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;
  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 static struct winsize           winsize;
  82 struct symbol_conf              symbol_conf;
  83
  84 /*
  85  * Source
  86  */
  87
  88 struct source_line {
  89         u64                     eip;
  90         unsigned long           count[MAX_COUNTERS];
  91         char                    *line;
  92         struct source_line      *next;
  93 };
  94
  95 static char                     *sym_filter                     =   NULL;
  96 struct sym_entry                *sym_filter_entry               =   NULL;
  97 static int                      sym_pcnt_filter                 =      5;
  98 static int                      sym_counter                     =      0;
  99 static int                      display_weighted                =     -1;
 100
 101 /*
 102  * Symbols
 103  */
 104
 105 struct sym_entry_source {
 106         struct source_line      *source;
 107         struct source_line      *lines;
 108         struct source_line      **lines_tail;
 109         pthread_mutex_t         lock;
 110 };
 111
 112 struct sym_entry {
 113         struct rb_node          rb_node;
 114         struct list_head        node;
 115         unsigned long           snap_count;
 116         double                  weight;
 117         int                     skip;
 118         u16                     name_len;
 119         u8                      origin;
 120         struct map              *map;
 121         struct sym_entry_source *src;
 122         unsigned long           count[0];
 123 };
 124
 125 /*
 126  * Source functions
 127  */
 128
 129 static inline struct symbol *sym_entry__symbol(struct sym_entry *self)
 130 {
 131        return ((void *)self) + symbol_conf.priv_size;
 132 }
 133
 134 static void get_term_dimensions(struct winsize *ws)
 135 {
 136         char *s = getenv("LINES");
 137
 138         if (s != NULL) {
 139                 ws->ws_row = atoi(s);
 140                 s = getenv("COLUMNS");
 141                 if (s != NULL) {
 142                         ws->ws_col = atoi(s);
 143                         if (ws->ws_row && ws->ws_col)
 144                                 return;
 145                 }
 146         }
 147 #ifdef TIOCGWINSZ
 148         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
 149             ws->ws_row && ws->ws_col)
 150                 return;
 151 #endif
 152         ws->ws_row = 25;
 153         ws->ws_col = 80;
 154 }
 155
 156 static void update_print_entries(struct winsize *ws)
 157 {
 158         print_entries = ws->ws_row;
 159
 160         if (print_entries > 9)
 161                 print_entries -= 9;
 162 }
 163
 164 static void sig_winch_handler(int sig __used)
 165 {
 166         get_term_dimensions(&winsize);
 167         update_print_entries(&winsize);
 168 }
 169
 170 static void parse_source(struct sym_entry *syme)
 171 {
 172         struct symbol *sym;
 173         struct sym_entry_source *source;
 174         struct map *map;
 175         FILE *file;
 176         char command[PATH_MAX*2];
 177         const char *path;
 178         u64 len;
 179
 180         if (!syme)
 181                 return;
 182
 183         if (syme->src == NULL) {
 184                 syme->src = zalloc(sizeof(*source));
 185                 if (syme->src == NULL)
 186                         return;
 187                 pthread_mutex_init(&syme->src->lock, NULL);
 188         }
 189
 190         source = syme->src;
 191
 192         if (source->lines) {
 193                 pthread_mutex_lock(&source->lock);
 194                 goto out_assign;
 195         }
 196
 197         sym = sym_entry__symbol(syme);
 198         map = syme->map;
 199         path = map->dso->long_name;
 200
 201         len = sym->end - sym->start;
 202
 203         sprintf(command,
 204                 "objdump --start-address=0x%016Lx "
 205                          "--stop-address=0x%016Lx -dS %s",
 206                 map->unmap_ip(map, sym->start),
 207                 map->unmap_ip(map, sym->end), path);
 208
 209         file = popen(command, "r");
 210         if (!file)
 211                 return;
 212
 213         pthread_mutex_lock(&source->lock);
 214         source->lines_tail = &source->lines;
 215         while (!feof(file)) {
 216                 struct source_line *src;
 217                 size_t dummy = 0;
 218                 char *c;
 219
 220                 src = malloc(sizeof(struct source_line));
 221                 assert(src != NULL);
 222                 memset(src, 0, sizeof(struct source_line));
 223
 224                 if (getline(&src->line, &dummy, file) < 0)
 225                         break;
 226                 if (!src->line)
 227                         break;
 228
 229                 c = strchr(src->line, '\n');
 230                 if (c)
 231                         *c = 0;
 232
 233                 src->next = NULL;
 234                 *source->lines_tail = src;
 235                 source->lines_tail = &src->next;
 236
 237                 if (strlen(src->line)>8 && src->line[8] == ':') {
 238                         src->eip = strtoull(src->line, NULL, 16);
 239                         src->eip = map->unmap_ip(map, src->eip);
 240                 }
 241                 if (strlen(src->line)>8 && src->line[16] == ':') {
 242                         src->eip = strtoull(src->line, NULL, 16);
 243                         src->eip = map->unmap_ip(map, src->eip);
 244                 }
 245         }
 246         pclose(file);
 247 out_assign:
 248         sym_filter_entry = syme;
 249         pthread_mutex_unlock(&source->lock);
 250 }
 251
 252 static void __zero_source_counters(struct sym_entry *syme)
 253 {
 254         int i;
 255         struct source_line *line;
 256
 257         line = syme->src->lines;
 258         while (line) {
 259                 for (i = 0; i < nr_counters; i++)
 260                         line->count[i] = 0;
 261                 line = line->next;
 262         }
 263 }
 264
 265 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
 266 {
 267         struct source_line *line;
 268
 269         if (syme != sym_filter_entry)
 270                 return;
 271
 272         if (pthread_mutex_trylock(&syme->src->lock))
 273                 return;
 274
 275         if (syme->src == NULL || syme->src->source == NULL)
 276                 goto out_unlock;
 277
 278         for (line = syme->src->lines; line; line = line->next) {
 279                 if (line->eip == ip) {
 280                         line->count[counter]++;
 281                         break;
 282                 }
 283                 if (line->eip > ip)
 284                         break;
 285         }
 286 out_unlock:
 287         pthread_mutex_unlock(&syme->src->lock);
 288 }
 289
 290 static void lookup_sym_source(struct sym_entry *syme)
 291 {
 292         struct symbol *symbol = sym_entry__symbol(syme);
 293         struct source_line *line;
 294         char pattern[PATH_MAX];
 295
 296         sprintf(pattern, "<%s>:", symbol->name);
 297
 298         pthread_mutex_lock(&syme->src->lock);
 299         for (line = syme->src->lines; line; line = line->next) {
 300                 if (strstr(line->line, pattern)) {
 301                         syme->src->source = line;
 302                         break;
 303                 }
 304         }
 305         pthread_mutex_unlock(&syme->src->lock);
 306 }
 307
 308 static void show_lines(struct source_line *queue, int count, int total)
 309 {
 310         int i;
 311         struct source_line *line;
 312
 313         line = queue;
 314         for (i = 0; i < count; i++) {
 315                 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
 316
 317                 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
 318                 line = line->next;
 319         }
 320 }
 321
 322 #define TRACE_COUNT     3
 323
 324 static void show_details(struct sym_entry *syme)
 325 {
 326         struct symbol *symbol;
 327         struct source_line *line;
 328         struct source_line *line_queue = NULL;
 329         int displayed = 0;
 330         int line_queue_count = 0, total = 0, more = 0;
 331
 332         if (!syme)
 333                 return;
 334
 335         if (!syme->src->source)
 336                 lookup_sym_source(syme);
 337
 338         if (!syme->src->source)
 339                 return;
 340
 341         symbol = sym_entry__symbol(syme);
 342         printf("Showing %s for %s\n", event_name(sym_counter), symbol->name);
 343         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
 344
 345         pthread_mutex_lock(&syme->src->lock);
 346         line = syme->src->source;
 347         while (line) {
 348                 total += line->count[sym_counter];
 349                 line = line->next;
 350         }
 351
 352         line = syme->src->source;
 353         while (line) {
 354                 float pcnt = 0.0;
 355
 356                 if (!line_queue_count)
 357                         line_queue = line;
 358                 line_queue_count++;
 359
 360                 if (line->count[sym_counter])
 361                         pcnt = 100.0 * line->count[sym_counter] / (float)total;
 362                 if (pcnt >= (float)sym_pcnt_filter) {
 363                         if (displayed <= print_entries)
 364                                 show_lines(line_queue, line_queue_count, total);
 365                         else more++;
 366                         displayed += line_queue_count;
 367                         line_queue_count = 0;
 368                         line_queue = NULL;
 369                 } else if (line_queue_count > TRACE_COUNT) {
 370                         line_queue = line_queue->next;
 371                         line_queue_count--;
 372                 }
 373
 374                 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
 375                 line = line->next;
 376         }
 377         pthread_mutex_unlock(&syme->src->lock);
 378         if (more)
 379                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 380 }
 381
 382 /*
 383  * Symbols will be added here in event__process_sample and will get out
 384  * after decayed.
 385  */
 386 static LIST_HEAD(active_symbols);
 387 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
 388
 389 /*
 390  * Ordering weight: count-1 * count-2 * ... / count-n
 391  */
 392 static double sym_weight(const struct sym_entry *sym)
 393 {
 394         double weight = sym->snap_count;
 395         int counter;
 396
 397         if (!display_weighted)
 398                 return weight;
 399
 400         for (counter = 1; counter < nr_counters-1; counter++)
 401                 weight *= sym->count[counter];
 402
 403         weight /= (sym->count[counter] + 1);
 404
 405         return weight;
 406 }
 407
 408 static long                     samples;
 409 static long                     userspace_samples;
 410 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
 411
 412 static void __list_insert_active_sym(struct sym_entry *syme)
 413 {
 414         list_add(&syme->node, &active_symbols);
 415 }
 416
 417 static void list_remove_active_sym(struct sym_entry *syme)
 418 {
 419         pthread_mutex_lock(&active_symbols_lock);
 420         list_del_init(&syme->node);
 421         pthread_mutex_unlock(&active_symbols_lock);
 422 }
 423
 424 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
 425 {
 426         struct rb_node **p = &tree->rb_node;
 427         struct rb_node *parent = NULL;
 428         struct sym_entry *iter;
 429
 430         while (*p != NULL) {
 431                 parent = *p;
 432                 iter = rb_entry(parent, struct sym_entry, rb_node);
 433
 434                 if (se->weight > iter->weight)
 435                         p = &(*p)->rb_left;
 436                 else
 437                         p = &(*p)->rb_right;
 438         }
 439
 440         rb_link_node(&se->rb_node, parent, p);
 441         rb_insert_color(&se->rb_node, tree);
 442 }
 443
 444 static void print_sym_table(void)
 445 {
 446         int printed = 0, j;
 447         int counter, snap = !display_weighted ? sym_counter : 0;
 448         float samples_per_sec = samples/delay_secs;
 449         float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
 450         float sum_ksamples = 0.0;
 451         struct sym_entry *syme, *n;
 452         struct rb_root tmp = RB_ROOT;
 453         struct rb_node *nd;
 454         int sym_width = 0, dso_width = 0, max_dso_width;
 455         const int win_width = winsize.ws_col - 1;
 456
 457         samples = userspace_samples = 0;
 458
 459         /* Sort the active symbols */
 460         pthread_mutex_lock(&active_symbols_lock);
 461         syme = list_entry(active_symbols.next, struct sym_entry, node);
 462         pthread_mutex_unlock(&active_symbols_lock);
 463
 464         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
 465                 syme->snap_count = syme->count[snap];
 466                 if (syme->snap_count != 0) {
 467
 468                         if ((hide_user_symbols &&
 469                              syme->origin == PERF_RECORD_MISC_USER) ||
 470                             (hide_kernel_symbols &&
 471                              syme->origin == PERF_RECORD_MISC_KERNEL)) {
 472                                 list_remove_active_sym(syme);
 473                                 continue;
 474                         }
 475                         syme->weight = sym_weight(syme);
 476                         rb_insert_active_sym(&tmp, syme);
 477                         sum_ksamples += syme->snap_count;
 478
 479                         for (j = 0; j < nr_counters; j++)
 480                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
 481                 } else
 482                         list_remove_active_sym(syme);
 483         }
 484
 485         puts(CONSOLE_CLEAR);
 486
 487         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 488         printf( "   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% [",
 489                 samples_per_sec,
 490                 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
 491
 492         if (nr_counters == 1 || !display_weighted) {
 493                 printf("%Ld", (u64)attrs[0].sample_period);
 494                 if (freq)
 495                         printf("Hz ");
 496                 else
 497                         printf(" ");
 498         }
 499
 500         if (!display_weighted)
 501                 printf("%s", event_name(sym_counter));
 502         else for (counter = 0; counter < nr_counters; counter++) {
 503                 if (counter)
 504                         printf("/");
 505
 506                 printf("%s", event_name(counter));
 507         }
 508
 509         printf( "], ");
 510
 511         if (target_pid != -1)
 512                 printf(" (target_pid: %d", target_pid);
 513         else
 514                 printf(" (all");
 515
 516         if (profile_cpu != -1)
 517                 printf(", cpu: %d)\n", profile_cpu);
 518         else {
 519                 if (target_pid != -1)
 520                         printf(")\n");
 521                 else
 522                         printf(", %d CPUs)\n", nr_cpus);
 523         }
 524
 525         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 526
 527         if (sym_filter_entry) {
 528                 show_details(sym_filter_entry);
 529                 return;
 530         }
 531
 532         /*
 533          * Find the longest symbol name that will be displayed
 534          */
 535         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
 536                 syme = rb_entry(nd, struct sym_entry, rb_node);
 537                 if (++printed > print_entries ||
 538                     (int)syme->snap_count < count_filter)
 539                         continue;
 540
 541                 if (syme->map->dso->long_name_len > dso_width)
 542                         dso_width = syme->map->dso->long_name_len;
 543
 544                 if (syme->name_len > sym_width)
 545                         sym_width = syme->name_len;
 546         }
 547
 548         printed = 0;
 549
 550         max_dso_width = winsize.ws_col - sym_width - 29;
 551         if (dso_width > max_dso_width)
 552                 dso_width = max_dso_width;
 553         putchar('\n');
 554         if (nr_counters == 1)
 555                 printf("             samples  pcnt");
 556         else
 557                 printf("   weight    samples  pcnt");
 558
 559         if (verbose)
 560                 printf("         RIP       ");
 561         printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
 562         printf("   %s    _______ _____",
 563                nr_counters == 1 ? "      " : "______");
 564         if (verbose)
 565                 printf(" ________________");
 566         printf(" %-*.*s", sym_width, sym_width, graph_line);
 567         printf(" %-*.*s", dso_width, dso_width, graph_line);
 568         puts("\n");
 569
 570         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
 571                 struct symbol *sym;
 572                 double pcnt;
 573
 574                 syme = rb_entry(nd, struct sym_entry, rb_node);
 575                 sym = sym_entry__symbol(syme);
 576
 577                 if (++printed > print_entries || (int)syme->snap_count < count_filter)
 578                         continue;
 579
 580                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
 581                                          sum_ksamples));
 582
 583                 if (nr_counters == 1 || !display_weighted)
 584                         printf("%20.2f ", syme->weight);
 585                 else
 586                         printf("%9.1f %10ld ", syme->weight, syme->snap_count);
 587
 588                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
 589                 if (verbose)
 590                         printf(" %016llx", sym->start);
 591                 printf(" %-*.*s", sym_width, sym_width, sym->name);
 592                 printf(" %-*.*s\n", dso_width, dso_width,
 593                        dso_width >= syme->map->dso->long_name_len ?
 594                                         syme->map->dso->long_name :
 595                                         syme->map->dso->short_name);
 596         }
 597 }
 598
 599 static void prompt_integer(int *target, const char *msg)
 600 {
 601         char *buf = malloc(0), *p;
 602         size_t dummy = 0;
 603         int tmp;
 604
 605         fprintf(stdout, "\n%s: ", msg);
 606         if (getline(&buf, &dummy, stdin) < 0)
 607                 return;
 608
 609         p = strchr(buf, '\n');
 610         if (p)
 611                 *p = 0;
 612
 613         p = buf;
 614         while(*p) {
 615                 if (!isdigit(*p))
 616                         goto out_free;
 617                 p++;
 618         }
 619         tmp = strtoul(buf, NULL, 10);
 620         *target = tmp;
 621 out_free:
 622         free(buf);
 623 }
 624
 625 static void prompt_percent(int *target, const char *msg)
 626 {
 627         int tmp = 0;
 628
 629         prompt_integer(&tmp, msg);
 630         if (tmp >= 0 && tmp <= 100)
 631                 *target = tmp;
 632 }
 633
 634 static void prompt_symbol(struct sym_entry **target, const char *msg)
 635 {
 636         char *buf = malloc(0), *p;
 637         struct sym_entry *syme = *target, *n, *found = NULL;
 638         size_t dummy = 0;
 639
 640         /* zero counters of active symbol */
 641         if (syme) {
 642                 pthread_mutex_lock(&syme->src->lock);
 643                 __zero_source_counters(syme);
 644                 *target = NULL;
 645                 pthread_mutex_unlock(&syme->src->lock);
 646         }
 647
 648         fprintf(stdout, "\n%s: ", msg);
 649         if (getline(&buf, &dummy, stdin) < 0)
 650                 goto out_free;
 651
 652         p = strchr(buf, '\n');
 653         if (p)
 654                 *p = 0;
 655
 656         pthread_mutex_lock(&active_symbols_lock);
 657         syme = list_entry(active_symbols.next, struct sym_entry, node);
 658         pthread_mutex_unlock(&active_symbols_lock);
 659
 660         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
 661                 struct symbol *sym = sym_entry__symbol(syme);
 662
 663                 if (!strcmp(buf, sym->name)) {
 664                         found = syme;
 665                         break;
 666                 }
 667         }
 668
 669         if (!found) {
 670                 fprintf(stderr, "Sorry, %s is not active.\n", sym_filter);
 671                 sleep(1);
 672                 return;
 673         } else
 674                 parse_source(found);
 675
 676 out_free:
 677         free(buf);
 678 }
 679
 680 static void print_mapped_keys(void)
 681 {
 682         char *name = NULL;
 683
 684         if (sym_filter_entry) {
 685                 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
 686                 name = sym->name;
 687         }
 688
 689         fprintf(stdout, "\nMapped keys:\n");
 690         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", delay_secs);
 691         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", print_entries);
 692
 693         if (nr_counters > 1)
 694                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(sym_counter));
 695
 696         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", count_filter);
 697
 698         if (symbol_conf.vmlinux_name) {
 699                 fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
 700                 fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 701                 fprintf(stdout, "\t[S]     stop annotation.\n");
 702         }
 703
 704         if (nr_counters > 1)
 705                 fprintf(stdout, "\t[w]     toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
 706
 707         fprintf(stdout,
 708                 "\t[K]     hide kernel_symbols symbols.             \t(%s)\n",
 709                 hide_kernel_symbols ? "yes" : "no");
 710         fprintf(stdout,
 711                 "\t[U]     hide user symbols.               \t(%s)\n",
 712                 hide_user_symbols ? "yes" : "no");
 713         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", zero ? 1 : 0);
 714         fprintf(stdout, "\t[qQ]    quit.\n");
 715 }
 716
 717 static int key_mapped(int c)
 718 {
 719         switch (c) {
 720                 case 'd':
 721                 case 'e':
 722                 case 'f':
 723                 case 'z':
 724                 case 'q':
 725                 case 'Q':
 726                 case 'K':
 727                 case 'U':
 728                         return 1;
 729                 case 'E':
 730                 case 'w':
 731                         return nr_counters > 1 ? 1 : 0;
 732                 case 'F':
 733                 case 's':
 734                 case 'S':
 735                         return symbol_conf.vmlinux_name ? 1 : 0;
 736                 default:
 737                         break;
 738         }
 739
 740         return 0;
 741 }
 742
 743 static void handle_keypress(int c)
 744 {
 745         if (!key_mapped(c)) {
 746                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 747                 struct termios tc, save;
 748
 749                 print_mapped_keys();
 750                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 751                 fflush(stdout);
 752
 753                 tcgetattr(0, &save);
 754                 tc = save;
 755                 tc.c_lflag &= ~(ICANON | ECHO);
 756                 tc.c_cc[VMIN] = 0;
 757                 tc.c_cc[VTIME] = 0;
 758                 tcsetattr(0, TCSANOW, &tc);
 759
 760                 poll(&stdin_poll, 1, -1);
 761                 c = getc(stdin);
 762
 763                 tcsetattr(0, TCSAFLUSH, &save);
 764                 if (!key_mapped(c))
 765                         return;
 766         }
 767
 768         switch (c) {
 769                 case 'd':
 770                         prompt_integer(&delay_secs, "Enter display delay");
 771                         if (delay_secs < 1)
 772                                 delay_secs = 1;
 773                         break;
 774                 case 'e':
 775                         prompt_integer(&print_entries, "Enter display entries (lines)");
 776                         if (print_entries == 0) {
 777                                 sig_winch_handler(SIGWINCH);
 778                                 signal(SIGWINCH, sig_winch_handler);
 779                         } else
 780                                 signal(SIGWINCH, SIG_DFL);
 781                         break;
 782                 case 'E':
 783                         if (nr_counters > 1) {
 784                                 int i;
 785
 786                                 fprintf(stderr, "\nAvailable events:");
 787                                 for (i = 0; i < nr_counters; i++)
 788                                         fprintf(stderr, "\n\t%d %s", i, event_name(i));
 789
 790                                 prompt_integer(&sym_counter, "Enter details event counter");
 791
 792                                 if (sym_counter >= nr_counters) {
 793                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
 794                                         sym_counter = 0;
 795                                         sleep(1);
 796                                 }
 797                         } else sym_counter = 0;
 798                         break;
 799                 case 'f':
 800                         prompt_integer(&count_filter, "Enter display event count filter");
 801                         break;
 802                 case 'F':
 803                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
 804                         break;
 805                 case 'K':
 806                         hide_kernel_symbols = !hide_kernel_symbols;
 807                         break;
 808                 case 'q':
 809                 case 'Q':
 810                         printf("exiting.\n");
 811                         if (dump_symtab)
 812                                 dsos__fprintf(stderr);
 813                         exit(0);
 814                 case 's':
 815                         prompt_symbol(&sym_filter_entry, "Enter details symbol");
 816                         break;
 817                 case 'S':
 818                         if (!sym_filter_entry)
 819                                 break;
 820                         else {
 821                                 struct sym_entry *syme = sym_filter_entry;
 822
 823                                 pthread_mutex_lock(&syme->src->lock);
 824                                 sym_filter_entry = NULL;
 825                                 __zero_source_counters(syme);
 826                                 pthread_mutex_unlock(&syme->src->lock);
 827                         }
 828                         break;
 829                 case 'U':
 830                         hide_user_symbols = !hide_user_symbols;
 831                         break;
 832                 case 'w':
 833                         display_weighted = ~display_weighted;
 834                         break;
 835                 case 'z':
 836                         zero = ~zero;
 837                         break;
 838                 default:
 839                         break;
 840         }
 841 }
 842
 843 static void *display_thread(void *arg __used)
 844 {
 845         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 846         struct termios tc, save;
 847         int delay_msecs, c;
 848
 849         tcgetattr(0, &save);
 850         tc = save;
 851         tc.c_lflag &= ~(ICANON | ECHO);
 852         tc.c_cc[VMIN] = 0;
 853         tc.c_cc[VTIME] = 0;
 854
 855 repeat:
 856         delay_msecs = delay_secs * 1000;
 857         tcsetattr(0, TCSANOW, &tc);
 858         /* trash return*/
 859         getc(stdin);
 860
 861         do {
 862                 print_sym_table();
 863         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
 864
 865         c = getc(stdin);
 866         tcsetattr(0, TCSAFLUSH, &save);
 867
 868         handle_keypress(c);
 869         goto repeat;
 870
 871         return NULL;
 872 }
 873
 874 /* Tag samples to be skipped. */
 875 static const char *skip_symbols[] = {
 876         "default_idle",
 877         "cpu_idle",
 878         "enter_idle",
 879         "exit_idle",
 880         "mwait_idle",
 881         "mwait_idle_with_hints",
 882         "poll_idle",
 883         "ppc64_runlatch_off",
 884         "pseries_dedicated_idle_sleep",
 885         NULL
 886 };
 887
 888 static int symbol_filter(struct map *map, struct symbol *sym)
 889 {
 890         struct sym_entry *syme;
 891         const char *name = sym->name;
 892         int i;
 893
 894         /*
 895          * ppc64 uses function descriptors and appends a '.' to the
 896          * start of every instruction address. Remove it.
 897          */
 898         if (name[0] == '.')
 899                 name++;
 900
 901         if (!strcmp(name, "_text") ||
 902             !strcmp(name, "_etext") ||
 903             !strcmp(name, "_sinittext") ||
 904             !strncmp("init_module", name, 11) ||
 905             !strncmp("cleanup_module", name, 14) ||
 906             strstr(name, "_text_start") ||
 907             strstr(name, "_text_end"))
 908                 return 1;
 909
 910         syme = symbol__priv(sym);
 911         syme->map = map;
 912         syme->src = NULL;
 913         if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter))
 914                 sym_filter_entry = syme;
 915
 916         for (i = 0; skip_symbols[i]; i++) {
 917                 if (!strcmp(skip_symbols[i], name)) {
 918                         syme->skip = 1;
 919                         break;
 920                 }
 921         }
 922
 923         if (!syme->skip)
 924                 syme->name_len = strlen(sym->name);
 925
 926         return 0;
 927 }
 928
 929 static void event__process_sample(const event_t *self, int counter)
 930 {
 931         u64 ip = self->ip.ip;
 932         struct map *map;
 933         struct sym_entry *syme;
 934         struct symbol *sym;
 935         u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 936
 937         switch (origin) {
 938         case PERF_RECORD_MISC_USER: {
 939                 struct thread *thread;
 940
 941                 if (hide_user_symbols)
 942                         return;
 943
 944                 thread = threads__findnew(self->ip.pid);
 945                 if (thread == NULL)
 946                         return;
 947
 948                 map = thread__find_map(thread, MAP__FUNCTION, ip);
 949                 if (map != NULL) {
 950                         ip = map->map_ip(map, ip);
 951                         sym = map__find_symbol(map, ip, symbol_filter);
 952                         if (sym == NULL)
 953                                 return;
 954                         userspace_samples++;
 955                         break;
 956                 }
 957         }
 958                 /*
 959                  * If this is outside of all known maps,
 960                  * and is a negative address, try to look it
 961                  * up in the kernel dso, as it might be a
 962                  * vsyscall or vdso (which executes in user-mode).
 963                  */
 964                 if ((long long)ip >= 0)
 965                         return;
 966                 /* Fall thru */
 967         case PERF_RECORD_MISC_KERNEL:
 968                 if (hide_kernel_symbols)
 969                         return;
 970
 971                 sym = kernel_maps__find_function(ip, &map, symbol_filter);
 972                 if (sym == NULL)
 973                         return;
 974                 break;
 975         default:
 976                 return;
 977         }
 978
 979         syme = symbol__priv(sym);
 980
 981         if (!syme->skip) {
 982                 syme->count[counter]++;
 983                 syme->origin = origin;
 984                 record_precise_ip(syme, counter, ip);
 985                 pthread_mutex_lock(&active_symbols_lock);
 986                 if (list_empty(&syme->node) || !syme->node.next)
 987                         __list_insert_active_sym(syme);
 988                 pthread_mutex_unlock(&active_symbols_lock);
 989                 ++samples;
 990                 return;
 991         }
 992 }
 993
 994 static void event__process_mmap(event_t *self)
 995 {
 996         struct thread *thread = threads__findnew(self->mmap.pid);
 997
 998         if (thread != NULL) {
 999                 struct map *map = map__new(&self->mmap, MAP__FUNCTION, NULL, 0);
1000                 if (map != NULL)
1001                         thread__insert_map(thread, map);
1002         }
1003 }
1004
1005 static void event__process_comm(event_t *self)
1006 {
1007         struct thread *thread = threads__findnew(self->comm.pid);
1008
1009         if (thread != NULL)
1010                 thread__set_comm(thread, self->comm.comm);
1011 }
1012
1013 static int event__process(event_t *event)
1014 {
1015         switch (event->header.type) {
1016         case PERF_RECORD_COMM:
1017                 event__process_comm(event);
1018                 break;
1019         case PERF_RECORD_MMAP:
1020                 event__process_mmap(event);
1021                 break;
1022         default:
1023                 break;
1024         }
1025
1026         return 0;
1027 }
1028
1029 struct mmap_data {
1030         int                     counter;
1031         void                    *base;
1032         int                     mask;
1033         unsigned int            prev;
1034 };
1035
1036 static unsigned int mmap_read_head(struct mmap_data *md)
1037 {
1038         struct perf_event_mmap_page *pc = md->base;
1039         int head;
1040
1041         head = pc->data_head;
1042         rmb();
1043
1044         return head;
1045 }
1046
1047 static void mmap_read_counter(struct mmap_data *md)
1048 {
1049         unsigned int head = mmap_read_head(md);
1050         unsigned int old = md->prev;
1051         unsigned char *data = md->base + page_size;
1052         int diff;
1053
1054         /*
1055          * If we're further behind than half the buffer, there's a chance
1056          * the writer will bite our tail and mess up the samples under us.
1057          *
1058          * If we somehow ended up ahead of the head, we got messed up.
1059          *
1060          * In either case, truncate and restart at head.
1061          */
1062         diff = head - old;
1063         if (diff > md->mask / 2 || diff < 0) {
1064                 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
1065
1066                 /*
1067                  * head points to a known good entry, start there.
1068                  */
1069                 old = head;
1070         }
1071
1072         for (; old != head;) {
1073                 event_t *event = (event_t *)&data[old & md->mask];
1074
1075                 event_t event_copy;
1076
1077                 size_t size = event->header.size;
1078
1079                 /*
1080                  * Event straddles the mmap boundary -- header should always
1081                  * be inside due to u64 alignment of output.
1082                  */
1083                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
1084                         unsigned int offset = old;
1085                         unsigned int len = min(sizeof(*event), size), cpy;
1086                         void *dst = &event_copy;
1087
1088                         do {
1089                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
1090                                 memcpy(dst, &data[offset & md->mask], cpy);
1091                                 offset += cpy;
1092                                 dst += cpy;
1093                                 len -= cpy;
1094                         } while (len);
1095
1096                         event = &event_copy;
1097                 }
1098
1099                 if (event->header.type == PERF_RECORD_SAMPLE)
1100                         event__process_sample(event, md->counter);
1101                 else
1102                         event__process(event);
1103                 old += size;
1104         }
1105
1106         md->prev = old;
1107 }
1108
1109 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
1110 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
1111
1112 static void mmap_read(void)
1113 {
1114         int i, counter;
1115
1116         for (i = 0; i < nr_cpus; i++) {
1117                 for (counter = 0; counter < nr_counters; counter++)
1118                         mmap_read_counter(&mmap_array[i][counter]);
1119         }
1120 }
1121
1122 int nr_poll;
1123 int group_fd;
1124
1125 static void start_counter(int i, int counter)
1126 {
1127         struct perf_event_attr *attr;
1128         int cpu;
1129
1130         cpu = profile_cpu;
1131         if (target_pid == -1 && profile_cpu == -1)
1132                 cpu = i;
1133
1134         attr = attrs + counter;
1135
1136         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1137
1138         if (freq) {
1139                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
1140                 attr->freq              = 1;
1141                 attr->sample_freq       = freq;
1142         }
1143
1144         attr->inherit           = (cpu < 0) && inherit;
1145         attr->mmap              = 1;
1146
1147 try_again:
1148         fd[i][counter] = sys_perf_event_open(attr, target_pid, cpu, group_fd, 0);
1149
1150         if (fd[i][counter] < 0) {
1151                 int err = errno;
1152
1153                 if (err == EPERM || err == EACCES)
1154                         die("No permission - are you root?\n");
1155                 /*
1156                  * If it's cycles then fall back to hrtimer
1157                  * based cpu-clock-tick sw counter, which
1158                  * is always available even if no PMU support:
1159                  */
1160                 if (attr->type == PERF_TYPE_HARDWARE
1161                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1162
1163                         if (verbose)
1164                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
1165
1166                         attr->type = PERF_TYPE_SOFTWARE;
1167                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
1168                         goto try_again;
1169                 }
1170                 printf("\n");
1171                 error("perfcounter syscall returned with %d (%s)\n",
1172                         fd[i][counter], strerror(err));
1173                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1174                 exit(-1);
1175         }
1176         assert(fd[i][counter] >= 0);
1177         fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
1178
1179         /*
1180          * First counter acts as the group leader:
1181          */
1182         if (group && group_fd == -1)
1183                 group_fd = fd[i][counter];
1184
1185         event_array[nr_poll].fd = fd[i][counter];
1186         event_array[nr_poll].events = POLLIN;
1187         nr_poll++;
1188
1189         mmap_array[i][counter].counter = counter;
1190         mmap_array[i][counter].prev = 0;
1191         mmap_array[i][counter].mask = mmap_pages*page_size - 1;
1192         mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
1193                         PROT_READ, MAP_SHARED, fd[i][counter], 0);
1194         if (mmap_array[i][counter].base == MAP_FAILED)
1195                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1196 }
1197
1198 static int __cmd_top(void)
1199 {
1200         pthread_t thread;
1201         int i, counter;
1202         int ret;
1203
1204         if (target_pid != -1)
1205                 event__synthesize_thread(target_pid, event__process);
1206         else
1207                 event__synthesize_threads(event__process);
1208
1209         for (i = 0; i < nr_cpus; i++) {
1210                 group_fd = -1;
1211                 for (counter = 0; counter < nr_counters; counter++)
1212                         start_counter(i, counter);
1213         }
1214
1215         /* Wait for a minimal set of events before starting the snapshot */
1216         poll(event_array, nr_poll, 100);
1217
1218         mmap_read();
1219
1220         if (pthread_create(&thread, NULL, display_thread, NULL)) {
1221                 printf("Could not create display thread.\n");
1222                 exit(-1);
1223         }
1224
1225         if (realtime_prio) {
1226                 struct sched_param param;
1227
1228                 param.sched_priority = realtime_prio;
1229                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1230                         printf("Could not set realtime priority.\n");
1231                         exit(-1);
1232                 }
1233         }
1234
1235         while (1) {
1236                 int hits = samples;
1237
1238                 mmap_read();
1239
1240                 if (hits == samples)
1241                         ret = poll(event_array, nr_poll, 100);
1242         }
1243
1244         return 0;
1245 }
1246
1247 static const char * const top_usage[] = {
1248         "perf top [<options>]",
1249         NULL
1250 };
1251
1252 static const struct option options[] = {
1253         OPT_CALLBACK('e', "event", NULL, "event",
1254                      "event selector. use 'perf list' to list available events",
1255                      parse_events),
1256         OPT_INTEGER('c', "count", &default_interval,
1257                     "event period to sample"),
1258         OPT_INTEGER('p', "pid", &target_pid,
1259                     "profile events on existing pid"),
1260         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1261                             "system-wide collection from all CPUs"),
1262         OPT_INTEGER('C', "CPU", &profile_cpu,
1263                     "CPU to profile on"),
1264         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1265                    "file", "vmlinux pathname"),
1266         OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols,
1267                     "hide kernel symbols"),
1268         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
1269                     "number of mmap data pages"),
1270         OPT_INTEGER('r', "realtime", &realtime_prio,
1271                     "collect data with this RT SCHED_FIFO priority"),
1272         OPT_INTEGER('d', "delay", &delay_secs,
1273                     "number of seconds to delay between refreshes"),
1274         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1275                             "dump the symbol table used for profiling"),
1276         OPT_INTEGER('f', "count-filter", &count_filter,
1277                     "only display functions with more events than this"),
1278         OPT_BOOLEAN('g', "group", &group,
1279                             "put the counters into a counter group"),
1280         OPT_BOOLEAN('i', "inherit", &inherit,
1281                     "child tasks inherit counters"),
1282         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1283                     "symbol to annotate - requires -k option"),
1284         OPT_BOOLEAN('z', "zero", &zero,
1285                     "zero history across updates"),
1286         OPT_INTEGER('F', "freq", &freq,
1287                     "profile at this frequency"),
1288         OPT_INTEGER('E', "entries", &print_entries,
1289                     "display this many functions"),
1290         OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols,
1291                     "hide user symbols"),
1292         OPT_BOOLEAN('v', "verbose", &verbose,
1293                     "be more verbose (show counter open errors, etc)"),
1294         OPT_END()
1295 };
1296
1297 int cmd_top(int argc, const char **argv, const char *prefix __used)
1298 {
1299         int counter;
1300
1301         page_size = sysconf(_SC_PAGE_SIZE);
1302
1303         argc = parse_options(argc, argv, options, top_usage, 0);
1304         if (argc)
1305                 usage_with_options(top_usage, options);
1306
1307         /* CPU and PID are mutually exclusive */
1308         if (target_pid != -1 && profile_cpu != -1) {
1309                 printf("WARNING: PID switch overriding CPU\n");
1310                 sleep(1);
1311                 profile_cpu = -1;
1312         }
1313
1314         if (!nr_counters)
1315                 nr_counters = 1;
1316
1317         symbol_conf.priv_size = (sizeof(struct sym_entry) +
1318                                  (nr_counters + 1) * sizeof(unsigned long));
1319         if (symbol_conf.vmlinux_name == NULL)
1320                 symbol_conf.try_vmlinux_path = true;
1321         if (symbol__init(&symbol_conf) < 0)
1322                 return -1;
1323
1324         if (delay_secs < 1)
1325                 delay_secs = 1;
1326
1327         parse_source(sym_filter_entry);
1328
1329         /*
1330          * User specified count overrides default frequency.
1331          */
1332         if (default_interval)
1333                 freq = 0;
1334         else if (freq) {
1335                 default_interval = freq;
1336         } else {
1337                 fprintf(stderr, "frequency and count are zero, aborting\n");
1338                 exit(EXIT_FAILURE);
1339         }
1340
1341         /*
1342          * Fill in the ones not specifically initialized via -c:
1343          */
1344         for (counter = 0; counter < nr_counters; counter++) {
1345                 if (attrs[counter].sample_period)
1346                         continue;
1347
1348                 attrs[counter].sample_period = default_interval;
1349         }
1350
1351         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1352         assert(nr_cpus <= MAX_NR_CPUS);
1353         assert(nr_cpus >= 0);
1354
1355         if (target_pid != -1 || profile_cpu != -1)
1356                 nr_cpus = 1;
1357
1358         get_term_dimensions(&winsize);
1359         if (print_entries == 0) {
1360                 update_print_entries(&winsize);
1361                 signal(SIGWINCH, sig_winch_handler);
1362         }
1363
1364         return __cmd_top();
1365 }