tools/perf/builtin-top.c

   1 /*
   2  * builtin-top.c
   3  *
   4  * Builtin top command: Display a continuously updated profile of
   5  * any workload, CPU or specific PID.
   6  *
   7  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
   8  *               2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   9  *
  10  * Improvements and fixes by:
  11  *
  12  *   Arjan van de Ven <arjan@linux.intel.com>
  13  *   Yanmin Zhang <yanmin.zhang@intel.com>
  14  *   Wu Fengguang <fengguang.wu@intel.com>
  15  *   Mike Galbraith <efault@gmx.de>
  16  *   Paul Mackerras <paulus@samba.org>
  17  *
  18  * Released under the GPL v2. (and only v2, not any later version)
  19  */
  20 #include "builtin.h"
  21
  22 #include "perf.h"
  23
  24 #include "util/annotate.h"
  25 #include "util/cache.h"
  26 #include "util/color.h"
  27 #include "util/evlist.h"
  28 #include "util/evsel.h"
  29 #include "util/machine.h"
  30 #include "util/session.h"
  31 #include "util/symbol.h"
  32 #include "util/thread.h"
  33 #include "util/thread_map.h"
  34 #include "util/top.h"
  35 #include "util/util.h"
  36 #include <linux/rbtree.h>
  37 #include "util/parse-options.h"
  38 #include "util/parse-events.h"
  39 #include "util/cpumap.h"
  40 #include "util/xyarray.h"
  41 #include "util/sort.h"
  42 #include "util/intlist.h"
  43 #include "arch/common.h"
  44
  45 #include "util/debug.h"
  46
  47 #include <assert.h>
  48 #include <elf.h>
  49 #include <fcntl.h>
  50
  51 #include <stdio.h>
  52 #include <termios.h>
  53 #include <unistd.h>
  54 #include <inttypes.h>
  55
  56 #include <errno.h>
  57 #include <time.h>
  58 #include <sched.h>
  59
  60 #include <sys/syscall.h>
  61 #include <sys/ioctl.h>
  62 #include <poll.h>
  63 #include <sys/prctl.h>
  64 #include <sys/wait.h>
  65 #include <sys/uio.h>
  66 #include <sys/utsname.h>
  67 #include <sys/mman.h>
  68
  69 #include <linux/types.h>
  70
  71 static volatile int done;
  72
  73 #define HEADER_LINE_NR  5
  74
  75 static void perf_top__update_print_entries(struct perf_top *top)
  76 {
  77         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
  78 }
  79
  80 static void perf_top__sig_winch(int sig __maybe_unused,
  81                                 siginfo_t *info __maybe_unused, void *arg)
  82 {
  83         struct perf_top *top = arg;
  84
  85         get_term_dimensions(&top->winsize);
  86         perf_top__update_print_entries(top);
  87 }
  88
  89 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
  90 {
  91         struct symbol *sym;
  92         struct annotation *notes;
  93         struct map *map;
  94         int err = -1;
  95
  96         if (!he || !he->ms.sym)
  97                 return -1;
  98
  99         sym = he->ms.sym;
 100         map = he->ms.map;
 101
 102         /*
 103          * We can't annotate with just /proc/kallsyms
 104          */
 105         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 106             !dso__is_kcore(map->dso)) {
 107                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 108                        "path\n", sym->name);
 109                 sleep(1);
 110                 return -1;
 111         }
 112
 113         notes = symbol__annotation(sym);
 114         if (notes->src != NULL) {
 115                 pthread_mutex_lock(&notes->lock);
 116                 goto out_assign;
 117         }
 118
 119         pthread_mutex_lock(&notes->lock);
 120
 121         if (symbol__alloc_hist(sym) < 0) {
 122                 pthread_mutex_unlock(&notes->lock);
 123                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 124                        sym->name);
 125                 sleep(1);
 126                 return err;
 127         }
 128
 129         err = symbol__annotate(sym, map, 0);
 130         if (err == 0) {
 131 out_assign:
 132                 top->sym_filter_entry = he;
 133         }
 134
 135         pthread_mutex_unlock(&notes->lock);
 136         return err;
 137 }
 138
 139 static void __zero_source_counters(struct hist_entry *he)
 140 {
 141         struct symbol *sym = he->ms.sym;
 142         symbol__annotate_zero_histograms(sym);
 143 }
 144
 145 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 146 {
 147         struct utsname uts;
 148         int err = uname(&uts);
 149
 150         ui__warning("Out of bounds address found:\n\n"
 151                     "Addr:   %" PRIx64 "\n"
 152                     "DSO:    %s %c\n"
 153                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
 154                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 155                     "Arch:   %s\n"
 156                     "Kernel: %s\n"
 157                     "Tools:  %s\n\n"
 158                     "Not all samples will be on the annotation output.\n\n"
 159                     "Please report to linux-kernel@vger.kernel.org\n",
 160                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
 161                     map->start, map->end, sym->start, sym->end,
 162                     sym->binding == STB_GLOBAL ? 'g' :
 163                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 164                     err ? "[unknown]" : uts.machine,
 165                     err ? "[unknown]" : uts.release, perf_version_string);
 166         if (use_browser <= 0)
 167                 sleep(5);
 168
 169         map->erange_warned = true;
 170 }
 171
 172 static void perf_top__record_precise_ip(struct perf_top *top,
 173                                         struct hist_entry *he,
 174                                         int counter, u64 ip)
 175 {
 176         struct annotation *notes;
 177         struct symbol *sym;
 178         int err = 0;
 179
 180         if (he == NULL || he->ms.sym == NULL ||
 181             ((top->sym_filter_entry == NULL ||
 182               top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
 183                 return;
 184
 185         sym = he->ms.sym;
 186         notes = symbol__annotation(sym);
 187
 188         if (pthread_mutex_trylock(&notes->lock))
 189                 return;
 190
 191         ip = he->ms.map->map_ip(he->ms.map, ip);
 192
 193         if (ui__has_annotation())
 194                 err = hist_entry__inc_addr_samples(he, counter, ip);
 195
 196         pthread_mutex_unlock(&notes->lock);
 197
 198         /*
 199          * This function is now called with he->hists->lock held.
 200          * Release it before going to sleep.
 201          */
 202         pthread_mutex_unlock(&he->hists->lock);
 203
 204         if (err == -ERANGE && !he->ms.map->erange_warned)
 205                 ui__warn_map_erange(he->ms.map, sym, ip);
 206         else if (err == -ENOMEM) {
 207                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 208                        sym->name);
 209                 sleep(1);
 210         }
 211
 212         pthread_mutex_lock(&he->hists->lock);
 213 }
 214
 215 static void perf_top__show_details(struct perf_top *top)
 216 {
 217         struct hist_entry *he = top->sym_filter_entry;
 218         struct annotation *notes;
 219         struct symbol *symbol;
 220         int more;
 221
 222         if (!he)
 223                 return;
 224
 225         symbol = he->ms.sym;
 226         notes = symbol__annotation(symbol);
 227
 228         pthread_mutex_lock(&notes->lock);
 229
 230         if (notes->src == NULL)
 231                 goto out_unlock;
 232
 233         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
 234         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
 235
 236         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel,
 237                                        0, top->sym_pcnt_filter, top->print_entries, 4);
 238
 239         if (top->evlist->enabled) {
 240                 if (top->zero)
 241                         symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 242                 else
 243                         symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 244         }
 245         if (more != 0)
 246                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 247 out_unlock:
 248         pthread_mutex_unlock(&notes->lock);
 249 }
 250
 251 static void perf_top__print_sym_table(struct perf_top *top)
 252 {
 253         char bf[160];
 254         int printed = 0;
 255         const int win_width = top->winsize.ws_col - 1;
 256         struct hists *hists = evsel__hists(top->sym_evsel);
 257
 258         puts(CONSOLE_CLEAR);
 259
 260         perf_top__header_snprintf(top, bf, sizeof(bf));
 261         printf("%s\n", bf);
 262
 263         perf_top__reset_sample_counters(top);
 264
 265         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 266
 267         if (hists->stats.nr_lost_warned !=
 268             hists->stats.nr_events[PERF_RECORD_LOST]) {
 269                 hists->stats.nr_lost_warned =
 270                               hists->stats.nr_events[PERF_RECORD_LOST];
 271                 color_fprintf(stdout, PERF_COLOR_RED,
 272                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 273                               hists->stats.nr_lost_warned);
 274                 ++printed;
 275         }
 276
 277         if (top->sym_filter_entry) {
 278                 perf_top__show_details(top);
 279                 return;
 280         }
 281
 282         if (top->evlist->enabled) {
 283                 if (top->zero) {
 284                         hists__delete_entries(hists);
 285                 } else {
 286                         hists__decay_entries(hists, top->hide_user_symbols,
 287                                              top->hide_kernel_symbols);
 288                 }
 289         }
 290
 291         hists__collapse_resort(hists, NULL);
 292         hists__output_resort(hists, NULL);
 293
 294         hists__output_recalc_col_len(hists, top->print_entries - printed);
 295         putchar('\n');
 296         hists__fprintf(hists, false, top->print_entries - printed, win_width,
 297                        top->min_percent, stdout);
 298 }
 299
 300 static void prompt_integer(int *target, const char *msg)
 301 {
 302         char *buf = malloc(0), *p;
 303         size_t dummy = 0;
 304         int tmp;
 305
 306         fprintf(stdout, "\n%s: ", msg);
 307         if (getline(&buf, &dummy, stdin) < 0)
 308                 return;
 309
 310         p = strchr(buf, '\n');
 311         if (p)
 312                 *p = 0;
 313
 314         p = buf;
 315         while(*p) {
 316                 if (!isdigit(*p))
 317                         goto out_free;
 318                 p++;
 319         }
 320         tmp = strtoul(buf, NULL, 10);
 321         *target = tmp;
 322 out_free:
 323         free(buf);
 324 }
 325
 326 static void prompt_percent(int *target, const char *msg)
 327 {
 328         int tmp = 0;
 329
 330         prompt_integer(&tmp, msg);
 331         if (tmp >= 0 && tmp <= 100)
 332                 *target = tmp;
 333 }
 334
 335 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 336 {
 337         char *buf = malloc(0), *p;
 338         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 339         struct hists *hists = evsel__hists(top->sym_evsel);
 340         struct rb_node *next;
 341         size_t dummy = 0;
 342
 343         /* zero counters of active symbol */
 344         if (syme) {
 345                 __zero_source_counters(syme);
 346                 top->sym_filter_entry = NULL;
 347         }
 348
 349         fprintf(stdout, "\n%s: ", msg);
 350         if (getline(&buf, &dummy, stdin) < 0)
 351                 goto out_free;
 352
 353         p = strchr(buf, '\n');
 354         if (p)
 355                 *p = 0;
 356
 357         next = rb_first(&hists->entries);
 358         while (next) {
 359                 n = rb_entry(next, struct hist_entry, rb_node);
 360                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 361                         found = n;
 362                         break;
 363                 }
 364                 next = rb_next(&n->rb_node);
 365         }
 366
 367         if (!found) {
 368                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 369                 sleep(1);
 370         } else
 371                 perf_top__parse_source(top, found);
 372
 373 out_free:
 374         free(buf);
 375 }
 376
 377 static void perf_top__print_mapped_keys(struct perf_top *top)
 378 {
 379         char *name = NULL;
 380
 381         if (top->sym_filter_entry) {
 382                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 383                 name = sym->name;
 384         }
 385
 386         fprintf(stdout, "\nMapped keys:\n");
 387         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 388         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 389
 390         if (top->evlist->nr_entries > 1)
 391                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 392
 393         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 394
 395         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
 396         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 397         fprintf(stdout, "\t[S]     stop annotation.\n");
 398
 399         fprintf(stdout,
 400                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 401                 top->hide_kernel_symbols ? "yes" : "no");
 402         fprintf(stdout,
 403                 "\t[U]     hide user symbols.               \t(%s)\n",
 404                 top->hide_user_symbols ? "yes" : "no");
 405         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 406         fprintf(stdout, "\t[qQ]    quit.\n");
 407 }
 408
 409 static int perf_top__key_mapped(struct perf_top *top, int c)
 410 {
 411         switch (c) {
 412                 case 'd':
 413                 case 'e':
 414                 case 'f':
 415                 case 'z':
 416                 case 'q':
 417                 case 'Q':
 418                 case 'K':
 419                 case 'U':
 420                 case 'F':
 421                 case 's':
 422                 case 'S':
 423                         return 1;
 424                 case 'E':
 425                         return top->evlist->nr_entries > 1 ? 1 : 0;
 426                 default:
 427                         break;
 428         }
 429
 430         return 0;
 431 }
 432
 433 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 434 {
 435         bool ret = true;
 436
 437         if (!perf_top__key_mapped(top, c)) {
 438                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 439                 struct termios save;
 440
 441                 perf_top__print_mapped_keys(top);
 442                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 443                 fflush(stdout);
 444
 445                 set_term_quiet_input(&save);
 446
 447                 poll(&stdin_poll, 1, -1);
 448                 c = getc(stdin);
 449
 450                 tcsetattr(0, TCSAFLUSH, &save);
 451                 if (!perf_top__key_mapped(top, c))
 452                         return ret;
 453         }
 454
 455         switch (c) {
 456                 case 'd':
 457                         prompt_integer(&top->delay_secs, "Enter display delay");
 458                         if (top->delay_secs < 1)
 459                                 top->delay_secs = 1;
 460                         break;
 461                 case 'e':
 462                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 463                         if (top->print_entries == 0) {
 464                                 struct sigaction act = {
 465                                         .sa_sigaction = perf_top__sig_winch,
 466                                         .sa_flags     = SA_SIGINFO,
 467                                 };
 468                                 perf_top__sig_winch(SIGWINCH, NULL, top);
 469                                 sigaction(SIGWINCH, &act, NULL);
 470                         } else {
 471                                 signal(SIGWINCH, SIG_DFL);
 472                         }
 473                         break;
 474                 case 'E':
 475                         if (top->evlist->nr_entries > 1) {
 476                                 /* Select 0 as the default event: */
 477                                 int counter = 0;
 478
 479                                 fprintf(stderr, "\nAvailable events:");
 480
 481                                 evlist__for_each(top->evlist, top->sym_evsel)
 482                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
 483
 484                                 prompt_integer(&counter, "Enter details event counter");
 485
 486                                 if (counter >= top->evlist->nr_entries) {
 487                                         top->sym_evsel = perf_evlist__first(top->evlist);
 488                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
 489                                         sleep(1);
 490                                         break;
 491                                 }
 492                                 evlist__for_each(top->evlist, top->sym_evsel)
 493                                         if (top->sym_evsel->idx == counter)
 494                                                 break;
 495                         } else
 496                                 top->sym_evsel = perf_evlist__first(top->evlist);
 497                         break;
 498                 case 'f':
 499                         prompt_integer(&top->count_filter, "Enter display event count filter");
 500                         break;
 501                 case 'F':
 502                         prompt_percent(&top->sym_pcnt_filter,
 503                                        "Enter details display event filter (percent)");
 504                         break;
 505                 case 'K':
 506                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 507                         break;
 508                 case 'q':
 509                 case 'Q':
 510                         printf("exiting.\n");
 511                         if (top->dump_symtab)
 512                                 perf_session__fprintf_dsos(top->session, stderr);
 513                         ret = false;
 514                         break;
 515                 case 's':
 516                         perf_top__prompt_symbol(top, "Enter details symbol");
 517                         break;
 518                 case 'S':
 519                         if (!top->sym_filter_entry)
 520                                 break;
 521                         else {
 522                                 struct hist_entry *syme = top->sym_filter_entry;
 523
 524                                 top->sym_filter_entry = NULL;
 525                                 __zero_source_counters(syme);
 526                         }
 527                         break;
 528                 case 'U':
 529                         top->hide_user_symbols = !top->hide_user_symbols;
 530                         break;
 531                 case 'z':
 532                         top->zero = !top->zero;
 533                         break;
 534                 default:
 535                         break;
 536         }
 537
 538         return ret;
 539 }
 540
 541 static void perf_top__sort_new_samples(void *arg)
 542 {
 543         struct perf_top *t = arg;
 544         struct hists *hists;
 545
 546         perf_top__reset_sample_counters(t);
 547
 548         if (t->evlist->selected != NULL)
 549                 t->sym_evsel = t->evlist->selected;
 550
 551         hists = evsel__hists(t->sym_evsel);
 552
 553         if (t->evlist->enabled) {
 554                 if (t->zero) {
 555                         hists__delete_entries(hists);
 556                 } else {
 557                         hists__decay_entries(hists, t->hide_user_symbols,
 558                                              t->hide_kernel_symbols);
 559                 }
 560         }
 561
 562         hists__collapse_resort(hists, NULL);
 563         hists__output_resort(hists, NULL);
 564 }
 565
 566 static void *display_thread_tui(void *arg)
 567 {
 568         struct perf_evsel *pos;
 569         struct perf_top *top = arg;
 570         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 571         struct hist_browser_timer hbt = {
 572                 .timer          = perf_top__sort_new_samples,
 573                 .arg            = top,
 574                 .refresh        = top->delay_secs,
 575         };
 576
 577         perf_top__sort_new_samples(top);
 578
 579         /*
 580          * Initialize the uid_filter_str, in the future the TUI will allow
 581          * Zooming in/out UIDs. For now juse use whatever the user passed
 582          * via --uid.
 583          */
 584         evlist__for_each(top->evlist, pos) {
 585                 struct hists *hists = evsel__hists(pos);
 586                 hists->uid_filter_str = top->record_opts.target.uid_str;
 587         }
 588
 589         perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
 590                                       top->min_percent,
 591                                       &top->session->header.env);
 592
 593         done = 1;
 594         return NULL;
 595 }
 596
 597 static void display_sig(int sig __maybe_unused)
 598 {
 599         done = 1;
 600 }
 601
 602 static void display_setup_sig(void)
 603 {
 604         signal(SIGSEGV, sighandler_dump_stack);
 605         signal(SIGFPE, sighandler_dump_stack);
 606         signal(SIGINT,  display_sig);
 607         signal(SIGQUIT, display_sig);
 608         signal(SIGTERM, display_sig);
 609 }
 610
 611 static void *display_thread(void *arg)
 612 {
 613         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 614         struct termios save;
 615         struct perf_top *top = arg;
 616         int delay_msecs, c;
 617
 618         display_setup_sig();
 619         pthread__unblock_sigwinch();
 620 repeat:
 621         delay_msecs = top->delay_secs * 1000;
 622         set_term_quiet_input(&save);
 623         /* trash return*/
 624         getc(stdin);
 625
 626         while (!done) {
 627                 perf_top__print_sym_table(top);
 628                 /*
 629                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 630                  * refresh screen in both cases.
 631                  */
 632                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 633                 case 0:
 634                         continue;
 635                 case -1:
 636                         if (errno == EINTR)
 637                                 continue;
 638                         /* Fall trhu */
 639                 default:
 640                         c = getc(stdin);
 641                         tcsetattr(0, TCSAFLUSH, &save);
 642
 643                         if (perf_top__handle_keypress(top, c))
 644                                 goto repeat;
 645                         done = 1;
 646                 }
 647         }
 648
 649         tcsetattr(0, TCSAFLUSH, &save);
 650         return NULL;
 651 }
 652
 653 static int symbol_filter(struct map *map, struct symbol *sym)
 654 {
 655         const char *name = sym->name;
 656
 657         if (!map->dso->kernel)
 658                 return 0;
 659         /*
 660          * ppc64 uses function descriptors and appends a '.' to the
 661          * start of every instruction address. Remove it.
 662          */
 663         if (name[0] == '.')
 664                 name++;
 665
 666         if (!strcmp(name, "_text") ||
 667             !strcmp(name, "_etext") ||
 668             !strcmp(name, "_sinittext") ||
 669             !strncmp("init_module", name, 11) ||
 670             !strncmp("cleanup_module", name, 14) ||
 671             strstr(name, "_text_start") ||
 672             strstr(name, "_text_end"))
 673                 return 1;
 674
 675         if (symbol__is_idle(sym))
 676                 sym->ignore = true;
 677
 678         return 0;
 679 }
 680
 681 static int hist_iter__top_callback(struct hist_entry_iter *iter,
 682                                    struct addr_location *al, bool single,
 683                                    void *arg)
 684 {
 685         struct perf_top *top = arg;
 686         struct hist_entry *he = iter->he;
 687         struct perf_evsel *evsel = iter->evsel;
 688
 689         if (sort__has_sym && single) {
 690                 u64 ip = al->addr;
 691
 692                 if (al->map)
 693                         ip = al->map->unmap_ip(al->map, ip);
 694
 695                 perf_top__record_precise_ip(top, he, evsel->idx, ip);
 696         }
 697
 698         return 0;
 699 }
 700
 701 static void perf_event__process_sample(struct perf_tool *tool,
 702                                        const union perf_event *event,
 703                                        struct perf_evsel *evsel,
 704                                        struct perf_sample *sample,
 705                                        struct machine *machine)
 706 {
 707         struct perf_top *top = container_of(tool, struct perf_top, tool);
 708         struct addr_location al;
 709         int err;
 710
 711         if (!machine && perf_guest) {
 712                 static struct intlist *seen;
 713
 714                 if (!seen)
 715                         seen = intlist__new(NULL);
 716
 717                 if (!intlist__has_entry(seen, sample->pid)) {
 718                         pr_err("Can't find guest [%d]'s kernel information\n",
 719                                 sample->pid);
 720                         intlist__add(seen, sample->pid);
 721                 }
 722                 return;
 723         }
 724
 725         if (!machine) {
 726                 pr_err("%u unprocessable samples recorded.\r",
 727                        top->session->evlist->stats.nr_unprocessable_samples++);
 728                 return;
 729         }
 730
 731         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 732                 top->exact_samples++;
 733
 734         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0)
 735                 return;
 736
 737         if (!top->kptr_restrict_warned &&
 738             symbol_conf.kptr_restrict &&
 739             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 740                 ui__warning(
 741 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 742 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 743 "Kernel%s samples will not be resolved.\n",
 744                           al.map && !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 745                           " modules" : "");
 746                 if (use_browser <= 0)
 747                         sleep(5);
 748                 top->kptr_restrict_warned = true;
 749         }
 750
 751         if (al.sym == NULL) {
 752                 const char *msg = "Kernel samples will not be resolved.\n";
 753                 /*
 754                  * As we do lazy loading of symtabs we only will know if the
 755                  * specified vmlinux file is invalid when we actually have a
 756                  * hit in kernel space and then try to load it. So if we get
 757                  * here and there are _no_ symbols in the DSO backing the
 758                  * kernel map, bail out.
 759                  *
 760                  * We may never get here, for instance, if we use -K/
 761                  * --hide-kernel-symbols, even if the user specifies an
 762                  * invalid --vmlinux ;-)
 763                  */
 764                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
 765                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 766                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 767                         if (symbol_conf.vmlinux_name) {
 768                                 char serr[256];
 769                                 dso__strerror_load(al.map->dso, serr, sizeof(serr));
 770                                 ui__warning("The %s file can't be used: %s\n%s",
 771                                             symbol_conf.vmlinux_name, serr, msg);
 772                         } else {
 773                                 ui__warning("A vmlinux file was not found.\n%s",
 774                                             msg);
 775                         }
 776
 777                         if (use_browser <= 0)
 778                                 sleep(5);
 779                         top->vmlinux_warned = true;
 780                 }
 781         }
 782
 783         if (al.sym == NULL || !al.sym->ignore) {
 784                 struct hists *hists = evsel__hists(evsel);
 785                 struct hist_entry_iter iter = {
 786                         .evsel          = evsel,
 787                         .sample         = sample,
 788                         .add_entry_cb   = hist_iter__top_callback,
 789                 };
 790
 791                 if (symbol_conf.cumulate_callchain)
 792                         iter.ops = &hist_iter_cumulative;
 793                 else
 794                         iter.ops = &hist_iter_normal;
 795
 796                 pthread_mutex_lock(&hists->lock);
 797
 798                 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
 799                 if (err < 0)
 800                         pr_err("Problem incrementing symbol period, skipping event\n");
 801
 802                 pthread_mutex_unlock(&hists->lock);
 803         }
 804
 805         addr_location__put(&al);
 806 }
 807
 808 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 809 {
 810         struct perf_sample sample;
 811         struct perf_evsel *evsel;
 812         struct perf_session *session = top->session;
 813         union perf_event *event;
 814         struct machine *machine;
 815         u8 origin;
 816         int ret;
 817
 818         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
 819                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
 820                 if (ret) {
 821                         pr_err("Can't parse sample, err = %d\n", ret);
 822                         goto next_event;
 823                 }
 824
 825                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 826                 assert(evsel != NULL);
 827
 828                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 829
 830                 if (event->header.type == PERF_RECORD_SAMPLE)
 831                         ++top->samples;
 832
 833                 switch (origin) {
 834                 case PERF_RECORD_MISC_USER:
 835                         ++top->us_samples;
 836                         if (top->hide_user_symbols)
 837                                 goto next_event;
 838                         machine = &session->machines.host;
 839                         break;
 840                 case PERF_RECORD_MISC_KERNEL:
 841                         ++top->kernel_samples;
 842                         if (top->hide_kernel_symbols)
 843                                 goto next_event;
 844                         machine = &session->machines.host;
 845                         break;
 846                 case PERF_RECORD_MISC_GUEST_KERNEL:
 847                         ++top->guest_kernel_samples;
 848                         machine = perf_session__find_machine(session,
 849                                                              sample.pid);
 850                         break;
 851                 case PERF_RECORD_MISC_GUEST_USER:
 852                         ++top->guest_us_samples;
 853                         /*
 854                          * TODO: we don't process guest user from host side
 855                          * except simple counting.
 856                          */
 857                         /* Fall thru */
 858                 default:
 859                         goto next_event;
 860                 }
 861
 862
 863                 if (event->header.type == PERF_RECORD_SAMPLE) {
 864                         perf_event__process_sample(&top->tool, event, evsel,
 865                                                    &sample, machine);
 866                 } else if (event->header.type < PERF_RECORD_MAX) {
 867                         hists__inc_nr_events(evsel__hists(evsel), event->header.type);
 868                         machine__process_event(machine, event, &sample);
 869                 } else
 870                         ++session->evlist->stats.nr_unknown_events;
 871 next_event:
 872                 perf_evlist__mmap_consume(top->evlist, idx);
 873         }
 874 }
 875
 876 static void perf_top__mmap_read(struct perf_top *top)
 877 {
 878         int i;
 879
 880         for (i = 0; i < top->evlist->nr_mmaps; i++)
 881                 perf_top__mmap_read_idx(top, i);
 882 }
 883
 884 static int perf_top__start_counters(struct perf_top *top)
 885 {
 886         char msg[512];
 887         struct perf_evsel *counter;
 888         struct perf_evlist *evlist = top->evlist;
 889         struct record_opts *opts = &top->record_opts;
 890
 891         perf_evlist__config(evlist, opts);
 892
 893         evlist__for_each(evlist, counter) {
 894 try_again:
 895                 if (perf_evsel__open(counter, top->evlist->cpus,
 896                                      top->evlist->threads) < 0) {
 897                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 898                                 if (verbose)
 899                                         ui__warning("%s\n", msg);
 900                                 goto try_again;
 901                         }
 902
 903                         perf_evsel__open_strerror(counter, &opts->target,
 904                                                   errno, msg, sizeof(msg));
 905                         ui__error("%s\n", msg);
 906                         goto out_err;
 907                 }
 908         }
 909
 910         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 911                 ui__error("Failed to mmap with %d (%s)\n",
 912                             errno, strerror_r(errno, msg, sizeof(msg)));
 913                 goto out_err;
 914         }
 915
 916         return 0;
 917
 918 out_err:
 919         return -1;
 920 }
 921
 922 static int perf_top__setup_sample_type(struct perf_top *top __maybe_unused)
 923 {
 924         if (!sort__has_sym) {
 925                 if (symbol_conf.use_callchain) {
 926                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
 927                         return -EINVAL;
 928                 }
 929         } else if (callchain_param.mode != CHAIN_NONE) {
 930                 if (callchain_register_param(&callchain_param) < 0) {
 931                         ui__error("Can't register callchain params.\n");
 932                         return -EINVAL;
 933                 }
 934         }
 935
 936         return 0;
 937 }
 938
 939 static int __cmd_top(struct perf_top *top)
 940 {
 941         struct record_opts *opts = &top->record_opts;
 942         pthread_t thread;
 943         int ret;
 944
 945         top->session = perf_session__new(NULL, false, NULL);
 946         if (top->session == NULL)
 947                 return -1;
 948
 949         machines__set_symbol_filter(&top->session->machines, symbol_filter);
 950
 951         if (!objdump_path) {
 952                 ret = perf_session_env__lookup_objdump(&top->session->header.env);
 953                 if (ret)
 954                         goto out_delete;
 955         }
 956
 957         ret = perf_top__setup_sample_type(top);
 958         if (ret)
 959                 goto out_delete;
 960
 961         machine__synthesize_threads(&top->session->machines.host, &opts->target,
 962                                     top->evlist->threads, false, opts->proc_map_timeout);
 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 (!target__none(&opts->target))
 979                 perf_evlist__enable(top->evlist);
 980
 981         /* Wait for a minimal set of events before starting the snapshot */
 982         perf_evlist__poll(top->evlist, 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_join;
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 = perf_evlist__poll(top->evlist, 100);
1010         }
1011
1012         ret = 0;
1013 out_join:
1014         pthread_join(thread, NULL);
1015 out_delete:
1016         perf_session__delete(top->session);
1017         top->session = NULL;
1018
1019         return ret;
1020 }
1021
1022 static int
1023 callchain_opt(const struct option *opt, const char *arg, int unset)
1024 {
1025         symbol_conf.use_callchain = true;
1026         return record_callchain_opt(opt, arg, unset);
1027 }
1028
1029 static int
1030 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1031 {
1032         symbol_conf.use_callchain = true;
1033         return record_parse_callchain_opt(opt, arg, unset);
1034 }
1035
1036 static int perf_top_config(const char *var, const char *value, void *cb)
1037 {
1038         if (!strcmp(var, "top.call-graph"))
1039                 var = "call-graph.record-mode"; /* fall-through */
1040         if (!strcmp(var, "top.children")) {
1041                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1042                 return 0;
1043         }
1044
1045         return perf_default_config(var, value, cb);
1046 }
1047
1048 static int
1049 parse_percent_limit(const struct option *opt, const char *arg,
1050                     int unset __maybe_unused)
1051 {
1052         struct perf_top *top = opt->value;
1053
1054         top->min_percent = strtof(arg, NULL);
1055         return 0;
1056 }
1057
1058 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1059 {
1060         char errbuf[BUFSIZ];
1061         struct perf_top top = {
1062                 .count_filter        = 5,
1063                 .delay_secs          = 2,
1064                 .record_opts = {
1065                         .mmap_pages     = UINT_MAX,
1066                         .user_freq      = UINT_MAX,
1067                         .user_interval  = ULLONG_MAX,
1068                         .freq           = 4000, /* 4 KHz */
1069                         .target         = {
1070                                 .uses_mmap   = true,
1071                         },
1072                         .proc_map_timeout    = 500,
1073                 },
1074                 .max_stack           = PERF_MAX_STACK_DEPTH,
1075                 .sym_pcnt_filter     = 5,
1076         };
1077         struct record_opts *opts = &top.record_opts;
1078         struct target *target = &opts->target;
1079         const struct option options[] = {
1080         OPT_CALLBACK('e', "event", &top.evlist, "event",
1081                      "event selector. use 'perf list' to list available events",
1082                      parse_events_option),
1083         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1084         OPT_STRING('p', "pid", &target->pid, "pid",
1085                     "profile events on existing process id"),
1086         OPT_STRING('t', "tid", &target->tid, "tid",
1087                     "profile events on existing thread id"),
1088         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1089                             "system-wide collection from all CPUs"),
1090         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1091                     "list of cpus to monitor"),
1092         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1093                    "file", "vmlinux pathname"),
1094         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1095                     "don't load vmlinux even if found"),
1096         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1097                     "hide kernel symbols"),
1098         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1099                      "number of mmap data pages",
1100                      perf_evlist__parse_mmap_pages),
1101         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1102                     "collect data with this RT SCHED_FIFO priority"),
1103         OPT_INTEGER('d', "delay", &top.delay_secs,
1104                     "number of seconds to delay between refreshes"),
1105         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1106                             "dump the symbol table used for profiling"),
1107         OPT_INTEGER('f', "count-filter", &top.count_filter,
1108                     "only display functions with more events than this"),
1109         OPT_BOOLEAN(0, "group", &opts->group,
1110                             "put the counters into a counter group"),
1111         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1112                     "child tasks do not inherit counters"),
1113         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1114                     "symbol to annotate"),
1115         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1116         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1117         OPT_INTEGER('E', "entries", &top.print_entries,
1118                     "display this many functions"),
1119         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1120                     "hide user symbols"),
1121         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1122         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1123         OPT_INCR('v', "verbose", &verbose,
1124                     "be more verbose (show counter open errors, etc)"),
1125         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1126                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1127                    " Please refer the man page for the complete list."),
1128         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1129                    "output field(s): overhead, period, sample plus all of sort keys"),
1130         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1131                     "Show a column with the number of samples"),
1132         OPT_CALLBACK_NOOPT('g', NULL, &top.record_opts,
1133                            NULL, "enables call-graph recording",
1134                            &callchain_opt),
1135         OPT_CALLBACK(0, "call-graph", &top.record_opts,
1136                      "mode[,dump_size]", record_callchain_help,
1137                      &parse_callchain_opt),
1138         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1139                     "Accumulate callchains of children and show total overhead as well"),
1140         OPT_INTEGER(0, "max-stack", &top.max_stack,
1141                     "Set the maximum stack depth when parsing the callchain. "
1142                     "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
1143         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1144                    "ignore callees of these functions in call graphs",
1145                    report_parse_ignore_callees_opt),
1146         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1147                     "Show a column with the sum of periods"),
1148         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1149                    "only consider symbols in these dsos"),
1150         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1151                    "only consider symbols in these comms"),
1152         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1153                    "only consider these symbols"),
1154         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1155                     "Interleave source code with assembly code (default)"),
1156         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1157                     "Display raw encoding of assembly instructions (default)"),
1158         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1159                     "Enable kernel symbol demangling"),
1160         OPT_STRING(0, "objdump", &objdump_path, "path",
1161                     "objdump binary to use for disassembly and annotations"),
1162         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1163                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1164         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1165         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1166                      "Don't show entries under that percent", parse_percent_limit),
1167         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1168                      "How to display percentage of filtered entries", parse_filter_percentage),
1169         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1170                    "width[,width...]",
1171                    "don't try to adjust column width, use these fixed values"),
1172         OPT_UINTEGER(0, "proc-map-timeout", &opts->proc_map_timeout,
1173                         "per thread proc mmap processing timeout in ms"),
1174         OPT_END()
1175         };
1176         const char * const top_usage[] = {
1177                 "perf top [<options>]",
1178                 NULL
1179         };
1180         int status = hists__init();
1181
1182         if (status < 0)
1183                 return status;
1184
1185         top.evlist = perf_evlist__new();
1186         if (top.evlist == NULL)
1187                 return -ENOMEM;
1188
1189         perf_config(perf_top_config, &top);
1190
1191         argc = parse_options(argc, argv, options, top_usage, 0);
1192         if (argc)
1193                 usage_with_options(top_usage, options);
1194
1195         sort__mode = SORT_MODE__TOP;
1196         /* display thread wants entries to be collapsed in a different tree */
1197         sort__need_collapse = 1;
1198
1199         if (setup_sorting() < 0) {
1200                 if (sort_order)
1201                         parse_options_usage(top_usage, options, "s", 1);
1202                 if (field_order)
1203                         parse_options_usage(sort_order ? NULL : top_usage,
1204                                             options, "fields", 0);
1205                 goto out_delete_evlist;
1206         }
1207
1208         if (top.use_stdio)
1209                 use_browser = 0;
1210         else if (top.use_tui)
1211                 use_browser = 1;
1212
1213         setup_browser(false);
1214
1215         status = target__validate(target);
1216         if (status) {
1217                 target__strerror(target, status, errbuf, BUFSIZ);
1218                 ui__warning("%s\n", errbuf);
1219         }
1220
1221         status = target__parse_uid(target);
1222         if (status) {
1223                 int saved_errno = errno;
1224
1225                 target__strerror(target, status, errbuf, BUFSIZ);
1226                 ui__error("%s\n", errbuf);
1227
1228                 status = -saved_errno;
1229                 goto out_delete_evlist;
1230         }
1231
1232         if (target__none(target))
1233                 target->system_wide = true;
1234
1235         if (perf_evlist__create_maps(top.evlist, target) < 0)
1236                 usage_with_options(top_usage, options);
1237
1238         if (!top.evlist->nr_entries &&
1239             perf_evlist__add_default(top.evlist) < 0) {
1240                 ui__error("Not enough memory for event selector list\n");
1241                 goto out_delete_evlist;
1242         }
1243
1244         symbol_conf.nr_events = top.evlist->nr_entries;
1245
1246         if (top.delay_secs < 1)
1247                 top.delay_secs = 1;
1248
1249         if (record_opts__config(opts)) {
1250                 status = -EINVAL;
1251                 goto out_delete_evlist;
1252         }
1253
1254         top.sym_evsel = perf_evlist__first(top.evlist);
1255
1256         if (!symbol_conf.use_callchain) {
1257                 symbol_conf.cumulate_callchain = false;
1258                 perf_hpp__cancel_cumulate();
1259         }
1260
1261         symbol_conf.priv_size = sizeof(struct annotation);
1262
1263         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1264         if (symbol__init(NULL) < 0)
1265                 return -1;
1266
1267         sort__setup_elide(stdout);
1268
1269         get_term_dimensions(&top.winsize);
1270         if (top.print_entries == 0) {
1271                 struct sigaction act = {
1272                         .sa_sigaction = perf_top__sig_winch,
1273                         .sa_flags     = SA_SIGINFO,
1274                 };
1275                 perf_top__update_print_entries(&top);
1276                 sigaction(SIGWINCH, &act, NULL);
1277         }
1278
1279         status = __cmd_top(&top);
1280
1281 out_delete_evlist:
1282         perf_evlist__delete(top.evlist);
1283
1284         return status;
1285 }