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
  44 #include "util/debug.h"
  45
  46 #include <assert.h>
  47 #include <elf.h>
  48 #include <fcntl.h>
  49
  50 #include <stdio.h>
  51 #include <termios.h>
  52 #include <unistd.h>
  53 #include <inttypes.h>
  54
  55 #include <errno.h>
  56 #include <time.h>
  57 #include <sched.h>
  58
  59 #include <sys/syscall.h>
  60 #include <sys/ioctl.h>
  61 #include <sys/poll.h>
  62 #include <sys/prctl.h>
  63 #include <sys/wait.h>
  64 #include <sys/uio.h>
  65 #include <sys/utsname.h>
  66 #include <sys/mman.h>
  67
  68 #include <linux/unistd.h>
  69 #include <linux/types.h>
  70
  71 static volatile int done;
  72
  73 static void perf_top__update_print_entries(struct perf_top *top)
  74 {
  75         if (top->print_entries > 9)
  76                 top->print_entries -= 9;
  77 }
  78
  79 static void perf_top__sig_winch(int sig __maybe_unused,
  80                                 siginfo_t *info __maybe_unused, void *arg)
  81 {
  82         struct perf_top *top = arg;
  83
  84         get_term_dimensions(&top->winsize);
  85         if (!top->print_entries
  86             || (top->print_entries+4) > top->winsize.ws_row) {
  87                 top->print_entries = top->winsize.ws_row;
  88         } else {
  89                 top->print_entries += 4;
  90                 top->winsize.ws_row = top->print_entries;
  91         }
  92         perf_top__update_print_entries(top);
  93 }
  94
  95 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
  96 {
  97         struct symbol *sym;
  98         struct annotation *notes;
  99         struct map *map;
 100         int err = -1;
 101
 102         if (!he || !he->ms.sym)
 103                 return -1;
 104
 105         sym = he->ms.sym;
 106         map = he->ms.map;
 107
 108         /*
 109          * We can't annotate with just /proc/kallsyms
 110          */
 111         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS) {
 112                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 113                        "path\n", sym->name);
 114                 sleep(1);
 115                 return -1;
 116         }
 117
 118         notes = symbol__annotation(sym);
 119         if (notes->src != NULL) {
 120                 pthread_mutex_lock(&notes->lock);
 121                 goto out_assign;
 122         }
 123
 124         pthread_mutex_lock(&notes->lock);
 125
 126         if (symbol__alloc_hist(sym) < 0) {
 127                 pthread_mutex_unlock(&notes->lock);
 128                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 129                        sym->name);
 130                 sleep(1);
 131                 return err;
 132         }
 133
 134         err = symbol__annotate(sym, map, 0);
 135         if (err == 0) {
 136 out_assign:
 137                 top->sym_filter_entry = he;
 138         }
 139
 140         pthread_mutex_unlock(&notes->lock);
 141         return err;
 142 }
 143
 144 static void __zero_source_counters(struct hist_entry *he)
 145 {
 146         struct symbol *sym = he->ms.sym;
 147         symbol__annotate_zero_histograms(sym);
 148 }
 149
 150 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 151 {
 152         struct utsname uts;
 153         int err = uname(&uts);
 154
 155         ui__warning("Out of bounds address found:\n\n"
 156                     "Addr:   %" PRIx64 "\n"
 157                     "DSO:    %s %c\n"
 158                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
 159                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 160                     "Arch:   %s\n"
 161                     "Kernel: %s\n"
 162                     "Tools:  %s\n\n"
 163                     "Not all samples will be on the annotation output.\n\n"
 164                     "Please report to linux-kernel@vger.kernel.org\n",
 165                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
 166                     map->start, map->end, sym->start, sym->end,
 167                     sym->binding == STB_GLOBAL ? 'g' :
 168                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 169                     err ? "[unknown]" : uts.machine,
 170                     err ? "[unknown]" : uts.release, perf_version_string);
 171         if (use_browser <= 0)
 172                 sleep(5);
 173
 174         map->erange_warned = true;
 175 }
 176
 177 static void perf_top__record_precise_ip(struct perf_top *top,
 178                                         struct hist_entry *he,
 179                                         int counter, u64 ip)
 180 {
 181         struct annotation *notes;
 182         struct symbol *sym;
 183         int err;
 184
 185         if (he == NULL || he->ms.sym == NULL ||
 186             ((top->sym_filter_entry == NULL ||
 187               top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
 188                 return;
 189
 190         sym = he->ms.sym;
 191         notes = symbol__annotation(sym);
 192
 193         if (pthread_mutex_trylock(&notes->lock))
 194                 return;
 195
 196         if (notes->src == NULL && symbol__alloc_hist(sym) < 0) {
 197                 pthread_mutex_unlock(&notes->lock);
 198                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 199                        sym->name);
 200                 sleep(1);
 201                 return;
 202         }
 203
 204         ip = he->ms.map->map_ip(he->ms.map, ip);
 205         err = symbol__inc_addr_samples(sym, he->ms.map, counter, ip);
 206
 207         pthread_mutex_unlock(&notes->lock);
 208
 209         if (err == -ERANGE && !he->ms.map->erange_warned)
 210                 ui__warn_map_erange(he->ms.map, sym, ip);
 211 }
 212
 213 static void perf_top__show_details(struct perf_top *top)
 214 {
 215         struct hist_entry *he = top->sym_filter_entry;
 216         struct annotation *notes;
 217         struct symbol *symbol;
 218         int more;
 219
 220         if (!he)
 221                 return;
 222
 223         symbol = he->ms.sym;
 224         notes = symbol__annotation(symbol);
 225
 226         pthread_mutex_lock(&notes->lock);
 227
 228         if (notes->src == NULL)
 229                 goto out_unlock;
 230
 231         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
 232         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
 233
 234         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel->idx,
 235                                        0, top->sym_pcnt_filter, top->print_entries, 4);
 236         if (top->zero)
 237                 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 238         else
 239                 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 240         if (more != 0)
 241                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 242 out_unlock:
 243         pthread_mutex_unlock(&notes->lock);
 244 }
 245
 246 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
 247
 248 static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
 249                                                      struct addr_location *al,
 250                                                      struct perf_sample *sample)
 251 {
 252         struct hist_entry *he;
 253
 254         he = __hists__add_entry(&evsel->hists, al, NULL, sample->period);
 255         if (he == NULL)
 256                 return NULL;
 257
 258         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 259         return he;
 260 }
 261
 262 static void perf_top__print_sym_table(struct perf_top *top)
 263 {
 264         char bf[160];
 265         int printed = 0;
 266         const int win_width = top->winsize.ws_col - 1;
 267
 268         puts(CONSOLE_CLEAR);
 269
 270         perf_top__header_snprintf(top, bf, sizeof(bf));
 271         printf("%s\n", bf);
 272
 273         perf_top__reset_sample_counters(top);
 274
 275         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 276
 277         if (top->sym_evsel->hists.stats.nr_lost_warned !=
 278             top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
 279                 top->sym_evsel->hists.stats.nr_lost_warned =
 280                         top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
 281                 color_fprintf(stdout, PERF_COLOR_RED,
 282                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 283                               top->sym_evsel->hists.stats.nr_lost_warned);
 284                 ++printed;
 285         }
 286
 287         if (top->sym_filter_entry) {
 288                 perf_top__show_details(top);
 289                 return;
 290         }
 291
 292         hists__collapse_resort_threaded(&top->sym_evsel->hists);
 293         hists__output_resort_threaded(&top->sym_evsel->hists);
 294         hists__decay_entries_threaded(&top->sym_evsel->hists,
 295                                       top->hide_user_symbols,
 296                                       top->hide_kernel_symbols);
 297         hists__output_recalc_col_len(&top->sym_evsel->hists,
 298                                      top->winsize.ws_row - 3);
 299         putchar('\n');
 300         hists__fprintf(&top->sym_evsel->hists, false,
 301                        top->winsize.ws_row - 4 - printed, win_width, stdout);
 302 }
 303
 304 static void prompt_integer(int *target, const char *msg)
 305 {
 306         char *buf = malloc(0), *p;
 307         size_t dummy = 0;
 308         int tmp;
 309
 310         fprintf(stdout, "\n%s: ", msg);
 311         if (getline(&buf, &dummy, stdin) < 0)
 312                 return;
 313
 314         p = strchr(buf, '\n');
 315         if (p)
 316                 *p = 0;
 317
 318         p = buf;
 319         while(*p) {
 320                 if (!isdigit(*p))
 321                         goto out_free;
 322                 p++;
 323         }
 324         tmp = strtoul(buf, NULL, 10);
 325         *target = tmp;
 326 out_free:
 327         free(buf);
 328 }
 329
 330 static void prompt_percent(int *target, const char *msg)
 331 {
 332         int tmp = 0;
 333
 334         prompt_integer(&tmp, msg);
 335         if (tmp >= 0 && tmp <= 100)
 336                 *target = tmp;
 337 }
 338
 339 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 340 {
 341         char *buf = malloc(0), *p;
 342         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 343         struct rb_node *next;
 344         size_t dummy = 0;
 345
 346         /* zero counters of active symbol */
 347         if (syme) {
 348                 __zero_source_counters(syme);
 349                 top->sym_filter_entry = NULL;
 350         }
 351
 352         fprintf(stdout, "\n%s: ", msg);
 353         if (getline(&buf, &dummy, stdin) < 0)
 354                 goto out_free;
 355
 356         p = strchr(buf, '\n');
 357         if (p)
 358                 *p = 0;
 359
 360         next = rb_first(&top->sym_evsel->hists.entries);
 361         while (next) {
 362                 n = rb_entry(next, struct hist_entry, rb_node);
 363                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 364                         found = n;
 365                         break;
 366                 }
 367                 next = rb_next(&n->rb_node);
 368         }
 369
 370         if (!found) {
 371                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 372                 sleep(1);
 373         } else
 374                 perf_top__parse_source(top, found);
 375
 376 out_free:
 377         free(buf);
 378 }
 379
 380 static void perf_top__print_mapped_keys(struct perf_top *top)
 381 {
 382         char *name = NULL;
 383
 384         if (top->sym_filter_entry) {
 385                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 386                 name = sym->name;
 387         }
 388
 389         fprintf(stdout, "\nMapped keys:\n");
 390         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 391         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 392
 393         if (top->evlist->nr_entries > 1)
 394                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 395
 396         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 397
 398         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
 399         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 400         fprintf(stdout, "\t[S]     stop annotation.\n");
 401
 402         fprintf(stdout,
 403                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 404                 top->hide_kernel_symbols ? "yes" : "no");
 405         fprintf(stdout,
 406                 "\t[U]     hide user symbols.               \t(%s)\n",
 407                 top->hide_user_symbols ? "yes" : "no");
 408         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 409         fprintf(stdout, "\t[qQ]    quit.\n");
 410 }
 411
 412 static int perf_top__key_mapped(struct perf_top *top, int c)
 413 {
 414         switch (c) {
 415                 case 'd':
 416                 case 'e':
 417                 case 'f':
 418                 case 'z':
 419                 case 'q':
 420                 case 'Q':
 421                 case 'K':
 422                 case 'U':
 423                 case 'F':
 424                 case 's':
 425                 case 'S':
 426                         return 1;
 427                 case 'E':
 428                         return top->evlist->nr_entries > 1 ? 1 : 0;
 429                 default:
 430                         break;
 431         }
 432
 433         return 0;
 434 }
 435
 436 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 437 {
 438         bool ret = true;
 439
 440         if (!perf_top__key_mapped(top, c)) {
 441                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 442                 struct termios tc, save;
 443
 444                 perf_top__print_mapped_keys(top);
 445                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 446                 fflush(stdout);
 447
 448                 tcgetattr(0, &save);
 449                 tc = save;
 450                 tc.c_lflag &= ~(ICANON | ECHO);
 451                 tc.c_cc[VMIN] = 0;
 452                 tc.c_cc[VTIME] = 0;
 453                 tcsetattr(0, TCSANOW, &tc);
 454
 455                 poll(&stdin_poll, 1, -1);
 456                 c = getc(stdin);
 457
 458                 tcsetattr(0, TCSAFLUSH, &save);
 459                 if (!perf_top__key_mapped(top, c))
 460                         return ret;
 461         }
 462
 463         switch (c) {
 464                 case 'd':
 465                         prompt_integer(&top->delay_secs, "Enter display delay");
 466                         if (top->delay_secs < 1)
 467                                 top->delay_secs = 1;
 468                         break;
 469                 case 'e':
 470                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 471                         if (top->print_entries == 0) {
 472                                 struct sigaction act = {
 473                                         .sa_sigaction = perf_top__sig_winch,
 474                                         .sa_flags     = SA_SIGINFO,
 475                                 };
 476                                 perf_top__sig_winch(SIGWINCH, NULL, top);
 477                                 sigaction(SIGWINCH, &act, NULL);
 478                         } else {
 479                                 perf_top__sig_winch(SIGWINCH, NULL, top);
 480                                 signal(SIGWINCH, SIG_DFL);
 481                         }
 482                         break;
 483                 case 'E':
 484                         if (top->evlist->nr_entries > 1) {
 485                                 /* Select 0 as the default event: */
 486                                 int counter = 0;
 487
 488                                 fprintf(stderr, "\nAvailable events:");
 489
 490                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
 491                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
 492
 493                                 prompt_integer(&counter, "Enter details event counter");
 494
 495                                 if (counter >= top->evlist->nr_entries) {
 496                                         top->sym_evsel = perf_evlist__first(top->evlist);
 497                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
 498                                         sleep(1);
 499                                         break;
 500                                 }
 501                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
 502                                         if (top->sym_evsel->idx == counter)
 503                                                 break;
 504                         } else
 505                                 top->sym_evsel = perf_evlist__first(top->evlist);
 506                         break;
 507                 case 'f':
 508                         prompt_integer(&top->count_filter, "Enter display event count filter");
 509                         break;
 510                 case 'F':
 511                         prompt_percent(&top->sym_pcnt_filter,
 512                                        "Enter details display event filter (percent)");
 513                         break;
 514                 case 'K':
 515                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 516                         break;
 517                 case 'q':
 518                 case 'Q':
 519                         printf("exiting.\n");
 520                         if (top->dump_symtab)
 521                                 perf_session__fprintf_dsos(top->session, stderr);
 522                         ret = false;
 523                         break;
 524                 case 's':
 525                         perf_top__prompt_symbol(top, "Enter details symbol");
 526                         break;
 527                 case 'S':
 528                         if (!top->sym_filter_entry)
 529                                 break;
 530                         else {
 531                                 struct hist_entry *syme = top->sym_filter_entry;
 532
 533                                 top->sym_filter_entry = NULL;
 534                                 __zero_source_counters(syme);
 535                         }
 536                         break;
 537                 case 'U':
 538                         top->hide_user_symbols = !top->hide_user_symbols;
 539                         break;
 540                 case 'z':
 541                         top->zero = !top->zero;
 542                         break;
 543                 default:
 544                         break;
 545         }
 546
 547         return ret;
 548 }
 549
 550 static void perf_top__sort_new_samples(void *arg)
 551 {
 552         struct perf_top *t = arg;
 553         perf_top__reset_sample_counters(t);
 554
 555         if (t->evlist->selected != NULL)
 556                 t->sym_evsel = t->evlist->selected;
 557
 558         hists__collapse_resort_threaded(&t->sym_evsel->hists);
 559         hists__output_resort_threaded(&t->sym_evsel->hists);
 560         hists__decay_entries_threaded(&t->sym_evsel->hists,
 561                                       t->hide_user_symbols,
 562                                       t->hide_kernel_symbols);
 563 }
 564
 565 static void *display_thread_tui(void *arg)
 566 {
 567         struct perf_evsel *pos;
 568         struct perf_top *top = arg;
 569         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 570         struct hist_browser_timer hbt = {
 571                 .timer          = perf_top__sort_new_samples,
 572                 .arg            = top,
 573                 .refresh        = top->delay_secs,
 574         };
 575
 576         perf_top__sort_new_samples(top);
 577
 578         /*
 579          * Initialize the uid_filter_str, in the future the TUI will allow
 580          * Zooming in/out UIDs. For now juse use whatever the user passed
 581          * via --uid.
 582          */
 583         list_for_each_entry(pos, &top->evlist->entries, node)
 584                 pos->hists.uid_filter_str = top->record_opts.target.uid_str;
 585
 586         perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
 587                                       &top->session->header.env);
 588
 589         done = 1;
 590         return NULL;
 591 }
 592
 593 static void *display_thread(void *arg)
 594 {
 595         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 596         struct termios tc, save;
 597         struct perf_top *top = arg;
 598         int delay_msecs, c;
 599
 600         tcgetattr(0, &save);
 601         tc = save;
 602         tc.c_lflag &= ~(ICANON | ECHO);
 603         tc.c_cc[VMIN] = 0;
 604         tc.c_cc[VTIME] = 0;
 605
 606         pthread__unblock_sigwinch();
 607 repeat:
 608         delay_msecs = top->delay_secs * 1000;
 609         tcsetattr(0, TCSANOW, &tc);
 610         /* trash return*/
 611         getc(stdin);
 612
 613         while (!done) {
 614                 perf_top__print_sym_table(top);
 615                 /*
 616                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 617                  * refresh screen in both cases.
 618                  */
 619                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 620                 case 0:
 621                         continue;
 622                 case -1:
 623                         if (errno == EINTR)
 624                                 continue;
 625                         /* Fall trhu */
 626                 default:
 627                         c = getc(stdin);
 628                         tcsetattr(0, TCSAFLUSH, &save);
 629
 630                         if (perf_top__handle_keypress(top, c))
 631                                 goto repeat;
 632                         done = 1;
 633                 }
 634         }
 635
 636         return NULL;
 637 }
 638
 639 /* Tag samples to be skipped. */
 640 static const char *skip_symbols[] = {
 641         "intel_idle",
 642         "default_idle",
 643         "native_safe_halt",
 644         "cpu_idle",
 645         "enter_idle",
 646         "exit_idle",
 647         "mwait_idle",
 648         "mwait_idle_with_hints",
 649         "poll_idle",
 650         "ppc64_runlatch_off",
 651         "pseries_dedicated_idle_sleep",
 652         NULL
 653 };
 654
 655 static int symbol_filter(struct map *map __maybe_unused, struct symbol *sym)
 656 {
 657         const char *name = sym->name;
 658         int i;
 659
 660         /*
 661          * ppc64 uses function descriptors and appends a '.' to the
 662          * start of every instruction address. Remove it.
 663          */
 664         if (name[0] == '.')
 665                 name++;
 666
 667         if (!strcmp(name, "_text") ||
 668             !strcmp(name, "_etext") ||
 669             !strcmp(name, "_sinittext") ||
 670             !strncmp("init_module", name, 11) ||
 671             !strncmp("cleanup_module", name, 14) ||
 672             strstr(name, "_text_start") ||
 673             strstr(name, "_text_end"))
 674                 return 1;
 675
 676         for (i = 0; skip_symbols[i]; i++) {
 677                 if (!strcmp(skip_symbols[i], name)) {
 678                         sym->ignore = true;
 679                         break;
 680                 }
 681         }
 682
 683         return 0;
 684 }
 685
 686 static void perf_event__process_sample(struct perf_tool *tool,
 687                                        const union perf_event *event,
 688                                        struct perf_evsel *evsel,
 689                                        struct perf_sample *sample,
 690                                        struct machine *machine)
 691 {
 692         struct perf_top *top = container_of(tool, struct perf_top, tool);
 693         struct symbol *parent = NULL;
 694         u64 ip = event->ip.ip;
 695         struct addr_location al;
 696         int err;
 697
 698         if (!machine && perf_guest) {
 699                 static struct intlist *seen;
 700
 701                 if (!seen)
 702                         seen = intlist__new(NULL);
 703
 704                 if (!intlist__has_entry(seen, event->ip.pid)) {
 705                         pr_err("Can't find guest [%d]'s kernel information\n",
 706                                 event->ip.pid);
 707                         intlist__add(seen, event->ip.pid);
 708                 }
 709                 return;
 710         }
 711
 712         if (!machine) {
 713                 pr_err("%u unprocessable samples recorded.\r",
 714                        top->session->stats.nr_unprocessable_samples++);
 715                 return;
 716         }
 717
 718         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 719                 top->exact_samples++;
 720
 721         if (perf_event__preprocess_sample(event, machine, &al, sample,
 722                                           symbol_filter) < 0 ||
 723             al.filtered)
 724                 return;
 725
 726         if (!top->kptr_restrict_warned &&
 727             symbol_conf.kptr_restrict &&
 728             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 729                 ui__warning(
 730 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 731 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 732 "Kernel%s samples will not be resolved.\n",
 733                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 734                           " modules" : "");
 735                 if (use_browser <= 0)
 736                         sleep(5);
 737                 top->kptr_restrict_warned = true;
 738         }
 739
 740         if (al.sym == NULL) {
 741                 const char *msg = "Kernel samples will not be resolved.\n";
 742                 /*
 743                  * As we do lazy loading of symtabs we only will know if the
 744                  * specified vmlinux file is invalid when we actually have a
 745                  * hit in kernel space and then try to load it. So if we get
 746                  * here and there are _no_ symbols in the DSO backing the
 747                  * kernel map, bail out.
 748                  *
 749                  * We may never get here, for instance, if we use -K/
 750                  * --hide-kernel-symbols, even if the user specifies an
 751                  * invalid --vmlinux ;-)
 752                  */
 753                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
 754                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 755                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 756                         if (symbol_conf.vmlinux_name) {
 757                                 ui__warning("The %s file can't be used.\n%s",
 758                                             symbol_conf.vmlinux_name, msg);
 759                         } else {
 760                                 ui__warning("A vmlinux file was not found.\n%s",
 761                                             msg);
 762                         }
 763
 764                         if (use_browser <= 0)
 765                                 sleep(5);
 766                         top->vmlinux_warned = true;
 767                 }
 768         }
 769
 770         if (al.sym == NULL || !al.sym->ignore) {
 771                 struct hist_entry *he;
 772
 773                 if ((sort__has_parent || symbol_conf.use_callchain) &&
 774                     sample->callchain) {
 775                         err = machine__resolve_callchain(machine, evsel,
 776                                                          al.thread, sample,
 777                                                          &parent);
 778
 779                         if (err)
 780                                 return;
 781                 }
 782
 783                 he = perf_evsel__add_hist_entry(evsel, &al, sample);
 784                 if (he == NULL) {
 785                         pr_err("Problem incrementing symbol period, skipping event\n");
 786                         return;
 787                 }
 788
 789                 if (symbol_conf.use_callchain) {
 790                         err = callchain_append(he->callchain, &callchain_cursor,
 791                                                sample->period);
 792                         if (err)
 793                                 return;
 794                 }
 795
 796                 if (top->sort_has_symbols)
 797                         perf_top__record_precise_ip(top, he, evsel->idx, ip);
 798         }
 799
 800         return;
 801 }
 802
 803 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 804 {
 805         struct perf_sample sample;
 806         struct perf_evsel *evsel;
 807         struct perf_session *session = top->session;
 808         union perf_event *event;
 809         struct machine *machine;
 810         u8 origin;
 811         int ret;
 812
 813         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
 814                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
 815                 if (ret) {
 816                         pr_err("Can't parse sample, err = %d\n", ret);
 817                         continue;
 818                 }
 819
 820                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 821                 assert(evsel != NULL);
 822
 823                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 824
 825                 if (event->header.type == PERF_RECORD_SAMPLE)
 826                         ++top->samples;
 827
 828                 switch (origin) {
 829                 case PERF_RECORD_MISC_USER:
 830                         ++top->us_samples;
 831                         if (top->hide_user_symbols)
 832                                 continue;
 833                         machine = &session->machines.host;
 834                         break;
 835                 case PERF_RECORD_MISC_KERNEL:
 836                         ++top->kernel_samples;
 837                         if (top->hide_kernel_symbols)
 838                                 continue;
 839                         machine = &session->machines.host;
 840                         break;
 841                 case PERF_RECORD_MISC_GUEST_KERNEL:
 842                         ++top->guest_kernel_samples;
 843                         machine = perf_session__find_machine(session, event->ip.pid);
 844                         break;
 845                 case PERF_RECORD_MISC_GUEST_USER:
 846                         ++top->guest_us_samples;
 847                         /*
 848                          * TODO: we don't process guest user from host side
 849                          * except simple counting.
 850                          */
 851                         /* Fall thru */
 852                 default:
 853                         continue;
 854                 }
 855
 856
 857                 if (event->header.type == PERF_RECORD_SAMPLE) {
 858                         perf_event__process_sample(&top->tool, event, evsel,
 859                                                    &sample, machine);
 860                 } else if (event->header.type < PERF_RECORD_MAX) {
 861                         hists__inc_nr_events(&evsel->hists, event->header.type);
 862                         machine__process_event(machine, event);
 863                 } else
 864                         ++session->stats.nr_unknown_events;
 865         }
 866 }
 867
 868 static void perf_top__mmap_read(struct perf_top *top)
 869 {
 870         int i;
 871
 872         for (i = 0; i < top->evlist->nr_mmaps; i++)
 873                 perf_top__mmap_read_idx(top, i);
 874 }
 875
 876 static int perf_top__start_counters(struct perf_top *top)
 877 {
 878         char msg[512];
 879         struct perf_evsel *counter;
 880         struct perf_evlist *evlist = top->evlist;
 881         struct perf_record_opts *opts = &top->record_opts;
 882
 883         perf_evlist__config(evlist, opts);
 884
 885         list_for_each_entry(counter, &evlist->entries, node) {
 886 try_again:
 887                 if (perf_evsel__open(counter, top->evlist->cpus,
 888                                      top->evlist->threads) < 0) {
 889                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 890                                 if (verbose)
 891                                         ui__warning("%s\n", msg);
 892                                 goto try_again;
 893                         }
 894
 895                         perf_evsel__open_strerror(counter, &opts->target,
 896                                                   errno, msg, sizeof(msg));
 897                         ui__error("%s\n", msg);
 898                         goto out_err;
 899                 }
 900         }
 901
 902         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 903                 ui__error("Failed to mmap with %d (%s)\n",
 904                             errno, strerror(errno));
 905                 goto out_err;
 906         }
 907
 908         return 0;
 909
 910 out_err:
 911         return -1;
 912 }
 913
 914 static int perf_top__setup_sample_type(struct perf_top *top)
 915 {
 916         if (!top->sort_has_symbols) {
 917                 if (symbol_conf.use_callchain) {
 918                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
 919                         return -EINVAL;
 920                 }
 921         } else if (callchain_param.mode != CHAIN_NONE) {
 922                 if (callchain_register_param(&callchain_param) < 0) {
 923                         ui__error("Can't register callchain params.\n");
 924                         return -EINVAL;
 925                 }
 926         }
 927
 928         return 0;
 929 }
 930
 931 static int __cmd_top(struct perf_top *top)
 932 {
 933         struct perf_record_opts *opts = &top->record_opts;
 934         pthread_t thread;
 935         int ret;
 936         /*
 937          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
 938          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
 939          */
 940         top->session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
 941         if (top->session == NULL)
 942                 return -ENOMEM;
 943
 944         ret = perf_top__setup_sample_type(top);
 945         if (ret)
 946                 goto out_delete;
 947
 948         if (perf_target__has_task(&opts->target))
 949                 perf_event__synthesize_thread_map(&top->tool, top->evlist->threads,
 950                                                   perf_event__process,
 951                                                   &top->session->machines.host);
 952         else
 953                 perf_event__synthesize_threads(&top->tool, perf_event__process,
 954                                                &top->session->machines.host);
 955
 956         ret = perf_top__start_counters(top);
 957         if (ret)
 958                 goto out_delete;
 959
 960         top->session->evlist = top->evlist;
 961         perf_session__set_id_hdr_size(top->session);
 962
 963         /*
 964          * When perf is starting the traced process, all the events (apart from
 965          * group members) have enable_on_exec=1 set, so don't spoil it by
 966          * prematurely enabling them.
 967          *
 968          * XXX 'top' still doesn't start workloads like record, trace, but should,
 969          * so leave the check here.
 970          */
 971         if (!perf_target__none(&opts->target))
 972                 perf_evlist__enable(top->evlist);
 973
 974         /* Wait for a minimal set of events before starting the snapshot */
 975         poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
 976
 977         perf_top__mmap_read(top);
 978
 979         ret = -1;
 980         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 981                                                             display_thread), top)) {
 982                 ui__error("Could not create display thread.\n");
 983                 goto out_delete;
 984         }
 985
 986         if (top->realtime_prio) {
 987                 struct sched_param param;
 988
 989                 param.sched_priority = top->realtime_prio;
 990                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 991                         ui__error("Could not set realtime priority.\n");
 992                         goto out_delete;
 993                 }
 994         }
 995
 996         while (!done) {
 997                 u64 hits = top->samples;
 998
 999                 perf_top__mmap_read(top);
1000
1001                 if (hits == top->samples)
1002                         ret = poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
1003         }
1004
1005         ret = 0;
1006 out_delete:
1007         perf_session__delete(top->session);
1008         top->session = NULL;
1009
1010         return ret;
1011 }
1012
1013 static int
1014 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1015 {
1016         /*
1017          * --no-call-graph
1018          */
1019         if (unset)
1020                 return 0;
1021
1022         symbol_conf.use_callchain = true;
1023
1024         return record_parse_callchain_opt(opt, arg, unset);
1025 }
1026
1027 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1028 {
1029         int status;
1030         char errbuf[BUFSIZ];
1031         struct perf_top top = {
1032                 .count_filter        = 5,
1033                 .delay_secs          = 2,
1034                 .record_opts = {
1035                         .mmap_pages     = UINT_MAX,
1036                         .user_freq      = UINT_MAX,
1037                         .user_interval  = ULLONG_MAX,
1038                         .freq           = 4000, /* 4 KHz */
1039                         .target              = {
1040                                 .uses_mmap   = true,
1041                         },
1042                 },
1043                 .sym_pcnt_filter     = 5,
1044         };
1045         struct perf_record_opts *opts = &top.record_opts;
1046         struct perf_target *target = &opts->target;
1047         const struct option options[] = {
1048         OPT_CALLBACK('e', "event", &top.evlist, "event",
1049                      "event selector. use 'perf list' to list available events",
1050                      parse_events_option),
1051         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1052         OPT_STRING('p', "pid", &target->pid, "pid",
1053                     "profile events on existing process id"),
1054         OPT_STRING('t', "tid", &target->tid, "tid",
1055                     "profile events on existing thread id"),
1056         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1057                             "system-wide collection from all CPUs"),
1058         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1059                     "list of cpus to monitor"),
1060         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1061                    "file", "vmlinux pathname"),
1062         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1063                     "hide kernel symbols"),
1064         OPT_UINTEGER('m', "mmap-pages", &opts->mmap_pages,
1065                      "number of mmap data pages"),
1066         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1067                     "collect data with this RT SCHED_FIFO priority"),
1068         OPT_INTEGER('d', "delay", &top.delay_secs,
1069                     "number of seconds to delay between refreshes"),
1070         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1071                             "dump the symbol table used for profiling"),
1072         OPT_INTEGER('f', "count-filter", &top.count_filter,
1073                     "only display functions with more events than this"),
1074         OPT_BOOLEAN('g', "group", &opts->group,
1075                             "put the counters into a counter group"),
1076         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1077                     "child tasks do not inherit counters"),
1078         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1079                     "symbol to annotate"),
1080         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1081         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1082         OPT_INTEGER('E', "entries", &top.print_entries,
1083                     "display this many functions"),
1084         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1085                     "hide user symbols"),
1086         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1087         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1088         OPT_INCR('v', "verbose", &verbose,
1089                     "be more verbose (show counter open errors, etc)"),
1090         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1091                    "sort by key(s): pid, comm, dso, symbol, parent"),
1092         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1093                     "Show a column with the number of samples"),
1094         OPT_CALLBACK_DEFAULT('G', "call-graph", &top.record_opts,
1095                              "mode[,dump_size]", record_callchain_help,
1096                              &parse_callchain_opt, "fp"),
1097         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1098                     "Show a column with the sum of periods"),
1099         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1100                    "only consider symbols in these dsos"),
1101         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1102                    "only consider symbols in these comms"),
1103         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1104                    "only consider these symbols"),
1105         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1106                     "Interleave source code with assembly code (default)"),
1107         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1108                     "Display raw encoding of assembly instructions (default)"),
1109         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1110                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1111         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1112         OPT_END()
1113         };
1114         const char * const top_usage[] = {
1115                 "perf top [<options>]",
1116                 NULL
1117         };
1118
1119         top.evlist = perf_evlist__new(NULL, NULL);
1120         if (top.evlist == NULL)
1121                 return -ENOMEM;
1122
1123         symbol_conf.exclude_other = false;
1124
1125         argc = parse_options(argc, argv, options, top_usage, 0);
1126         if (argc)
1127                 usage_with_options(top_usage, options);
1128
1129         if (sort_order == default_sort_order)
1130                 sort_order = "dso,symbol";
1131
1132         if (setup_sorting() < 0)
1133                 usage_with_options(top_usage, options);
1134
1135         if (top.use_stdio)
1136                 use_browser = 0;
1137         else if (top.use_tui)
1138                 use_browser = 1;
1139
1140         setup_browser(false);
1141
1142         status = perf_target__validate(target);
1143         if (status) {
1144                 perf_target__strerror(target, status, errbuf, BUFSIZ);
1145                 ui__warning("%s", errbuf);
1146         }
1147
1148         status = perf_target__parse_uid(target);
1149         if (status) {
1150                 int saved_errno = errno;
1151
1152                 perf_target__strerror(target, status, errbuf, BUFSIZ);
1153                 ui__error("%s", errbuf);
1154
1155                 status = -saved_errno;
1156                 goto out_delete_evlist;
1157         }
1158
1159         if (perf_target__none(target))
1160                 target->system_wide = true;
1161
1162         if (perf_evlist__create_maps(top.evlist, target) < 0)
1163                 usage_with_options(top_usage, options);
1164
1165         if (!top.evlist->nr_entries &&
1166             perf_evlist__add_default(top.evlist) < 0) {
1167                 ui__error("Not enough memory for event selector list\n");
1168                 goto out_delete_maps;
1169         }
1170
1171         symbol_conf.nr_events = top.evlist->nr_entries;
1172
1173         if (top.delay_secs < 1)
1174                 top.delay_secs = 1;
1175
1176         if (opts->user_interval != ULLONG_MAX)
1177                 opts->default_interval = opts->user_interval;
1178         if (opts->user_freq != UINT_MAX)
1179                 opts->freq = opts->user_freq;
1180
1181         /*
1182          * User specified count overrides default frequency.
1183          */
1184         if (opts->default_interval)
1185                 opts->freq = 0;
1186         else if (opts->freq) {
1187                 opts->default_interval = opts->freq;
1188         } else {
1189                 ui__error("frequency and count are zero, aborting\n");
1190                 status = -EINVAL;
1191                 goto out_delete_maps;
1192         }
1193
1194         top.sym_evsel = perf_evlist__first(top.evlist);
1195
1196         symbol_conf.priv_size = sizeof(struct annotation);
1197
1198         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1199         if (symbol__init() < 0)
1200                 return -1;
1201
1202         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1203         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1204         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1205
1206         /*
1207          * Avoid annotation data structures overhead when symbols aren't on the
1208          * sort list.
1209          */
1210         top.sort_has_symbols = sort_sym.list.next != NULL;
1211
1212         get_term_dimensions(&top.winsize);
1213         if (top.print_entries == 0) {
1214                 struct sigaction act = {
1215                         .sa_sigaction = perf_top__sig_winch,
1216                         .sa_flags     = SA_SIGINFO,
1217                 };
1218                 perf_top__update_print_entries(&top);
1219                 sigaction(SIGWINCH, &act, NULL);
1220         }
1221
1222         status = __cmd_top(&top);
1223
1224 out_delete_maps:
1225         perf_evlist__delete_maps(top.evlist);
1226 out_delete_evlist:
1227         perf_evlist__delete(top.evlist);
1228
1229         return status;
1230 }