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         while (true)  {
 590                 int key = perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
 591                                                         top->min_percent,
 592                                                         &top->session->header.env);
 593
 594                 if (key != 'f')
 595                         break;
 596
 597                 perf_evlist__toggle_enable(top->evlist);
 598                 /*
 599                  * No need to refresh, resort/decay histogram entries
 600                  * if we are not collecting samples:
 601                  */
 602                 if (top->evlist->enabled) {
 603                         hbt.refresh = top->delay_secs;
 604                         help = "Press 'f' to disable the events or 'h' to see other hotkeys";
 605                 } else {
 606                         help = "Press 'f' again to re-enable the events";
 607                         hbt.refresh = 0;
 608                 }
 609         }
 610
 611         done = 1;
 612         return NULL;
 613 }
 614
 615 static void display_sig(int sig __maybe_unused)
 616 {
 617         done = 1;
 618 }
 619
 620 static void display_setup_sig(void)
 621 {
 622         signal(SIGSEGV, display_sig);
 623         signal(SIGFPE,  display_sig);
 624         signal(SIGINT,  display_sig);
 625         signal(SIGQUIT, display_sig);
 626         signal(SIGTERM, display_sig);
 627 }
 628
 629 static void *display_thread(void *arg)
 630 {
 631         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 632         struct termios save;
 633         struct perf_top *top = arg;
 634         int delay_msecs, c;
 635
 636         display_setup_sig();
 637         pthread__unblock_sigwinch();
 638 repeat:
 639         delay_msecs = top->delay_secs * 1000;
 640         set_term_quiet_input(&save);
 641         /* trash return*/
 642         getc(stdin);
 643
 644         while (!done) {
 645                 perf_top__print_sym_table(top);
 646                 /*
 647                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 648                  * refresh screen in both cases.
 649                  */
 650                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 651                 case 0:
 652                         continue;
 653                 case -1:
 654                         if (errno == EINTR)
 655                                 continue;
 656                         /* Fall trhu */
 657                 default:
 658                         c = getc(stdin);
 659                         tcsetattr(0, TCSAFLUSH, &save);
 660
 661                         if (perf_top__handle_keypress(top, c))
 662                                 goto repeat;
 663                         done = 1;
 664                 }
 665         }
 666
 667         tcsetattr(0, TCSAFLUSH, &save);
 668         return NULL;
 669 }
 670
 671 static int symbol_filter(struct map *map, struct symbol *sym)
 672 {
 673         const char *name = sym->name;
 674
 675         if (!map->dso->kernel)
 676                 return 0;
 677         /*
 678          * ppc64 uses function descriptors and appends a '.' to the
 679          * start of every instruction address. Remove it.
 680          */
 681         if (name[0] == '.')
 682                 name++;
 683
 684         if (!strcmp(name, "_text") ||
 685             !strcmp(name, "_etext") ||
 686             !strcmp(name, "_sinittext") ||
 687             !strncmp("init_module", name, 11) ||
 688             !strncmp("cleanup_module", name, 14) ||
 689             strstr(name, "_text_start") ||
 690             strstr(name, "_text_end"))
 691                 return 1;
 692
 693         if (symbol__is_idle(sym))
 694                 sym->ignore = true;
 695
 696         return 0;
 697 }
 698
 699 static int hist_iter__top_callback(struct hist_entry_iter *iter,
 700                                    struct addr_location *al, bool single,
 701                                    void *arg)
 702 {
 703         struct perf_top *top = arg;
 704         struct hist_entry *he = iter->he;
 705         struct perf_evsel *evsel = iter->evsel;
 706
 707         if (sort__has_sym && single) {
 708                 u64 ip = al->addr;
 709
 710                 if (al->map)
 711                         ip = al->map->unmap_ip(al->map, ip);
 712
 713                 perf_top__record_precise_ip(top, he, evsel->idx, ip);
 714         }
 715
 716         return 0;
 717 }
 718
 719 static void perf_event__process_sample(struct perf_tool *tool,
 720                                        const union perf_event *event,
 721                                        struct perf_evsel *evsel,
 722                                        struct perf_sample *sample,
 723                                        struct machine *machine)
 724 {
 725         struct perf_top *top = container_of(tool, struct perf_top, tool);
 726         struct addr_location al;
 727         int err;
 728
 729         if (!machine && perf_guest) {
 730                 static struct intlist *seen;
 731
 732                 if (!seen)
 733                         seen = intlist__new(NULL);
 734
 735                 if (!intlist__has_entry(seen, sample->pid)) {
 736                         pr_err("Can't find guest [%d]'s kernel information\n",
 737                                 sample->pid);
 738                         intlist__add(seen, sample->pid);
 739                 }
 740                 return;
 741         }
 742
 743         if (!machine) {
 744                 pr_err("%u unprocessable samples recorded.\r",
 745                        top->session->evlist->stats.nr_unprocessable_samples++);
 746                 return;
 747         }
 748
 749         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 750                 top->exact_samples++;
 751
 752         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0)
 753                 return;
 754
 755         if (!top->kptr_restrict_warned &&
 756             symbol_conf.kptr_restrict &&
 757             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 758                 ui__warning(
 759 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 760 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 761 "Kernel%s samples will not be resolved.\n",
 762                           al.map && !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 763                           " modules" : "");
 764                 if (use_browser <= 0)
 765                         sleep(5);
 766                 top->kptr_restrict_warned = true;
 767         }
 768
 769         if (al.sym == NULL) {
 770                 const char *msg = "Kernel samples will not be resolved.\n";
 771                 /*
 772                  * As we do lazy loading of symtabs we only will know if the
 773                  * specified vmlinux file is invalid when we actually have a
 774                  * hit in kernel space and then try to load it. So if we get
 775                  * here and there are _no_ symbols in the DSO backing the
 776                  * kernel map, bail out.
 777                  *
 778                  * We may never get here, for instance, if we use -K/
 779                  * --hide-kernel-symbols, even if the user specifies an
 780                  * invalid --vmlinux ;-)
 781                  */
 782                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
 783                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 784                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 785                         if (symbol_conf.vmlinux_name) {
 786                                 char serr[256];
 787                                 dso__strerror_load(al.map->dso, serr, sizeof(serr));
 788                                 ui__warning("The %s file can't be used: %s\n%s",
 789                                             symbol_conf.vmlinux_name, serr, msg);
 790                         } else {
 791                                 ui__warning("A vmlinux file was not found.\n%s",
 792                                             msg);
 793                         }
 794
 795                         if (use_browser <= 0)
 796                                 sleep(5);
 797                         top->vmlinux_warned = true;
 798                 }
 799         }
 800
 801         if (al.sym == NULL || !al.sym->ignore) {
 802                 struct hists *hists = evsel__hists(evsel);
 803                 struct hist_entry_iter iter = {
 804                         .evsel          = evsel,
 805                         .sample         = sample,
 806                         .add_entry_cb   = hist_iter__top_callback,
 807                 };
 808
 809                 if (symbol_conf.cumulate_callchain)
 810                         iter.ops = &hist_iter_cumulative;
 811                 else
 812                         iter.ops = &hist_iter_normal;
 813
 814                 pthread_mutex_lock(&hists->lock);
 815
 816                 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
 817                 if (err < 0)
 818                         pr_err("Problem incrementing symbol period, skipping event\n");
 819
 820                 pthread_mutex_unlock(&hists->lock);
 821         }
 822
 823         addr_location__put(&al);
 824 }
 825
 826 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 827 {
 828         struct perf_sample sample;
 829         struct perf_evsel *evsel;
 830         struct perf_session *session = top->session;
 831         union perf_event *event;
 832         struct machine *machine;
 833         u8 origin;
 834         int ret;
 835
 836         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
 837                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
 838                 if (ret) {
 839                         pr_err("Can't parse sample, err = %d\n", ret);
 840                         goto next_event;
 841                 }
 842
 843                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 844                 assert(evsel != NULL);
 845
 846                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 847
 848                 if (event->header.type == PERF_RECORD_SAMPLE)
 849                         ++top->samples;
 850
 851                 switch (origin) {
 852                 case PERF_RECORD_MISC_USER:
 853                         ++top->us_samples;
 854                         if (top->hide_user_symbols)
 855                                 goto next_event;
 856                         machine = &session->machines.host;
 857                         break;
 858                 case PERF_RECORD_MISC_KERNEL:
 859                         ++top->kernel_samples;
 860                         if (top->hide_kernel_symbols)
 861                                 goto next_event;
 862                         machine = &session->machines.host;
 863                         break;
 864                 case PERF_RECORD_MISC_GUEST_KERNEL:
 865                         ++top->guest_kernel_samples;
 866                         machine = perf_session__find_machine(session,
 867                                                              sample.pid);
 868                         break;
 869                 case PERF_RECORD_MISC_GUEST_USER:
 870                         ++top->guest_us_samples;
 871                         /*
 872                          * TODO: we don't process guest user from host side
 873                          * except simple counting.
 874                          */
 875                         /* Fall thru */
 876                 default:
 877                         goto next_event;
 878                 }
 879
 880
 881                 if (event->header.type == PERF_RECORD_SAMPLE) {
 882                         perf_event__process_sample(&top->tool, event, evsel,
 883                                                    &sample, machine);
 884                 } else if (event->header.type < PERF_RECORD_MAX) {
 885                         hists__inc_nr_events(evsel__hists(evsel), event->header.type);
 886                         machine__process_event(machine, event, &sample);
 887                 } else
 888                         ++session->evlist->stats.nr_unknown_events;
 889 next_event:
 890                 perf_evlist__mmap_consume(top->evlist, idx);
 891         }
 892 }
 893
 894 static void perf_top__mmap_read(struct perf_top *top)
 895 {
 896         int i;
 897
 898         for (i = 0; i < top->evlist->nr_mmaps; i++)
 899                 perf_top__mmap_read_idx(top, i);
 900 }
 901
 902 static int perf_top__start_counters(struct perf_top *top)
 903 {
 904         char msg[512];
 905         struct perf_evsel *counter;
 906         struct perf_evlist *evlist = top->evlist;
 907         struct record_opts *opts = &top->record_opts;
 908
 909         perf_evlist__config(evlist, opts);
 910
 911         evlist__for_each(evlist, counter) {
 912 try_again:
 913                 if (perf_evsel__open(counter, top->evlist->cpus,
 914                                      top->evlist->threads) < 0) {
 915                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 916                                 if (verbose)
 917                                         ui__warning("%s\n", msg);
 918                                 goto try_again;
 919                         }
 920
 921                         perf_evsel__open_strerror(counter, &opts->target,
 922                                                   errno, msg, sizeof(msg));
 923                         ui__error("%s\n", msg);
 924                         goto out_err;
 925                 }
 926         }
 927
 928         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 929                 ui__error("Failed to mmap with %d (%s)\n",
 930                             errno, strerror_r(errno, msg, sizeof(msg)));
 931                 goto out_err;
 932         }
 933
 934         return 0;
 935
 936 out_err:
 937         return -1;
 938 }
 939
 940 static int perf_top__setup_sample_type(struct perf_top *top __maybe_unused)
 941 {
 942         if (!sort__has_sym) {
 943                 if (symbol_conf.use_callchain) {
 944                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
 945                         return -EINVAL;
 946                 }
 947         } else if (callchain_param.mode != CHAIN_NONE) {
 948                 if (callchain_register_param(&callchain_param) < 0) {
 949                         ui__error("Can't register callchain params.\n");
 950                         return -EINVAL;
 951                 }
 952         }
 953
 954         return 0;
 955 }
 956
 957 static int __cmd_top(struct perf_top *top)
 958 {
 959         struct record_opts *opts = &top->record_opts;
 960         pthread_t thread;
 961         int ret;
 962
 963         top->session = perf_session__new(NULL, false, NULL);
 964         if (top->session == NULL)
 965                 return -1;
 966
 967         machines__set_symbol_filter(&top->session->machines, symbol_filter);
 968
 969         if (!objdump_path) {
 970                 ret = perf_session_env__lookup_objdump(&top->session->header.env);
 971                 if (ret)
 972                         goto out_delete;
 973         }
 974
 975         ret = perf_top__setup_sample_type(top);
 976         if (ret)
 977                 goto out_delete;
 978
 979         machine__synthesize_threads(&top->session->machines.host, &opts->target,
 980                                     top->evlist->threads, false, opts->proc_map_timeout);
 981         ret = perf_top__start_counters(top);
 982         if (ret)
 983                 goto out_delete;
 984
 985         top->session->evlist = top->evlist;
 986         perf_session__set_id_hdr_size(top->session);
 987
 988         /*
 989          * When perf is starting the traced process, all the events (apart from
 990          * group members) have enable_on_exec=1 set, so don't spoil it by
 991          * prematurely enabling them.
 992          *
 993          * XXX 'top' still doesn't start workloads like record, trace, but should,
 994          * so leave the check here.
 995          */
 996         if (!target__none(&opts->target))
 997                 perf_evlist__enable(top->evlist);
 998
 999         /* Wait for a minimal set of events before starting the snapshot */
1000         perf_evlist__poll(top->evlist, 100);
1001
1002         perf_top__mmap_read(top);
1003
1004         ret = -1;
1005         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1006                                                             display_thread), top)) {
1007                 ui__error("Could not create display thread.\n");
1008                 goto out_delete;
1009         }
1010
1011         if (top->realtime_prio) {
1012                 struct sched_param param;
1013
1014                 param.sched_priority = top->realtime_prio;
1015                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1016                         ui__error("Could not set realtime priority.\n");
1017                         goto out_join;
1018                 }
1019         }
1020
1021         while (!done) {
1022                 u64 hits = top->samples;
1023
1024                 perf_top__mmap_read(top);
1025
1026                 if (hits == top->samples)
1027                         ret = perf_evlist__poll(top->evlist, 100);
1028         }
1029
1030         ret = 0;
1031 out_join:
1032         pthread_join(thread, NULL);
1033 out_delete:
1034         perf_session__delete(top->session);
1035         top->session = NULL;
1036
1037         return ret;
1038 }
1039
1040 static int
1041 callchain_opt(const struct option *opt, const char *arg, int unset)
1042 {
1043         symbol_conf.use_callchain = true;
1044         return record_callchain_opt(opt, arg, unset);
1045 }
1046
1047 static int
1048 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1049 {
1050         symbol_conf.use_callchain = true;
1051         return record_parse_callchain_opt(opt, arg, unset);
1052 }
1053
1054 static int perf_top_config(const char *var, const char *value, void *cb)
1055 {
1056         if (!strcmp(var, "top.call-graph"))
1057                 var = "call-graph.record-mode"; /* fall-through */
1058         if (!strcmp(var, "top.children")) {
1059                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1060                 return 0;
1061         }
1062
1063         return perf_default_config(var, value, cb);
1064 }
1065
1066 static int
1067 parse_percent_limit(const struct option *opt, const char *arg,
1068                     int unset __maybe_unused)
1069 {
1070         struct perf_top *top = opt->value;
1071
1072         top->min_percent = strtof(arg, NULL);
1073         return 0;
1074 }
1075
1076 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1077 {
1078         char errbuf[BUFSIZ];
1079         struct perf_top top = {
1080                 .count_filter        = 5,
1081                 .delay_secs          = 2,
1082                 .record_opts = {
1083                         .mmap_pages     = UINT_MAX,
1084                         .user_freq      = UINT_MAX,
1085                         .user_interval  = ULLONG_MAX,
1086                         .freq           = 4000, /* 4 KHz */
1087                         .target         = {
1088                                 .uses_mmap   = true,
1089                         },
1090                         .proc_map_timeout    = 500,
1091                 },
1092                 .max_stack           = PERF_MAX_STACK_DEPTH,
1093                 .sym_pcnt_filter     = 5,
1094         };
1095         struct record_opts *opts = &top.record_opts;
1096         struct target *target = &opts->target;
1097         const struct option options[] = {
1098         OPT_CALLBACK('e', "event", &top.evlist, "event",
1099                      "event selector. use 'perf list' to list available events",
1100                      parse_events_option),
1101         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1102         OPT_STRING('p', "pid", &target->pid, "pid",
1103                     "profile events on existing process id"),
1104         OPT_STRING('t', "tid", &target->tid, "tid",
1105                     "profile events on existing thread id"),
1106         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1107                             "system-wide collection from all CPUs"),
1108         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1109                     "list of cpus to monitor"),
1110         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1111                    "file", "vmlinux pathname"),
1112         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1113                     "don't load vmlinux even if found"),
1114         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1115                     "hide kernel symbols"),
1116         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1117                      "number of mmap data pages",
1118                      perf_evlist__parse_mmap_pages),
1119         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1120                     "collect data with this RT SCHED_FIFO priority"),
1121         OPT_INTEGER('d', "delay", &top.delay_secs,
1122                     "number of seconds to delay between refreshes"),
1123         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1124                             "dump the symbol table used for profiling"),
1125         OPT_INTEGER('f', "count-filter", &top.count_filter,
1126                     "only display functions with more events than this"),
1127         OPT_BOOLEAN(0, "group", &opts->group,
1128                             "put the counters into a counter group"),
1129         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1130                     "child tasks do not inherit counters"),
1131         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1132                     "symbol to annotate"),
1133         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1134         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1135         OPT_INTEGER('E', "entries", &top.print_entries,
1136                     "display this many functions"),
1137         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1138                     "hide user symbols"),
1139         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1140         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1141         OPT_INCR('v', "verbose", &verbose,
1142                     "be more verbose (show counter open errors, etc)"),
1143         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1144                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1145                    " Please refer the man page for the complete list."),
1146         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1147                    "output field(s): overhead, period, sample plus all of sort keys"),
1148         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1149                     "Show a column with the number of samples"),
1150         OPT_CALLBACK_NOOPT('g', NULL, &top.record_opts,
1151                            NULL, "enables call-graph recording",
1152                            &callchain_opt),
1153         OPT_CALLBACK(0, "call-graph", &top.record_opts,
1154                      "mode[,dump_size]", record_callchain_help,
1155                      &parse_callchain_opt),
1156         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1157                     "Accumulate callchains of children and show total overhead as well"),
1158         OPT_INTEGER(0, "max-stack", &top.max_stack,
1159                     "Set the maximum stack depth when parsing the callchain. "
1160                     "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
1161         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1162                    "ignore callees of these functions in call graphs",
1163                    report_parse_ignore_callees_opt),
1164         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1165                     "Show a column with the sum of periods"),
1166         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1167                    "only consider symbols in these dsos"),
1168         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1169                    "only consider symbols in these comms"),
1170         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1171                    "only consider these symbols"),
1172         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1173                     "Interleave source code with assembly code (default)"),
1174         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1175                     "Display raw encoding of assembly instructions (default)"),
1176         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1177                     "Enable kernel symbol demangling"),
1178         OPT_STRING(0, "objdump", &objdump_path, "path",
1179                     "objdump binary to use for disassembly and annotations"),
1180         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1181                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1182         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1183         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1184                      "Don't show entries under that percent", parse_percent_limit),
1185         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1186                      "How to display percentage of filtered entries", parse_filter_percentage),
1187         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1188                    "width[,width...]",
1189                    "don't try to adjust column width, use these fixed values"),
1190         OPT_UINTEGER(0, "proc-map-timeout", &opts->proc_map_timeout,
1191                         "per thread proc mmap processing timeout in ms"),
1192         OPT_END()
1193         };
1194         const char * const top_usage[] = {
1195                 "perf top [<options>]",
1196                 NULL
1197         };
1198         int status = hists__init();
1199
1200         if (status < 0)
1201                 return status;
1202
1203         top.evlist = perf_evlist__new();
1204         if (top.evlist == NULL)
1205                 return -ENOMEM;
1206
1207         perf_config(perf_top_config, &top);
1208
1209         argc = parse_options(argc, argv, options, top_usage, 0);
1210         if (argc)
1211                 usage_with_options(top_usage, options);
1212
1213         sort__mode = SORT_MODE__TOP;
1214         /* display thread wants entries to be collapsed in a different tree */
1215         sort__need_collapse = 1;
1216
1217         if (setup_sorting() < 0) {
1218                 if (sort_order)
1219                         parse_options_usage(top_usage, options, "s", 1);
1220                 if (field_order)
1221                         parse_options_usage(sort_order ? NULL : top_usage,
1222                                             options, "fields", 0);
1223                 goto out_delete_evlist;
1224         }
1225
1226         if (top.use_stdio)
1227                 use_browser = 0;
1228         else if (top.use_tui)
1229                 use_browser = 1;
1230
1231         setup_browser(false);
1232
1233         status = target__validate(target);
1234         if (status) {
1235                 target__strerror(target, status, errbuf, BUFSIZ);
1236                 ui__warning("%s\n", errbuf);
1237         }
1238
1239         status = target__parse_uid(target);
1240         if (status) {
1241                 int saved_errno = errno;
1242
1243                 target__strerror(target, status, errbuf, BUFSIZ);
1244                 ui__error("%s\n", errbuf);
1245
1246                 status = -saved_errno;
1247                 goto out_delete_evlist;
1248         }
1249
1250         if (target__none(target))
1251                 target->system_wide = true;
1252
1253         if (perf_evlist__create_maps(top.evlist, target) < 0)
1254                 usage_with_options(top_usage, options);
1255
1256         if (!top.evlist->nr_entries &&
1257             perf_evlist__add_default(top.evlist) < 0) {
1258                 ui__error("Not enough memory for event selector list\n");
1259                 goto out_delete_evlist;
1260         }
1261
1262         symbol_conf.nr_events = top.evlist->nr_entries;
1263
1264         if (top.delay_secs < 1)
1265                 top.delay_secs = 1;
1266
1267         if (record_opts__config(opts)) {
1268                 status = -EINVAL;
1269                 goto out_delete_evlist;
1270         }
1271
1272         top.sym_evsel = perf_evlist__first(top.evlist);
1273
1274         if (!symbol_conf.use_callchain) {
1275                 symbol_conf.cumulate_callchain = false;
1276                 perf_hpp__cancel_cumulate();
1277         }
1278
1279         symbol_conf.priv_size = sizeof(struct annotation);
1280
1281         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1282         if (symbol__init(NULL) < 0)
1283                 return -1;
1284
1285         sort__setup_elide(stdout);
1286
1287         get_term_dimensions(&top.winsize);
1288         if (top.print_entries == 0) {
1289                 struct sigaction act = {
1290                         .sa_sigaction = perf_top__sig_winch,
1291                         .sa_flags     = SA_SIGINFO,
1292                 };
1293                 perf_top__update_print_entries(&top);
1294                 sigaction(SIGWINCH, &act, NULL);
1295         }
1296
1297         status = __cmd_top(&top);
1298
1299 out_delete_evlist:
1300         perf_evlist__delete(top.evlist);
1301
1302         return status;
1303 }