Pileus Git - ~andy/linux/blob - 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  *
   9  * Improvements and fixes by:
  10  *
  11  *   Arjan van de Ven <arjan@linux.intel.com>
  12  *   Yanmin Zhang <yanmin.zhang@intel.com>
  13  *   Wu Fengguang <fengguang.wu@intel.com>
  14  *   Mike Galbraith <efault@gmx.de>
  15  *   Paul Mackerras <paulus@samba.org>
  16  *
  17  * Released under the GPL v2. (and only v2, not any later version)
  18  */
  19 #include "builtin.h"
  20
  21 #include "perf.h"
  22
  23 #include "util/symbol.h"
  24 #include "util/color.h"
  25 #include "util/thread.h"
  26 #include "util/util.h"
  27 #include <linux/rbtree.h>
  28 #include "util/parse-options.h"
  29 #include "util/parse-events.h"
  30
  31 #include "util/debug.h"
  32
  33 #include <assert.h>
  34 #include <fcntl.h>
  35
  36 #include <stdio.h>
  37 #include <termios.h>
  38 #include <unistd.h>
  39
  40 #include <errno.h>
  41 #include <time.h>
  42 #include <sched.h>
  43 #include <pthread.h>
  44
  45 #include <sys/syscall.h>
  46 #include <sys/ioctl.h>
  47 #include <sys/poll.h>
  48 #include <sys/prctl.h>
  49 #include <sys/wait.h>
  50 #include <sys/uio.h>
  51 #include <sys/mman.h>
  52
  53 #include <linux/unistd.h>
  54 #include <linux/types.h>
  55
  56 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  57
  58 static int                      system_wide                     =      0;
  59
  60 static int                      default_interval                =      0;
  61
  62 static int                      count_filter                    =      5;
  63 static int                      print_entries;
  64
  65 static int                      target_pid                      =     -1;
  66 static int                      inherit                         =      0;
  67 static int                      profile_cpu                     =     -1;
  68 static int                      nr_cpus                         =      0;
  69 static unsigned int             realtime_prio                   =      0;
  70 static int                      group                           =      0;
  71 static unsigned int             page_size;
  72 static unsigned int             mmap_pages                      =     16;
  73 static int                      freq                            =   1000; /* 1 KHz */
  74
  75 static int                      delay_secs                      =      2;
  76 static int                      zero                            =      0;
  77 static int                      dump_symtab                     =      0;
  78
  79 static bool                     hide_kernel_symbols             =  false;
  80 static bool                     hide_user_symbols               =  false;
  81
  82 /*
  83  * Source
  84  */
  85
  86 struct source_line {
  87         u64                     eip;
  88         unsigned long           count[MAX_COUNTERS];
  89         char                    *line;
  90         struct source_line      *next;
  91 };
  92
  93 static char                     *sym_filter                     =   NULL;
  94 struct sym_entry                *sym_filter_entry               =   NULL;
  95 static int                      sym_pcnt_filter                 =      5;
  96 static int                      sym_counter                     =      0;
  97 static int                      display_weighted                =     -1;
  98
  99 /*
 100  * Symbols
 101  */
 102
 103 struct sym_entry {
 104         struct rb_node          rb_node;
 105         struct list_head        node;
 106         unsigned long           count[MAX_COUNTERS];
 107         unsigned long           snap_count;
 108         double                  weight;
 109         int                     skip;
 110         u8                      origin;
 111         struct map              *map;
 112         struct source_line      *source;
 113         struct source_line      *lines;
 114         struct source_line      **lines_tail;
 115         pthread_mutex_t         source_lock;
 116 };
 117
 118 /*
 119  * Source functions
 120  */
 121
 122 /* most GUI terminals set LINES (although some don't export it) */
 123 static int term_rows(void)
 124 {
 125         char *lines_string = getenv("LINES");
 126         int n_lines;
 127
 128         if (lines_string && (n_lines = atoi(lines_string)) > 0)
 129                 return n_lines;
 130 #ifdef TIOCGWINSZ
 131         else {
 132                 struct winsize ws;
 133                 if (!ioctl(1, TIOCGWINSZ, &ws) && ws.ws_row)
 134                         return ws.ws_row;
 135         }
 136 #endif
 137         return 25;
 138 }
 139
 140 static void update_print_entries(void)
 141 {
 142         print_entries = term_rows();
 143         if (print_entries > 9)
 144                 print_entries -= 9;
 145 }
 146
 147 static void sig_winch_handler(int sig __used)
 148 {
 149         update_print_entries();
 150 }
 151
 152 static void parse_source(struct sym_entry *syme)
 153 {
 154         struct symbol *sym;
 155         struct map *map;
 156         FILE *file;
 157         char command[PATH_MAX*2];
 158         const char *path;
 159         u64 len;
 160
 161         if (!syme)
 162                 return;
 163
 164         if (syme->lines) {
 165                 pthread_mutex_lock(&syme->source_lock);
 166                 goto out_assign;
 167         }
 168
 169         sym = (struct symbol *)(syme + 1);
 170         map = syme->map;
 171         path = map->dso->long_name;
 172
 173         len = sym->end - sym->start;
 174
 175         sprintf(command,
 176                 "objdump --start-address=0x%016Lx "
 177                          "--stop-address=0x%016Lx -dS %s",
 178                 map->unmap_ip(map, sym->start),
 179                 map->unmap_ip(map, sym->end), path);
 180
 181         file = popen(command, "r");
 182         if (!file)
 183                 return;
 184
 185         pthread_mutex_lock(&syme->source_lock);
 186         syme->lines_tail = &syme->lines;
 187         while (!feof(file)) {
 188                 struct source_line *src;
 189                 size_t dummy = 0;
 190                 char *c;
 191
 192                 src = malloc(sizeof(struct source_line));
 193                 assert(src != NULL);
 194                 memset(src, 0, sizeof(struct source_line));
 195
 196                 if (getline(&src->line, &dummy, file) < 0)
 197                         break;
 198                 if (!src->line)
 199                         break;
 200
 201                 c = strchr(src->line, '\n');
 202                 if (c)
 203                         *c = 0;
 204
 205                 src->next = NULL;
 206                 *syme->lines_tail = src;
 207                 syme->lines_tail = &src->next;
 208
 209                 if (strlen(src->line)>8 && src->line[8] == ':') {
 210                         src->eip = strtoull(src->line, NULL, 16);
 211                         src->eip = map->unmap_ip(map, src->eip);
 212                 }
 213                 if (strlen(src->line)>8 && src->line[16] == ':') {
 214                         src->eip = strtoull(src->line, NULL, 16);
 215                         src->eip = map->unmap_ip(map, src->eip);
 216                 }
 217         }
 218         pclose(file);
 219 out_assign:
 220         sym_filter_entry = syme;
 221         pthread_mutex_unlock(&syme->source_lock);
 222 }
 223
 224 static void __zero_source_counters(struct sym_entry *syme)
 225 {
 226         int i;
 227         struct source_line *line;
 228
 229         line = syme->lines;
 230         while (line) {
 231                 for (i = 0; i < nr_counters; i++)
 232                         line->count[i] = 0;
 233                 line = line->next;
 234         }
 235 }
 236
 237 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
 238 {
 239         struct source_line *line;
 240
 241         if (syme != sym_filter_entry)
 242                 return;
 243
 244         if (pthread_mutex_trylock(&syme->source_lock))
 245                 return;
 246
 247         if (!syme->source)
 248                 goto out_unlock;
 249
 250         for (line = syme->lines; line; line = line->next) {
 251                 if (line->eip == ip) {
 252                         line->count[counter]++;
 253                         break;
 254                 }
 255                 if (line->eip > ip)
 256                         break;
 257         }
 258 out_unlock:
 259         pthread_mutex_unlock(&syme->source_lock);
 260 }
 261
 262 static void lookup_sym_source(struct sym_entry *syme)
 263 {
 264         struct symbol *symbol = (struct symbol *)(syme + 1);
 265         struct source_line *line;
 266         char pattern[PATH_MAX];
 267
 268         sprintf(pattern, "<%s>:", symbol->name);
 269
 270         pthread_mutex_lock(&syme->source_lock);
 271         for (line = syme->lines; line; line = line->next) {
 272                 if (strstr(line->line, pattern)) {
 273                         syme->source = line;
 274                         break;
 275                 }
 276         }
 277         pthread_mutex_unlock(&syme->source_lock);
 278 }
 279
 280 static void show_lines(struct source_line *queue, int count, int total)
 281 {
 282         int i;
 283         struct source_line *line;
 284
 285         line = queue;
 286         for (i = 0; i < count; i++) {
 287                 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
 288
 289                 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
 290                 line = line->next;
 291         }
 292 }
 293
 294 #define TRACE_COUNT     3
 295
 296 static void show_details(struct sym_entry *syme)
 297 {
 298         struct symbol *symbol;
 299         struct source_line *line;
 300         struct source_line *line_queue = NULL;
 301         int displayed = 0;
 302         int line_queue_count = 0, total = 0, more = 0;
 303
 304         if (!syme)
 305                 return;
 306
 307         if (!syme->source)
 308                 lookup_sym_source(syme);
 309
 310         if (!syme->source)
 311                 return;
 312
 313         symbol = (struct symbol *)(syme + 1);
 314         printf("Showing %s for %s\n", event_name(sym_counter), symbol->name);
 315         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
 316
 317         pthread_mutex_lock(&syme->source_lock);
 318         line = syme->source;
 319         while (line) {
 320                 total += line->count[sym_counter];
 321                 line = line->next;
 322         }
 323
 324         line = syme->source;
 325         while (line) {
 326                 float pcnt = 0.0;
 327
 328                 if (!line_queue_count)
 329                         line_queue = line;
 330                 line_queue_count++;
 331
 332                 if (line->count[sym_counter])
 333                         pcnt = 100.0 * line->count[sym_counter] / (float)total;
 334                 if (pcnt >= (float)sym_pcnt_filter) {
 335                         if (displayed <= print_entries)
 336                                 show_lines(line_queue, line_queue_count, total);
 337                         else more++;
 338                         displayed += line_queue_count;
 339                         line_queue_count = 0;
 340                         line_queue = NULL;
 341                 } else if (line_queue_count > TRACE_COUNT) {
 342                         line_queue = line_queue->next;
 343                         line_queue_count--;
 344                 }
 345
 346                 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
 347                 line = line->next;
 348         }
 349         pthread_mutex_unlock(&syme->source_lock);
 350         if (more)
 351                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 352 }
 353
 354 /*
 355  * Symbols will be added here in event__process_sample and will get out
 356  * after decayed.
 357  */
 358 static LIST_HEAD(active_symbols);
 359 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
 360
 361 /*
 362  * Ordering weight: count-1 * count-2 * ... / count-n
 363  */
 364 static double sym_weight(const struct sym_entry *sym)
 365 {
 366         double weight = sym->snap_count;
 367         int counter;
 368
 369         if (!display_weighted)
 370                 return weight;
 371
 372         for (counter = 1; counter < nr_counters-1; counter++)
 373                 weight *= sym->count[counter];
 374
 375         weight /= (sym->count[counter] + 1);
 376
 377         return weight;
 378 }
 379
 380 static long                     samples;
 381 static long                     userspace_samples;
 382 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
 383
 384 static void __list_insert_active_sym(struct sym_entry *syme)
 385 {
 386         list_add(&syme->node, &active_symbols);
 387 }
 388
 389 static void list_remove_active_sym(struct sym_entry *syme)
 390 {
 391         pthread_mutex_lock(&active_symbols_lock);
 392         list_del_init(&syme->node);
 393         pthread_mutex_unlock(&active_symbols_lock);
 394 }
 395
 396 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
 397 {
 398         struct rb_node **p = &tree->rb_node;
 399         struct rb_node *parent = NULL;
 400         struct sym_entry *iter;
 401
 402         while (*p != NULL) {
 403                 parent = *p;
 404                 iter = rb_entry(parent, struct sym_entry, rb_node);
 405
 406                 if (se->weight > iter->weight)
 407                         p = &(*p)->rb_left;
 408                 else
 409                         p = &(*p)->rb_right;
 410         }
 411
 412         rb_link_node(&se->rb_node, parent, p);
 413         rb_insert_color(&se->rb_node, tree);
 414 }
 415
 416 static void print_sym_table(void)
 417 {
 418         int printed = 0, j;
 419         int counter, snap = !display_weighted ? sym_counter : 0;
 420         float samples_per_sec = samples/delay_secs;
 421         float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
 422         float sum_ksamples = 0.0;
 423         struct sym_entry *syme, *n;
 424         struct rb_root tmp = RB_ROOT;
 425         struct rb_node *nd;
 426
 427         samples = userspace_samples = 0;
 428
 429         /* Sort the active symbols */
 430         pthread_mutex_lock(&active_symbols_lock);
 431         syme = list_entry(active_symbols.next, struct sym_entry, node);
 432         pthread_mutex_unlock(&active_symbols_lock);
 433
 434         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
 435                 syme->snap_count = syme->count[snap];
 436                 if (syme->snap_count != 0) {
 437                         if ((hide_user_symbols &&
 438                              syme->origin == PERF_RECORD_MISC_USER) ||
 439                             (hide_kernel_symbols &&
 440                              syme->origin == PERF_RECORD_MISC_KERNEL)) {
 441                                 list_remove_active_sym(syme);
 442                                 continue;
 443                         }
 444                         syme->weight = sym_weight(syme);
 445                         rb_insert_active_sym(&tmp, syme);
 446                         sum_ksamples += syme->snap_count;
 447
 448                         for (j = 0; j < nr_counters; j++)
 449                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
 450                 } else
 451                         list_remove_active_sym(syme);
 452         }
 453
 454         puts(CONSOLE_CLEAR);
 455
 456         printf(
 457 "------------------------------------------------------------------------------\n");
 458         printf( "   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% [",
 459                 samples_per_sec,
 460                 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
 461
 462         if (nr_counters == 1 || !display_weighted) {
 463                 printf("%Ld", (u64)attrs[0].sample_period);
 464                 if (freq)
 465                         printf("Hz ");
 466                 else
 467                         printf(" ");
 468         }
 469
 470         if (!display_weighted)
 471                 printf("%s", event_name(sym_counter));
 472         else for (counter = 0; counter < nr_counters; counter++) {
 473                 if (counter)
 474                         printf("/");
 475
 476                 printf("%s", event_name(counter));
 477         }
 478
 479         printf( "], ");
 480
 481         if (target_pid != -1)
 482                 printf(" (target_pid: %d", target_pid);
 483         else
 484                 printf(" (all");
 485
 486         if (profile_cpu != -1)
 487                 printf(", cpu: %d)\n", profile_cpu);
 488         else {
 489                 if (target_pid != -1)
 490                         printf(")\n");
 491                 else
 492                         printf(", %d CPUs)\n", nr_cpus);
 493         }
 494
 495         printf("------------------------------------------------------------------------------\n\n");
 496
 497         if (sym_filter_entry) {
 498                 show_details(sym_filter_entry);
 499                 return;
 500         }
 501
 502         if (nr_counters == 1)
 503                 printf("             samples  pcnt");
 504         else
 505                 printf("   weight    samples  pcnt");
 506
 507         if (verbose)
 508                 printf("         RIP       ");
 509         printf(" function                                 DSO\n");
 510         printf("   %s    _______ _____",
 511                nr_counters == 1 ? "      " : "______");
 512         if (verbose)
 513                 printf(" ________________");
 514         printf(" ________________________________ ________________\n\n");
 515
 516         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
 517                 struct symbol *sym;
 518                 double pcnt;
 519
 520                 syme = rb_entry(nd, struct sym_entry, rb_node);
 521                 sym = (struct symbol *)(syme + 1);
 522
 523                 if (++printed > print_entries || (int)syme->snap_count < count_filter)
 524                         continue;
 525
 526                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
 527                                          sum_ksamples));
 528
 529                 if (nr_counters == 1 || !display_weighted)
 530                         printf("%20.2f ", syme->weight);
 531                 else
 532                         printf("%9.1f %10ld ", syme->weight, syme->snap_count);
 533
 534                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
 535                 if (verbose)
 536                         printf(" %016llx", sym->start);
 537                 printf(" %-32s", sym->name);
 538                 printf(" %s", syme->map->dso->short_name);
 539                 printf("\n");
 540         }
 541 }
 542
 543 static void prompt_integer(int *target, const char *msg)
 544 {
 545         char *buf = malloc(0), *p;
 546         size_t dummy = 0;
 547         int tmp;
 548
 549         fprintf(stdout, "\n%s: ", msg);
 550         if (getline(&buf, &dummy, stdin) < 0)
 551                 return;
 552
 553         p = strchr(buf, '\n');
 554         if (p)
 555                 *p = 0;
 556
 557         p = buf;
 558         while(*p) {
 559                 if (!isdigit(*p))
 560                         goto out_free;
 561                 p++;
 562         }
 563         tmp = strtoul(buf, NULL, 10);
 564         *target = tmp;
 565 out_free:
 566         free(buf);
 567 }
 568
 569 static void prompt_percent(int *target, const char *msg)
 570 {
 571         int tmp = 0;
 572
 573         prompt_integer(&tmp, msg);
 574         if (tmp >= 0 && tmp <= 100)
 575                 *target = tmp;
 576 }
 577
 578 static void prompt_symbol(struct sym_entry **target, const char *msg)
 579 {
 580         char *buf = malloc(0), *p;
 581         struct sym_entry *syme = *target, *n, *found = NULL;
 582         size_t dummy = 0;
 583
 584         /* zero counters of active symbol */
 585         if (syme) {
 586                 pthread_mutex_lock(&syme->source_lock);
 587                 __zero_source_counters(syme);
 588                 *target = NULL;
 589                 pthread_mutex_unlock(&syme->source_lock);
 590         }
 591
 592         fprintf(stdout, "\n%s: ", msg);
 593         if (getline(&buf, &dummy, stdin) < 0)
 594                 goto out_free;
 595
 596         p = strchr(buf, '\n');
 597         if (p)
 598                 *p = 0;
 599
 600         pthread_mutex_lock(&active_symbols_lock);
 601         syme = list_entry(active_symbols.next, struct sym_entry, node);
 602         pthread_mutex_unlock(&active_symbols_lock);
 603
 604         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
 605                 struct symbol *sym = (struct symbol *)(syme + 1);
 606
 607                 if (!strcmp(buf, sym->name)) {
 608                         found = syme;
 609                         break;
 610                 }
 611         }
 612
 613         if (!found) {
 614                 fprintf(stderr, "Sorry, %s is not active.\n", sym_filter);
 615                 sleep(1);
 616                 return;
 617         } else
 618                 parse_source(found);
 619
 620 out_free:
 621         free(buf);
 622 }
 623
 624 static void print_mapped_keys(void)
 625 {
 626         char *name = NULL;
 627
 628         if (sym_filter_entry) {
 629                 struct symbol *sym = (struct symbol *)(sym_filter_entry+1);
 630                 name = sym->name;
 631         }
 632
 633         fprintf(stdout, "\nMapped keys:\n");
 634         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", delay_secs);
 635         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", print_entries);
 636
 637         if (nr_counters > 1)
 638                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(sym_counter));
 639
 640         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", count_filter);
 641
 642         if (vmlinux_name) {
 643                 fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
 644                 fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 645                 fprintf(stdout, "\t[S]     stop annotation.\n");
 646         }
 647
 648         if (nr_counters > 1)
 649                 fprintf(stdout, "\t[w]     toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
 650
 651         fprintf(stdout,
 652                 "\t[K]     hide kernel_symbols symbols.             \t(%s)\n",
 653                 hide_kernel_symbols ? "yes" : "no");
 654         fprintf(stdout,
 655                 "\t[U]     hide user symbols.               \t(%s)\n",
 656                 hide_user_symbols ? "yes" : "no");
 657         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", zero ? 1 : 0);
 658         fprintf(stdout, "\t[qQ]    quit.\n");
 659 }
 660
 661 static int key_mapped(int c)
 662 {
 663         switch (c) {
 664                 case 'd':
 665                 case 'e':
 666                 case 'f':
 667                 case 'z':
 668                 case 'q':
 669                 case 'Q':
 670                 case 'K':
 671                 case 'U':
 672                         return 1;
 673                 case 'E':
 674                 case 'w':
 675                         return nr_counters > 1 ? 1 : 0;
 676                 case 'F':
 677                 case 's':
 678                 case 'S':
 679                         return vmlinux_name ? 1 : 0;
 680                 default:
 681                         break;
 682         }
 683
 684         return 0;
 685 }
 686
 687 static void handle_keypress(int c)
 688 {
 689         if (!key_mapped(c)) {
 690                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 691                 struct termios tc, save;
 692
 693                 print_mapped_keys();
 694                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 695                 fflush(stdout);
 696
 697                 tcgetattr(0, &save);
 698                 tc = save;
 699                 tc.c_lflag &= ~(ICANON | ECHO);
 700                 tc.c_cc[VMIN] = 0;
 701                 tc.c_cc[VTIME] = 0;
 702                 tcsetattr(0, TCSANOW, &tc);
 703
 704                 poll(&stdin_poll, 1, -1);
 705                 c = getc(stdin);
 706
 707                 tcsetattr(0, TCSAFLUSH, &save);
 708                 if (!key_mapped(c))
 709                         return;
 710         }
 711
 712         switch (c) {
 713                 case 'd':
 714                         prompt_integer(&delay_secs, "Enter display delay");
 715                         if (delay_secs < 1)
 716                                 delay_secs = 1;
 717                         break;
 718                 case 'e':
 719                         prompt_integer(&print_entries, "Enter display entries (lines)");
 720                         if (print_entries == 0) {
 721                                 update_print_entries();
 722                                 signal(SIGWINCH, sig_winch_handler);
 723                         } else
 724                                 signal(SIGWINCH, SIG_DFL);
 725                         break;
 726                 case 'E':
 727                         if (nr_counters > 1) {
 728                                 int i;
 729
 730                                 fprintf(stderr, "\nAvailable events:");
 731                                 for (i = 0; i < nr_counters; i++)
 732                                         fprintf(stderr, "\n\t%d %s", i, event_name(i));
 733
 734                                 prompt_integer(&sym_counter, "Enter details event counter");
 735
 736                                 if (sym_counter >= nr_counters) {
 737                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
 738                                         sym_counter = 0;
 739                                         sleep(1);
 740                                 }
 741                         } else sym_counter = 0;
 742                         break;
 743                 case 'f':
 744                         prompt_integer(&count_filter, "Enter display event count filter");
 745                         break;
 746                 case 'F':
 747                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
 748                         break;
 749                 case 'K':
 750                         hide_kernel_symbols = !hide_kernel_symbols;
 751                         break;
 752                 case 'q':
 753                 case 'Q':
 754                         printf("exiting.\n");
 755                         exit(0);
 756                 case 's':
 757                         prompt_symbol(&sym_filter_entry, "Enter details symbol");
 758                         break;
 759                 case 'S':
 760                         if (!sym_filter_entry)
 761                                 break;
 762                         else {
 763                                 struct sym_entry *syme = sym_filter_entry;
 764
 765                                 pthread_mutex_lock(&syme->source_lock);
 766                                 sym_filter_entry = NULL;
 767                                 __zero_source_counters(syme);
 768                                 pthread_mutex_unlock(&syme->source_lock);
 769                         }
 770                         break;
 771                 case 'U':
 772                         hide_user_symbols = !hide_user_symbols;
 773                         break;
 774                 case 'w':
 775                         display_weighted = ~display_weighted;
 776                         break;
 777                 case 'z':
 778                         zero = ~zero;
 779                         break;
 780                 default:
 781                         break;
 782         }
 783 }
 784
 785 static void *display_thread(void *arg __used)
 786 {
 787         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 788         struct termios tc, save;
 789         int delay_msecs, c;
 790
 791         tcgetattr(0, &save);
 792         tc = save;
 793         tc.c_lflag &= ~(ICANON | ECHO);
 794         tc.c_cc[VMIN] = 0;
 795         tc.c_cc[VTIME] = 0;
 796
 797 repeat:
 798         delay_msecs = delay_secs * 1000;
 799         tcsetattr(0, TCSANOW, &tc);
 800         /* trash return*/
 801         getc(stdin);
 802
 803         do {
 804                 print_sym_table();
 805         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
 806
 807         c = getc(stdin);
 808         tcsetattr(0, TCSAFLUSH, &save);
 809
 810         handle_keypress(c);
 811         goto repeat;
 812
 813         return NULL;
 814 }
 815
 816 /* Tag samples to be skipped. */
 817 static const char *skip_symbols[] = {
 818         "default_idle",
 819         "cpu_idle",
 820         "enter_idle",
 821         "exit_idle",
 822         "mwait_idle",
 823         "mwait_idle_with_hints",
 824         "poll_idle",
 825         "ppc64_runlatch_off",
 826         "pseries_dedicated_idle_sleep",
 827         NULL
 828 };
 829
 830 static int symbol_filter(struct map *map, struct symbol *sym)
 831 {
 832         struct sym_entry *syme;
 833         const char *name = sym->name;
 834         int i;
 835
 836         /*
 837          * ppc64 uses function descriptors and appends a '.' to the
 838          * start of every instruction address. Remove it.
 839          */
 840         if (name[0] == '.')
 841                 name++;
 842
 843         if (!strcmp(name, "_text") ||
 844             !strcmp(name, "_etext") ||
 845             !strcmp(name, "_sinittext") ||
 846             !strncmp("init_module", name, 11) ||
 847             !strncmp("cleanup_module", name, 14) ||
 848             strstr(name, "_text_start") ||
 849             strstr(name, "_text_end"))
 850                 return 1;
 851
 852         syme = symbol__priv(sym);
 853         syme->map = map;
 854         pthread_mutex_init(&syme->source_lock, NULL);
 855         if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter))
 856                 sym_filter_entry = syme;
 857
 858         for (i = 0; skip_symbols[i]; i++) {
 859                 if (!strcmp(skip_symbols[i], name)) {
 860                         syme->skip = 1;
 861                         break;
 862                 }
 863         }
 864
 865         return 0;
 866 }
 867
 868 static int parse_symbols(void)
 869 {
 870         if (dsos__load_kernel(vmlinux_name, symbol_filter, 1) <= 0)
 871                 return -1;
 872
 873         if (dump_symtab)
 874                 dsos__fprintf(stderr);
 875
 876         return 0;
 877 }
 878
 879 static void event__process_sample(const event_t *self, int counter)
 880 {
 881         u64 ip = self->ip.ip;
 882         struct map *map;
 883         struct sym_entry *syme;
 884         struct symbol *sym;
 885         u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 886
 887         switch (origin) {
 888         case PERF_RECORD_MISC_USER: {
 889                 struct thread *thread;
 890
 891                 if (hide_user_symbols)
 892                         return;
 893
 894                 thread = threads__findnew(self->ip.pid);
 895                 if (thread == NULL)
 896                         return;
 897
 898                 map = thread__find_map(thread, ip);
 899                 if (map != NULL) {
 900                         ip = map->map_ip(map, ip);
 901                         sym = map__find_symbol(map, ip, symbol_filter);
 902                         if (sym == NULL)
 903                                 return;
 904                         userspace_samples++;
 905                         break;
 906                 }
 907         }
 908                 /*
 909                  * If this is outside of all known maps,
 910                  * and is a negative address, try to look it
 911                  * up in the kernel dso, as it might be a
 912                  * vsyscall or vdso (which executes in user-mode).
 913                  */
 914                 if ((long long)ip >= 0)
 915                         return;
 916                 /* Fall thru */
 917         case PERF_RECORD_MISC_KERNEL:
 918                 if (hide_kernel_symbols)
 919                         return;
 920
 921                 sym = kernel_maps__find_symbol(ip, &map);
 922                 if (sym == NULL)
 923                         return;
 924                 break;
 925         default:
 926                 return;
 927         }
 928
 929         syme = symbol__priv(sym);
 930
 931         if (!syme->skip) {
 932                 syme->count[counter]++;
 933                 syme->origin = origin;
 934                 record_precise_ip(syme, counter, ip);
 935                 pthread_mutex_lock(&active_symbols_lock);
 936                 if (list_empty(&syme->node) || !syme->node.next)
 937                         __list_insert_active_sym(syme);
 938                 pthread_mutex_unlock(&active_symbols_lock);
 939                 ++samples;
 940                 return;
 941         }
 942 }
 943
 944 static void event__process_mmap(event_t *self)
 945 {
 946         struct thread *thread = threads__findnew(self->mmap.pid);
 947
 948         if (thread != NULL) {
 949                 struct map *map = map__new(&self->mmap, NULL, 0);
 950                 if (map != NULL)
 951                         thread__insert_map(thread, map);
 952         }
 953 }
 954
 955 static void event__process_comm(event_t *self)
 956 {
 957         struct thread *thread = threads__findnew(self->comm.pid);
 958
 959         if (thread != NULL)
 960                 thread__set_comm(thread, self->comm.comm);
 961 }
 962
 963 static int event__process(event_t *event)
 964 {
 965         switch (event->header.type) {
 966         case PERF_RECORD_COMM:
 967                 event__process_comm(event);
 968                 break;
 969         case PERF_RECORD_MMAP:
 970                 event__process_mmap(event);
 971                 break;
 972         default:
 973                 break;
 974         }
 975
 976         return 0;
 977 }
 978
 979 struct mmap_data {
 980         int                     counter;
 981         void                    *base;
 982         int                     mask;
 983         unsigned int            prev;
 984 };
 985
 986 static unsigned int mmap_read_head(struct mmap_data *md)
 987 {
 988         struct perf_event_mmap_page *pc = md->base;
 989         int head;
 990
 991         head = pc->data_head;
 992         rmb();
 993
 994         return head;
 995 }
 996
 997 static void mmap_read_counter(struct mmap_data *md)
 998 {
 999         unsigned int head = mmap_read_head(md);
1000         unsigned int old = md->prev;
1001         unsigned char *data = md->base + page_size;
1002         int diff;
1003
1004         /*
1005          * If we're further behind than half the buffer, there's a chance
1006          * the writer will bite our tail and mess up the samples under us.
1007          *
1008          * If we somehow ended up ahead of the head, we got messed up.
1009          *
1010          * In either case, truncate and restart at head.
1011          */
1012         diff = head - old;
1013         if (diff > md->mask / 2 || diff < 0) {
1014                 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
1015
1016                 /*
1017                  * head points to a known good entry, start there.
1018                  */
1019                 old = head;
1020         }
1021
1022         for (; old != head;) {
1023                 event_t *event = (event_t *)&data[old & md->mask];
1024
1025                 event_t event_copy;
1026
1027                 size_t size = event->header.size;
1028
1029                 /*
1030                  * Event straddles the mmap boundary -- header should always
1031                  * be inside due to u64 alignment of output.
1032                  */
1033                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
1034                         unsigned int offset = old;
1035                         unsigned int len = min(sizeof(*event), size), cpy;
1036                         void *dst = &event_copy;
1037
1038                         do {
1039                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
1040                                 memcpy(dst, &data[offset & md->mask], cpy);
1041                                 offset += cpy;
1042                                 dst += cpy;
1043                                 len -= cpy;
1044                         } while (len);
1045
1046                         event = &event_copy;
1047                 }
1048
1049                 if (event->header.type == PERF_RECORD_SAMPLE)
1050                         event__process_sample(event, md->counter);
1051                 else
1052                         event__process(event);
1053                 old += size;
1054         }
1055
1056         md->prev = old;
1057 }
1058
1059 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
1060 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
1061
1062 static void mmap_read(void)
1063 {
1064         int i, counter;
1065
1066         for (i = 0; i < nr_cpus; i++) {
1067                 for (counter = 0; counter < nr_counters; counter++)
1068                         mmap_read_counter(&mmap_array[i][counter]);
1069         }
1070 }
1071
1072 int nr_poll;
1073 int group_fd;
1074
1075 static void start_counter(int i, int counter)
1076 {
1077         struct perf_event_attr *attr;
1078         int cpu;
1079
1080         cpu = profile_cpu;
1081         if (target_pid == -1 && profile_cpu == -1)
1082                 cpu = i;
1083
1084         attr = attrs + counter;
1085
1086         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1087
1088         if (freq) {
1089                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
1090                 attr->freq              = 1;
1091                 attr->sample_freq       = freq;
1092         }
1093
1094         attr->inherit           = (cpu < 0) && inherit;
1095         attr->mmap              = 1;
1096
1097 try_again:
1098         fd[i][counter] = sys_perf_event_open(attr, target_pid, cpu, group_fd, 0);
1099
1100         if (fd[i][counter] < 0) {
1101                 int err = errno;
1102
1103                 if (err == EPERM || err == EACCES)
1104                         die("No permission - are you root?\n");
1105                 /*
1106                  * If it's cycles then fall back to hrtimer
1107                  * based cpu-clock-tick sw counter, which
1108                  * is always available even if no PMU support:
1109                  */
1110                 if (attr->type == PERF_TYPE_HARDWARE
1111                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1112
1113                         if (verbose)
1114                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
1115
1116                         attr->type = PERF_TYPE_SOFTWARE;
1117                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
1118                         goto try_again;
1119                 }
1120                 printf("\n");
1121                 error("perfcounter syscall returned with %d (%s)\n",
1122                         fd[i][counter], strerror(err));
1123                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1124                 exit(-1);
1125         }
1126         assert(fd[i][counter] >= 0);
1127         fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
1128
1129         /*
1130          * First counter acts as the group leader:
1131          */
1132         if (group && group_fd == -1)
1133                 group_fd = fd[i][counter];
1134
1135         event_array[nr_poll].fd = fd[i][counter];
1136         event_array[nr_poll].events = POLLIN;
1137         nr_poll++;
1138
1139         mmap_array[i][counter].counter = counter;
1140         mmap_array[i][counter].prev = 0;
1141         mmap_array[i][counter].mask = mmap_pages*page_size - 1;
1142         mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
1143                         PROT_READ, MAP_SHARED, fd[i][counter], 0);
1144         if (mmap_array[i][counter].base == MAP_FAILED)
1145                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1146 }
1147
1148 static int __cmd_top(void)
1149 {
1150         pthread_t thread;
1151         int i, counter;
1152         int ret;
1153
1154         if (target_pid != -1)
1155                 event__synthesize_thread(target_pid, event__process);
1156         else
1157                 event__synthesize_threads(event__process);
1158
1159         for (i = 0; i < nr_cpus; i++) {
1160                 group_fd = -1;
1161                 for (counter = 0; counter < nr_counters; counter++)
1162                         start_counter(i, counter);
1163         }
1164
1165         /* Wait for a minimal set of events before starting the snapshot */
1166         poll(event_array, nr_poll, 100);
1167
1168         mmap_read();
1169
1170         if (pthread_create(&thread, NULL, display_thread, NULL)) {
1171                 printf("Could not create display thread.\n");
1172                 exit(-1);
1173         }
1174
1175         if (realtime_prio) {
1176                 struct sched_param param;
1177
1178                 param.sched_priority = realtime_prio;
1179                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1180                         printf("Could not set realtime priority.\n");
1181                         exit(-1);
1182                 }
1183         }
1184
1185         while (1) {
1186                 int hits = samples;
1187
1188                 mmap_read();
1189
1190                 if (hits == samples)
1191                         ret = poll(event_array, nr_poll, 100);
1192         }
1193
1194         return 0;
1195 }
1196
1197 static const char * const top_usage[] = {
1198         "perf top [<options>]",
1199         NULL
1200 };
1201
1202 static const struct option options[] = {
1203         OPT_CALLBACK('e', "event", NULL, "event",
1204                      "event selector. use 'perf list' to list available events",
1205                      parse_events),
1206         OPT_INTEGER('c', "count", &default_interval,
1207                     "event period to sample"),
1208         OPT_INTEGER('p', "pid", &target_pid,
1209                     "profile events on existing pid"),
1210         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1211                             "system-wide collection from all CPUs"),
1212         OPT_INTEGER('C', "CPU", &profile_cpu,
1213                     "CPU to profile on"),
1214         OPT_STRING('k', "vmlinux", &vmlinux_name, "file", "vmlinux pathname"),
1215         OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols,
1216                     "hide kernel symbols"),
1217         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
1218                     "number of mmap data pages"),
1219         OPT_INTEGER('r', "realtime", &realtime_prio,
1220                     "collect data with this RT SCHED_FIFO priority"),
1221         OPT_INTEGER('d', "delay", &delay_secs,
1222                     "number of seconds to delay between refreshes"),
1223         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1224                             "dump the symbol table used for profiling"),
1225         OPT_INTEGER('f', "count-filter", &count_filter,
1226                     "only display functions with more events than this"),
1227         OPT_BOOLEAN('g', "group", &group,
1228                             "put the counters into a counter group"),
1229         OPT_BOOLEAN('i', "inherit", &inherit,
1230                     "child tasks inherit counters"),
1231         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1232                     "symbol to annotate - requires -k option"),
1233         OPT_BOOLEAN('z', "zero", &zero,
1234                     "zero history across updates"),
1235         OPT_INTEGER('F', "freq", &freq,
1236                     "profile at this frequency"),
1237         OPT_INTEGER('E', "entries", &print_entries,
1238                     "display this many functions"),
1239         OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols,
1240                     "hide user symbols"),
1241         OPT_BOOLEAN('v', "verbose", &verbose,
1242                     "be more verbose (show counter open errors, etc)"),
1243         OPT_END()
1244 };
1245
1246 int cmd_top(int argc, const char **argv, const char *prefix __used)
1247 {
1248         int counter;
1249
1250         symbol__init(sizeof(struct sym_entry));
1251
1252         page_size = sysconf(_SC_PAGE_SIZE);
1253
1254         argc = parse_options(argc, argv, options, top_usage, 0);
1255         if (argc)
1256                 usage_with_options(top_usage, options);
1257
1258         /* CPU and PID are mutually exclusive */
1259         if (target_pid != -1 && profile_cpu != -1) {
1260                 printf("WARNING: PID switch overriding CPU\n");
1261                 sleep(1);
1262                 profile_cpu = -1;
1263         }
1264
1265         if (!nr_counters)
1266                 nr_counters = 1;
1267
1268         if (delay_secs < 1)
1269                 delay_secs = 1;
1270
1271         parse_symbols();
1272         parse_source(sym_filter_entry);
1273
1274
1275         /*
1276          * User specified count overrides default frequency.
1277          */
1278         if (default_interval)
1279                 freq = 0;
1280         else if (freq) {
1281                 default_interval = freq;
1282         } else {
1283                 fprintf(stderr, "frequency and count are zero, aborting\n");
1284                 exit(EXIT_FAILURE);
1285         }
1286
1287         /*
1288          * Fill in the ones not specifically initialized via -c:
1289          */
1290         for (counter = 0; counter < nr_counters; counter++) {
1291                 if (attrs[counter].sample_period)
1292                         continue;
1293
1294                 attrs[counter].sample_period = default_interval;
1295         }
1296
1297         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1298         assert(nr_cpus <= MAX_NR_CPUS);
1299         assert(nr_cpus >= 0);
1300
1301         if (target_pid != -1 || profile_cpu != -1)
1302                 nr_cpus = 1;
1303
1304         if (print_entries == 0) {
1305                 update_print_entries();
1306                 signal(SIGWINCH, sig_winch_handler);
1307         }
1308
1309         return __cmd_top();
1310 }