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