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/evsel.h" 25 #include "util/session.h" 26 #include "util/symbol.h" 27 #include "util/thread.h" 28 #include "util/util.h" 29 #include <linux/rbtree.h> 30 #include "util/parse-options.h" 31 #include "util/parse-events.h" 32 #include "util/cpumap.h" 33 #include "util/xyarray.h" 34 35 #include "util/debug.h" 36 37 #include <assert.h> 38 #include <fcntl.h> 39 40 #include <stdio.h> 41 #include <termios.h> 42 #include <unistd.h> 43 44 #include <errno.h> 45 #include <time.h> 46 #include <sched.h> 47 #include <pthread.h> 48 49 #include <sys/syscall.h> 50 #include <sys/ioctl.h> 51 #include <sys/poll.h> 52 #include <sys/prctl.h> 53 #include <sys/wait.h> 54 #include <sys/uio.h> 55 #include <sys/mman.h> 56 57 #include <linux/unistd.h> 58 #include <linux/types.h> 59 60 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y)) 61 62 static bool system_wide = false; 63 64 static int default_interval = 0; 65 66 static int count_filter = 5; 67 static int print_entries; 68 69 static int target_pid = -1; 70 static int target_tid = -1; 71 static struct thread_map *threads; 72 static bool inherit = false; 73 static struct cpu_map *cpus; 74 static int realtime_prio = 0; 75 static bool group = false; 76 static unsigned int page_size; 77 static unsigned int mmap_pages = 16; 78 static int freq = 1000; /* 1 KHz */ 79 80 static int delay_secs = 2; 81 static bool zero = false; 82 static bool dump_symtab = false; 83 84 static bool hide_kernel_symbols = false; 85 static bool hide_user_symbols = false; 86 static struct winsize winsize; 87 88 /* 89 * Source 90 */ 91 92 struct source_line { 93 u64 eip; 94 unsigned long count[MAX_COUNTERS]; 95 char *line; 96 struct source_line *next; 97 }; 98 99 static const char *sym_filter = NULL; 100 struct sym_entry *sym_filter_entry = NULL; 101 struct sym_entry *sym_filter_entry_sched = NULL; 102 static int sym_pcnt_filter = 5; 103 static int sym_counter = 0; 104 static struct perf_evsel *sym_evsel = NULL; 105 static int display_weighted = -1; 106 static const char *cpu_list; 107 108 /* 109 * Symbols 110 */ 111 112 struct sym_entry_source { 113 struct source_line *source; 114 struct source_line *lines; 115 struct source_line **lines_tail; 116 pthread_mutex_t lock; 117 }; 118 119 struct sym_entry { 120 struct rb_node rb_node; 121 struct list_head node; 122 unsigned long snap_count; 123 double weight; 124 int skip; 125 u16 name_len; 126 u8 origin; 127 struct map *map; 128 struct sym_entry_source *src; 129 unsigned long count[0]; 130 }; 131 132 /* 133 * Source functions 134 */ 135 136 static inline struct symbol *sym_entry__symbol(struct sym_entry *self) 137 { 138 return ((void *)self) + symbol_conf.priv_size; 139 } 140 141 void get_term_dimensions(struct winsize *ws) 142 { 143 char *s = getenv("LINES"); 144 145 if (s != NULL) { 146 ws->ws_row = atoi(s); 147 s = getenv("COLUMNS"); 148 if (s != NULL) { 149 ws->ws_col = atoi(s); 150 if (ws->ws_row && ws->ws_col) 151 return; 152 } 153 } 154 #ifdef TIOCGWINSZ 155 if (ioctl(1, TIOCGWINSZ, ws) == 0 && 156 ws->ws_row && ws->ws_col) 157 return; 158 #endif 159 ws->ws_row = 25; 160 ws->ws_col = 80; 161 } 162 163 static void update_print_entries(struct winsize *ws) 164 { 165 print_entries = ws->ws_row; 166 167 if (print_entries > 9) 168 print_entries -= 9; 169 } 170 171 static void sig_winch_handler(int sig __used) 172 { 173 get_term_dimensions(&winsize); 174 update_print_entries(&winsize); 175 } 176 177 static int parse_source(struct sym_entry *syme) 178 { 179 struct symbol *sym; 180 struct sym_entry_source *source; 181 struct map *map; 182 FILE *file; 183 char command[PATH_MAX*2]; 184 const char *path; 185 u64 len; 186 187 if (!syme) 188 return -1; 189 190 sym = sym_entry__symbol(syme); 191 map = syme->map; 192 193 /* 194 * We can't annotate with just /proc/kallsyms 195 */ 196 if (map->dso->origin == DSO__ORIG_KERNEL) 197 return -1; 198 199 if (syme->src == NULL) { 200 syme->src = zalloc(sizeof(*source)); 201 if (syme->src == NULL) 202 return -1; 203 pthread_mutex_init(&syme->src->lock, NULL); 204 } 205 206 source = syme->src; 207 208 if (source->lines) { 209 pthread_mutex_lock(&source->lock); 210 goto out_assign; 211 } 212 path = map->dso->long_name; 213 214 len = sym->end - sym->start; 215 216 sprintf(command, 217 "objdump --start-address=%#0*Lx --stop-address=%#0*Lx -dS %s", 218 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->start), 219 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->end), path); 220 221 file = popen(command, "r"); 222 if (!file) 223 return -1; 224 225 pthread_mutex_lock(&source->lock); 226 source->lines_tail = &source->lines; 227 while (!feof(file)) { 228 struct source_line *src; 229 size_t dummy = 0; 230 char *c, *sep; 231 232 src = malloc(sizeof(struct source_line)); 233 assert(src != NULL); 234 memset(src, 0, sizeof(struct source_line)); 235 236 if (getline(&src->line, &dummy, file) < 0) 237 break; 238 if (!src->line) 239 break; 240 241 c = strchr(src->line, '\n'); 242 if (c) 243 *c = 0; 244 245 src->next = NULL; 246 *source->lines_tail = src; 247 source->lines_tail = &src->next; 248 249 src->eip = strtoull(src->line, &sep, 16); 250 if (*sep == ':') 251 src->eip = map__objdump_2ip(map, src->eip); 252 else /* this line has no ip info (e.g. source line) */ 253 src->eip = 0; 254 } 255 pclose(file); 256 out_assign: 257 sym_filter_entry = syme; 258 pthread_mutex_unlock(&source->lock); 259 return 0; 260 } 261 262 static void __zero_source_counters(struct sym_entry *syme) 263 { 264 int i; 265 struct source_line *line; 266 267 line = syme->src->lines; 268 while (line) { 269 for (i = 0; i < nr_counters; i++) 270 line->count[i] = 0; 271 line = line->next; 272 } 273 } 274 275 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip) 276 { 277 struct source_line *line; 278 279 if (syme != sym_filter_entry) 280 return; 281 282 if (pthread_mutex_trylock(&syme->src->lock)) 283 return; 284 285 if (syme->src == NULL || syme->src->source == NULL) 286 goto out_unlock; 287 288 for (line = syme->src->lines; line; line = line->next) { 289 /* skip lines without IP info */ 290 if (line->eip == 0) 291 continue; 292 if (line->eip == ip) { 293 line->count[counter]++; 294 break; 295 } 296 if (line->eip > ip) 297 break; 298 } 299 out_unlock: 300 pthread_mutex_unlock(&syme->src->lock); 301 } 302 303 #define PATTERN_LEN (BITS_PER_LONG / 4 + 2) 304 305 static void lookup_sym_source(struct sym_entry *syme) 306 { 307 struct symbol *symbol = sym_entry__symbol(syme); 308 struct source_line *line; 309 char pattern[PATTERN_LEN + 1]; 310 311 sprintf(pattern, "%0*Lx <", BITS_PER_LONG / 4, 312 map__rip_2objdump(syme->map, symbol->start)); 313 314 pthread_mutex_lock(&syme->src->lock); 315 for (line = syme->src->lines; line; line = line->next) { 316 if (memcmp(line->line, pattern, PATTERN_LEN) == 0) { 317 syme->src->source = line; 318 break; 319 } 320 } 321 pthread_mutex_unlock(&syme->src->lock); 322 } 323 324 static void show_lines(struct source_line *queue, int count, int total) 325 { 326 int i; 327 struct source_line *line; 328 329 line = queue; 330 for (i = 0; i < count; i++) { 331 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total; 332 333 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line); 334 line = line->next; 335 } 336 } 337 338 #define TRACE_COUNT 3 339 340 static void show_details(struct sym_entry *syme) 341 { 342 struct symbol *symbol; 343 struct source_line *line; 344 struct source_line *line_queue = NULL; 345 int displayed = 0; 346 int line_queue_count = 0, total = 0, more = 0; 347 348 if (!syme) 349 return; 350 351 if (!syme->src->source) 352 lookup_sym_source(syme); 353 354 if (!syme->src->source) 355 return; 356 357 symbol = sym_entry__symbol(syme); 358 printf("Showing %s for %s\n", event_name(sym_evsel), symbol->name); 359 printf(" Events Pcnt (>=%d%%)\n", sym_pcnt_filter); 360 361 pthread_mutex_lock(&syme->src->lock); 362 line = syme->src->source; 363 while (line) { 364 total += line->count[sym_counter]; 365 line = line->next; 366 } 367 368 line = syme->src->source; 369 while (line) { 370 float pcnt = 0.0; 371 372 if (!line_queue_count) 373 line_queue = line; 374 line_queue_count++; 375 376 if (line->count[sym_counter]) 377 pcnt = 100.0 * line->count[sym_counter] / (float)total; 378 if (pcnt >= (float)sym_pcnt_filter) { 379 if (displayed <= print_entries) 380 show_lines(line_queue, line_queue_count, total); 381 else more++; 382 displayed += line_queue_count; 383 line_queue_count = 0; 384 line_queue = NULL; 385 } else if (line_queue_count > TRACE_COUNT) { 386 line_queue = line_queue->next; 387 line_queue_count--; 388 } 389 390 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8; 391 line = line->next; 392 } 393 pthread_mutex_unlock(&syme->src->lock); 394 if (more) 395 printf("%d lines not displayed, maybe increase display entries [e]\n", more); 396 } 397 398 /* 399 * Symbols will be added here in event__process_sample and will get out 400 * after decayed. 401 */ 402 static LIST_HEAD(active_symbols); 403 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER; 404 405 /* 406 * Ordering weight: count-1 * count-2 * ... / count-n 407 */ 408 static double sym_weight(const struct sym_entry *sym) 409 { 410 double weight = sym->snap_count; 411 int counter; 412 413 if (!display_weighted) 414 return weight; 415 416 for (counter = 1; counter < nr_counters-1; counter++) 417 weight *= sym->count[counter]; 418 419 weight /= (sym->count[counter] + 1); 420 421 return weight; 422 } 423 424 static long samples; 425 static long kernel_samples, us_samples; 426 static long exact_samples; 427 static long guest_us_samples, guest_kernel_samples; 428 static const char CONSOLE_CLEAR[] = "[H[2J"; 429 430 static void __list_insert_active_sym(struct sym_entry *syme) 431 { 432 list_add(&syme->node, &active_symbols); 433 } 434 435 static void list_remove_active_sym(struct sym_entry *syme) 436 { 437 pthread_mutex_lock(&active_symbols_lock); 438 list_del_init(&syme->node); 439 pthread_mutex_unlock(&active_symbols_lock); 440 } 441 442 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se) 443 { 444 struct rb_node **p = &tree->rb_node; 445 struct rb_node *parent = NULL; 446 struct sym_entry *iter; 447 448 while (*p != NULL) { 449 parent = *p; 450 iter = rb_entry(parent, struct sym_entry, rb_node); 451 452 if (se->weight > iter->weight) 453 p = &(*p)->rb_left; 454 else 455 p = &(*p)->rb_right; 456 } 457 458 rb_link_node(&se->rb_node, parent, p); 459 rb_insert_color(&se->rb_node, tree); 460 } 461 462 static void print_sym_table(void) 463 { 464 int printed = 0, j; 465 struct perf_evsel *counter; 466 int snap = !display_weighted ? sym_counter : 0; 467 float samples_per_sec = samples/delay_secs; 468 float ksamples_per_sec = kernel_samples/delay_secs; 469 float us_samples_per_sec = (us_samples)/delay_secs; 470 float guest_kernel_samples_per_sec = (guest_kernel_samples)/delay_secs; 471 float guest_us_samples_per_sec = (guest_us_samples)/delay_secs; 472 float esamples_percent = (100.0*exact_samples)/samples; 473 float sum_ksamples = 0.0; 474 struct sym_entry *syme, *n; 475 struct rb_root tmp = RB_ROOT; 476 struct rb_node *nd; 477 int sym_width = 0, dso_width = 0, dso_short_width = 0; 478 const int win_width = winsize.ws_col - 1; 479 480 samples = us_samples = kernel_samples = exact_samples = 0; 481 guest_kernel_samples = guest_us_samples = 0; 482 483 /* Sort the active symbols */ 484 pthread_mutex_lock(&active_symbols_lock); 485 syme = list_entry(active_symbols.next, struct sym_entry, node); 486 pthread_mutex_unlock(&active_symbols_lock); 487 488 list_for_each_entry_safe_from(syme, n, &active_symbols, node) { 489 syme->snap_count = syme->count[snap]; 490 if (syme->snap_count != 0) { 491 492 if ((hide_user_symbols && 493 syme->origin == PERF_RECORD_MISC_USER) || 494 (hide_kernel_symbols && 495 syme->origin == PERF_RECORD_MISC_KERNEL)) { 496 list_remove_active_sym(syme); 497 continue; 498 } 499 syme->weight = sym_weight(syme); 500 rb_insert_active_sym(&tmp, syme); 501 sum_ksamples += syme->snap_count; 502 503 for (j = 0; j < nr_counters; j++) 504 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8; 505 } else 506 list_remove_active_sym(syme); 507 } 508 509 puts(CONSOLE_CLEAR); 510 511 printf("%-*.*s\n", win_width, win_width, graph_dotted_line); 512 if (!perf_guest) { 513 printf(" PerfTop:%8.0f irqs/sec kernel:%4.1f%%" 514 " exact: %4.1f%% [", 515 samples_per_sec, 516 100.0 - (100.0 * ((samples_per_sec - ksamples_per_sec) / 517 samples_per_sec)), 518 esamples_percent); 519 } else { 520 printf(" PerfTop:%8.0f irqs/sec kernel:%4.1f%% us:%4.1f%%" 521 " guest kernel:%4.1f%% guest us:%4.1f%%" 522 " exact: %4.1f%% [", 523 samples_per_sec, 524 100.0 - (100.0 * ((samples_per_sec-ksamples_per_sec) / 525 samples_per_sec)), 526 100.0 - (100.0 * ((samples_per_sec-us_samples_per_sec) / 527 samples_per_sec)), 528 100.0 - (100.0 * ((samples_per_sec - 529 guest_kernel_samples_per_sec) / 530 samples_per_sec)), 531 100.0 - (100.0 * ((samples_per_sec - 532 guest_us_samples_per_sec) / 533 samples_per_sec)), 534 esamples_percent); 535 } 536 537 if (nr_counters == 1 || !display_weighted) { 538 struct perf_evsel *first; 539 first = list_entry(evsel_list.next, struct perf_evsel, node); 540 printf("%Ld", first->attr.sample_period); 541 if (freq) 542 printf("Hz "); 543 else 544 printf(" "); 545 } 546 547 if (!display_weighted) 548 printf("%s", event_name(sym_evsel)); 549 else list_for_each_entry(counter, &evsel_list, node) { 550 if (counter->idx) 551 printf("/"); 552 553 printf("%s", event_name(counter)); 554 } 555 556 printf( "], "); 557 558 if (target_pid != -1) 559 printf(" (target_pid: %d", target_pid); 560 else if (target_tid != -1) 561 printf(" (target_tid: %d", target_tid); 562 else 563 printf(" (all"); 564 565 if (cpu_list) 566 printf(", CPU%s: %s)\n", cpus->nr > 1 ? "s" : "", cpu_list); 567 else { 568 if (target_tid != -1) 569 printf(")\n"); 570 else 571 printf(", %d CPU%s)\n", cpus->nr, cpus->nr > 1 ? "s" : ""); 572 } 573 574 printf("%-*.*s\n", win_width, win_width, graph_dotted_line); 575 576 if (sym_filter_entry) { 577 show_details(sym_filter_entry); 578 return; 579 } 580 581 /* 582 * Find the longest symbol name that will be displayed 583 */ 584 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) { 585 syme = rb_entry(nd, struct sym_entry, rb_node); 586 if (++printed > print_entries || 587 (int)syme->snap_count < count_filter) 588 continue; 589 590 if (syme->map->dso->long_name_len > dso_width) 591 dso_width = syme->map->dso->long_name_len; 592 593 if (syme->map->dso->short_name_len > dso_short_width) 594 dso_short_width = syme->map->dso->short_name_len; 595 596 if (syme->name_len > sym_width) 597 sym_width = syme->name_len; 598 } 599 600 printed = 0; 601 602 if (sym_width + dso_width > winsize.ws_col - 29) { 603 dso_width = dso_short_width; 604 if (sym_width + dso_width > winsize.ws_col - 29) 605 sym_width = winsize.ws_col - dso_width - 29; 606 } 607 putchar('\n'); 608 if (nr_counters == 1) 609 printf(" samples pcnt"); 610 else 611 printf(" weight samples pcnt"); 612 613 if (verbose) 614 printf(" RIP "); 615 printf(" %-*.*s DSO\n", sym_width, sym_width, "function"); 616 printf(" %s _______ _____", 617 nr_counters == 1 ? " " : "______"); 618 if (verbose) 619 printf(" ________________"); 620 printf(" %-*.*s", sym_width, sym_width, graph_line); 621 printf(" %-*.*s", dso_width, dso_width, graph_line); 622 puts("\n"); 623 624 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) { 625 struct symbol *sym; 626 double pcnt; 627 628 syme = rb_entry(nd, struct sym_entry, rb_node); 629 sym = sym_entry__symbol(syme); 630 if (++printed > print_entries || (int)syme->snap_count < count_filter) 631 continue; 632 633 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) / 634 sum_ksamples)); 635 636 if (nr_counters == 1 || !display_weighted) 637 printf("%20.2f ", syme->weight); 638 else 639 printf("%9.1f %10ld ", syme->weight, syme->snap_count); 640 641 percent_color_fprintf(stdout, "%4.1f%%", pcnt); 642 if (verbose) 643 printf(" %016llx", sym->start); 644 printf(" %-*.*s", sym_width, sym_width, sym->name); 645 printf(" %-*.*s\n", dso_width, dso_width, 646 dso_width >= syme->map->dso->long_name_len ? 647 syme->map->dso->long_name : 648 syme->map->dso->short_name); 649 } 650 } 651 652 static void prompt_integer(int *target, const char *msg) 653 { 654 char *buf = malloc(0), *p; 655 size_t dummy = 0; 656 int tmp; 657 658 fprintf(stdout, "\n%s: ", msg); 659 if (getline(&buf, &dummy, stdin) < 0) 660 return; 661 662 p = strchr(buf, '\n'); 663 if (p) 664 *p = 0; 665 666 p = buf; 667 while(*p) { 668 if (!isdigit(*p)) 669 goto out_free; 670 p++; 671 } 672 tmp = strtoul(buf, NULL, 10); 673 *target = tmp; 674 out_free: 675 free(buf); 676 } 677 678 static void prompt_percent(int *target, const char *msg) 679 { 680 int tmp = 0; 681 682 prompt_integer(&tmp, msg); 683 if (tmp >= 0 && tmp <= 100) 684 *target = tmp; 685 } 686 687 static void prompt_symbol(struct sym_entry **target, const char *msg) 688 { 689 char *buf = malloc(0), *p; 690 struct sym_entry *syme = *target, *n, *found = NULL; 691 size_t dummy = 0; 692 693 /* zero counters of active symbol */ 694 if (syme) { 695 pthread_mutex_lock(&syme->src->lock); 696 __zero_source_counters(syme); 697 *target = NULL; 698 pthread_mutex_unlock(&syme->src->lock); 699 } 700 701 fprintf(stdout, "\n%s: ", msg); 702 if (getline(&buf, &dummy, stdin) < 0) 703 goto out_free; 704 705 p = strchr(buf, '\n'); 706 if (p) 707 *p = 0; 708 709 pthread_mutex_lock(&active_symbols_lock); 710 syme = list_entry(active_symbols.next, struct sym_entry, node); 711 pthread_mutex_unlock(&active_symbols_lock); 712 713 list_for_each_entry_safe_from(syme, n, &active_symbols, node) { 714 struct symbol *sym = sym_entry__symbol(syme); 715 716 if (!strcmp(buf, sym->name)) { 717 found = syme; 718 break; 719 } 720 } 721 722 if (!found) { 723 fprintf(stderr, "Sorry, %s is not active.\n", buf); 724 sleep(1); 725 return; 726 } else 727 parse_source(found); 728 729 out_free: 730 free(buf); 731 } 732 733 static void print_mapped_keys(void) 734 { 735 char *name = NULL; 736 737 if (sym_filter_entry) { 738 struct symbol *sym = sym_entry__symbol(sym_filter_entry); 739 name = sym->name; 740 } 741 742 fprintf(stdout, "\nMapped keys:\n"); 743 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", delay_secs); 744 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", print_entries); 745 746 if (nr_counters > 1) 747 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", event_name(sym_evsel)); 748 749 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", count_filter); 750 751 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter); 752 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL"); 753 fprintf(stdout, "\t[S] stop annotation.\n"); 754 755 if (nr_counters > 1) 756 fprintf(stdout, "\t[w] toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0); 757 758 fprintf(stdout, 759 "\t[K] hide kernel_symbols symbols. \t(%s)\n", 760 hide_kernel_symbols ? "yes" : "no"); 761 fprintf(stdout, 762 "\t[U] hide user symbols. \t(%s)\n", 763 hide_user_symbols ? "yes" : "no"); 764 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", zero ? 1 : 0); 765 fprintf(stdout, "\t[qQ] quit.\n"); 766 } 767 768 static int key_mapped(int c) 769 { 770 switch (c) { 771 case 'd': 772 case 'e': 773 case 'f': 774 case 'z': 775 case 'q': 776 case 'Q': 777 case 'K': 778 case 'U': 779 case 'F': 780 case 's': 781 case 'S': 782 return 1; 783 case 'E': 784 case 'w': 785 return nr_counters > 1 ? 1 : 0; 786 default: 787 break; 788 } 789 790 return 0; 791 } 792 793 static void handle_keypress(struct perf_session *session, int c) 794 { 795 if (!key_mapped(c)) { 796 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN }; 797 struct termios tc, save; 798 799 print_mapped_keys(); 800 fprintf(stdout, "\nEnter selection, or unmapped key to continue: "); 801 fflush(stdout); 802 803 tcgetattr(0, &save); 804 tc = save; 805 tc.c_lflag &= ~(ICANON | ECHO); 806 tc.c_cc[VMIN] = 0; 807 tc.c_cc[VTIME] = 0; 808 tcsetattr(0, TCSANOW, &tc); 809 810 poll(&stdin_poll, 1, -1); 811 c = getc(stdin); 812 813 tcsetattr(0, TCSAFLUSH, &save); 814 if (!key_mapped(c)) 815 return; 816 } 817 818 switch (c) { 819 case 'd': 820 prompt_integer(&delay_secs, "Enter display delay"); 821 if (delay_secs < 1) 822 delay_secs = 1; 823 break; 824 case 'e': 825 prompt_integer(&print_entries, "Enter display entries (lines)"); 826 if (print_entries == 0) { 827 sig_winch_handler(SIGWINCH); 828 signal(SIGWINCH, sig_winch_handler); 829 } else 830 signal(SIGWINCH, SIG_DFL); 831 break; 832 case 'E': 833 if (nr_counters > 1) { 834 fprintf(stderr, "\nAvailable events:"); 835 836 list_for_each_entry(sym_evsel, &evsel_list, node) 837 fprintf(stderr, "\n\t%d %s", sym_evsel->idx, event_name(sym_evsel)); 838 839 prompt_integer(&sym_counter, "Enter details event counter"); 840 841 if (sym_counter >= nr_counters) { 842 sym_evsel = list_entry(evsel_list.next, struct perf_evsel, node); 843 sym_counter = 0; 844 fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(sym_evsel)); 845 sleep(1); 846 break; 847 } 848 list_for_each_entry(sym_evsel, &evsel_list, node) 849 if (sym_evsel->idx == sym_counter) 850 break; 851 } else sym_counter = 0; 852 break; 853 case 'f': 854 prompt_integer(&count_filter, "Enter display event count filter"); 855 break; 856 case 'F': 857 prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)"); 858 break; 859 case 'K': 860 hide_kernel_symbols = !hide_kernel_symbols; 861 break; 862 case 'q': 863 case 'Q': 864 printf("exiting.\n"); 865 if (dump_symtab) 866 perf_session__fprintf_dsos(session, stderr); 867 exit(0); 868 case 's': 869 prompt_symbol(&sym_filter_entry, "Enter details symbol"); 870 break; 871 case 'S': 872 if (!sym_filter_entry) 873 break; 874 else { 875 struct sym_entry *syme = sym_filter_entry; 876 877 pthread_mutex_lock(&syme->src->lock); 878 sym_filter_entry = NULL; 879 __zero_source_counters(syme); 880 pthread_mutex_unlock(&syme->src->lock); 881 } 882 break; 883 case 'U': 884 hide_user_symbols = !hide_user_symbols; 885 break; 886 case 'w': 887 display_weighted = ~display_weighted; 888 break; 889 case 'z': 890 zero = !zero; 891 break; 892 default: 893 break; 894 } 895 } 896 897 static void *display_thread(void *arg __used) 898 { 899 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN }; 900 struct termios tc, save; 901 int delay_msecs, c; 902 struct perf_session *session = (struct perf_session *) arg; 903 904 tcgetattr(0, &save); 905 tc = save; 906 tc.c_lflag &= ~(ICANON | ECHO); 907 tc.c_cc[VMIN] = 0; 908 tc.c_cc[VTIME] = 0; 909 910 repeat: 911 delay_msecs = delay_secs * 1000; 912 tcsetattr(0, TCSANOW, &tc); 913 /* trash return*/ 914 getc(stdin); 915 916 do { 917 print_sym_table(); 918 } while (!poll(&stdin_poll, 1, delay_msecs) == 1); 919 920 c = getc(stdin); 921 tcsetattr(0, TCSAFLUSH, &save); 922 923 handle_keypress(session, c); 924 goto repeat; 925 926 return NULL; 927 } 928 929 /* Tag samples to be skipped. */ 930 static const char *skip_symbols[] = { 931 "default_idle", 932 "cpu_idle", 933 "enter_idle", 934 "exit_idle", 935 "mwait_idle", 936 "mwait_idle_with_hints", 937 "poll_idle", 938 "ppc64_runlatch_off", 939 "pseries_dedicated_idle_sleep", 940 NULL 941 }; 942 943 static int symbol_filter(struct map *map, struct symbol *sym) 944 { 945 struct sym_entry *syme; 946 const char *name = sym->name; 947 int i; 948 949 /* 950 * ppc64 uses function descriptors and appends a '.' to the 951 * start of every instruction address. Remove it. 952 */ 953 if (name[0] == '.') 954 name++; 955 956 if (!strcmp(name, "_text") || 957 !strcmp(name, "_etext") || 958 !strcmp(name, "_sinittext") || 959 !strncmp("init_module", name, 11) || 960 !strncmp("cleanup_module", name, 14) || 961 strstr(name, "_text_start") || 962 strstr(name, "_text_end")) 963 return 1; 964 965 syme = symbol__priv(sym); 966 syme->map = map; 967 syme->src = NULL; 968 969 if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) { 970 /* schedule initial sym_filter_entry setup */ 971 sym_filter_entry_sched = syme; 972 sym_filter = NULL; 973 } 974 975 for (i = 0; skip_symbols[i]; i++) { 976 if (!strcmp(skip_symbols[i], name)) { 977 syme->skip = 1; 978 break; 979 } 980 } 981 982 if (!syme->skip) 983 syme->name_len = strlen(sym->name); 984 985 return 0; 986 } 987 988 static void event__process_sample(const event_t *self, 989 struct sample_data *sample, 990 struct perf_session *session, 991 struct perf_evsel *evsel) 992 { 993 u64 ip = self->ip.ip; 994 struct sym_entry *syme; 995 struct addr_location al; 996 struct machine *machine; 997 u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK; 998 999 ++samples; 1000 1001 switch (origin) { 1002 case PERF_RECORD_MISC_USER: 1003 ++us_samples; 1004 if (hide_user_symbols) 1005 return; 1006 machine = perf_session__find_host_machine(session); 1007 break; 1008 case PERF_RECORD_MISC_KERNEL: 1009 ++kernel_samples; 1010 if (hide_kernel_symbols) 1011 return; 1012 machine = perf_session__find_host_machine(session); 1013 break; 1014 case PERF_RECORD_MISC_GUEST_KERNEL: 1015 ++guest_kernel_samples; 1016 machine = perf_session__find_machine(session, self->ip.pid); 1017 break; 1018 case PERF_RECORD_MISC_GUEST_USER: 1019 ++guest_us_samples; 1020 /* 1021 * TODO: we don't process guest user from host side 1022 * except simple counting. 1023 */ 1024 return; 1025 default: 1026 return; 1027 } 1028 1029 if (!machine && perf_guest) { 1030 pr_err("Can't find guest [%d]'s kernel information\n", 1031 self->ip.pid); 1032 return; 1033 } 1034 1035 if (self->header.misc & PERF_RECORD_MISC_EXACT_IP) 1036 exact_samples++; 1037 1038 if (event__preprocess_sample(self, session, &al, sample, 1039 symbol_filter) < 0 || 1040 al.filtered) 1041 return; 1042 1043 if (al.sym == NULL) { 1044 /* 1045 * As we do lazy loading of symtabs we only will know if the 1046 * specified vmlinux file is invalid when we actually have a 1047 * hit in kernel space and then try to load it. So if we get 1048 * here and there are _no_ symbols in the DSO backing the 1049 * kernel map, bail out. 1050 * 1051 * We may never get here, for instance, if we use -K/ 1052 * --hide-kernel-symbols, even if the user specifies an 1053 * invalid --vmlinux ;-) 1054 */ 1055 if (al.map == machine->vmlinux_maps[MAP__FUNCTION] && 1056 RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) { 1057 pr_err("The %s file can't be used\n", 1058 symbol_conf.vmlinux_name); 1059 exit(1); 1060 } 1061 1062 return; 1063 } 1064 1065 /* let's see, whether we need to install initial sym_filter_entry */ 1066 if (sym_filter_entry_sched) { 1067 sym_filter_entry = sym_filter_entry_sched; 1068 sym_filter_entry_sched = NULL; 1069 if (parse_source(sym_filter_entry) < 0) { 1070 struct symbol *sym = sym_entry__symbol(sym_filter_entry); 1071 1072 pr_err("Can't annotate %s", sym->name); 1073 if (sym_filter_entry->map->dso->origin == DSO__ORIG_KERNEL) { 1074 pr_err(": No vmlinux file was found in the path:\n"); 1075 machine__fprintf_vmlinux_path(machine, stderr); 1076 } else 1077 pr_err(".\n"); 1078 exit(1); 1079 } 1080 } 1081 1082 syme = symbol__priv(al.sym); 1083 if (!syme->skip) { 1084 syme->count[evsel->idx]++; 1085 syme->origin = origin; 1086 record_precise_ip(syme, evsel->idx, ip); 1087 pthread_mutex_lock(&active_symbols_lock); 1088 if (list_empty(&syme->node) || !syme->node.next) 1089 __list_insert_active_sym(syme); 1090 pthread_mutex_unlock(&active_symbols_lock); 1091 } 1092 } 1093 1094 struct mmap_data { 1095 void *base; 1096 int mask; 1097 unsigned int prev; 1098 }; 1099 1100 static int perf_evsel__alloc_mmap_per_thread(struct perf_evsel *evsel, 1101 int ncpus, int nthreads) 1102 { 1103 evsel->priv = xyarray__new(ncpus, nthreads, sizeof(struct mmap_data)); 1104 return evsel->priv != NULL ? 0 : -ENOMEM; 1105 } 1106 1107 static void perf_evsel__free_mmap(struct perf_evsel *evsel) 1108 { 1109 xyarray__delete(evsel->priv); 1110 evsel->priv = NULL; 1111 } 1112 1113 static unsigned int mmap_read_head(struct mmap_data *md) 1114 { 1115 struct perf_event_mmap_page *pc = md->base; 1116 int head; 1117 1118 head = pc->data_head; 1119 rmb(); 1120 1121 return head; 1122 } 1123 1124 static void perf_session__mmap_read_counter(struct perf_session *self, 1125 struct perf_evsel *evsel, 1126 int cpu, int thread_idx) 1127 { 1128 struct xyarray *mmap_array = evsel->priv; 1129 struct mmap_data *md = xyarray__entry(mmap_array, cpu, thread_idx); 1130 unsigned int head = mmap_read_head(md); 1131 unsigned int old = md->prev; 1132 unsigned char *data = md->base + page_size; 1133 struct sample_data sample; 1134 int diff; 1135 1136 /* 1137 * If we're further behind than half the buffer, there's a chance 1138 * the writer will bite our tail and mess up the samples under us. 1139 * 1140 * If we somehow ended up ahead of the head, we got messed up. 1141 * 1142 * In either case, truncate and restart at head. 1143 */ 1144 diff = head - old; 1145 if (diff > md->mask / 2 || diff < 0) { 1146 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n"); 1147 1148 /* 1149 * head points to a known good entry, start there. 1150 */ 1151 old = head; 1152 } 1153 1154 for (; old != head;) { 1155 event_t *event = (event_t *)&data[old & md->mask]; 1156 1157 event_t event_copy; 1158 1159 size_t size = event->header.size; 1160 1161 /* 1162 * Event straddles the mmap boundary -- header should always 1163 * be inside due to u64 alignment of output. 1164 */ 1165 if ((old & md->mask) + size != ((old + size) & md->mask)) { 1166 unsigned int offset = old; 1167 unsigned int len = min(sizeof(*event), size), cpy; 1168 void *dst = &event_copy; 1169 1170 do { 1171 cpy = min(md->mask + 1 - (offset & md->mask), len); 1172 memcpy(dst, &data[offset & md->mask], cpy); 1173 offset += cpy; 1174 dst += cpy; 1175 len -= cpy; 1176 } while (len); 1177 1178 event = &event_copy; 1179 } 1180 1181 event__parse_sample(event, self, &sample); 1182 if (event->header.type == PERF_RECORD_SAMPLE) 1183 event__process_sample(event, &sample, self, evsel); 1184 else 1185 event__process(event, &sample, self); 1186 old += size; 1187 } 1188 1189 md->prev = old; 1190 } 1191 1192 static struct pollfd *event_array; 1193 1194 static void perf_session__mmap_read(struct perf_session *self) 1195 { 1196 struct perf_evsel *counter; 1197 int i, thread_index; 1198 1199 for (i = 0; i < cpus->nr; i++) { 1200 list_for_each_entry(counter, &evsel_list, node) { 1201 for (thread_index = 0; 1202 thread_index < threads->nr; 1203 thread_index++) { 1204 perf_session__mmap_read_counter(self, 1205 counter, i, thread_index); 1206 } 1207 } 1208 } 1209 } 1210 1211 int nr_poll; 1212 int group_fd; 1213 1214 static void start_counter(int i, struct perf_evsel *evsel) 1215 { 1216 struct xyarray *mmap_array = evsel->priv; 1217 struct mmap_data *mm; 1218 struct perf_event_attr *attr; 1219 int cpu = -1; 1220 int thread_index; 1221 1222 if (target_tid == -1) 1223 cpu = cpus->map[i]; 1224 1225 attr = &evsel->attr; 1226 1227 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID; 1228 1229 if (freq) { 1230 attr->sample_type |= PERF_SAMPLE_PERIOD; 1231 attr->freq = 1; 1232 attr->sample_freq = freq; 1233 } 1234 1235 attr->inherit = (cpu < 0) && inherit; 1236 attr->mmap = 1; 1237 1238 for (thread_index = 0; thread_index < threads->nr; thread_index++) { 1239 try_again: 1240 FD(evsel, i, thread_index) = sys_perf_event_open(attr, 1241 threads->map[thread_index], cpu, group_fd, 0); 1242 1243 if (FD(evsel, i, thread_index) < 0) { 1244 int err = errno; 1245 1246 if (err == EPERM || err == EACCES) 1247 die("Permission error - are you root?\n" 1248 "\t Consider tweaking" 1249 " /proc/sys/kernel/perf_event_paranoid.\n"); 1250 /* 1251 * If it's cycles then fall back to hrtimer 1252 * based cpu-clock-tick sw counter, which 1253 * is always available even if no PMU support: 1254 */ 1255 if (attr->type == PERF_TYPE_HARDWARE 1256 && attr->config == PERF_COUNT_HW_CPU_CYCLES) { 1257 1258 if (verbose) 1259 warning(" ... trying to fall back to cpu-clock-ticks\n"); 1260 1261 attr->type = PERF_TYPE_SOFTWARE; 1262 attr->config = PERF_COUNT_SW_CPU_CLOCK; 1263 goto try_again; 1264 } 1265 printf("\n"); 1266 error("sys_perf_event_open() syscall returned with %d (%s). /bin/dmesg may provide additional information.\n", 1267 FD(evsel, i, thread_index), strerror(err)); 1268 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n"); 1269 exit(-1); 1270 } 1271 assert(FD(evsel, i, thread_index) >= 0); 1272 fcntl(FD(evsel, i, thread_index), F_SETFL, O_NONBLOCK); 1273 1274 /* 1275 * First counter acts as the group leader: 1276 */ 1277 if (group && group_fd == -1) 1278 group_fd = FD(evsel, i, thread_index); 1279 1280 event_array[nr_poll].fd = FD(evsel, i, thread_index); 1281 event_array[nr_poll].events = POLLIN; 1282 nr_poll++; 1283 1284 mm = xyarray__entry(mmap_array, i, thread_index); 1285 mm->prev = 0; 1286 mm->mask = mmap_pages*page_size - 1; 1287 mm->base = mmap(NULL, (mmap_pages+1)*page_size, 1288 PROT_READ, MAP_SHARED, FD(evsel, i, thread_index), 0); 1289 if (mm->base == MAP_FAILED) 1290 die("failed to mmap with %d (%s)\n", errno, strerror(errno)); 1291 } 1292 } 1293 1294 static int __cmd_top(void) 1295 { 1296 pthread_t thread; 1297 struct perf_evsel *counter; 1298 int i, ret; 1299 /* 1300 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this 1301 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now. 1302 */ 1303 struct perf_session *session = perf_session__new(NULL, O_WRONLY, false, false, NULL); 1304 if (session == NULL) 1305 return -ENOMEM; 1306 1307 if (target_tid != -1) 1308 event__synthesize_thread(target_tid, event__process, session); 1309 else 1310 event__synthesize_threads(event__process, session); 1311 1312 for (i = 0; i < cpus->nr; i++) { 1313 group_fd = -1; 1314 list_for_each_entry(counter, &evsel_list, node) 1315 start_counter(i, counter); 1316 } 1317 1318 /* Wait for a minimal set of events before starting the snapshot */ 1319 poll(&event_array[0], nr_poll, 100); 1320 1321 perf_session__mmap_read(session); 1322 1323 if (pthread_create(&thread, NULL, display_thread, session)) { 1324 printf("Could not create display thread.\n"); 1325 exit(-1); 1326 } 1327 1328 if (realtime_prio) { 1329 struct sched_param param; 1330 1331 param.sched_priority = realtime_prio; 1332 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) { 1333 printf("Could not set realtime priority.\n"); 1334 exit(-1); 1335 } 1336 } 1337 1338 while (1) { 1339 int hits = samples; 1340 1341 perf_session__mmap_read(session); 1342 1343 if (hits == samples) 1344 ret = poll(event_array, nr_poll, 100); 1345 } 1346 1347 return 0; 1348 } 1349 1350 static const char * const top_usage[] = { 1351 "perf top [<options>]", 1352 NULL 1353 }; 1354 1355 static const struct option options[] = { 1356 OPT_CALLBACK('e', "event", NULL, "event", 1357 "event selector. use 'perf list' to list available events", 1358 parse_events), 1359 OPT_INTEGER('c', "count", &default_interval, 1360 "event period to sample"), 1361 OPT_INTEGER('p', "pid", &target_pid, 1362 "profile events on existing process id"), 1363 OPT_INTEGER('t', "tid", &target_tid, 1364 "profile events on existing thread id"), 1365 OPT_BOOLEAN('a', "all-cpus", &system_wide, 1366 "system-wide collection from all CPUs"), 1367 OPT_STRING('C', "cpu", &cpu_list, "cpu", 1368 "list of cpus to monitor"), 1369 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name, 1370 "file", "vmlinux pathname"), 1371 OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols, 1372 "hide kernel symbols"), 1373 OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"), 1374 OPT_INTEGER('r', "realtime", &realtime_prio, 1375 "collect data with this RT SCHED_FIFO priority"), 1376 OPT_INTEGER('d', "delay", &delay_secs, 1377 "number of seconds to delay between refreshes"), 1378 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab, 1379 "dump the symbol table used for profiling"), 1380 OPT_INTEGER('f', "count-filter", &count_filter, 1381 "only display functions with more events than this"), 1382 OPT_BOOLEAN('g', "group", &group, 1383 "put the counters into a counter group"), 1384 OPT_BOOLEAN('i', "inherit", &inherit, 1385 "child tasks inherit counters"), 1386 OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name", 1387 "symbol to annotate"), 1388 OPT_BOOLEAN('z', "zero", &zero, 1389 "zero history across updates"), 1390 OPT_INTEGER('F', "freq", &freq, 1391 "profile at this frequency"), 1392 OPT_INTEGER('E', "entries", &print_entries, 1393 "display this many functions"), 1394 OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols, 1395 "hide user symbols"), 1396 OPT_INCR('v', "verbose", &verbose, 1397 "be more verbose (show counter open errors, etc)"), 1398 OPT_END() 1399 }; 1400 1401 int cmd_top(int argc, const char **argv, const char *prefix __used) 1402 { 1403 struct perf_evsel *pos; 1404 int status = -ENOMEM; 1405 1406 page_size = sysconf(_SC_PAGE_SIZE); 1407 1408 argc = parse_options(argc, argv, options, top_usage, 0); 1409 if (argc) 1410 usage_with_options(top_usage, options); 1411 1412 if (target_pid != -1) 1413 target_tid = target_pid; 1414 1415 threads = thread_map__new(target_pid, target_tid); 1416 if (threads == NULL) { 1417 pr_err("Problems finding threads of monitor\n"); 1418 usage_with_options(top_usage, options); 1419 } 1420 1421 event_array = malloc((sizeof(struct pollfd) * 1422 MAX_NR_CPUS * MAX_COUNTERS * threads->nr)); 1423 if (!event_array) 1424 return -ENOMEM; 1425 1426 /* CPU and PID are mutually exclusive */ 1427 if (target_tid > 0 && cpu_list) { 1428 printf("WARNING: PID switch overriding CPU\n"); 1429 sleep(1); 1430 cpu_list = NULL; 1431 } 1432 1433 if (!nr_counters && perf_evsel_list__create_default() < 0) { 1434 pr_err("Not enough memory for event selector list\n"); 1435 return -ENOMEM; 1436 } 1437 1438 if (delay_secs < 1) 1439 delay_secs = 1; 1440 1441 /* 1442 * User specified count overrides default frequency. 1443 */ 1444 if (default_interval) 1445 freq = 0; 1446 else if (freq) { 1447 default_interval = freq; 1448 } else { 1449 fprintf(stderr, "frequency and count are zero, aborting\n"); 1450 exit(EXIT_FAILURE); 1451 } 1452 1453 if (target_tid != -1) 1454 cpus = cpu_map__dummy_new(); 1455 else 1456 cpus = cpu_map__new(cpu_list); 1457 1458 if (cpus == NULL) 1459 usage_with_options(top_usage, options); 1460 1461 list_for_each_entry(pos, &evsel_list, node) { 1462 if (perf_evsel__alloc_mmap_per_thread(pos, cpus->nr, threads->nr) < 0 || 1463 perf_evsel__alloc_fd(pos, cpus->nr, threads->nr) < 0) 1464 goto out_free_fd; 1465 /* 1466 * Fill in the ones not specifically initialized via -c: 1467 */ 1468 if (pos->attr.sample_period) 1469 continue; 1470 1471 pos->attr.sample_period = default_interval; 1472 } 1473 1474 symbol_conf.priv_size = (sizeof(struct sym_entry) + 1475 (nr_counters + 1) * sizeof(unsigned long)); 1476 1477 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL); 1478 if (symbol__init() < 0) 1479 return -1; 1480 1481 get_term_dimensions(&winsize); 1482 if (print_entries == 0) { 1483 update_print_entries(&winsize); 1484 signal(SIGWINCH, sig_winch_handler); 1485 } 1486 1487 status = __cmd_top(); 1488 out_free_fd: 1489 list_for_each_entry(pos, &evsel_list, node) 1490 perf_evsel__free_mmap(pos); 1491 1492 return status; 1493 } 1494