xref: /linux/tools/perf/builtin-top.c (revision 6000fc4d6f3e55ad52cce8d76317187fe01af2aa)
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 <linux/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 #include <termios.h>
35 #include <unistd.h>
36 
37 #include <errno.h>
38 #include <time.h>
39 #include <sched.h>
40 #include <pthread.h>
41 
42 #include <sys/syscall.h>
43 #include <sys/ioctl.h>
44 #include <sys/poll.h>
45 #include <sys/prctl.h>
46 #include <sys/wait.h>
47 #include <sys/uio.h>
48 #include <sys/mman.h>
49 
50 #include <linux/unistd.h>
51 #include <linux/types.h>
52 
53 static int			fd[MAX_NR_CPUS][MAX_COUNTERS];
54 
55 static int			system_wide			=  0;
56 
57 static int			default_interval		= 100000;
58 
59 static int			count_filter			=  5;
60 static int			print_entries			= 15;
61 
62 static int			target_pid			= -1;
63 static int			inherit				=  0;
64 static int			profile_cpu			= -1;
65 static int			nr_cpus				=  0;
66 static unsigned int		realtime_prio			=  0;
67 static int			group				=  0;
68 static unsigned int		page_size;
69 static unsigned int		mmap_pages			= 16;
70 static int			freq				=  0;
71 static int			verbose				=  0;
72 static char			*vmlinux			=  NULL;
73 
74 static int			delay_secs			=  2;
75 static int			zero;
76 static int			dump_symtab;
77 
78 /*
79  * Source
80  */
81 
82 struct source_line {
83 	u64			eip;
84 	unsigned long		count[MAX_COUNTERS];
85 	char			*line;
86 	struct source_line	*next;
87 };
88 
89 static char			*sym_filter			=  NULL;
90 struct sym_entry		*sym_filter_entry		=  NULL;
91 static int			sym_pcnt_filter			=  5;
92 static int			sym_counter			=  0;
93 static int			display_weighted		= -1;
94 
95 /*
96  * Symbols
97  */
98 
99 static u64			min_ip;
100 static u64			max_ip = -1ll;
101 
102 struct sym_entry {
103 	struct rb_node		rb_node;
104 	struct list_head	node;
105 	unsigned long		count[MAX_COUNTERS];
106 	unsigned long		snap_count;
107 	double			weight;
108 	int			skip;
109 	struct source_line	*source;
110 	struct source_line	*lines;
111 	struct source_line	**lines_tail;
112 	pthread_mutex_t		source_lock;
113 };
114 
115 /*
116  * Source functions
117  */
118 
119 static void parse_source(struct sym_entry *syme)
120 {
121 	struct symbol *sym;
122 	struct module *module;
123 	struct section *section = NULL;
124 	FILE *file;
125 	char command[PATH_MAX*2], *path = vmlinux;
126 	u64 start, end, len;
127 
128 	if (!syme)
129 		return;
130 
131 	if (syme->lines) {
132 		pthread_mutex_lock(&syme->source_lock);
133 		goto out_assign;
134 	}
135 
136 	sym = (struct symbol *)(syme + 1);
137 	module = sym->module;
138 
139 	if (module)
140 		path = module->path;
141 	if (!path)
142 		return;
143 
144 	start = sym->obj_start;
145 	if (!start)
146 		start = sym->start;
147 
148 	if (module) {
149 		section = module->sections->find_section(module->sections, ".text");
150 		if (section)
151 			start -= section->vma;
152 	}
153 
154 	end = start + sym->end - sym->start + 1;
155 	len = sym->end - sym->start;
156 
157 	sprintf(command, "objdump --start-address=0x%016Lx --stop-address=0x%016Lx -dS %s", start, end, path);
158 
159 	file = popen(command, "r");
160 	if (!file)
161 		return;
162 
163 	pthread_mutex_lock(&syme->source_lock);
164 	syme->lines_tail = &syme->lines;
165 	while (!feof(file)) {
166 		struct source_line *src;
167 		size_t dummy = 0;
168 		char *c;
169 
170 		src = malloc(sizeof(struct source_line));
171 		assert(src != NULL);
172 		memset(src, 0, sizeof(struct source_line));
173 
174 		if (getline(&src->line, &dummy, file) < 0)
175 			break;
176 		if (!src->line)
177 			break;
178 
179 		c = strchr(src->line, '\n');
180 		if (c)
181 			*c = 0;
182 
183 		src->next = NULL;
184 		*syme->lines_tail = src;
185 		syme->lines_tail = &src->next;
186 
187 		if (strlen(src->line)>8 && src->line[8] == ':') {
188 			src->eip = strtoull(src->line, NULL, 16);
189 			if (section)
190 				src->eip += section->vma;
191 		}
192 		if (strlen(src->line)>8 && src->line[16] == ':') {
193 			src->eip = strtoull(src->line, NULL, 16);
194 			if (section)
195 				src->eip += section->vma;
196 		}
197 	}
198 	pclose(file);
199 out_assign:
200 	sym_filter_entry = syme;
201 	pthread_mutex_unlock(&syme->source_lock);
202 }
203 
204 static void __zero_source_counters(struct sym_entry *syme)
205 {
206 	int i;
207 	struct source_line *line;
208 
209 	line = syme->lines;
210 	while (line) {
211 		for (i = 0; i < nr_counters; i++)
212 			line->count[i] = 0;
213 		line = line->next;
214 	}
215 }
216 
217 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
218 {
219 	struct source_line *line;
220 
221 	if (syme != sym_filter_entry)
222 		return;
223 
224 	if (pthread_mutex_trylock(&syme->source_lock))
225 		return;
226 
227 	if (!syme->source)
228 		goto out_unlock;
229 
230 	for (line = syme->lines; line; line = line->next) {
231 		if (line->eip == ip) {
232 			line->count[counter]++;
233 			break;
234 		}
235 		if (line->eip > ip)
236 			break;
237 	}
238 out_unlock:
239 	pthread_mutex_unlock(&syme->source_lock);
240 }
241 
242 static void lookup_sym_source(struct sym_entry *syme)
243 {
244 	struct symbol *symbol = (struct symbol *)(syme + 1);
245 	struct source_line *line;
246 	char pattern[PATH_MAX];
247 	char *idx;
248 
249 	sprintf(pattern, "<%s>:", symbol->name);
250 
251 	if (symbol->module) {
252 		idx = strstr(pattern, "\t");
253 		if (idx)
254 			*idx = 0;
255 	}
256 
257 	pthread_mutex_lock(&syme->source_lock);
258 	for (line = syme->lines; line; line = line->next) {
259 		if (strstr(line->line, pattern)) {
260 			syme->source = line;
261 			break;
262 		}
263 	}
264 	pthread_mutex_unlock(&syme->source_lock);
265 }
266 
267 static void show_lines(struct source_line *queue, int count, int total)
268 {
269 	int i;
270 	struct source_line *line;
271 
272 	line = queue;
273 	for (i = 0; i < count; i++) {
274 		float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
275 
276 		printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
277 		line = line->next;
278 	}
279 }
280 
281 #define TRACE_COUNT     3
282 
283 static void show_details(struct sym_entry *syme)
284 {
285 	struct symbol *symbol;
286 	struct source_line *line;
287 	struct source_line *line_queue = NULL;
288 	int displayed = 0;
289 	int line_queue_count = 0, total = 0, more = 0;
290 
291 	if (!syme)
292 		return;
293 
294 	if (!syme->source)
295 		lookup_sym_source(syme);
296 
297 	if (!syme->source)
298 		return;
299 
300 	symbol = (struct symbol *)(syme + 1);
301 	printf("Showing %s for %s\n", event_name(sym_counter), symbol->name);
302 	printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
303 
304 	pthread_mutex_lock(&syme->source_lock);
305 	line = syme->source;
306 	while (line) {
307 		total += line->count[sym_counter];
308 		line = line->next;
309 	}
310 
311 	line = syme->source;
312 	while (line) {
313 		float pcnt = 0.0;
314 
315 		if (!line_queue_count)
316 			line_queue = line;
317 		line_queue_count++;
318 
319 		if (line->count[sym_counter])
320 			pcnt = 100.0 * line->count[sym_counter] / (float)total;
321 		if (pcnt >= (float)sym_pcnt_filter) {
322 			if (displayed <= print_entries)
323 				show_lines(line_queue, line_queue_count, total);
324 			else more++;
325 			displayed += line_queue_count;
326 			line_queue_count = 0;
327 			line_queue = NULL;
328 		} else if (line_queue_count > TRACE_COUNT) {
329 			line_queue = line_queue->next;
330 			line_queue_count--;
331 		}
332 
333 		line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
334 		line = line->next;
335 	}
336 	pthread_mutex_unlock(&syme->source_lock);
337 	if (more)
338 		printf("%d lines not displayed, maybe increase display entries [e]\n", more);
339 }
340 
341 struct dso			*kernel_dso;
342 
343 /*
344  * Symbols will be added here in record_ip and will get out
345  * after decayed.
346  */
347 static LIST_HEAD(active_symbols);
348 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
349 
350 /*
351  * Ordering weight: count-1 * count-2 * ... / count-n
352  */
353 static double sym_weight(const struct sym_entry *sym)
354 {
355 	double weight = sym->snap_count;
356 	int counter;
357 
358 	if (!display_weighted)
359 		return weight;
360 
361 	for (counter = 1; counter < nr_counters-1; counter++)
362 		weight *= sym->count[counter];
363 
364 	weight /= (sym->count[counter] + 1);
365 
366 	return weight;
367 }
368 
369 static long			samples;
370 static long			userspace_samples;
371 static const char		CONSOLE_CLEAR[] = "";
372 
373 static void __list_insert_active_sym(struct sym_entry *syme)
374 {
375 	list_add(&syme->node, &active_symbols);
376 }
377 
378 static void list_remove_active_sym(struct sym_entry *syme)
379 {
380 	pthread_mutex_lock(&active_symbols_lock);
381 	list_del_init(&syme->node);
382 	pthread_mutex_unlock(&active_symbols_lock);
383 }
384 
385 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
386 {
387 	struct rb_node **p = &tree->rb_node;
388 	struct rb_node *parent = NULL;
389 	struct sym_entry *iter;
390 
391 	while (*p != NULL) {
392 		parent = *p;
393 		iter = rb_entry(parent, struct sym_entry, rb_node);
394 
395 		if (se->weight > iter->weight)
396 			p = &(*p)->rb_left;
397 		else
398 			p = &(*p)->rb_right;
399 	}
400 
401 	rb_link_node(&se->rb_node, parent, p);
402 	rb_insert_color(&se->rb_node, tree);
403 }
404 
405 static void print_sym_table(void)
406 {
407 	int printed = 0, j;
408 	int counter, snap = !display_weighted ? sym_counter : 0;
409 	float samples_per_sec = samples/delay_secs;
410 	float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
411 	float sum_ksamples = 0.0;
412 	struct sym_entry *syme, *n;
413 	struct rb_root tmp = RB_ROOT;
414 	struct rb_node *nd;
415 
416 	samples = userspace_samples = 0;
417 
418 	/* Sort the active symbols */
419 	pthread_mutex_lock(&active_symbols_lock);
420 	syme = list_entry(active_symbols.next, struct sym_entry, node);
421 	pthread_mutex_unlock(&active_symbols_lock);
422 
423 	list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
424 		syme->snap_count = syme->count[snap];
425 		if (syme->snap_count != 0) {
426 			syme->weight = sym_weight(syme);
427 			rb_insert_active_sym(&tmp, syme);
428 			sum_ksamples += syme->snap_count;
429 
430 			for (j = 0; j < nr_counters; j++)
431 				syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
432 		} else
433 			list_remove_active_sym(syme);
434 	}
435 
436 	puts(CONSOLE_CLEAR);
437 
438 	printf(
439 "------------------------------------------------------------------------------\n");
440 	printf( "   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% [",
441 		samples_per_sec,
442 		100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
443 
444 	if (nr_counters == 1 || !display_weighted) {
445 		printf("%Ld", (u64)attrs[0].sample_period);
446 		if (freq)
447 			printf("Hz ");
448 		else
449 			printf(" ");
450 	}
451 
452 	if (!display_weighted)
453 		printf("%s", event_name(sym_counter));
454 	else for (counter = 0; counter < nr_counters; counter++) {
455 		if (counter)
456 			printf("/");
457 
458 		printf("%s", event_name(counter));
459 	}
460 
461 	printf( "], ");
462 
463 	if (target_pid != -1)
464 		printf(" (target_pid: %d", target_pid);
465 	else
466 		printf(" (all");
467 
468 	if (profile_cpu != -1)
469 		printf(", cpu: %d)\n", profile_cpu);
470 	else {
471 		if (target_pid != -1)
472 			printf(")\n");
473 		else
474 			printf(", %d CPUs)\n", nr_cpus);
475 	}
476 
477 	printf("------------------------------------------------------------------------------\n\n");
478 
479 	if (sym_filter_entry) {
480 		show_details(sym_filter_entry);
481 		return;
482 	}
483 
484 	if (nr_counters == 1)
485 		printf("             samples    pcnt");
486 	else
487 		printf("  weight     samples    pcnt");
488 
489 	printf("         RIP          kernel function\n"
490 	       	       "  ______     _______   _____   ________________   _______________\n\n"
491 	);
492 
493 	for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
494 		struct sym_entry *syme = rb_entry(nd, struct sym_entry, rb_node);
495 		struct symbol *sym = (struct symbol *)(syme + 1);
496 		double pcnt;
497 
498 		if (++printed > print_entries || (int)syme->snap_count < count_filter)
499 			continue;
500 
501 		pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
502 					 sum_ksamples));
503 
504 		if (nr_counters == 1 || !display_weighted)
505 			printf("%20.2f - ", syme->weight);
506 		else
507 			printf("%9.1f %10ld - ", syme->weight, syme->snap_count);
508 
509 		percent_color_fprintf(stdout, "%4.1f%%", pcnt);
510 		printf(" - %016llx : %s", sym->start, sym->name);
511 		if (sym->module)
512 			printf("\t[%s]", sym->module->name);
513 		printf("\n");
514 	}
515 }
516 
517 static void prompt_integer(int *target, const char *msg)
518 {
519 	char *buf = malloc(0), *p;
520 	size_t dummy = 0;
521 	int tmp;
522 
523 	fprintf(stdout, "\n%s: ", msg);
524 	if (getline(&buf, &dummy, stdin) < 0)
525 		return;
526 
527 	p = strchr(buf, '\n');
528 	if (p)
529 		*p = 0;
530 
531 	p = buf;
532 	while(*p) {
533 		if (!isdigit(*p))
534 			goto out_free;
535 		p++;
536 	}
537 	tmp = strtoul(buf, NULL, 10);
538 	*target = tmp;
539 out_free:
540 	free(buf);
541 }
542 
543 static void prompt_percent(int *target, const char *msg)
544 {
545 	int tmp = 0;
546 
547 	prompt_integer(&tmp, msg);
548 	if (tmp >= 0 && tmp <= 100)
549 		*target = tmp;
550 }
551 
552 static void prompt_symbol(struct sym_entry **target, const char *msg)
553 {
554 	char *buf = malloc(0), *p;
555 	struct sym_entry *syme = *target, *n, *found = NULL;
556 	size_t dummy = 0;
557 
558 	/* zero counters of active symbol */
559 	if (syme) {
560 		pthread_mutex_lock(&syme->source_lock);
561 		__zero_source_counters(syme);
562 		*target = NULL;
563 		pthread_mutex_unlock(&syme->source_lock);
564 	}
565 
566 	fprintf(stdout, "\n%s: ", msg);
567 	if (getline(&buf, &dummy, stdin) < 0)
568 		goto out_free;
569 
570 	p = strchr(buf, '\n');
571 	if (p)
572 		*p = 0;
573 
574 	pthread_mutex_lock(&active_symbols_lock);
575 	syme = list_entry(active_symbols.next, struct sym_entry, node);
576 	pthread_mutex_unlock(&active_symbols_lock);
577 
578 	list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
579 		struct symbol *sym = (struct symbol *)(syme + 1);
580 
581 		if (!strcmp(buf, sym->name)) {
582 			found = syme;
583 			break;
584 		}
585 	}
586 
587 	if (!found) {
588 		fprintf(stderr, "Sorry, %s is not active.\n", sym_filter);
589 		sleep(1);
590 		return;
591 	} else
592 		parse_source(found);
593 
594 out_free:
595 	free(buf);
596 }
597 
598 static void print_mapped_keys(void)
599 {
600 	char *name = NULL;
601 
602 	if (sym_filter_entry) {
603 		struct symbol *sym = (struct symbol *)(sym_filter_entry+1);
604 		name = sym->name;
605 	}
606 
607 	fprintf(stdout, "\nMapped keys:\n");
608 	fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", delay_secs);
609 	fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", print_entries);
610 
611 	if (nr_counters > 1)
612 		fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(sym_counter));
613 
614 	fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", count_filter);
615 
616 	if (vmlinux) {
617 		fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
618 		fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
619 		fprintf(stdout, "\t[S]     stop annotation.\n");
620 	}
621 
622 	if (nr_counters > 1)
623 		fprintf(stdout, "\t[w]     toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
624 
625 	fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", zero ? 1 : 0);
626 	fprintf(stdout, "\t[qQ]    quit.\n");
627 }
628 
629 static int key_mapped(int c)
630 {
631 	switch (c) {
632 		case 'd':
633 		case 'e':
634 		case 'f':
635 		case 'z':
636 		case 'q':
637 		case 'Q':
638 			return 1;
639 		case 'E':
640 		case 'w':
641 			return nr_counters > 1 ? 1 : 0;
642 		case 'F':
643 		case 's':
644 		case 'S':
645 			return vmlinux ? 1 : 0;
646 	}
647 
648 	return 0;
649 }
650 
651 static void handle_keypress(int c)
652 {
653 	if (!key_mapped(c)) {
654 		struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
655 		struct termios tc, save;
656 
657 		print_mapped_keys();
658 		fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
659 		fflush(stdout);
660 
661 		tcgetattr(0, &save);
662 		tc = save;
663 		tc.c_lflag &= ~(ICANON | ECHO);
664 		tc.c_cc[VMIN] = 0;
665 		tc.c_cc[VTIME] = 0;
666 		tcsetattr(0, TCSANOW, &tc);
667 
668 		poll(&stdin_poll, 1, -1);
669 		c = getc(stdin);
670 
671 		tcsetattr(0, TCSAFLUSH, &save);
672 		if (!key_mapped(c))
673 			return;
674 	}
675 
676 	switch (c) {
677 		case 'd':
678 			prompt_integer(&delay_secs, "Enter display delay");
679 			break;
680 		case 'e':
681 			prompt_integer(&print_entries, "Enter display entries (lines)");
682 			break;
683 		case 'E':
684 			if (nr_counters > 1) {
685 				int i;
686 
687 				fprintf(stderr, "\nAvailable events:");
688 				for (i = 0; i < nr_counters; i++)
689 					fprintf(stderr, "\n\t%d %s", i, event_name(i));
690 
691 				prompt_integer(&sym_counter, "Enter details event counter");
692 
693 				if (sym_counter >= nr_counters) {
694 					fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
695 					sym_counter = 0;
696 					sleep(1);
697 				}
698 			} else sym_counter = 0;
699 			break;
700 		case 'f':
701 			prompt_integer(&count_filter, "Enter display event count filter");
702 			break;
703 		case 'F':
704 			prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
705 			break;
706 		case 'q':
707 		case 'Q':
708 			printf("exiting.\n");
709 			exit(0);
710 		case 's':
711 			prompt_symbol(&sym_filter_entry, "Enter details symbol");
712 			break;
713 		case 'S':
714 			if (!sym_filter_entry)
715 				break;
716 			else {
717 				struct sym_entry *syme = sym_filter_entry;
718 
719 				pthread_mutex_lock(&syme->source_lock);
720 				sym_filter_entry = NULL;
721 				__zero_source_counters(syme);
722 				pthread_mutex_unlock(&syme->source_lock);
723 			}
724 			break;
725 		case 'w':
726 			display_weighted = ~display_weighted;
727 			break;
728 		case 'z':
729 			zero = ~zero;
730 			break;
731 	}
732 }
733 
734 static void *display_thread(void *arg __used)
735 {
736 	struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
737 	struct termios tc, save;
738 	int delay_msecs, c;
739 
740 	tcgetattr(0, &save);
741 	tc = save;
742 	tc.c_lflag &= ~(ICANON | ECHO);
743 	tc.c_cc[VMIN] = 0;
744 	tc.c_cc[VTIME] = 0;
745 
746 repeat:
747 	delay_msecs = delay_secs * 1000;
748 	tcsetattr(0, TCSANOW, &tc);
749 	/* trash return*/
750 	getc(stdin);
751 
752 	do {
753 		print_sym_table();
754 	} while (!poll(&stdin_poll, 1, delay_msecs) == 1);
755 
756 	c = getc(stdin);
757 	tcsetattr(0, TCSAFLUSH, &save);
758 
759 	handle_keypress(c);
760 	goto repeat;
761 
762 	return NULL;
763 }
764 
765 /* Tag samples to be skipped. */
766 static const char *skip_symbols[] = {
767 	"default_idle",
768 	"cpu_idle",
769 	"enter_idle",
770 	"exit_idle",
771 	"mwait_idle",
772 	"mwait_idle_with_hints",
773 	"ppc64_runlatch_off",
774 	"pseries_dedicated_idle_sleep",
775 	NULL
776 };
777 
778 static int symbol_filter(struct dso *self, struct symbol *sym)
779 {
780 	struct sym_entry *syme;
781 	const char *name = sym->name;
782 	int i;
783 
784 	/*
785 	 * ppc64 uses function descriptors and appends a '.' to the
786 	 * start of every instruction address. Remove it.
787 	 */
788 	if (name[0] == '.')
789 		name++;
790 
791 	if (!strcmp(name, "_text") ||
792 	    !strcmp(name, "_etext") ||
793 	    !strcmp(name, "_sinittext") ||
794 	    !strncmp("init_module", name, 11) ||
795 	    !strncmp("cleanup_module", name, 14) ||
796 	    strstr(name, "_text_start") ||
797 	    strstr(name, "_text_end"))
798 		return 1;
799 
800 	syme = dso__sym_priv(self, sym);
801 	pthread_mutex_init(&syme->source_lock, NULL);
802 	if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter))
803 		sym_filter_entry = syme;
804 
805 	for (i = 0; skip_symbols[i]; i++) {
806 		if (!strcmp(skip_symbols[i], name)) {
807 			syme->skip = 1;
808 			break;
809 		}
810 	}
811 
812 	return 0;
813 }
814 
815 static int parse_symbols(void)
816 {
817 	struct rb_node *node;
818 	struct symbol  *sym;
819 	int modules = vmlinux ? 1 : 0;
820 
821 	kernel_dso = dso__new("[kernel]", sizeof(struct sym_entry));
822 	if (kernel_dso == NULL)
823 		return -1;
824 
825 	if (dso__load_kernel(kernel_dso, vmlinux, symbol_filter, verbose, modules) <= 0)
826 		goto out_delete_dso;
827 
828 	node = rb_first(&kernel_dso->syms);
829 	sym = rb_entry(node, struct symbol, rb_node);
830 	min_ip = sym->start;
831 
832 	node = rb_last(&kernel_dso->syms);
833 	sym = rb_entry(node, struct symbol, rb_node);
834 	max_ip = sym->end;
835 
836 	if (dump_symtab)
837 		dso__fprintf(kernel_dso, stderr);
838 
839 	return 0;
840 
841 out_delete_dso:
842 	dso__delete(kernel_dso);
843 	kernel_dso = NULL;
844 	return -1;
845 }
846 
847 /*
848  * Binary search in the histogram table and record the hit:
849  */
850 static void record_ip(u64 ip, int counter)
851 {
852 	struct symbol *sym = dso__find_symbol(kernel_dso, ip);
853 
854 	if (sym != NULL) {
855 		struct sym_entry *syme = dso__sym_priv(kernel_dso, sym);
856 
857 		if (!syme->skip) {
858 			syme->count[counter]++;
859 			record_precise_ip(syme, counter, ip);
860 			pthread_mutex_lock(&active_symbols_lock);
861 			if (list_empty(&syme->node) || !syme->node.next)
862 				__list_insert_active_sym(syme);
863 			pthread_mutex_unlock(&active_symbols_lock);
864 			return;
865 		}
866 	}
867 
868 	samples--;
869 }
870 
871 static void process_event(u64 ip, int counter, int user)
872 {
873 	samples++;
874 
875 	if (user) {
876 		userspace_samples++;
877 		return;
878 	}
879 
880 	record_ip(ip, counter);
881 }
882 
883 struct mmap_data {
884 	int			counter;
885 	void			*base;
886 	int			mask;
887 	unsigned int		prev;
888 };
889 
890 static unsigned int mmap_read_head(struct mmap_data *md)
891 {
892 	struct perf_counter_mmap_page *pc = md->base;
893 	int head;
894 
895 	head = pc->data_head;
896 	rmb();
897 
898 	return head;
899 }
900 
901 struct timeval last_read, this_read;
902 
903 static void mmap_read_counter(struct mmap_data *md)
904 {
905 	unsigned int head = mmap_read_head(md);
906 	unsigned int old = md->prev;
907 	unsigned char *data = md->base + page_size;
908 	int diff;
909 
910 	gettimeofday(&this_read, NULL);
911 
912 	/*
913 	 * If we're further behind than half the buffer, there's a chance
914 	 * the writer will bite our tail and mess up the samples under us.
915 	 *
916 	 * If we somehow ended up ahead of the head, we got messed up.
917 	 *
918 	 * In either case, truncate and restart at head.
919 	 */
920 	diff = head - old;
921 	if (diff > md->mask / 2 || diff < 0) {
922 		struct timeval iv;
923 		unsigned long msecs;
924 
925 		timersub(&this_read, &last_read, &iv);
926 		msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
927 
928 		fprintf(stderr, "WARNING: failed to keep up with mmap data."
929 				"  Last read %lu msecs ago.\n", msecs);
930 
931 		/*
932 		 * head points to a known good entry, start there.
933 		 */
934 		old = head;
935 	}
936 
937 	last_read = this_read;
938 
939 	for (; old != head;) {
940 		struct ip_event {
941 			struct perf_event_header header;
942 			u64 ip;
943 			u32 pid, target_pid;
944 		};
945 		struct mmap_event {
946 			struct perf_event_header header;
947 			u32 pid, target_pid;
948 			u64 start;
949 			u64 len;
950 			u64 pgoff;
951 			char filename[PATH_MAX];
952 		};
953 
954 		typedef union event_union {
955 			struct perf_event_header header;
956 			struct ip_event ip;
957 			struct mmap_event mmap;
958 		} event_t;
959 
960 		event_t *event = (event_t *)&data[old & md->mask];
961 
962 		event_t event_copy;
963 
964 		size_t size = event->header.size;
965 
966 		/*
967 		 * Event straddles the mmap boundary -- header should always
968 		 * be inside due to u64 alignment of output.
969 		 */
970 		if ((old & md->mask) + size != ((old + size) & md->mask)) {
971 			unsigned int offset = old;
972 			unsigned int len = min(sizeof(*event), size), cpy;
973 			void *dst = &event_copy;
974 
975 			do {
976 				cpy = min(md->mask + 1 - (offset & md->mask), len);
977 				memcpy(dst, &data[offset & md->mask], cpy);
978 				offset += cpy;
979 				dst += cpy;
980 				len -= cpy;
981 			} while (len);
982 
983 			event = &event_copy;
984 		}
985 
986 		old += size;
987 
988 		if (event->header.type == PERF_EVENT_SAMPLE) {
989 			int user =
990 	(event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK) == PERF_EVENT_MISC_USER;
991 			process_event(event->ip.ip, md->counter, user);
992 		}
993 	}
994 
995 	md->prev = old;
996 }
997 
998 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
999 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
1000 
1001 static void mmap_read(void)
1002 {
1003 	int i, counter;
1004 
1005 	for (i = 0; i < nr_cpus; i++) {
1006 		for (counter = 0; counter < nr_counters; counter++)
1007 			mmap_read_counter(&mmap_array[i][counter]);
1008 	}
1009 }
1010 
1011 int nr_poll;
1012 int group_fd;
1013 
1014 static void start_counter(int i, int counter)
1015 {
1016 	struct perf_counter_attr *attr;
1017 	int cpu;
1018 
1019 	cpu = profile_cpu;
1020 	if (target_pid == -1 && profile_cpu == -1)
1021 		cpu = i;
1022 
1023 	attr = attrs + counter;
1024 
1025 	attr->sample_type	= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1026 	attr->freq		= freq;
1027 	attr->inherit		= (cpu < 0) && inherit;
1028 
1029 try_again:
1030 	fd[i][counter] = sys_perf_counter_open(attr, target_pid, cpu, group_fd, 0);
1031 
1032 	if (fd[i][counter] < 0) {
1033 		int err = errno;
1034 
1035 		if (err == EPERM)
1036 			die("No permission - are you root?\n");
1037 		/*
1038 		 * If it's cycles then fall back to hrtimer
1039 		 * based cpu-clock-tick sw counter, which
1040 		 * is always available even if no PMU support:
1041 		 */
1042 		if (attr->type == PERF_TYPE_HARDWARE
1043 			&& attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1044 
1045 			if (verbose)
1046 				warning(" ... trying to fall back to cpu-clock-ticks\n");
1047 
1048 			attr->type = PERF_TYPE_SOFTWARE;
1049 			attr->config = PERF_COUNT_SW_CPU_CLOCK;
1050 			goto try_again;
1051 		}
1052 		printf("\n");
1053 		error("perfcounter syscall returned with %d (%s)\n",
1054 			fd[i][counter], strerror(err));
1055 		die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
1056 		exit(-1);
1057 	}
1058 	assert(fd[i][counter] >= 0);
1059 	fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
1060 
1061 	/*
1062 	 * First counter acts as the group leader:
1063 	 */
1064 	if (group && group_fd == -1)
1065 		group_fd = fd[i][counter];
1066 
1067 	event_array[nr_poll].fd = fd[i][counter];
1068 	event_array[nr_poll].events = POLLIN;
1069 	nr_poll++;
1070 
1071 	mmap_array[i][counter].counter = counter;
1072 	mmap_array[i][counter].prev = 0;
1073 	mmap_array[i][counter].mask = mmap_pages*page_size - 1;
1074 	mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
1075 			PROT_READ, MAP_SHARED, fd[i][counter], 0);
1076 	if (mmap_array[i][counter].base == MAP_FAILED)
1077 		die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1078 }
1079 
1080 static int __cmd_top(void)
1081 {
1082 	pthread_t thread;
1083 	int i, counter;
1084 	int ret;
1085 
1086 	for (i = 0; i < nr_cpus; i++) {
1087 		group_fd = -1;
1088 		for (counter = 0; counter < nr_counters; counter++)
1089 			start_counter(i, counter);
1090 	}
1091 
1092 	/* Wait for a minimal set of events before starting the snapshot */
1093 	poll(event_array, nr_poll, 100);
1094 
1095 	mmap_read();
1096 
1097 	if (pthread_create(&thread, NULL, display_thread, NULL)) {
1098 		printf("Could not create display thread.\n");
1099 		exit(-1);
1100 	}
1101 
1102 	if (realtime_prio) {
1103 		struct sched_param param;
1104 
1105 		param.sched_priority = realtime_prio;
1106 		if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1107 			printf("Could not set realtime priority.\n");
1108 			exit(-1);
1109 		}
1110 	}
1111 
1112 	while (1) {
1113 		int hits = samples;
1114 
1115 		mmap_read();
1116 
1117 		if (hits == samples)
1118 			ret = poll(event_array, nr_poll, 100);
1119 	}
1120 
1121 	return 0;
1122 }
1123 
1124 static const char * const top_usage[] = {
1125 	"perf top [<options>]",
1126 	NULL
1127 };
1128 
1129 static const struct option options[] = {
1130 	OPT_CALLBACK('e', "event", NULL, "event",
1131 		     "event selector. use 'perf list' to list available events",
1132 		     parse_events),
1133 	OPT_INTEGER('c', "count", &default_interval,
1134 		    "event period to sample"),
1135 	OPT_INTEGER('p', "pid", &target_pid,
1136 		    "profile events on existing pid"),
1137 	OPT_BOOLEAN('a', "all-cpus", &system_wide,
1138 			    "system-wide collection from all CPUs"),
1139 	OPT_INTEGER('C', "CPU", &profile_cpu,
1140 		    "CPU to profile on"),
1141 	OPT_STRING('k', "vmlinux", &vmlinux, "file", "vmlinux pathname"),
1142 	OPT_INTEGER('m', "mmap-pages", &mmap_pages,
1143 		    "number of mmap data pages"),
1144 	OPT_INTEGER('r', "realtime", &realtime_prio,
1145 		    "collect data with this RT SCHED_FIFO priority"),
1146 	OPT_INTEGER('d', "delay", &delay_secs,
1147 		    "number of seconds to delay between refreshes"),
1148 	OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1149 			    "dump the symbol table used for profiling"),
1150 	OPT_INTEGER('f', "count-filter", &count_filter,
1151 		    "only display functions with more events than this"),
1152 	OPT_BOOLEAN('g', "group", &group,
1153 			    "put the counters into a counter group"),
1154 	OPT_BOOLEAN('i', "inherit", &inherit,
1155 		    "child tasks inherit counters"),
1156 	OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1157 		    "symbol to annotate - requires -k option"),
1158 	OPT_BOOLEAN('z', "zero", &zero,
1159 		    "zero history across updates"),
1160 	OPT_INTEGER('F', "freq", &freq,
1161 		    "profile at this frequency"),
1162 	OPT_INTEGER('E', "entries", &print_entries,
1163 		    "display this many functions"),
1164 	OPT_BOOLEAN('v', "verbose", &verbose,
1165 		    "be more verbose (show counter open errors, etc)"),
1166 	OPT_END()
1167 };
1168 
1169 int cmd_top(int argc, const char **argv, const char *prefix __used)
1170 {
1171 	int counter;
1172 
1173 	symbol__init();
1174 
1175 	page_size = sysconf(_SC_PAGE_SIZE);
1176 
1177 	argc = parse_options(argc, argv, options, top_usage, 0);
1178 	if (argc)
1179 		usage_with_options(top_usage, options);
1180 
1181 	if (freq) {
1182 		default_interval = freq;
1183 		freq = 1;
1184 	}
1185 
1186 	/* CPU and PID are mutually exclusive */
1187 	if (target_pid != -1 && profile_cpu != -1) {
1188 		printf("WARNING: PID switch overriding CPU\n");
1189 		sleep(1);
1190 		profile_cpu = -1;
1191 	}
1192 
1193 	if (!nr_counters)
1194 		nr_counters = 1;
1195 
1196 	if (delay_secs < 1)
1197 		delay_secs = 1;
1198 
1199 	parse_symbols();
1200 	parse_source(sym_filter_entry);
1201 
1202 	/*
1203 	 * Fill in the ones not specifically initialized via -c:
1204 	 */
1205 	for (counter = 0; counter < nr_counters; counter++) {
1206 		if (attrs[counter].sample_period)
1207 			continue;
1208 
1209 		attrs[counter].sample_period = default_interval;
1210 	}
1211 
1212 	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1213 	assert(nr_cpus <= MAX_NR_CPUS);
1214 	assert(nr_cpus >= 0);
1215 
1216 	if (target_pid != -1 || profile_cpu != -1)
1217 		nr_cpus = 1;
1218 
1219 	return __cmd_top();
1220 }
1221