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/color.h"
  24 #include "util/session.h"
  25 #include "util/symbol.h"
  26 #include "util/thread.h"
  27 #include "util/util.h"
  28 #include <linux/rbtree.h>
  29 #include "util/parse-options.h"
  30 #include "util/parse-events.h"
  31
  32 #include "util/debug.h"
  33
  34 #include <assert.h>
  35 #include <fcntl.h>
  36
  37 #include <stdio.h>
  38 #include <termios.h>
  39 #include <unistd.h>
  40
  41 #include <errno.h>
  42 #include <time.h>
  43 #include <sched.h>
  44 #include <pthread.h>
  45
  46 #include <sys/syscall.h>
  47 #include <sys/ioctl.h>
  48 #include <sys/poll.h>
  49 #include <sys/prctl.h>
  50 #include <sys/wait.h>
  51 #include <sys/uio.h>
  52 #include <sys/mman.h>
  53
  54 #include <linux/unistd.h>
  55 #include <linux/types.h>
  56
  57 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  58
  59 static int                      system_wide                     =      0;
  60
  61 static int                      default_interval                =      0;
  62
  63 static int                      count_filter                    =      5;
  64 static int                      print_entries;
  65
  66 static int                      target_pid                      =     -1;
  67 static int                      inherit                         =      0;
  68 static int                      profile_cpu                     =     -1;
  69 static int                      nr_cpus                         =      0;
  70 static unsigned int             realtime_prio                   =      0;
  71 static int                      group                           =      0;
  72 static unsigned int             page_size;
  73 static unsigned int             mmap_pages                      =     16;
  74 static int                      freq                            =   1000; /* 1 KHz */
  75
  76 static int                      delay_secs                      =      2;
  77 static int                      zero                            =      0;
  78 static int                      dump_symtab                     =      0;
  79
  80 static bool                     hide_kernel_symbols             =  false;
  81 static bool                     hide_user_symbols               =  false;
  82 static struct winsize           winsize;
  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 struct sym_entry                *sym_filter_entry_sched         =   NULL;
  98 static int                      sym_pcnt_filter                 =      5;
  99 static int                      sym_counter                     =      0;
 100 static int                      display_weighted                =     -1;
 101
 102 /*
 103  * Symbols
 104  */
 105
 106 struct sym_entry_source {
 107         struct source_line      *source;
 108         struct source_line      *lines;
 109         struct source_line      **lines_tail;
 110         pthread_mutex_t         lock;
 111 };
 112
 113 struct sym_entry {
 114         struct rb_node          rb_node;
 115         struct list_head        node;
 116         unsigned long           snap_count;
 117         double                  weight;
 118         int                     skip;
 119         u16                     name_len;
 120         u8                      origin;
 121         struct map              *map;
 122         struct sym_entry_source *src;
 123         unsigned long           count[0];
 124 };
 125
 126 /*
 127  * Source functions
 128  */
 129
 130 static inline struct symbol *sym_entry__symbol(struct sym_entry *self)
 131 {
 132        return ((void *)self) + symbol_conf.priv_size;
 133 }
 134
 135 static void get_term_dimensions(struct winsize *ws)
 136 {
 137         char *s = getenv("LINES");
 138
 139         if (s != NULL) {
 140                 ws->ws_row = atoi(s);
 141                 s = getenv("COLUMNS");
 142                 if (s != NULL) {
 143                         ws->ws_col = atoi(s);
 144                         if (ws->ws_row && ws->ws_col)
 145                                 return;
 146                 }
 147         }
 148 #ifdef TIOCGWINSZ
 149         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
 150             ws->ws_row && ws->ws_col)
 151                 return;
 152 #endif
 153         ws->ws_row = 25;
 154         ws->ws_col = 80;
 155 }
 156
 157 static void update_print_entries(struct winsize *ws)
 158 {
 159         print_entries = ws->ws_row;
 160
 161         if (print_entries > 9)
 162                 print_entries -= 9;
 163 }
 164
 165 static void sig_winch_handler(int sig __used)
 166 {
 167         get_term_dimensions(&winsize);
 168         update_print_entries(&winsize);
 169 }
 170
 171 static void parse_source(struct sym_entry *syme)
 172 {
 173         struct symbol *sym;
 174         struct sym_entry_source *source;
 175         struct map *map;
 176         FILE *file;
 177         char command[PATH_MAX*2];
 178         const char *path;
 179         u64 len;
 180
 181         if (!syme)
 182                 return;
 183
 184         if (syme->src == NULL) {
 185                 syme->src = zalloc(sizeof(*source));
 186                 if (syme->src == NULL)
 187                         return;
 188                 pthread_mutex_init(&syme->src->lock, NULL);
 189         }
 190
 191         source = syme->src;
 192
 193         if (source->lines) {
 194                 pthread_mutex_lock(&source->lock);
 195                 goto out_assign;
 196         }
 197
 198         sym = sym_entry__symbol(syme);
 199         map = syme->map;
 200         path = map->dso->long_name;
 201
 202         len = sym->end - sym->start;
 203
 204         sprintf(command,
 205                 "objdump --start-address=%#0*Lx --stop-address=%#0*Lx -dS %s",
 206                 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->start),
 207                 BITS_PER_LONG / 4, map__rip_2objdump(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, *sep;
 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                 src->eip = strtoull(src->line, &sep, 16);
 238                 if (*sep == ':')
 239                         src->eip = map__objdump_2ip(map, src->eip);
 240                 else /* this line has no ip info (e.g. source line) */
 241                         src->eip = 0;
 242         }
 243         pclose(file);
 244 out_assign:
 245         sym_filter_entry = syme;
 246         pthread_mutex_unlock(&source->lock);
 247 }
 248
 249 static void __zero_source_counters(struct sym_entry *syme)
 250 {
 251         int i;
 252         struct source_line *line;
 253
 254         line = syme->src->lines;
 255         while (line) {
 256                 for (i = 0; i < nr_counters; i++)
 257                         line->count[i] = 0;
 258                 line = line->next;
 259         }
 260 }
 261
 262 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
 263 {
 264         struct source_line *line;
 265
 266         if (syme != sym_filter_entry)
 267                 return;
 268
 269         if (pthread_mutex_trylock(&syme->src->lock))
 270                 return;
 271
 272         if (syme->src == NULL || syme->src->source == NULL)
 273                 goto out_unlock;
 274
 275         for (line = syme->src->lines; line; line = line->next) {
 276                 /* skip lines without IP info */
 277                 if (line->eip == 0)
 278                         continue;
 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         const size_t pattern_len = BITS_PER_LONG / 4 + 2;
 295         char pattern[pattern_len + 1];
 296
 297         sprintf(pattern, "%0*Lx <", BITS_PER_LONG / 4,
 298                 map__rip_2objdump(syme->map, symbol->start));
 299
 300         pthread_mutex_lock(&syme->src->lock);
 301         for (line = syme->src->lines; line; line = line->next) {
 302                 if (memcmp(line->line, pattern, pattern_len) == 0) {
 303                         syme->src->source = line;
 304                         break;
 305                 }
 306         }
 307         pthread_mutex_unlock(&syme->src->lock);
 308 }
 309
 310 static void show_lines(struct source_line *queue, int count, int total)
 311 {
 312         int i;
 313         struct source_line *line;
 314
 315         line = queue;
 316         for (i = 0; i < count; i++) {
 317                 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
 318
 319                 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
 320                 line = line->next;
 321         }
 322 }
 323
 324 #define TRACE_COUNT     3
 325
 326 static void show_details(struct sym_entry *syme)
 327 {
 328         struct symbol *symbol;
 329         struct source_line *line;
 330         struct source_line *line_queue = NULL;
 331         int displayed = 0;
 332         int line_queue_count = 0, total = 0, more = 0;
 333
 334         if (!syme)
 335                 return;
 336
 337         if (!syme->src->source)
 338                 lookup_sym_source(syme);
 339
 340         if (!syme->src->source)
 341                 return;
 342
 343         symbol = sym_entry__symbol(syme);
 344         printf("Showing %s for %s\n", event_name(sym_counter), symbol->name);
 345         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
 346
 347         pthread_mutex_lock(&syme->src->lock);
 348         line = syme->src->source;
 349         while (line) {
 350                 total += line->count[sym_counter];
 351                 line = line->next;
 352         }
 353
 354         line = syme->src->source;
 355         while (line) {
 356                 float pcnt = 0.0;
 357
 358                 if (!line_queue_count)
 359                         line_queue = line;
 360                 line_queue_count++;
 361
 362                 if (line->count[sym_counter])
 363                         pcnt = 100.0 * line->count[sym_counter] / (float)total;
 364                 if (pcnt >= (float)sym_pcnt_filter) {
 365                         if (displayed <= print_entries)
 366                                 show_lines(line_queue, line_queue_count, total);
 367                         else more++;
 368                         displayed += line_queue_count;
 369                         line_queue_count = 0;
 370                         line_queue = NULL;
 371                 } else if (line_queue_count > TRACE_COUNT) {
 372                         line_queue = line_queue->next;
 373                         line_queue_count--;
 374                 }
 375
 376                 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
 377                 line = line->next;
 378         }
 379         pthread_mutex_unlock(&syme->src->lock);
 380         if (more)
 381                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 382 }
 383
 384 /*
 385  * Symbols will be added here in event__process_sample and will get out
 386  * after decayed.
 387  */
 388 static LIST_HEAD(active_symbols);
 389 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
 390
 391 /*
 392  * Ordering weight: count-1 * count-2 * ... / count-n
 393  */
 394 static double sym_weight(const struct sym_entry *sym)
 395 {
 396         double weight = sym->snap_count;
 397         int counter;
 398
 399         if (!display_weighted)
 400                 return weight;
 401
 402         for (counter = 1; counter < nr_counters-1; counter++)
 403                 weight *= sym->count[counter];
 404
 405         weight /= (sym->count[counter] + 1);
 406
 407         return weight;
 408 }
 409
 410 static long                     samples;
 411 static long                     userspace_samples;
 412 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
 413
 414 static void __list_insert_active_sym(struct sym_entry *syme)
 415 {
 416         list_add(&syme->node, &active_symbols);
 417 }
 418
 419 static void list_remove_active_sym(struct sym_entry *syme)
 420 {
 421         pthread_mutex_lock(&active_symbols_lock);
 422         list_del_init(&syme->node);
 423         pthread_mutex_unlock(&active_symbols_lock);
 424 }
 425
 426 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
 427 {
 428         struct rb_node **p = &tree->rb_node;
 429         struct rb_node *parent = NULL;
 430         struct sym_entry *iter;
 431
 432         while (*p != NULL) {
 433                 parent = *p;
 434                 iter = rb_entry(parent, struct sym_entry, rb_node);
 435
 436                 if (se->weight > iter->weight)
 437                         p = &(*p)->rb_left;
 438                 else
 439                         p = &(*p)->rb_right;
 440         }
 441
 442         rb_link_node(&se->rb_node, parent, p);
 443         rb_insert_color(&se->rb_node, tree);
 444 }
 445
 446 static void print_sym_table(void)
 447 {
 448         int printed = 0, j;
 449         int counter, snap = !display_weighted ? sym_counter : 0;
 450         float samples_per_sec = samples/delay_secs;
 451         float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
 452         float sum_ksamples = 0.0;
 453         struct sym_entry *syme, *n;
 454         struct rb_root tmp = RB_ROOT;
 455         struct rb_node *nd;
 456         int sym_width = 0, dso_width = 0, max_dso_width;
 457         const int win_width = winsize.ws_col - 1;
 458
 459         samples = userspace_samples = 0;
 460
 461         /* Sort the active symbols */
 462         pthread_mutex_lock(&active_symbols_lock);
 463         syme = list_entry(active_symbols.next, struct sym_entry, node);
 464         pthread_mutex_unlock(&active_symbols_lock);
 465
 466         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
 467                 syme->snap_count = syme->count[snap];
 468                 if (syme->snap_count != 0) {
 469
 470                         if ((hide_user_symbols &&
 471                              syme->origin == PERF_RECORD_MISC_USER) ||
 472                             (hide_kernel_symbols &&
 473                              syme->origin == PERF_RECORD_MISC_KERNEL)) {
 474                                 list_remove_active_sym(syme);
 475                                 continue;
 476                         }
 477                         syme->weight = sym_weight(syme);
 478                         rb_insert_active_sym(&tmp, syme);
 479                         sum_ksamples += syme->snap_count;
 480
 481                         for (j = 0; j < nr_counters; j++)
 482                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
 483                 } else
 484                         list_remove_active_sym(syme);
 485         }
 486
 487         puts(CONSOLE_CLEAR);
 488
 489         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 490         printf( "   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% [",
 491                 samples_per_sec,
 492                 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
 493
 494         if (nr_counters == 1 || !display_weighted) {
 495                 printf("%Ld", (u64)attrs[0].sample_period);
 496                 if (freq)
 497                         printf("Hz ");
 498                 else
 499                         printf(" ");
 500         }
 501
 502         if (!display_weighted)
 503                 printf("%s", event_name(sym_counter));
 504         else for (counter = 0; counter < nr_counters; counter++) {
 505                 if (counter)
 506                         printf("/");
 507
 508                 printf("%s", event_name(counter));
 509         }
 510
 511         printf( "], ");
 512
 513         if (target_pid != -1)
 514                 printf(" (target_pid: %d", target_pid);
 515         else
 516                 printf(" (all");
 517
 518         if (profile_cpu != -1)
 519                 printf(", cpu: %d)\n", profile_cpu);
 520         else {
 521                 if (target_pid != -1)
 522                         printf(")\n");
 523                 else
 524                         printf(", %d CPUs)\n", nr_cpus);
 525         }
 526
 527         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 528
 529         if (sym_filter_entry) {
 530                 show_details(sym_filter_entry);
 531                 return;
 532         }
 533
 534         /*
 535          * Find the longest symbol name that will be displayed
 536          */
 537         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
 538                 syme = rb_entry(nd, struct sym_entry, rb_node);
 539                 if (++printed > print_entries ||
 540                     (int)syme->snap_count < count_filter)
 541                         continue;
 542
 543                 if (syme->map->dso->long_name_len > dso_width)
 544                         dso_width = syme->map->dso->long_name_len;
 545
 546                 if (syme->name_len > sym_width)
 547                         sym_width = syme->name_len;
 548         }
 549
 550         printed = 0;
 551
 552         max_dso_width = winsize.ws_col - sym_width - 29;
 553         if (dso_width > max_dso_width)
 554                 dso_width = max_dso_width;
 555         putchar('\n');
 556         if (nr_counters == 1)
 557                 printf("             samples  pcnt");
 558         else
 559                 printf("   weight    samples  pcnt");
 560
 561         if (verbose)
 562                 printf("         RIP       ");
 563         printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
 564         printf("   %s    _______ _____",
 565                nr_counters == 1 ? "      " : "______");
 566         if (verbose)
 567                 printf(" ________________");
 568         printf(" %-*.*s", sym_width, sym_width, graph_line);
 569         printf(" %-*.*s", dso_width, dso_width, graph_line);
 570         puts("\n");
 571
 572         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
 573                 struct symbol *sym;
 574                 double pcnt;
 575
 576                 syme = rb_entry(nd, struct sym_entry, rb_node);
 577                 sym = sym_entry__symbol(syme);
 578
 579                 if (++printed > print_entries || (int)syme->snap_count < count_filter)
 580                         continue;
 581
 582                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
 583                                          sum_ksamples));
 584
 585                 if (nr_counters == 1 || !display_weighted)
 586                         printf("%20.2f ", syme->weight);
 587                 else
 588                         printf("%9.1f %10ld ", syme->weight, syme->snap_count);
 589
 590                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
 591                 if (verbose)
 592                         printf(" %016llx", sym->start);
 593                 printf(" %-*.*s", sym_width, sym_width, sym->name);
 594                 printf(" %-*.*s\n", dso_width, dso_width,
 595                        dso_width >= syme->map->dso->long_name_len ?
 596                                         syme->map->dso->long_name :
 597                                         syme->map->dso->short_name);
 598         }
 599 }
 600
 601 static void prompt_integer(int *target, const char *msg)
 602 {
 603         char *buf = malloc(0), *p;
 604         size_t dummy = 0;
 605         int tmp;
 606
 607         fprintf(stdout, "\n%s: ", msg);
 608         if (getline(&buf, &dummy, stdin) < 0)
 609                 return;
 610
 611         p = strchr(buf, '\n');
 612         if (p)
 613                 *p = 0;
 614
 615         p = buf;
 616         while(*p) {
 617                 if (!isdigit(*p))
 618                         goto out_free;
 619                 p++;
 620         }
 621         tmp = strtoul(buf, NULL, 10);
 622         *target = tmp;
 623 out_free:
 624         free(buf);
 625 }
 626
 627 static void prompt_percent(int *target, const char *msg)
 628 {
 629         int tmp = 0;
 630
 631         prompt_integer(&tmp, msg);
 632         if (tmp >= 0 && tmp <= 100)
 633                 *target = tmp;
 634 }
 635
 636 static void prompt_symbol(struct sym_entry **target, const char *msg)
 637 {
 638         char *buf = malloc(0), *p;
 639         struct sym_entry *syme = *target, *n, *found = NULL;
 640         size_t dummy = 0;
 641
 642         /* zero counters of active symbol */
 643         if (syme) {
 644                 pthread_mutex_lock(&syme->src->lock);
 645                 __zero_source_counters(syme);
 646                 *target = NULL;
 647                 pthread_mutex_unlock(&syme->src->lock);
 648         }
 649
 650         fprintf(stdout, "\n%s: ", msg);
 651         if (getline(&buf, &dummy, stdin) < 0)
 652                 goto out_free;
 653
 654         p = strchr(buf, '\n');
 655         if (p)
 656                 *p = 0;
 657
 658         pthread_mutex_lock(&active_symbols_lock);
 659         syme = list_entry(active_symbols.next, struct sym_entry, node);
 660         pthread_mutex_unlock(&active_symbols_lock);
 661
 662         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
 663                 struct symbol *sym = sym_entry__symbol(syme);
 664
 665                 if (!strcmp(buf, sym->name)) {
 666                         found = syme;
 667                         break;
 668                 }
 669         }
 670
 671         if (!found) {
 672                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 673                 sleep(1);
 674                 return;
 675         } else
 676                 parse_source(found);
 677
 678 out_free:
 679         free(buf);
 680 }
 681
 682 static void print_mapped_keys(void)
 683 {
 684         char *name = NULL;
 685
 686         if (sym_filter_entry) {
 687                 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
 688                 name = sym->name;
 689         }
 690
 691         fprintf(stdout, "\nMapped keys:\n");
 692         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", delay_secs);
 693         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", print_entries);
 694
 695         if (nr_counters > 1)
 696                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(sym_counter));
 697
 698         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", count_filter);
 699
 700         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
 701         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 702         fprintf(stdout, "\t[S]     stop annotation.\n");
 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                 case 'F':
 729                 case 's':
 730                 case 'S':
 731                         return 1;
 732                 case 'E':
 733                 case 'w':
 734                         return nr_counters > 1 ? 1 : 0;
 735                 default:
 736                         break;
 737         }
 738
 739         return 0;
 740 }
 741
 742 static void handle_keypress(int c)
 743 {
 744         if (!key_mapped(c)) {
 745                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 746                 struct termios tc, save;
 747
 748                 print_mapped_keys();
 749                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 750                 fflush(stdout);
 751
 752                 tcgetattr(0, &save);
 753                 tc = save;
 754                 tc.c_lflag &= ~(ICANON | ECHO);
 755                 tc.c_cc[VMIN] = 0;
 756                 tc.c_cc[VTIME] = 0;
 757                 tcsetattr(0, TCSANOW, &tc);
 758
 759                 poll(&stdin_poll, 1, -1);
 760                 c = getc(stdin);
 761
 762                 tcsetattr(0, TCSAFLUSH, &save);
 763                 if (!key_mapped(c))
 764                         return;
 765         }
 766
 767         switch (c) {
 768                 case 'd':
 769                         prompt_integer(&delay_secs, "Enter display delay");
 770                         if (delay_secs < 1)
 771                                 delay_secs = 1;
 772                         break;
 773                 case 'e':
 774                         prompt_integer(&print_entries, "Enter display entries (lines)");
 775                         if (print_entries == 0) {
 776                                 sig_winch_handler(SIGWINCH);
 777                                 signal(SIGWINCH, sig_winch_handler);
 778                         } else
 779                                 signal(SIGWINCH, SIG_DFL);
 780                         break;
 781                 case 'E':
 782                         if (nr_counters > 1) {
 783                                 int i;
 784
 785                                 fprintf(stderr, "\nAvailable events:");
 786                                 for (i = 0; i < nr_counters; i++)
 787                                         fprintf(stderr, "\n\t%d %s", i, event_name(i));
 788
 789                                 prompt_integer(&sym_counter, "Enter details event counter");
 790
 791                                 if (sym_counter >= nr_counters) {
 792                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
 793                                         sym_counter = 0;
 794                                         sleep(1);
 795                                 }
 796                         } else sym_counter = 0;
 797                         break;
 798                 case 'f':
 799                         prompt_integer(&count_filter, "Enter display event count filter");
 800                         break;
 801                 case 'F':
 802                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
 803                         break;
 804                 case 'K':
 805                         hide_kernel_symbols = !hide_kernel_symbols;
 806                         break;
 807                 case 'q':
 808                 case 'Q':
 809                         printf("exiting.\n");
 810                         if (dump_symtab)
 811                                 dsos__fprintf(stderr);
 812                         exit(0);
 813                 case 's':
 814                         prompt_symbol(&sym_filter_entry, "Enter details symbol");
 815                         break;
 816                 case 'S':
 817                         if (!sym_filter_entry)
 818                                 break;
 819                         else {
 820                                 struct sym_entry *syme = sym_filter_entry;
 821
 822                                 pthread_mutex_lock(&syme->src->lock);
 823                                 sym_filter_entry = NULL;
 824                                 __zero_source_counters(syme);
 825                                 pthread_mutex_unlock(&syme->src->lock);
 826                         }
 827                         break;
 828                 case 'U':
 829                         hide_user_symbols = !hide_user_symbols;
 830                         break;
 831                 case 'w':
 832                         display_weighted = ~display_weighted;
 833                         break;
 834                 case 'z':
 835                         zero = ~zero;
 836                         break;
 837                 default:
 838                         break;
 839         }
 840 }
 841
 842 static void *display_thread(void *arg __used)
 843 {
 844         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 845         struct termios tc, save;
 846         int delay_msecs, c;
 847
 848         tcgetattr(0, &save);
 849         tc = save;
 850         tc.c_lflag &= ~(ICANON | ECHO);
 851         tc.c_cc[VMIN] = 0;
 852         tc.c_cc[VTIME] = 0;
 853
 854 repeat:
 855         delay_msecs = delay_secs * 1000;
 856         tcsetattr(0, TCSANOW, &tc);
 857         /* trash return*/
 858         getc(stdin);
 859
 860         do {
 861                 print_sym_table();
 862         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
 863
 864         c = getc(stdin);
 865         tcsetattr(0, TCSAFLUSH, &save);
 866
 867         handle_keypress(c);
 868         goto repeat;
 869
 870         return NULL;
 871 }
 872
 873 /* Tag samples to be skipped. */
 874 static const char *skip_symbols[] = {
 875         "default_idle",
 876         "cpu_idle",
 877         "enter_idle",
 878         "exit_idle",
 879         "mwait_idle",
 880         "mwait_idle_with_hints",
 881         "poll_idle",
 882         "ppc64_runlatch_off",
 883         "pseries_dedicated_idle_sleep",
 884         NULL
 885 };
 886
 887 static int symbol_filter(struct map *map, struct symbol *sym)
 888 {
 889         struct sym_entry *syme;
 890         const char *name = sym->name;
 891         int i;
 892
 893         /*
 894          * ppc64 uses function descriptors and appends a '.' to the
 895          * start of every instruction address. Remove it.
 896          */
 897         if (name[0] == '.')
 898                 name++;
 899
 900         if (!strcmp(name, "_text") ||
 901             !strcmp(name, "_etext") ||
 902             !strcmp(name, "_sinittext") ||
 903             !strncmp("init_module", name, 11) ||
 904             !strncmp("cleanup_module", name, 14) ||
 905             strstr(name, "_text_start") ||
 906             strstr(name, "_text_end"))
 907                 return 1;
 908
 909         syme = symbol__priv(sym);
 910         syme->map = map;
 911         syme->src = NULL;
 912
 913         if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) {
 914                 /* schedule initial sym_filter_entry setup */
 915                 sym_filter_entry_sched = syme;
 916                 sym_filter = NULL;
 917         }
 918
 919         for (i = 0; skip_symbols[i]; i++) {
 920                 if (!strcmp(skip_symbols[i], name)) {
 921                         syme->skip = 1;
 922                         break;
 923                 }
 924         }
 925
 926         if (!syme->skip)
 927                 syme->name_len = strlen(sym->name);
 928
 929         return 0;
 930 }
 931
 932 static void event__process_sample(const event_t *self,
 933                                  struct perf_session *session, int counter)
 934 {
 935         u64 ip = self->ip.ip;
 936         struct sym_entry *syme;
 937         struct addr_location al;
 938         u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 939
 940         ++samples;
 941
 942         switch (origin) {
 943         case PERF_RECORD_MISC_USER:
 944                 ++userspace_samples;
 945                 if (hide_user_symbols)
 946                         return;
 947                 break;
 948         case PERF_RECORD_MISC_KERNEL:
 949                 if (hide_kernel_symbols)
 950                         return;
 951                 break;
 952         default:
 953                 return;
 954         }
 955
 956         if (event__preprocess_sample(self, session, &al, symbol_filter) < 0 ||
 957             al.filtered)
 958                 return;
 959
 960         if (al.sym == NULL) {
 961                 /*
 962                  * As we do lazy loading of symtabs we only will know if the
 963                  * specified vmlinux file is invalid when we actually have a
 964                  * hit in kernel space and then try to load it. So if we get
 965                  * here and there are _no_ symbols in the DSO backing the
 966                  * kernel map, bail out.
 967                  *
 968                  * We may never get here, for instance, if we use -K/
 969                  * --hide-kernel-symbols, even if the user specifies an
 970                  * invalid --vmlinux ;-)
 971                  */
 972                 if (al.map == session->vmlinux_maps[MAP__FUNCTION] &&
 973                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 974                         pr_err("The %s file can't be used\n",
 975                                symbol_conf.vmlinux_name);
 976                         exit(1);
 977                 }
 978
 979                 return;
 980         }
 981
 982         /* let's see, whether we need to install initial sym_filter_entry */
 983         if (sym_filter_entry_sched) {
 984                 sym_filter_entry = sym_filter_entry_sched;
 985                 sym_filter_entry_sched = NULL;
 986                 parse_source(sym_filter_entry);
 987         }
 988
 989         syme = symbol__priv(al.sym);
 990         if (!syme->skip) {
 991                 syme->count[counter]++;
 992                 syme->origin = origin;
 993                 record_precise_ip(syme, counter, ip);
 994                 pthread_mutex_lock(&active_symbols_lock);
 995                 if (list_empty(&syme->node) || !syme->node.next)
 996                         __list_insert_active_sym(syme);
 997                 pthread_mutex_unlock(&active_symbols_lock);
 998         }
 999 }
1000
1001 static int event__process(event_t *event, struct perf_session *session)
1002 {
1003         switch (event->header.type) {
1004         case PERF_RECORD_COMM:
1005                 event__process_comm(event, session);
1006                 break;
1007         case PERF_RECORD_MMAP:
1008                 event__process_mmap(event, session);
1009                 break;
1010         case PERF_RECORD_FORK:
1011         case PERF_RECORD_EXIT:
1012                 event__process_task(event, session);
1013                 break;
1014         default:
1015                 break;
1016         }
1017
1018         return 0;
1019 }
1020
1021 struct mmap_data {
1022         int                     counter;
1023         void                    *base;
1024         int                     mask;
1025         unsigned int            prev;
1026 };
1027
1028 static unsigned int mmap_read_head(struct mmap_data *md)
1029 {
1030         struct perf_event_mmap_page *pc = md->base;
1031         int head;
1032
1033         head = pc->data_head;
1034         rmb();
1035
1036         return head;
1037 }
1038
1039 static void perf_session__mmap_read_counter(struct perf_session *self,
1040                                             struct mmap_data *md)
1041 {
1042         unsigned int head = mmap_read_head(md);
1043         unsigned int old = md->prev;
1044         unsigned char *data = md->base + page_size;
1045         int diff;
1046
1047         /*
1048          * If we're further behind than half the buffer, there's a chance
1049          * the writer will bite our tail and mess up the samples under us.
1050          *
1051          * If we somehow ended up ahead of the head, we got messed up.
1052          *
1053          * In either case, truncate and restart at head.
1054          */
1055         diff = head - old;
1056         if (diff > md->mask / 2 || diff < 0) {
1057                 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
1058
1059                 /*
1060                  * head points to a known good entry, start there.
1061                  */
1062                 old = head;
1063         }
1064
1065         for (; old != head;) {
1066                 event_t *event = (event_t *)&data[old & md->mask];
1067
1068                 event_t event_copy;
1069
1070                 size_t size = event->header.size;
1071
1072                 /*
1073                  * Event straddles the mmap boundary -- header should always
1074                  * be inside due to u64 alignment of output.
1075                  */
1076                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
1077                         unsigned int offset = old;
1078                         unsigned int len = min(sizeof(*event), size), cpy;
1079                         void *dst = &event_copy;
1080
1081                         do {
1082                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
1083                                 memcpy(dst, &data[offset & md->mask], cpy);
1084                                 offset += cpy;
1085                                 dst += cpy;
1086                                 len -= cpy;
1087                         } while (len);
1088
1089                         event = &event_copy;
1090                 }
1091
1092                 if (event->header.type == PERF_RECORD_SAMPLE)
1093                         event__process_sample(event, self, md->counter);
1094                 else
1095                         event__process(event, self);
1096                 old += size;
1097         }
1098
1099         md->prev = old;
1100 }
1101
1102 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
1103 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
1104
1105 static void perf_session__mmap_read(struct perf_session *self)
1106 {
1107         int i, counter;
1108
1109         for (i = 0; i < nr_cpus; i++) {
1110                 for (counter = 0; counter < nr_counters; counter++)
1111                         perf_session__mmap_read_counter(self, &mmap_array[i][counter]);
1112         }
1113 }
1114
1115 int nr_poll;
1116 int group_fd;
1117
1118 static void start_counter(int i, int counter)
1119 {
1120         struct perf_event_attr *attr;
1121         int cpu;
1122
1123         cpu = profile_cpu;
1124         if (target_pid == -1 && profile_cpu == -1)
1125                 cpu = i;
1126
1127         attr = attrs + counter;
1128
1129         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1130
1131         if (freq) {
1132                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
1133                 attr->freq              = 1;
1134                 attr->sample_freq       = freq;
1135         }
1136
1137         attr->inherit           = (cpu < 0) && inherit;
1138         attr->mmap              = 1;
1139
1140 try_again:
1141         fd[i][counter] = sys_perf_event_open(attr, target_pid, cpu, group_fd, 0);
1142
1143         if (fd[i][counter] < 0) {
1144                 int err = errno;
1145
1146                 if (err == EPERM || err == EACCES)
1147                         die("No permission - are you root?\n");
1148                 /*
1149                  * If it's cycles then fall back to hrtimer
1150                  * based cpu-clock-tick sw counter, which
1151                  * is always available even if no PMU support:
1152                  */
1153                 if (attr->type == PERF_TYPE_HARDWARE
1154                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1155
1156                         if (verbose)
1157                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
1158
1159                         attr->type = PERF_TYPE_SOFTWARE;
1160                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
1161                         goto try_again;
1162                 }
1163                 printf("\n");
1164                 error("perfcounter syscall returned with %d (%s)\n",
1165                         fd[i][counter], strerror(err));
1166                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1167                 exit(-1);
1168         }
1169         assert(fd[i][counter] >= 0);
1170         fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
1171
1172         /*
1173          * First counter acts as the group leader:
1174          */
1175         if (group && group_fd == -1)
1176                 group_fd = fd[i][counter];
1177
1178         event_array[nr_poll].fd = fd[i][counter];
1179         event_array[nr_poll].events = POLLIN;
1180         nr_poll++;
1181
1182         mmap_array[i][counter].counter = counter;
1183         mmap_array[i][counter].prev = 0;
1184         mmap_array[i][counter].mask = mmap_pages*page_size - 1;
1185         mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
1186                         PROT_READ, MAP_SHARED, fd[i][counter], 0);
1187         if (mmap_array[i][counter].base == MAP_FAILED)
1188                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1189 }
1190
1191 static int __cmd_top(void)
1192 {
1193         pthread_t thread;
1194         int i, counter;
1195         int ret;
1196         /*
1197          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1198          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1199          */
1200         struct perf_session *session = perf_session__new(NULL, O_WRONLY, false);
1201         if (session == NULL)
1202                 return -ENOMEM;
1203
1204         if (target_pid != -1)
1205                 event__synthesize_thread(target_pid, event__process, session);
1206         else
1207                 event__synthesize_threads(event__process, session);
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         perf_session__mmap_read(session);
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                 perf_session__mmap_read(session);
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"),
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
1320         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1321         if (symbol__init() < 0)
1322                 return -1;
1323
1324         if (delay_secs < 1)
1325                 delay_secs = 1;
1326
1327         /*
1328          * User specified count overrides default frequency.
1329          */
1330         if (default_interval)
1331                 freq = 0;
1332         else if (freq) {
1333                 default_interval = freq;
1334         } else {
1335                 fprintf(stderr, "frequency and count are zero, aborting\n");
1336                 exit(EXIT_FAILURE);
1337         }
1338
1339         /*
1340          * Fill in the ones not specifically initialized via -c:
1341          */
1342         for (counter = 0; counter < nr_counters; counter++) {
1343                 if (attrs[counter].sample_period)
1344                         continue;
1345
1346                 attrs[counter].sample_period = default_interval;
1347         }
1348
1349         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1350         assert(nr_cpus <= MAX_NR_CPUS);
1351         assert(nr_cpus >= 0);
1352
1353         if (target_pid != -1 || profile_cpu != -1)
1354                 nr_cpus = 1;
1355
1356         get_term_dimensions(&winsize);
1357         if (print_entries == 0) {
1358                 update_print_entries(&winsize);
1359                 signal(SIGWINCH, sig_winch_handler);
1360         }
1361
1362         return __cmd_top();
1363 }