc)"),
1083         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1084                    "sort by key(s): pid, comm, dso, symbol, parent, weight, local_weight,"
1085                    " abort, in_tx, transaction"),
1086         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1087                     "Show a column with the number of samples"),
1088         OPT_CALLBACK_NOOPT('g', NULL, &top.record_opts,
1089                            NULL, "enables call-graph recording",
1090                            &callchain_opt),
1091         OPT_CALLBACK(0, "call-graph", &top.record_opts,
1092                      "mode[,dump_size]", record_callchain_help,
1093                      &parse_callchain_opt),
1094         OPT_INTEGER(0, "max-stack", &top.max_stack,
1095                     "Set the maximum stack depth when parsing the callchain. "
1096                     "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
1097         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1098                    "ignore callees of these functions in call graphs",
1099                    report_parse_ignore_callees_opt),
1100         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1101                     "Show a column with the sum of periods"),
1102         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1103                    "only consider symbols in these dsos"),
1104         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1105                    "only consider symbols in these comms"),
1106         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1107                    "only consider these symbols"),
1108         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1109                     "Interleave source code with assembly code (default)"),
1110         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1111                     "Display raw encoding of assembly instructions (default)"),
1112         OPT_STRING(0, "objdump", &objdump_path, "path",
1113                     "objdump binary to use for disassembly and annotations"),
1114         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1115                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1116         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1117         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1118                      "Don't show entries under that percent", parse_percent_limit),
1119         OPT_END()
1120         };
1121         const char * const top_usage[] = {
1122                 "perf top [<options>]",
1123                 NULL
1124         };
1125
1126         top.evlist = perf_evlist__new();
1127         if (top.evlist == NULL)
1128                 return -ENOMEM;
1129
1130         perf_config(perf_top_config, &top);
1131
1132         argc = parse_options(argc, argv, options, top_usage, 0);
1133         if (argc)
1134                 usage_with_options(top_usage, options);
1135
1136         if (sort_order == default_sort_order)
1137                 sort_order = "dso,symbol";
1138
1139         if (setup_sorting() < 0) {
1140                 parse_options_usage(top_usage, options, "s", 1);
1141                 goto out_delete_evlist;
1142         }
1143
1144         /* display thread wants entries to be collapsed in a different tree */
1145         sort__need_collapse = 1;
1146
1147         if (top.use_stdio)
1148                 use_browser = 0;
1149         else if (top.use_tui)
1150                 use_browser = 1;
1151
1152         setup_browser(false);
1153
1154         status = target__validate(target);
1155         if (status) {
1156                 target__strerror(target, status, errbuf, BUFSIZ);
1157                 ui__warning("%s\n", errbuf);
1158         }
1159
1160         status = target__parse_uid(target);
1161         if (status) {
1162                 int saved_errno = errno;
1163
1164                 target__strerror(target, status, errbuf, BUFSIZ);
1165                 ui__error("%s\n", errbuf);
1166
1167                 status = -saved_errno;
1168                 goto out_delete_evlist;
1169         }
1170
1171         if (target__none(target))
1172                 target->system_wide = true;
1173
1174         if (perf_evlist__create_maps(top.evlist, target) < 0)
1175                 usage_with_options(top_usage, options);
1176
1177         if (!top.evlist->nr_entries &&
1178             perf_evlist__add_default(top.evlist) < 0) {
1179                 ui__error("Not enough memory for event selector list\n");
1180                 goto out_delete_evlist;
1181         }
1182
1183         symbol_conf.nr_events = top.evlist->nr_entries;
1184
1185         if (top.delay_secs < 1)
1186                 top.delay_secs = 1;
1187
1188         if (record_opts__config(opts)) {
1189                 status = -EINVAL;
1190                 goto out_delete_evlist;
1191         }
1192
1193         top.sym_evsel = perf_evlist__first(top.evlist);
1194
1195         symbol_conf.priv_size = sizeof(struct annotation);
1196
1197         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1198         if (symbol__init() < 0)
1199                 return -1;
1200
1201         sort__setup_elide(stdout);
1202
1203         get_term_dimensions(&top.winsize);
1204         if (top.print_entries == 0) {
1205                 struct sigaction act = {
1206                         .sa_sigaction = perf_top__sig_winch,
1207                         .sa_flags     = SA_SIGINFO,
1208                 };
1209                 perf_top__update_print_entries(&top);
1210                 sigaction(SIGWINCH, &act, NULL);
1211         }
1212
1213         status = __cmd_top(&top);
1214
1215 out_delete_evlist:
1216         perf_evlist__delete(top.evlist);
1217
1218         return status;
1219 }