xref: /linux/tools/perf/builtin-top.c (revision c1a604dff486399ae0be95e6396e0158df95ad5d)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * builtin-top.c
4  *
5  * Builtin top command: Display a continuously updated profile of
6  * any workload, CPU or specific PID.
7  *
8  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
9  *		 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
10  *
11  * Improvements and fixes by:
12  *
13  *   Arjan van de Ven <arjan@linux.intel.com>
14  *   Yanmin Zhang <yanmin.zhang@intel.com>
15  *   Wu Fengguang <fengguang.wu@intel.com>
16  *   Mike Galbraith <efault@gmx.de>
17  *   Paul Mackerras <paulus@samba.org>
18  */
19 #include "builtin.h"
20 
21 #include "perf.h"
22 
23 #include "util/annotate.h"
24 #include "util/bpf-event.h"
25 #include "util/config.h"
26 #include "util/color.h"
27 #include "util/evlist.h"
28 #include "util/evsel.h"
29 #include "util/event.h"
30 #include "util/machine.h"
31 #include "util/map.h"
32 #include "util/session.h"
33 #include "util/symbol.h"
34 #include "util/thread.h"
35 #include "util/thread_map.h"
36 #include "util/top.h"
37 #include "util/util.h"
38 #include <linux/rbtree.h>
39 #include <subcmd/parse-options.h>
40 #include "util/parse-events.h"
41 #include "util/cpumap.h"
42 #include "util/sort.h"
43 #include "util/string2.h"
44 #include "util/term.h"
45 #include "util/intlist.h"
46 #include "util/parse-branch-options.h"
47 #include "arch/common.h"
48 
49 #include "util/debug.h"
50 #include "util/ordered-events.h"
51 
52 #include <assert.h>
53 #include <elf.h>
54 #include <fcntl.h>
55 
56 #include <stdio.h>
57 #include <termios.h>
58 #include <unistd.h>
59 #include <inttypes.h>
60 
61 #include <errno.h>
62 #include <time.h>
63 #include <sched.h>
64 #include <signal.h>
65 
66 #include <sys/syscall.h>
67 #include <sys/ioctl.h>
68 #include <poll.h>
69 #include <sys/prctl.h>
70 #include <sys/wait.h>
71 #include <sys/uio.h>
72 #include <sys/utsname.h>
73 #include <sys/mman.h>
74 
75 #include <linux/stringify.h>
76 #include <linux/time64.h>
77 #include <linux/types.h>
78 
79 #include <linux/ctype.h>
80 
81 static volatile int done;
82 static volatile int resize;
83 
84 #define HEADER_LINE_NR  5
85 
86 static void perf_top__update_print_entries(struct perf_top *top)
87 {
88 	top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
89 }
90 
91 static void winch_sig(int sig __maybe_unused)
92 {
93 	resize = 1;
94 }
95 
96 static void perf_top__resize(struct perf_top *top)
97 {
98 	get_term_dimensions(&top->winsize);
99 	perf_top__update_print_entries(top);
100 }
101 
102 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
103 {
104 	struct evsel *evsel;
105 	struct symbol *sym;
106 	struct annotation *notes;
107 	struct map *map;
108 	int err = -1;
109 
110 	if (!he || !he->ms.sym)
111 		return -1;
112 
113 	evsel = hists_to_evsel(he->hists);
114 
115 	sym = he->ms.sym;
116 	map = he->ms.map;
117 
118 	/*
119 	 * We can't annotate with just /proc/kallsyms
120 	 */
121 	if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
122 	    !dso__is_kcore(map->dso)) {
123 		pr_err("Can't annotate %s: No vmlinux file was found in the "
124 		       "path\n", sym->name);
125 		sleep(1);
126 		return -1;
127 	}
128 
129 	notes = symbol__annotation(sym);
130 	pthread_mutex_lock(&notes->lock);
131 
132 	if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
133 		pthread_mutex_unlock(&notes->lock);
134 		pr_err("Not enough memory for annotating '%s' symbol!\n",
135 		       sym->name);
136 		sleep(1);
137 		return err;
138 	}
139 
140 	err = symbol__annotate(sym, map, evsel, 0, &top->annotation_opts, NULL);
141 	if (err == 0) {
142 		top->sym_filter_entry = he;
143 	} else {
144 		char msg[BUFSIZ];
145 		symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
146 		pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
147 	}
148 
149 	pthread_mutex_unlock(&notes->lock);
150 	return err;
151 }
152 
153 static void __zero_source_counters(struct hist_entry *he)
154 {
155 	struct symbol *sym = he->ms.sym;
156 	symbol__annotate_zero_histograms(sym);
157 }
158 
159 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
160 {
161 	struct utsname uts;
162 	int err = uname(&uts);
163 
164 	ui__warning("Out of bounds address found:\n\n"
165 		    "Addr:   %" PRIx64 "\n"
166 		    "DSO:    %s %c\n"
167 		    "Map:    %" PRIx64 "-%" PRIx64 "\n"
168 		    "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
169 		    "Arch:   %s\n"
170 		    "Kernel: %s\n"
171 		    "Tools:  %s\n\n"
172 		    "Not all samples will be on the annotation output.\n\n"
173 		    "Please report to linux-kernel@vger.kernel.org\n",
174 		    ip, map->dso->long_name, dso__symtab_origin(map->dso),
175 		    map->start, map->end, sym->start, sym->end,
176 		    sym->binding == STB_GLOBAL ? 'g' :
177 		    sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
178 		    err ? "[unknown]" : uts.machine,
179 		    err ? "[unknown]" : uts.release, perf_version_string);
180 	if (use_browser <= 0)
181 		sleep(5);
182 
183 	map->erange_warned = true;
184 }
185 
186 static void perf_top__record_precise_ip(struct perf_top *top,
187 					struct hist_entry *he,
188 					struct perf_sample *sample,
189 					struct evsel *evsel, u64 ip)
190 {
191 	struct annotation *notes;
192 	struct symbol *sym = he->ms.sym;
193 	int err = 0;
194 
195 	if (sym == NULL || (use_browser == 0 &&
196 			    (top->sym_filter_entry == NULL ||
197 			     top->sym_filter_entry->ms.sym != sym)))
198 		return;
199 
200 	notes = symbol__annotation(sym);
201 
202 	if (pthread_mutex_trylock(&notes->lock))
203 		return;
204 
205 	err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
206 
207 	pthread_mutex_unlock(&notes->lock);
208 
209 	if (unlikely(err)) {
210 		/*
211 		 * This function is now called with he->hists->lock held.
212 		 * Release it before going to sleep.
213 		 */
214 		pthread_mutex_unlock(&he->hists->lock);
215 
216 		if (err == -ERANGE && !he->ms.map->erange_warned)
217 			ui__warn_map_erange(he->ms.map, sym, ip);
218 		else if (err == -ENOMEM) {
219 			pr_err("Not enough memory for annotating '%s' symbol!\n",
220 			       sym->name);
221 			sleep(1);
222 		}
223 
224 		pthread_mutex_lock(&he->hists->lock);
225 	}
226 }
227 
228 static void perf_top__show_details(struct perf_top *top)
229 {
230 	struct hist_entry *he = top->sym_filter_entry;
231 	struct evsel *evsel;
232 	struct annotation *notes;
233 	struct symbol *symbol;
234 	int more;
235 
236 	if (!he)
237 		return;
238 
239 	evsel = hists_to_evsel(he->hists);
240 
241 	symbol = he->ms.sym;
242 	notes = symbol__annotation(symbol);
243 
244 	pthread_mutex_lock(&notes->lock);
245 
246 	symbol__calc_percent(symbol, evsel);
247 
248 	if (notes->src == NULL)
249 		goto out_unlock;
250 
251 	printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
252 	printf("  Events  Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
253 
254 	more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel, &top->annotation_opts);
255 
256 	if (top->evlist->enabled) {
257 		if (top->zero)
258 			symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
259 		else
260 			symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
261 	}
262 	if (more != 0)
263 		printf("%d lines not displayed, maybe increase display entries [e]\n", more);
264 out_unlock:
265 	pthread_mutex_unlock(&notes->lock);
266 }
267 
268 static void perf_top__resort_hists(struct perf_top *t)
269 {
270 	struct evlist *evlist = t->evlist;
271 	struct evsel *pos;
272 
273 	evlist__for_each_entry(evlist, pos) {
274 		struct hists *hists = evsel__hists(pos);
275 
276 		/*
277 		 * unlink existing entries so that they can be linked
278 		 * in a correct order in hists__match() below.
279 		 */
280 		hists__unlink(hists);
281 
282 		if (evlist->enabled) {
283 			if (t->zero) {
284 				hists__delete_entries(hists);
285 			} else {
286 				hists__decay_entries(hists, t->hide_user_symbols,
287 						     t->hide_kernel_symbols);
288 			}
289 		}
290 
291 		hists__collapse_resort(hists, NULL);
292 
293 		/* Non-group events are considered as leader */
294 		if (symbol_conf.event_group &&
295 		    !perf_evsel__is_group_leader(pos)) {
296 			struct hists *leader_hists = evsel__hists(pos->leader);
297 
298 			hists__match(leader_hists, hists);
299 			hists__link(leader_hists, hists);
300 		}
301 	}
302 
303 	evlist__for_each_entry(evlist, pos) {
304 		perf_evsel__output_resort(pos, NULL);
305 	}
306 }
307 
308 static void perf_top__print_sym_table(struct perf_top *top)
309 {
310 	char bf[160];
311 	int printed = 0;
312 	const int win_width = top->winsize.ws_col - 1;
313 	struct evsel *evsel = top->sym_evsel;
314 	struct hists *hists = evsel__hists(evsel);
315 
316 	puts(CONSOLE_CLEAR);
317 
318 	perf_top__header_snprintf(top, bf, sizeof(bf));
319 	printf("%s\n", bf);
320 
321 	printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
322 
323 	if (!top->record_opts.overwrite &&
324 	    (hists->stats.nr_lost_warned !=
325 	    hists->stats.nr_events[PERF_RECORD_LOST])) {
326 		hists->stats.nr_lost_warned =
327 			      hists->stats.nr_events[PERF_RECORD_LOST];
328 		color_fprintf(stdout, PERF_COLOR_RED,
329 			      "WARNING: LOST %d chunks, Check IO/CPU overload",
330 			      hists->stats.nr_lost_warned);
331 		++printed;
332 	}
333 
334 	if (top->sym_filter_entry) {
335 		perf_top__show_details(top);
336 		return;
337 	}
338 
339 	perf_top__resort_hists(top);
340 
341 	hists__output_recalc_col_len(hists, top->print_entries - printed);
342 	putchar('\n');
343 	hists__fprintf(hists, false, top->print_entries - printed, win_width,
344 		       top->min_percent, stdout, !symbol_conf.use_callchain);
345 }
346 
347 static void prompt_integer(int *target, const char *msg)
348 {
349 	char *buf = malloc(0), *p;
350 	size_t dummy = 0;
351 	int tmp;
352 
353 	fprintf(stdout, "\n%s: ", msg);
354 	if (getline(&buf, &dummy, stdin) < 0)
355 		return;
356 
357 	p = strchr(buf, '\n');
358 	if (p)
359 		*p = 0;
360 
361 	p = buf;
362 	while(*p) {
363 		if (!isdigit(*p))
364 			goto out_free;
365 		p++;
366 	}
367 	tmp = strtoul(buf, NULL, 10);
368 	*target = tmp;
369 out_free:
370 	free(buf);
371 }
372 
373 static void prompt_percent(int *target, const char *msg)
374 {
375 	int tmp = 0;
376 
377 	prompt_integer(&tmp, msg);
378 	if (tmp >= 0 && tmp <= 100)
379 		*target = tmp;
380 }
381 
382 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
383 {
384 	char *buf = malloc(0), *p;
385 	struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
386 	struct hists *hists = evsel__hists(top->sym_evsel);
387 	struct rb_node *next;
388 	size_t dummy = 0;
389 
390 	/* zero counters of active symbol */
391 	if (syme) {
392 		__zero_source_counters(syme);
393 		top->sym_filter_entry = NULL;
394 	}
395 
396 	fprintf(stdout, "\n%s: ", msg);
397 	if (getline(&buf, &dummy, stdin) < 0)
398 		goto out_free;
399 
400 	p = strchr(buf, '\n');
401 	if (p)
402 		*p = 0;
403 
404 	next = rb_first_cached(&hists->entries);
405 	while (next) {
406 		n = rb_entry(next, struct hist_entry, rb_node);
407 		if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
408 			found = n;
409 			break;
410 		}
411 		next = rb_next(&n->rb_node);
412 	}
413 
414 	if (!found) {
415 		fprintf(stderr, "Sorry, %s is not active.\n", buf);
416 		sleep(1);
417 	} else
418 		perf_top__parse_source(top, found);
419 
420 out_free:
421 	free(buf);
422 }
423 
424 static void perf_top__print_mapped_keys(struct perf_top *top)
425 {
426 	char *name = NULL;
427 
428 	if (top->sym_filter_entry) {
429 		struct symbol *sym = top->sym_filter_entry->ms.sym;
430 		name = sym->name;
431 	}
432 
433 	fprintf(stdout, "\nMapped keys:\n");
434 	fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
435 	fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
436 
437 	if (top->evlist->core.nr_entries > 1)
438 		fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
439 
440 	fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
441 
442 	fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
443 	fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
444 	fprintf(stdout, "\t[S]     stop annotation.\n");
445 
446 	fprintf(stdout,
447 		"\t[K]     hide kernel symbols.             \t(%s)\n",
448 		top->hide_kernel_symbols ? "yes" : "no");
449 	fprintf(stdout,
450 		"\t[U]     hide user symbols.               \t(%s)\n",
451 		top->hide_user_symbols ? "yes" : "no");
452 	fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
453 	fprintf(stdout, "\t[qQ]    quit.\n");
454 }
455 
456 static int perf_top__key_mapped(struct perf_top *top, int c)
457 {
458 	switch (c) {
459 		case 'd':
460 		case 'e':
461 		case 'f':
462 		case 'z':
463 		case 'q':
464 		case 'Q':
465 		case 'K':
466 		case 'U':
467 		case 'F':
468 		case 's':
469 		case 'S':
470 			return 1;
471 		case 'E':
472 			return top->evlist->core.nr_entries > 1 ? 1 : 0;
473 		default:
474 			break;
475 	}
476 
477 	return 0;
478 }
479 
480 static bool perf_top__handle_keypress(struct perf_top *top, int c)
481 {
482 	bool ret = true;
483 
484 	if (!perf_top__key_mapped(top, c)) {
485 		struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
486 		struct termios save;
487 
488 		perf_top__print_mapped_keys(top);
489 		fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
490 		fflush(stdout);
491 
492 		set_term_quiet_input(&save);
493 
494 		poll(&stdin_poll, 1, -1);
495 		c = getc(stdin);
496 
497 		tcsetattr(0, TCSAFLUSH, &save);
498 		if (!perf_top__key_mapped(top, c))
499 			return ret;
500 	}
501 
502 	switch (c) {
503 		case 'd':
504 			prompt_integer(&top->delay_secs, "Enter display delay");
505 			if (top->delay_secs < 1)
506 				top->delay_secs = 1;
507 			break;
508 		case 'e':
509 			prompt_integer(&top->print_entries, "Enter display entries (lines)");
510 			if (top->print_entries == 0) {
511 				perf_top__resize(top);
512 				signal(SIGWINCH, winch_sig);
513 			} else {
514 				signal(SIGWINCH, SIG_DFL);
515 			}
516 			break;
517 		case 'E':
518 			if (top->evlist->core.nr_entries > 1) {
519 				/* Select 0 as the default event: */
520 				int counter = 0;
521 
522 				fprintf(stderr, "\nAvailable events:");
523 
524 				evlist__for_each_entry(top->evlist, top->sym_evsel)
525 					fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
526 
527 				prompt_integer(&counter, "Enter details event counter");
528 
529 				if (counter >= top->evlist->core.nr_entries) {
530 					top->sym_evsel = perf_evlist__first(top->evlist);
531 					fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
532 					sleep(1);
533 					break;
534 				}
535 				evlist__for_each_entry(top->evlist, top->sym_evsel)
536 					if (top->sym_evsel->idx == counter)
537 						break;
538 			} else
539 				top->sym_evsel = perf_evlist__first(top->evlist);
540 			break;
541 		case 'f':
542 			prompt_integer(&top->count_filter, "Enter display event count filter");
543 			break;
544 		case 'F':
545 			prompt_percent(&top->annotation_opts.min_pcnt,
546 				       "Enter details display event filter (percent)");
547 			break;
548 		case 'K':
549 			top->hide_kernel_symbols = !top->hide_kernel_symbols;
550 			break;
551 		case 'q':
552 		case 'Q':
553 			printf("exiting.\n");
554 			if (top->dump_symtab)
555 				perf_session__fprintf_dsos(top->session, stderr);
556 			ret = false;
557 			break;
558 		case 's':
559 			perf_top__prompt_symbol(top, "Enter details symbol");
560 			break;
561 		case 'S':
562 			if (!top->sym_filter_entry)
563 				break;
564 			else {
565 				struct hist_entry *syme = top->sym_filter_entry;
566 
567 				top->sym_filter_entry = NULL;
568 				__zero_source_counters(syme);
569 			}
570 			break;
571 		case 'U':
572 			top->hide_user_symbols = !top->hide_user_symbols;
573 			break;
574 		case 'z':
575 			top->zero = !top->zero;
576 			break;
577 		default:
578 			break;
579 	}
580 
581 	return ret;
582 }
583 
584 static void perf_top__sort_new_samples(void *arg)
585 {
586 	struct perf_top *t = arg;
587 
588 	if (t->evlist->selected != NULL)
589 		t->sym_evsel = t->evlist->selected;
590 
591 	perf_top__resort_hists(t);
592 
593 	if (t->lost || t->drop)
594 		pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
595 }
596 
597 static void stop_top(void)
598 {
599 	session_done = 1;
600 	done = 1;
601 }
602 
603 static void *display_thread_tui(void *arg)
604 {
605 	struct evsel *pos;
606 	struct perf_top *top = arg;
607 	const char *help = "For a higher level overview, try: perf top --sort comm,dso";
608 	struct hist_browser_timer hbt = {
609 		.timer		= perf_top__sort_new_samples,
610 		.arg		= top,
611 		.refresh	= top->delay_secs,
612 	};
613 
614 	/* In order to read symbols from other namespaces perf to  needs to call
615 	 * setns(2).  This isn't permitted if the struct_fs has multiple users.
616 	 * unshare(2) the fs so that we may continue to setns into namespaces
617 	 * that we're observing.
618 	 */
619 	unshare(CLONE_FS);
620 
621 	prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
622 
623 	perf_top__sort_new_samples(top);
624 
625 	/*
626 	 * Initialize the uid_filter_str, in the future the TUI will allow
627 	 * Zooming in/out UIDs. For now just use whatever the user passed
628 	 * via --uid.
629 	 */
630 	evlist__for_each_entry(top->evlist, pos) {
631 		struct hists *hists = evsel__hists(pos);
632 		hists->uid_filter_str = top->record_opts.target.uid_str;
633 	}
634 
635 	perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
636 				      top->min_percent,
637 				      &top->session->header.env,
638 				      !top->record_opts.overwrite,
639 				      &top->annotation_opts);
640 
641 	stop_top();
642 	return NULL;
643 }
644 
645 static void display_sig(int sig __maybe_unused)
646 {
647 	stop_top();
648 }
649 
650 static void display_setup_sig(void)
651 {
652 	signal(SIGSEGV, sighandler_dump_stack);
653 	signal(SIGFPE, sighandler_dump_stack);
654 	signal(SIGINT,  display_sig);
655 	signal(SIGQUIT, display_sig);
656 	signal(SIGTERM, display_sig);
657 }
658 
659 static void *display_thread(void *arg)
660 {
661 	struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
662 	struct termios save;
663 	struct perf_top *top = arg;
664 	int delay_msecs, c;
665 
666 	/* In order to read symbols from other namespaces perf to  needs to call
667 	 * setns(2).  This isn't permitted if the struct_fs has multiple users.
668 	 * unshare(2) the fs so that we may continue to setns into namespaces
669 	 * that we're observing.
670 	 */
671 	unshare(CLONE_FS);
672 
673 	prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
674 
675 	display_setup_sig();
676 	pthread__unblock_sigwinch();
677 repeat:
678 	delay_msecs = top->delay_secs * MSEC_PER_SEC;
679 	set_term_quiet_input(&save);
680 	/* trash return*/
681 	getc(stdin);
682 
683 	while (!done) {
684 		perf_top__print_sym_table(top);
685 		/*
686 		 * Either timeout expired or we got an EINTR due to SIGWINCH,
687 		 * refresh screen in both cases.
688 		 */
689 		switch (poll(&stdin_poll, 1, delay_msecs)) {
690 		case 0:
691 			continue;
692 		case -1:
693 			if (errno == EINTR)
694 				continue;
695 			__fallthrough;
696 		default:
697 			c = getc(stdin);
698 			tcsetattr(0, TCSAFLUSH, &save);
699 
700 			if (perf_top__handle_keypress(top, c))
701 				goto repeat;
702 			stop_top();
703 		}
704 	}
705 
706 	tcsetattr(0, TCSAFLUSH, &save);
707 	return NULL;
708 }
709 
710 static int hist_iter__top_callback(struct hist_entry_iter *iter,
711 				   struct addr_location *al, bool single,
712 				   void *arg)
713 {
714 	struct perf_top *top = arg;
715 	struct hist_entry *he = iter->he;
716 	struct evsel *evsel = iter->evsel;
717 
718 	if (perf_hpp_list.sym && single)
719 		perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
720 
721 	hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
722 		     !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
723 	return 0;
724 }
725 
726 static void perf_event__process_sample(struct perf_tool *tool,
727 				       const union perf_event *event,
728 				       struct evsel *evsel,
729 				       struct perf_sample *sample,
730 				       struct machine *machine)
731 {
732 	struct perf_top *top = container_of(tool, struct perf_top, tool);
733 	struct addr_location al;
734 	int err;
735 
736 	if (!machine && perf_guest) {
737 		static struct intlist *seen;
738 
739 		if (!seen)
740 			seen = intlist__new(NULL);
741 
742 		if (!intlist__has_entry(seen, sample->pid)) {
743 			pr_err("Can't find guest [%d]'s kernel information\n",
744 				sample->pid);
745 			intlist__add(seen, sample->pid);
746 		}
747 		return;
748 	}
749 
750 	if (!machine) {
751 		pr_err("%u unprocessable samples recorded.\r",
752 		       top->session->evlist->stats.nr_unprocessable_samples++);
753 		return;
754 	}
755 
756 	if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
757 		top->exact_samples++;
758 
759 	if (machine__resolve(machine, &al, sample) < 0)
760 		return;
761 
762 	if (!machine->kptr_restrict_warned &&
763 	    symbol_conf.kptr_restrict &&
764 	    al.cpumode == PERF_RECORD_MISC_KERNEL) {
765 		if (!perf_evlist__exclude_kernel(top->session->evlist)) {
766 			ui__warning(
767 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
768 "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
769 "Kernel%s samples will not be resolved.\n",
770 			  al.map && map__has_symbols(al.map) ?
771 			  " modules" : "");
772 			if (use_browser <= 0)
773 				sleep(5);
774 		}
775 		machine->kptr_restrict_warned = true;
776 	}
777 
778 	if (al.sym == NULL && al.map != NULL) {
779 		const char *msg = "Kernel samples will not be resolved.\n";
780 		/*
781 		 * As we do lazy loading of symtabs we only will know if the
782 		 * specified vmlinux file is invalid when we actually have a
783 		 * hit in kernel space and then try to load it. So if we get
784 		 * here and there are _no_ symbols in the DSO backing the
785 		 * kernel map, bail out.
786 		 *
787 		 * We may never get here, for instance, if we use -K/
788 		 * --hide-kernel-symbols, even if the user specifies an
789 		 * invalid --vmlinux ;-)
790 		 */
791 		if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
792 		    __map__is_kernel(al.map) && map__has_symbols(al.map)) {
793 			if (symbol_conf.vmlinux_name) {
794 				char serr[256];
795 				dso__strerror_load(al.map->dso, serr, sizeof(serr));
796 				ui__warning("The %s file can't be used: %s\n%s",
797 					    symbol_conf.vmlinux_name, serr, msg);
798 			} else {
799 				ui__warning("A vmlinux file was not found.\n%s",
800 					    msg);
801 			}
802 
803 			if (use_browser <= 0)
804 				sleep(5);
805 			top->vmlinux_warned = true;
806 		}
807 	}
808 
809 	if (al.sym == NULL || !al.sym->idle) {
810 		struct hists *hists = evsel__hists(evsel);
811 		struct hist_entry_iter iter = {
812 			.evsel		= evsel,
813 			.sample 	= sample,
814 			.add_entry_cb 	= hist_iter__top_callback,
815 		};
816 
817 		if (symbol_conf.cumulate_callchain)
818 			iter.ops = &hist_iter_cumulative;
819 		else
820 			iter.ops = &hist_iter_normal;
821 
822 		pthread_mutex_lock(&hists->lock);
823 
824 		err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
825 		if (err < 0)
826 			pr_err("Problem incrementing symbol period, skipping event\n");
827 
828 		pthread_mutex_unlock(&hists->lock);
829 	}
830 
831 	addr_location__put(&al);
832 }
833 
834 static void
835 perf_top__process_lost(struct perf_top *top, union perf_event *event,
836 		       struct evsel *evsel)
837 {
838 	struct hists *hists = evsel__hists(evsel);
839 
840 	top->lost += event->lost.lost;
841 	top->lost_total += event->lost.lost;
842 	hists->stats.total_lost += event->lost.lost;
843 }
844 
845 static void
846 perf_top__process_lost_samples(struct perf_top *top,
847 			       union perf_event *event,
848 			       struct evsel *evsel)
849 {
850 	struct hists *hists = evsel__hists(evsel);
851 
852 	top->lost += event->lost_samples.lost;
853 	top->lost_total += event->lost_samples.lost;
854 	hists->stats.total_lost_samples += event->lost_samples.lost;
855 }
856 
857 static u64 last_timestamp;
858 
859 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
860 {
861 	struct record_opts *opts = &top->record_opts;
862 	struct evlist *evlist = top->evlist;
863 	struct perf_mmap *md;
864 	union perf_event *event;
865 
866 	md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
867 	if (perf_mmap__read_init(md) < 0)
868 		return;
869 
870 	while ((event = perf_mmap__read_event(md)) != NULL) {
871 		int ret;
872 
873 		ret = perf_evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
874 		if (ret && ret != -1)
875 			break;
876 
877 		ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0);
878 		if (ret)
879 			break;
880 
881 		perf_mmap__consume(md);
882 
883 		if (top->qe.rotate) {
884 			pthread_mutex_lock(&top->qe.mutex);
885 			top->qe.rotate = false;
886 			pthread_cond_signal(&top->qe.cond);
887 			pthread_mutex_unlock(&top->qe.mutex);
888 		}
889 	}
890 
891 	perf_mmap__read_done(md);
892 }
893 
894 static void perf_top__mmap_read(struct perf_top *top)
895 {
896 	bool overwrite = top->record_opts.overwrite;
897 	struct evlist *evlist = top->evlist;
898 	int i;
899 
900 	if (overwrite)
901 		perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
902 
903 	for (i = 0; i < top->evlist->nr_mmaps; i++)
904 		perf_top__mmap_read_idx(top, i);
905 
906 	if (overwrite) {
907 		perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
908 		perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
909 	}
910 }
911 
912 /*
913  * Check per-event overwrite term.
914  * perf top should support consistent term for all events.
915  * - All events don't have per-event term
916  *   E.g. "cpu/cpu-cycles/,cpu/instructions/"
917  *   Nothing change, return 0.
918  * - All events have same per-event term
919  *   E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
920  *   Using the per-event setting to replace the opts->overwrite if
921  *   they are different, then return 0.
922  * - Events have different per-event term
923  *   E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
924  *   Return -1
925  * - Some of the event set per-event term, but some not.
926  *   E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
927  *   Return -1
928  */
929 static int perf_top__overwrite_check(struct perf_top *top)
930 {
931 	struct record_opts *opts = &top->record_opts;
932 	struct evlist *evlist = top->evlist;
933 	struct perf_evsel_config_term *term;
934 	struct list_head *config_terms;
935 	struct evsel *evsel;
936 	int set, overwrite = -1;
937 
938 	evlist__for_each_entry(evlist, evsel) {
939 		set = -1;
940 		config_terms = &evsel->config_terms;
941 		list_for_each_entry(term, config_terms, list) {
942 			if (term->type == PERF_EVSEL__CONFIG_TERM_OVERWRITE)
943 				set = term->val.overwrite ? 1 : 0;
944 		}
945 
946 		/* no term for current and previous event (likely) */
947 		if ((overwrite < 0) && (set < 0))
948 			continue;
949 
950 		/* has term for both current and previous event, compare */
951 		if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
952 			return -1;
953 
954 		/* no term for current event but has term for previous one */
955 		if ((overwrite >= 0) && (set < 0))
956 			return -1;
957 
958 		/* has term for current event */
959 		if ((overwrite < 0) && (set >= 0)) {
960 			/* if it's first event, set overwrite */
961 			if (evsel == perf_evlist__first(evlist))
962 				overwrite = set;
963 			else
964 				return -1;
965 		}
966 	}
967 
968 	if ((overwrite >= 0) && (opts->overwrite != overwrite))
969 		opts->overwrite = overwrite;
970 
971 	return 0;
972 }
973 
974 static int perf_top_overwrite_fallback(struct perf_top *top,
975 				       struct evsel *evsel)
976 {
977 	struct record_opts *opts = &top->record_opts;
978 	struct evlist *evlist = top->evlist;
979 	struct evsel *counter;
980 
981 	if (!opts->overwrite)
982 		return 0;
983 
984 	/* only fall back when first event fails */
985 	if (evsel != perf_evlist__first(evlist))
986 		return 0;
987 
988 	evlist__for_each_entry(evlist, counter)
989 		counter->core.attr.write_backward = false;
990 	opts->overwrite = false;
991 	pr_debug2("fall back to non-overwrite mode\n");
992 	return 1;
993 }
994 
995 static int perf_top__start_counters(struct perf_top *top)
996 {
997 	char msg[BUFSIZ];
998 	struct evsel *counter;
999 	struct evlist *evlist = top->evlist;
1000 	struct record_opts *opts = &top->record_opts;
1001 
1002 	if (perf_top__overwrite_check(top)) {
1003 		ui__error("perf top only support consistent per-event "
1004 			  "overwrite setting for all events\n");
1005 		goto out_err;
1006 	}
1007 
1008 	perf_evlist__config(evlist, opts, &callchain_param);
1009 
1010 	evlist__for_each_entry(evlist, counter) {
1011 try_again:
1012 		if (evsel__open(counter, top->evlist->core.cpus,
1013 				     top->evlist->core.threads) < 0) {
1014 
1015 			/*
1016 			 * Specially handle overwrite fall back.
1017 			 * Because perf top is the only tool which has
1018 			 * overwrite mode by default, support
1019 			 * both overwrite and non-overwrite mode, and
1020 			 * require consistent mode for all events.
1021 			 *
1022 			 * May move it to generic code with more tools
1023 			 * have similar attribute.
1024 			 */
1025 			if (perf_missing_features.write_backward &&
1026 			    perf_top_overwrite_fallback(top, counter))
1027 				goto try_again;
1028 
1029 			if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
1030 				if (verbose > 0)
1031 					ui__warning("%s\n", msg);
1032 				goto try_again;
1033 			}
1034 
1035 			perf_evsel__open_strerror(counter, &opts->target,
1036 						  errno, msg, sizeof(msg));
1037 			ui__error("%s\n", msg);
1038 			goto out_err;
1039 		}
1040 	}
1041 
1042 	if (perf_evlist__mmap(evlist, opts->mmap_pages) < 0) {
1043 		ui__error("Failed to mmap with %d (%s)\n",
1044 			    errno, str_error_r(errno, msg, sizeof(msg)));
1045 		goto out_err;
1046 	}
1047 
1048 	return 0;
1049 
1050 out_err:
1051 	return -1;
1052 }
1053 
1054 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1055 {
1056 	if (callchain->mode != CHAIN_NONE) {
1057 		if (callchain_register_param(callchain) < 0) {
1058 			ui__error("Can't register callchain params.\n");
1059 			return -EINVAL;
1060 		}
1061 	}
1062 
1063 	return 0;
1064 }
1065 
1066 static struct ordered_events *rotate_queues(struct perf_top *top)
1067 {
1068 	struct ordered_events *in = top->qe.in;
1069 
1070 	if (top->qe.in == &top->qe.data[1])
1071 		top->qe.in = &top->qe.data[0];
1072 	else
1073 		top->qe.in = &top->qe.data[1];
1074 
1075 	return in;
1076 }
1077 
1078 static void *process_thread(void *arg)
1079 {
1080 	struct perf_top *top = arg;
1081 
1082 	while (!done) {
1083 		struct ordered_events *out, *in = top->qe.in;
1084 
1085 		if (!in->nr_events) {
1086 			usleep(100);
1087 			continue;
1088 		}
1089 
1090 		out = rotate_queues(top);
1091 
1092 		pthread_mutex_lock(&top->qe.mutex);
1093 		top->qe.rotate = true;
1094 		pthread_cond_wait(&top->qe.cond, &top->qe.mutex);
1095 		pthread_mutex_unlock(&top->qe.mutex);
1096 
1097 		if (ordered_events__flush(out, OE_FLUSH__TOP))
1098 			pr_err("failed to process events\n");
1099 	}
1100 
1101 	return NULL;
1102 }
1103 
1104 /*
1105  * Allow only 'top->delay_secs' seconds behind samples.
1106  */
1107 static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1108 {
1109 	union perf_event *event = qevent->event;
1110 	u64 delay_timestamp;
1111 
1112 	if (event->header.type != PERF_RECORD_SAMPLE)
1113 		return false;
1114 
1115 	delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1116 	return delay_timestamp < last_timestamp;
1117 }
1118 
1119 static int deliver_event(struct ordered_events *qe,
1120 			 struct ordered_event *qevent)
1121 {
1122 	struct perf_top *top = qe->data;
1123 	struct evlist *evlist = top->evlist;
1124 	struct perf_session *session = top->session;
1125 	union perf_event *event = qevent->event;
1126 	struct perf_sample sample;
1127 	struct evsel *evsel;
1128 	struct machine *machine;
1129 	int ret = -1;
1130 
1131 	if (should_drop(qevent, top)) {
1132 		top->drop++;
1133 		top->drop_total++;
1134 		return 0;
1135 	}
1136 
1137 	ret = perf_evlist__parse_sample(evlist, event, &sample);
1138 	if (ret) {
1139 		pr_err("Can't parse sample, err = %d\n", ret);
1140 		goto next_event;
1141 	}
1142 
1143 	evsel = perf_evlist__id2evsel(session->evlist, sample.id);
1144 	assert(evsel != NULL);
1145 
1146 	if (event->header.type == PERF_RECORD_SAMPLE) {
1147 		if (evswitch__discard(&top->evswitch, evsel))
1148 			return 0;
1149 		++top->samples;
1150 	}
1151 
1152 	switch (sample.cpumode) {
1153 	case PERF_RECORD_MISC_USER:
1154 		++top->us_samples;
1155 		if (top->hide_user_symbols)
1156 			goto next_event;
1157 		machine = &session->machines.host;
1158 		break;
1159 	case PERF_RECORD_MISC_KERNEL:
1160 		++top->kernel_samples;
1161 		if (top->hide_kernel_symbols)
1162 			goto next_event;
1163 		machine = &session->machines.host;
1164 		break;
1165 	case PERF_RECORD_MISC_GUEST_KERNEL:
1166 		++top->guest_kernel_samples;
1167 		machine = perf_session__find_machine(session,
1168 						     sample.pid);
1169 		break;
1170 	case PERF_RECORD_MISC_GUEST_USER:
1171 		++top->guest_us_samples;
1172 		/*
1173 		 * TODO: we don't process guest user from host side
1174 		 * except simple counting.
1175 		 */
1176 		goto next_event;
1177 	default:
1178 		if (event->header.type == PERF_RECORD_SAMPLE)
1179 			goto next_event;
1180 		machine = &session->machines.host;
1181 		break;
1182 	}
1183 
1184 	if (event->header.type == PERF_RECORD_SAMPLE) {
1185 		perf_event__process_sample(&top->tool, event, evsel,
1186 					   &sample, machine);
1187 	} else if (event->header.type == PERF_RECORD_LOST) {
1188 		perf_top__process_lost(top, event, evsel);
1189 	} else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1190 		perf_top__process_lost_samples(top, event, evsel);
1191 	} else if (event->header.type < PERF_RECORD_MAX) {
1192 		hists__inc_nr_events(evsel__hists(evsel), event->header.type);
1193 		machine__process_event(machine, event, &sample);
1194 	} else
1195 		++session->evlist->stats.nr_unknown_events;
1196 
1197 	ret = 0;
1198 next_event:
1199 	return ret;
1200 }
1201 
1202 static void init_process_thread(struct perf_top *top)
1203 {
1204 	ordered_events__init(&top->qe.data[0], deliver_event, top);
1205 	ordered_events__init(&top->qe.data[1], deliver_event, top);
1206 	ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1207 	ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1208 	top->qe.in = &top->qe.data[0];
1209 	pthread_mutex_init(&top->qe.mutex, NULL);
1210 	pthread_cond_init(&top->qe.cond, NULL);
1211 }
1212 
1213 static int __cmd_top(struct perf_top *top)
1214 {
1215 	struct record_opts *opts = &top->record_opts;
1216 	pthread_t thread, thread_process;
1217 	int ret;
1218 
1219 	if (!top->annotation_opts.objdump_path) {
1220 		ret = perf_env__lookup_objdump(&top->session->header.env,
1221 					       &top->annotation_opts.objdump_path);
1222 		if (ret)
1223 			return ret;
1224 	}
1225 
1226 	ret = callchain_param__setup_sample_type(&callchain_param);
1227 	if (ret)
1228 		return ret;
1229 
1230 	if (perf_session__register_idle_thread(top->session) < 0)
1231 		return ret;
1232 
1233 	if (top->nr_threads_synthesize > 1)
1234 		perf_set_multithreaded();
1235 
1236 	init_process_thread(top);
1237 
1238 	if (opts->record_namespaces)
1239 		top->tool.namespace_events = true;
1240 
1241 	ret = perf_event__synthesize_bpf_events(top->session, perf_event__process,
1242 						&top->session->machines.host,
1243 						&top->record_opts);
1244 	if (ret < 0)
1245 		pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n");
1246 
1247 	machine__synthesize_threads(&top->session->machines.host, &opts->target,
1248 				    top->evlist->core.threads, false,
1249 				    top->nr_threads_synthesize);
1250 
1251 	if (top->nr_threads_synthesize > 1)
1252 		perf_set_singlethreaded();
1253 
1254 	if (perf_hpp_list.socket) {
1255 		ret = perf_env__read_cpu_topology_map(&perf_env);
1256 		if (ret < 0) {
1257 			char errbuf[BUFSIZ];
1258 			const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1259 
1260 			ui__error("Could not read the CPU topology map: %s\n", err);
1261 			return ret;
1262 		}
1263 	}
1264 
1265 	ret = perf_top__start_counters(top);
1266 	if (ret)
1267 		return ret;
1268 
1269 	top->session->evlist = top->evlist;
1270 	perf_session__set_id_hdr_size(top->session);
1271 
1272 	/*
1273 	 * When perf is starting the traced process, all the events (apart from
1274 	 * group members) have enable_on_exec=1 set, so don't spoil it by
1275 	 * prematurely enabling them.
1276 	 *
1277 	 * XXX 'top' still doesn't start workloads like record, trace, but should,
1278 	 * so leave the check here.
1279 	 */
1280         if (!target__none(&opts->target))
1281 		evlist__enable(top->evlist);
1282 
1283 	ret = -1;
1284 	if (pthread_create(&thread_process, NULL, process_thread, top)) {
1285 		ui__error("Could not create process thread.\n");
1286 		return ret;
1287 	}
1288 
1289 	if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1290 							    display_thread), top)) {
1291 		ui__error("Could not create display thread.\n");
1292 		goto out_join_thread;
1293 	}
1294 
1295 	if (top->realtime_prio) {
1296 		struct sched_param param;
1297 
1298 		param.sched_priority = top->realtime_prio;
1299 		if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1300 			ui__error("Could not set realtime priority.\n");
1301 			goto out_join;
1302 		}
1303 	}
1304 
1305 	/* Wait for a minimal set of events before starting the snapshot */
1306 	perf_evlist__poll(top->evlist, 100);
1307 
1308 	perf_top__mmap_read(top);
1309 
1310 	while (!done) {
1311 		u64 hits = top->samples;
1312 
1313 		perf_top__mmap_read(top);
1314 
1315 		if (opts->overwrite || (hits == top->samples))
1316 			ret = perf_evlist__poll(top->evlist, 100);
1317 
1318 		if (resize) {
1319 			perf_top__resize(top);
1320 			resize = 0;
1321 		}
1322 	}
1323 
1324 	ret = 0;
1325 out_join:
1326 	pthread_join(thread, NULL);
1327 out_join_thread:
1328 	pthread_cond_signal(&top->qe.cond);
1329 	pthread_join(thread_process, NULL);
1330 	return ret;
1331 }
1332 
1333 static int
1334 callchain_opt(const struct option *opt, const char *arg, int unset)
1335 {
1336 	symbol_conf.use_callchain = true;
1337 	return record_callchain_opt(opt, arg, unset);
1338 }
1339 
1340 static int
1341 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1342 {
1343 	struct callchain_param *callchain = opt->value;
1344 
1345 	callchain->enabled = !unset;
1346 	callchain->record_mode = CALLCHAIN_FP;
1347 
1348 	/*
1349 	 * --no-call-graph
1350 	 */
1351 	if (unset) {
1352 		symbol_conf.use_callchain = false;
1353 		callchain->record_mode = CALLCHAIN_NONE;
1354 		return 0;
1355 	}
1356 
1357 	return parse_callchain_top_opt(arg);
1358 }
1359 
1360 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1361 {
1362 	if (!strcmp(var, "top.call-graph")) {
1363 		var = "call-graph.record-mode";
1364 		return perf_default_config(var, value, cb);
1365 	}
1366 	if (!strcmp(var, "top.children")) {
1367 		symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1368 		return 0;
1369 	}
1370 
1371 	return 0;
1372 }
1373 
1374 static int
1375 parse_percent_limit(const struct option *opt, const char *arg,
1376 		    int unset __maybe_unused)
1377 {
1378 	struct perf_top *top = opt->value;
1379 
1380 	top->min_percent = strtof(arg, NULL);
1381 	return 0;
1382 }
1383 
1384 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1385 	"\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1386 
1387 int cmd_top(int argc, const char **argv)
1388 {
1389 	char errbuf[BUFSIZ];
1390 	struct perf_top top = {
1391 		.count_filter	     = 5,
1392 		.delay_secs	     = 2,
1393 		.record_opts = {
1394 			.mmap_pages	= UINT_MAX,
1395 			.user_freq	= UINT_MAX,
1396 			.user_interval	= ULLONG_MAX,
1397 			.freq		= 4000, /* 4 KHz */
1398 			.target		= {
1399 				.uses_mmap   = true,
1400 			},
1401 			/*
1402 			 * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1403 			 * when we pause, fix that and reenable. Probably using a
1404 			 * separate evlist with a dummy event, i.e. a non-overwrite
1405 			 * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1406 			 * stays in overwrite mode. -acme
1407 			 * */
1408 			.overwrite	= 0,
1409 			.sample_time	= true,
1410 			.sample_time_set = true,
1411 		},
1412 		.max_stack	     = sysctl__max_stack(),
1413 		.annotation_opts     = annotation__default_options,
1414 		.nr_threads_synthesize = UINT_MAX,
1415 	};
1416 	struct record_opts *opts = &top.record_opts;
1417 	struct target *target = &opts->target;
1418 	const struct option options[] = {
1419 	OPT_CALLBACK('e', "event", &top.evlist, "event",
1420 		     "event selector. use 'perf list' to list available events",
1421 		     parse_events_option),
1422 	OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1423 	OPT_STRING('p', "pid", &target->pid, "pid",
1424 		    "profile events on existing process id"),
1425 	OPT_STRING('t', "tid", &target->tid, "tid",
1426 		    "profile events on existing thread id"),
1427 	OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1428 			    "system-wide collection from all CPUs"),
1429 	OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1430 		    "list of cpus to monitor"),
1431 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1432 		   "file", "vmlinux pathname"),
1433 	OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1434 		    "don't load vmlinux even if found"),
1435 	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1436 		   "file", "kallsyms pathname"),
1437 	OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1438 		    "hide kernel symbols"),
1439 	OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1440 		     "number of mmap data pages",
1441 		     perf_evlist__parse_mmap_pages),
1442 	OPT_INTEGER('r', "realtime", &top.realtime_prio,
1443 		    "collect data with this RT SCHED_FIFO priority"),
1444 	OPT_INTEGER('d', "delay", &top.delay_secs,
1445 		    "number of seconds to delay between refreshes"),
1446 	OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1447 			    "dump the symbol table used for profiling"),
1448 	OPT_INTEGER('f', "count-filter", &top.count_filter,
1449 		    "only display functions with more events than this"),
1450 	OPT_BOOLEAN(0, "group", &opts->group,
1451 			    "put the counters into a counter group"),
1452 	OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1453 		    "child tasks do not inherit counters"),
1454 	OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1455 		    "symbol to annotate"),
1456 	OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1457 	OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1458 		     "profile at this frequency",
1459 		      record__parse_freq),
1460 	OPT_INTEGER('E', "entries", &top.print_entries,
1461 		    "display this many functions"),
1462 	OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1463 		    "hide user symbols"),
1464 	OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1465 	OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1466 	OPT_INCR('v', "verbose", &verbose,
1467 		    "be more verbose (show counter open errors, etc)"),
1468 	OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1469 		   "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1470 		   " Please refer the man page for the complete list."),
1471 	OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1472 		   "output field(s): overhead, period, sample plus all of sort keys"),
1473 	OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1474 		    "Show a column with the number of samples"),
1475 	OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1476 			   NULL, "enables call-graph recording and display",
1477 			   &callchain_opt),
1478 	OPT_CALLBACK(0, "call-graph", &callchain_param,
1479 		     "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1480 		     top_callchain_help, &parse_callchain_opt),
1481 	OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1482 		    "Accumulate callchains of children and show total overhead as well"),
1483 	OPT_INTEGER(0, "max-stack", &top.max_stack,
1484 		    "Set the maximum stack depth when parsing the callchain. "
1485 		    "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1486 	OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1487 		   "ignore callees of these functions in call graphs",
1488 		   report_parse_ignore_callees_opt),
1489 	OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1490 		    "Show a column with the sum of periods"),
1491 	OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1492 		   "only consider symbols in these dsos"),
1493 	OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1494 		   "only consider symbols in these comms"),
1495 	OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1496 		   "only consider these symbols"),
1497 	OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1498 		    "Interleave source code with assembly code (default)"),
1499 	OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1500 		    "Display raw encoding of assembly instructions (default)"),
1501 	OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1502 		    "Enable kernel symbol demangling"),
1503 	OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"),
1504 	OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path",
1505 		    "objdump binary to use for disassembly and annotations"),
1506 	OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1507 		   "Specify disassembler style (e.g. -M intel for intel syntax)"),
1508 	OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1509 	OPT_CALLBACK(0, "percent-limit", &top, "percent",
1510 		     "Don't show entries under that percent", parse_percent_limit),
1511 	OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1512 		     "How to display percentage of filtered entries", parse_filter_percentage),
1513 	OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1514 		   "width[,width...]",
1515 		   "don't try to adjust column width, use these fixed values"),
1516 	OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1517 			"per thread proc mmap processing timeout in ms"),
1518 	OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1519 		     "branch any", "sample any taken branches",
1520 		     parse_branch_stack),
1521 	OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1522 		     "branch filter mask", "branch stack filter modes",
1523 		     parse_branch_stack),
1524 	OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1525 		    "Show raw trace event output (do not use print fmt or plugins)"),
1526 	OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1527 		    "Show entries in a hierarchy"),
1528 	OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1529 		    "Use a backward ring buffer, default: no"),
1530 	OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1531 	OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1532 			"number of thread to run event synthesize"),
1533 	OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces,
1534 		    "Record namespaces events"),
1535 	OPTS_EVSWITCH(&top.evswitch),
1536 	OPT_END()
1537 	};
1538 	struct evlist *sb_evlist = NULL;
1539 	const char * const top_usage[] = {
1540 		"perf top [<options>]",
1541 		NULL
1542 	};
1543 	int status = hists__init();
1544 
1545 	if (status < 0)
1546 		return status;
1547 
1548 	top.annotation_opts.min_pcnt = 5;
1549 	top.annotation_opts.context  = 4;
1550 
1551 	top.evlist = evlist__new();
1552 	if (top.evlist == NULL)
1553 		return -ENOMEM;
1554 
1555 	status = perf_config(perf_top_config, &top);
1556 	if (status)
1557 		return status;
1558 
1559 	argc = parse_options(argc, argv, options, top_usage, 0);
1560 	if (argc)
1561 		usage_with_options(top_usage, options);
1562 
1563 	if (!top.evlist->core.nr_entries &&
1564 	    perf_evlist__add_default(top.evlist) < 0) {
1565 		pr_err("Not enough memory for event selector list\n");
1566 		goto out_delete_evlist;
1567 	}
1568 
1569 	status = evswitch__init(&top.evswitch, top.evlist, stderr);
1570 	if (status)
1571 		goto out_delete_evlist;
1572 
1573 	if (symbol_conf.report_hierarchy) {
1574 		/* disable incompatible options */
1575 		symbol_conf.event_group = false;
1576 		symbol_conf.cumulate_callchain = false;
1577 
1578 		if (field_order) {
1579 			pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1580 			parse_options_usage(top_usage, options, "fields", 0);
1581 			parse_options_usage(NULL, options, "hierarchy", 0);
1582 			goto out_delete_evlist;
1583 		}
1584 	}
1585 
1586 	if (opts->branch_stack && callchain_param.enabled)
1587 		symbol_conf.show_branchflag_count = true;
1588 
1589 	sort__mode = SORT_MODE__TOP;
1590 	/* display thread wants entries to be collapsed in a different tree */
1591 	perf_hpp_list.need_collapse = 1;
1592 
1593 	if (top.use_stdio)
1594 		use_browser = 0;
1595 	else if (top.use_tui)
1596 		use_browser = 1;
1597 
1598 	setup_browser(false);
1599 
1600 	if (setup_sorting(top.evlist) < 0) {
1601 		if (sort_order)
1602 			parse_options_usage(top_usage, options, "s", 1);
1603 		if (field_order)
1604 			parse_options_usage(sort_order ? NULL : top_usage,
1605 					    options, "fields", 0);
1606 		goto out_delete_evlist;
1607 	}
1608 
1609 	status = target__validate(target);
1610 	if (status) {
1611 		target__strerror(target, status, errbuf, BUFSIZ);
1612 		ui__warning("%s\n", errbuf);
1613 	}
1614 
1615 	status = target__parse_uid(target);
1616 	if (status) {
1617 		int saved_errno = errno;
1618 
1619 		target__strerror(target, status, errbuf, BUFSIZ);
1620 		ui__error("%s\n", errbuf);
1621 
1622 		status = -saved_errno;
1623 		goto out_delete_evlist;
1624 	}
1625 
1626 	if (target__none(target))
1627 		target->system_wide = true;
1628 
1629 	if (perf_evlist__create_maps(top.evlist, target) < 0) {
1630 		ui__error("Couldn't create thread/CPU maps: %s\n",
1631 			  errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1632 		goto out_delete_evlist;
1633 	}
1634 
1635 	if (top.delay_secs < 1)
1636 		top.delay_secs = 1;
1637 
1638 	if (record_opts__config(opts)) {
1639 		status = -EINVAL;
1640 		goto out_delete_evlist;
1641 	}
1642 
1643 	top.sym_evsel = perf_evlist__first(top.evlist);
1644 
1645 	if (!callchain_param.enabled) {
1646 		symbol_conf.cumulate_callchain = false;
1647 		perf_hpp__cancel_cumulate();
1648 	}
1649 
1650 	if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1651 		callchain_param.order = ORDER_CALLER;
1652 
1653 	status = symbol__annotation_init();
1654 	if (status < 0)
1655 		goto out_delete_evlist;
1656 
1657 	annotation_config__init();
1658 
1659 	symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1660 	status = symbol__init(NULL);
1661 	if (status < 0)
1662 		goto out_delete_evlist;
1663 
1664 	sort__setup_elide(stdout);
1665 
1666 	get_term_dimensions(&top.winsize);
1667 	if (top.print_entries == 0) {
1668 		perf_top__update_print_entries(&top);
1669 		signal(SIGWINCH, winch_sig);
1670 	}
1671 
1672 	top.session = perf_session__new(NULL, false, NULL);
1673 	if (top.session == NULL) {
1674 		status = -1;
1675 		goto out_delete_evlist;
1676 	}
1677 
1678 	if (!top.record_opts.no_bpf_event)
1679 		bpf_event__add_sb_event(&sb_evlist, &perf_env);
1680 
1681 	if (perf_evlist__start_sb_thread(sb_evlist, target)) {
1682 		pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1683 		opts->no_bpf_event = true;
1684 	}
1685 
1686 	status = __cmd_top(&top);
1687 
1688 	if (!opts->no_bpf_event)
1689 		perf_evlist__stop_sb_thread(sb_evlist);
1690 
1691 out_delete_evlist:
1692 	evlist__delete(top.evlist);
1693 	perf_session__delete(top.session);
1694 
1695 	return status;
1696 }
1697