SAFE public projects git trees. - safe/jmp/linux-2.6/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                   =     15;
  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 static void parse_source(struct sym_entry *syme)
 119 {
 120         struct symbol *sym;
 121         struct map *map;
 122         FILE *file;
 123         char command[PATH_MAX*2];
 124         const char *path;
 125         u64 len;
 126
 127         if (!syme)
 128                 return;
 129
 130         if (syme->lines) {
 131                 pthread_mutex_lock(&syme->source_lock);
 132                 goto out_assign;
 133         }
 134
 135         sym = (struct symbol *)(syme + 1);
 136         map = syme->map;
 137         path = map->dso->long_name;
 138
 139         len = sym->end - sym->start;
 140
 141         sprintf(command,
 142                 "objdump --start-address=0x%016Lx "
 143                          "--stop-address=0x%016Lx -dS %s",
 144                 map->unmap_ip(map, sym->start),
 145                 map->unmap_ip(map, sym->end), path);
 146
 147         file = popen(command, "r");
 148         if (!file)
 149                 return;
 150
 151         pthread_mutex_lock(&syme->source_lock);
 152         syme->lines_tail = &syme->lines;
 153         while (!feof(file)) {
 154                 struct source_line *src;
 155                 size_t dummy = 0;
 156                 char *c;
 157
 158                 src = malloc(sizeof(struct source_line));
 159                 assert(src != NULL);
 160                 memset(src, 0, sizeof(struct source_line));
 161
 162                 if (getline(&src->line, &dummy, file) < 0)
 163                         break;
 164                 if (!src->line)
 165                         break;
 166
 167                 c = strchr(src->line, '\n');
 168                 if (c)
 169                         *c = 0;
 170
 171                 src->next = NULL;
 172                 *syme->lines_tail = src;
 173                 syme->lines_tail = &src->next;
 174
 175                 if (strlen(src->line)>8 && src->line[8] == ':') {
 176                         src->eip = strtoull(src->line, NULL, 16);
 177                         src->eip = map->unmap_ip(map, src->eip);
 178                 }
 179                 if (strlen(src->line)>8 && src->line[16] == ':') {
 180                         src->eip = strtoull(src->line, NULL, 16);
 181                         src->eip = map->unmap_ip(map, src->eip);
 182                 }
 183         }
 184         pclose(file);
 185 out_assign:
 186         sym_filter_entry = syme;
 187         pthread_mutex_unlock(&syme->source_lock);
 188 }
 189
 190 static void __zero_source_counters(struct sym_entry *syme)
 191 {
 192         int i;
 193         struct source_line *line;
 194
 195         line = syme->lines;
 196         while (line) {
 197                 for (i = 0; i < nr_counters; i++)
 198                         line->count[i] = 0;
 199                 line = line->next;
 200         }
 201 }
 202
 203 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
 204 {
 205         struct source_line *line;
 206
 207         if (syme != sym_filter_entry)
 208                 return;
 209
 210         if (pthread_mutex_trylock(&syme->source_lock))
 211                 return;
 212
 213         if (!syme->source)
 214                 goto out_unlock;
 215
 216         for (line = syme->lines; line; line = line->next) {
 217                 if (line->eip == ip) {
 218                         line->count[counter]++;
 219                         break;
 220                 }
 221                 if (line->eip > ip)
 222                         break;
 223         }
 224 out_unlock:
 225         pthread_mutex_unlock(&syme->source_lock);
 226 }
 227
 228 static void lookup_sym_source(struct sym_entry *syme)
 229 {
 230         struct symbol *symbol = (struct symbol *)(syme + 1);
 231         struct source_line *line;
 232         char pattern[PATH_MAX];
 233
 234         sprintf(pattern, "<%s>:", symbol->name);
 235
 236         pthread_mutex_lock(&syme->source_lock);
 237         for (line = syme->lines; line; line = line->next) {
 238                 if (strstr(line->line, pattern)) {
 239                         syme->source = line;
 240                         break;
 241                 }
 242         }
 243         pthread_mutex_unlock(&syme->source_lock);
 244 }
 245
 246 static void show_lines(struct source_line *queue, int count, int total)
 247 {
 248         int i;
 249         struct source_line *line;
 250
 251         line = queue;
 252         for (i = 0; i < count; i++) {
 253                 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
 254
 255                 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
 256                 line = line->next;
 257         }
 258 }
 259
 260 #define TRACE_COUNT     3
 261
 262 static void show_details(struct sym_entry *syme)
 263 {
 264         struct symbol *symbol;
 265         struct source_line *line;
 266         struct source_line *line_queue = NULL;
 267         int displayed = 0;
 268         int line_queue_count = 0, total = 0, more = 0;
 269
 270         if (!syme)
 271                 return;
 272
 273         if (!syme->source)
 274                 lookup_sym_source(syme);
 275
 276         if (!syme->source)
 277                 return;
 278
 279         symbol = (struct symbol *)(syme + 1);
 280         printf("Showing %s for %s\n", event_name(sym_counter), symbol->name);
 281         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
 282
 283         pthread_mutex_lock(&syme->source_lock);
 284         line = syme->source;
 285         while (line) {
 286                 total += line->count[sym_counter];
 287                 line = line->next;
 288         }
 289
 290         line = syme->source;
 291         while (line) {
 292                 float pcnt = 0.0;
 293
 294                 if (!line_queue_count)
 295                         line_queue = line;
 296                 line_queue_count++;
 297
 298                 if (line->count[sym_counter])
 299                         pcnt = 100.0 * line->count[sym_counter] / (float)total;
 300                 if (pcnt >= (float)sym_pcnt_filter) {
 301                         if (displayed <= print_entries)
 302                                 show_lines(line_queue, line_queue_count, total);
 303                         else more++;
 304                         displayed += line_queue_count;
 305                         line_queue_count = 0;
 306                         line_queue = NULL;
 307                 } else if (line_queue_count > TRACE_COUNT) {
 308                         line_queue = line_queue->next;
 309                         line_queue_count--;
 310                 }
 311
 312                 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
 313                 line = line->next;
 314         }
 315         pthread_mutex_unlock(&syme->source_lock);
 316         if (more)
 317                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 318 }
 319
 320 /*
 321  * Symbols will be added here in event__process_sample and will get out
 322  * after decayed.
 323  */
 324 static LIST_HEAD(active_symbols);
 325 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
 326
 327 /*
 328  * Ordering weight: count-1 * count-2 * ... / count-n
 329  */
 330 static double sym_weight(const struct sym_entry *sym)
 331 {
 332         double weight = sym->snap_count;
 333         int counter;
 334
 335         if (!display_weighted)
 336                 return weight;
 337
 338         for (counter = 1; counter < nr_counters-1; counter++)
 339                 weight *= sym->count[counter];
 340
 341         weight /= (sym->count[counter] + 1);
 342
 343         return weight;
 344 }
 345
 346 static long                     samples;
 347 static long                     userspace_samples;
 348 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
 349
 350 static void __list_insert_active_sym(struct sym_entry *syme)
 351 {
 352         list_add(&syme->node, &active_symbols);
 353 }
 354
 355 static void list_remove_active_sym(struct sym_entry *syme)
 356 {
 357         pthread_mutex_lock(&active_symbols_lock);
 358         list_del_init(&syme->node);
 359         pthread_mutex_unlock(&active_symbols_lock);
 360 }
 361
 362 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
 363 {
 364         struct rb_node **p = &tree->rb_node;
 365         struct rb_node *parent = NULL;
 366         struct sym_entry *iter;
 367
 368         while (*p != NULL) {
 369                 parent = *p;
 370                 iter = rb_entry(parent, struct sym_entry, rb_node);
 371
 372                 if (se->weight > iter->weight)
 373                         p = &(*p)->rb_left;
 374                 else
 375                         p = &(*p)->rb_right;
 376         }
 377
 378         rb_link_node(&se->rb_node, parent, p);
 379         rb_insert_color(&se->rb_node, tree);
 380 }
 381
 382 static void print_sym_table(void)
 383 {
 384         int printed = 0, j;
 385         int counter, snap = !display_weighted ? sym_counter : 0;
 386         float samples_per_sec = samples/delay_secs;
 387         float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
 388         float sum_ksamples = 0.0;
 389         struct sym_entry *syme, *n;
 390         struct rb_root tmp = RB_ROOT;
 391         struct rb_node *nd;
 392
 393         samples = userspace_samples = 0;
 394
 395         /* Sort the active symbols */
 396         pthread_mutex_lock(&active_symbols_lock);
 397         syme = list_entry(active_symbols.next, struct sym_entry, node);
 398         pthread_mutex_unlock(&active_symbols_lock);
 399
 400         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
 401                 syme->snap_count = syme->count[snap];
 402                 if (syme->snap_count != 0) {
 403                         syme->weight = sym_weight(syme);
 404                         rb_insert_active_sym(&tmp, syme);
 405                         sum_ksamples += syme->snap_count;
 406
 407                         for (j = 0; j < nr_counters; j++)
 408                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
 409                 } else
 410                         list_remove_active_sym(syme);
 411         }
 412
 413         puts(CONSOLE_CLEAR);
 414
 415         printf(
 416 "------------------------------------------------------------------------------\n");
 417         printf( "   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% [",
 418                 samples_per_sec,
 419                 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
 420
 421         if (nr_counters == 1 || !display_weighted) {
 422                 printf("%Ld", (u64)attrs[0].sample_period);
 423                 if (freq)
 424                         printf("Hz ");
 425                 else
 426                         printf(" ");
 427         }
 428
 429         if (!display_weighted)
 430                 printf("%s", event_name(sym_counter));
 431         else for (counter = 0; counter < nr_counters; counter++) {
 432                 if (counter)
 433                         printf("/");
 434
 435                 printf("%s", event_name(counter));
 436         }
 437
 438         printf( "], ");
 439
 440         if (target_pid != -1)
 441                 printf(" (target_pid: %d", target_pid);
 442         else
 443                 printf(" (all");
 444
 445         if (profile_cpu != -1)
 446                 printf(", cpu: %d)\n", profile_cpu);
 447         else {
 448                 if (target_pid != -1)
 449                         printf(")\n");
 450                 else
 451                         printf(", %d CPUs)\n", nr_cpus);
 452         }
 453
 454         printf("------------------------------------------------------------------------------\n\n");
 455
 456         if (sym_filter_entry) {
 457                 show_details(sym_filter_entry);
 458                 return;
 459         }
 460
 461         if (nr_counters == 1)
 462                 printf("             samples  pcnt");
 463         else
 464                 printf("   weight    samples  pcnt");
 465
 466         if (verbose)
 467                 printf("         RIP       ");
 468         printf(" function                                 DSO\n");
 469         printf("   %s    _______ _____",
 470                nr_counters == 1 ? "      " : "______");
 471         if (verbose)
 472                 printf(" ________________");
 473         printf(" ________________________________ ________________\n\n");
 474
 475         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
 476                 struct symbol *sym;
 477                 double pcnt;
 478
 479                 syme = rb_entry(nd, struct sym_entry, rb_node);
 480                 sym = (struct symbol *)(syme + 1);
 481
 482                 if (++printed > print_entries || (int)syme->snap_count < count_filter)
 483                         continue;
 484
 485                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
 486                                          sum_ksamples));
 487
 488                 if (nr_counters == 1 || !display_weighted)
 489                         printf("%20.2f ", syme->weight);
 490                 else
 491                         printf("%9.1f %10ld ", syme->weight, syme->snap_count);
 492
 493                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
 494                 if (verbose)
 495                         printf(" %016llx", sym->start);
 496                 printf(" %-32s", sym->name);
 497                 printf(" %s", syme->map->dso->short_name);
 498                 printf("\n");
 499         }
 500 }
 501
 502 static void prompt_integer(int *target, const char *msg)
 503 {
 504         char *buf = malloc(0), *p;
 505         size_t dummy = 0;
 506         int tmp;
 507
 508         fprintf(stdout, "\n%s: ", msg);
 509         if (getline(&buf, &dummy, stdin) < 0)
 510                 return;
 511
 512         p = strchr(buf, '\n');
 513         if (p)
 514                 *p = 0;
 515
 516         p = buf;
 517         while(*p) {
 518                 if (!isdigit(*p))
 519                         goto out_free;
 520                 p++;
 521         }
 522         tmp = strtoul(buf, NULL, 10);
 523         *target = tmp;
 524 out_free:
 525         free(buf);
 526 }
 527
 528 static void prompt_percent(int *target, const char *msg)
 529 {
 530         int tmp = 0;
 531
 532         prompt_integer(&tmp, msg);
 533         if (tmp >= 0 && tmp <= 100)
 534                 *target = tmp;
 535 }
 536
 537 static void prompt_symbol(struct sym_entry **target, const char *msg)
 538 {
 539         char *buf = malloc(0), *p;
 540         struct sym_entry *syme = *target, *n, *found = NULL;
 541         size_t dummy = 0;
 542
 543         /* zero counters of active symbol */
 544         if (syme) {
 545                 pthread_mutex_lock(&syme->source_lock);
 546                 __zero_source_counters(syme);
 547                 *target = NULL;
 548                 pthread_mutex_unlock(&syme->source_lock);
 549         }
 550
 551         fprintf(stdout, "\n%s: ", msg);
 552         if (getline(&buf, &dummy, stdin) < 0)
 553                 goto out_free;
 554
 555         p = strchr(buf, '\n');
 556         if (p)
 557                 *p = 0;
 558
 559         pthread_mutex_lock(&active_symbols_lock);
 560         syme = list_entry(active_symbols.next, struct sym_entry, node);
 561         pthread_mutex_unlock(&active_symbols_lock);
 562
 563         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
 564                 struct symbol *sym = (struct symbol *)(syme + 1);
 565
 566                 if (!strcmp(buf, sym->name)) {
 567                         found = syme;
 568                         break;
 569                 }
 570         }
 571
 572         if (!found) {
 573                 fprintf(stderr, "Sorry, %s is not active.\n", sym_filter);
 574                 sleep(1);
 575                 return;
 576         } else
 577                 parse_source(found);
 578
 579 out_free:
 580         free(buf);
 581 }
 582
 583 static void print_mapped_keys(void)
 584 {
 585         char *name = NULL;
 586
 587         if (sym_filter_entry) {
 588                 struct symbol *sym = (struct symbol *)(sym_filter_entry+1);
 589                 name = sym->name;
 590         }
 591
 592         fprintf(stdout, "\nMapped keys:\n");
 593         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", delay_secs);
 594         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", print_entries);
 595
 596         if (nr_counters > 1)
 597                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(sym_counter));
 598
 599         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", count_filter);
 600
 601         if (vmlinux_name) {
 602                 fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
 603                 fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 604                 fprintf(stdout, "\t[S]     stop annotation.\n");
 605         }
 606
 607         if (nr_counters > 1)
 608                 fprintf(stdout, "\t[w]     toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
 609
 610         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", zero ? 1 : 0);
 611         fprintf(stdout, "\t[qQ]    quit.\n");
 612 }
 613
 614 static int key_mapped(int c)
 615 {
 616         switch (c) {
 617                 case 'd':
 618                 case 'e':
 619                 case 'f':
 620                 case 'z':
 621                 case 'q':
 622                 case 'Q':
 623                         return 1;
 624                 case 'E':
 625                 case 'w':
 626                         return nr_counters > 1 ? 1 : 0;
 627                 case 'F':
 628                 case 's':
 629                 case 'S':
 630                         return vmlinux_name ? 1 : 0;
 631                 default:
 632                         break;
 633         }
 634
 635         return 0;
 636 }
 637
 638 static void handle_keypress(int c)
 639 {
 640         if (!key_mapped(c)) {
 641                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 642                 struct termios tc, save;
 643
 644                 print_mapped_keys();
 645                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 646                 fflush(stdout);
 647
 648                 tcgetattr(0, &save);
 649                 tc = save;
 650                 tc.c_lflag &= ~(ICANON | ECHO);
 651                 tc.c_cc[VMIN] = 0;
 652                 tc.c_cc[VTIME] = 0;
 653                 tcsetattr(0, TCSANOW, &tc);
 654
 655                 poll(&stdin_poll, 1, -1);
 656                 c = getc(stdin);
 657
 658                 tcsetattr(0, TCSAFLUSH, &save);
 659                 if (!key_mapped(c))
 660                         return;
 661         }
 662
 663         switch (c) {
 664                 case 'd':
 665                         prompt_integer(&delay_secs, "Enter display delay");
 666                         if (delay_secs < 1)
 667                                 delay_secs = 1;
 668                         break;
 669                 case 'e':
 670                         prompt_integer(&print_entries, "Enter display entries (lines)");
 671                         break;
 672                 case 'E':
 673                         if (nr_counters > 1) {
 674                                 int i;
 675
 676                                 fprintf(stderr, "\nAvailable events:");
 677                                 for (i = 0; i < nr_counters; i++)
 678                                         fprintf(stderr, "\n\t%d %s", i, event_name(i));
 679
 680                                 prompt_integer(&sym_counter, "Enter details event counter");
 681
 682                                 if (sym_counter >= nr_counters) {
 683                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
 684                                         sym_counter = 0;
 685                                         sleep(1);
 686                                 }
 687                         } else sym_counter = 0;
 688                         break;
 689                 case 'f':
 690                         prompt_integer(&count_filter, "Enter display event count filter");
 691                         break;
 692                 case 'F':
 693                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
 694                         break;
 695                 case 'q':
 696                 case 'Q':
 697                         printf("exiting.\n");
 698                         exit(0);
 699                 case 's':
 700                         prompt_symbol(&sym_filter_entry, "Enter details symbol");
 701                         break;
 702                 case 'S':
 703                         if (!sym_filter_entry)
 704                                 break;
 705                         else {
 706                                 struct sym_entry *syme = sym_filter_entry;
 707
 708                                 pthread_mutex_lock(&syme->source_lock);
 709                                 sym_filter_entry = NULL;
 710                                 __zero_source_counters(syme);
 711                                 pthread_mutex_unlock(&syme->source_lock);
 712                         }
 713                         break;
 714                 case 'w':
 715                         display_weighted = ~display_weighted;
 716                         break;
 717                 case 'z':
 718                         zero = ~zero;
 719                         break;
 720                 default:
 721                         break;
 722         }
 723 }
 724
 725 static void *display_thread(void *arg __used)
 726 {
 727         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 728         struct termios tc, save;
 729         int delay_msecs, c;
 730
 731         tcgetattr(0, &save);
 732         tc = save;
 733         tc.c_lflag &= ~(ICANON | ECHO);
 734         tc.c_cc[VMIN] = 0;
 735         tc.c_cc[VTIME] = 0;
 736
 737 repeat:
 738         delay_msecs = delay_secs * 1000;
 739         tcsetattr(0, TCSANOW, &tc);
 740         /* trash return*/
 741         getc(stdin);
 742
 743         do {
 744                 print_sym_table();
 745         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
 746
 747         c = getc(stdin);
 748         tcsetattr(0, TCSAFLUSH, &save);
 749
 750         handle_keypress(c);
 751         goto repeat;
 752
 753         return NULL;
 754 }
 755
 756 /* Tag samples to be skipped. */
 757 static const char *skip_symbols[] = {
 758         "default_idle",
 759         "cpu_idle",
 760         "enter_idle",
 761         "exit_idle",
 762         "mwait_idle",
 763         "mwait_idle_with_hints",
 764         "poll_idle",
 765         "ppc64_runlatch_off",
 766         "pseries_dedicated_idle_sleep",
 767         NULL
 768 };
 769
 770 static int symbol_filter(struct map *map, struct symbol *sym)
 771 {
 772         struct sym_entry *syme;
 773         const char *name = sym->name;
 774         int i;
 775
 776         /*
 777          * ppc64 uses function descriptors and appends a '.' to the
 778          * start of every instruction address. Remove it.
 779          */
 780         if (name[0] == '.')
 781                 name++;
 782
 783         if (!strcmp(name, "_text") ||
 784             !strcmp(name, "_etext") ||
 785             !strcmp(name, "_sinittext") ||
 786             !strncmp("init_module", name, 11) ||
 787             !strncmp("cleanup_module", name, 14) ||
 788             strstr(name, "_text_start") ||
 789             strstr(name, "_text_end"))
 790                 return 1;
 791
 792         syme = dso__sym_priv(map->dso, sym);
 793         syme->map = map;
 794         pthread_mutex_init(&syme->source_lock, NULL);
 795         if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter))
 796                 sym_filter_entry = syme;
 797
 798         for (i = 0; skip_symbols[i]; i++) {
 799                 if (!strcmp(skip_symbols[i], name)) {
 800                         syme->skip = 1;
 801                         break;
 802                 }
 803         }
 804
 805         return 0;
 806 }
 807
 808 static int parse_symbols(void)
 809 {
 810         if (dsos__load_kernel(vmlinux_name, sizeof(struct sym_entry),
 811                               symbol_filter, 1) <= 0)
 812                 return -1;
 813
 814         if (dump_symtab)
 815                 dsos__fprintf(stderr);
 816
 817         return 0;
 818 }
 819
 820 static void event__process_sample(const event_t *self, int counter)
 821 {
 822         u64 ip = self->ip.ip;
 823         struct map *map;
 824         struct sym_entry *syme;
 825         struct symbol *sym;
 826
 827         switch (self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK) {
 828         case PERF_RECORD_MISC_USER: {
 829                 struct thread *thread = threads__findnew(self->ip.pid);
 830
 831                 if (thread == NULL)
 832                         return;
 833
 834                 map = thread__find_map(thread, ip);
 835                 if (map != NULL) {
 836                         ip = map->map_ip(map, ip);
 837                         sym = map->dso->find_symbol(map->dso, ip);
 838                         if (sym == NULL)
 839                                 return;
 840                         userspace_samples++;
 841                         break;
 842                 }
 843         }
 844                 /*
 845                  * If this is outside of all known maps,
 846                  * and is a negative address, try to look it
 847                  * up in the kernel dso, as it might be a
 848                  * vsyscall or vdso (which executes in user-mode).
 849                  */
 850                 if ((long long)ip >= 0)
 851                         return;
 852                 /* Fall thru */
 853         case PERF_RECORD_MISC_KERNEL:
 854                 sym = kernel_maps__find_symbol(ip, &map);
 855                 if (sym == NULL)
 856                         return;
 857                 break;
 858         default:
 859                 return;
 860         }
 861
 862         syme = dso__sym_priv(map->dso, sym);
 863
 864         if (!syme->skip) {
 865                 syme->count[counter]++;
 866                 record_precise_ip(syme, counter, ip);
 867                 pthread_mutex_lock(&active_symbols_lock);
 868                 if (list_empty(&syme->node) || !syme->node.next)
 869                         __list_insert_active_sym(syme);
 870                 pthread_mutex_unlock(&active_symbols_lock);
 871                 ++samples;
 872                 return;
 873         }
 874 }
 875
 876 static void event__process_mmap(event_t *self)
 877 {
 878         struct thread *thread = threads__findnew(self->mmap.pid);
 879
 880         if (thread != NULL) {
 881                 struct map *map = map__new(&self->mmap, NULL, 0,
 882                                            sizeof(struct sym_entry),
 883                                            symbol_filter);
 884                 if (map != NULL)
 885                         thread__insert_map(thread, map);
 886         }
 887 }
 888
 889 static void event__process_comm(event_t *self)
 890 {
 891         struct thread *thread = threads__findnew(self->comm.pid);
 892
 893         if (thread != NULL)
 894                 thread__set_comm(thread, self->comm.comm);
 895 }
 896
 897 static int event__process(event_t *event)
 898 {
 899         switch (event->header.type) {
 900         case PERF_RECORD_COMM:
 901                 event__process_comm(event);
 902                 break;
 903         case PERF_RECORD_MMAP:
 904                 event__process_mmap(event);
 905                 break;
 906         default:
 907                 break;
 908         }
 909
 910         return 0;
 911 }
 912
 913 struct mmap_data {
 914         int                     counter;
 915         void                    *base;
 916         int                     mask;
 917         unsigned int            prev;
 918 };
 919
 920 static unsigned int mmap_read_head(struct mmap_data *md)
 921 {
 922         struct perf_event_mmap_page *pc = md->base;
 923         int head;
 924
 925         head = pc->data_head;
 926         rmb();
 927
 928         return head;
 929 }
 930
 931 static void mmap_read_counter(struct mmap_data *md)
 932 {
 933         unsigned int head = mmap_read_head(md);
 934         unsigned int old = md->prev;
 935         unsigned char *data = md->base + page_size;
 936         int diff;
 937
 938         /*
 939          * If we're further behind than half the buffer, there's a chance
 940          * the writer will bite our tail and mess up the samples under us.
 941          *
 942          * If we somehow ended up ahead of the head, we got messed up.
 943          *
 944          * In either case, truncate and restart at head.
 945          */
 946         diff = head - old;
 947         if (diff > md->mask / 2 || diff < 0) {
 948                 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
 949
 950                 /*
 951                  * head points to a known good entry, start there.
 952                  */
 953                 old = head;
 954         }
 955
 956         for (; old != head;) {
 957                 event_t *event = (event_t *)&data[old & md->mask];
 958
 959                 event_t event_copy;
 960
 961                 size_t size = event->header.size;
 962
 963                 /*
 964                  * Event straddles the mmap boundary -- header should always
 965                  * be inside due to u64 alignment of output.
 966                  */
 967                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
 968                         unsigned int offset = old;
 969                         unsigned int len = min(sizeof(*event), size), cpy;
 970                         void *dst = &event_copy;
 971
 972                         do {
 973                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
 974                                 memcpy(dst, &data[offset & md->mask], cpy);
 975                                 offset += cpy;
 976                                 dst += cpy;
 977                                 len -= cpy;
 978                         } while (len);
 979
 980                         event = &event_copy;
 981                 }
 982
 983                 if (event->header.type == PERF_RECORD_SAMPLE)
 984                         event__process_sample(event, md->counter);
 985                 else
 986                         event__process(event);
 987                 old += size;
 988         }
 989
 990         md->prev = old;
 991 }
 992
 993 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
 994 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
 995
 996 static void mmap_read(void)
 997 {
 998         int i, counter;
 999
1000         for (i = 0; i < nr_cpus; i++) {
1001                 for (counter = 0; counter < nr_counters; counter++)
1002                         mmap_read_counter(&mmap_array[i][counter]);
1003         }
1004 }
1005
1006 int nr_poll;
1007 int group_fd;
1008
1009 static void start_counter(int i, int counter)
1010 {
1011         struct perf_event_attr *attr;
1012         int cpu;
1013
1014         cpu = profile_cpu;
1015         if (target_pid == -1 && profile_cpu == -1)
1016                 cpu = i;
1017
1018         attr = attrs + counter;
1019
1020         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1021
1022         if (freq) {
1023                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
1024                 attr->freq              = 1;
1025                 attr->sample_freq       = freq;
1026         }
1027
1028         attr->inherit           = (cpu < 0) && inherit;
1029         attr->mmap              = 1;
1030
1031 try_again:
1032         fd[i][counter] = sys_perf_event_open(attr, target_pid, cpu, group_fd, 0);
1033
1034         if (fd[i][counter] < 0) {
1035                 int err = errno;
1036
1037                 if (err == EPERM)
1038                         die("No permission - are you root?\n");
1039                 /*
1040                  * If it's cycles then fall back to hrtimer
1041                  * based cpu-clock-tick sw counter, which
1042                  * is always available even if no PMU support:
1043                  */
1044                 if (attr->type == PERF_TYPE_HARDWARE
1045                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1046
1047                         if (verbose)
1048                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
1049
1050                         attr->type = PERF_TYPE_SOFTWARE;
1051                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
1052                         goto try_again;
1053                 }
1054                 printf("\n");
1055                 error("perfcounter syscall returned with %d (%s)\n",
1056                         fd[i][counter], strerror(err));
1057                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1058                 exit(-1);
1059         }
1060         assert(fd[i][counter] >= 0);
1061         fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
1062
1063         /*
1064          * First counter acts as the group leader:
1065          */
1066         if (group && group_fd == -1)
1067                 group_fd = fd[i][counter];
1068
1069         event_array[nr_poll].fd = fd[i][counter];
1070         event_array[nr_poll].events = POLLIN;
1071         nr_poll++;
1072
1073         mmap_array[i][counter].counter = counter;
1074         mmap_array[i][counter].prev = 0;
1075         mmap_array[i][counter].mask = mmap_pages*page_size - 1;
1076         mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
1077                         PROT_READ, MAP_SHARED, fd[i][counter], 0);
1078         if (mmap_array[i][counter].base == MAP_FAILED)
1079                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1080 }
1081
1082 static int __cmd_top(void)
1083 {
1084         pthread_t thread;
1085         int i, counter;
1086         int ret;
1087
1088         if (target_pid != -1)
1089                 event__synthesize_thread(target_pid, event__process);
1090         else
1091                 event__synthesize_threads(event__process);
1092
1093         for (i = 0; i < nr_cpus; i++) {
1094                 group_fd = -1;
1095                 for (counter = 0; counter < nr_counters; counter++)
1096                         start_counter(i, counter);
1097         }
1098
1099         /* Wait for a minimal set of events before starting the snapshot */
1100         poll(event_array, nr_poll, 100);
1101
1102         mmap_read();
1103
1104         if (pthread_create(&thread, NULL, display_thread, NULL)) {
1105                 printf("Could not create display thread.\n");
1106                 exit(-1);
1107         }
1108
1109         if (realtime_prio) {
1110                 struct sched_param param;
1111
1112                 param.sched_priority = realtime_prio;
1113                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1114                         printf("Could not set realtime priority.\n");
1115                         exit(-1);
1116                 }
1117         }
1118
1119         while (1) {
1120                 int hits = samples;
1121
1122                 mmap_read();
1123
1124                 if (hits == samples)
1125                         ret = poll(event_array, nr_poll, 100);
1126         }
1127
1128         return 0;
1129 }
1130
1131 static const char * const top_usage[] = {
1132         "perf top [<options>]",
1133         NULL
1134 };
1135
1136 static const struct option options[] = {
1137         OPT_CALLBACK('e', "event", NULL, "event",
1138                      "event selector. use 'perf list' to list available events",
1139                      parse_events),
1140         OPT_INTEGER('c', "count", &default_interval,
1141                     "event period to sample"),
1142         OPT_INTEGER('p', "pid", &target_pid,
1143                     "profile events on existing pid"),
1144         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1145                             "system-wide collection from all CPUs"),
1146         OPT_INTEGER('C', "CPU", &profile_cpu,
1147                     "CPU to profile on"),
1148         OPT_STRING('k', "vmlinux", &vmlinux_name, "file", "vmlinux pathname"),
1149         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
1150                     "number of mmap data pages"),
1151         OPT_INTEGER('r', "realtime", &realtime_prio,
1152                     "collect data with this RT SCHED_FIFO priority"),
1153         OPT_INTEGER('d', "delay", &delay_secs,
1154                     "number of seconds to delay between refreshes"),
1155         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1156                             "dump the symbol table used for profiling"),
1157         OPT_INTEGER('f', "count-filter", &count_filter,
1158                     "only display functions with more events than this"),
1159         OPT_BOOLEAN('g', "group", &group,
1160                             "put the counters into a counter group"),
1161         OPT_BOOLEAN('i', "inherit", &inherit,
1162                     "child tasks inherit counters"),
1163         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1164                     "symbol to annotate - requires -k option"),
1165         OPT_BOOLEAN('z', "zero", &zero,
1166                     "zero history across updates"),
1167         OPT_INTEGER('F', "freq", &freq,
1168                     "profile at this frequency"),
1169         OPT_INTEGER('E', "entries", &print_entries,
1170                     "display this many functions"),
1171         OPT_BOOLEAN('v', "verbose", &verbose,
1172                     "be more verbose (show counter open errors, etc)"),
1173         OPT_END()
1174 };
1175
1176 int cmd_top(int argc, const char **argv, const char *prefix __used)
1177 {
1178         int counter;
1179
1180         symbol__init();
1181
1182         page_size = sysconf(_SC_PAGE_SIZE);
1183
1184         argc = parse_options(argc, argv, options, top_usage, 0);
1185         if (argc)
1186                 usage_with_options(top_usage, options);
1187
1188         /* CPU and PID are mutually exclusive */
1189         if (target_pid != -1 && profile_cpu != -1) {
1190                 printf("WARNING: PID switch overriding CPU\n");
1191                 sleep(1);
1192                 profile_cpu = -1;
1193         }
1194
1195         if (!nr_counters)
1196                 nr_counters = 1;
1197
1198         if (delay_secs < 1)
1199                 delay_secs = 1;
1200
1201         parse_symbols();
1202         parse_source(sym_filter_entry);
1203
1204
1205         /*
1206          * User specified count overrides default frequency.
1207          */
1208         if (default_interval)
1209                 freq = 0;
1210         else if (freq) {
1211                 default_interval = freq;
1212         } else {
1213                 fprintf(stderr, "frequency and count are zero, aborting\n");
1214                 exit(EXIT_FAILURE);
1215         }
1216
1217         /*
1218          * Fill in the ones not specifically initialized via -c:
1219          */
1220         for (counter = 0; counter < nr_counters; counter++) {
1221                 if (attrs[counter].sample_period)
1222                         continue;
1223
1224                 attrs[counter].sample_period = default_interval;
1225         }
1226
1227         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1228         assert(nr_cpus <= MAX_NR_CPUS);
1229         assert(nr_cpus >= 0);
1230
1231         if (target_pid != -1 || profile_cpu != -1)
1232                 nr_cpus = 1;
1233
1234         return __cmd_top();
1235 }