Pileus Git - ~andy/linux/blob - tools/perf/builtin-top.c

   1 /*
   2  * builtin-top.c
   3  *
   4  * Builtin top command: Display a continuously updated profile of
   5  * any workload, CPU or specific PID.
   6  *
   7  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
   8  *
   9  * Improvements and fixes by:
  10  *
  11  *   Arjan van de Ven <arjan@linux.intel.com>
  12  *   Yanmin Zhang <yanmin.zhang@intel.com>
  13  *   Wu Fengguang <fengguang.wu@intel.com>
  14  *   Mike Galbraith <efault@gmx.de>
  15  *   Paul Mackerras <paulus@samba.org>
  16  *
  17  * Released under the GPL v2. (and only v2, not any later version)
  18  */
  19 #include "builtin.h"
  20
  21 #include "perf.h"
  22
  23 #include "util/symbol.h"
  24 #include "util/color.h"
  25 #include "util/thread.h"
  26 #include "util/util.h"
  27 #include <linux/rbtree.h>
  28 #include "util/parse-options.h"
  29 #include "util/parse-events.h"
  30
  31 #include "util/debug.h"
  32
  33 #include <assert.h>
  34 #include <fcntl.h>
  35
  36 #include <stdio.h>
  37 #include <termios.h>
  38 #include <unistd.h>
  39
  40 #include <errno.h>
  41 #include <time.h>
  42 #include <sched.h>
  43 #include <pthread.h>
  44
  45 #include <sys/syscall.h>
  46 #include <sys/ioctl.h>
  47 #include <sys/poll.h>
  48 #include <sys/prctl.h>
  49 #include <sys/wait.h>
  50 #include <sys/uio.h>
  51 #include <sys/mman.h>
  52
  53 #include <linux/unistd.h>
  54 #include <linux/types.h>
  55
  56 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  57
  58 static int                      system_wide                     =      0;
  59
  60 static int                      default_interval                =      0;
  61
  62 static int                      count_filter                    =      5;
  63 static int                      print_entries;
  64
  65 static int                      target_pid                      =     -1;
  66 static int                      inherit                         =      0;
  67 static int                      profile_cpu                     =     -1;
  68 static int                      nr_cpus                         =      0;
  69 static unsigned int             realtime_prio                   =      0;
  70 static int                      group                           =      0;
  71 static unsigned int             page_size;
  72 static unsigned int             mmap_pages                      =     16;
  73 static int                      freq                            =   1000; /* 1 KHz */
  74
  75 static int                      delay_secs                      =      2;
  76 static int                      zero                            =      0;
  77 static int                      dump_symtab                     =      0;
  78
  79 static bool                     hide_kernel_symbols             =  false;
  80 static bool                     hide_user_symbols               =  false;
  81 static struct winsize           winsize;
  82 const char                      *vmlinux_name;
  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__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 = calloc(1, 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;
 455         const int win_width = winsize.ws_col - 1;
 456         struct dso *unique_dso = NULL, *first_dso = NULL;
 457
 458         samples = userspace_samples = 0;
 459
 460         /* Sort the active symbols */
 461         pthread_mutex_lock(&active_symbols_lock);
 462         syme = list_entry(active_symbols.next, struct sym_entry, node);
 463         pthread_mutex_unlock(&active_symbols_lock);
 464
 465         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
 466                 syme->snap_count = syme->count[snap];
 467                 if (syme->snap_count != 0) {
 468
 469                         if ((hide_user_symbols &&
 470                              syme->origin == PERF_RECORD_MISC_USER) ||
 471                             (hide_kernel_symbols &&
 472                              syme->origin == PERF_RECORD_MISC_KERNEL)) {
 473                                 list_remove_active_sym(syme);
 474                                 continue;
 475                         }
 476                         syme->weight = sym_weight(syme);
 477                         rb_insert_active_sym(&tmp, syme);
 478                         sum_ksamples += syme->snap_count;
 479
 480                         for (j = 0; j < nr_counters; j++)
 481                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
 482                 } else
 483                         list_remove_active_sym(syme);
 484         }
 485
 486         puts(CONSOLE_CLEAR);
 487
 488         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 489         printf( "   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% [",
 490                 samples_per_sec,
 491                 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
 492
 493         if (nr_counters == 1 || !display_weighted) {
 494                 printf("%Ld", (u64)attrs[0].sample_period);
 495                 if (freq)
 496                         printf("Hz ");
 497                 else
 498                         printf(" ");
 499         }
 500
 501         if (!display_weighted)
 502                 printf("%s", event_name(sym_counter));
 503         else for (counter = 0; counter < nr_counters; counter++) {
 504                 if (counter)
 505                         printf("/");
 506
 507                 printf("%s", event_name(counter));
 508         }
 509
 510         printf( "], ");
 511
 512         if (target_pid != -1)
 513                 printf(" (target_pid: %d", target_pid);
 514         else
 515                 printf(" (all");
 516
 517         if (profile_cpu != -1)
 518                 printf(", cpu: %d)\n", profile_cpu);
 519         else {
 520                 if (target_pid != -1)
 521                         printf(")\n");
 522                 else
 523                         printf(", %d CPUs)\n", nr_cpus);
 524         }
 525
 526         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 527
 528         if (sym_filter_entry) {
 529                 show_details(sym_filter_entry);
 530                 return;
 531         }
 532
 533         /*
 534          * Find the longest symbol name that will be displayed
 535          */
 536         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
 537                 syme = rb_entry(nd, struct sym_entry, rb_node);
 538                 if (++printed > print_entries ||
 539                     (int)syme->snap_count < count_filter)
 540                         continue;
 541
 542                 if (first_dso == NULL)
 543                         unique_dso = first_dso = syme->map->dso;
 544                 else if (syme->map->dso != first_dso)
 545                         unique_dso = NULL;
 546
 547                 if (syme->map->dso->long_name_len > dso_width)
 548                         dso_width = syme->map->dso->long_name_len;
 549
 550                 if (syme->name_len > sym_width)
 551                         sym_width = syme->name_len;
 552         }
 553
 554         printed = 0;
 555
 556         if (unique_dso)
 557                 printf("DSO: %s\n", unique_dso->long_name);
 558         else {
 559                 int max_dso_width = winsize.ws_col - sym_width - 29;
 560                 if (dso_width > max_dso_width)
 561                         dso_width = max_dso_width;
 562                 putchar('\n');
 563         }
 564         if (nr_counters == 1)
 565                 printf("             samples  pcnt");
 566         else
 567                 printf("   weight    samples  pcnt");
 568
 569         if (verbose)
 570                 printf("         RIP       ");
 571         printf(" %-*.*s", sym_width, sym_width, "function");
 572         if (!unique_dso)
 573                 printf(" DSO");
 574         putchar('\n');
 575         printf("   %s    _______ _____",
 576                nr_counters == 1 ? "      " : "______");
 577         if (verbose)
 578                 printf(" ________________");
 579         printf(" %-*.*s", sym_width, sym_width, graph_line);
 580         if (!unique_dso)
 581                 printf(" %-*.*s", dso_width, dso_width, graph_line);
 582         puts("\n");
 583
 584         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
 585                 struct symbol *sym;
 586                 double pcnt;
 587
 588                 syme = rb_entry(nd, struct sym_entry, rb_node);
 589                 sym = sym_entry__symbol(syme);
 590
 591                 if (++printed > print_entries || (int)syme->snap_count < count_filter)
 592                         continue;
 593
 594                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
 595                                          sum_ksamples));
 596
 597                 if (nr_counters == 1 || !display_weighted)
 598                         printf("%20.2f ", syme->weight);
 599                 else
 600                         printf("%9.1f %10ld ", syme->weight, syme->snap_count);
 601
 602                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
 603                 if (verbose)
 604                         printf(" %016llx", sym->start);
 605                 printf(" %-*.*s", sym_width, sym_width, sym->name);
 606                 if (!unique_dso)
 607                         printf(" %-*.*s", dso_width, dso_width,
 608                                dso_width >= syme->map->dso->long_name_len ?
 609                                                 syme->map->dso->long_name :
 610                                                 syme->map->dso->short_name);
 611                 printf("\n");
 612         }
 613 }
 614
 615 static void prompt_integer(int *target, const char *msg)
 616 {
 617         char *buf = malloc(0), *p;
 618         size_t dummy = 0;
 619         int tmp;
 620
 621         fprintf(stdout, "\n%s: ", msg);
 622         if (getline(&buf, &dummy, stdin) < 0)
 623                 return;
 624
 625         p = strchr(buf, '\n');
 626         if (p)
 627                 *p = 0;
 628
 629         p = buf;
 630         while(*p) {
 631                 if (!isdigit(*p))
 632                         goto out_free;
 633                 p++;
 634         }
 635         tmp = strtoul(buf, NULL, 10);
 636         *target = tmp;
 637 out_free:
 638         free(buf);
 639 }
 640
 641 static void prompt_percent(int *target, const char *msg)
 642 {
 643         int tmp = 0;
 644
 645         prompt_integer(&tmp, msg);
 646         if (tmp >= 0 && tmp <= 100)
 647                 *target = tmp;
 648 }
 649
 650 static void prompt_symbol(struct sym_entry **target, const char *msg)
 651 {
 652         char *buf = malloc(0), *p;
 653         struct sym_entry *syme = *target, *n, *found = NULL;
 654         size_t dummy = 0;
 655
 656         /* zero counters of active symbol */
 657         if (syme) {
 658                 pthread_mutex_lock(&syme->src->lock);
 659                 __zero_source_counters(syme);
 660                 *target = NULL;
 661                 pthread_mutex_unlock(&syme->src->lock);
 662         }
 663
 664         fprintf(stdout, "\n%s: ", msg);
 665         if (getline(&buf, &dummy, stdin) < 0)
 666                 goto out_free;
 667
 668         p = strchr(buf, '\n');
 669         if (p)
 670                 *p = 0;
 671
 672         pthread_mutex_lock(&active_symbols_lock);
 673         syme = list_entry(active_symbols.next, struct sym_entry, node);
 674         pthread_mutex_unlock(&active_symbols_lock);
 675
 676         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
 677                 struct symbol *sym = sym_entry__symbol(syme);
 678
 679                 if (!strcmp(buf, sym->name)) {
 680                         found = syme;
 681                         break;
 682                 }
 683         }
 684
 685         if (!found) {
 686                 fprintf(stderr, "Sorry, %s is not active.\n", sym_filter);
 687                 sleep(1);
 688                 return;
 689         } else
 690                 parse_source(found);
 691
 692 out_free:
 693         free(buf);
 694 }
 695
 696 static void print_mapped_keys(void)
 697 {
 698         char *name = NULL;
 699
 700         if (sym_filter_entry) {
 701                 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
 702                 name = sym->name;
 703         }
 704
 705         fprintf(stdout, "\nMapped keys:\n");
 706         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", delay_secs);
 707         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", print_entries);
 708
 709         if (nr_counters > 1)
 710                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(sym_counter));
 711
 712         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", count_filter);
 713
 714         if (vmlinux_name) {
 715                 fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
 716                 fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 717                 fprintf(stdout, "\t[S]     stop annotation.\n");
 718         }
 719
 720         if (nr_counters > 1)
 721                 fprintf(stdout, "\t[w]     toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
 722
 723         fprintf(stdout,
 724                 "\t[K]     hide kernel_symbols symbols.             \t(%s)\n",
 725                 hide_kernel_symbols ? "yes" : "no");
 726         fprintf(stdout,
 727                 "\t[U]     hide user symbols.               \t(%s)\n",
 728                 hide_user_symbols ? "yes" : "no");
 729         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", zero ? 1 : 0);
 730         fprintf(stdout, "\t[qQ]    quit.\n");
 731 }
 732
 733 static int key_mapped(int c)
 734 {
 735         switch (c) {
 736                 case 'd':
 737                 case 'e':
 738                 case 'f':
 739                 case 'z':
 740                 case 'q':
 741                 case 'Q':
 742                 case 'K':
 743                 case 'U':
 744                         return 1;
 745                 case 'E':
 746                 case 'w':
 747                         return nr_counters > 1 ? 1 : 0;
 748                 case 'F':
 749                 case 's':
 750                 case 'S':
 751                         return vmlinux_name ? 1 : 0;
 752                 default:
 753                         break;
 754         }
 755
 756         return 0;
 757 }
 758
 759 static void handle_keypress(int c)
 760 {
 761         if (!key_mapped(c)) {
 762                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 763                 struct termios tc, save;
 764
 765                 print_mapped_keys();
 766                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 767                 fflush(stdout);
 768
 769                 tcgetattr(0, &save);
 770                 tc = save;
 771                 tc.c_lflag &= ~(ICANON | ECHO);
 772                 tc.c_cc[VMIN] = 0;
 773                 tc.c_cc[VTIME] = 0;
 774                 tcsetattr(0, TCSANOW, &tc);
 775
 776                 poll(&stdin_poll, 1, -1);
 777                 c = getc(stdin);
 778
 779                 tcsetattr(0, TCSAFLUSH, &save);
 780                 if (!key_mapped(c))
 781                         return;
 782         }
 783
 784         switch (c) {
 785                 case 'd':
 786                         prompt_integer(&delay_secs, "Enter display delay");
 787                         if (delay_secs < 1)
 788                                 delay_secs = 1;
 789                         break;
 790                 case 'e':
 791                         prompt_integer(&print_entries, "Enter display entries (lines)");
 792                         if (print_entries == 0) {
 793                                 sig_winch_handler(SIGWINCH);
 794                                 signal(SIGWINCH, sig_winch_handler);
 795                         } else
 796                                 signal(SIGWINCH, SIG_DFL);
 797                         break;
 798                 case 'E':
 799                         if (nr_counters > 1) {
 800                                 int i;
 801
 802                                 fprintf(stderr, "\nAvailable events:");
 803                                 for (i = 0; i < nr_counters; i++)
 804                                         fprintf(stderr, "\n\t%d %s", i, event_name(i));
 805
 806                                 prompt_integer(&sym_counter, "Enter details event counter");
 807
 808                                 if (sym_counter >= nr_counters) {
 809                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
 810                                         sym_counter = 0;
 811                                         sleep(1);
 812                                 }
 813                         } else sym_counter = 0;
 814                         break;
 815                 case 'f':
 816                         prompt_integer(&count_filter, "Enter display event count filter");
 817                         break;
 818                 case 'F':
 819                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
 820                         break;
 821                 case 'K':
 822                         hide_kernel_symbols = !hide_kernel_symbols;
 823                         break;
 824                 case 'q':
 825                 case 'Q':
 826                         printf("exiting.\n");
 827                         if (dump_symtab)
 828                                 dsos__fprintf(stderr);
 829                         exit(0);
 830                 case 's':
 831                         prompt_symbol(&sym_filter_entry, "Enter details symbol");
 832                         break;
 833                 case 'S':
 834                         if (!sym_filter_entry)
 835                                 break;
 836                         else {
 837                                 struct sym_entry *syme = sym_filter_entry;
 838
 839                                 pthread_mutex_lock(&syme->src->lock);
 840                                 sym_filter_entry = NULL;
 841                                 __zero_source_counters(syme);
 842                                 pthread_mutex_unlock(&syme->src->lock);
 843                         }
 844                         break;
 845                 case 'U':
 846                         hide_user_symbols = !hide_user_symbols;
 847                         break;
 848                 case 'w':
 849                         display_weighted = ~display_weighted;
 850                         break;
 851                 case 'z':
 852                         zero = ~zero;
 853                         break;
 854                 default:
 855                         break;
 856         }
 857 }
 858
 859 static void *display_thread(void *arg __used)
 860 {
 861         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 862         struct termios tc, save;
 863         int delay_msecs, c;
 864
 865         tcgetattr(0, &save);
 866         tc = save;
 867         tc.c_lflag &= ~(ICANON | ECHO);
 868         tc.c_cc[VMIN] = 0;
 869         tc.c_cc[VTIME] = 0;
 870
 871 repeat:
 872         delay_msecs = delay_secs * 1000;
 873         tcsetattr(0, TCSANOW, &tc);
 874         /* trash return*/
 875         getc(stdin);
 876
 877         do {
 878                 print_sym_table();
 879         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
 880
 881         c = getc(stdin);
 882         tcsetattr(0, TCSAFLUSH, &save);
 883
 884         handle_keypress(c);
 885         goto repeat;
 886
 887         return NULL;
 888 }
 889
 890 /* Tag samples to be skipped. */
 891 static const char *skip_symbols[] = {
 892         "default_idle",
 893         "cpu_idle",
 894         "enter_idle",
 895         "exit_idle",
 896         "mwait_idle",
 897         "mwait_idle_with_hints",
 898         "poll_idle",
 899         "ppc64_runlatch_off",
 900         "pseries_dedicated_idle_sleep",
 901         NULL
 902 };
 903
 904 static int symbol_filter(struct map *map, struct symbol *sym)
 905 {
 906         struct sym_entry *syme;
 907         const char *name = sym->name;
 908         int i;
 909
 910         /*
 911          * ppc64 uses function descriptors and appends a '.' to the
 912          * start of every instruction address. Remove it.
 913          */
 914         if (name[0] == '.')
 915                 name++;
 916
 917         if (!strcmp(name, "_text") ||
 918             !strcmp(name, "_etext") ||
 919             !strcmp(name, "_sinittext") ||
 920             !strncmp("init_module", name, 11) ||
 921             !strncmp("cleanup_module", name, 14) ||
 922             strstr(name, "_text_start") ||
 923             strstr(name, "_text_end"))
 924                 return 1;
 925
 926         syme = symbol__priv(sym);
 927         syme->map = map;
 928         syme->src = NULL;
 929         if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter))
 930                 sym_filter_entry = syme;
 931
 932         for (i = 0; skip_symbols[i]; i++) {
 933                 if (!strcmp(skip_symbols[i], name)) {
 934                         syme->skip = 1;
 935                         break;
 936                 }
 937         }
 938
 939         if (!syme->skip)
 940                 syme->name_len = strlen(sym->name);
 941
 942         return 0;
 943 }
 944
 945 static void event__process_sample(const event_t *self, int counter)
 946 {
 947         u64 ip = self->ip.ip;
 948         struct map *map;
 949         struct sym_entry *syme;
 950         struct symbol *sym;
 951         u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 952
 953         switch (origin) {
 954         case PERF_RECORD_MISC_USER: {
 955                 struct thread *thread;
 956
 957                 if (hide_user_symbols)
 958                         return;
 959
 960                 thread = threads__findnew(self->ip.pid);
 961                 if (thread == NULL)
 962                         return;
 963
 964                 map = thread__find_map(thread, ip);
 965                 if (map != NULL) {
 966                         ip = map->map_ip(map, ip);
 967                         sym = map__find_symbol(map, ip, symbol_filter);
 968                         if (sym == NULL)
 969                                 return;
 970                         userspace_samples++;
 971                         break;
 972                 }
 973         }
 974                 /*
 975                  * If this is outside of all known maps,
 976                  * and is a negative address, try to look it
 977                  * up in the kernel dso, as it might be a
 978                  * vsyscall or vdso (which executes in user-mode).
 979                  */
 980                 if ((long long)ip >= 0)
 981                         return;
 982                 /* Fall thru */
 983         case PERF_RECORD_MISC_KERNEL:
 984                 if (hide_kernel_symbols)
 985                         return;
 986
 987                 sym = kernel_maps__find_symbol(ip, &map, symbol_filter);
 988                 if (sym == NULL)
 989                         return;
 990                 break;
 991         default:
 992                 return;
 993         }
 994
 995         syme = symbol__priv(sym);
 996
 997         if (!syme->skip) {
 998                 syme->count[counter]++;
 999                 syme->origin = origin;
1000                 record_precise_ip(syme, counter, ip);
1001                 pthread_mutex_lock(&active_symbols_lock);
1002                 if (list_empty(&syme->node) || !syme->node.next)
1003                         __list_insert_active_sym(syme);
1004                 pthread_mutex_unlock(&active_symbols_lock);
1005                 ++samples;
1006                 return;
1007         }
1008 }
1009
1010 static void event__process_mmap(event_t *self)
1011 {
1012         struct thread *thread = threads__findnew(self->mmap.pid);
1013
1014         if (thread != NULL) {
1015                 struct map *map = map__new(&self->mmap, NULL, 0);
1016                 if (map != NULL)
1017                         thread__insert_map(thread, map);
1018         }
1019 }
1020
1021 static void event__process_comm(event_t *self)
1022 {
1023         struct thread *thread = threads__findnew(self->comm.pid);
1024
1025         if (thread != NULL)
1026                 thread__set_comm(thread, self->comm.comm);
1027 }
1028
1029 static int event__process(event_t *event)
1030 {
1031         switch (event->header.type) {
1032         case PERF_RECORD_COMM:
1033                 event__process_comm(event);
1034                 break;
1035         case PERF_RECORD_MMAP:
1036                 event__process_mmap(event);
1037                 break;
1038         default:
1039                 break;
1040         }
1041
1042         return 0;
1043 }
1044
1045 struct mmap_data {
1046         int                     counter;
1047         void                    *base;
1048         int                     mask;
1049         unsigned int            prev;
1050 };
1051
1052 static unsigned int mmap_read_head(struct mmap_data *md)
1053 {
1054         struct perf_event_mmap_page *pc = md->base;
1055         int head;
1056
1057         head = pc->data_head;
1058         rmb();
1059
1060         return head;
1061 }
1062
1063 static void mmap_read_counter(struct mmap_data *md)
1064 {
1065         unsigned int head = mmap_read_head(md);
1066         unsigned int old = md->prev;
1067         unsigned char *data = md->base + page_size;
1068         int diff;
1069
1070         /*
1071          * If we're further behind than half the buffer, there's a chance
1072          * the writer will bite our tail and mess up the samples under us.
1073          *
1074          * If we somehow ended up ahead of the head, we got messed up.
1075          *
1076          * In either case, truncate and restart at head.
1077          */
1078         diff = head - old;
1079         if (diff > md->mask / 2 || diff < 0) {
1080                 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
1081
1082                 /*
1083                  * head points to a known good entry, start there.
1084                  */
1085                 old = head;
1086         }
1087
1088         for (; old != head;) {
1089                 event_t *event = (event_t *)&data[old & md->mask];
1090
1091                 event_t event_copy;
1092
1093                 size_t size = event->header.size;
1094
1095                 /*
1096                  * Event straddles the mmap boundary -- header should always
1097                  * be inside due to u64 alignment of output.
1098                  */
1099                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
1100                         unsigned int offset = old;
1101                         unsigned int len = min(sizeof(*event), size), cpy;
1102                         void *dst = &event_copy;
1103
1104                         do {
1105                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
1106                                 memcpy(dst, &data[offset & md->mask], cpy);
1107                                 offset += cpy;
1108                                 dst += cpy;
1109                                 len -= cpy;
1110                         } while (len);
1111
1112                         event = &event_copy;
1113                 }
1114
1115                 if (event->header.type == PERF_RECORD_SAMPLE)
1116                         event__process_sample(event, md->counter);
1117                 else
1118                         event__process(event);
1119                 old += size;
1120         }
1121
1122         md->prev = old;
1123 }
1124
1125 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
1126 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
1127
1128 static void mmap_read(void)
1129 {
1130         int i, counter;
1131
1132         for (i = 0; i < nr_cpus; i++) {
1133                 for (counter = 0; counter < nr_counters; counter++)
1134                         mmap_read_counter(&mmap_array[i][counter]);
1135         }
1136 }
1137
1138 int nr_poll;
1139 int group_fd;
1140
1141 static void start_counter(int i, int counter)
1142 {
1143         struct perf_event_attr *attr;
1144         int cpu;
1145
1146         cpu = profile_cpu;
1147         if (target_pid == -1 && profile_cpu == -1)
1148                 cpu = i;
1149
1150         attr = attrs + counter;
1151
1152         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1153
1154         if (freq) {
1155                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
1156                 attr->freq              = 1;
1157                 attr->sample_freq       = freq;
1158         }
1159
1160         attr->inherit           = (cpu < 0) && inherit;
1161         attr->mmap              = 1;
1162
1163 try_again:
1164         fd[i][counter] = sys_perf_event_open(attr, target_pid, cpu, group_fd, 0);
1165
1166         if (fd[i][counter] < 0) {
1167                 int err = errno;
1168
1169                 if (err == EPERM || err == EACCES)
1170                         die("No permission - are you root?\n");
1171                 /*
1172                  * If it's cycles then fall back to hrtimer
1173                  * based cpu-clock-tick sw counter, which
1174                  * is always available even if no PMU support:
1175                  */
1176                 if (attr->type == PERF_TYPE_HARDWARE
1177                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1178
1179                         if (verbose)
1180                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
1181
1182                         attr->type = PERF_TYPE_SOFTWARE;
1183                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
1184                         goto try_again;
1185                 }
1186                 printf("\n");
1187                 error("perfcounter syscall returned with %d (%s)\n",
1188                         fd[i][counter], strerror(err));
1189                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1190                 exit(-1);
1191         }
1192         assert(fd[i][counter] >= 0);
1193         fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
1194
1195         /*
1196          * First counter acts as the group leader:
1197          */
1198         if (group && group_fd == -1)
1199                 group_fd = fd[i][counter];
1200
1201         event_array[nr_poll].fd = fd[i][counter];
1202         event_array[nr_poll].events = POLLIN;
1203         nr_poll++;
1204
1205         mmap_array[i][counter].counter = counter;
1206         mmap_array[i][counter].prev = 0;
1207         mmap_array[i][counter].mask = mmap_pages*page_size - 1;
1208         mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
1209                         PROT_READ, MAP_SHARED, fd[i][counter], 0);
1210         if (mmap_array[i][counter].base == MAP_FAILED)
1211                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1212 }
1213
1214 static int __cmd_top(void)
1215 {
1216         pthread_t thread;
1217         int i, counter;
1218         int ret;
1219
1220         if (target_pid != -1)
1221                 event__synthesize_thread(target_pid, event__process);
1222         else
1223                 event__synthesize_threads(event__process);
1224
1225         for (i = 0; i < nr_cpus; i++) {
1226                 group_fd = -1;
1227                 for (counter = 0; counter < nr_counters; counter++)
1228                         start_counter(i, counter);
1229         }
1230
1231         /* Wait for a minimal set of events before starting the snapshot */
1232         poll(event_array, nr_poll, 100);
1233
1234         mmap_read();
1235
1236         if (pthread_create(&thread, NULL, display_thread, NULL)) {
1237                 printf("Could not create display thread.\n");
1238                 exit(-1);
1239         }
1240
1241         if (realtime_prio) {
1242                 struct sched_param param;
1243
1244                 param.sched_priority = realtime_prio;
1245                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1246                         printf("Could not set realtime priority.\n");
1247                         exit(-1);
1248                 }
1249         }
1250
1251         while (1) {
1252                 int hits = samples;
1253
1254                 mmap_read();
1255
1256                 if (hits == samples)
1257                         ret = poll(event_array, nr_poll, 100);
1258         }
1259
1260         return 0;
1261 }
1262
1263 static const char * const top_usage[] = {
1264         "perf top [<options>]",
1265         NULL
1266 };
1267
1268 static const struct option options[] = {
1269         OPT_CALLBACK('e', "event", NULL, "event",
1270                      "event selector. use 'perf list' to list available events",
1271                      parse_events),
1272         OPT_INTEGER('c', "count", &default_interval,
1273                     "event period to sample"),
1274         OPT_INTEGER('p', "pid", &target_pid,
1275                     "profile events on existing pid"),
1276         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1277                             "system-wide collection from all CPUs"),
1278         OPT_INTEGER('C', "CPU", &profile_cpu,
1279                     "CPU to profile on"),
1280         OPT_STRING('k', "vmlinux", &vmlinux_name, "file", "vmlinux pathname"),
1281         OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols,
1282                     "hide kernel symbols"),
1283         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
1284                     "number of mmap data pages"),
1285         OPT_INTEGER('r', "realtime", &realtime_prio,
1286                     "collect data with this RT SCHED_FIFO priority"),
1287         OPT_INTEGER('d', "delay", &delay_secs,
1288                     "number of seconds to delay between refreshes"),
1289         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1290                             "dump the symbol table used for profiling"),
1291         OPT_INTEGER('f', "count-filter", &count_filter,
1292                     "only display functions with more events than this"),
1293         OPT_BOOLEAN('g', "group", &group,
1294                             "put the counters into a counter group"),
1295         OPT_BOOLEAN('i', "inherit", &inherit,
1296                     "child tasks inherit counters"),
1297         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1298                     "symbol to annotate - requires -k option"),
1299         OPT_BOOLEAN('z', "zero", &zero,
1300                     "zero history across updates"),
1301         OPT_INTEGER('F', "freq", &freq,
1302                     "profile at this frequency"),
1303         OPT_INTEGER('E', "entries", &print_entries,
1304                     "display this many functions"),
1305         OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols,
1306                     "hide user symbols"),
1307         OPT_BOOLEAN('v', "verbose", &verbose,
1308                     "be more verbose (show counter open errors, etc)"),
1309         OPT_END()
1310 };
1311
1312 int cmd_top(int argc, const char **argv, const char *prefix __used)
1313 {
1314         int counter, err;
1315
1316         page_size = sysconf(_SC_PAGE_SIZE);
1317
1318         argc = parse_options(argc, argv, options, top_usage, 0);
1319         if (argc)
1320                 usage_with_options(top_usage, options);
1321
1322         /* CPU and PID are mutually exclusive */
1323         if (target_pid != -1 && profile_cpu != -1) {
1324                 printf("WARNING: PID switch overriding CPU\n");
1325                 sleep(1);
1326                 profile_cpu = -1;
1327         }
1328
1329         if (!nr_counters)
1330                 nr_counters = 1;
1331
1332         symbol__init(sizeof(struct sym_entry) +
1333                      (nr_counters + 1) * sizeof(unsigned long));
1334
1335         if (delay_secs < 1)
1336                 delay_secs = 1;
1337
1338         err = kernel_maps__init(vmlinux_name, !vmlinux_name, true);
1339         if (err < 0)
1340                 return err;
1341         parse_source(sym_filter_entry);
1342
1343         /*
1344          * User specified count overrides default frequency.
1345          */
1346         if (default_interval)
1347                 freq = 0;
1348         else if (freq) {
1349                 default_interval = freq;
1350         } else {
1351                 fprintf(stderr, "frequency and count are zero, aborting\n");
1352                 exit(EXIT_FAILURE);
1353         }
1354
1355         /*
1356          * Fill in the ones not specifically initialized via -c:
1357          */
1358         for (counter = 0; counter < nr_counters; counter++) {
1359                 if (attrs[counter].sample_period)
1360                         continue;
1361
1362                 attrs[counter].sample_period = default_interval;
1363         }
1364
1365         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1366         assert(nr_cpus <= MAX_NR_CPUS);
1367         assert(nr_cpus >= 0);
1368
1369         if (target_pid != -1 || profile_cpu != -1)
1370                 nr_cpus = 1;
1371
1372         get_term_dimensions(&winsize);
1373         if (print_entries == 0) {
1374                 update_print_entries(&winsize);
1375                 signal(SIGWINCH, sig_winch_handler);
1376         }
1377
1378         return __cmd_top();
1379 }