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 #include <inttypes.h> 44 45 #include <errno.h> 46 #include <time.h> 47 #include <sched.h> 48 #include <pthread.h> 49 50 #include <sys/syscall.h> 51 #include <sys/ioctl.h> 52 #include <sys/poll.h> 53 #include <sys/prctl.h> 54 #include <sys/wait.h> 55 #include <sys/uio.h> 56 #include <sys/mman.h> 57 58 #include <linux/unistd.h> 59 #include <linux/types.h> 60 61 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y)) 62 63 static bool system_wide = false; 64 65 static int default_interval = 0; 66 67 static int count_filter = 5; 68 static int print_entries; 69 70 static int target_pid = -1; 71 static int target_tid = -1; 72 static struct thread_map *threads; 73 static bool inherit = false; 74 static struct cpu_map *cpus; 75 static int realtime_prio = 0; 76 static bool group = false; 77 static unsigned int page_size; 78 static unsigned int mmap_pages = 16; 79 static int freq = 1000; /* 1 KHz */ 80 81 static int delay_secs = 2; 82 static bool zero = false; 83 static bool dump_symtab = false; 84 85 static bool hide_kernel_symbols = false; 86 static bool hide_user_symbols = false; 87 static struct winsize winsize; 88 89 /* 90 * Source 91 */ 92 93 struct source_line { 94 u64 eip; 95 unsigned long count[MAX_COUNTERS]; 96 char *line; 97 struct source_line *next; 98 }; 99 100 static const char *sym_filter = NULL; 101 struct sym_entry *sym_filter_entry = NULL; 102 struct sym_entry *sym_filter_entry_sched = NULL; 103 static int sym_pcnt_filter = 5; 104 static int sym_counter = 0; 105 static struct perf_evsel *sym_evsel = NULL; 106 static int display_weighted = -1; 107 static const char *cpu_list; 108 109 /* 110 * Symbols 111 */ 112 113 struct sym_entry_source { 114 struct source_line *source; 115 struct source_line *lines; 116 struct source_line **lines_tail; 117 pthread_mutex_t lock; 118 }; 119 120 struct sym_entry { 121 struct rb_node rb_node; 122 struct list_head node; 123 unsigned long snap_count; 124 double weight; 125 int skip; 126 u16 name_len; 127 u8 origin; 128 struct map *map; 129 struct sym_entry_source *src; 130 unsigned long count[0]; 131 }; 132 133 /* 134 * Source functions 135 */ 136 137 static inline struct symbol *sym_entry__symbol(struct sym_entry *self) 138 { 139 return ((void *)self) + symbol_conf.priv_size; 140 } 141 142 void get_term_dimensions(struct winsize *ws) 143 { 144 char *s = getenv("LINES"); 145 146 if (s != NULL) { 147 ws->ws_row = atoi(s); 148 s = getenv("COLUMNS"); 149 if (s != NULL) { 150 ws->ws_col = atoi(s); 151 if (ws->ws_row && ws->ws_col) 152 return; 153 } 154 } 155 #ifdef TIOCGWINSZ 156 if (ioctl(1, TIOCGWINSZ, ws) == 0 && 157 ws->ws_row && ws->ws_col) 158 return; 159 #endif 160 ws->ws_row = 25; 161 ws->ws_col = 80; 162 } 163 164 static void update_print_entries(struct winsize *ws) 165 { 166 print_entries = ws->ws_row; 167 168 if (print_entries > 9) 169 print_entries -= 9; 170 } 171 172 static void sig_winch_handler(int sig __used) 173 { 174 get_term_dimensions(&winsize); 175 update_print_entries(&winsize); 176 } 177 178 static int parse_source(struct sym_entry *syme) 179 { 180 struct symbol *sym; 181 struct sym_entry_source *source; 182 struct map *map; 183 FILE *file; 184 char command[PATH_MAX*2]; 185 const char *path; 186 u64 len; 187 188 if (!syme) 189 return -1; 190 191 sym = sym_entry__symbol(syme); 192 map = syme->map; 193 194 /* 195 * We can't annotate with just /proc/kallsyms 196 */ 197 if (map->dso->origin == DSO__ORIG_KERNEL) 198 return -1; 199 200 if (syme->src == NULL) { 201 syme->src = zalloc(sizeof(*source)); 202 if (syme->src == NULL) 203 return -1; 204 pthread_mutex_init(&syme->src->lock, NULL); 205 } 206 207 source = syme->src; 208 209 if (source->lines) { 210 pthread_mutex_lock(&source->lock); 211 goto out_assign; 212 } 213 path = map->dso->long_name; 214 215 len = sym->end - sym->start; 216 217 sprintf(command, 218 "objdump --start-address=%#0*" PRIx64 " --stop-address=%#0*" PRIx64 " -dS %s", 219 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->start), 220 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->end), path); 221 222 file = popen(command, "r"); 223 if (!file) 224 return -1; 225 226 pthread_mutex_lock(&source->lock); 227 source->lines_tail = &source->lines; 228 while (!feof(file)) { 229 struct source_line *src; 230 size_t dummy = 0; 231 char *c, *sep; 232 233 src = malloc(sizeof(struct source_line)); 234 assert(src != NULL); 235 memset(src, 0, sizeof(struct source_line)); 236 237 if (getline(&src->line, &dummy, file) < 0) 238 break; 239 if (!src->line) 240 break; 241 242 c = strchr(src->line, '\n'); 243 if (c) 244 *c = 0; 245 246 src->next = NULL; 247 *source->lines_tail = src; 248 source->lines_tail = &src->next; 249 250 src->eip = strtoull(src->line, &sep, 16); 251 if (*sep == ':') 252 src->eip = map__objdump_2ip(map, src->eip); 253 else /* this line has no ip info (e.g. source line) */ 254 src->eip = 0; 255 } 256 pclose(file); 257 out_assign: 258 sym_filter_entry = syme; 259 pthread_mutex_unlock(&source->lock); 260 return 0; 261 } 262 263 static void __zero_source_counters(struct sym_entry *syme) 264 { 265 int i; 266 struct source_line *line; 267 268 line = syme->src->lines; 269 while (line) { 270 for (i = 0; i < nr_counters; i++) 271 line->count[i] = 0; 272 line = line->next; 273 } 274 } 275 276 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip) 277 { 278 struct source_line *line; 279 280 if (syme != sym_filter_entry) 281 return; 282 283 if (pthread_mutex_trylock(&syme->src->lock)) 284 return; 285 286 if (syme->src == NULL || syme->src->source == NULL) 287 goto out_unlock; 288 289 for (line = syme->src->lines; line; line = line->next) { 290 /* skip lines without IP info */ 291 if (line->eip == 0) 292 continue; 293 if (line->eip == ip) { 294 line->count[counter]++; 295 break; 296 } 297 if (line->eip > ip) 298 break; 299 } 300 out_unlock: 301 pthread_mutex_unlock(&syme->src->lock); 302 } 303 304 #define PATTERN_LEN (BITS_PER_LONG / 4 + 2) 305 306 static void lookup_sym_source(struct sym_entry *syme) 307 { 308 struct symbol *symbol = sym_entry__symbol(syme); 309 struct source_line *line; 310 char pattern[PATTERN_LEN + 1]; 311 312 sprintf(pattern, "%0*" PRIx64 " <", BITS_PER_LONG / 4, 313 map__rip_2objdump(syme->map, symbol->start)); 314 315 pthread_mutex_lock(&syme->src->lock); 316 for (line = syme->src->lines; line; line = line->next) { 317 if (memcmp(line->line, pattern, PATTERN_LEN) == 0) { 318 syme->src->source = line; 319 break; 320 } 321 } 322 pthread_mutex_unlock(&syme->src->lock); 323 } 324 325 static void show_lines(struct source_line *queue, int count, int total) 326 { 327 int i; 328 struct source_line *line; 329 330 line = queue; 331 for (i = 0; i < count; i++) { 332 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total; 333 334 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line); 335 line = line->next; 336 } 337 } 338 339 #define TRACE_COUNT 3 340 341 static void show_details(struct sym_entry *syme) 342 { 343 struct symbol *symbol; 344 struct source_line *line; 345 struct source_line *line_queue = NULL; 346 int displayed = 0; 347 int line_queue_count = 0, total = 0, more = 0; 348 349 if (!syme) 350 return; 351 352 if (!syme->src->source) 353 lookup_sym_source(syme); 354 355 if (!syme->src->source) 356 return; 357 358 symbol = sym_entry__symbol(syme); 359 printf("Showing %s for %s\n", event_name(sym_evsel), symbol->name); 360 printf(" Events Pcnt (>=%d%%)\n", sym_pcnt_filter); 361 362 pthread_mutex_lock(&syme->src->lock); 363 line = syme->src->source; 364 while (line) { 365 total += line->count[sym_counter]; 366 line = line->next; 367 } 368 369 line = syme->src->source; 370 while (line) { 371 float pcnt = 0.0; 372 373 if (!line_queue_count) 374 line_queue = line; 375 line_queue_count++; 376 377 if (line->count[sym_counter]) 378 pcnt = 100.0 * line->count[sym_counter] / (float)total; 379 if (pcnt >= (float)sym_pcnt_filter) { 380 if (displayed <= print_entries) 381 show_lines(line_queue, line_queue_count, total); 382 else more++; 383 displayed += line_queue_count; 384 line_queue_count = 0; 385 line_queue = NULL; 386 } else if (line_queue_count > TRACE_COUNT) { 387 line_queue = line_queue->next; 388 line_queue_count--; 389 } 390 391 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8; 392 line = line->next; 393 } 394 pthread_mutex_unlock(&syme->src->lock); 395 if (more) 396 printf("%d lines not displayed, maybe increase display entries [e]\n", more); 397 } 398 399 /* 400 * Symbols will be added here in event__process_sample and will get out 401 * after decayed. 402 */ 403 static LIST_HEAD(active_symbols); 404 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER; 405 406 /* 407 * Ordering weight: count-1 * count-2 * ... / count-n 408 */ 409 static double sym_weight(const struct sym_entry *sym) 410 { 411 double weight = sym->snap_count; 412 int counter; 413 414 if (!display_weighted) 415 return weight; 416 417 for (counter = 1; counter < nr_counters-1; counter++) 418 weight *= sym->count[counter]; 419 420 weight /= (sym->count[counter] + 1); 421 422 return weight; 423 } 424 425 static long samples; 426 static long kernel_samples, us_samples; 427 static long exact_samples; 428 static long guest_us_samples, guest_kernel_samples; 429 static const char CONSOLE_CLEAR[] = "[H[2J"; 430 431 static void __list_insert_active_sym(struct sym_entry *syme) 432 { 433 list_add(&syme->node, &active_symbols); 434 } 435 436 static void list_remove_active_sym(struct sym_entry *syme) 437 { 438 pthread_mutex_lock(&active_symbols_lock); 439 list_del_init(&syme->node); 440 pthread_mutex_unlock(&active_symbols_lock); 441 } 442 443 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se) 444 { 445 struct rb_node **p = &tree->rb_node; 446 struct rb_node *parent = NULL; 447 struct sym_entry *iter; 448 449 while (*p != NULL) { 450 parent = *p; 451 iter = rb_entry(parent, struct sym_entry, rb_node); 452 453 if (se->weight > iter->weight) 454 p = &(*p)->rb_left; 455 else 456 p = &(*p)->rb_right; 457 } 458 459 rb_link_node(&se->rb_node, parent, p); 460 rb_insert_color(&se->rb_node, tree); 461 } 462 463 static void print_sym_table(void) 464 { 465 int printed = 0, j; 466 struct perf_evsel *counter; 467 int snap = !display_weighted ? sym_counter : 0; 468 float samples_per_sec = samples/delay_secs; 469 float ksamples_per_sec = kernel_samples/delay_secs; 470 float us_samples_per_sec = (us_samples)/delay_secs; 471 float guest_kernel_samples_per_sec = (guest_kernel_samples)/delay_secs; 472 float guest_us_samples_per_sec = (guest_us_samples)/delay_secs; 473 float esamples_percent = (100.0*exact_samples)/samples; 474 float sum_ksamples = 0.0; 475 struct sym_entry *syme, *n; 476 struct rb_root tmp = RB_ROOT; 477 struct rb_node *nd; 478 int sym_width = 0, dso_width = 0, dso_short_width = 0; 479 const int win_width = winsize.ws_col - 1; 480 481 samples = us_samples = kernel_samples = exact_samples = 0; 482 guest_kernel_samples = guest_us_samples = 0; 483 484 /* Sort the active symbols */ 485 pthread_mutex_lock(&active_symbols_lock); 486 syme = list_entry(active_symbols.next, struct sym_entry, node); 487 pthread_mutex_unlock(&active_symbols_lock); 488 489 list_for_each_entry_safe_from(syme, n, &active_symbols, node) { 490 syme->snap_count = syme->count[snap]; 491 if (syme->snap_count != 0) { 492 493 if ((hide_user_symbols && 494 syme->origin == PERF_RECORD_MISC_USER) || 495 (hide_kernel_symbols && 496 syme->origin == PERF_RECORD_MISC_KERNEL)) { 497 list_remove_active_sym(syme); 498 continue; 499 } 500 syme->weight = sym_weight(syme); 501 rb_insert_active_sym(&tmp, syme); 502 sum_ksamples += syme->snap_count; 503 504 for (j = 0; j < nr_counters; j++) 505 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8; 506 } else 507 list_remove_active_sym(syme); 508 } 509 510 puts(CONSOLE_CLEAR); 511 512 printf("%-*.*s\n", win_width, win_width, graph_dotted_line); 513 if (!perf_guest) { 514 printf(" PerfTop:%8.0f irqs/sec kernel:%4.1f%%" 515 " exact: %4.1f%% [", 516 samples_per_sec, 517 100.0 - (100.0 * ((samples_per_sec - ksamples_per_sec) / 518 samples_per_sec)), 519 esamples_percent); 520 } else { 521 printf(" PerfTop:%8.0f irqs/sec kernel:%4.1f%% us:%4.1f%%" 522 " guest kernel:%4.1f%% guest us:%4.1f%%" 523 " exact: %4.1f%% [", 524 samples_per_sec, 525 100.0 - (100.0 * ((samples_per_sec-ksamples_per_sec) / 526 samples_per_sec)), 527 100.0 - (100.0 * ((samples_per_sec-us_samples_per_sec) / 528 samples_per_sec)), 529 100.0 - (100.0 * ((samples_per_sec - 530 guest_kernel_samples_per_sec) / 531 samples_per_sec)), 532 100.0 - (100.0 * ((samples_per_sec - 533 guest_us_samples_per_sec) / 534 samples_per_sec)), 535 esamples_percent); 536 } 537 538 if (nr_counters == 1 || !display_weighted) { 539 struct perf_evsel *first; 540 first = list_entry(evsel_list.next, struct perf_evsel, node); 541 printf("%" PRIu64, (uint64_t)first->attr.sample_period); 542 if (freq) 543 printf("Hz "); 544 else 545 printf(" "); 546 } 547 548 if (!display_weighted) 549 printf("%s", event_name(sym_evsel)); 550 else list_for_each_entry(counter, &evsel_list, node) { 551 if (counter->idx) 552 printf("/"); 553 554 printf("%s", event_name(counter)); 555 } 556 557 printf( "], "); 558 559 if (target_pid != -1) 560 printf(" (target_pid: %d", target_pid); 561 else if (target_tid != -1) 562 printf(" (target_tid: %d", target_tid); 563 else 564 printf(" (all"); 565 566 if (cpu_list) 567 printf(", CPU%s: %s)\n", cpus->nr > 1 ? "s" : "", cpu_list); 568 else { 569 if (target_tid != -1) 570 printf(")\n"); 571 else 572 printf(", %d CPU%s)\n", cpus->nr, cpus->nr > 1 ? "s" : ""); 573 } 574 575 printf("%-*.*s\n", win_width, win_width, graph_dotted_line); 576 577 if (sym_filter_entry) { 578 show_details(sym_filter_entry); 579 return; 580 } 581 582 /* 583 * Find the longest symbol name that will be displayed 584 */ 585 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) { 586 syme = rb_entry(nd, struct sym_entry, rb_node); 587 if (++printed > print_entries || 588 (int)syme->snap_count < count_filter) 589 continue; 590 591 if (syme->map->dso->long_name_len > dso_width) 592 dso_width = syme->map->dso->long_name_len; 593 594 if (syme->map->dso->short_name_len > dso_short_width) 595 dso_short_width = syme->map->dso->short_name_len; 596 597 if (syme->name_len > sym_width) 598 sym_width = syme->name_len; 599 } 600 601 printed = 0; 602 603 if (sym_width + dso_width > winsize.ws_col - 29) { 604 dso_width = dso_short_width; 605 if (sym_width + dso_width > winsize.ws_col - 29) 606 sym_width = winsize.ws_col - dso_width - 29; 607 } 608 putchar('\n'); 609 if (nr_counters == 1) 610 printf(" samples pcnt"); 611 else 612 printf(" weight samples pcnt"); 613 614 if (verbose) 615 printf(" RIP "); 616 printf(" %-*.*s DSO\n", sym_width, sym_width, "function"); 617 printf(" %s _______ _____", 618 nr_counters == 1 ? " " : "______"); 619 if (verbose) 620 printf(" ________________"); 621 printf(" %-*.*s", sym_width, sym_width, graph_line); 622 printf(" %-*.*s", dso_width, dso_width, graph_line); 623 puts("\n"); 624 625 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) { 626 struct symbol *sym; 627 double pcnt; 628 629 syme = rb_entry(nd, struct sym_entry, rb_node); 630 sym = sym_entry__symbol(syme); 631 if (++printed > print_entries || (int)syme->snap_count < count_filter) 632 continue; 633 634 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) / 635 sum_ksamples)); 636 637 if (nr_counters == 1 || !display_weighted) 638 printf("%20.2f ", syme->weight); 639 else 640 printf("%9.1f %10ld ", syme->weight, syme->snap_count); 641 642 percent_color_fprintf(stdout, "%4.1f%%", pcnt); 643 if (verbose) 644 printf(" %016" PRIx64, sym->start); 645 printf(" %-*.*s", sym_width, sym_width, sym->name); 646 printf(" %-*.*s\n", dso_width, dso_width, 647 dso_width >= syme->map->dso->long_name_len ? 648 syme->map->dso->long_name : 649 syme->map->dso->short_name); 650 } 651 } 652 653 static void prompt_integer(int *target, const char *msg) 654 { 655 char *buf = malloc(0), *p; 656 size_t dummy = 0; 657 int tmp; 658 659 fprintf(stdout, "\n%s: ", msg); 660 if (getline(&buf, &dummy, stdin) < 0) 661 return; 662 663 p = strchr(buf, '\n'); 664 if (p) 665 *p = 0; 666 667 p = buf; 668 while(*p) { 669 if (!isdigit(*p)) 670 goto out_free; 671 p++; 672 } 673 tmp = strtoul(buf, NULL, 10); 674 *target = tmp; 675 out_free: 676 free(buf); 677 } 678 679 static void prompt_percent(int *target, const char *msg) 680 { 681 int tmp = 0; 682 683 prompt_integer(&tmp, msg); 684 if (tmp >= 0 && tmp <= 100) 685 *target = tmp; 686 } 687 688 static void prompt_symbol(struct sym_entry **target, const char *msg) 689 { 690 char *buf = malloc(0), *p; 691 struct sym_entry *syme = *target, *n, *found = NULL; 692 size_t dummy = 0; 693 694 /* zero counters of active symbol */ 695 if (syme) { 696 pthread_mutex_lock(&syme->src->lock); 697 __zero_source_counters(syme); 698 *target = NULL; 699 pthread_mutex_unlock(&syme->src->lock); 700 } 701 702 fprintf(stdout, "\n%s: ", msg); 703 if (getline(&buf, &dummy, stdin) < 0) 704 goto out_free; 705 706 p = strchr(buf, '\n'); 707 if (p) 708 *p = 0; 709 710 pthread_mutex_lock(&active_symbols_lock); 711 syme = list_entry(active_symbols.next, struct sym_entry, node); 712 pthread_mutex_unlock(&active_symbols_lock); 713 714 list_for_each_entry_safe_from(syme, n, &active_symbols, node) { 715 struct symbol *sym = sym_entry__symbol(syme); 716 717 if (!strcmp(buf, sym->name)) { 718 found = syme; 719 break; 720 } 721 } 722 723 if (!found) { 724 fprintf(stderr, "Sorry, %s is not active.\n", buf); 725 sleep(1); 726 return; 727 } else 728 parse_source(found); 729 730 out_free: 731 free(buf); 732 } 733 734 static void print_mapped_keys(void) 735 { 736 char *name = NULL; 737 738 if (sym_filter_entry) { 739 struct symbol *sym = sym_entry__symbol(sym_filter_entry); 740 name = sym->name; 741 } 742 743 fprintf(stdout, "\nMapped keys:\n"); 744 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", delay_secs); 745 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", print_entries); 746 747 if (nr_counters > 1) 748 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", event_name(sym_evsel)); 749 750 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", count_filter); 751 752 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter); 753 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL"); 754 fprintf(stdout, "\t[S] stop annotation.\n"); 755 756 if (nr_counters > 1) 757 fprintf(stdout, "\t[w] toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0); 758 759 fprintf(stdout, 760 "\t[K] hide kernel_symbols symbols. \t(%s)\n", 761 hide_kernel_symbols ? "yes" : "no"); 762 fprintf(stdout, 763 "\t[U] hide user symbols. \t(%s)\n", 764 hide_user_symbols ? "yes" : "no"); 765 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", zero ? 1 : 0); 766 fprintf(stdout, "\t[qQ] quit.\n"); 767 } 768 769 static int key_mapped(int c) 770 { 771 switch (c) { 772 case 'd': 773 case 'e': 774 case 'f': 775 case 'z': 776 case 'q': 777 case 'Q': 778 case 'K': 779 case 'U': 780 case 'F': 781 case 's': 782 case 'S': 783 return 1; 784 case 'E': 785 case 'w': 786 return nr_counters > 1 ? 1 : 0; 787 default: 788 break; 789 } 790 791 return 0; 792 } 793 794 static void handle_keypress(struct perf_session *session, int c) 795 { 796 if (!key_mapped(c)) { 797 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN }; 798 struct termios tc, save; 799 800 print_mapped_keys(); 801 fprintf(stdout, "\nEnter selection, or unmapped key to continue: "); 802 fflush(stdout); 803 804 tcgetattr(0, &save); 805 tc = save; 806 tc.c_lflag &= ~(ICANON | ECHO); 807 tc.c_cc[VMIN] = 0; 808 tc.c_cc[VTIME] = 0; 809 tcsetattr(0, TCSANOW, &tc); 810 811 poll(&stdin_poll, 1, -1); 812 c = getc(stdin); 813 814 tcsetattr(0, TCSAFLUSH, &save); 815 if (!key_mapped(c)) 816 return; 817 } 818 819 switch (c) { 820 case 'd': 821 prompt_integer(&delay_secs, "Enter display delay"); 822 if (delay_secs < 1) 823 delay_secs = 1; 824 break; 825 case 'e': 826 prompt_integer(&print_entries, "Enter display entries (lines)"); 827 if (print_entries == 0) { 828 sig_winch_handler(SIGWINCH); 829 signal(SIGWINCH, sig_winch_handler); 830 } else 831 signal(SIGWINCH, SIG_DFL); 832 break; 833 case 'E': 834 if (nr_counters > 1) { 835 fprintf(stderr, "\nAvailable events:"); 836 837 list_for_each_entry(sym_evsel, &evsel_list, node) 838 fprintf(stderr, "\n\t%d %s", sym_evsel->idx, event_name(sym_evsel)); 839 840 prompt_integer(&sym_counter, "Enter details event counter"); 841 842 if (sym_counter >= nr_counters) { 843 sym_evsel = list_entry(evsel_list.next, struct perf_evsel, node); 844 sym_counter = 0; 845 fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(sym_evsel)); 846 sleep(1); 847 break; 848 } 849 list_for_each_entry(sym_evsel, &evsel_list, node) 850 if (sym_evsel->idx == sym_counter) 851 break; 852 } else sym_counter = 0; 853 break; 854 case 'f': 855 prompt_integer(&count_filter, "Enter display event count filter"); 856 break; 857 case 'F': 858 prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)"); 859 break; 860 case 'K': 861 hide_kernel_symbols = !hide_kernel_symbols; 862 break; 863 case 'q': 864 case 'Q': 865 printf("exiting.\n"); 866 if (dump_symtab) 867 perf_session__fprintf_dsos(session, stderr); 868 exit(0); 869 case 's': 870 prompt_symbol(&sym_filter_entry, "Enter details symbol"); 871 break; 872 case 'S': 873 if (!sym_filter_entry) 874 break; 875 else { 876 struct sym_entry *syme = sym_filter_entry; 877 878 pthread_mutex_lock(&syme->src->lock); 879 sym_filter_entry = NULL; 880 __zero_source_counters(syme); 881 pthread_mutex_unlock(&syme->src->lock); 882 } 883 break; 884 case 'U': 885 hide_user_symbols = !hide_user_symbols; 886 break; 887 case 'w': 888 display_weighted = ~display_weighted; 889 break; 890 case 'z': 891 zero = !zero; 892 break; 893 default: 894 break; 895 } 896 } 897 898 static void *display_thread(void *arg __used) 899 { 900 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN }; 901 struct termios tc, save; 902 int delay_msecs, c; 903 struct perf_session *session = (struct perf_session *) arg; 904 905 tcgetattr(0, &save); 906 tc = save; 907 tc.c_lflag &= ~(ICANON | ECHO); 908 tc.c_cc[VMIN] = 0; 909 tc.c_cc[VTIME] = 0; 910 911 repeat: 912 delay_msecs = delay_secs * 1000; 913 tcsetattr(0, TCSANOW, &tc); 914 /* trash return*/ 915 getc(stdin); 916 917 do { 918 print_sym_table(); 919 } while (!poll(&stdin_poll, 1, delay_msecs) == 1); 920 921 c = getc(stdin); 922 tcsetattr(0, TCSAFLUSH, &save); 923 924 handle_keypress(session, c); 925 goto repeat; 926 927 return NULL; 928 } 929 930 /* Tag samples to be skipped. */ 931 static const char *skip_symbols[] = { 932 "default_idle", 933 "cpu_idle", 934 "enter_idle", 935 "exit_idle", 936 "mwait_idle", 937 "mwait_idle_with_hints", 938 "poll_idle", 939 "ppc64_runlatch_off", 940 "pseries_dedicated_idle_sleep", 941 NULL 942 }; 943 944 static int symbol_filter(struct map *map, struct symbol *sym) 945 { 946 struct sym_entry *syme; 947 const char *name = sym->name; 948 int i; 949 950 /* 951 * ppc64 uses function descriptors and appends a '.' to the 952 * start of every instruction address. Remove it. 953 */ 954 if (name[0] == '.') 955 name++; 956 957 if (!strcmp(name, "_text") || 958 !strcmp(name, "_etext") || 959 !strcmp(name, "_sinittext") || 960 !strncmp("init_module", name, 11) || 961 !strncmp("cleanup_module", name, 14) || 962 strstr(name, "_text_start") || 963 strstr(name, "_text_end")) 964 return 1; 965 966 syme = symbol__priv(sym); 967 syme->map = map; 968 syme->src = NULL; 969 970 if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) { 971 /* schedule initial sym_filter_entry setup */ 972 sym_filter_entry_sched = syme; 973 sym_filter = NULL; 974 } 975 976 for (i = 0; skip_symbols[i]; i++) { 977 if (!strcmp(skip_symbols[i], name)) { 978 syme->skip = 1; 979 break; 980 } 981 } 982 983 if (!syme->skip) 984 syme->name_len = strlen(sym->name); 985 986 return 0; 987 } 988 989 static void event__process_sample(const event_t *self, 990 struct sample_data *sample, 991 struct perf_session *session, 992 struct perf_evsel *evsel) 993 { 994 u64 ip = self->ip.ip; 995 struct sym_entry *syme; 996 struct addr_location al; 997 struct machine *machine; 998 u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK; 999 1000 ++samples; 1001 1002 switch (origin) { 1003 case PERF_RECORD_MISC_USER: 1004 ++us_samples; 1005 if (hide_user_symbols) 1006 return; 1007 machine = perf_session__find_host_machine(session); 1008 break; 1009 case PERF_RECORD_MISC_KERNEL: 1010 ++kernel_samples; 1011 if (hide_kernel_symbols) 1012 return; 1013 machine = perf_session__find_host_machine(session); 1014 break; 1015 case PERF_RECORD_MISC_GUEST_KERNEL: 1016 ++guest_kernel_samples; 1017 machine = perf_session__find_machine(session, self->ip.pid); 1018 break; 1019 case PERF_RECORD_MISC_GUEST_USER: 1020 ++guest_us_samples; 1021 /* 1022 * TODO: we don't process guest user from host side 1023 * except simple counting. 1024 */ 1025 return; 1026 default: 1027 return; 1028 } 1029 1030 if (!machine && perf_guest) { 1031 pr_err("Can't find guest [%d]'s kernel information\n", 1032 self->ip.pid); 1033 return; 1034 } 1035 1036 if (self->header.misc & PERF_RECORD_MISC_EXACT_IP) 1037 exact_samples++; 1038 1039 if (event__preprocess_sample(self, session, &al, sample, 1040 symbol_filter) < 0 || 1041 al.filtered) 1042 return; 1043 1044 if (al.sym == NULL) { 1045 /* 1046 * As we do lazy loading of symtabs we only will know if the 1047 * specified vmlinux file is invalid when we actually have a 1048 * hit in kernel space and then try to load it. So if we get 1049 * here and there are _no_ symbols in the DSO backing the 1050 * kernel map, bail out. 1051 * 1052 * We may never get here, for instance, if we use -K/ 1053 * --hide-kernel-symbols, even if the user specifies an 1054 * invalid --vmlinux ;-) 1055 */ 1056 if (al.map == machine->vmlinux_maps[MAP__FUNCTION] && 1057 RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) { 1058 pr_err("The %s file can't be used\n", 1059 symbol_conf.vmlinux_name); 1060 exit(1); 1061 } 1062 1063 return; 1064 } 1065 1066 /* let's see, whether we need to install initial sym_filter_entry */ 1067 if (sym_filter_entry_sched) { 1068 sym_filter_entry = sym_filter_entry_sched; 1069 sym_filter_entry_sched = NULL; 1070 if (parse_source(sym_filter_entry) < 0) { 1071 struct symbol *sym = sym_entry__symbol(sym_filter_entry); 1072 1073 pr_err("Can't annotate %s", sym->name); 1074 if (sym_filter_entry->map->dso->origin == DSO__ORIG_KERNEL) { 1075 pr_err(": No vmlinux file was found in the path:\n"); 1076 machine__fprintf_vmlinux_path(machine, stderr); 1077 } else 1078 pr_err(".\n"); 1079 exit(1); 1080 } 1081 } 1082 1083 syme = symbol__priv(al.sym); 1084 if (!syme->skip) { 1085 syme->count[evsel->idx]++; 1086 syme->origin = origin; 1087 record_precise_ip(syme, evsel->idx, ip); 1088 pthread_mutex_lock(&active_symbols_lock); 1089 if (list_empty(&syme->node) || !syme->node.next) 1090 __list_insert_active_sym(syme); 1091 pthread_mutex_unlock(&active_symbols_lock); 1092 } 1093 } 1094 1095 struct mmap_data { 1096 void *base; 1097 int mask; 1098 unsigned int prev; 1099 }; 1100 1101 static int perf_evsel__alloc_mmap_per_thread(struct perf_evsel *evsel, 1102 int ncpus, int nthreads) 1103 { 1104 evsel->priv = xyarray__new(ncpus, nthreads, sizeof(struct mmap_data)); 1105 return evsel->priv != NULL ? 0 : -ENOMEM; 1106 } 1107 1108 static void perf_evsel__free_mmap(struct perf_evsel *evsel) 1109 { 1110 xyarray__delete(evsel->priv); 1111 evsel->priv = NULL; 1112 } 1113 1114 static unsigned int mmap_read_head(struct mmap_data *md) 1115 { 1116 struct perf_event_mmap_page *pc = md->base; 1117 int head; 1118 1119 head = pc->data_head; 1120 rmb(); 1121 1122 return head; 1123 } 1124 1125 static void perf_session__mmap_read_counter(struct perf_session *self, 1126 struct perf_evsel *evsel, 1127 int cpu, int thread_idx) 1128 { 1129 struct xyarray *mmap_array = evsel->priv; 1130 struct mmap_data *md = xyarray__entry(mmap_array, cpu, thread_idx); 1131 unsigned int head = mmap_read_head(md); 1132 unsigned int old = md->prev; 1133 unsigned char *data = md->base + page_size; 1134 struct sample_data sample; 1135 int diff; 1136 1137 /* 1138 * If we're further behind than half the buffer, there's a chance 1139 * the writer will bite our tail and mess up the samples under us. 1140 * 1141 * If we somehow ended up ahead of the head, we got messed up. 1142 * 1143 * In either case, truncate and restart at head. 1144 */ 1145 diff = head - old; 1146 if (diff > md->mask / 2 || diff < 0) { 1147 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n"); 1148 1149 /* 1150 * head points to a known good entry, start there. 1151 */ 1152 old = head; 1153 } 1154 1155 for (; old != head;) { 1156 event_t *event = (event_t *)&data[old & md->mask]; 1157 1158 event_t event_copy; 1159 1160 size_t size = event->header.size; 1161 1162 /* 1163 * Event straddles the mmap boundary -- header should always 1164 * be inside due to u64 alignment of output. 1165 */ 1166 if ((old & md->mask) + size != ((old + size) & md->mask)) { 1167 unsigned int offset = old; 1168 unsigned int len = min(sizeof(*event), size), cpy; 1169 void *dst = &event_copy; 1170 1171 do { 1172 cpy = min(md->mask + 1 - (offset & md->mask), len); 1173 memcpy(dst, &data[offset & md->mask], cpy); 1174 offset += cpy; 1175 dst += cpy; 1176 len -= cpy; 1177 } while (len); 1178 1179 event = &event_copy; 1180 } 1181 1182 event__parse_sample(event, self, &sample); 1183 if (event->header.type == PERF_RECORD_SAMPLE) 1184 event__process_sample(event, &sample, self, evsel); 1185 else 1186 event__process(event, &sample, self); 1187 old += size; 1188 } 1189 1190 md->prev = old; 1191 } 1192 1193 static struct pollfd *event_array; 1194 1195 static void perf_session__mmap_read(struct perf_session *self) 1196 { 1197 struct perf_evsel *counter; 1198 int i, thread_index; 1199 1200 for (i = 0; i < cpus->nr; i++) { 1201 list_for_each_entry(counter, &evsel_list, node) { 1202 for (thread_index = 0; 1203 thread_index < threads->nr; 1204 thread_index++) { 1205 perf_session__mmap_read_counter(self, 1206 counter, i, thread_index); 1207 } 1208 } 1209 } 1210 } 1211 1212 int nr_poll; 1213 int group_fd; 1214 1215 static void start_counter(int i, struct perf_evsel *evsel) 1216 { 1217 struct xyarray *mmap_array = evsel->priv; 1218 struct mmap_data *mm; 1219 struct perf_event_attr *attr; 1220 int cpu = -1; 1221 int thread_index; 1222 1223 if (target_tid == -1) 1224 cpu = cpus->map[i]; 1225 1226 attr = &evsel->attr; 1227 1228 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID; 1229 1230 if (freq) { 1231 attr->sample_type |= PERF_SAMPLE_PERIOD; 1232 attr->freq = 1; 1233 attr->sample_freq = freq; 1234 } 1235 1236 attr->inherit = (cpu < 0) && inherit; 1237 attr->mmap = 1; 1238 1239 for (thread_index = 0; thread_index < threads->nr; thread_index++) { 1240 try_again: 1241 FD(evsel, i, thread_index) = sys_perf_event_open(attr, 1242 threads->map[thread_index], cpu, group_fd, 0); 1243 1244 if (FD(evsel, i, thread_index) < 0) { 1245 int err = errno; 1246 1247 if (err == EPERM || err == EACCES) 1248 die("Permission error - are you root?\n" 1249 "\t Consider tweaking" 1250 " /proc/sys/kernel/perf_event_paranoid.\n"); 1251 /* 1252 * If it's cycles then fall back to hrtimer 1253 * based cpu-clock-tick sw counter, which 1254 * is always available even if no PMU support: 1255 */ 1256 if (attr->type == PERF_TYPE_HARDWARE 1257 && attr->config == PERF_COUNT_HW_CPU_CYCLES) { 1258 1259 if (verbose) 1260 warning(" ... trying to fall back to cpu-clock-ticks\n"); 1261 1262 attr->type = PERF_TYPE_SOFTWARE; 1263 attr->config = PERF_COUNT_SW_CPU_CLOCK; 1264 goto try_again; 1265 } 1266 printf("\n"); 1267 error("sys_perf_event_open() syscall returned with %d (%s). /bin/dmesg may provide additional information.\n", 1268 FD(evsel, i, thread_index), strerror(err)); 1269 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n"); 1270 exit(-1); 1271 } 1272 assert(FD(evsel, i, thread_index) >= 0); 1273 fcntl(FD(evsel, i, thread_index), F_SETFL, O_NONBLOCK); 1274 1275 /* 1276 * First counter acts as the group leader: 1277 */ 1278 if (group && group_fd == -1) 1279 group_fd = FD(evsel, i, thread_index); 1280 1281 event_array[nr_poll].fd = FD(evsel, i, thread_index); 1282 event_array[nr_poll].events = POLLIN; 1283 nr_poll++; 1284 1285 mm = xyarray__entry(mmap_array, i, thread_index); 1286 mm->prev = 0; 1287 mm->mask = mmap_pages*page_size - 1; 1288 mm->base = mmap(NULL, (mmap_pages+1)*page_size, 1289 PROT_READ, MAP_SHARED, FD(evsel, i, thread_index), 0); 1290 if (mm->base == MAP_FAILED) 1291 die("failed to mmap with %d (%s)\n", errno, strerror(errno)); 1292 } 1293 } 1294 1295 static int __cmd_top(void) 1296 { 1297 pthread_t thread; 1298 struct perf_evsel *counter; 1299 int i, ret; 1300 /* 1301 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this 1302 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now. 1303 */ 1304 struct perf_session *session = perf_session__new(NULL, O_WRONLY, false, false, NULL); 1305 if (session == NULL) 1306 return -ENOMEM; 1307 1308 if (target_tid != -1) 1309 event__synthesize_thread_map(threads, event__process, session); 1310 else 1311 event__synthesize_threads(event__process, session); 1312 1313 for (i = 0; i < cpus->nr; i++) { 1314 group_fd = -1; 1315 list_for_each_entry(counter, &evsel_list, node) 1316 start_counter(i, counter); 1317 } 1318 1319 /* Wait for a minimal set of events before starting the snapshot */ 1320 poll(&event_array[0], nr_poll, 100); 1321 1322 perf_session__mmap_read(session); 1323 1324 if (pthread_create(&thread, NULL, display_thread, session)) { 1325 printf("Could not create display thread.\n"); 1326 exit(-1); 1327 } 1328 1329 if (realtime_prio) { 1330 struct sched_param param; 1331 1332 param.sched_priority = realtime_prio; 1333 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) { 1334 printf("Could not set realtime priority.\n"); 1335 exit(-1); 1336 } 1337 } 1338 1339 while (1) { 1340 int hits = samples; 1341 1342 perf_session__mmap_read(session); 1343 1344 if (hits == samples) 1345 ret = poll(event_array, nr_poll, 100); 1346 } 1347 1348 return 0; 1349 } 1350 1351 static const char * const top_usage[] = { 1352 "perf top [<options>]", 1353 NULL 1354 }; 1355 1356 static const struct option options[] = { 1357 OPT_CALLBACK('e', "event", NULL, "event", 1358 "event selector. use 'perf list' to list available events", 1359 parse_events), 1360 OPT_INTEGER('c', "count", &default_interval, 1361 "event period to sample"), 1362 OPT_INTEGER('p', "pid", &target_pid, 1363 "profile events on existing process id"), 1364 OPT_INTEGER('t', "tid", &target_tid, 1365 "profile events on existing thread id"), 1366 OPT_BOOLEAN('a', "all-cpus", &system_wide, 1367 "system-wide collection from all CPUs"), 1368 OPT_STRING('C', "cpu", &cpu_list, "cpu", 1369 "list of cpus to monitor"), 1370 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name, 1371 "file", "vmlinux pathname"), 1372 OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols, 1373 "hide kernel symbols"), 1374 OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"), 1375 OPT_INTEGER('r', "realtime", &realtime_prio, 1376 "collect data with this RT SCHED_FIFO priority"), 1377 OPT_INTEGER('d', "delay", &delay_secs, 1378 "number of seconds to delay between refreshes"), 1379 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab, 1380 "dump the symbol table used for profiling"), 1381 OPT_INTEGER('f', "count-filter", &count_filter, 1382 "only display functions with more events than this"), 1383 OPT_BOOLEAN('g', "group", &group, 1384 "put the counters into a counter group"), 1385 OPT_BOOLEAN('i', "inherit", &inherit, 1386 "child tasks inherit counters"), 1387 OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name", 1388 "symbol to annotate"), 1389 OPT_BOOLEAN('z', "zero", &zero, 1390 "zero history across updates"), 1391 OPT_INTEGER('F', "freq", &freq, 1392 "profile at this frequency"), 1393 OPT_INTEGER('E', "entries", &print_entries, 1394 "display this many functions"), 1395 OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols, 1396 "hide user symbols"), 1397 OPT_INCR('v', "verbose", &verbose, 1398 "be more verbose (show counter open errors, etc)"), 1399 OPT_END() 1400 }; 1401 1402 int cmd_top(int argc, const char **argv, const char *prefix __used) 1403 { 1404 struct perf_evsel *pos; 1405 int status = -ENOMEM; 1406 1407 page_size = sysconf(_SC_PAGE_SIZE); 1408 1409 argc = parse_options(argc, argv, options, top_usage, 0); 1410 if (argc) 1411 usage_with_options(top_usage, options); 1412 1413 if (target_pid != -1) 1414 target_tid = target_pid; 1415 1416 threads = thread_map__new(target_pid, target_tid); 1417 if (threads == NULL) { 1418 pr_err("Problems finding threads of monitor\n"); 1419 usage_with_options(top_usage, options); 1420 } 1421 1422 event_array = malloc((sizeof(struct pollfd) * 1423 MAX_NR_CPUS * MAX_COUNTERS * threads->nr)); 1424 if (!event_array) 1425 return -ENOMEM; 1426 1427 /* CPU and PID are mutually exclusive */ 1428 if (target_tid > 0 && cpu_list) { 1429 printf("WARNING: PID switch overriding CPU\n"); 1430 sleep(1); 1431 cpu_list = NULL; 1432 } 1433 1434 if (!nr_counters && perf_evsel_list__create_default() < 0) { 1435 pr_err("Not enough memory for event selector list\n"); 1436 return -ENOMEM; 1437 } 1438 1439 if (delay_secs < 1) 1440 delay_secs = 1; 1441 1442 /* 1443 * User specified count overrides default frequency. 1444 */ 1445 if (default_interval) 1446 freq = 0; 1447 else if (freq) { 1448 default_interval = freq; 1449 } else { 1450 fprintf(stderr, "frequency and count are zero, aborting\n"); 1451 exit(EXIT_FAILURE); 1452 } 1453 1454 if (target_tid != -1) 1455 cpus = cpu_map__dummy_new(); 1456 else 1457 cpus = cpu_map__new(cpu_list); 1458 1459 if (cpus == NULL) 1460 usage_with_options(top_usage, options); 1461 1462 list_for_each_entry(pos, &evsel_list, node) { 1463 if (perf_evsel__alloc_mmap_per_thread(pos, cpus->nr, threads->nr) < 0 || 1464 perf_evsel__alloc_fd(pos, cpus->nr, threads->nr) < 0) 1465 goto out_free_fd; 1466 /* 1467 * Fill in the ones not specifically initialized via -c: 1468 */ 1469 if (pos->attr.sample_period) 1470 continue; 1471 1472 pos->attr.sample_period = default_interval; 1473 } 1474 1475 sym_evsel = list_entry(evsel_list.next, struct perf_evsel, node); 1476 1477 symbol_conf.priv_size = (sizeof(struct sym_entry) + 1478 (nr_counters + 1) * sizeof(unsigned long)); 1479 1480 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL); 1481 if (symbol__init() < 0) 1482 return -1; 1483 1484 get_term_dimensions(&winsize); 1485 if (print_entries == 0) { 1486 update_print_entries(&winsize); 1487 signal(SIGWINCH, sig_winch_handler); 1488 } 1489 1490 status = __cmd_top(); 1491 out_free_fd: 1492 list_for_each_entry(pos, &evsel_list, node) 1493 perf_evsel__free_mmap(pos); 1494 perf_evsel_list__delete(); 1495 1496 return status; 1497 } 1498