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 #include "util/cpumap.h"
  32
  33 #include "util/debug.h"
  34
  35 #include <assert.h>
  36 #include <fcntl.h>
  37
  38 #include <stdio.h>
  39 #include <termios.h>
  40 #include <unistd.h>
  41
  42 #include <errno.h>
  43 #include <time.h>
  44 #include <sched.h>
  45 #include <pthread.h>
  46
  47 #include <sys/syscall.h>
  48 #include <sys/ioctl.h>
  49 #include <sys/poll.h>
  50 #include <sys/prctl.h>
  51 #include <sys/wait.h>
  52 #include <sys/uio.h>
  53 #include <sys/mman.h>
  54
  55 #include <linux/unistd.h>
  56 #include <linux/types.h>
  57
  58 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  59
  60 static int                      system_wide                     =      0;
  61
  62 static int                      default_interval                =      0;
  63
  64 static int                      count_filter                    =      5;
  65 static int                      print_entries;
  66
  67 static int                      target_pid                      =     -1;
  68 static int                      inherit                         =      0;
  69 static int                      profile_cpu                     =     -1;
  70 static int                      nr_cpus                         =      0;
  71 static unsigned int             realtime_prio                   =      0;
  72 static int                      group                           =      0;
  73 static unsigned int             page_size;
  74 static unsigned int             mmap_pages                      =     16;
  75 static int                      freq                            =   1000; /* 1 KHz */
  76
  77 static int                      delay_secs                      =      2;
  78 static int                      zero                            =      0;
  79 static int                      dump_symtab                     =      0;
  80
  81 static bool                     hide_kernel_symbols             =  false;
  82 static bool                     hide_user_symbols               =  false;
  83 static struct winsize           winsize;
  84
  85 /*
  86  * Source
  87  */
  88
  89 struct source_line {
  90         u64                     eip;
  91         unsigned long           count[MAX_COUNTERS];
  92         char                    *line;
  93         struct source_line      *next;
  94 };
  95
  96 static char                     *sym_filter                     =   NULL;
  97 struct sym_entry                *sym_filter_entry               =   NULL;
  98 struct sym_entry                *sym_filter_entry_sched         =   NULL;
  99 static int                      sym_pcnt_filter                 =      5;
 100 static int                      sym_counter                     =      0;
 101 static int                      display_weighted                =     -1;
 102
 103 /*
 104  * Symbols
 105  */
 106
 107 struct sym_entry_source {
 108         struct source_line      *source;
 109         struct source_line      *lines;
 110         struct source_line      **lines_tail;
 111         pthread_mutex_t         lock;
 112 };
 113
 114 struct sym_entry {
 115         struct rb_node          rb_node;
 116         struct list_head        node;
 117         unsigned long           snap_count;
 118         double                  weight;
 119         int                     skip;
 120         u16                     name_len;
 121         u8                      origin;
 122         struct map              *map;
 123         struct sym_entry_source *src;
 124         unsigned long           count[0];
 125 };
 126
 127 /*
 128  * Source functions
 129  */
 130
 131 static inline struct symbol *sym_entry__symbol(struct sym_entry *self)
 132 {
 133        return ((void *)self) + symbol_conf.priv_size;
 134 }
 135
 136 static void get_term_dimensions(struct winsize *ws)
 137 {
 138         char *s = getenv("LINES");
 139
 140         if (s != NULL) {
 141                 ws->ws_row = atoi(s);
 142                 s = getenv("COLUMNS");
 143                 if (s != NULL) {
 144                         ws->ws_col = atoi(s);
 145                         if (ws->ws_row && ws->ws_col)
 146                                 return;
 147                 }
 148         }
 149 #ifdef TIOCGWINSZ
 150         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
 151             ws->ws_row && ws->ws_col)
 152                 return;
 153 #endif
 154         ws->ws_row = 25;
 155         ws->ws_col = 80;
 156 }
 157
 158 static void update_print_entries(struct winsize *ws)
 159 {
 160         print_entries = ws->ws_row;
 161
 162         if (print_entries > 9)
 163                 print_entries -= 9;
 164 }
 165
 166 static void sig_winch_handler(int sig __used)
 167 {
 168         get_term_dimensions(&winsize);
 169         update_print_entries(&winsize);
 170 }
 171
 172 static void parse_source(struct sym_entry *syme)
 173 {
 174         struct symbol *sym;
 175         struct sym_entry_source *source;
 176         struct map *map;
 177         FILE *file;
 178         char command[PATH_MAX*2];
 179         const char *path;
 180         u64 len;
 181
 182         if (!syme)
 183                 return;
 184
 185         if (syme->src == NULL) {
 186                 syme->src = zalloc(sizeof(*source));
 187                 if (syme->src == NULL)
 188                         return;
 189                 pthread_mutex_init(&syme->src->lock, NULL);
 190         }
 191
 192         source = syme->src;
 193
 194         if (source->lines) {
 195                 pthread_mutex_lock(&source->lock);
 196                 goto out_assign;
 197         }
 198
 199         sym = sym_entry__symbol(syme);
 200         map = syme->map;
 201         path = map->dso->long_name;
 202
 203         len = sym->end - sym->start;
 204
 205         sprintf(command,
 206                 "objdump --start-address=%#0*Lx --stop-address=%#0*Lx -dS %s",
 207                 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->start),
 208                 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->end), path);
 209
 210         file = popen(command, "r");
 211         if (!file)
 212                 return;
 213
 214         pthread_mutex_lock(&source->lock);
 215         source->lines_tail = &source->lines;
 216         while (!feof(file)) {
 217                 struct source_line *src;
 218                 size_t dummy = 0;
 219                 char *c, *sep;
 220
 221                 src = malloc(sizeof(struct source_line));
 222                 assert(src != NULL);
 223                 memset(src, 0, sizeof(struct source_line));
 224
 225                 if (getline(&src->line, &dummy, file) < 0)
 226                         break;
 227                 if (!src->line)
 228                         break;
 229
 230                 c = strchr(src->line, '\n');
 231                 if (c)
 232                         *c = 0;
 233
 234                 src->next = NULL;
 235                 *source->lines_tail = src;
 236                 source->lines_tail = &src->next;
 237
 238                 src->eip = strtoull(src->line, &sep, 16);
 239                 if (*sep == ':')
 240                         src->eip = map__objdump_2ip(map, src->eip);
 241                 else /* this line has no ip info (e.g. source line) */
 242                         src->eip = 0;
 243         }
 244         pclose(file);
 245 out_assign:
 246         sym_filter_entry = syme;
 247         pthread_mutex_unlock(&source->lock);
 248 }
 249
 250 static void __zero_source_counters(struct sym_entry *syme)
 251 {
 252         int i;
 253         struct source_line *line;
 254
 255         line = syme->src->lines;
 256         while (line) {
 257                 for (i = 0; i < nr_counters; i++)
 258                         line->count[i] = 0;
 259                 line = line->next;
 260         }
 261 }
 262
 263 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
 264 {
 265         struct source_line *line;
 266
 267         if (syme != sym_filter_entry)
 268                 return;
 269
 270         if (pthread_mutex_trylock(&syme->src->lock))
 271                 return;
 272
 273         if (syme->src == NULL || syme->src->source == NULL)
 274                 goto out_unlock;
 275
 276         for (line = syme->src->lines; line; line = line->next) {
 277                 /* skip lines without IP info */
 278                 if (line->eip == 0)
 279                         continue;
 280                 if (line->eip == ip) {
 281                         line->count[counter]++;
 282                         break;
 283                 }
 284                 if (line->eip > ip)
 285                         break;
 286         }
 287 out_unlock:
 288         pthread_mutex_unlock(&syme->src->lock);
 289 }
 290
 291 #define PATTERN_LEN             (BITS_PER_LONG / 4 + 2)
 292
 293 static void lookup_sym_source(struct sym_entry *syme)
 294 {
 295         struct symbol *symbol = sym_entry__symbol(syme);
 296         struct source_line *line;
 297         char pattern[PATTERN_LEN + 1];
 298
 299         sprintf(pattern, "%0*Lx <", BITS_PER_LONG / 4,
 300                 map__rip_2objdump(syme->map, symbol->start));
 301
 302         pthread_mutex_lock(&syme->src->lock);
 303         for (line = syme->src->lines; line; line = line->next) {
 304                 if (memcmp(line->line, pattern, PATTERN_LEN) == 0) {
 305                         syme->src->source = line;
 306                         break;
 307                 }
 308         }
 309         pthread_mutex_unlock(&syme->src->lock);
 310 }
 311
 312 static void show_lines(struct source_line *queue, int count, int total)
 313 {
 314         int i;
 315         struct source_line *line;
 316
 317         line = queue;
 318         for (i = 0; i < count; i++) {
 319                 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
 320
 321                 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
 322                 line = line->next;
 323         }
 324 }
 325
 326 #define TRACE_COUNT     3
 327
 328 static void show_details(struct sym_entry *syme)
 329 {
 330         struct symbol *symbol;
 331         struct source_line *line;
 332         struct source_line *line_queue = NULL;
 333         int displayed = 0;
 334         int line_queue_count = 0, total = 0, more = 0;
 335
 336         if (!syme)
 337                 return;
 338
 339         if (!syme->src->source)
 340                 lookup_sym_source(syme);
 341
 342         if (!syme->src->source)
 343                 return;
 344
 345         symbol = sym_entry__symbol(syme);
 346         printf("Showing %s for %s\n", event_name(sym_counter), symbol->name);
 347         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
 348
 349         pthread_mutex_lock(&syme->src->lock);
 350         line = syme->src->source;
 351         while (line) {
 352                 total += line->count[sym_counter];
 353                 line = line->next;
 354         }
 355
 356         line = syme->src->source;
 357         while (line) {
 358                 float pcnt = 0.0;
 359
 360                 if (!line_queue_count)
 361                         line_queue = line;
 362                 line_queue_count++;
 363
 364                 if (line->count[sym_counter])
 365                         pcnt = 100.0 * line->count[sym_counter] / (float)total;
 366                 if (pcnt >= (float)sym_pcnt_filter) {
 367                         if (displayed <= print_entries)
 368                                 show_lines(line_queue, line_queue_count, total);
 369                         else more++;
 370                         displayed += line_queue_count;
 371                         line_queue_count = 0;
 372                         line_queue = NULL;
 373                 } else if (line_queue_count > TRACE_COUNT) {
 374                         line_queue = line_queue->next;
 375                         line_queue_count--;
 376                 }
 377
 378                 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
 379                 line = line->next;
 380         }
 381         pthread_mutex_unlock(&syme->src->lock);
 382         if (more)
 383                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 384 }
 385
 386 /*
 387  * Symbols will be added here in event__process_sample and will get out
 388  * after decayed.
 389  */
 390 static LIST_HEAD(active_symbols);
 391 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
 392
 393 /*
 394  * Ordering weight: count-1 * count-2 * ... / count-n
 395  */
 396 static double sym_weight(const struct sym_entry *sym)
 397 {
 398         double weight = sym->snap_count;
 399         int counter;
 400
 401         if (!display_weighted)
 402                 return weight;
 403
 404         for (counter = 1; counter < nr_counters-1; counter++)
 405                 weight *= sym->count[counter];
 406
 407         weight /= (sym->count[counter] + 1);
 408
 409         return weight;
 410 }
 411
 412 static long                     samples;
 413 static long                     userspace_samples;
 414 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
 415
 416 static void __list_insert_active_sym(struct sym_entry *syme)
 417 {
 418         list_add(&syme->node, &active_symbols);
 419 }
 420
 421 static void list_remove_active_sym(struct sym_entry *syme)
 422 {
 423         pthread_mutex_lock(&active_symbols_lock);
 424         list_del_init(&syme->node);
 425         pthread_mutex_unlock(&active_symbols_lock);
 426 }
 427
 428 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
 429 {
 430         struct rb_node **p = &tree->rb_node;
 431         struct rb_node *parent = NULL;
 432         struct sym_entry *iter;
 433
 434         while (*p != NULL) {
 435                 parent = *p;
 436                 iter = rb_entry(parent, struct sym_entry, rb_node);
 437
 438                 if (se->weight > iter->weight)
 439                         p = &(*p)->rb_left;
 440                 else
 441                         p = &(*p)->rb_right;
 442         }
 443
 444         rb_link_node(&se->rb_node, parent, p);
 445         rb_insert_color(&se->rb_node, tree);
 446 }
 447
 448 static void print_sym_table(void)
 449 {
 450         int printed = 0, j;
 451         int counter, snap = !display_weighted ? sym_counter : 0;
 452         float samples_per_sec = samples/delay_secs;
 453         float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
 454         float sum_ksamples = 0.0;
 455         struct sym_entry *syme, *n;
 456         struct rb_root tmp = RB_ROOT;
 457         struct rb_node *nd;
 458         int sym_width = 0, dso_width = 0, dso_short_width;
 459         const int win_width = winsize.ws_col - 1;
 460
 461         samples = userspace_samples = 0;
 462
 463         /* Sort the active symbols */
 464         pthread_mutex_lock(&active_symbols_lock);
 465         syme = list_entry(active_symbols.next, struct sym_entry, node);
 466         pthread_mutex_unlock(&active_symbols_lock);
 467
 468         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
 469                 syme->snap_count = syme->count[snap];
 470                 if (syme->snap_count != 0) {
 471
 472                         if ((hide_user_symbols &&
 473                              syme->origin == PERF_RECORD_MISC_USER) ||
 474                             (hide_kernel_symbols &&
 475                              syme->origin == PERF_RECORD_MISC_KERNEL)) {
 476                                 list_remove_active_sym(syme);
 477                                 continue;
 478                         }
 479                         syme->weight = sym_weight(syme);
 480                         rb_insert_active_sym(&tmp, syme);
 481                         sum_ksamples += syme->snap_count;
 482
 483                         for (j = 0; j < nr_counters; j++)
 484                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
 485                 } else
 486                         list_remove_active_sym(syme);
 487         }
 488
 489         puts(CONSOLE_CLEAR);
 490
 491         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 492         printf( "   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% [",
 493                 samples_per_sec,
 494                 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
 495
 496         if (nr_counters == 1 || !display_weighted) {
 497                 printf("%Ld", (u64)attrs[0].sample_period);
 498                 if (freq)
 499                         printf("Hz ");
 500                 else
 501                         printf(" ");
 502         }
 503
 504         if (!display_weighted)
 505                 printf("%s", event_name(sym_counter));
 506         else for (counter = 0; counter < nr_counters; counter++) {
 507                 if (counter)
 508                         printf("/");
 509
 510                 printf("%s", event_name(counter));
 511         }
 512
 513         printf( "], ");
 514
 515         if (target_pid != -1)
 516                 printf(" (target_pid: %d", target_pid);
 517         else
 518                 printf(" (all");
 519
 520         if (profile_cpu != -1)
 521                 printf(", cpu: %d)\n", profile_cpu);
 522         else {
 523                 if (target_pid != -1)
 524                         printf(")\n");
 525                 else
 526                         printf(", %d CPUs)\n", nr_cpus);
 527         }
 528
 529         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 530
 531         if (sym_filter_entry) {
 532                 show_details(sym_filter_entry);
 533                 return;
 534         }
 535
 536         /*
 537          * Find the longest symbol name that will be displayed
 538          */
 539         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
 540                 syme = rb_entry(nd, struct sym_entry, rb_node);
 541                 if (++printed > print_entries ||
 542                     (int)syme->snap_count < count_filter)
 543                         continue;
 544
 545                 if (syme->map->dso->long_name_len > dso_width)
 546                         dso_width = syme->map->dso->long_name_len;
 547
 548                 if (syme->map->dso->short_name_len > dso_short_width)
 549                         dso_short_width = syme->map->dso->short_name_len;
 550
 551                 if (syme->name_len > sym_width)
 552                         sym_width = syme->name_len;
 553         }
 554
 555         printed = 0;
 556
 557         if (sym_width + dso_width > winsize.ws_col - 29) {
 558                 dso_width = dso_short_width;
 559                 if (sym_width + dso_width > winsize.ws_col - 29)
 560                         sym_width = winsize.ws_col - dso_width - 29;
 561         }
 562         putchar('\n');
 563         if (nr_counters == 1)
 564                 printf("             samples  pcnt");
 565         else
 566                 printf("   weight    samples  pcnt");
 567
 568         if (verbose)
 569                 printf("         RIP       ");
 570         printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
 571         printf("   %s    _______ _____",
 572                nr_counters == 1 ? "      " : "______");
 573         if (verbose)
 574                 printf(" ________________");
 575         printf(" %-*.*s", sym_width, sym_width, graph_line);
 576         printf(" %-*.*s", dso_width, dso_width, graph_line);
 577         puts("\n");
 578
 579         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
 580                 struct symbol *sym;
 581                 double pcnt;
 582
 583                 syme = rb_entry(nd, struct sym_entry, rb_node);
 584                 sym = sym_entry__symbol(syme);
 585
 586                 if (++printed > print_entries || (int)syme->snap_count < count_filter)
 587                         continue;
 588
 589                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
 590                                          sum_ksamples));
 591
 592                 if (nr_counters == 1 || !display_weighted)
 593                         printf("%20.2f ", syme->weight);
 594                 else
 595                         printf("%9.1f %10ld ", syme->weight, syme->snap_count);
 596
 597                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
 598                 if (verbose)
 599                         printf(" %016llx", sym->start);
 600                 printf(" %-*.*s", sym_width, sym_width, sym->name);
 601                 printf(" %-*.*s\n", dso_width, dso_width,
 602                        dso_width >= syme->map->dso->long_name_len ?
 603                                         syme->map->dso->long_name :
 604                                         syme->map->dso->short_name);
 605         }
 606 }
 607
 608 static void prompt_integer(int *target, const char *msg)
 609 {
 610         char *buf = malloc(0), *p;
 611         size_t dummy = 0;
 612         int tmp;
 613
 614         fprintf(stdout, "\n%s: ", msg);
 615         if (getline(&buf, &dummy, stdin) < 0)
 616                 return;
 617
 618         p = strchr(buf, '\n');
 619         if (p)
 620                 *p = 0;
 621
 622         p = buf;
 623         while(*p) {
 624                 if (!isdigit(*p))
 625                         goto out_free;
 626                 p++;
 627         }
 628         tmp = strtoul(buf, NULL, 10);
 629         *target = tmp;
 630 out_free:
 631         free(buf);
 632 }
 633
 634 static void prompt_percent(int *target, const char *msg)
 635 {
 636         int tmp = 0;
 637
 638         prompt_integer(&tmp, msg);
 639         if (tmp >= 0 && tmp <= 100)
 640                 *target = tmp;
 641 }
 642
 643 static void prompt_symbol(struct sym_entry **target, const char *msg)
 644 {
 645         char *buf = malloc(0), *p;
 646         struct sym_entry *syme = *target, *n, *found = NULL;
 647         size_t dummy = 0;
 648
 649         /* zero counters of active symbol */
 650         if (syme) {
 651                 pthread_mutex_lock(&syme->src->lock);
 652                 __zero_source_counters(syme);
 653                 *target = NULL;
 654                 pthread_mutex_unlock(&syme->src->lock);
 655         }
 656
 657         fprintf(stdout, "\n%s: ", msg);
 658         if (getline(&buf, &dummy, stdin) < 0)
 659                 goto out_free;
 660
 661         p = strchr(buf, '\n');
 662         if (p)
 663                 *p = 0;
 664
 665         pthread_mutex_lock(&active_symbols_lock);
 666         syme = list_entry(active_symbols.next, struct sym_entry, node);
 667         pthread_mutex_unlock(&active_symbols_lock);
 668
 669         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
 670                 struct symbol *sym = sym_entry__symbol(syme);
 671
 672                 if (!strcmp(buf, sym->name)) {
 673                         found = syme;
 674                         break;
 675                 }
 676         }
 677
 678         if (!found) {
 679                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 680                 sleep(1);
 681                 return;
 682         } else
 683                 parse_source(found);
 684
 685 out_free:
 686         free(buf);
 687 }
 688
 689 static void print_mapped_keys(void)
 690 {
 691         char *name = NULL;
 692
 693         if (sym_filter_entry) {
 694                 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
 695                 name = sym->name;
 696         }
 697
 698         fprintf(stdout, "\nMapped keys:\n");
 699         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", delay_secs);
 700         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", print_entries);
 701
 702         if (nr_counters > 1)
 703                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(sym_counter));
 704
 705         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", count_filter);
 706
 707         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
 708         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 709         fprintf(stdout, "\t[S]     stop annotation.\n");
 710
 711         if (nr_counters > 1)
 712                 fprintf(stdout, "\t[w]     toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
 713
 714         fprintf(stdout,
 715                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 716                 hide_kernel_symbols ? "yes" : "no");
 717         fprintf(stdout,
 718                 "\t[U]     hide user symbols.               \t(%s)\n",
 719                 hide_user_symbols ? "yes" : "no");
 720         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", zero ? 1 : 0);
 721         fprintf(stdout, "\t[qQ]    quit.\n");
 722 }
 723
 724 static int key_mapped(int c)
 725 {
 726         switch (c) {
 727                 case 'd':
 728                 case 'e':
 729                 case 'f':
 730                 case 'z':
 731                 case 'q':
 732                 case 'Q':
 733                 case 'K':
 734                 case 'U':
 735                 case 'F':
 736                 case 's':
 737                 case 'S':
 738                         return 1;
 739                 case 'E':
 740                 case 'w':
 741                         return nr_counters > 1 ? 1 : 0;
 742                 default:
 743                         break;
 744         }
 745
 746         return 0;
 747 }
 748
 749 static void handle_keypress(int c)
 750 {
 751         if (!key_mapped(c)) {
 752                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 753                 struct termios tc, save;
 754
 755                 print_mapped_keys();
 756                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 757                 fflush(stdout);
 758
 759                 tcgetattr(0, &save);
 760                 tc = save;
 761                 tc.c_lflag &= ~(ICANON | ECHO);
 762                 tc.c_cc[VMIN] = 0;
 763                 tc.c_cc[VTIME] = 0;
 764                 tcsetattr(0, TCSANOW, &tc);
 765
 766                 poll(&stdin_poll, 1, -1);
 767                 c = getc(stdin);
 768
 769                 tcsetattr(0, TCSAFLUSH, &save);
 770                 if (!key_mapped(c))
 771                         return;
 772         }
 773
 774         switch (c) {
 775                 case 'd':
 776                         prompt_integer(&delay_secs, "Enter display delay");
 777                         if (delay_secs < 1)
 778                                 delay_secs = 1;
 779                         break;
 780                 case 'e':
 781                         prompt_integer(&print_entries, "Enter display entries (lines)");
 782                         if (print_entries == 0) {
 783                                 sig_winch_handler(SIGWINCH);
 784                                 signal(SIGWINCH, sig_winch_handler);
 785                         } else
 786                                 signal(SIGWINCH, SIG_DFL);
 787                         break;
 788                 case 'E':
 789                         if (nr_counters > 1) {
 790                                 int i;
 791
 792                                 fprintf(stderr, "\nAvailable events:");
 793                                 for (i = 0; i < nr_counters; i++)
 794                                         fprintf(stderr, "\n\t%d %s", i, event_name(i));
 795
 796                                 prompt_integer(&sym_counter, "Enter details event counter");
 797
 798                                 if (sym_counter >= nr_counters) {
 799                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
 800                                         sym_counter = 0;
 801                                         sleep(1);
 802                                 }
 803                         } else sym_counter = 0;
 804                         break;
 805                 case 'f':
 806                         prompt_integer(&count_filter, "Enter display event count filter");
 807                         break;
 808                 case 'F':
 809                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
 810                         break;
 811                 case 'K':
 812                         hide_kernel_symbols = !hide_kernel_symbols;
 813                         break;
 814                 case 'q':
 815                 case 'Q':
 816                         printf("exiting.\n");
 817                         if (dump_symtab)
 818                                 dsos__fprintf(stderr);
 819                         exit(0);
 820                 case 's':
 821                         prompt_symbol(&sym_filter_entry, "Enter details symbol");
 822                         break;
 823                 case 'S':
 824                         if (!sym_filter_entry)
 825                                 break;
 826                         else {
 827                                 struct sym_entry *syme = sym_filter_entry;
 828
 829                                 pthread_mutex_lock(&syme->src->lock);
 830                                 sym_filter_entry = NULL;
 831                                 __zero_source_counters(syme);
 832                                 pthread_mutex_unlock(&syme->src->lock);
 833                         }
 834                         break;
 835                 case 'U':
 836                         hide_user_symbols = !hide_user_symbols;
 837                         break;
 838                 case 'w':
 839                         display_weighted = ~display_weighted;
 840                         break;
 841                 case 'z':
 842                         zero = ~zero;
 843                         break;
 844                 default:
 845                         break;
 846         }
 847 }
 848
 849 static void *display_thread(void *arg __used)
 850 {
 851         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 852         struct termios tc, save;
 853         int delay_msecs, c;
 854
 855         tcgetattr(0, &save);
 856         tc = save;
 857         tc.c_lflag &= ~(ICANON | ECHO);
 858         tc.c_cc[VMIN] = 0;
 859         tc.c_cc[VTIME] = 0;
 860
 861 repeat:
 862         delay_msecs = delay_secs * 1000;
 863         tcsetattr(0, TCSANOW, &tc);
 864         /* trash return*/
 865         getc(stdin);
 866
 867         do {
 868                 print_sym_table();
 869         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
 870
 871         c = getc(stdin);
 872         tcsetattr(0, TCSAFLUSH, &save);
 873
 874         handle_keypress(c);
 875         goto repeat;
 876
 877         return NULL;
 878 }
 879
 880 /* Tag samples to be skipped. */
 881 static const char *skip_symbols[] = {
 882         "default_idle",
 883         "cpu_idle",
 884         "enter_idle",
 885         "exit_idle",
 886         "mwait_idle",
 887         "mwait_idle_with_hints",
 888         "poll_idle",
 889         "ppc64_runlatch_off",
 890         "pseries_dedicated_idle_sleep",
 891         NULL
 892 };
 893
 894 static int symbol_filter(struct map *map, struct symbol *sym)
 895 {
 896         struct sym_entry *syme;
 897         const char *name = sym->name;
 898         int i;
 899
 900         /*
 901          * ppc64 uses function descriptors and appends a '.' to the
 902          * start of every instruction address. Remove it.
 903          */
 904         if (name[0] == '.')
 905                 name++;
 906
 907         if (!strcmp(name, "_text") ||
 908             !strcmp(name, "_etext") ||
 909             !strcmp(name, "_sinittext") ||
 910             !strncmp("init_module", name, 11) ||
 911             !strncmp("cleanup_module", name, 14) ||
 912             strstr(name, "_text_start") ||
 913             strstr(name, "_text_end"))
 914                 return 1;
 915
 916         syme = symbol__priv(sym);
 917         syme->map = map;
 918         syme->src = NULL;
 919
 920         if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) {
 921                 /* schedule initial sym_filter_entry setup */
 922                 sym_filter_entry_sched = syme;
 923                 sym_filter = NULL;
 924         }
 925
 926         for (i = 0; skip_symbols[i]; i++) {
 927                 if (!strcmp(skip_symbols[i], name)) {
 928                         syme->skip = 1;
 929                         break;
 930                 }
 931         }
 932
 933         if (!syme->skip)
 934                 syme->name_len = strlen(sym->name);
 935
 936         return 0;
 937 }
 938
 939 static void event__process_sample(const event_t *self,
 940                                  struct perf_session *session, int counter)
 941 {
 942         u64 ip = self->ip.ip;
 943         struct sym_entry *syme;
 944         struct addr_location al;
 945         u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 946
 947         ++samples;
 948
 949         switch (origin) {
 950         case PERF_RECORD_MISC_USER:
 951                 ++userspace_samples;
 952                 if (hide_user_symbols)
 953                         return;
 954                 break;
 955         case PERF_RECORD_MISC_KERNEL:
 956                 if (hide_kernel_symbols)
 957                         return;
 958                 break;
 959         default:
 960                 return;
 961         }
 962
 963         if (event__preprocess_sample(self, session, &al, symbol_filter) < 0 ||
 964             al.filtered)
 965                 return;
 966
 967         if (al.sym == NULL) {
 968                 /*
 969                  * As we do lazy loading of symtabs we only will know if the
 970                  * specified vmlinux file is invalid when we actually have a
 971                  * hit in kernel space and then try to load it. So if we get
 972                  * here and there are _no_ symbols in the DSO backing the
 973                  * kernel map, bail out.
 974                  *
 975                  * We may never get here, for instance, if we use -K/
 976                  * --hide-kernel-symbols, even if the user specifies an
 977                  * invalid --vmlinux ;-)
 978                  */
 979                 if (al.map == session->vmlinux_maps[MAP__FUNCTION] &&
 980                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 981                         pr_err("The %s file can't be used\n",
 982                                symbol_conf.vmlinux_name);
 983                         exit(1);
 984                 }
 985
 986                 return;
 987         }
 988
 989         /* let's see, whether we need to install initial sym_filter_entry */
 990         if (sym_filter_entry_sched) {
 991                 sym_filter_entry = sym_filter_entry_sched;
 992                 sym_filter_entry_sched = NULL;
 993                 parse_source(sym_filter_entry);
 994         }
 995
 996         syme = symbol__priv(al.sym);
 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         }
1006 }
1007
1008 static int event__process(event_t *event, struct perf_session *session)
1009 {
1010         switch (event->header.type) {
1011         case PERF_RECORD_COMM:
1012                 event__process_comm(event, session);
1013                 break;
1014         case PERF_RECORD_MMAP:
1015                 event__process_mmap(event, session);
1016                 break;
1017         case PERF_RECORD_FORK:
1018         case PERF_RECORD_EXIT:
1019                 event__process_task(event, session);
1020                 break;
1021         default:
1022                 break;
1023         }
1024
1025         return 0;
1026 }
1027
1028 struct mmap_data {
1029         int                     counter;
1030         void                    *base;
1031         int                     mask;
1032         unsigned int            prev;
1033 };
1034
1035 static unsigned int mmap_read_head(struct mmap_data *md)
1036 {
1037         struct perf_event_mmap_page *pc = md->base;
1038         int head;
1039
1040         head = pc->data_head;
1041         rmb();
1042
1043         return head;
1044 }
1045
1046 static void perf_session__mmap_read_counter(struct perf_session *self,
1047                                             struct mmap_data *md)
1048 {
1049         unsigned int head = mmap_read_head(md);
1050         unsigned int old = md->prev;
1051         unsigned char *data = md->base + page_size;
1052         int diff;
1053
1054         /*
1055          * If we're further behind than half the buffer, there's a chance
1056          * the writer will bite our tail and mess up the samples under us.
1057          *
1058          * If we somehow ended up ahead of the head, we got messed up.
1059          *
1060          * In either case, truncate and restart at head.
1061          */
1062         diff = head - old;
1063         if (diff > md->mask / 2 || diff < 0) {
1064                 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
1065
1066                 /*
1067                  * head points to a known good entry, start there.
1068                  */
1069                 old = head;
1070         }
1071
1072         for (; old != head;) {
1073                 event_t *event = (event_t *)&data[old & md->mask];
1074
1075                 event_t event_copy;
1076
1077                 size_t size = event->header.size;
1078
1079                 /*
1080                  * Event straddles the mmap boundary -- header should always
1081                  * be inside due to u64 alignment of output.
1082                  */
1083                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
1084                         unsigned int offset = old;
1085                         unsigned int len = min(sizeof(*event), size), cpy;
1086                         void *dst = &event_copy;
1087
1088                         do {
1089                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
1090                                 memcpy(dst, &data[offset & md->mask], cpy);
1091                                 offset += cpy;
1092                                 dst += cpy;
1093                                 len -= cpy;
1094                         } while (len);
1095
1096                         event = &event_copy;
1097                 }
1098
1099                 if (event->header.type == PERF_RECORD_SAMPLE)
1100                         event__process_sample(event, self, md->counter);
1101                 else
1102                         event__process(event, self);
1103                 old += size;
1104         }
1105
1106         md->prev = old;
1107 }
1108
1109 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
1110 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
1111
1112 static void perf_session__mmap_read(struct perf_session *self)
1113 {
1114         int i, counter;
1115
1116         for (i = 0; i < nr_cpus; i++) {
1117                 for (counter = 0; counter < nr_counters; counter++)
1118                         perf_session__mmap_read_counter(self, &mmap_array[i][counter]);
1119         }
1120 }
1121
1122 int nr_poll;
1123 int group_fd;
1124
1125 static void start_counter(int i, int counter)
1126 {
1127         struct perf_event_attr *attr;
1128         int cpu;
1129
1130         cpu = profile_cpu;
1131         if (target_pid == -1 && profile_cpu == -1)
1132                 cpu = cpumap[i];
1133
1134         attr = attrs + counter;
1135
1136         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1137
1138         if (freq) {
1139                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
1140                 attr->freq              = 1;
1141                 attr->sample_freq       = freq;
1142         }
1143
1144         attr->inherit           = (cpu < 0) && inherit;
1145         attr->mmap              = 1;
1146
1147 try_again:
1148         fd[i][counter] = sys_perf_event_open(attr, target_pid, cpu, group_fd, 0);
1149
1150         if (fd[i][counter] < 0) {
1151                 int err = errno;
1152
1153                 if (err == EPERM || err == EACCES)
1154                         die("No permission - are you root?\n");
1155                 /*
1156                  * If it's cycles then fall back to hrtimer
1157                  * based cpu-clock-tick sw counter, which
1158                  * is always available even if no PMU support:
1159                  */
1160                 if (attr->type == PERF_TYPE_HARDWARE
1161                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1162
1163                         if (verbose)
1164                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
1165
1166                         attr->type = PERF_TYPE_SOFTWARE;
1167                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
1168                         goto try_again;
1169                 }
1170                 printf("\n");
1171                 error("perfcounter syscall returned with %d (%s)\n",
1172                         fd[i][counter], strerror(err));
1173                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1174                 exit(-1);
1175         }
1176         assert(fd[i][counter] >= 0);
1177         fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
1178
1179         /*
1180          * First counter acts as the group leader:
1181          */
1182         if (group && group_fd == -1)
1183                 group_fd = fd[i][counter];
1184
1185         event_array[nr_poll].fd = fd[i][counter];
1186         event_array[nr_poll].events = POLLIN;
1187         nr_poll++;
1188
1189         mmap_array[i][counter].counter = counter;
1190         mmap_array[i][counter].prev = 0;
1191         mmap_array[i][counter].mask = mmap_pages*page_size - 1;
1192         mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
1193                         PROT_READ, MAP_SHARED, fd[i][counter], 0);
1194         if (mmap_array[i][counter].base == MAP_FAILED)
1195                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1196 }
1197
1198 static int __cmd_top(void)
1199 {
1200         pthread_t thread;
1201         int i, counter;
1202         int ret;
1203         /*
1204          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1205          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1206          */
1207         struct perf_session *session = perf_session__new(NULL, O_WRONLY, false);
1208         if (session == NULL)
1209                 return -ENOMEM;
1210
1211         if (target_pid != -1)
1212                 event__synthesize_thread(target_pid, event__process, session);
1213         else
1214                 event__synthesize_threads(event__process, session);
1215
1216         for (i = 0; i < nr_cpus; i++) {
1217                 group_fd = -1;
1218                 for (counter = 0; counter < nr_counters; counter++)
1219                         start_counter(i, counter);
1220         }
1221
1222         /* Wait for a minimal set of events before starting the snapshot */
1223         poll(event_array, nr_poll, 100);
1224
1225         perf_session__mmap_read(session);
1226
1227         if (pthread_create(&thread, NULL, display_thread, NULL)) {
1228                 printf("Could not create display thread.\n");
1229                 exit(-1);
1230         }
1231
1232         if (realtime_prio) {
1233                 struct sched_param param;
1234
1235                 param.sched_priority = realtime_prio;
1236                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1237                         printf("Could not set realtime priority.\n");
1238                         exit(-1);
1239                 }
1240         }
1241
1242         while (1) {
1243                 int hits = samples;
1244
1245                 perf_session__mmap_read(session);
1246
1247                 if (hits == samples)
1248                         ret = poll(event_array, nr_poll, 100);
1249         }
1250
1251         return 0;
1252 }
1253
1254 static const char * const top_usage[] = {
1255         "perf top [<options>]",
1256         NULL
1257 };
1258
1259 static const struct option options[] = {
1260         OPT_CALLBACK('e', "event", NULL, "event",
1261                      "event selector. use 'perf list' to list available events",
1262                      parse_events),
1263         OPT_INTEGER('c', "count", &default_interval,
1264                     "event period to sample"),
1265         OPT_INTEGER('p', "pid", &target_pid,
1266                     "profile events on existing pid"),
1267         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1268                             "system-wide collection from all CPUs"),
1269         OPT_INTEGER('C', "CPU", &profile_cpu,
1270                     "CPU to profile on"),
1271         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1272                    "file", "vmlinux pathname"),
1273         OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols,
1274                     "hide kernel symbols"),
1275         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
1276                     "number of mmap data pages"),
1277         OPT_INTEGER('r', "realtime", &realtime_prio,
1278                     "collect data with this RT SCHED_FIFO priority"),
1279         OPT_INTEGER('d', "delay", &delay_secs,
1280                     "number of seconds to delay between refreshes"),
1281         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1282                             "dump the symbol table used for profiling"),
1283         OPT_INTEGER('f', "count-filter", &count_filter,
1284                     "only display functions with more events than this"),
1285         OPT_BOOLEAN('g', "group", &group,
1286                             "put the counters into a counter group"),
1287         OPT_BOOLEAN('i', "inherit", &inherit,
1288                     "child tasks inherit counters"),
1289         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1290                     "symbol to annotate"),
1291         OPT_BOOLEAN('z', "zero", &zero,
1292                     "zero history across updates"),
1293         OPT_INTEGER('F', "freq", &freq,
1294                     "profile at this frequency"),
1295         OPT_INTEGER('E', "entries", &print_entries,
1296                     "display this many functions"),
1297         OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols,
1298                     "hide user symbols"),
1299         OPT_BOOLEAN('v', "verbose", &verbose,
1300                     "be more verbose (show counter open errors, etc)"),
1301         OPT_END()
1302 };
1303
1304 int cmd_top(int argc, const char **argv, const char *prefix __used)
1305 {
1306         int counter;
1307
1308         page_size = sysconf(_SC_PAGE_SIZE);
1309
1310         argc = parse_options(argc, argv, options, top_usage, 0);
1311         if (argc)
1312                 usage_with_options(top_usage, options);
1313
1314         /* CPU and PID are mutually exclusive */
1315         if (target_pid != -1 && profile_cpu != -1) {
1316                 printf("WARNING: PID switch overriding CPU\n");
1317                 sleep(1);
1318                 profile_cpu = -1;
1319         }
1320
1321         if (!nr_counters)
1322                 nr_counters = 1;
1323
1324         symbol_conf.priv_size = (sizeof(struct sym_entry) +
1325                                  (nr_counters + 1) * sizeof(unsigned long));
1326
1327         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1328         if (symbol__init() < 0)
1329                 return -1;
1330
1331         if (delay_secs < 1)
1332                 delay_secs = 1;
1333
1334         /*
1335          * User specified count overrides default frequency.
1336          */
1337         if (default_interval)
1338                 freq = 0;
1339         else if (freq) {
1340                 default_interval = freq;
1341         } else {
1342                 fprintf(stderr, "frequency and count are zero, aborting\n");
1343                 exit(EXIT_FAILURE);
1344         }
1345
1346         /*
1347          * Fill in the ones not specifically initialized via -c:
1348          */
1349         for (counter = 0; counter < nr_counters; counter++) {
1350                 if (attrs[counter].sample_period)
1351                         continue;
1352
1353                 attrs[counter].sample_period = default_interval;
1354         }
1355
1356         if (target_pid != -1 || profile_cpu != -1)
1357                 nr_cpus = 1;
1358         else
1359                 nr_cpus = read_cpu_map();
1360
1361         get_term_dimensions(&winsize);
1362         if (print_entries == 0) {
1363                 update_print_entries(&winsize);
1364                 signal(SIGWINCH, sig_winch_handler);
1365         }
1366
1367         return __cmd_top();
1368 }