xref: /linux/tools/perf/builtin-top.c (revision ec2212088c42ff7d1362629ec26dda4f3e8bdad3)
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  *		 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
9  *
10  * Improvements and fixes by:
11  *
12  *   Arjan van de Ven <arjan@linux.intel.com>
13  *   Yanmin Zhang <yanmin.zhang@intel.com>
14  *   Wu Fengguang <fengguang.wu@intel.com>
15  *   Mike Galbraith <efault@gmx.de>
16  *   Paul Mackerras <paulus@samba.org>
17  *
18  * Released under the GPL v2. (and only v2, not any later version)
19  */
20 #include "builtin.h"
21 
22 #include "perf.h"
23 
24 #include "util/annotate.h"
25 #include "util/cache.h"
26 #include "util/color.h"
27 #include "util/evlist.h"
28 #include "util/evsel.h"
29 #include "util/session.h"
30 #include "util/symbol.h"
31 #include "util/thread.h"
32 #include "util/thread_map.h"
33 #include "util/top.h"
34 #include "util/util.h"
35 #include <linux/rbtree.h>
36 #include "util/parse-options.h"
37 #include "util/parse-events.h"
38 #include "util/cpumap.h"
39 #include "util/xyarray.h"
40 #include "util/sort.h"
41 
42 #include "util/debug.h"
43 
44 #include <assert.h>
45 #include <fcntl.h>
46 
47 #include <stdio.h>
48 #include <termios.h>
49 #include <unistd.h>
50 #include <inttypes.h>
51 
52 #include <errno.h>
53 #include <time.h>
54 #include <sched.h>
55 
56 #include <sys/syscall.h>
57 #include <sys/ioctl.h>
58 #include <sys/poll.h>
59 #include <sys/prctl.h>
60 #include <sys/wait.h>
61 #include <sys/uio.h>
62 #include <sys/mman.h>
63 
64 #include <linux/unistd.h>
65 #include <linux/types.h>
66 
67 void get_term_dimensions(struct winsize *ws)
68 {
69 	char *s = getenv("LINES");
70 
71 	if (s != NULL) {
72 		ws->ws_row = atoi(s);
73 		s = getenv("COLUMNS");
74 		if (s != NULL) {
75 			ws->ws_col = atoi(s);
76 			if (ws->ws_row && ws->ws_col)
77 				return;
78 		}
79 	}
80 #ifdef TIOCGWINSZ
81 	if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
82 	    ws->ws_row && ws->ws_col)
83 		return;
84 #endif
85 	ws->ws_row = 25;
86 	ws->ws_col = 80;
87 }
88 
89 static void perf_top__update_print_entries(struct perf_top *top)
90 {
91 	if (top->print_entries > 9)
92 		top->print_entries -= 9;
93 }
94 
95 static void perf_top__sig_winch(int sig __used, siginfo_t *info __used, void *arg)
96 {
97 	struct perf_top *top = arg;
98 
99 	get_term_dimensions(&top->winsize);
100 	if (!top->print_entries
101 	    || (top->print_entries+4) > top->winsize.ws_row) {
102 		top->print_entries = top->winsize.ws_row;
103 	} else {
104 		top->print_entries += 4;
105 		top->winsize.ws_row = top->print_entries;
106 	}
107 	perf_top__update_print_entries(top);
108 }
109 
110 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
111 {
112 	struct symbol *sym;
113 	struct annotation *notes;
114 	struct map *map;
115 	int err = -1;
116 
117 	if (!he || !he->ms.sym)
118 		return -1;
119 
120 	sym = he->ms.sym;
121 	map = he->ms.map;
122 
123 	/*
124 	 * We can't annotate with just /proc/kallsyms
125 	 */
126 	if (map->dso->symtab_type == SYMTAB__KALLSYMS) {
127 		pr_err("Can't annotate %s: No vmlinux file was found in the "
128 		       "path\n", sym->name);
129 		sleep(1);
130 		return -1;
131 	}
132 
133 	notes = symbol__annotation(sym);
134 	if (notes->src != NULL) {
135 		pthread_mutex_lock(&notes->lock);
136 		goto out_assign;
137 	}
138 
139 	pthread_mutex_lock(&notes->lock);
140 
141 	if (symbol__alloc_hist(sym) < 0) {
142 		pthread_mutex_unlock(&notes->lock);
143 		pr_err("Not enough memory for annotating '%s' symbol!\n",
144 		       sym->name);
145 		sleep(1);
146 		return err;
147 	}
148 
149 	err = symbol__annotate(sym, map, 0);
150 	if (err == 0) {
151 out_assign:
152 		top->sym_filter_entry = he;
153 	}
154 
155 	pthread_mutex_unlock(&notes->lock);
156 	return err;
157 }
158 
159 static void __zero_source_counters(struct hist_entry *he)
160 {
161 	struct symbol *sym = he->ms.sym;
162 	symbol__annotate_zero_histograms(sym);
163 }
164 
165 static void perf_top__record_precise_ip(struct perf_top *top,
166 					struct hist_entry *he,
167 					int counter, u64 ip)
168 {
169 	struct annotation *notes;
170 	struct symbol *sym;
171 
172 	if (he == NULL || he->ms.sym == NULL ||
173 	    ((top->sym_filter_entry == NULL ||
174 	      top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
175 		return;
176 
177 	sym = he->ms.sym;
178 	notes = symbol__annotation(sym);
179 
180 	if (pthread_mutex_trylock(&notes->lock))
181 		return;
182 
183 	if (notes->src == NULL && symbol__alloc_hist(sym) < 0) {
184 		pthread_mutex_unlock(&notes->lock);
185 		pr_err("Not enough memory for annotating '%s' symbol!\n",
186 		       sym->name);
187 		sleep(1);
188 		return;
189 	}
190 
191 	ip = he->ms.map->map_ip(he->ms.map, ip);
192 	symbol__inc_addr_samples(sym, he->ms.map, counter, ip);
193 
194 	pthread_mutex_unlock(&notes->lock);
195 }
196 
197 static void perf_top__show_details(struct perf_top *top)
198 {
199 	struct hist_entry *he = top->sym_filter_entry;
200 	struct annotation *notes;
201 	struct symbol *symbol;
202 	int more;
203 
204 	if (!he)
205 		return;
206 
207 	symbol = he->ms.sym;
208 	notes = symbol__annotation(symbol);
209 
210 	pthread_mutex_lock(&notes->lock);
211 
212 	if (notes->src == NULL)
213 		goto out_unlock;
214 
215 	printf("Showing %s for %s\n", event_name(top->sym_evsel), symbol->name);
216 	printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
217 
218 	more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel->idx,
219 				       0, top->sym_pcnt_filter, top->print_entries, 4);
220 	if (top->zero)
221 		symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
222 	else
223 		symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
224 	if (more != 0)
225 		printf("%d lines not displayed, maybe increase display entries [e]\n", more);
226 out_unlock:
227 	pthread_mutex_unlock(&notes->lock);
228 }
229 
230 static const char		CONSOLE_CLEAR[] = "";
231 
232 static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
233 						     struct addr_location *al,
234 						     struct perf_sample *sample)
235 {
236 	struct hist_entry *he;
237 
238 	he = __hists__add_entry(&evsel->hists, al, NULL, sample->period);
239 	if (he == NULL)
240 		return NULL;
241 
242 	hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
243 	return he;
244 }
245 
246 static void perf_top__print_sym_table(struct perf_top *top)
247 {
248 	char bf[160];
249 	int printed = 0;
250 	const int win_width = top->winsize.ws_col - 1;
251 
252 	puts(CONSOLE_CLEAR);
253 
254 	perf_top__header_snprintf(top, bf, sizeof(bf));
255 	printf("%s\n", bf);
256 
257 	perf_top__reset_sample_counters(top);
258 
259 	printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
260 
261 	if (top->sym_evsel->hists.stats.nr_lost_warned !=
262 	    top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
263 		top->sym_evsel->hists.stats.nr_lost_warned =
264 			top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
265 		color_fprintf(stdout, PERF_COLOR_RED,
266 			      "WARNING: LOST %d chunks, Check IO/CPU overload",
267 			      top->sym_evsel->hists.stats.nr_lost_warned);
268 		++printed;
269 	}
270 
271 	if (top->sym_filter_entry) {
272 		perf_top__show_details(top);
273 		return;
274 	}
275 
276 	hists__collapse_resort_threaded(&top->sym_evsel->hists);
277 	hists__output_resort_threaded(&top->sym_evsel->hists);
278 	hists__decay_entries_threaded(&top->sym_evsel->hists,
279 				      top->hide_user_symbols,
280 				      top->hide_kernel_symbols);
281 	hists__output_recalc_col_len(&top->sym_evsel->hists,
282 				     top->winsize.ws_row - 3);
283 	putchar('\n');
284 	hists__fprintf(&top->sym_evsel->hists, NULL, false, false,
285 		       top->winsize.ws_row - 4 - printed, win_width, stdout);
286 }
287 
288 static void prompt_integer(int *target, const char *msg)
289 {
290 	char *buf = malloc(0), *p;
291 	size_t dummy = 0;
292 	int tmp;
293 
294 	fprintf(stdout, "\n%s: ", msg);
295 	if (getline(&buf, &dummy, stdin) < 0)
296 		return;
297 
298 	p = strchr(buf, '\n');
299 	if (p)
300 		*p = 0;
301 
302 	p = buf;
303 	while(*p) {
304 		if (!isdigit(*p))
305 			goto out_free;
306 		p++;
307 	}
308 	tmp = strtoul(buf, NULL, 10);
309 	*target = tmp;
310 out_free:
311 	free(buf);
312 }
313 
314 static void prompt_percent(int *target, const char *msg)
315 {
316 	int tmp = 0;
317 
318 	prompt_integer(&tmp, msg);
319 	if (tmp >= 0 && tmp <= 100)
320 		*target = tmp;
321 }
322 
323 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
324 {
325 	char *buf = malloc(0), *p;
326 	struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
327 	struct rb_node *next;
328 	size_t dummy = 0;
329 
330 	/* zero counters of active symbol */
331 	if (syme) {
332 		__zero_source_counters(syme);
333 		top->sym_filter_entry = NULL;
334 	}
335 
336 	fprintf(stdout, "\n%s: ", msg);
337 	if (getline(&buf, &dummy, stdin) < 0)
338 		goto out_free;
339 
340 	p = strchr(buf, '\n');
341 	if (p)
342 		*p = 0;
343 
344 	next = rb_first(&top->sym_evsel->hists.entries);
345 	while (next) {
346 		n = rb_entry(next, struct hist_entry, rb_node);
347 		if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
348 			found = n;
349 			break;
350 		}
351 		next = rb_next(&n->rb_node);
352 	}
353 
354 	if (!found) {
355 		fprintf(stderr, "Sorry, %s is not active.\n", buf);
356 		sleep(1);
357 	} else
358 		perf_top__parse_source(top, found);
359 
360 out_free:
361 	free(buf);
362 }
363 
364 static void perf_top__print_mapped_keys(struct perf_top *top)
365 {
366 	char *name = NULL;
367 
368 	if (top->sym_filter_entry) {
369 		struct symbol *sym = top->sym_filter_entry->ms.sym;
370 		name = sym->name;
371 	}
372 
373 	fprintf(stdout, "\nMapped keys:\n");
374 	fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
375 	fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
376 
377 	if (top->evlist->nr_entries > 1)
378 		fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(top->sym_evsel));
379 
380 	fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
381 
382 	fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
383 	fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
384 	fprintf(stdout, "\t[S]     stop annotation.\n");
385 
386 	fprintf(stdout,
387 		"\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
388 		top->hide_kernel_symbols ? "yes" : "no");
389 	fprintf(stdout,
390 		"\t[U]     hide user symbols.               \t(%s)\n",
391 		top->hide_user_symbols ? "yes" : "no");
392 	fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
393 	fprintf(stdout, "\t[qQ]    quit.\n");
394 }
395 
396 static int perf_top__key_mapped(struct perf_top *top, int c)
397 {
398 	switch (c) {
399 		case 'd':
400 		case 'e':
401 		case 'f':
402 		case 'z':
403 		case 'q':
404 		case 'Q':
405 		case 'K':
406 		case 'U':
407 		case 'F':
408 		case 's':
409 		case 'S':
410 			return 1;
411 		case 'E':
412 			return top->evlist->nr_entries > 1 ? 1 : 0;
413 		default:
414 			break;
415 	}
416 
417 	return 0;
418 }
419 
420 static void perf_top__handle_keypress(struct perf_top *top, int c)
421 {
422 	if (!perf_top__key_mapped(top, c)) {
423 		struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
424 		struct termios tc, save;
425 
426 		perf_top__print_mapped_keys(top);
427 		fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
428 		fflush(stdout);
429 
430 		tcgetattr(0, &save);
431 		tc = save;
432 		tc.c_lflag &= ~(ICANON | ECHO);
433 		tc.c_cc[VMIN] = 0;
434 		tc.c_cc[VTIME] = 0;
435 		tcsetattr(0, TCSANOW, &tc);
436 
437 		poll(&stdin_poll, 1, -1);
438 		c = getc(stdin);
439 
440 		tcsetattr(0, TCSAFLUSH, &save);
441 		if (!perf_top__key_mapped(top, c))
442 			return;
443 	}
444 
445 	switch (c) {
446 		case 'd':
447 			prompt_integer(&top->delay_secs, "Enter display delay");
448 			if (top->delay_secs < 1)
449 				top->delay_secs = 1;
450 			break;
451 		case 'e':
452 			prompt_integer(&top->print_entries, "Enter display entries (lines)");
453 			if (top->print_entries == 0) {
454 				struct sigaction act = {
455 					.sa_sigaction = perf_top__sig_winch,
456 					.sa_flags     = SA_SIGINFO,
457 				};
458 				perf_top__sig_winch(SIGWINCH, NULL, top);
459 				sigaction(SIGWINCH, &act, NULL);
460 			} else {
461 				perf_top__sig_winch(SIGWINCH, NULL, top);
462 				signal(SIGWINCH, SIG_DFL);
463 			}
464 			break;
465 		case 'E':
466 			if (top->evlist->nr_entries > 1) {
467 				/* Select 0 as the default event: */
468 				int counter = 0;
469 
470 				fprintf(stderr, "\nAvailable events:");
471 
472 				list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
473 					fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, event_name(top->sym_evsel));
474 
475 				prompt_integer(&counter, "Enter details event counter");
476 
477 				if (counter >= top->evlist->nr_entries) {
478 					top->sym_evsel = list_entry(top->evlist->entries.next, struct perf_evsel, node);
479 					fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(top->sym_evsel));
480 					sleep(1);
481 					break;
482 				}
483 				list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
484 					if (top->sym_evsel->idx == counter)
485 						break;
486 			} else
487 				top->sym_evsel = list_entry(top->evlist->entries.next, struct perf_evsel, node);
488 			break;
489 		case 'f':
490 			prompt_integer(&top->count_filter, "Enter display event count filter");
491 			break;
492 		case 'F':
493 			prompt_percent(&top->sym_pcnt_filter,
494 				       "Enter details display event filter (percent)");
495 			break;
496 		case 'K':
497 			top->hide_kernel_symbols = !top->hide_kernel_symbols;
498 			break;
499 		case 'q':
500 		case 'Q':
501 			printf("exiting.\n");
502 			if (top->dump_symtab)
503 				perf_session__fprintf_dsos(top->session, stderr);
504 			exit(0);
505 		case 's':
506 			perf_top__prompt_symbol(top, "Enter details symbol");
507 			break;
508 		case 'S':
509 			if (!top->sym_filter_entry)
510 				break;
511 			else {
512 				struct hist_entry *syme = top->sym_filter_entry;
513 
514 				top->sym_filter_entry = NULL;
515 				__zero_source_counters(syme);
516 			}
517 			break;
518 		case 'U':
519 			top->hide_user_symbols = !top->hide_user_symbols;
520 			break;
521 		case 'z':
522 			top->zero = !top->zero;
523 			break;
524 		default:
525 			break;
526 	}
527 }
528 
529 static void perf_top__sort_new_samples(void *arg)
530 {
531 	struct perf_top *t = arg;
532 	perf_top__reset_sample_counters(t);
533 
534 	if (t->evlist->selected != NULL)
535 		t->sym_evsel = t->evlist->selected;
536 
537 	hists__collapse_resort_threaded(&t->sym_evsel->hists);
538 	hists__output_resort_threaded(&t->sym_evsel->hists);
539 	hists__decay_entries_threaded(&t->sym_evsel->hists,
540 				      t->hide_user_symbols,
541 				      t->hide_kernel_symbols);
542 }
543 
544 static void *display_thread_tui(void *arg)
545 {
546 	struct perf_evsel *pos;
547 	struct perf_top *top = arg;
548 	const char *help = "For a higher level overview, try: perf top --sort comm,dso";
549 
550 	perf_top__sort_new_samples(top);
551 
552 	/*
553 	 * Initialize the uid_filter_str, in the future the TUI will allow
554 	 * Zooming in/out UIDs. For now juse use whatever the user passed
555 	 * via --uid.
556 	 */
557 	list_for_each_entry(pos, &top->evlist->entries, node)
558 		pos->hists.uid_filter_str = top->uid_str;
559 
560 	perf_evlist__tui_browse_hists(top->evlist, help,
561 				      perf_top__sort_new_samples,
562 				      top, top->delay_secs);
563 
564 	exit_browser(0);
565 	exit(0);
566 	return NULL;
567 }
568 
569 static void *display_thread(void *arg)
570 {
571 	struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
572 	struct termios tc, save;
573 	struct perf_top *top = arg;
574 	int delay_msecs, c;
575 
576 	tcgetattr(0, &save);
577 	tc = save;
578 	tc.c_lflag &= ~(ICANON | ECHO);
579 	tc.c_cc[VMIN] = 0;
580 	tc.c_cc[VTIME] = 0;
581 
582 	pthread__unblock_sigwinch();
583 repeat:
584 	delay_msecs = top->delay_secs * 1000;
585 	tcsetattr(0, TCSANOW, &tc);
586 	/* trash return*/
587 	getc(stdin);
588 
589 	while (1) {
590 		perf_top__print_sym_table(top);
591 		/*
592 		 * Either timeout expired or we got an EINTR due to SIGWINCH,
593 		 * refresh screen in both cases.
594 		 */
595 		switch (poll(&stdin_poll, 1, delay_msecs)) {
596 		case 0:
597 			continue;
598 		case -1:
599 			if (errno == EINTR)
600 				continue;
601 			/* Fall trhu */
602 		default:
603 			goto process_hotkey;
604 		}
605 	}
606 process_hotkey:
607 	c = getc(stdin);
608 	tcsetattr(0, TCSAFLUSH, &save);
609 
610 	perf_top__handle_keypress(top, c);
611 	goto repeat;
612 
613 	return NULL;
614 }
615 
616 /* Tag samples to be skipped. */
617 static const char *skip_symbols[] = {
618 	"default_idle",
619 	"native_safe_halt",
620 	"cpu_idle",
621 	"enter_idle",
622 	"exit_idle",
623 	"mwait_idle",
624 	"mwait_idle_with_hints",
625 	"poll_idle",
626 	"ppc64_runlatch_off",
627 	"pseries_dedicated_idle_sleep",
628 	NULL
629 };
630 
631 static int symbol_filter(struct map *map __used, struct symbol *sym)
632 {
633 	const char *name = sym->name;
634 	int i;
635 
636 	/*
637 	 * ppc64 uses function descriptors and appends a '.' to the
638 	 * start of every instruction address. Remove it.
639 	 */
640 	if (name[0] == '.')
641 		name++;
642 
643 	if (!strcmp(name, "_text") ||
644 	    !strcmp(name, "_etext") ||
645 	    !strcmp(name, "_sinittext") ||
646 	    !strncmp("init_module", name, 11) ||
647 	    !strncmp("cleanup_module", name, 14) ||
648 	    strstr(name, "_text_start") ||
649 	    strstr(name, "_text_end"))
650 		return 1;
651 
652 	for (i = 0; skip_symbols[i]; i++) {
653 		if (!strcmp(skip_symbols[i], name)) {
654 			sym->ignore = true;
655 			break;
656 		}
657 	}
658 
659 	return 0;
660 }
661 
662 static void perf_event__process_sample(struct perf_tool *tool,
663 				       const union perf_event *event,
664 				       struct perf_evsel *evsel,
665 				       struct perf_sample *sample,
666 				       struct machine *machine)
667 {
668 	struct perf_top *top = container_of(tool, struct perf_top, tool);
669 	struct symbol *parent = NULL;
670 	u64 ip = event->ip.ip;
671 	struct addr_location al;
672 	int err;
673 
674 	if (!machine && perf_guest) {
675 		pr_err("Can't find guest [%d]'s kernel information\n",
676 			event->ip.pid);
677 		return;
678 	}
679 
680 	if (!machine) {
681 		pr_err("%u unprocessable samples recorded.",
682 		       top->session->hists.stats.nr_unprocessable_samples++);
683 		return;
684 	}
685 
686 	if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
687 		top->exact_samples++;
688 
689 	if (perf_event__preprocess_sample(event, machine, &al, sample,
690 					  symbol_filter) < 0 ||
691 	    al.filtered)
692 		return;
693 
694 	if (!top->kptr_restrict_warned &&
695 	    symbol_conf.kptr_restrict &&
696 	    al.cpumode == PERF_RECORD_MISC_KERNEL) {
697 		ui__warning(
698 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
699 "Check /proc/sys/kernel/kptr_restrict.\n\n"
700 "Kernel%s samples will not be resolved.\n",
701 			  !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
702 			  " modules" : "");
703 		if (use_browser <= 0)
704 			sleep(5);
705 		top->kptr_restrict_warned = true;
706 	}
707 
708 	if (al.sym == NULL) {
709 		const char *msg = "Kernel samples will not be resolved.\n";
710 		/*
711 		 * As we do lazy loading of symtabs we only will know if the
712 		 * specified vmlinux file is invalid when we actually have a
713 		 * hit in kernel space and then try to load it. So if we get
714 		 * here and there are _no_ symbols in the DSO backing the
715 		 * kernel map, bail out.
716 		 *
717 		 * We may never get here, for instance, if we use -K/
718 		 * --hide-kernel-symbols, even if the user specifies an
719 		 * invalid --vmlinux ;-)
720 		 */
721 		if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
722 		    al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
723 		    RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
724 			if (symbol_conf.vmlinux_name) {
725 				ui__warning("The %s file can't be used.\n%s",
726 					    symbol_conf.vmlinux_name, msg);
727 			} else {
728 				ui__warning("A vmlinux file was not found.\n%s",
729 					    msg);
730 			}
731 
732 			if (use_browser <= 0)
733 				sleep(5);
734 			top->vmlinux_warned = true;
735 		}
736 	}
737 
738 	if (al.sym == NULL || !al.sym->ignore) {
739 		struct hist_entry *he;
740 
741 		if ((sort__has_parent || symbol_conf.use_callchain) &&
742 		    sample->callchain) {
743 			err = machine__resolve_callchain(machine, evsel, al.thread,
744 							 sample->callchain, &parent);
745 			if (err)
746 				return;
747 		}
748 
749 		he = perf_evsel__add_hist_entry(evsel, &al, sample);
750 		if (he == NULL) {
751 			pr_err("Problem incrementing symbol period, skipping event\n");
752 			return;
753 		}
754 
755 		if (symbol_conf.use_callchain) {
756 			err = callchain_append(he->callchain, &evsel->hists.callchain_cursor,
757 					       sample->period);
758 			if (err)
759 				return;
760 		}
761 
762 		if (top->sort_has_symbols)
763 			perf_top__record_precise_ip(top, he, evsel->idx, ip);
764 	}
765 
766 	return;
767 }
768 
769 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
770 {
771 	struct perf_sample sample;
772 	struct perf_evsel *evsel;
773 	struct perf_session *session = top->session;
774 	union perf_event *event;
775 	struct machine *machine;
776 	u8 origin;
777 	int ret;
778 
779 	while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
780 		ret = perf_session__parse_sample(session, event, &sample);
781 		if (ret) {
782 			pr_err("Can't parse sample, err = %d\n", ret);
783 			continue;
784 		}
785 
786 		evsel = perf_evlist__id2evsel(session->evlist, sample.id);
787 		assert(evsel != NULL);
788 
789 		origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
790 
791 		if (event->header.type == PERF_RECORD_SAMPLE)
792 			++top->samples;
793 
794 		switch (origin) {
795 		case PERF_RECORD_MISC_USER:
796 			++top->us_samples;
797 			if (top->hide_user_symbols)
798 				continue;
799 			machine = perf_session__find_host_machine(session);
800 			break;
801 		case PERF_RECORD_MISC_KERNEL:
802 			++top->kernel_samples;
803 			if (top->hide_kernel_symbols)
804 				continue;
805 			machine = perf_session__find_host_machine(session);
806 			break;
807 		case PERF_RECORD_MISC_GUEST_KERNEL:
808 			++top->guest_kernel_samples;
809 			machine = perf_session__find_machine(session, event->ip.pid);
810 			break;
811 		case PERF_RECORD_MISC_GUEST_USER:
812 			++top->guest_us_samples;
813 			/*
814 			 * TODO: we don't process guest user from host side
815 			 * except simple counting.
816 			 */
817 			/* Fall thru */
818 		default:
819 			continue;
820 		}
821 
822 
823 		if (event->header.type == PERF_RECORD_SAMPLE) {
824 			perf_event__process_sample(&top->tool, event, evsel,
825 						   &sample, machine);
826 		} else if (event->header.type < PERF_RECORD_MAX) {
827 			hists__inc_nr_events(&evsel->hists, event->header.type);
828 			perf_event__process(&top->tool, event, &sample, machine);
829 		} else
830 			++session->hists.stats.nr_unknown_events;
831 	}
832 }
833 
834 static void perf_top__mmap_read(struct perf_top *top)
835 {
836 	int i;
837 
838 	for (i = 0; i < top->evlist->nr_mmaps; i++)
839 		perf_top__mmap_read_idx(top, i);
840 }
841 
842 static void perf_top__start_counters(struct perf_top *top)
843 {
844 	struct perf_evsel *counter, *first;
845 	struct perf_evlist *evlist = top->evlist;
846 
847 	first = list_entry(evlist->entries.next, struct perf_evsel, node);
848 
849 	list_for_each_entry(counter, &evlist->entries, node) {
850 		struct perf_event_attr *attr = &counter->attr;
851 		struct xyarray *group_fd = NULL;
852 
853 		if (top->group && counter != first)
854 			group_fd = first->fd;
855 
856 		attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
857 
858 		if (top->freq) {
859 			attr->sample_type |= PERF_SAMPLE_PERIOD;
860 			attr->freq	  = 1;
861 			attr->sample_freq = top->freq;
862 		}
863 
864 		if (evlist->nr_entries > 1) {
865 			attr->sample_type |= PERF_SAMPLE_ID;
866 			attr->read_format |= PERF_FORMAT_ID;
867 		}
868 
869 		if (symbol_conf.use_callchain)
870 			attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
871 
872 		attr->mmap = 1;
873 		attr->comm = 1;
874 		attr->inherit = top->inherit;
875 fallback_missing_features:
876 		if (top->exclude_guest_missing)
877 			attr->exclude_guest = attr->exclude_host = 0;
878 retry_sample_id:
879 		attr->sample_id_all = top->sample_id_all_missing ? 0 : 1;
880 try_again:
881 		if (perf_evsel__open(counter, top->evlist->cpus,
882 				     top->evlist->threads, top->group,
883 				     group_fd) < 0) {
884 			int err = errno;
885 
886 			if (err == EPERM || err == EACCES) {
887 				ui__error_paranoid();
888 				goto out_err;
889 			} else if (err == EINVAL) {
890 				if (!top->exclude_guest_missing &&
891 				    (attr->exclude_guest || attr->exclude_host)) {
892 					pr_debug("Old kernel, cannot exclude "
893 						 "guest or host samples.\n");
894 					top->exclude_guest_missing = true;
895 					goto fallback_missing_features;
896 				} else if (!top->sample_id_all_missing) {
897 					/*
898 					 * Old kernel, no attr->sample_id_type_all field
899 					 */
900 					top->sample_id_all_missing = true;
901 					goto retry_sample_id;
902 				}
903 			}
904 			/*
905 			 * If it's cycles then fall back to hrtimer
906 			 * based cpu-clock-tick sw counter, which
907 			 * is always available even if no PMU support:
908 			 */
909 			if (attr->type == PERF_TYPE_HARDWARE &&
910 			    attr->config == PERF_COUNT_HW_CPU_CYCLES) {
911 				if (verbose)
912 					ui__warning("Cycles event not supported,\n"
913 						    "trying to fall back to cpu-clock-ticks\n");
914 
915 				attr->type = PERF_TYPE_SOFTWARE;
916 				attr->config = PERF_COUNT_SW_CPU_CLOCK;
917 				goto try_again;
918 			}
919 
920 			if (err == ENOENT) {
921 				ui__warning("The %s event is not supported.\n",
922 					    event_name(counter));
923 				goto out_err;
924 			} else if (err == EMFILE) {
925 				ui__warning("Too many events are opened.\n"
926 					    "Try again after reducing the number of events\n");
927 				goto out_err;
928 			}
929 
930 			ui__warning("The sys_perf_event_open() syscall "
931 				    "returned with %d (%s).  /bin/dmesg "
932 				    "may provide additional information.\n"
933 				    "No CONFIG_PERF_EVENTS=y kernel support "
934 				    "configured?\n", err, strerror(err));
935 			goto out_err;
936 		}
937 	}
938 
939 	if (perf_evlist__mmap(evlist, top->mmap_pages, false) < 0) {
940 		ui__warning("Failed to mmap with %d (%s)\n",
941 			    errno, strerror(errno));
942 		goto out_err;
943 	}
944 
945 	return;
946 
947 out_err:
948 	exit_browser(0);
949 	exit(0);
950 }
951 
952 static int perf_top__setup_sample_type(struct perf_top *top)
953 {
954 	if (!top->sort_has_symbols) {
955 		if (symbol_conf.use_callchain) {
956 			ui__warning("Selected -g but \"sym\" not present in --sort/-s.");
957 			return -EINVAL;
958 		}
959 	} else if (!top->dont_use_callchains && callchain_param.mode != CHAIN_NONE) {
960 		if (callchain_register_param(&callchain_param) < 0) {
961 			ui__warning("Can't register callchain params.\n");
962 			return -EINVAL;
963 		}
964 	}
965 
966 	return 0;
967 }
968 
969 static int __cmd_top(struct perf_top *top)
970 {
971 	pthread_t thread;
972 	int ret;
973 	/*
974 	 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
975 	 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
976 	 */
977 	top->session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
978 	if (top->session == NULL)
979 		return -ENOMEM;
980 
981 	ret = perf_top__setup_sample_type(top);
982 	if (ret)
983 		goto out_delete;
984 
985 	if (top->target_tid || top->uid != UINT_MAX)
986 		perf_event__synthesize_thread_map(&top->tool, top->evlist->threads,
987 						  perf_event__process,
988 						  &top->session->host_machine);
989 	else
990 		perf_event__synthesize_threads(&top->tool, perf_event__process,
991 					       &top->session->host_machine);
992 	perf_top__start_counters(top);
993 	top->session->evlist = top->evlist;
994 	perf_session__update_sample_type(top->session);
995 
996 	/* Wait for a minimal set of events before starting the snapshot */
997 	poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
998 
999 	perf_top__mmap_read(top);
1000 
1001 	if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1002 							    display_thread), top)) {
1003 		printf("Could not create display thread.\n");
1004 		exit(-1);
1005 	}
1006 
1007 	if (top->realtime_prio) {
1008 		struct sched_param param;
1009 
1010 		param.sched_priority = top->realtime_prio;
1011 		if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1012 			printf("Could not set realtime priority.\n");
1013 			exit(-1);
1014 		}
1015 	}
1016 
1017 	while (1) {
1018 		u64 hits = top->samples;
1019 
1020 		perf_top__mmap_read(top);
1021 
1022 		if (hits == top->samples)
1023 			ret = poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
1024 	}
1025 
1026 out_delete:
1027 	perf_session__delete(top->session);
1028 	top->session = NULL;
1029 
1030 	return 0;
1031 }
1032 
1033 static int
1034 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1035 {
1036 	struct perf_top *top = (struct perf_top *)opt->value;
1037 	char *tok, *tok2;
1038 	char *endptr;
1039 
1040 	/*
1041 	 * --no-call-graph
1042 	 */
1043 	if (unset) {
1044 		top->dont_use_callchains = true;
1045 		return 0;
1046 	}
1047 
1048 	symbol_conf.use_callchain = true;
1049 
1050 	if (!arg)
1051 		return 0;
1052 
1053 	tok = strtok((char *)arg, ",");
1054 	if (!tok)
1055 		return -1;
1056 
1057 	/* get the output mode */
1058 	if (!strncmp(tok, "graph", strlen(arg)))
1059 		callchain_param.mode = CHAIN_GRAPH_ABS;
1060 
1061 	else if (!strncmp(tok, "flat", strlen(arg)))
1062 		callchain_param.mode = CHAIN_FLAT;
1063 
1064 	else if (!strncmp(tok, "fractal", strlen(arg)))
1065 		callchain_param.mode = CHAIN_GRAPH_REL;
1066 
1067 	else if (!strncmp(tok, "none", strlen(arg))) {
1068 		callchain_param.mode = CHAIN_NONE;
1069 		symbol_conf.use_callchain = false;
1070 
1071 		return 0;
1072 	} else
1073 		return -1;
1074 
1075 	/* get the min percentage */
1076 	tok = strtok(NULL, ",");
1077 	if (!tok)
1078 		goto setup;
1079 
1080 	callchain_param.min_percent = strtod(tok, &endptr);
1081 	if (tok == endptr)
1082 		return -1;
1083 
1084 	/* get the print limit */
1085 	tok2 = strtok(NULL, ",");
1086 	if (!tok2)
1087 		goto setup;
1088 
1089 	if (tok2[0] != 'c') {
1090 		callchain_param.print_limit = strtod(tok2, &endptr);
1091 		tok2 = strtok(NULL, ",");
1092 		if (!tok2)
1093 			goto setup;
1094 	}
1095 
1096 	/* get the call chain order */
1097 	if (!strcmp(tok2, "caller"))
1098 		callchain_param.order = ORDER_CALLER;
1099 	else if (!strcmp(tok2, "callee"))
1100 		callchain_param.order = ORDER_CALLEE;
1101 	else
1102 		return -1;
1103 setup:
1104 	if (callchain_register_param(&callchain_param) < 0) {
1105 		fprintf(stderr, "Can't register callchain params\n");
1106 		return -1;
1107 	}
1108 	return 0;
1109 }
1110 
1111 static const char * const top_usage[] = {
1112 	"perf top [<options>]",
1113 	NULL
1114 };
1115 
1116 int cmd_top(int argc, const char **argv, const char *prefix __used)
1117 {
1118 	struct perf_evsel *pos;
1119 	int status = -ENOMEM;
1120 	struct perf_top top = {
1121 		.count_filter	     = 5,
1122 		.delay_secs	     = 2,
1123 		.uid		     = UINT_MAX,
1124 		.freq		     = 1000, /* 1 KHz */
1125 		.mmap_pages	     = 128,
1126 		.sym_pcnt_filter     = 5,
1127 	};
1128 	char callchain_default_opt[] = "fractal,0.5,callee";
1129 	const struct option options[] = {
1130 	OPT_CALLBACK('e', "event", &top.evlist, "event",
1131 		     "event selector. use 'perf list' to list available events",
1132 		     parse_events_option),
1133 	OPT_INTEGER('c', "count", &top.default_interval,
1134 		    "event period to sample"),
1135 	OPT_STRING('p', "pid", &top.target_pid, "pid",
1136 		    "profile events on existing process id"),
1137 	OPT_STRING('t', "tid", &top.target_tid, "tid",
1138 		    "profile events on existing thread id"),
1139 	OPT_BOOLEAN('a', "all-cpus", &top.system_wide,
1140 			    "system-wide collection from all CPUs"),
1141 	OPT_STRING('C', "cpu", &top.cpu_list, "cpu",
1142 		    "list of cpus to monitor"),
1143 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1144 		   "file", "vmlinux pathname"),
1145 	OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1146 		    "hide kernel symbols"),
1147 	OPT_UINTEGER('m', "mmap-pages", &top.mmap_pages, "number of mmap data pages"),
1148 	OPT_INTEGER('r', "realtime", &top.realtime_prio,
1149 		    "collect data with this RT SCHED_FIFO priority"),
1150 	OPT_INTEGER('d', "delay", &top.delay_secs,
1151 		    "number of seconds to delay between refreshes"),
1152 	OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1153 			    "dump the symbol table used for profiling"),
1154 	OPT_INTEGER('f', "count-filter", &top.count_filter,
1155 		    "only display functions with more events than this"),
1156 	OPT_BOOLEAN('g', "group", &top.group,
1157 			    "put the counters into a counter group"),
1158 	OPT_BOOLEAN('i', "inherit", &top.inherit,
1159 		    "child tasks inherit counters"),
1160 	OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1161 		    "symbol to annotate"),
1162 	OPT_BOOLEAN('z', "zero", &top.zero,
1163 		    "zero history across updates"),
1164 	OPT_INTEGER('F', "freq", &top.freq,
1165 		    "profile at this frequency"),
1166 	OPT_INTEGER('E', "entries", &top.print_entries,
1167 		    "display this many functions"),
1168 	OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1169 		    "hide user symbols"),
1170 	OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1171 	OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1172 	OPT_INCR('v', "verbose", &verbose,
1173 		    "be more verbose (show counter open errors, etc)"),
1174 	OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1175 		   "sort by key(s): pid, comm, dso, symbol, parent"),
1176 	OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1177 		    "Show a column with the number of samples"),
1178 	OPT_CALLBACK_DEFAULT('G', "call-graph", &top, "output_type,min_percent, call_order",
1179 		     "Display callchains using output_type (graph, flat, fractal, or none), min percent threshold and callchain order. "
1180 		     "Default: fractal,0.5,callee", &parse_callchain_opt,
1181 		     callchain_default_opt),
1182 	OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1183 		    "Show a column with the sum of periods"),
1184 	OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1185 		   "only consider symbols in these dsos"),
1186 	OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1187 		   "only consider symbols in these comms"),
1188 	OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1189 		   "only consider these symbols"),
1190 	OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1191 		    "Interleave source code with assembly code (default)"),
1192 	OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1193 		    "Display raw encoding of assembly instructions (default)"),
1194 	OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1195 		   "Specify disassembler style (e.g. -M intel for intel syntax)"),
1196 	OPT_STRING('u', "uid", &top.uid_str, "user", "user to profile"),
1197 	OPT_END()
1198 	};
1199 
1200 	top.evlist = perf_evlist__new(NULL, NULL);
1201 	if (top.evlist == NULL)
1202 		return -ENOMEM;
1203 
1204 	symbol_conf.exclude_other = false;
1205 
1206 	argc = parse_options(argc, argv, options, top_usage, 0);
1207 	if (argc)
1208 		usage_with_options(top_usage, options);
1209 
1210 	if (sort_order == default_sort_order)
1211 		sort_order = "dso,symbol";
1212 
1213 	setup_sorting(top_usage, options);
1214 
1215 	if (top.use_stdio)
1216 		use_browser = 0;
1217 	else if (top.use_tui)
1218 		use_browser = 1;
1219 
1220 	setup_browser(false);
1221 
1222 	top.uid = parse_target_uid(top.uid_str, top.target_tid, top.target_pid);
1223 	if (top.uid_str != NULL && top.uid == UINT_MAX - 1)
1224 		goto out_delete_evlist;
1225 
1226 	/* CPU and PID are mutually exclusive */
1227 	if (top.target_tid && top.cpu_list) {
1228 		printf("WARNING: PID switch overriding CPU\n");
1229 		sleep(1);
1230 		top.cpu_list = NULL;
1231 	}
1232 
1233 	if (top.target_pid)
1234 		top.target_tid = top.target_pid;
1235 
1236 	if (perf_evlist__create_maps(top.evlist, top.target_pid,
1237 				     top.target_tid, top.uid, top.cpu_list) < 0)
1238 		usage_with_options(top_usage, options);
1239 
1240 	if (!top.evlist->nr_entries &&
1241 	    perf_evlist__add_default(top.evlist) < 0) {
1242 		pr_err("Not enough memory for event selector list\n");
1243 		return -ENOMEM;
1244 	}
1245 
1246 	symbol_conf.nr_events = top.evlist->nr_entries;
1247 
1248 	if (top.delay_secs < 1)
1249 		top.delay_secs = 1;
1250 
1251 	/*
1252 	 * User specified count overrides default frequency.
1253 	 */
1254 	if (top.default_interval)
1255 		top.freq = 0;
1256 	else if (top.freq) {
1257 		top.default_interval = top.freq;
1258 	} else {
1259 		fprintf(stderr, "frequency and count are zero, aborting\n");
1260 		exit(EXIT_FAILURE);
1261 	}
1262 
1263 	list_for_each_entry(pos, &top.evlist->entries, node) {
1264 		/*
1265 		 * Fill in the ones not specifically initialized via -c:
1266 		 */
1267 		if (!pos->attr.sample_period)
1268 			pos->attr.sample_period = top.default_interval;
1269 	}
1270 
1271 	top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1272 
1273 	symbol_conf.priv_size = sizeof(struct annotation);
1274 
1275 	symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1276 	if (symbol__init() < 0)
1277 		return -1;
1278 
1279 	sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1280 	sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1281 	sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1282 
1283 	/*
1284 	 * Avoid annotation data structures overhead when symbols aren't on the
1285 	 * sort list.
1286 	 */
1287 	top.sort_has_symbols = sort_sym.list.next != NULL;
1288 
1289 	get_term_dimensions(&top.winsize);
1290 	if (top.print_entries == 0) {
1291 		struct sigaction act = {
1292 			.sa_sigaction = perf_top__sig_winch,
1293 			.sa_flags     = SA_SIGINFO,
1294 		};
1295 		perf_top__update_print_entries(&top);
1296 		sigaction(SIGWINCH, &act, NULL);
1297 	}
1298 
1299 	status = __cmd_top(&top);
1300 
1301 out_delete_evlist:
1302 	perf_evlist__delete(top.evlist);
1303 
1304 	return status;
1305 }
1306