| 1 | /* |
| 2 | * builtin-top.c |
| 3 | * |
| 4 | * Builtin top command: Display a continuously updated profile of |
| 5 | * any workload, CPU or specific PID. |
| 6 | * |
| 7 | * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com> |
| 8 | * |
| 9 | * Improvements and fixes by: |
| 10 | * |
| 11 | * Arjan van de Ven <arjan@linux.intel.com> |
| 12 | * Yanmin Zhang <yanmin.zhang@intel.com> |
| 13 | * Wu Fengguang <fengguang.wu@intel.com> |
| 14 | * Mike Galbraith <efault@gmx.de> |
| 15 | * Paul Mackerras <paulus@samba.org> |
| 16 | * |
| 17 | * Released under the GPL v2. (and only v2, not any later version) |
| 18 | */ |
| 19 | #include "builtin.h" |
| 20 | |
| 21 | #include "perf.h" |
| 22 | |
| 23 | #include "util/symbol.h" |
| 24 | #include "util/color.h" |
| 25 | #include "util/util.h" |
| 26 | #include "util/rbtree.h" |
| 27 | #include "util/parse-options.h" |
| 28 | #include "util/parse-events.h" |
| 29 | |
| 30 | #include <assert.h> |
| 31 | #include <fcntl.h> |
| 32 | |
| 33 | #include <stdio.h> |
| 34 | |
| 35 | #include <errno.h> |
| 36 | #include <time.h> |
| 37 | #include <sched.h> |
| 38 | #include <pthread.h> |
| 39 | |
| 40 | #include <sys/syscall.h> |
| 41 | #include <sys/ioctl.h> |
| 42 | #include <sys/poll.h> |
| 43 | #include <sys/prctl.h> |
| 44 | #include <sys/wait.h> |
| 45 | #include <sys/uio.h> |
| 46 | #include <sys/mman.h> |
| 47 | |
| 48 | #include <linux/unistd.h> |
| 49 | #include <linux/types.h> |
| 50 | |
| 51 | static int system_wide = 0; |
| 52 | |
| 53 | static __u64 default_event_id[MAX_COUNTERS] = { |
| 54 | EID(PERF_TYPE_SOFTWARE, PERF_COUNT_TASK_CLOCK), |
| 55 | EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CONTEXT_SWITCHES), |
| 56 | EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_MIGRATIONS), |
| 57 | EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS), |
| 58 | |
| 59 | EID(PERF_TYPE_HARDWARE, PERF_COUNT_CPU_CYCLES), |
| 60 | EID(PERF_TYPE_HARDWARE, PERF_COUNT_INSTRUCTIONS), |
| 61 | EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_REFERENCES), |
| 62 | EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_MISSES), |
| 63 | }; |
| 64 | static int default_interval = 100000; |
| 65 | static int event_count[MAX_COUNTERS]; |
| 66 | static int fd[MAX_NR_CPUS][MAX_COUNTERS]; |
| 67 | |
| 68 | static __u64 count_filter = 5; |
| 69 | static int print_entries = 15; |
| 70 | |
| 71 | static int target_pid = -1; |
| 72 | static int profile_cpu = -1; |
| 73 | static int nr_cpus = 0; |
| 74 | static unsigned int realtime_prio = 0; |
| 75 | static int group = 0; |
| 76 | static unsigned int page_size; |
| 77 | static unsigned int mmap_pages = 16; |
| 78 | static int freq = 0; |
| 79 | |
| 80 | static char *sym_filter; |
| 81 | static unsigned long filter_start; |
| 82 | static unsigned long filter_end; |
| 83 | |
| 84 | static int delay_secs = 2; |
| 85 | static int zero; |
| 86 | static int dump_symtab; |
| 87 | |
| 88 | static const unsigned int default_count[] = { |
| 89 | 1000000, |
| 90 | 1000000, |
| 91 | 10000, |
| 92 | 10000, |
| 93 | 1000000, |
| 94 | 10000, |
| 95 | }; |
| 96 | |
| 97 | /* |
| 98 | * Symbols |
| 99 | */ |
| 100 | |
| 101 | static uint64_t min_ip; |
| 102 | static uint64_t max_ip = -1ll; |
| 103 | |
| 104 | struct sym_entry { |
| 105 | struct rb_node rb_node; |
| 106 | struct list_head node; |
| 107 | unsigned long count[MAX_COUNTERS]; |
| 108 | unsigned long snap_count; |
| 109 | double weight; |
| 110 | int skip; |
| 111 | }; |
| 112 | |
| 113 | struct sym_entry *sym_filter_entry; |
| 114 | |
| 115 | struct dso *kernel_dso; |
| 116 | |
| 117 | /* |
| 118 | * Symbols will be added here in record_ip and will get out |
| 119 | * after decayed. |
| 120 | */ |
| 121 | static LIST_HEAD(active_symbols); |
| 122 | static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER; |
| 123 | |
| 124 | /* |
| 125 | * Ordering weight: count-1 * count-2 * ... / count-n |
| 126 | */ |
| 127 | static double sym_weight(const struct sym_entry *sym) |
| 128 | { |
| 129 | double weight = sym->snap_count; |
| 130 | int counter; |
| 131 | |
| 132 | for (counter = 1; counter < nr_counters-1; counter++) |
| 133 | weight *= sym->count[counter]; |
| 134 | |
| 135 | weight /= (sym->count[counter] + 1); |
| 136 | |
| 137 | return weight; |
| 138 | } |
| 139 | |
| 140 | static long events; |
| 141 | static long userspace_events; |
| 142 | static const char CONSOLE_CLEAR[] = "\e[H\e[2J"; |
| 143 | |
| 144 | static void __list_insert_active_sym(struct sym_entry *syme) |
| 145 | { |
| 146 | list_add(&syme->node, &active_symbols); |
| 147 | } |
| 148 | |
| 149 | static void list_remove_active_sym(struct sym_entry *syme) |
| 150 | { |
| 151 | pthread_mutex_lock(&active_symbols_lock); |
| 152 | list_del_init(&syme->node); |
| 153 | pthread_mutex_unlock(&active_symbols_lock); |
| 154 | } |
| 155 | |
| 156 | static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se) |
| 157 | { |
| 158 | struct rb_node **p = &tree->rb_node; |
| 159 | struct rb_node *parent = NULL; |
| 160 | struct sym_entry *iter; |
| 161 | |
| 162 | while (*p != NULL) { |
| 163 | parent = *p; |
| 164 | iter = rb_entry(parent, struct sym_entry, rb_node); |
| 165 | |
| 166 | if (se->weight > iter->weight) |
| 167 | p = &(*p)->rb_left; |
| 168 | else |
| 169 | p = &(*p)->rb_right; |
| 170 | } |
| 171 | |
| 172 | rb_link_node(&se->rb_node, parent, p); |
| 173 | rb_insert_color(&se->rb_node, tree); |
| 174 | } |
| 175 | |
| 176 | static void print_sym_table(void) |
| 177 | { |
| 178 | int printed = 0, j; |
| 179 | int counter; |
| 180 | float events_per_sec = events/delay_secs; |
| 181 | float kevents_per_sec = (events-userspace_events)/delay_secs; |
| 182 | float sum_kevents = 0.0; |
| 183 | struct sym_entry *syme, *n; |
| 184 | struct rb_root tmp = RB_ROOT; |
| 185 | struct rb_node *nd; |
| 186 | |
| 187 | events = userspace_events = 0; |
| 188 | |
| 189 | /* Sort the active symbols */ |
| 190 | pthread_mutex_lock(&active_symbols_lock); |
| 191 | syme = list_entry(active_symbols.next, struct sym_entry, node); |
| 192 | pthread_mutex_unlock(&active_symbols_lock); |
| 193 | |
| 194 | list_for_each_entry_safe_from(syme, n, &active_symbols, node) { |
| 195 | syme->snap_count = syme->count[0]; |
| 196 | if (syme->snap_count != 0) { |
| 197 | syme->weight = sym_weight(syme); |
| 198 | rb_insert_active_sym(&tmp, syme); |
| 199 | sum_kevents += syme->snap_count; |
| 200 | |
| 201 | for (j = 0; j < nr_counters; j++) |
| 202 | syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8; |
| 203 | } else |
| 204 | list_remove_active_sym(syme); |
| 205 | } |
| 206 | |
| 207 | puts(CONSOLE_CLEAR); |
| 208 | |
| 209 | printf( |
| 210 | "------------------------------------------------------------------------------\n"); |
| 211 | printf( " PerfTop:%8.0f irqs/sec kernel:%4.1f%% [", |
| 212 | events_per_sec, |
| 213 | 100.0 - (100.0*((events_per_sec-kevents_per_sec)/events_per_sec))); |
| 214 | |
| 215 | if (nr_counters == 1) |
| 216 | printf("%d ", event_count[0]); |
| 217 | |
| 218 | for (counter = 0; counter < nr_counters; counter++) { |
| 219 | if (counter) |
| 220 | printf("/"); |
| 221 | |
| 222 | printf("%s", event_name(counter)); |
| 223 | } |
| 224 | |
| 225 | printf( "], "); |
| 226 | |
| 227 | if (target_pid != -1) |
| 228 | printf(" (target_pid: %d", target_pid); |
| 229 | else |
| 230 | printf(" (all"); |
| 231 | |
| 232 | if (profile_cpu != -1) |
| 233 | printf(", cpu: %d)\n", profile_cpu); |
| 234 | else { |
| 235 | if (target_pid != -1) |
| 236 | printf(")\n"); |
| 237 | else |
| 238 | printf(", %d CPUs)\n", nr_cpus); |
| 239 | } |
| 240 | |
| 241 | printf("------------------------------------------------------------------------------\n\n"); |
| 242 | |
| 243 | if (nr_counters == 1) |
| 244 | printf(" events pcnt"); |
| 245 | else |
| 246 | printf(" weight events pcnt"); |
| 247 | |
| 248 | printf(" RIP kernel function\n" |
| 249 | " ______ ______ _____ ________________ _______________\n\n" |
| 250 | ); |
| 251 | |
| 252 | for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) { |
| 253 | struct sym_entry *syme = rb_entry(nd, struct sym_entry, rb_node); |
| 254 | struct symbol *sym = (struct symbol *)(syme + 1); |
| 255 | char *color = PERF_COLOR_NORMAL; |
| 256 | double pcnt; |
| 257 | |
| 258 | if (++printed > print_entries || syme->snap_count < count_filter) |
| 259 | continue; |
| 260 | |
| 261 | pcnt = 100.0 - (100.0 * ((sum_kevents - syme->snap_count) / |
| 262 | sum_kevents)); |
| 263 | |
| 264 | /* |
| 265 | * We color high-overhead entries in red, low-overhead |
| 266 | * entries in green - and keep the middle ground normal: |
| 267 | */ |
| 268 | if (pcnt >= 5.0) |
| 269 | color = PERF_COLOR_RED; |
| 270 | if (pcnt < 0.5) |
| 271 | color = PERF_COLOR_GREEN; |
| 272 | |
| 273 | if (nr_counters == 1) |
| 274 | printf("%19.2f - ", syme->weight); |
| 275 | else |
| 276 | printf("%8.1f %10ld - ", syme->weight, syme->snap_count); |
| 277 | |
| 278 | color_fprintf(stdout, color, "%4.1f%%", pcnt); |
| 279 | printf(" - %016llx : %s\n", sym->start, sym->name); |
| 280 | } |
| 281 | } |
| 282 | |
| 283 | static void *display_thread(void *arg) |
| 284 | { |
| 285 | struct pollfd stdin_poll = { .fd = 0, .events = POLLIN }; |
| 286 | int delay_msecs = delay_secs * 1000; |
| 287 | |
| 288 | printf("PerfTop refresh period: %d seconds\n", delay_secs); |
| 289 | |
| 290 | do { |
| 291 | print_sym_table(); |
| 292 | } while (!poll(&stdin_poll, 1, delay_msecs) == 1); |
| 293 | |
| 294 | printf("key pressed - exiting.\n"); |
| 295 | exit(0); |
| 296 | |
| 297 | return NULL; |
| 298 | } |
| 299 | |
| 300 | static int symbol_filter(struct dso *self, struct symbol *sym) |
| 301 | { |
| 302 | static int filter_match; |
| 303 | struct sym_entry *syme; |
| 304 | const char *name = sym->name; |
| 305 | |
| 306 | if (!strcmp(name, "_text") || |
| 307 | !strcmp(name, "_etext") || |
| 308 | !strcmp(name, "_sinittext") || |
| 309 | !strncmp("init_module", name, 11) || |
| 310 | !strncmp("cleanup_module", name, 14) || |
| 311 | strstr(name, "_text_start") || |
| 312 | strstr(name, "_text_end")) |
| 313 | return 1; |
| 314 | |
| 315 | syme = dso__sym_priv(self, sym); |
| 316 | /* Tag events to be skipped. */ |
| 317 | if (!strcmp("default_idle", name) || |
| 318 | !strcmp("cpu_idle", name) || |
| 319 | !strcmp("enter_idle", name) || |
| 320 | !strcmp("exit_idle", name) || |
| 321 | !strcmp("mwait_idle", name)) |
| 322 | syme->skip = 1; |
| 323 | |
| 324 | if (filter_match == 1) { |
| 325 | filter_end = sym->start; |
| 326 | filter_match = -1; |
| 327 | if (filter_end - filter_start > 10000) { |
| 328 | fprintf(stderr, |
| 329 | "hm, too large filter symbol <%s> - skipping.\n", |
| 330 | sym_filter); |
| 331 | fprintf(stderr, "symbol filter start: %016lx\n", |
| 332 | filter_start); |
| 333 | fprintf(stderr, " end: %016lx\n", |
| 334 | filter_end); |
| 335 | filter_end = filter_start = 0; |
| 336 | sym_filter = NULL; |
| 337 | sleep(1); |
| 338 | } |
| 339 | } |
| 340 | |
| 341 | if (filter_match == 0 && sym_filter && !strcmp(name, sym_filter)) { |
| 342 | filter_match = 1; |
| 343 | filter_start = sym->start; |
| 344 | } |
| 345 | |
| 346 | |
| 347 | return 0; |
| 348 | } |
| 349 | |
| 350 | static int parse_symbols(void) |
| 351 | { |
| 352 | struct rb_node *node; |
| 353 | struct symbol *sym; |
| 354 | |
| 355 | kernel_dso = dso__new("[kernel]", sizeof(struct sym_entry)); |
| 356 | if (kernel_dso == NULL) |
| 357 | return -1; |
| 358 | |
| 359 | if (dso__load_kernel(kernel_dso, NULL, symbol_filter, 1) != 0) |
| 360 | goto out_delete_dso; |
| 361 | |
| 362 | node = rb_first(&kernel_dso->syms); |
| 363 | sym = rb_entry(node, struct symbol, rb_node); |
| 364 | min_ip = sym->start; |
| 365 | |
| 366 | node = rb_last(&kernel_dso->syms); |
| 367 | sym = rb_entry(node, struct symbol, rb_node); |
| 368 | max_ip = sym->end; |
| 369 | |
| 370 | if (dump_symtab) |
| 371 | dso__fprintf(kernel_dso, stderr); |
| 372 | |
| 373 | return 0; |
| 374 | |
| 375 | out_delete_dso: |
| 376 | dso__delete(kernel_dso); |
| 377 | kernel_dso = NULL; |
| 378 | return -1; |
| 379 | } |
| 380 | |
| 381 | #define TRACE_COUNT 3 |
| 382 | |
| 383 | /* |
| 384 | * Binary search in the histogram table and record the hit: |
| 385 | */ |
| 386 | static void record_ip(uint64_t ip, int counter) |
| 387 | { |
| 388 | struct symbol *sym = dso__find_symbol(kernel_dso, ip); |
| 389 | |
| 390 | if (sym != NULL) { |
| 391 | struct sym_entry *syme = dso__sym_priv(kernel_dso, sym); |
| 392 | |
| 393 | if (!syme->skip) { |
| 394 | syme->count[counter]++; |
| 395 | pthread_mutex_lock(&active_symbols_lock); |
| 396 | if (list_empty(&syme->node) || !syme->node.next) |
| 397 | __list_insert_active_sym(syme); |
| 398 | pthread_mutex_unlock(&active_symbols_lock); |
| 399 | return; |
| 400 | } |
| 401 | } |
| 402 | |
| 403 | events--; |
| 404 | } |
| 405 | |
| 406 | static void process_event(uint64_t ip, int counter) |
| 407 | { |
| 408 | events++; |
| 409 | |
| 410 | if (ip < min_ip || ip > max_ip) { |
| 411 | userspace_events++; |
| 412 | return; |
| 413 | } |
| 414 | |
| 415 | record_ip(ip, counter); |
| 416 | } |
| 417 | |
| 418 | struct mmap_data { |
| 419 | int counter; |
| 420 | void *base; |
| 421 | unsigned int mask; |
| 422 | unsigned int prev; |
| 423 | }; |
| 424 | |
| 425 | static unsigned int mmap_read_head(struct mmap_data *md) |
| 426 | { |
| 427 | struct perf_counter_mmap_page *pc = md->base; |
| 428 | int head; |
| 429 | |
| 430 | head = pc->data_head; |
| 431 | rmb(); |
| 432 | |
| 433 | return head; |
| 434 | } |
| 435 | |
| 436 | struct timeval last_read, this_read; |
| 437 | |
| 438 | static void mmap_read(struct mmap_data *md) |
| 439 | { |
| 440 | unsigned int head = mmap_read_head(md); |
| 441 | unsigned int old = md->prev; |
| 442 | unsigned char *data = md->base + page_size; |
| 443 | int diff; |
| 444 | |
| 445 | gettimeofday(&this_read, NULL); |
| 446 | |
| 447 | /* |
| 448 | * If we're further behind than half the buffer, there's a chance |
| 449 | * the writer will bite our tail and screw up the events under us. |
| 450 | * |
| 451 | * If we somehow ended up ahead of the head, we got messed up. |
| 452 | * |
| 453 | * In either case, truncate and restart at head. |
| 454 | */ |
| 455 | diff = head - old; |
| 456 | if (diff > md->mask / 2 || diff < 0) { |
| 457 | struct timeval iv; |
| 458 | unsigned long msecs; |
| 459 | |
| 460 | timersub(&this_read, &last_read, &iv); |
| 461 | msecs = iv.tv_sec*1000 + iv.tv_usec/1000; |
| 462 | |
| 463 | fprintf(stderr, "WARNING: failed to keep up with mmap data." |
| 464 | " Last read %lu msecs ago.\n", msecs); |
| 465 | |
| 466 | /* |
| 467 | * head points to a known good entry, start there. |
| 468 | */ |
| 469 | old = head; |
| 470 | } |
| 471 | |
| 472 | last_read = this_read; |
| 473 | |
| 474 | for (; old != head;) { |
| 475 | struct ip_event { |
| 476 | struct perf_event_header header; |
| 477 | __u64 ip; |
| 478 | __u32 pid, target_pid; |
| 479 | }; |
| 480 | struct mmap_event { |
| 481 | struct perf_event_header header; |
| 482 | __u32 pid, target_pid; |
| 483 | __u64 start; |
| 484 | __u64 len; |
| 485 | __u64 pgoff; |
| 486 | char filename[PATH_MAX]; |
| 487 | }; |
| 488 | |
| 489 | typedef union event_union { |
| 490 | struct perf_event_header header; |
| 491 | struct ip_event ip; |
| 492 | struct mmap_event mmap; |
| 493 | } event_t; |
| 494 | |
| 495 | event_t *event = (event_t *)&data[old & md->mask]; |
| 496 | |
| 497 | event_t event_copy; |
| 498 | |
| 499 | size_t size = event->header.size; |
| 500 | |
| 501 | /* |
| 502 | * Event straddles the mmap boundary -- header should always |
| 503 | * be inside due to u64 alignment of output. |
| 504 | */ |
| 505 | if ((old & md->mask) + size != ((old + size) & md->mask)) { |
| 506 | unsigned int offset = old; |
| 507 | unsigned int len = min(sizeof(*event), size), cpy; |
| 508 | void *dst = &event_copy; |
| 509 | |
| 510 | do { |
| 511 | cpy = min(md->mask + 1 - (offset & md->mask), len); |
| 512 | memcpy(dst, &data[offset & md->mask], cpy); |
| 513 | offset += cpy; |
| 514 | dst += cpy; |
| 515 | len -= cpy; |
| 516 | } while (len); |
| 517 | |
| 518 | event = &event_copy; |
| 519 | } |
| 520 | |
| 521 | old += size; |
| 522 | |
| 523 | if (event->header.misc & PERF_EVENT_MISC_OVERFLOW) { |
| 524 | if (event->header.type & PERF_SAMPLE_IP) |
| 525 | process_event(event->ip.ip, md->counter); |
| 526 | } |
| 527 | } |
| 528 | |
| 529 | md->prev = old; |
| 530 | } |
| 531 | |
| 532 | static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS]; |
| 533 | static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS]; |
| 534 | |
| 535 | static int __cmd_top(void) |
| 536 | { |
| 537 | struct perf_counter_attr attr; |
| 538 | pthread_t thread; |
| 539 | int i, counter, group_fd, nr_poll = 0; |
| 540 | unsigned int cpu; |
| 541 | int ret; |
| 542 | |
| 543 | for (i = 0; i < nr_cpus; i++) { |
| 544 | group_fd = -1; |
| 545 | for (counter = 0; counter < nr_counters; counter++) { |
| 546 | |
| 547 | cpu = profile_cpu; |
| 548 | if (target_pid == -1 && profile_cpu == -1) |
| 549 | cpu = i; |
| 550 | |
| 551 | memset(&attr, 0, sizeof(attr)); |
| 552 | attr.config = event_id[counter]; |
| 553 | attr.sample_period = event_count[counter]; |
| 554 | attr.sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID; |
| 555 | attr.freq = freq; |
| 556 | |
| 557 | fd[i][counter] = sys_perf_counter_open(&attr, target_pid, cpu, group_fd, 0); |
| 558 | if (fd[i][counter] < 0) { |
| 559 | int err = errno; |
| 560 | |
| 561 | error("syscall returned with %d (%s)\n", |
| 562 | fd[i][counter], strerror(err)); |
| 563 | if (err == EPERM) |
| 564 | printf("Are you root?\n"); |
| 565 | exit(-1); |
| 566 | } |
| 567 | assert(fd[i][counter] >= 0); |
| 568 | fcntl(fd[i][counter], F_SETFL, O_NONBLOCK); |
| 569 | |
| 570 | /* |
| 571 | * First counter acts as the group leader: |
| 572 | */ |
| 573 | if (group && group_fd == -1) |
| 574 | group_fd = fd[i][counter]; |
| 575 | |
| 576 | event_array[nr_poll].fd = fd[i][counter]; |
| 577 | event_array[nr_poll].events = POLLIN; |
| 578 | nr_poll++; |
| 579 | |
| 580 | mmap_array[i][counter].counter = counter; |
| 581 | mmap_array[i][counter].prev = 0; |
| 582 | mmap_array[i][counter].mask = mmap_pages*page_size - 1; |
| 583 | mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size, |
| 584 | PROT_READ, MAP_SHARED, fd[i][counter], 0); |
| 585 | if (mmap_array[i][counter].base == MAP_FAILED) |
| 586 | die("failed to mmap with %d (%s)\n", errno, strerror(errno)); |
| 587 | } |
| 588 | } |
| 589 | |
| 590 | if (pthread_create(&thread, NULL, display_thread, NULL)) { |
| 591 | printf("Could not create display thread.\n"); |
| 592 | exit(-1); |
| 593 | } |
| 594 | |
| 595 | if (realtime_prio) { |
| 596 | struct sched_param param; |
| 597 | |
| 598 | param.sched_priority = realtime_prio; |
| 599 | if (sched_setscheduler(0, SCHED_FIFO, ¶m)) { |
| 600 | printf("Could not set realtime priority.\n"); |
| 601 | exit(-1); |
| 602 | } |
| 603 | } |
| 604 | |
| 605 | while (1) { |
| 606 | int hits = events; |
| 607 | |
| 608 | for (i = 0; i < nr_cpus; i++) { |
| 609 | for (counter = 0; counter < nr_counters; counter++) |
| 610 | mmap_read(&mmap_array[i][counter]); |
| 611 | } |
| 612 | |
| 613 | if (hits == events) |
| 614 | ret = poll(event_array, nr_poll, 100); |
| 615 | } |
| 616 | |
| 617 | return 0; |
| 618 | } |
| 619 | |
| 620 | static const char * const top_usage[] = { |
| 621 | "perf top [<options>]", |
| 622 | NULL |
| 623 | }; |
| 624 | |
| 625 | static char events_help_msg[EVENTS_HELP_MAX]; |
| 626 | |
| 627 | static const struct option options[] = { |
| 628 | OPT_CALLBACK('e', "event", NULL, "event", |
| 629 | events_help_msg, parse_events), |
| 630 | OPT_INTEGER('c', "count", &default_interval, |
| 631 | "event period to sample"), |
| 632 | OPT_INTEGER('p', "pid", &target_pid, |
| 633 | "profile events on existing pid"), |
| 634 | OPT_BOOLEAN('a', "all-cpus", &system_wide, |
| 635 | "system-wide collection from all CPUs"), |
| 636 | OPT_INTEGER('C', "CPU", &profile_cpu, |
| 637 | "CPU to profile on"), |
| 638 | OPT_INTEGER('m', "mmap-pages", &mmap_pages, |
| 639 | "number of mmap data pages"), |
| 640 | OPT_INTEGER('r', "realtime", &realtime_prio, |
| 641 | "collect data with this RT SCHED_FIFO priority"), |
| 642 | OPT_INTEGER('d', "delay", &delay_secs, |
| 643 | "number of seconds to delay between refreshes"), |
| 644 | OPT_BOOLEAN('D', "dump-symtab", &dump_symtab, |
| 645 | "dump the symbol table used for profiling"), |
| 646 | OPT_INTEGER('f', "count-filter", &count_filter, |
| 647 | "only display functions with more events than this"), |
| 648 | OPT_BOOLEAN('g', "group", &group, |
| 649 | "put the counters into a counter group"), |
| 650 | OPT_STRING('s', "sym-filter", &sym_filter, "pattern", |
| 651 | "only display symbols matchig this pattern"), |
| 652 | OPT_BOOLEAN('z', "zero", &group, |
| 653 | "zero history across updates"), |
| 654 | OPT_INTEGER('F', "freq", &freq, |
| 655 | "profile at this frequency"), |
| 656 | OPT_INTEGER('E', "entries", &print_entries, |
| 657 | "display this many functions"), |
| 658 | OPT_END() |
| 659 | }; |
| 660 | |
| 661 | int cmd_top(int argc, const char **argv, const char *prefix) |
| 662 | { |
| 663 | int counter; |
| 664 | |
| 665 | page_size = sysconf(_SC_PAGE_SIZE); |
| 666 | |
| 667 | create_events_help(events_help_msg); |
| 668 | memcpy(event_id, default_event_id, sizeof(default_event_id)); |
| 669 | |
| 670 | argc = parse_options(argc, argv, options, top_usage, 0); |
| 671 | if (argc) |
| 672 | usage_with_options(top_usage, options); |
| 673 | |
| 674 | if (freq) { |
| 675 | default_interval = freq; |
| 676 | freq = 1; |
| 677 | } |
| 678 | |
| 679 | /* CPU and PID are mutually exclusive */ |
| 680 | if (target_pid != -1 && profile_cpu != -1) { |
| 681 | printf("WARNING: PID switch overriding CPU\n"); |
| 682 | sleep(1); |
| 683 | profile_cpu = -1; |
| 684 | } |
| 685 | |
| 686 | if (!nr_counters) { |
| 687 | nr_counters = 1; |
| 688 | event_id[0] = 0; |
| 689 | } |
| 690 | |
| 691 | for (counter = 0; counter < nr_counters; counter++) { |
| 692 | if (event_count[counter]) |
| 693 | continue; |
| 694 | |
| 695 | event_count[counter] = default_interval; |
| 696 | } |
| 697 | |
| 698 | nr_cpus = sysconf(_SC_NPROCESSORS_ONLN); |
| 699 | assert(nr_cpus <= MAX_NR_CPUS); |
| 700 | assert(nr_cpus >= 0); |
| 701 | |
| 702 | if (target_pid != -1 || profile_cpu != -1) |
| 703 | nr_cpus = 1; |
| 704 | |
| 705 | parse_symbols(); |
| 706 | |
| 707 | return __cmd_top(); |
| 708 | } |