xref: /linux/tools/perf/builtin-top.c (revision e78f70bad29c5ae1e1076698b690b15794e9b81e)
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/cgroup.h"
26 #include "util/config.h"
27 #include "util/color.h"
28 #include "util/dso.h"
29 #include "util/evlist.h"
30 #include "util/evsel.h"
31 #include "util/evsel_config.h"
32 #include "util/event.h"
33 #include "util/machine.h"
34 #include "util/map.h"
35 #include "util/mmap.h"
36 #include "util/session.h"
37 #include "util/thread.h"
38 #include "util/stat.h"
39 #include "util/symbol.h"
40 #include "util/synthetic-events.h"
41 #include "util/top.h"
42 #include "util/util.h"
43 #include <linux/rbtree.h>
44 #include <subcmd/parse-options.h>
45 #include "util/parse-events.h"
46 #include "util/callchain.h"
47 #include "util/cpumap.h"
48 #include "util/sort.h"
49 #include "util/string2.h"
50 #include "util/term.h"
51 #include "util/intlist.h"
52 #include "util/parse-branch-options.h"
53 #include "arch/common.h"
54 #include "ui/ui.h"
55 
56 #include "util/debug.h"
57 #include "util/ordered-events.h"
58 #include "util/pfm.h"
59 
60 #include <assert.h>
61 #include <elf.h>
62 #include <fcntl.h>
63 
64 #include <stdio.h>
65 #include <termios.h>
66 #include <unistd.h>
67 #include <inttypes.h>
68 
69 #include <errno.h>
70 #include <time.h>
71 #include <sched.h>
72 #include <signal.h>
73 
74 #include <sys/syscall.h>
75 #include <sys/ioctl.h>
76 #include <poll.h>
77 #include <sys/prctl.h>
78 #include <sys/wait.h>
79 #include <sys/uio.h>
80 #include <sys/utsname.h>
81 #include <sys/mman.h>
82 
83 #include <linux/stringify.h>
84 #include <linux/time64.h>
85 #include <linux/types.h>
86 #include <linux/err.h>
87 
88 #include <linux/ctype.h>
89 #include <perf/mmap.h>
90 
91 static volatile sig_atomic_t done;
92 static volatile sig_atomic_t resize;
93 
94 #define HEADER_LINE_NR  5
95 
96 static void perf_top__update_print_entries(struct perf_top *top)
97 {
98 	top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
99 }
100 
101 static void winch_sig(int sig __maybe_unused)
102 {
103 	resize = 1;
104 }
105 
106 static void perf_top__resize(struct perf_top *top)
107 {
108 	get_term_dimensions(&top->winsize);
109 	perf_top__update_print_entries(top);
110 }
111 
112 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
113 {
114 	struct evsel *evsel;
115 	struct symbol *sym;
116 	struct annotation *notes;
117 	struct map *map;
118 	struct dso *dso;
119 	int err = -1;
120 
121 	if (!he || !he->ms.sym)
122 		return -1;
123 
124 	evsel = hists_to_evsel(he->hists);
125 
126 	sym = he->ms.sym;
127 	map = he->ms.map;
128 	dso = map__dso(map);
129 
130 	/*
131 	 * We can't annotate with just /proc/kallsyms
132 	 */
133 	if (dso__symtab_type(dso) == DSO_BINARY_TYPE__KALLSYMS && !dso__is_kcore(dso)) {
134 		pr_err("Can't annotate %s: No vmlinux file was found in the "
135 		       "path\n", sym->name);
136 		sleep(1);
137 		return -1;
138 	}
139 
140 	notes = symbol__annotation(sym);
141 	annotation__lock(notes);
142 
143 	if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
144 		annotation__unlock(notes);
145 		pr_err("Not enough memory for annotating '%s' symbol!\n",
146 		       sym->name);
147 		sleep(1);
148 		return err;
149 	}
150 
151 	err = symbol__annotate(&he->ms, evsel, NULL);
152 	if (err == 0) {
153 		top->sym_filter_entry = he;
154 	} else {
155 		char msg[BUFSIZ];
156 		symbol__strerror_disassemble(&he->ms, err, msg, sizeof(msg));
157 		pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
158 	}
159 
160 	annotation__unlock(notes);
161 	return err;
162 }
163 
164 static void __zero_source_counters(struct hist_entry *he)
165 {
166 	struct symbol *sym = he->ms.sym;
167 	symbol__annotate_zero_histograms(sym);
168 }
169 
170 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
171 {
172 	struct utsname uts;
173 	int err = uname(&uts);
174 	struct dso *dso = map__dso(map);
175 
176 	ui__warning("Out of bounds address found:\n\n"
177 		    "Addr:   %" PRIx64 "\n"
178 		    "DSO:    %s %c\n"
179 		    "Map:    %" PRIx64 "-%" PRIx64 "\n"
180 		    "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
181 		    "Arch:   %s\n"
182 		    "Kernel: %s\n"
183 		    "Tools:  %s\n\n"
184 		    "Not all samples will be on the annotation output.\n\n"
185 		    "Please report to linux-kernel@vger.kernel.org\n",
186 		    ip, dso__long_name(dso), dso__symtab_origin(dso),
187 		    map__start(map), map__end(map), sym->start, sym->end,
188 		    sym->binding == STB_GLOBAL ? 'g' :
189 		    sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
190 		    err ? "[unknown]" : uts.machine,
191 		    err ? "[unknown]" : uts.release, perf_version_string);
192 	if (use_browser <= 0)
193 		sleep(5);
194 
195 	map__set_erange_warned(map);
196 }
197 
198 static void perf_top__record_precise_ip(struct perf_top *top,
199 					struct hist_entry *he,
200 					struct perf_sample *sample,
201 					struct evsel *evsel, u64 ip)
202 	EXCLUSIVE_LOCKS_REQUIRED(he->hists->lock)
203 {
204 	struct annotation *notes;
205 	struct symbol *sym = he->ms.sym;
206 	int err = 0;
207 
208 	if (sym == NULL || (use_browser == 0 &&
209 			    (top->sym_filter_entry == NULL ||
210 			     top->sym_filter_entry->ms.sym != sym)))
211 		return;
212 
213 	notes = symbol__annotation(sym);
214 
215 	if (!annotation__trylock(notes))
216 		return;
217 
218 	err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
219 
220 	annotation__unlock(notes);
221 
222 	if (unlikely(err)) {
223 		/*
224 		 * This function is now called with he->hists->lock held.
225 		 * Release it before going to sleep.
226 		 */
227 		mutex_unlock(&he->hists->lock);
228 
229 		if (err == -ERANGE && !map__erange_warned(he->ms.map))
230 			ui__warn_map_erange(he->ms.map, sym, ip);
231 		else if (err == -ENOMEM) {
232 			pr_err("Not enough memory for annotating '%s' symbol!\n",
233 			       sym->name);
234 			sleep(1);
235 		}
236 
237 		mutex_lock(&he->hists->lock);
238 	}
239 }
240 
241 static void perf_top__show_details(struct perf_top *top)
242 {
243 	struct hist_entry *he = top->sym_filter_entry;
244 	struct evsel *evsel;
245 	struct annotation *notes;
246 	struct symbol *symbol;
247 	int more;
248 
249 	if (!he)
250 		return;
251 
252 	evsel = hists_to_evsel(he->hists);
253 
254 	symbol = he->ms.sym;
255 	notes = symbol__annotation(symbol);
256 
257 	annotation__lock(notes);
258 
259 	symbol__calc_percent(symbol, evsel);
260 
261 	if (notes->src == NULL)
262 		goto out_unlock;
263 
264 	printf("Showing %s for %s\n", evsel__name(top->sym_evsel), symbol->name);
265 	printf("  Events  Pcnt (>=%d%%)\n", annotate_opts.min_pcnt);
266 
267 	more = hist_entry__annotate_printf(he, top->sym_evsel);
268 
269 	if (top->evlist->enabled) {
270 		if (top->zero)
271 			symbol__annotate_zero_histogram(symbol, top->sym_evsel);
272 		else
273 			symbol__annotate_decay_histogram(symbol, top->sym_evsel);
274 	}
275 	if (more != 0)
276 		printf("%d lines not displayed, maybe increase display entries [e]\n", more);
277 out_unlock:
278 	annotation__unlock(notes);
279 }
280 
281 static void perf_top__resort_hists(struct perf_top *t)
282 {
283 	struct evlist *evlist = t->evlist;
284 	struct evsel *pos;
285 
286 	evlist__for_each_entry(evlist, pos) {
287 		struct hists *hists = evsel__hists(pos);
288 
289 		/*
290 		 * unlink existing entries so that they can be linked
291 		 * in a correct order in hists__match() below.
292 		 */
293 		hists__unlink(hists);
294 
295 		if (evlist->enabled) {
296 			if (t->zero) {
297 				hists__delete_entries(hists);
298 			} else {
299 				hists__decay_entries(hists, t->hide_user_symbols,
300 						     t->hide_kernel_symbols);
301 			}
302 		}
303 
304 		hists__collapse_resort(hists, NULL);
305 
306 		/* Non-group events are considered as leader */
307 		if (symbol_conf.event_group && !evsel__is_group_leader(pos)) {
308 			struct hists *leader_hists = evsel__hists(evsel__leader(pos));
309 
310 			hists__match(leader_hists, hists);
311 			hists__link(leader_hists, hists);
312 		}
313 	}
314 
315 	evlist__for_each_entry(evlist, pos) {
316 		evsel__output_resort(pos, NULL);
317 	}
318 }
319 
320 static void perf_top__print_sym_table(struct perf_top *top)
321 {
322 	char bf[160];
323 	int printed = 0;
324 	const int win_width = top->winsize.ws_col - 1;
325 	struct evsel *evsel = top->sym_evsel;
326 	struct hists *hists = evsel__hists(evsel);
327 
328 	puts(CONSOLE_CLEAR);
329 
330 	perf_top__header_snprintf(top, bf, sizeof(bf));
331 	printf("%s\n", bf);
332 
333 	printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
334 
335 	if (!top->record_opts.overwrite &&
336 	    (top->evlist->stats.nr_lost_warned !=
337 	     top->evlist->stats.nr_events[PERF_RECORD_LOST])) {
338 		top->evlist->stats.nr_lost_warned =
339 			      top->evlist->stats.nr_events[PERF_RECORD_LOST];
340 		color_fprintf(stdout, PERF_COLOR_RED,
341 			      "WARNING: LOST %d chunks, Check IO/CPU overload",
342 			      top->evlist->stats.nr_lost_warned);
343 		++printed;
344 	}
345 
346 	if (top->sym_filter_entry) {
347 		perf_top__show_details(top);
348 		return;
349 	}
350 
351 	perf_top__resort_hists(top);
352 
353 	hists__output_recalc_col_len(hists, top->print_entries - printed);
354 	putchar('\n');
355 	hists__fprintf(hists, false, top->print_entries - printed, win_width,
356 		       top->min_percent, stdout, !symbol_conf.use_callchain);
357 }
358 
359 static void prompt_integer(int *target, const char *msg)
360 {
361 	char *buf = NULL, *p;
362 	size_t dummy = 0;
363 	int tmp;
364 
365 	fprintf(stdout, "\n%s: ", msg);
366 	if (getline(&buf, &dummy, stdin) < 0)
367 		return;
368 
369 	p = strchr(buf, '\n');
370 	if (p)
371 		*p = 0;
372 
373 	p = buf;
374 	while(*p) {
375 		if (!isdigit(*p))
376 			goto out_free;
377 		p++;
378 	}
379 	tmp = strtoul(buf, NULL, 10);
380 	*target = tmp;
381 out_free:
382 	free(buf);
383 }
384 
385 static void prompt_percent(int *target, const char *msg)
386 {
387 	int tmp = 0;
388 
389 	prompt_integer(&tmp, msg);
390 	if (tmp >= 0 && tmp <= 100)
391 		*target = tmp;
392 }
393 
394 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
395 {
396 	char *buf = NULL, *p;
397 	struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
398 	struct hists *hists = evsel__hists(top->sym_evsel);
399 	struct rb_node *next;
400 	size_t dummy = 0;
401 
402 	/* zero counters of active symbol */
403 	if (syme) {
404 		__zero_source_counters(syme);
405 		top->sym_filter_entry = NULL;
406 	}
407 
408 	fprintf(stdout, "\n%s: ", msg);
409 	if (getline(&buf, &dummy, stdin) < 0)
410 		goto out_free;
411 
412 	p = strchr(buf, '\n');
413 	if (p)
414 		*p = 0;
415 
416 	next = rb_first_cached(&hists->entries);
417 	while (next) {
418 		n = rb_entry(next, struct hist_entry, rb_node);
419 		if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
420 			found = n;
421 			break;
422 		}
423 		next = rb_next(&n->rb_node);
424 	}
425 
426 	if (!found) {
427 		fprintf(stderr, "Sorry, %s is not active.\n", buf);
428 		sleep(1);
429 	} else
430 		perf_top__parse_source(top, found);
431 
432 out_free:
433 	free(buf);
434 }
435 
436 static void perf_top__print_mapped_keys(struct perf_top *top)
437 {
438 	char *name = NULL;
439 
440 	if (top->sym_filter_entry) {
441 		struct symbol *sym = top->sym_filter_entry->ms.sym;
442 		name = sym->name;
443 	}
444 
445 	fprintf(stdout, "\nMapped keys:\n");
446 	fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
447 	fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
448 
449 	if (top->evlist->core.nr_entries > 1)
450 		fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", evsel__name(top->sym_evsel));
451 
452 	fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
453 
454 	fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", annotate_opts.min_pcnt);
455 	fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
456 	fprintf(stdout, "\t[S]     stop annotation.\n");
457 
458 	fprintf(stdout,
459 		"\t[K]     hide kernel symbols.             \t(%s)\n",
460 		top->hide_kernel_symbols ? "yes" : "no");
461 	fprintf(stdout,
462 		"\t[U]     hide user symbols.               \t(%s)\n",
463 		top->hide_user_symbols ? "yes" : "no");
464 	fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
465 	fprintf(stdout, "\t[qQ]    quit.\n");
466 }
467 
468 static int perf_top__key_mapped(struct perf_top *top, int c)
469 {
470 	switch (c) {
471 		case 'd':
472 		case 'e':
473 		case 'f':
474 		case 'z':
475 		case 'q':
476 		case 'Q':
477 		case 'K':
478 		case 'U':
479 		case 'F':
480 		case 's':
481 		case 'S':
482 			return 1;
483 		case 'E':
484 			return top->evlist->core.nr_entries > 1 ? 1 : 0;
485 		default:
486 			break;
487 	}
488 
489 	return 0;
490 }
491 
492 static bool perf_top__handle_keypress(struct perf_top *top, int c)
493 {
494 	bool ret = true;
495 
496 	if (!perf_top__key_mapped(top, c)) {
497 		struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
498 		struct termios save;
499 
500 		perf_top__print_mapped_keys(top);
501 		fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
502 		fflush(stdout);
503 
504 		set_term_quiet_input(&save);
505 
506 		poll(&stdin_poll, 1, -1);
507 		c = getc(stdin);
508 
509 		tcsetattr(0, TCSAFLUSH, &save);
510 		if (!perf_top__key_mapped(top, c))
511 			return ret;
512 	}
513 
514 	switch (c) {
515 		case 'd':
516 			prompt_integer(&top->delay_secs, "Enter display delay");
517 			if (top->delay_secs < 1)
518 				top->delay_secs = 1;
519 			break;
520 		case 'e':
521 			prompt_integer(&top->print_entries, "Enter display entries (lines)");
522 			if (top->print_entries == 0) {
523 				perf_top__resize(top);
524 				signal(SIGWINCH, winch_sig);
525 			} else {
526 				signal(SIGWINCH, SIG_DFL);
527 			}
528 			break;
529 		case 'E':
530 			if (top->evlist->core.nr_entries > 1) {
531 				/* Select 0 as the default event: */
532 				int counter = 0;
533 
534 				fprintf(stderr, "\nAvailable events:");
535 
536 				evlist__for_each_entry(top->evlist, top->sym_evsel)
537 					fprintf(stderr, "\n\t%d %s", top->sym_evsel->core.idx, evsel__name(top->sym_evsel));
538 
539 				prompt_integer(&counter, "Enter details event counter");
540 
541 				if (counter >= top->evlist->core.nr_entries) {
542 					top->sym_evsel = evlist__first(top->evlist);
543 					fprintf(stderr, "Sorry, no such event, using %s.\n", evsel__name(top->sym_evsel));
544 					sleep(1);
545 					break;
546 				}
547 				evlist__for_each_entry(top->evlist, top->sym_evsel)
548 					if (top->sym_evsel->core.idx == counter)
549 						break;
550 			} else
551 				top->sym_evsel = evlist__first(top->evlist);
552 			break;
553 		case 'f':
554 			prompt_integer(&top->count_filter, "Enter display event count filter");
555 			break;
556 		case 'F':
557 			prompt_percent(&annotate_opts.min_pcnt,
558 				       "Enter details display event filter (percent)");
559 			break;
560 		case 'K':
561 			top->hide_kernel_symbols = !top->hide_kernel_symbols;
562 			break;
563 		case 'q':
564 		case 'Q':
565 			printf("exiting.\n");
566 			if (top->dump_symtab)
567 				perf_session__fprintf_dsos(top->session, stderr);
568 			ret = false;
569 			break;
570 		case 's':
571 			perf_top__prompt_symbol(top, "Enter details symbol");
572 			break;
573 		case 'S':
574 			if (!top->sym_filter_entry)
575 				break;
576 			else {
577 				struct hist_entry *syme = top->sym_filter_entry;
578 
579 				top->sym_filter_entry = NULL;
580 				__zero_source_counters(syme);
581 			}
582 			break;
583 		case 'U':
584 			top->hide_user_symbols = !top->hide_user_symbols;
585 			break;
586 		case 'z':
587 			top->zero = !top->zero;
588 			break;
589 		default:
590 			break;
591 	}
592 
593 	return ret;
594 }
595 
596 static void perf_top__sort_new_samples(void *arg)
597 {
598 	struct perf_top *t = arg;
599 
600 	if (t->evlist->selected != NULL)
601 		t->sym_evsel = t->evlist->selected;
602 
603 	perf_top__resort_hists(t);
604 
605 	if (t->lost || t->drop)
606 		pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
607 }
608 
609 static void stop_top(void)
610 {
611 	session_done = 1;
612 	done = 1;
613 }
614 
615 static void *display_thread_tui(void *arg)
616 {
617 	struct evsel *pos;
618 	struct perf_top *top = arg;
619 	const char *help = "For a higher level overview, try: perf top --sort comm,dso";
620 	struct hist_browser_timer hbt = {
621 		.timer		= perf_top__sort_new_samples,
622 		.arg		= top,
623 		.refresh	= top->delay_secs,
624 	};
625 	int ret;
626 
627 	/* In order to read symbols from other namespaces perf to  needs to call
628 	 * setns(2).  This isn't permitted if the struct_fs has multiple users.
629 	 * unshare(2) the fs so that we may continue to setns into namespaces
630 	 * that we're observing.
631 	 */
632 	unshare(CLONE_FS);
633 
634 	prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
635 
636 repeat:
637 	perf_top__sort_new_samples(top);
638 
639 	/*
640 	 * Initialize the uid_filter_str, in the future the TUI will allow
641 	 * Zooming in/out UIDs. For now just use whatever the user passed
642 	 * via --uid.
643 	 */
644 	evlist__for_each_entry(top->evlist, pos) {
645 		struct hists *hists = evsel__hists(pos);
646 		hists->uid_filter_str = top->record_opts.target.uid_str;
647 	}
648 
649 	ret = evlist__tui_browse_hists(top->evlist, help, &hbt, top->min_percent,
650 				       &top->session->header.env, !top->record_opts.overwrite);
651 	if (ret == K_RELOAD) {
652 		top->zero = true;
653 		goto repeat;
654 	} else
655 		stop_top();
656 
657 	return NULL;
658 }
659 
660 static void display_sig(int sig __maybe_unused)
661 {
662 	stop_top();
663 }
664 
665 static void display_setup_sig(void)
666 {
667 	signal(SIGSEGV, sighandler_dump_stack);
668 	signal(SIGFPE, sighandler_dump_stack);
669 	signal(SIGINT,  display_sig);
670 	signal(SIGQUIT, display_sig);
671 	signal(SIGTERM, display_sig);
672 }
673 
674 static void *display_thread(void *arg)
675 {
676 	struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
677 	struct termios save;
678 	struct perf_top *top = arg;
679 	int delay_msecs, c;
680 
681 	/* In order to read symbols from other namespaces perf to  needs to call
682 	 * setns(2).  This isn't permitted if the struct_fs has multiple users.
683 	 * unshare(2) the fs so that we may continue to setns into namespaces
684 	 * that we're observing.
685 	 */
686 	unshare(CLONE_FS);
687 
688 	prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
689 
690 	display_setup_sig();
691 	pthread__unblock_sigwinch();
692 repeat:
693 	delay_msecs = top->delay_secs * MSEC_PER_SEC;
694 	set_term_quiet_input(&save);
695 	/* trash return*/
696 	clearerr(stdin);
697 	if (poll(&stdin_poll, 1, 0) > 0)
698 		getc(stdin);
699 
700 	while (!done) {
701 		perf_top__print_sym_table(top);
702 		/*
703 		 * Either timeout expired or we got an EINTR due to SIGWINCH,
704 		 * refresh screen in both cases.
705 		 */
706 		switch (poll(&stdin_poll, 1, delay_msecs)) {
707 		case 0:
708 			continue;
709 		case -1:
710 			if (errno == EINTR)
711 				continue;
712 			fallthrough;
713 		default:
714 			c = getc(stdin);
715 			tcsetattr(0, TCSAFLUSH, &save);
716 
717 			if (perf_top__handle_keypress(top, c))
718 				goto repeat;
719 			stop_top();
720 		}
721 	}
722 
723 	tcsetattr(0, TCSAFLUSH, &save);
724 	return NULL;
725 }
726 
727 static int hist_iter__top_callback(struct hist_entry_iter *iter,
728 				   struct addr_location *al, bool single,
729 				   void *arg)
730 	EXCLUSIVE_LOCKS_REQUIRED(iter->he->hists->lock)
731 {
732 	struct perf_top *top = arg;
733 	struct evsel *evsel = iter->evsel;
734 
735 	if (perf_hpp_list.sym && single)
736 		perf_top__record_precise_ip(top, iter->he, iter->sample, evsel, al->addr);
737 
738 	hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
739 			     !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY),
740 			     NULL, evsel);
741 	return 0;
742 }
743 
744 static void perf_event__process_sample(const struct perf_tool *tool,
745 				       const union perf_event *event,
746 				       struct evsel *evsel,
747 				       struct perf_sample *sample,
748 				       struct machine *machine)
749 {
750 	struct perf_top *top = container_of(tool, struct perf_top, tool);
751 	struct addr_location al;
752 
753 	if (!machine && perf_guest) {
754 		static struct intlist *seen;
755 
756 		if (!seen)
757 			seen = intlist__new(NULL);
758 
759 		if (!intlist__has_entry(seen, sample->pid)) {
760 			pr_err("Can't find guest [%d]'s kernel information\n",
761 				sample->pid);
762 			intlist__add(seen, sample->pid);
763 		}
764 		return;
765 	}
766 
767 	if (!machine) {
768 		pr_err("%u unprocessable samples recorded.\r",
769 		       top->session->evlist->stats.nr_unprocessable_samples++);
770 		return;
771 	}
772 
773 	if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
774 		top->exact_samples++;
775 
776 	addr_location__init(&al);
777 	if (machine__resolve(machine, &al, sample) < 0)
778 		goto out;
779 
780 	if (top->stitch_lbr)
781 		thread__set_lbr_stitch_enable(al.thread, true);
782 
783 	if (!machine->kptr_restrict_warned &&
784 	    symbol_conf.kptr_restrict &&
785 	    al.cpumode == PERF_RECORD_MISC_KERNEL) {
786 		if (!evlist__exclude_kernel(top->session->evlist)) {
787 			ui__warning(
788 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
789 "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
790 "Kernel%s samples will not be resolved.\n",
791 			  al.map && map__has_symbols(al.map) ?
792 			  " modules" : "");
793 			if (use_browser <= 0)
794 				sleep(5);
795 		}
796 		machine->kptr_restrict_warned = true;
797 	}
798 
799 	if (al.sym == NULL && al.map != NULL) {
800 		const char *msg = "Kernel samples will not be resolved.\n";
801 		/*
802 		 * As we do lazy loading of symtabs we only will know if the
803 		 * specified vmlinux file is invalid when we actually have a
804 		 * hit in kernel space and then try to load it. So if we get
805 		 * here and there are _no_ symbols in the DSO backing the
806 		 * kernel map, bail out.
807 		 *
808 		 * We may never get here, for instance, if we use -K/
809 		 * --hide-kernel-symbols, even if the user specifies an
810 		 * invalid --vmlinux ;-)
811 		 */
812 		if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
813 		    __map__is_kernel(al.map) && !map__has_symbols(al.map)) {
814 			if (symbol_conf.vmlinux_name) {
815 				char serr[256];
816 
817 				dso__strerror_load(map__dso(al.map), serr, sizeof(serr));
818 				ui__warning("The %s file can't be used: %s\n%s",
819 					    symbol_conf.vmlinux_name, serr, msg);
820 			} else {
821 				ui__warning("A vmlinux file was not found.\n%s",
822 					    msg);
823 			}
824 
825 			if (use_browser <= 0)
826 				sleep(5);
827 			top->vmlinux_warned = true;
828 		}
829 	}
830 
831 	if (al.sym == NULL || !al.sym->idle) {
832 		struct hists *hists = evsel__hists(evsel);
833 		struct hist_entry_iter iter = {
834 			.evsel		= evsel,
835 			.sample 	= sample,
836 			.add_entry_cb 	= hist_iter__top_callback,
837 		};
838 
839 		if (symbol_conf.cumulate_callchain)
840 			iter.ops = &hist_iter_cumulative;
841 		else
842 			iter.ops = &hist_iter_normal;
843 
844 		mutex_lock(&hists->lock);
845 
846 		if (hist_entry_iter__add(&iter, &al, top->max_stack, top) < 0)
847 			pr_err("Problem incrementing symbol period, skipping event\n");
848 
849 		mutex_unlock(&hists->lock);
850 	}
851 
852 out:
853 	addr_location__exit(&al);
854 }
855 
856 static void
857 perf_top__process_lost(struct perf_top *top, union perf_event *event,
858 		       struct evsel *evsel)
859 {
860 	top->lost += event->lost.lost;
861 	top->lost_total += event->lost.lost;
862 	evsel->evlist->stats.total_lost += event->lost.lost;
863 }
864 
865 static void
866 perf_top__process_lost_samples(struct perf_top *top,
867 			       union perf_event *event,
868 			       struct evsel *evsel)
869 {
870 	top->lost += event->lost_samples.lost;
871 	top->lost_total += event->lost_samples.lost;
872 	evsel->evlist->stats.total_lost_samples += event->lost_samples.lost;
873 }
874 
875 static u64 last_timestamp;
876 
877 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
878 {
879 	struct record_opts *opts = &top->record_opts;
880 	struct evlist *evlist = top->evlist;
881 	struct mmap *md;
882 	union perf_event *event;
883 
884 	md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
885 	if (perf_mmap__read_init(&md->core) < 0)
886 		return;
887 
888 	while ((event = perf_mmap__read_event(&md->core)) != NULL) {
889 		int ret;
890 
891 		ret = evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
892 		if (ret && ret != -1)
893 			break;
894 
895 		ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0, NULL);
896 		if (ret)
897 			break;
898 
899 		perf_mmap__consume(&md->core);
900 
901 		if (top->qe.rotate) {
902 			mutex_lock(&top->qe.mutex);
903 			top->qe.rotate = false;
904 			cond_signal(&top->qe.cond);
905 			mutex_unlock(&top->qe.mutex);
906 		}
907 	}
908 
909 	perf_mmap__read_done(&md->core);
910 }
911 
912 static void perf_top__mmap_read(struct perf_top *top)
913 {
914 	bool overwrite = top->record_opts.overwrite;
915 	struct evlist *evlist = top->evlist;
916 	int i;
917 
918 	if (overwrite)
919 		evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
920 
921 	for (i = 0; i < top->evlist->core.nr_mmaps; i++)
922 		perf_top__mmap_read_idx(top, i);
923 
924 	if (overwrite) {
925 		evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
926 		evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
927 	}
928 }
929 
930 /*
931  * Check per-event overwrite term.
932  * perf top should support consistent term for all events.
933  * - All events don't have per-event term
934  *   E.g. "cpu/cpu-cycles/,cpu/instructions/"
935  *   Nothing change, return 0.
936  * - All events have same per-event term
937  *   E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
938  *   Using the per-event setting to replace the opts->overwrite if
939  *   they are different, then return 0.
940  * - Events have different per-event term
941  *   E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
942  *   Return -1
943  * - Some of the event set per-event term, but some not.
944  *   E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
945  *   Return -1
946  */
947 static int perf_top__overwrite_check(struct perf_top *top)
948 {
949 	struct record_opts *opts = &top->record_opts;
950 	struct evlist *evlist = top->evlist;
951 	struct evsel_config_term *term;
952 	struct list_head *config_terms;
953 	struct evsel *evsel;
954 	int set, overwrite = -1;
955 
956 	evlist__for_each_entry(evlist, evsel) {
957 		set = -1;
958 		config_terms = &evsel->config_terms;
959 		list_for_each_entry(term, config_terms, list) {
960 			if (term->type == EVSEL__CONFIG_TERM_OVERWRITE)
961 				set = term->val.overwrite ? 1 : 0;
962 		}
963 
964 		/* no term for current and previous event (likely) */
965 		if ((overwrite < 0) && (set < 0))
966 			continue;
967 
968 		/* has term for both current and previous event, compare */
969 		if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
970 			return -1;
971 
972 		/* no term for current event but has term for previous one */
973 		if ((overwrite >= 0) && (set < 0))
974 			return -1;
975 
976 		/* has term for current event */
977 		if ((overwrite < 0) && (set >= 0)) {
978 			/* if it's first event, set overwrite */
979 			if (evsel == evlist__first(evlist))
980 				overwrite = set;
981 			else
982 				return -1;
983 		}
984 	}
985 
986 	if ((overwrite >= 0) && (opts->overwrite != overwrite))
987 		opts->overwrite = overwrite;
988 
989 	return 0;
990 }
991 
992 static int perf_top_overwrite_fallback(struct perf_top *top,
993 				       struct evsel *evsel)
994 {
995 	struct record_opts *opts = &top->record_opts;
996 	struct evlist *evlist = top->evlist;
997 	struct evsel *counter;
998 
999 	if (!opts->overwrite)
1000 		return 0;
1001 
1002 	/* only fall back when first event fails */
1003 	if (evsel != evlist__first(evlist))
1004 		return 0;
1005 
1006 	evlist__for_each_entry(evlist, counter)
1007 		counter->core.attr.write_backward = false;
1008 	opts->overwrite = false;
1009 	pr_debug2("fall back to non-overwrite mode\n");
1010 	return 1;
1011 }
1012 
1013 static int perf_top__start_counters(struct perf_top *top)
1014 {
1015 	char msg[BUFSIZ];
1016 	struct evsel *counter;
1017 	struct evlist *evlist = top->evlist;
1018 	struct record_opts *opts = &top->record_opts;
1019 
1020 	if (perf_top__overwrite_check(top)) {
1021 		ui__error("perf top only support consistent per-event "
1022 			  "overwrite setting for all events\n");
1023 		goto out_err;
1024 	}
1025 
1026 	evlist__config(evlist, opts, &callchain_param);
1027 
1028 	evlist__for_each_entry(evlist, counter) {
1029 try_again:
1030 		if (evsel__open(counter, counter->core.cpus,
1031 				counter->core.threads) < 0) {
1032 
1033 			/*
1034 			 * Specially handle overwrite fall back.
1035 			 * Because perf top is the only tool which has
1036 			 * overwrite mode by default, support
1037 			 * both overwrite and non-overwrite mode, and
1038 			 * require consistent mode for all events.
1039 			 *
1040 			 * May move it to generic code with more tools
1041 			 * have similar attribute.
1042 			 */
1043 			if (perf_missing_features.write_backward &&
1044 			    perf_top_overwrite_fallback(top, counter))
1045 				goto try_again;
1046 
1047 			if (evsel__fallback(counter, &opts->target, errno, msg, sizeof(msg))) {
1048 				if (verbose > 0)
1049 					ui__warning("%s\n", msg);
1050 				goto try_again;
1051 			}
1052 
1053 			evsel__open_strerror(counter, &opts->target, errno, msg, sizeof(msg));
1054 			ui__error("%s\n", msg);
1055 			goto out_err;
1056 		}
1057 	}
1058 
1059 	if (evlist__apply_filters(evlist, &counter, &opts->target)) {
1060 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
1061 			counter->filter ?: "BPF", evsel__name(counter), errno,
1062 			str_error_r(errno, msg, sizeof(msg)));
1063 		goto out_err;
1064 	}
1065 
1066 	if (evlist__mmap(evlist, opts->mmap_pages) < 0) {
1067 		ui__error("Failed to mmap with %d (%s)\n",
1068 			    errno, str_error_r(errno, msg, sizeof(msg)));
1069 		goto out_err;
1070 	}
1071 
1072 	return 0;
1073 
1074 out_err:
1075 	return -1;
1076 }
1077 
1078 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1079 {
1080 	if (callchain->mode != CHAIN_NONE) {
1081 		if (callchain_register_param(callchain) < 0) {
1082 			ui__error("Can't register callchain params.\n");
1083 			return -EINVAL;
1084 		}
1085 	}
1086 
1087 	return 0;
1088 }
1089 
1090 static struct ordered_events *rotate_queues(struct perf_top *top)
1091 {
1092 	struct ordered_events *in = top->qe.in;
1093 
1094 	if (top->qe.in == &top->qe.data[1])
1095 		top->qe.in = &top->qe.data[0];
1096 	else
1097 		top->qe.in = &top->qe.data[1];
1098 
1099 	return in;
1100 }
1101 
1102 static void *process_thread(void *arg)
1103 {
1104 	struct perf_top *top = arg;
1105 
1106 	while (!done) {
1107 		struct ordered_events *out, *in = top->qe.in;
1108 
1109 		if (!in->nr_events) {
1110 			usleep(100);
1111 			continue;
1112 		}
1113 
1114 		out = rotate_queues(top);
1115 
1116 		mutex_lock(&top->qe.mutex);
1117 		top->qe.rotate = true;
1118 		cond_wait(&top->qe.cond, &top->qe.mutex);
1119 		mutex_unlock(&top->qe.mutex);
1120 
1121 		if (ordered_events__flush(out, OE_FLUSH__TOP))
1122 			pr_err("failed to process events\n");
1123 	}
1124 
1125 	return NULL;
1126 }
1127 
1128 /*
1129  * Allow only 'top->delay_secs' seconds behind samples.
1130  */
1131 static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1132 {
1133 	union perf_event *event = qevent->event;
1134 	u64 delay_timestamp;
1135 
1136 	if (event->header.type != PERF_RECORD_SAMPLE)
1137 		return false;
1138 
1139 	delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1140 	return delay_timestamp < last_timestamp;
1141 }
1142 
1143 static int deliver_event(struct ordered_events *qe,
1144 			 struct ordered_event *qevent)
1145 {
1146 	struct perf_top *top = qe->data;
1147 	struct evlist *evlist = top->evlist;
1148 	struct perf_session *session = top->session;
1149 	union perf_event *event = qevent->event;
1150 	struct perf_sample sample;
1151 	struct evsel *evsel;
1152 	struct machine *machine;
1153 	int ret = -1;
1154 
1155 	if (should_drop(qevent, top)) {
1156 		top->drop++;
1157 		top->drop_total++;
1158 		return 0;
1159 	}
1160 
1161 	perf_sample__init(&sample, /*all=*/false);
1162 	ret = evlist__parse_sample(evlist, event, &sample);
1163 	if (ret) {
1164 		pr_err("Can't parse sample, err = %d\n", ret);
1165 		goto next_event;
1166 	}
1167 
1168 	evsel = evlist__id2evsel(session->evlist, sample.id);
1169 	assert(evsel != NULL);
1170 
1171 	if (event->header.type == PERF_RECORD_SAMPLE) {
1172 		if (evswitch__discard(&top->evswitch, evsel)) {
1173 			ret = 0;
1174 			goto next_event;
1175 		}
1176 		++top->samples;
1177 	}
1178 
1179 	switch (sample.cpumode) {
1180 	case PERF_RECORD_MISC_USER:
1181 		++top->us_samples;
1182 		if (top->hide_user_symbols)
1183 			goto next_event;
1184 		machine = &session->machines.host;
1185 		break;
1186 	case PERF_RECORD_MISC_KERNEL:
1187 		++top->kernel_samples;
1188 		if (top->hide_kernel_symbols)
1189 			goto next_event;
1190 		machine = &session->machines.host;
1191 		break;
1192 	case PERF_RECORD_MISC_GUEST_KERNEL:
1193 		++top->guest_kernel_samples;
1194 		machine = perf_session__find_machine(session,
1195 						     sample.pid);
1196 		break;
1197 	case PERF_RECORD_MISC_GUEST_USER:
1198 		++top->guest_us_samples;
1199 		/*
1200 		 * TODO: we don't process guest user from host side
1201 		 * except simple counting.
1202 		 */
1203 		goto next_event;
1204 	default:
1205 		if (event->header.type == PERF_RECORD_SAMPLE)
1206 			goto next_event;
1207 		machine = &session->machines.host;
1208 		break;
1209 	}
1210 
1211 	if (event->header.type == PERF_RECORD_SAMPLE) {
1212 		perf_event__process_sample(&top->tool, event, evsel,
1213 					   &sample, machine);
1214 	} else if (event->header.type == PERF_RECORD_LOST) {
1215 		perf_top__process_lost(top, event, evsel);
1216 	} else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1217 		perf_top__process_lost_samples(top, event, evsel);
1218 	} else if (event->header.type < PERF_RECORD_MAX) {
1219 		events_stats__inc(&session->evlist->stats, event->header.type);
1220 		machine__process_event(machine, event, &sample);
1221 	} else
1222 		++session->evlist->stats.nr_unknown_events;
1223 
1224 	ret = 0;
1225 next_event:
1226 	perf_sample__exit(&sample);
1227 	return ret;
1228 }
1229 
1230 static void init_process_thread(struct perf_top *top)
1231 {
1232 	ordered_events__init(&top->qe.data[0], deliver_event, top);
1233 	ordered_events__init(&top->qe.data[1], deliver_event, top);
1234 	ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1235 	ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1236 	top->qe.in = &top->qe.data[0];
1237 	mutex_init(&top->qe.mutex);
1238 	cond_init(&top->qe.cond);
1239 }
1240 
1241 static void exit_process_thread(struct perf_top *top)
1242 {
1243 	ordered_events__free(&top->qe.data[0]);
1244 	ordered_events__free(&top->qe.data[1]);
1245 	mutex_destroy(&top->qe.mutex);
1246 	cond_destroy(&top->qe.cond);
1247 }
1248 
1249 static int __cmd_top(struct perf_top *top)
1250 {
1251 	struct record_opts *opts = &top->record_opts;
1252 	pthread_t thread, thread_process;
1253 	int ret;
1254 
1255 	if (!annotate_opts.objdump_path) {
1256 		ret = perf_env__lookup_objdump(&top->session->header.env,
1257 					       &annotate_opts.objdump_path);
1258 		if (ret)
1259 			return ret;
1260 	}
1261 
1262 	ret = callchain_param__setup_sample_type(&callchain_param);
1263 	if (ret)
1264 		return ret;
1265 
1266 	if (perf_session__register_idle_thread(top->session) < 0)
1267 		return ret;
1268 
1269 	if (top->nr_threads_synthesize > 1)
1270 		perf_set_multithreaded();
1271 
1272 	init_process_thread(top);
1273 
1274 	if (opts->record_namespaces)
1275 		top->tool.namespace_events = true;
1276 	if (opts->record_cgroup) {
1277 #ifdef HAVE_FILE_HANDLE
1278 		top->tool.cgroup_events = true;
1279 #else
1280 		pr_err("cgroup tracking is not supported.\n");
1281 		return -1;
1282 #endif
1283 	}
1284 
1285 	ret = perf_event__synthesize_bpf_events(top->session, perf_event__process,
1286 						&top->session->machines.host,
1287 						&top->record_opts);
1288 	if (ret < 0)
1289 		pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n");
1290 
1291 	ret = perf_event__synthesize_cgroups(&top->tool, perf_event__process,
1292 					     &top->session->machines.host);
1293 	if (ret < 0)
1294 		pr_debug("Couldn't synthesize cgroup events.\n");
1295 
1296 	machine__synthesize_threads(&top->session->machines.host, &opts->target,
1297 				    top->evlist->core.threads, true, false,
1298 				    top->nr_threads_synthesize);
1299 
1300 	perf_set_multithreaded();
1301 
1302 	if (perf_hpp_list.socket) {
1303 		ret = perf_env__read_cpu_topology_map(&perf_env);
1304 		if (ret < 0) {
1305 			char errbuf[BUFSIZ];
1306 			const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1307 
1308 			ui__error("Could not read the CPU topology map: %s\n", err);
1309 			return ret;
1310 		}
1311 	}
1312 
1313 	/*
1314 	 * Use global stat_config that is zero meaning aggr_mode is AGGR_NONE
1315 	 * and hybrid_merge is false.
1316 	 */
1317 	evlist__uniquify_evsel_names(top->evlist, &stat_config);
1318 	ret = perf_top__start_counters(top);
1319 	if (ret)
1320 		return ret;
1321 
1322 	top->session->evlist = top->evlist;
1323 	perf_session__set_id_hdr_size(top->session);
1324 
1325 	/*
1326 	 * When perf is starting the traced process, all the events (apart from
1327 	 * group members) have enable_on_exec=1 set, so don't spoil it by
1328 	 * prematurely enabling them.
1329 	 *
1330 	 * XXX 'top' still doesn't start workloads like record, trace, but should,
1331 	 * so leave the check here.
1332 	 */
1333         if (!target__none(&opts->target))
1334 		evlist__enable(top->evlist);
1335 
1336 	ret = -1;
1337 	if (pthread_create(&thread_process, NULL, process_thread, top)) {
1338 		ui__error("Could not create process thread.\n");
1339 		return ret;
1340 	}
1341 
1342 	if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1343 							    display_thread), top)) {
1344 		ui__error("Could not create display thread.\n");
1345 		goto out_join_thread;
1346 	}
1347 
1348 	if (top->realtime_prio) {
1349 		struct sched_param param;
1350 
1351 		param.sched_priority = top->realtime_prio;
1352 		if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1353 			ui__error("Could not set realtime priority.\n");
1354 			goto out_join;
1355 		}
1356 	}
1357 
1358 	/* Wait for a minimal set of events before starting the snapshot */
1359 	evlist__poll(top->evlist, 100);
1360 
1361 	perf_top__mmap_read(top);
1362 
1363 	while (!done) {
1364 		u64 hits = top->samples;
1365 
1366 		perf_top__mmap_read(top);
1367 
1368 		if (opts->overwrite || (hits == top->samples))
1369 			ret = evlist__poll(top->evlist, 100);
1370 
1371 		if (resize) {
1372 			perf_top__resize(top);
1373 			resize = 0;
1374 		}
1375 	}
1376 
1377 	ret = 0;
1378 out_join:
1379 	pthread_join(thread, NULL);
1380 out_join_thread:
1381 	cond_signal(&top->qe.cond);
1382 	pthread_join(thread_process, NULL);
1383 	perf_set_singlethreaded();
1384 	exit_process_thread(top);
1385 	return ret;
1386 }
1387 
1388 static int
1389 callchain_opt(const struct option *opt, const char *arg, int unset)
1390 {
1391 	symbol_conf.use_callchain = true;
1392 	return record_callchain_opt(opt, arg, unset);
1393 }
1394 
1395 static int
1396 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1397 {
1398 	struct callchain_param *callchain = opt->value;
1399 
1400 	callchain->enabled = !unset;
1401 	callchain->record_mode = CALLCHAIN_FP;
1402 
1403 	/*
1404 	 * --no-call-graph
1405 	 */
1406 	if (unset) {
1407 		symbol_conf.use_callchain = false;
1408 		callchain->record_mode = CALLCHAIN_NONE;
1409 		return 0;
1410 	}
1411 
1412 	return parse_callchain_top_opt(arg);
1413 }
1414 
1415 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1416 {
1417 	if (!strcmp(var, "top.call-graph")) {
1418 		var = "call-graph.record-mode";
1419 		return perf_default_config(var, value, cb);
1420 	}
1421 	if (!strcmp(var, "top.children")) {
1422 		symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1423 		return 0;
1424 	}
1425 
1426 	return 0;
1427 }
1428 
1429 static int
1430 parse_percent_limit(const struct option *opt, const char *arg,
1431 		    int unset __maybe_unused)
1432 {
1433 	struct perf_top *top = opt->value;
1434 
1435 	top->min_percent = strtof(arg, NULL);
1436 	return 0;
1437 }
1438 
1439 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1440 	"\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1441 
1442 int cmd_top(int argc, const char **argv)
1443 {
1444 	char errbuf[BUFSIZ];
1445 	struct perf_top top = {
1446 		.count_filter	     = 5,
1447 		.delay_secs	     = 2,
1448 		.record_opts = {
1449 			.mmap_pages	= UINT_MAX,
1450 			.user_freq	= UINT_MAX,
1451 			.user_interval	= ULLONG_MAX,
1452 			.freq		= 4000, /* 4 KHz */
1453 			.target		= {
1454 				.uses_mmap   = true,
1455 			},
1456 			/*
1457 			 * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1458 			 * when we pause, fix that and reenable. Probably using a
1459 			 * separate evlist with a dummy event, i.e. a non-overwrite
1460 			 * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1461 			 * stays in overwrite mode. -acme
1462 			 * */
1463 			.overwrite	= 0,
1464 			.sample_time	= true,
1465 			.sample_time_set = true,
1466 		},
1467 		.max_stack	     = sysctl__max_stack(),
1468 		.nr_threads_synthesize = UINT_MAX,
1469 	};
1470 	struct parse_events_option_args parse_events_option_args = {
1471 		.evlistp = &top.evlist,
1472 	};
1473 	bool branch_call_mode = false;
1474 	struct record_opts *opts = &top.record_opts;
1475 	struct target *target = &opts->target;
1476 	const char *disassembler_style = NULL, *objdump_path = NULL, *addr2line_path = NULL;
1477 	const struct option options[] = {
1478 	OPT_CALLBACK('e', "event", &parse_events_option_args, "event",
1479 		     "event selector. use 'perf list' to list available events",
1480 		     parse_events_option),
1481 	OPT_CALLBACK(0, "filter", &top.evlist, "filter",
1482 		     "event filter", parse_filter),
1483 	OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1484 	OPT_STRING('p', "pid", &target->pid, "pid",
1485 		    "profile events on existing process id"),
1486 	OPT_STRING('t', "tid", &target->tid, "tid",
1487 		    "profile events on existing thread id"),
1488 	OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1489 			    "system-wide collection from all CPUs"),
1490 	OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1491 		    "list of cpus to monitor"),
1492 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1493 		   "file", "vmlinux pathname"),
1494 	OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1495 		    "don't load vmlinux even if found"),
1496 	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1497 		   "file", "kallsyms pathname"),
1498 	OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1499 		    "hide kernel symbols"),
1500 	OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1501 		     "number of mmap data pages", evlist__parse_mmap_pages),
1502 	OPT_INTEGER('r', "realtime", &top.realtime_prio,
1503 		    "collect data with this RT SCHED_FIFO priority"),
1504 	OPT_INTEGER('d', "delay", &top.delay_secs,
1505 		    "number of seconds to delay between refreshes"),
1506 	OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1507 			    "dump the symbol table used for profiling"),
1508 	OPT_INTEGER('f', "count-filter", &top.count_filter,
1509 		    "only display functions with more events than this"),
1510 	OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1511 		    "child tasks do not inherit counters"),
1512 	OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1513 		    "symbol to annotate"),
1514 	OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1515 	OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1516 		     "profile at this frequency",
1517 		      record__parse_freq),
1518 	OPT_INTEGER('E', "entries", &top.print_entries,
1519 		    "display this many functions"),
1520 	OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1521 		    "hide user symbols"),
1522 #ifdef HAVE_SLANG_SUPPORT
1523 	OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1524 #endif
1525 	OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1526 	OPT_INCR('v', "verbose", &verbose,
1527 		    "be more verbose (show counter open errors, etc)"),
1528 	OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1529 		   "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1530 		   " Please refer the man page for the complete list."),
1531 	OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1532 		   "output field(s): overhead, period, sample plus all of sort keys"),
1533 	OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1534 		    "Show a column with the number of samples"),
1535 	OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1536 			   NULL, "enables call-graph recording and display",
1537 			   &callchain_opt),
1538 	OPT_CALLBACK(0, "call-graph", &callchain_param,
1539 		     "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1540 		     top_callchain_help, &parse_callchain_opt),
1541 	OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1542 		    "Accumulate callchains of children and show total overhead as well"),
1543 	OPT_INTEGER(0, "max-stack", &top.max_stack,
1544 		    "Set the maximum stack depth when parsing the callchain. "
1545 		    "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1546 	OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1547 		   "ignore callees of these functions in call graphs",
1548 		   report_parse_ignore_callees_opt),
1549 	OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1550 		    "Show a column with the sum of periods"),
1551 	OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1552 		   "only consider symbols in these dsos"),
1553 	OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1554 		   "only consider symbols in these comms"),
1555 	OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1556 		   "only consider these symbols"),
1557 	OPT_BOOLEAN(0, "source", &annotate_opts.annotate_src,
1558 		    "Interleave source code with assembly code (default)"),
1559 	OPT_BOOLEAN(0, "asm-raw", &annotate_opts.show_asm_raw,
1560 		    "Display raw encoding of assembly instructions (default)"),
1561 	OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1562 		    "Enable kernel symbol demangling"),
1563 	OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"),
1564 	OPT_STRING(0, "objdump", &objdump_path, "path",
1565 		    "objdump binary to use for disassembly and annotations"),
1566 	OPT_STRING(0, "addr2line", &addr2line_path, "path",
1567 		   "addr2line binary to use for line numbers"),
1568 	OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1569 		   "Specify disassembler style (e.g. -M intel for intel syntax)"),
1570 	OPT_STRING(0, "prefix", &annotate_opts.prefix, "prefix",
1571 		    "Add prefix to source file path names in programs (with --prefix-strip)"),
1572 	OPT_STRING(0, "prefix-strip", &annotate_opts.prefix_strip, "N",
1573 		    "Strip first N entries of source file path name in programs (with --prefix)"),
1574 	OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1575 	OPT_CALLBACK(0, "percent-limit", &top, "percent",
1576 		     "Don't show entries under that percent", parse_percent_limit),
1577 	OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1578 		     "How to display percentage of filtered entries", parse_filter_percentage),
1579 	OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1580 		   "width[,width...]",
1581 		   "don't try to adjust column width, use these fixed values"),
1582 	OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1583 			"per thread proc mmap processing timeout in ms"),
1584 	OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1585 		     "branch any", "sample any taken branches",
1586 		     parse_branch_stack),
1587 	OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1588 		     "branch filter mask", "branch stack filter modes",
1589 		     parse_branch_stack),
1590 	OPT_BOOLEAN(0, "branch-history", &branch_call_mode,
1591 		    "add last branch records to call history"),
1592 	OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1593 		    "Show raw trace event output (do not use print fmt or plugins)"),
1594 	OPT_BOOLEAN('H', "hierarchy", &symbol_conf.report_hierarchy,
1595 		    "Show entries in a hierarchy"),
1596 	OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1597 		    "Use a backward ring buffer, default: no"),
1598 	OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1599 	OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1600 			"number of thread to run event synthesize"),
1601 	OPT_CALLBACK('G', "cgroup", &top.evlist, "name",
1602 		     "monitor event in cgroup name only", parse_cgroups),
1603 	OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces,
1604 		    "Record namespaces events"),
1605 	OPT_BOOLEAN(0, "all-cgroups", &opts->record_cgroup,
1606 		    "Record cgroup events"),
1607 	OPT_INTEGER(0, "group-sort-idx", &symbol_conf.group_sort_idx,
1608 		    "Sort the output by the event at the index n in group. "
1609 		    "If n is invalid, sort by the first event. "
1610 		    "WARNING: should be used on grouped events."),
1611 	OPT_BOOLEAN(0, "stitch-lbr", &top.stitch_lbr,
1612 		    "Enable LBR callgraph stitching approach"),
1613 #ifdef HAVE_LIBPFM
1614 	OPT_CALLBACK(0, "pfm-events", &top.evlist, "event",
1615 		"libpfm4 event selector. use 'perf list' to list available events",
1616 		parse_libpfm_events_option),
1617 #endif
1618 	OPTS_EVSWITCH(&top.evswitch),
1619 	OPT_END()
1620 	};
1621 	const char * const top_usage[] = {
1622 		"perf top [<options>]",
1623 		NULL
1624 	};
1625 	int status = hists__init();
1626 
1627 	if (status < 0)
1628 		return status;
1629 
1630 	annotation_options__init();
1631 
1632 	annotate_opts.min_pcnt = 5;
1633 	annotate_opts.context  = 4;
1634 
1635 	top.evlist = evlist__new();
1636 	if (top.evlist == NULL)
1637 		return -ENOMEM;
1638 
1639 	status = perf_config(perf_top_config, &top);
1640 	if (status)
1641 		return status;
1642 	/*
1643 	 * Since the per arch annotation init routine may need the cpuid, read
1644 	 * it here, since we are not getting this from the perf.data header.
1645 	 */
1646 	status = perf_env__read_cpuid(&perf_env);
1647 	if (status) {
1648 		/*
1649 		 * Some arches do not provide a get_cpuid(), so just use pr_debug, otherwise
1650 		 * warn the user explicitly.
1651 		 */
1652 		eprintf(status == ENOSYS ? 1 : 0, verbose,
1653 			"Couldn't read the cpuid for this machine: %s\n",
1654 			str_error_r(errno, errbuf, sizeof(errbuf)));
1655 	}
1656 	top.evlist->env = &perf_env;
1657 
1658 	argc = parse_options(argc, argv, options, top_usage, 0);
1659 	if (argc)
1660 		usage_with_options(top_usage, options);
1661 
1662 	if (disassembler_style) {
1663 		annotate_opts.disassembler_style = strdup(disassembler_style);
1664 		if (!annotate_opts.disassembler_style)
1665 			return -ENOMEM;
1666 	}
1667 	if (objdump_path) {
1668 		annotate_opts.objdump_path = strdup(objdump_path);
1669 		if (!annotate_opts.objdump_path)
1670 			return -ENOMEM;
1671 	}
1672 	if (addr2line_path) {
1673 		symbol_conf.addr2line_path = strdup(addr2line_path);
1674 		if (!symbol_conf.addr2line_path)
1675 			return -ENOMEM;
1676 	}
1677 
1678 	status = symbol__validate_sym_arguments();
1679 	if (status)
1680 		goto out_delete_evlist;
1681 
1682 	if (annotate_check_args() < 0)
1683 		goto out_delete_evlist;
1684 
1685 	if (!top.evlist->core.nr_entries) {
1686 		bool can_profile_kernel = perf_event_paranoid_check(1);
1687 		int err = parse_event(top.evlist, can_profile_kernel ? "cycles:P" : "cycles:Pu");
1688 
1689 		if (err)
1690 			goto out_delete_evlist;
1691 	}
1692 
1693 	status = evswitch__init(&top.evswitch, top.evlist, stderr);
1694 	if (status)
1695 		goto out_delete_evlist;
1696 
1697 	if (symbol_conf.report_hierarchy) {
1698 		/* disable incompatible options */
1699 		symbol_conf.event_group = false;
1700 		symbol_conf.cumulate_callchain = false;
1701 
1702 		if (field_order) {
1703 			pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1704 			parse_options_usage(top_usage, options, "fields", 0);
1705 			parse_options_usage(NULL, options, "hierarchy", 0);
1706 			goto out_delete_evlist;
1707 		}
1708 	}
1709 
1710 	if (top.stitch_lbr && !(callchain_param.record_mode == CALLCHAIN_LBR)) {
1711 		pr_err("Error: --stitch-lbr must be used with --call-graph lbr\n");
1712 		goto out_delete_evlist;
1713 	}
1714 
1715 	if (nr_cgroups > 0 && opts->record_cgroup) {
1716 		pr_err("--cgroup and --all-cgroups cannot be used together\n");
1717 		goto out_delete_evlist;
1718 	}
1719 
1720 	if (branch_call_mode) {
1721 		if (!opts->branch_stack)
1722 			opts->branch_stack = PERF_SAMPLE_BRANCH_ANY;
1723 		symbol_conf.use_callchain = true;
1724 		callchain_param.key = CCKEY_ADDRESS;
1725 		callchain_param.branch_callstack = true;
1726 		callchain_param.enabled = true;
1727 		if (callchain_param.record_mode == CALLCHAIN_NONE)
1728 			callchain_param.record_mode = CALLCHAIN_FP;
1729 		callchain_register_param(&callchain_param);
1730 		if (!sort_order)
1731 			sort_order = "srcline,symbol,dso";
1732 	}
1733 
1734 	if (opts->branch_stack && callchain_param.enabled)
1735 		symbol_conf.show_branchflag_count = true;
1736 
1737 	sort__mode = SORT_MODE__TOP;
1738 	/* display thread wants entries to be collapsed in a different tree */
1739 	perf_hpp_list.need_collapse = 1;
1740 
1741 	if (top.use_stdio)
1742 		use_browser = 0;
1743 #ifdef HAVE_SLANG_SUPPORT
1744 	else if (top.use_tui)
1745 		use_browser = 1;
1746 #endif
1747 
1748 	setup_browser(false);
1749 
1750 	if (setup_sorting(top.evlist) < 0) {
1751 		if (sort_order)
1752 			parse_options_usage(top_usage, options, "s", 1);
1753 		if (field_order)
1754 			parse_options_usage(sort_order ? NULL : top_usage,
1755 					    options, "fields", 0);
1756 		goto out_delete_evlist;
1757 	}
1758 
1759 	status = target__validate(target);
1760 	if (status) {
1761 		target__strerror(target, status, errbuf, BUFSIZ);
1762 		ui__warning("%s\n", errbuf);
1763 	}
1764 
1765 	status = target__parse_uid(target);
1766 	if (status) {
1767 		int saved_errno = errno;
1768 
1769 		target__strerror(target, status, errbuf, BUFSIZ);
1770 		ui__error("%s\n", errbuf);
1771 
1772 		status = -saved_errno;
1773 		goto out_delete_evlist;
1774 	}
1775 
1776 	if (target__none(target))
1777 		target->system_wide = true;
1778 
1779 	if (evlist__create_maps(top.evlist, target) < 0) {
1780 		ui__error("Couldn't create thread/CPU maps: %s\n",
1781 			  errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1782 		status = -errno;
1783 		goto out_delete_evlist;
1784 	}
1785 
1786 	if (top.delay_secs < 1)
1787 		top.delay_secs = 1;
1788 
1789 	if (record_opts__config(opts)) {
1790 		status = -EINVAL;
1791 		goto out_delete_evlist;
1792 	}
1793 
1794 	top.sym_evsel = evlist__first(top.evlist);
1795 
1796 	if (!callchain_param.enabled) {
1797 		symbol_conf.cumulate_callchain = false;
1798 		perf_hpp__cancel_cumulate(top.evlist);
1799 	}
1800 
1801 	if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1802 		callchain_param.order = ORDER_CALLER;
1803 
1804 	status = symbol__annotation_init();
1805 	if (status < 0)
1806 		goto out_delete_evlist;
1807 
1808 	annotation_config__init();
1809 
1810 	symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1811 	status = symbol__init(NULL);
1812 	if (status < 0)
1813 		goto out_delete_evlist;
1814 
1815 	sort__setup_elide(stdout);
1816 
1817 	get_term_dimensions(&top.winsize);
1818 	if (top.print_entries == 0) {
1819 		perf_top__update_print_entries(&top);
1820 		signal(SIGWINCH, winch_sig);
1821 	}
1822 
1823 	top.session = perf_session__new(NULL, NULL);
1824 	if (IS_ERR(top.session)) {
1825 		status = PTR_ERR(top.session);
1826 		top.session = NULL;
1827 		goto out_delete_evlist;
1828 	}
1829 
1830 	if (!evlist__needs_bpf_sb_event(top.evlist))
1831 		top.record_opts.no_bpf_event = true;
1832 
1833 #ifdef HAVE_LIBBPF_SUPPORT
1834 	if (!top.record_opts.no_bpf_event) {
1835 		top.sb_evlist = evlist__new();
1836 
1837 		if (top.sb_evlist == NULL) {
1838 			pr_err("Couldn't create side band evlist.\n.");
1839 			status = -EINVAL;
1840 			goto out_delete_evlist;
1841 		}
1842 
1843 		if (evlist__add_bpf_sb_event(top.sb_evlist, &perf_env)) {
1844 			pr_err("Couldn't ask for PERF_RECORD_BPF_EVENT side band events.\n.");
1845 			status = -EINVAL;
1846 			goto out_delete_evlist;
1847 		}
1848 	}
1849 #endif
1850 
1851 	if (evlist__start_sb_thread(top.sb_evlist, target)) {
1852 		pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1853 		opts->no_bpf_event = true;
1854 	}
1855 
1856 	status = __cmd_top(&top);
1857 
1858 	if (!opts->no_bpf_event)
1859 		evlist__stop_sb_thread(top.sb_evlist);
1860 
1861 out_delete_evlist:
1862 	evlist__delete(top.evlist);
1863 	perf_session__delete(top.session);
1864 	annotation_options__exit();
1865 
1866 	return status;
1867 }
1868