tools/perf/builtin-top.c

   1 /*
   2  * builtin-top.c
   3  *
   4  * Builtin top command: Display a continuously updated profile of
   5  * any workload, CPU or specific PID.
   6  *
   7  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
   8  *               2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   9  *
  10  * Improvements and fixes by:
  11  *
  12  *   Arjan van de Ven <arjan@linux.intel.com>
  13  *   Yanmin Zhang <yanmin.zhang@intel.com>
  14  *   Wu Fengguang <fengguang.wu@intel.com>
  15  *   Mike Galbraith <efault@gmx.de>
  16  *   Paul Mackerras <paulus@samba.org>
  17  *
  18  * Released under the GPL v2. (and only v2, not any later version)
  19  */
  20 #include "builtin.h"
  21
  22 #include "perf.h"
  23
  24 #include "util/annotate.h"
  25 #include "util/cache.h"
  26 #include "util/color.h"
  27 #include "util/evlist.h"
  28 #include "util/evsel.h"
  29 #include "util/machine.h"
  30 #include "util/session.h"
  31 #include "util/symbol.h"
  32 #include "util/thread.h"
  33 #include "util/thread_map.h"
  34 #include "util/top.h"
  35 #include "util/util.h"
  36 #include <linux/rbtree.h>
  37 #include "util/parse-options.h"
  38 #include "util/parse-events.h"
  39 #include "util/cpumap.h"
  40 #include "util/xyarray.h"
  41 #include "util/sort.h"
  42 #include "util/intlist.h"
  43 #include "arch/common.h"
  44
  45 #include "util/debug.h"
  46
  47 #include <assert.h>
  48 #include <elf.h>
  49 #include <fcntl.h>
  50
  51 #include <stdio.h>
  52 #include <termios.h>
  53 #include <unistd.h>
  54 #include <inttypes.h>
  55
  56 #include <errno.h>
  57 #include <time.h>
  58 #include <sched.h>
  59
  60 #include <sys/syscall.h>
  61 #include <sys/ioctl.h>
  62 #include <sys/poll.h>
  63 #include <sys/prctl.h>
  64 #include <sys/wait.h>
  65 #include <sys/uio.h>
  66 #include <sys/utsname.h>
  67 #include <sys/mman.h>
  68
  69 #include <linux/unistd.h>
  70 #include <linux/types.h>
  71
  72 static volatile int done;
  73
  74 #define HEADER_LINE_NR  5
  75
  76 static void perf_top__update_print_entries(struct perf_top *top)
  77 {
  78         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
  79 }
  80
  81 static void perf_top__sig_winch(int sig __maybe_unused,
  82                                 siginfo_t *info __maybe_unused, void *arg)
  83 {
  84         struct perf_top *top = arg;
  85
  86         get_term_dimensions(&top->winsize);
  87         perf_top__update_print_entries(top);
  88 }
  89
  90 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
  91 {
  92         struct symbol *sym;
  93         struct annotation *notes;
  94         struct map *map;
  95         int err = -1;
  96
  97         if (!he || !he->ms.sym)
  98                 return -1;
  99
 100         sym = he->ms.sym;
 101         map = he->ms.map;
 102
 103         /*
 104          * We can't annotate with just /proc/kallsyms
 105          */
 106         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 107             !dso__is_kcore(map->dso)) {
 108                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 109                        "path\n", sym->name);
 110                 sleep(1);
 111                 return -1;
 112         }
 113
 114         notes = symbol__annotation(sym);
 115         if (notes->src != NULL) {
 116                 pthread_mutex_lock(&notes->lock);
 117                 goto out_assign;
 118         }
 119
 120         pthread_mutex_lock(&notes->lock);
 121
 122         if (symbol__alloc_hist(sym) < 0) {
 123                 pthread_mutex_unlock(&notes->lock);
 124                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 125                        sym->name);
 126                 sleep(1);
 127                 return err;
 128         }
 129
 130         err = symbol__annotate(sym, map, 0);
 131         if (err == 0) {
 132 out_assign:
 133                 top->sym_filter_entry = he;
 134         }
 135
 136         pthread_mutex_unlock(&notes->lock);
 137         return err;
 138 }
 139
 140 static void __zero_source_counters(struct hist_entry *he)
 141 {
 142         struct symbol *sym = he->ms.sym;
 143         symbol__annotate_zero_histograms(sym);
 144 }
 145
 146 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 147 {
 148         struct utsname uts;
 149         int err = uname(&uts);
 150
 151         ui__warning("Out of bounds address found:\n\n"
 152                     "Addr:   %" PRIx64 "\n"
 153                     "DSO:    %s %c\n"
 154                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
 155                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 156                     "Arch:   %s\n"
 157                     "Kernel: %s\n"
 158                     "Tools:  %s\n\n"
 159                     "Not all samples will be on the annotation output.\n\n"
 160                     "Please report to linux-kernel@vger.kernel.org\n",
 161                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
 162                     map->start, map->end, sym->start, sym->end,
 163                     sym->binding == STB_GLOBAL ? 'g' :
 164                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 165                     err ? "[unknown]" : uts.machine,
 166                     err ? "[unknown]" : uts.release, perf_version_string);
 167         if (use_browser <= 0)
 168                 sleep(5);
 169
 170         map->erange_warned = true;
 171 }
 172
 173 static void perf_top__record_precise_ip(struct perf_top *top,
 174                                         struct hist_entry *he,
 175                                         int counter, u64 ip)
 176 {
 177         struct annotation *notes;
 178         struct symbol *sym;
 179         int err = 0;
 180
 181         if (he == NULL || he->ms.sym == NULL ||
 182             ((top->sym_filter_entry == NULL ||
 183               top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
 184                 return;
 185
 186         sym = he->ms.sym;
 187         notes = symbol__annotation(sym);
 188
 189         if (pthread_mutex_trylock(&notes->lock))
 190                 return;
 191
 192         ip = he->ms.map->map_ip(he->ms.map, ip);
 193
 194         if (ui__has_annotation())
 195                 err = hist_entry__inc_addr_samples(he, counter, ip);
 196
 197         pthread_mutex_unlock(&notes->lock);
 198
 199         if (err == -ERANGE && !he->ms.map->erange_warned)
 200                 ui__warn_map_erange(he->ms.map, sym, ip);
 201         else if (err == -ENOMEM) {
 202                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 203                        sym->name);
 204                 sleep(1);
 205         }
 206 }
 207
 208 static void perf_top__show_details(struct perf_top *top)
 209 {
 210         struct hist_entry *he = top->sym_filter_entry;
 211         struct annotation *notes;
 212         struct symbol *symbol;
 213         int more;
 214
 215         if (!he)
 216                 return;
 217
 218         symbol = he->ms.sym;
 219         notes = symbol__annotation(symbol);
 220
 221         pthread_mutex_lock(&notes->lock);
 222
 223         if (notes->src == NULL)
 224                 goto out_unlock;
 225
 226         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
 227         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
 228
 229         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel,
 230                                        0, top->sym_pcnt_filter, top->print_entries, 4);
 231         if (top->zero)
 232                 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 233         else
 234                 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 235         if (more != 0)
 236                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 237 out_unlock:
 238         pthread_mutex_unlock(&notes->lock);
 239 }
 240
 241 static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
 242                                                      struct addr_location *al,
 243                                                      struct perf_sample *sample)
 244 {
 245         struct hist_entry *he;
 246
 247         pthread_mutex_lock(&evsel->hists.lock);
 248         he = __hists__add_entry(&evsel->hists, al, NULL, NULL, NULL,
 249                                 sample->period, sample->weight,
 250                                 sample->transaction);
 251         pthread_mutex_unlock(&evsel->hists.lock);
 252         if (he == NULL)
 253                 return NULL;
 254
 255         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 256         if (!he->filtered)
 257                 evsel->hists.stats.nr_non_filtered_samples++;
 258
 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(&top->sym_evsel->hists, NULL);
 293         hists__output_resort(&top->sym_evsel->hists);
 294         hists__decay_entries(&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->print_entries - printed);
 299         putchar('\n');
 300         hists__fprintf(&top->sym_evsel->hists, false,
 301                        top->print_entries - printed, win_width,
 302                        top->min_percent, stdout);
 303 }
 304
 305 static void prompt_integer(int *target, const char *msg)
 306 {
 307         char *buf = malloc(0), *p;
 308         size_t dummy = 0;
 309         int tmp;
 310
 311         fprintf(stdout, "\n%s: ", msg);
 312         if (getline(&buf, &dummy, stdin) < 0)
 313                 return;
 314
 315         p = strchr(buf, '\n');
 316         if (p)
 317                 *p = 0;
 318
 319         p = buf;
 320         while(*p) {
 321                 if (!isdigit(*p))
 322                         goto out_free;
 323                 p++;
 324         }
 325         tmp = strtoul(buf, NULL, 10);
 326         *target = tmp;
 327 out_free:
 328         free(buf);
 329 }
 330
 331 static void prompt_percent(int *target, const char *msg)
 332 {
 333         int tmp = 0;
 334
 335         prompt_integer(&tmp, msg);
 336         if (tmp >= 0 && tmp <= 100)
 337                 *target = tmp;
 338 }
 339
 340 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 341 {
 342         char *buf = malloc(0), *p;
 343         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 344         struct rb_node *next;
 345         size_t dummy = 0;
 346
 347         /* zero counters of active symbol */
 348         if (syme) {
 349                 __zero_source_counters(syme);
 350                 top->sym_filter_entry = NULL;
 351         }
 352
 353         fprintf(stdout, "\n%s: ", msg);
 354         if (getline(&buf, &dummy, stdin) < 0)
 355                 goto out_free;
 356
 357         p = strchr(buf, '\n');
 358         if (p)
 359                 *p = 0;
 360
 361         next = rb_first(&top->sym_evsel->hists.entries);
 362         while (next) {
 363                 n = rb_entry(next, struct hist_entry, rb_node);
 364                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 365                         found = n;
 366                         break;
 367                 }
 368                 next = rb_next(&n->rb_node);
 369         }
 370
 371         if (!found) {
 372                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 373                 sleep(1);
 374         } else
 375                 perf_top__parse_source(top, found);
 376
 377 out_free:
 378         free(buf);
 379 }
 380
 381 static void perf_top__print_mapped_keys(struct perf_top *top)
 382 {
 383         char *name = NULL;
 384
 385         if (top->sym_filter_entry) {
 386                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 387                 name = sym->name;
 388         }
 389
 390         fprintf(stdout, "\nMapped keys:\n");
 391         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 392         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 393
 394         if (top->evlist->nr_entries > 1)
 395                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 396
 397         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 398
 399         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
 400         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 401         fprintf(stdout, "\t[S]     stop annotation.\n");
 402
 403         fprintf(stdout,
 404                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 405                 top->hide_kernel_symbols ? "yes" : "no");
 406         fprintf(stdout,
 407                 "\t[U]     hide user symbols.               \t(%s)\n",
 408                 top->hide_user_symbols ? "yes" : "no");
 409         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 410         fprintf(stdout, "\t[qQ]    quit.\n");
 411 }
 412
 413 static int perf_top__key_mapped(struct perf_top *top, int c)
 414 {
 415         switch (c) {
 416                 case 'd':
 417                 case 'e':
 418                 case 'f':
 419                 case 'z':
 420                 case 'q':
 421                 case 'Q':
 422                 case 'K':
 423                 case 'U':
 424                 case 'F':
 425                 case 's':
 426                 case 'S':
 427                         return 1;
 428                 case 'E':
 429                         return top->evlist->nr_entries > 1 ? 1 : 0;
 430                 default:
 431                         break;
 432         }
 433
 434         return 0;
 435 }
 436
 437 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 438 {
 439         bool ret = true;
 440
 441         if (!perf_top__key_mapped(top, c)) {
 442                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 443                 struct termios tc, save;
 444
 445                 perf_top__print_mapped_keys(top);
 446                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 447                 fflush(stdout);
 448
 449                 tcgetattr(0, &save);
 450                 tc = save;
 451                 tc.c_lflag &= ~(ICANON | ECHO);
 452                 tc.c_cc[VMIN] = 0;
 453                 tc.c_cc[VTIME] = 0;
 454                 tcsetattr(0, TCSANOW, &tc);
 455
 456                 poll(&stdin_poll, 1, -1);
 457                 c = getc(stdin);
 458
 459                 tcsetattr(0, TCSAFLUSH, &save);
 460                 if (!perf_top__key_mapped(top, c))
 461                         return ret;
 462         }
 463
 464         switch (c) {
 465                 case 'd':
 466                         prompt_integer(&top->delay_secs, "Enter display delay");
 467                         if (top->delay_secs < 1)
 468                                 top->delay_secs = 1;
 469                         break;
 470                 case 'e':
 471                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 472                         if (top->print_entries == 0) {
 473                                 struct sigaction act = {
 474                                         .sa_sigaction = perf_top__sig_winch,
 475                                         .sa_flags     = SA_SIGINFO,
 476                                 };
 477                                 perf_top__sig_winch(SIGWINCH, NULL, top);
 478                                 sigaction(SIGWINCH, &act, NULL);
 479                         } else {
 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                                 evlist__for_each(top->evlist, top->sym_evsel)
 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                                 evlist__for_each(top->evlist, top->sym_evsel)
 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(&t->sym_evsel->hists, NULL);
 559         hists__output_resort(&t->sym_evsel->hists);
 560         hists__decay_entries(&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         evlist__for_each(top->evlist, pos)
 584                 pos->hists.uid_filter_str = top->record_opts.target.uid_str;
 585
 586         perf_evlist__tui_browse_hists(top->evlist, help, &hbt, top->min_percent,
 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 static int symbol_filter(struct map *map __maybe_unused, struct symbol *sym)
 640 {
 641         const char *name = sym->name;
 642
 643         /*
 644          * ppc64 uses function descriptors and appends a '.' to the
 645          * start of every instruction address. Remove it.
 646          */
 647         if (name[0] == '.')
 648                 name++;
 649
 650         if (!strcmp(name, "_text") ||
 651             !strcmp(name, "_etext") ||
 652             !strcmp(name, "_sinittext") ||
 653             !strncmp("init_module", name, 11) ||
 654             !strncmp("cleanup_module", name, 14) ||
 655             strstr(name, "_text_start") ||
 656             strstr(name, "_text_end"))
 657                 return 1;
 658
 659         if (symbol__is_idle(sym))
 660                 sym->ignore = true;
 661
 662         return 0;
 663 }
 664
 665 static void perf_event__process_sample(struct perf_tool *tool,
 666                                        const union perf_event *event,
 667                                        struct perf_evsel *evsel,
 668                                        struct perf_sample *sample,
 669                                        struct machine *machine)
 670 {
 671         struct perf_top *top = container_of(tool, struct perf_top, tool);
 672         struct symbol *parent = NULL;
 673         u64 ip = sample->ip;
 674         struct addr_location al;
 675         int err;
 676
 677         if (!machine && perf_guest) {
 678                 static struct intlist *seen;
 679
 680                 if (!seen)
 681                         seen = intlist__new(NULL);
 682
 683                 if (!intlist__has_entry(seen, sample->pid)) {
 684                         pr_err("Can't find guest [%d]'s kernel information\n",
 685                                 sample->pid);
 686                         intlist__add(seen, sample->pid);
 687                 }
 688                 return;
 689         }
 690
 691         if (!machine) {
 692                 pr_err("%u unprocessable samples recorded.\r",
 693                        top->session->stats.nr_unprocessable_samples++);
 694                 return;
 695         }
 696
 697         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 698                 top->exact_samples++;
 699
 700         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0)
 701                 return;
 702
 703         if (!top->kptr_restrict_warned &&
 704             symbol_conf.kptr_restrict &&
 705             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 706                 ui__warning(
 707 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 708 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 709 "Kernel%s samples will not be resolved.\n",
 710                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 711                           " modules" : "");
 712                 if (use_browser <= 0)
 713                         sleep(5);
 714                 top->kptr_restrict_warned = true;
 715         }
 716
 717         if (al.sym == NULL) {
 718                 const char *msg = "Kernel samples will not be resolved.\n";
 719                 /*
 720                  * As we do lazy loading of symtabs we only will know if the
 721                  * specified vmlinux file is invalid when we actually have a
 722                  * hit in kernel space and then try to load it. So if we get
 723                  * here and there are _no_ symbols in the DSO backing the
 724                  * kernel map, bail out.
 725                  *
 726                  * We may never get here, for instance, if we use -K/
 727                  * --hide-kernel-symbols, even if the user specifies an
 728                  * invalid --vmlinux ;-)
 729                  */
 730                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
 731                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 732                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 733                         if (symbol_conf.vmlinux_name) {
 734                                 ui__warning("The %s file can't be used.\n%s",
 735                                             symbol_conf.vmlinux_name, msg);
 736                         } else {
 737                                 ui__warning("A vmlinux file was not found.\n%s",
 738                                             msg);
 739                         }
 740
 741                         if (use_browser <= 0)
 742                                 sleep(5);
 743                         top->vmlinux_warned = true;
 744                 }
 745         }
 746
 747         if (al.sym == NULL || !al.sym->ignore) {
 748                 struct hist_entry *he;
 749
 750                 err = sample__resolve_callchain(sample, &parent, evsel, &al,
 751                                                 top->max_stack);
 752                 if (err)
 753                         return;
 754
 755                 he = perf_evsel__add_hist_entry(evsel, &al, sample);
 756                 if (he == NULL) {
 757                         pr_err("Problem incrementing symbol period, skipping event\n");
 758                         return;
 759                 }
 760
 761                 err = hist_entry__append_callchain(he, sample);
 762                 if (err)
 763                         return;
 764
 765                 if (sort__has_sym)
 766                         perf_top__record_precise_ip(top, he, evsel->idx, ip);
 767         }
 768
 769         return;
 770 }
 771
 772 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 773 {
 774         struct perf_sample sample;
 775         struct perf_evsel *evsel;
 776         struct perf_session *session = top->session;
 777         union perf_event *event;
 778         struct machine *machine;
 779         u8 origin;
 780         int ret;
 781
 782         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
 783                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
 784                 if (ret) {
 785                         pr_err("Can't parse sample, err = %d\n", ret);
 786                         goto next_event;
 787                 }
 788
 789                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 790                 assert(evsel != NULL);
 791
 792                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 793
 794                 if (event->header.type == PERF_RECORD_SAMPLE)
 795                         ++top->samples;
 796
 797                 switch (origin) {
 798                 case PERF_RECORD_MISC_USER:
 799                         ++top->us_samples;
 800                         if (top->hide_user_symbols)
 801                                 goto next_event;
 802                         machine = &session->machines.host;
 803                         break;
 804                 case PERF_RECORD_MISC_KERNEL:
 805                         ++top->kernel_samples;
 806                         if (top->hide_kernel_symbols)
 807                                 goto next_event;
 808                         machine = &session->machines.host;
 809                         break;
 810                 case PERF_RECORD_MISC_GUEST_KERNEL:
 811                         ++top->guest_kernel_samples;
 812                         machine = perf_session__find_machine(session,
 813                                                              sample.pid);
 814                         break;
 815                 case PERF_RECORD_MISC_GUEST_USER:
 816                         ++top->guest_us_samples;
 817                         /*
 818                          * TODO: we don't process guest user from host side
 819                          * except simple counting.
 820                          */
 821                         /* Fall thru */
 822                 default:
 823                         goto next_event;
 824                 }
 825
 826
 827                 if (event->header.type == PERF_RECORD_SAMPLE) {
 828                         perf_event__process_sample(&top->tool, event, evsel,
 829                                                    &sample, machine);
 830                 } else if (event->header.type < PERF_RECORD_MAX) {
 831                         hists__inc_nr_events(&evsel->hists, event->header.type);
 832                         machine__process_event(machine, event, &sample);
 833                 } else
 834                         ++session->stats.nr_unknown_events;
 835 next_event:
 836                 perf_evlist__mmap_consume(top->evlist, idx);
 837         }
 838 }
 839
 840 static void perf_top__mmap_read(struct perf_top *top)
 841 {
 842         int i;
 843
 844         for (i = 0; i < top->evlist->nr_mmaps; i++)
 845                 perf_top__mmap_read_idx(top, i);
 846 }
 847
 848 static int perf_top__start_counters(struct perf_top *top)
 849 {
 850         char msg[512];
 851         struct perf_evsel *counter;
 852         struct perf_evlist *evlist = top->evlist;
 853         struct record_opts *opts = &top->record_opts;
 854
 855         perf_evlist__config(evlist, opts);
 856
 857         evlist__for_each(evlist, counter) {
 858 try_again:
 859                 if (perf_evsel__open(counter, top->evlist->cpus,
 860                                      top->evlist->threads) < 0) {
 861                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 862                                 if (verbose)
 863                                         ui__warning("%s\n", msg);
 864                                 goto try_again;
 865                         }
 866
 867                         perf_evsel__open_strerror(counter, &opts->target,
 868                                                   errno, msg, sizeof(msg));
 869                         ui__error("%s\n", msg);
 870                         goto out_err;
 871                 }
 872         }
 873
 874         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 875                 ui__error("Failed to mmap with %d (%s)\n",
 876                             errno, strerror(errno));
 877                 goto out_err;
 878         }
 879
 880         return 0;
 881
 882 out_err:
 883         return -1;
 884 }
 885
 886 static int perf_top__setup_sample_type(struct perf_top *top __maybe_unused)
 887 {
 888         if (!sort__has_sym) {
 889                 if (symbol_conf.use_callchain) {
 890                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
 891                         return -EINVAL;
 892                 }
 893         } else if (callchain_param.mode != CHAIN_NONE) {
 894                 if (callchain_register_param(&callchain_param) < 0) {
 895                         ui__error("Can't register callchain params.\n");
 896                         return -EINVAL;
 897                 }
 898         }
 899
 900         return 0;
 901 }
 902
 903 static int __cmd_top(struct perf_top *top)
 904 {
 905         struct record_opts *opts = &top->record_opts;
 906         pthread_t thread;
 907         int ret;
 908
 909         top->session = perf_session__new(NULL, false, NULL);
 910         if (top->session == NULL)
 911                 return -ENOMEM;
 912
 913         machines__set_symbol_filter(&top->session->machines, symbol_filter);
 914
 915         if (!objdump_path) {
 916                 ret = perf_session_env__lookup_objdump(&top->session->header.env);
 917                 if (ret)
 918                         goto out_delete;
 919         }
 920
 921         ret = perf_top__setup_sample_type(top);
 922         if (ret)
 923                 goto out_delete;
 924
 925         machine__synthesize_threads(&top->session->machines.host, &opts->target,
 926                                     top->evlist->threads, false);
 927         ret = perf_top__start_counters(top);
 928         if (ret)
 929                 goto out_delete;
 930
 931         top->session->evlist = top->evlist;
 932         perf_session__set_id_hdr_size(top->session);
 933
 934         /*
 935          * When perf is starting the traced process, all the events (apart from
 936          * group members) have enable_on_exec=1 set, so don't spoil it by
 937          * prematurely enabling them.
 938          *
 939          * XXX 'top' still doesn't start workloads like record, trace, but should,
 940          * so leave the check here.
 941          */
 942         if (!target__none(&opts->target))
 943                 perf_evlist__enable(top->evlist);
 944
 945         /* Wait for a minimal set of events before starting the snapshot */
 946         poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
 947
 948         perf_top__mmap_read(top);
 949
 950         ret = -1;
 951         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 952                                                             display_thread), top)) {
 953                 ui__error("Could not create display thread.\n");
 954                 goto out_delete;
 955         }
 956
 957         if (top->realtime_prio) {
 958                 struct sched_param param;
 959
 960                 param.sched_priority = top->realtime_prio;
 961                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 962                         ui__error("Could not set realtime priority.\n");
 963                         goto out_delete;
 964                 }
 965         }
 966
 967         while (!done) {
 968                 u64 hits = top->samples;
 969
 970                 perf_top__mmap_read(top);
 971
 972                 if (hits == top->samples)
 973                         ret = poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
 974         }
 975
 976         ret = 0;
 977 out_delete:
 978         perf_session__delete(top->session);
 979         top->session = NULL;
 980
 981         return ret;
 982 }
 983
 984 static int
 985 callchain_opt(const struct option *opt, const char *arg, int unset)
 986 {
 987         symbol_conf.use_callchain = true;
 988         return record_callchain_opt(opt, arg, unset);
 989 }
 990
 991 static int
 992 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
 993 {
 994         symbol_conf.use_callchain = true;
 995         return record_parse_callchain_opt(opt, arg, unset);
 996 }
 997
 998 static int perf_top_config(const char *var, const char *value, void *cb)
 999 {
1000         struct perf_top *top = cb;
1001
1002         if (!strcmp(var, "top.call-graph"))
1003                 return record_parse_callchain(value, &top->record_opts);
1004
1005         return perf_default_config(var, value, cb);
1006 }
1007
1008 static int
1009 parse_percent_limit(const struct option *opt, const char *arg,
1010                     int unset __maybe_unused)
1011 {
1012         struct perf_top *top = opt->value;
1013
1014         top->min_percent = strtof(arg, NULL);
1015         return 0;
1016 }
1017
1018 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1019 {
1020         int status = -1;
1021         char errbuf[BUFSIZ];
1022         struct perf_top top = {
1023                 .count_filter        = 5,
1024                 .delay_secs          = 2,
1025                 .record_opts = {
1026                         .mmap_pages     = UINT_MAX,
1027                         .user_freq      = UINT_MAX,
1028                         .user_interval  = ULLONG_MAX,
1029                         .freq           = 4000, /* 4 KHz */
1030                         .target         = {
1031                                 .uses_mmap   = true,
1032                         },
1033                 },
1034                 .max_stack           = PERF_MAX_STACK_DEPTH,
1035                 .sym_pcnt_filter     = 5,
1036         };
1037         struct record_opts *opts = &top.record_opts;
1038         struct target *target = &opts->target;
1039         const struct option options[] = {
1040         OPT_CALLBACK('e', "event", &top.evlist, "event",
1041                      "event selector. use 'perf list' to list available events",
1042                      parse_events_option),
1043         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1044         OPT_STRING('p', "pid", &target->pid, "pid",
1045                     "profile events on existing process id"),
1046         OPT_STRING('t', "tid", &target->tid, "tid",
1047                     "profile events on existing thread id"),
1048         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1049                             "system-wide collection from all CPUs"),
1050         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1051                     "list of cpus to monitor"),
1052         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1053                    "file", "vmlinux pathname"),
1054         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1055                     "don't load vmlinux even if found"),
1056         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1057                     "hide kernel symbols"),
1058         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1059                      "number of mmap data pages",
1060                      perf_evlist__parse_mmap_pages),
1061         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1062                     "collect data with this RT SCHED_FIFO priority"),
1063         OPT_INTEGER('d', "delay", &top.delay_secs,
1064                     "number of seconds to delay between refreshes"),
1065         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1066                             "dump the symbol table used for profiling"),
1067         OPT_INTEGER('f', "count-filter", &top.count_filter,
1068                     "only display functions with more events than this"),
1069         OPT_BOOLEAN(0, "group", &opts->group,
1070                             "put the counters into a counter group"),
1071         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1072                     "child tasks do not inherit counters"),
1073         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1074                     "symbol to annotate"),
1075         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1076         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1077         OPT_INTEGER('E', "entries", &top.print_entries,
1078                     "display this many functions"),
1079         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1080                     "hide user symbols"),
1081         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1082         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1083         OPT_INCR('v', "verbose", &verbose,
1084                     "be more verbose (show counter open errors, etc)"),
1085         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1086                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1087                    " Please refer the man page for the complete list."),
1088         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1089                    "output field(s): overhead, period, sample plus all of sort keys"),
1090         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1091                     "Show a column with the number of samples"),
1092         OPT_CALLBACK_NOOPT('g', NULL, &top.record_opts,
1093                            NULL, "enables call-graph recording",
1094                            &callchain_opt),
1095         OPT_CALLBACK(0, "call-graph", &top.record_opts,
1096                      "mode[,dump_size]", record_callchain_help,
1097                      &parse_callchain_opt),
1098         OPT_INTEGER(0, "max-stack", &top.max_stack,
1099                     "Set the maximum stack depth when parsing the callchain. "
1100                     "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
1101         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1102                    "ignore callees of these functions in call graphs",
1103                    report_parse_ignore_callees_opt),
1104         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1105                     "Show a column with the sum of periods"),
1106         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1107                    "only consider symbols in these dsos"),
1108         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1109                    "only consider symbols in these comms"),
1110         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1111                    "only consider these symbols"),
1112         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1113                     "Interleave source code with assembly code (default)"),
1114         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1115                     "Display raw encoding of assembly instructions (default)"),
1116         OPT_STRING(0, "objdump", &objdump_path, "path",
1117                     "objdump binary to use for disassembly and annotations"),
1118         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1119                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1120         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1121         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1122                      "Don't show entries under that percent", parse_percent_limit),
1123         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1124                      "How to display percentage of filtered entries", parse_filter_percentage),
1125         OPT_END()
1126         };
1127         const char * const top_usage[] = {
1128                 "perf top [<options>]",
1129                 NULL
1130         };
1131
1132         top.evlist = perf_evlist__new();
1133         if (top.evlist == NULL)
1134                 return -ENOMEM;
1135
1136         perf_config(perf_top_config, &top);
1137
1138         argc = parse_options(argc, argv, options, top_usage, 0);
1139         if (argc)
1140                 usage_with_options(top_usage, options);
1141
1142         sort__mode = SORT_MODE__TOP;
1143         /* display thread wants entries to be collapsed in a different tree */
1144         sort__need_collapse = 1;
1145
1146         if (setup_sorting() < 0) {
1147                 if (sort_order)
1148                         parse_options_usage(top_usage, options, "s", 1);
1149                 if (field_order)
1150                         parse_options_usage(sort_order ? NULL : top_usage,
1151                                             options, "fields", 0);
1152                 goto out_delete_evlist;
1153         }
1154
1155         if (top.use_stdio)
1156                 use_browser = 0;
1157         else if (top.use_tui)
1158                 use_browser = 1;
1159
1160         setup_browser(false);
1161
1162         status = target__validate(target);
1163         if (status) {
1164                 target__strerror(target, status, errbuf, BUFSIZ);
1165                 ui__warning("%s\n", errbuf);
1166         }
1167
1168         status = target__parse_uid(target);
1169         if (status) {
1170                 int saved_errno = errno;
1171
1172                 target__strerror(target, status, errbuf, BUFSIZ);
1173                 ui__error("%s\n", errbuf);
1174
1175                 status = -saved_errno;
1176                 goto out_delete_evlist;
1177         }
1178
1179         if (target__none(target))
1180                 target->system_wide = true;
1181
1182         if (perf_evlist__create_maps(top.evlist, target) < 0)
1183                 usage_with_options(top_usage, options);
1184
1185         if (!top.evlist->nr_entries &&
1186             perf_evlist__add_default(top.evlist) < 0) {
1187                 ui__error("Not enough memory for event selector list\n");
1188                 goto out_delete_evlist;
1189         }
1190
1191         symbol_conf.nr_events = top.evlist->nr_entries;
1192
1193         if (top.delay_secs < 1)
1194                 top.delay_secs = 1;
1195
1196         if (record_opts__config(opts)) {
1197                 status = -EINVAL;
1198                 goto out_delete_evlist;
1199         }
1200
1201         top.sym_evsel = perf_evlist__first(top.evlist);
1202
1203         symbol_conf.priv_size = sizeof(struct annotation);
1204
1205         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1206         if (symbol__init() < 0)
1207                 return -1;
1208
1209         sort__setup_elide(stdout);
1210
1211         get_term_dimensions(&top.winsize);
1212         if (top.print_entries == 0) {
1213                 struct sigaction act = {
1214                         .sa_sigaction = perf_top__sig_winch,
1215                         .sa_flags     = SA_SIGINFO,
1216                 };
1217                 perf_top__update_print_entries(&top);
1218                 sigaction(SIGWINCH, &act, NULL);
1219         }
1220
1221         status = __cmd_top(&top);
1222
1223 out_delete_evlist:
1224         perf_evlist__delete(top.evlist);
1225
1226         return status;
1227 }