xref: /linux/tools/perf/builtin-kvm.c (revision c34e9ab9a612ee8b18273398ef75c207b01f516d)
1 // SPDX-License-Identifier: GPL-2.0
2 #include "builtin.h"
3 #include "perf.h"
4 
5 #include "util/build-id.h"
6 #include "util/evsel.h"
7 #include "util/evlist.h"
8 #include "util/mmap.h"
9 #include "util/term.h"
10 #include "util/symbol.h"
11 #include "util/thread.h"
12 #include "util/header.h"
13 #include "util/session.h"
14 #include "util/intlist.h"
15 #include <subcmd/pager.h>
16 #include <subcmd/parse-options.h>
17 #include "util/trace-event.h"
18 #include "util/debug.h"
19 #include "util/tool.h"
20 #include "util/stat.h"
21 #include "util/synthetic-events.h"
22 #include "util/top.h"
23 #include "util/data.h"
24 #include "util/ordered-events.h"
25 #include "util/kvm-stat.h"
26 #include "util/util.h"
27 #include "ui/browsers/hists.h"
28 #include "ui/progress.h"
29 #include "ui/ui.h"
30 #include "util/string2.h"
31 
32 #include <sys/prctl.h>
33 #ifdef HAVE_TIMERFD_SUPPORT
34 #include <sys/timerfd.h>
35 #endif
36 #include <sys/time.h>
37 #include <sys/types.h>
38 #include <sys/stat.h>
39 #include <fcntl.h>
40 
41 #include <linux/err.h>
42 #include <linux/kernel.h>
43 #include <linux/string.h>
44 #include <linux/time64.h>
45 #include <linux/zalloc.h>
46 #include <errno.h>
47 #include <inttypes.h>
48 #include <poll.h>
49 #include <termios.h>
50 #include <semaphore.h>
51 #include <signal.h>
52 #include <math.h>
53 #include <perf/mmap.h>
54 
55 #if defined(HAVE_KVM_STAT_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
56 #define GET_EVENT_KEY(func, field)					\
57 static u64 get_event_ ##func(struct kvm_event *event, int vcpu)		\
58 {									\
59 	if (vcpu == -1)							\
60 		return event->total.field;				\
61 									\
62 	if (vcpu >= event->max_vcpu)					\
63 		return 0;						\
64 									\
65 	return event->vcpu[vcpu].field;					\
66 }
67 
68 #define COMPARE_EVENT_KEY(func, field)					\
69 GET_EVENT_KEY(func, field)						\
70 static int64_t cmp_event_ ## func(struct kvm_event *one,		\
71 			      struct kvm_event *two, int vcpu)		\
72 {									\
73 	return get_event_ ##func(one, vcpu) -				\
74 	       get_event_ ##func(two, vcpu);				\
75 }
76 
77 COMPARE_EVENT_KEY(time, time);
78 COMPARE_EVENT_KEY(max, stats.max);
79 COMPARE_EVENT_KEY(min, stats.min);
80 COMPARE_EVENT_KEY(count, stats.n);
81 COMPARE_EVENT_KEY(mean, stats.mean);
82 
83 struct kvm_hists {
84 	struct hists		hists;
85 	struct perf_hpp_list	list;
86 };
87 
88 struct kvm_dimension {
89 	const char *name;
90 	const char *header;
91 	int width;
92 	int64_t (*cmp)(struct perf_hpp_fmt *fmt, struct hist_entry *left,
93 		       struct hist_entry *right);
94 	int (*entry)(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
95 		     struct hist_entry *he);
96 };
97 
98 struct kvm_fmt {
99 	struct perf_hpp_fmt	fmt;
100 	struct kvm_dimension	*dim;
101 };
102 
103 static struct kvm_hists kvm_hists;
104 
105 static int64_t ev_name_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
106 			   struct hist_entry *left,
107 			   struct hist_entry *right)
108 {
109 	/* Return opposite number for sorting in alphabetical order */
110 	return -strcmp(left->kvm_info->name, right->kvm_info->name);
111 }
112 
113 static int fmt_width(struct perf_hpp_fmt *fmt,
114 		     struct perf_hpp *hpp __maybe_unused,
115 		     struct hists *hists __maybe_unused);
116 
117 static int ev_name_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
118 			 struct hist_entry *he)
119 {
120 	int width = fmt_width(fmt, hpp, he->hists);
121 
122 	return scnprintf(hpp->buf, hpp->size, "%*s", width, he->kvm_info->name);
123 }
124 
125 static struct kvm_dimension dim_event = {
126 	.header		= "Event name",
127 	.name		= "ev_name",
128 	.cmp		= ev_name_cmp,
129 	.entry		= ev_name_entry,
130 	.width		= 40,
131 };
132 
133 #define EV_METRIC_CMP(metric)						\
134 static int64_t ev_cmp_##metric(struct perf_hpp_fmt *fmt __maybe_unused,	\
135 			       struct hist_entry *left,			\
136 			       struct hist_entry *right)		\
137 {									\
138 	struct kvm_event *event_left;					\
139 	struct kvm_event *event_right;					\
140 	struct perf_kvm_stat *perf_kvm;					\
141 									\
142 	event_left  = container_of(left, struct kvm_event, he);		\
143 	event_right = container_of(right, struct kvm_event, he);	\
144 									\
145 	perf_kvm = event_left->perf_kvm;				\
146 	return cmp_event_##metric(event_left, event_right,		\
147 				  perf_kvm->trace_vcpu);		\
148 }
149 
150 EV_METRIC_CMP(time)
151 EV_METRIC_CMP(count)
152 EV_METRIC_CMP(max)
153 EV_METRIC_CMP(min)
154 EV_METRIC_CMP(mean)
155 
156 #define EV_METRIC_ENTRY(metric)						\
157 static int ev_entry_##metric(struct perf_hpp_fmt *fmt,			\
158 			     struct perf_hpp *hpp,			\
159 			     struct hist_entry *he)			\
160 {									\
161 	struct kvm_event *event;					\
162 	int width = fmt_width(fmt, hpp, he->hists);			\
163 	struct perf_kvm_stat *perf_kvm;					\
164 									\
165 	event = container_of(he, struct kvm_event, he);			\
166 	perf_kvm = event->perf_kvm;					\
167 	return scnprintf(hpp->buf, hpp->size, "%*lu", width,		\
168 		get_event_##metric(event, perf_kvm->trace_vcpu));	\
169 }
170 
171 EV_METRIC_ENTRY(time)
172 EV_METRIC_ENTRY(count)
173 EV_METRIC_ENTRY(max)
174 EV_METRIC_ENTRY(min)
175 
176 static struct kvm_dimension dim_time = {
177 	.header		= "Time (ns)",
178 	.name		= "time",
179 	.cmp		= ev_cmp_time,
180 	.entry		= ev_entry_time,
181 	.width		= 12,
182 };
183 
184 static struct kvm_dimension dim_count = {
185 	.header		= "Samples",
186 	.name		= "sample",
187 	.cmp		= ev_cmp_count,
188 	.entry		= ev_entry_count,
189 	.width		= 12,
190 };
191 
192 static struct kvm_dimension dim_max_time = {
193 	.header		= "Max Time (ns)",
194 	.name		= "max_t",
195 	.cmp		= ev_cmp_max,
196 	.entry		= ev_entry_max,
197 	.width		= 14,
198 };
199 
200 static struct kvm_dimension dim_min_time = {
201 	.header		= "Min Time (ns)",
202 	.name		= "min_t",
203 	.cmp		= ev_cmp_min,
204 	.entry		= ev_entry_min,
205 	.width		= 14,
206 };
207 
208 static int ev_entry_mean(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
209 			 struct hist_entry *he)
210 {
211 	struct kvm_event *event;
212 	int width = fmt_width(fmt, hpp, he->hists);
213 	struct perf_kvm_stat *perf_kvm;
214 
215 	event = container_of(he, struct kvm_event, he);
216 	perf_kvm = event->perf_kvm;
217 	return scnprintf(hpp->buf, hpp->size, "%*lu", width,
218 			 get_event_mean(event, perf_kvm->trace_vcpu));
219 }
220 
221 static struct kvm_dimension dim_mean_time = {
222 	.header		= "Mean Time (ns)",
223 	.name		= "mean_t",
224 	.cmp		= ev_cmp_mean,
225 	.entry		= ev_entry_mean,
226 	.width		= 14,
227 };
228 
229 #define PERC_STR(__s, __v)				\
230 ({							\
231 	scnprintf(__s, sizeof(__s), "%.2F%%", __v);	\
232 	__s;						\
233 })
234 
235 static double percent(u64 st, u64 tot)
236 {
237 	return tot ? 100. * (double) st / (double) tot : 0;
238 }
239 
240 #define EV_METRIC_PERCENT(metric)					\
241 static int ev_percent_##metric(struct hist_entry *he)			\
242 {									\
243 	struct kvm_event *event;					\
244 	struct perf_kvm_stat *perf_kvm;					\
245 									\
246 	event = container_of(he, struct kvm_event, he);			\
247 	perf_kvm = event->perf_kvm;					\
248 									\
249 	return percent(get_event_##metric(event, perf_kvm->trace_vcpu),	\
250 		       perf_kvm->total_##metric);			\
251 }
252 
253 EV_METRIC_PERCENT(time)
254 EV_METRIC_PERCENT(count)
255 
256 static int ev_entry_time_precent(struct perf_hpp_fmt *fmt,
257 				 struct perf_hpp *hpp,
258 				 struct hist_entry *he)
259 {
260 	int width = fmt_width(fmt, hpp, he->hists);
261 	double per;
262 	char buf[10];
263 
264 	per = ev_percent_time(he);
265 	return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
266 }
267 
268 static int64_t
269 ev_cmp_time_precent(struct perf_hpp_fmt *fmt __maybe_unused,
270 		    struct hist_entry *left, struct hist_entry *right)
271 {
272 	double per_left;
273 	double per_right;
274 
275 	per_left  = ev_percent_time(left);
276 	per_right = ev_percent_time(right);
277 
278 	return per_left - per_right;
279 }
280 
281 static struct kvm_dimension dim_time_percent = {
282 	.header		= "Time%",
283 	.name		= "percent_time",
284 	.cmp		= ev_cmp_time_precent,
285 	.entry		= ev_entry_time_precent,
286 	.width		= 12,
287 };
288 
289 static int ev_entry_count_precent(struct perf_hpp_fmt *fmt,
290 				  struct perf_hpp *hpp,
291 				  struct hist_entry *he)
292 {
293 	int width = fmt_width(fmt, hpp, he->hists);
294 	double per;
295 	char buf[10];
296 
297 	per = ev_percent_count(he);
298 	return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
299 }
300 
301 static int64_t
302 ev_cmp_count_precent(struct perf_hpp_fmt *fmt __maybe_unused,
303 		     struct hist_entry *left, struct hist_entry *right)
304 {
305 	double per_left;
306 	double per_right;
307 
308 	per_left  = ev_percent_count(left);
309 	per_right = ev_percent_count(right);
310 
311 	return per_left - per_right;
312 }
313 
314 static struct kvm_dimension dim_count_percent = {
315 	.header		= "Sample%",
316 	.name		= "percent_sample",
317 	.cmp		= ev_cmp_count_precent,
318 	.entry		= ev_entry_count_precent,
319 	.width		= 12,
320 };
321 
322 static struct kvm_dimension *dimensions[] = {
323 	&dim_event,
324 	&dim_time,
325 	&dim_time_percent,
326 	&dim_count,
327 	&dim_count_percent,
328 	&dim_max_time,
329 	&dim_min_time,
330 	&dim_mean_time,
331 	NULL,
332 };
333 
334 static int fmt_width(struct perf_hpp_fmt *fmt,
335 		     struct perf_hpp *hpp __maybe_unused,
336 		     struct hists *hists __maybe_unused)
337 {
338 	struct kvm_fmt *kvm_fmt;
339 
340 	kvm_fmt = container_of(fmt, struct kvm_fmt, fmt);
341 	return kvm_fmt->dim->width;
342 }
343 
344 static int fmt_header(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
345 		      struct hists *hists, int line __maybe_unused,
346 		      int *span __maybe_unused)
347 {
348 	struct kvm_fmt *kvm_fmt;
349 	struct kvm_dimension *dim;
350 	int width = fmt_width(fmt, hpp, hists);
351 
352 	kvm_fmt = container_of(fmt, struct kvm_fmt, fmt);
353 	dim = kvm_fmt->dim;
354 
355 	return scnprintf(hpp->buf, hpp->size, "%*s", width, dim->header);
356 }
357 
358 static bool fmt_equal(struct perf_hpp_fmt *a, struct perf_hpp_fmt *b)
359 {
360 	struct kvm_fmt *kvm_fmt_a = container_of(a, struct kvm_fmt, fmt);
361 	struct kvm_fmt *kvm_fmt_b = container_of(b, struct kvm_fmt, fmt);
362 
363 	return kvm_fmt_a->dim == kvm_fmt_b->dim;
364 }
365 
366 static void fmt_free(struct perf_hpp_fmt *fmt)
367 {
368 	struct kvm_fmt *kvm_fmt;
369 
370 	kvm_fmt = container_of(fmt, struct kvm_fmt, fmt);
371 	free(kvm_fmt);
372 }
373 
374 static struct kvm_dimension *get_dimension(const char *name)
375 {
376 	unsigned int i;
377 
378 	for (i = 0; dimensions[i] != NULL; i++) {
379 		if (!strcmp(dimensions[i]->name, name))
380 			return dimensions[i];
381 	}
382 
383 	return NULL;
384 }
385 
386 static struct kvm_fmt *get_format(const char *name)
387 {
388 	struct kvm_dimension *dim = get_dimension(name);
389 	struct kvm_fmt *kvm_fmt;
390 	struct perf_hpp_fmt *fmt;
391 
392 	if (!dim)
393 		return NULL;
394 
395 	kvm_fmt = zalloc(sizeof(*kvm_fmt));
396 	if (!kvm_fmt)
397 		return NULL;
398 
399 	kvm_fmt->dim = dim;
400 
401 	fmt = &kvm_fmt->fmt;
402 	INIT_LIST_HEAD(&fmt->list);
403 	INIT_LIST_HEAD(&fmt->sort_list);
404 	fmt->cmp	= dim->cmp;
405 	fmt->sort	= dim->cmp;
406 	fmt->color	= NULL;
407 	fmt->entry	= dim->entry;
408 	fmt->header	= fmt_header;
409 	fmt->width	= fmt_width;
410 	fmt->collapse	= dim->cmp;
411 	fmt->equal	= fmt_equal;
412 	fmt->free	= fmt_free;
413 
414 	return kvm_fmt;
415 }
416 
417 static int kvm_hists__init_output(struct perf_hpp_list *hpp_list, char *name)
418 {
419 	struct kvm_fmt *kvm_fmt = get_format(name);
420 
421 	if (!kvm_fmt) {
422 		pr_warning("Fail to find format for output field %s.\n", name);
423 		return -EINVAL;
424 	}
425 
426 	perf_hpp_list__column_register(hpp_list, &kvm_fmt->fmt);
427 	return 0;
428 }
429 
430 static int kvm_hists__init_sort(struct perf_hpp_list *hpp_list, char *name)
431 {
432 	struct kvm_fmt *kvm_fmt = get_format(name);
433 
434 	if (!kvm_fmt) {
435 		pr_warning("Fail to find format for sorting %s.\n", name);
436 		return -EINVAL;
437 	}
438 
439 	perf_hpp_list__register_sort_field(hpp_list, &kvm_fmt->fmt);
440 	return 0;
441 }
442 
443 static int kvm_hpp_list__init(char *list,
444 			      struct perf_hpp_list *hpp_list,
445 			      int (*fn)(struct perf_hpp_list *hpp_list,
446 					char *name))
447 {
448 	char *tmp, *tok;
449 	int ret;
450 
451 	if (!list || !fn)
452 		return 0;
453 
454 	for (tok = strtok_r(list, ", ", &tmp); tok;
455 	     tok = strtok_r(NULL, ", ", &tmp)) {
456 		ret = fn(hpp_list, tok);
457 		if (!ret)
458 			continue;
459 
460 		/* Handle errors */
461 		if (ret == -EINVAL)
462 			pr_err("Invalid field key: '%s'", tok);
463 		else if (ret == -ESRCH)
464 			pr_err("Unknown field key: '%s'", tok);
465 		else
466 			pr_err("Fail to initialize for field key: '%s'", tok);
467 
468 		break;
469 	}
470 
471 	return ret;
472 }
473 
474 static int kvm_hpp_list__parse(struct perf_hpp_list *hpp_list,
475 			       const char *output_, const char *sort_)
476 {
477 	char *output = output_ ? strdup(output_) : NULL;
478 	char *sort = sort_ ? strdup(sort_) : NULL;
479 	int ret;
480 
481 	ret = kvm_hpp_list__init(output, hpp_list, kvm_hists__init_output);
482 	if (ret)
483 		goto out;
484 
485 	ret = kvm_hpp_list__init(sort, hpp_list, kvm_hists__init_sort);
486 	if (ret)
487 		goto out;
488 
489 	/* Copy sort keys to output fields */
490 	perf_hpp__setup_output_field(hpp_list);
491 
492 	/* and then copy output fields to sort keys */
493 	perf_hpp__append_sort_keys(hpp_list);
494 out:
495 	free(output);
496 	free(sort);
497 	return ret;
498 }
499 
500 static int kvm_hists__init(void)
501 {
502 	kvm_hists.list.nr_header_lines = 1;
503 	__hists__init(&kvm_hists.hists, &kvm_hists.list);
504 	perf_hpp_list__init(&kvm_hists.list);
505 	return kvm_hpp_list__parse(&kvm_hists.list, NULL, "ev_name");
506 }
507 
508 static int kvm_hists__reinit(const char *output, const char *sort)
509 {
510 	perf_hpp__reset_output_field(&kvm_hists.list);
511 	return kvm_hpp_list__parse(&kvm_hists.list, output, sort);
512 }
513 static void print_result(struct perf_kvm_stat *kvm);
514 
515 #ifdef HAVE_SLANG_SUPPORT
516 static void kvm_browser__update_nr_entries(struct hist_browser *hb)
517 {
518 	struct rb_node *nd = rb_first_cached(&hb->hists->entries);
519 	u64 nr_entries = 0;
520 
521 	for (; nd; nd = rb_next(nd)) {
522 		struct hist_entry *he = rb_entry(nd, struct hist_entry,
523 						 rb_node);
524 
525 		if (!he->filtered)
526 			nr_entries++;
527 	}
528 
529 	hb->nr_non_filtered_entries = nr_entries;
530 }
531 
532 static int kvm_browser__title(struct hist_browser *browser,
533 			      char *buf, size_t size)
534 {
535 	scnprintf(buf, size, "KVM event statistics (%lu entries)",
536 		  browser->nr_non_filtered_entries);
537 	return 0;
538 }
539 
540 static struct hist_browser*
541 perf_kvm_browser__new(struct hists *hists)
542 {
543 	struct hist_browser *browser = hist_browser__new(hists);
544 
545 	if (browser)
546 		browser->title = kvm_browser__title;
547 
548 	return browser;
549 }
550 
551 static int kvm__hists_browse(struct hists *hists)
552 {
553 	struct hist_browser *browser;
554 	int key = -1;
555 
556 	browser = perf_kvm_browser__new(hists);
557 	if (browser == NULL)
558 		return -1;
559 
560 	/* reset abort key so that it can get Ctrl-C as a key */
561 	SLang_reset_tty();
562 	SLang_init_tty(0, 0, 0);
563 
564 	kvm_browser__update_nr_entries(browser);
565 
566 	while (1) {
567 		key = hist_browser__run(browser, "? - help", true, 0);
568 
569 		switch (key) {
570 		case 'q':
571 			goto out;
572 		default:
573 			break;
574 		}
575 	}
576 
577 out:
578 	hist_browser__delete(browser);
579 	return 0;
580 }
581 
582 static void kvm_display(struct perf_kvm_stat *kvm)
583 {
584 	if (!use_browser)
585 		print_result(kvm);
586 	else
587 		kvm__hists_browse(&kvm_hists.hists);
588 }
589 
590 #else
591 
592 static void kvm_display(struct perf_kvm_stat *kvm)
593 {
594 	use_browser = 0;
595 	print_result(kvm);
596 }
597 
598 #endif /* HAVE_SLANG_SUPPORT */
599 
600 #endif // defined(HAVE_KVM_STAT_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
601 
602 static const char *get_filename_for_perf_kvm(void)
603 {
604 	const char *filename;
605 
606 	if (perf_host && !perf_guest)
607 		filename = strdup("perf.data.host");
608 	else if (!perf_host && perf_guest)
609 		filename = strdup("perf.data.guest");
610 	else
611 		filename = strdup("perf.data.kvm");
612 
613 	return filename;
614 }
615 
616 #if defined(HAVE_KVM_STAT_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
617 
618 void exit_event_get_key(struct evsel *evsel,
619 			struct perf_sample *sample,
620 			struct event_key *key)
621 {
622 	key->info = 0;
623 	key->key  = evsel__intval(evsel, sample, kvm_exit_reason);
624 }
625 
626 bool kvm_exit_event(struct evsel *evsel)
627 {
628 	return evsel__name_is(evsel, kvm_exit_trace);
629 }
630 
631 bool exit_event_begin(struct evsel *evsel,
632 		      struct perf_sample *sample, struct event_key *key)
633 {
634 	if (kvm_exit_event(evsel)) {
635 		exit_event_get_key(evsel, sample, key);
636 		return true;
637 	}
638 
639 	return false;
640 }
641 
642 bool kvm_entry_event(struct evsel *evsel)
643 {
644 	return evsel__name_is(evsel, kvm_entry_trace);
645 }
646 
647 bool exit_event_end(struct evsel *evsel,
648 		    struct perf_sample *sample __maybe_unused,
649 		    struct event_key *key __maybe_unused)
650 {
651 	return kvm_entry_event(evsel);
652 }
653 
654 static const char *get_exit_reason(struct perf_kvm_stat *kvm,
655 				   struct exit_reasons_table *tbl,
656 				   u64 exit_code)
657 {
658 	while (tbl->reason != NULL) {
659 		if (tbl->exit_code == exit_code)
660 			return tbl->reason;
661 		tbl++;
662 	}
663 
664 	pr_err("unknown kvm exit code:%lld on %s\n",
665 		(unsigned long long)exit_code, kvm->exit_reasons_isa);
666 	return "UNKNOWN";
667 }
668 
669 void exit_event_decode_key(struct perf_kvm_stat *kvm,
670 			   struct event_key *key,
671 			   char *decode)
672 {
673 	const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
674 						  key->key);
675 
676 	scnprintf(decode, KVM_EVENT_NAME_LEN, "%s", exit_reason);
677 }
678 
679 static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
680 {
681 	struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
682 
683 	for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
684 		if (!strcmp(events_ops->name, kvm->report_event)) {
685 			kvm->events_ops = events_ops->ops;
686 			return true;
687 		}
688 	}
689 
690 	return false;
691 }
692 
693 struct vcpu_event_record {
694 	int vcpu_id;
695 	u64 start_time;
696 	struct kvm_event *last_event;
697 };
698 
699 #ifdef HAVE_TIMERFD_SUPPORT
700 static void clear_events_cache_stats(void)
701 {
702 	struct rb_root_cached *root;
703 	struct rb_node *nd;
704 	struct kvm_event *event;
705 	int i;
706 
707 	if (hists__has(&kvm_hists.hists, need_collapse))
708 		root = &kvm_hists.hists.entries_collapsed;
709 	else
710 		root = kvm_hists.hists.entries_in;
711 
712 	for (nd = rb_first_cached(root); nd; nd = rb_next(nd)) {
713 		struct hist_entry *he;
714 
715 		he = rb_entry(nd, struct hist_entry, rb_node_in);
716 		event = container_of(he, struct kvm_event, he);
717 
718 		/* reset stats for event */
719 		event->total.time = 0;
720 		init_stats(&event->total.stats);
721 
722 		for (i = 0; i < event->max_vcpu; ++i) {
723 			event->vcpu[i].time = 0;
724 			init_stats(&event->vcpu[i].stats);
725 		}
726 	}
727 }
728 #endif
729 
730 static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
731 {
732 	int old_max_vcpu = event->max_vcpu;
733 	void *prev;
734 
735 	if (vcpu_id < event->max_vcpu)
736 		return true;
737 
738 	while (event->max_vcpu <= vcpu_id)
739 		event->max_vcpu += DEFAULT_VCPU_NUM;
740 
741 	prev = event->vcpu;
742 	event->vcpu = realloc(event->vcpu,
743 			      event->max_vcpu * sizeof(*event->vcpu));
744 	if (!event->vcpu) {
745 		free(prev);
746 		pr_err("Not enough memory\n");
747 		return false;
748 	}
749 
750 	memset(event->vcpu + old_max_vcpu, 0,
751 	       (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
752 	return true;
753 }
754 
755 static void *kvm_he_zalloc(size_t size)
756 {
757 	struct kvm_event *kvm_ev;
758 
759 	kvm_ev = zalloc(size + sizeof(*kvm_ev));
760 	if (!kvm_ev)
761 		return NULL;
762 
763 	init_stats(&kvm_ev->total.stats);
764 	hists__inc_nr_samples(&kvm_hists.hists, 0);
765 	return &kvm_ev->he;
766 }
767 
768 static void kvm_he_free(void *he)
769 {
770 	struct kvm_event *kvm_ev;
771 
772 	kvm_ev = container_of(he, struct kvm_event, he);
773 	free(kvm_ev);
774 }
775 
776 static struct hist_entry_ops kvm_ev_entry_ops = {
777 	.new	= kvm_he_zalloc,
778 	.free	= kvm_he_free,
779 };
780 
781 static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
782 					       struct event_key *key,
783 					       struct perf_sample *sample)
784 {
785 	struct kvm_event *event;
786 	struct hist_entry *he;
787 	struct kvm_info *ki;
788 
789 	BUG_ON(key->key == INVALID_KEY);
790 
791 	ki = kvm_info__new();
792 	if (!ki) {
793 		pr_err("Failed to allocate kvm info\n");
794 		return NULL;
795 	}
796 
797 	kvm->events_ops->decode_key(kvm, key, ki->name);
798 	he = hists__add_entry_ops(&kvm_hists.hists, &kvm_ev_entry_ops,
799 				  &kvm->al, NULL, NULL, NULL, ki, sample, true);
800 	if (he == NULL) {
801 		pr_err("Failed to allocate hist entry\n");
802 		free(ki);
803 		return NULL;
804 	}
805 
806 	event = container_of(he, struct kvm_event, he);
807 	if (!event->perf_kvm) {
808 		event->perf_kvm = kvm;
809 		event->key = *key;
810 	}
811 
812 	return event;
813 }
814 
815 static bool handle_begin_event(struct perf_kvm_stat *kvm,
816 			       struct vcpu_event_record *vcpu_record,
817 			       struct event_key *key,
818 			       struct perf_sample *sample)
819 {
820 	struct kvm_event *event = NULL;
821 
822 	if (key->key != INVALID_KEY)
823 		event = find_create_kvm_event(kvm, key, sample);
824 
825 	vcpu_record->last_event = event;
826 	vcpu_record->start_time = sample->time;
827 	return true;
828 }
829 
830 static void
831 kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
832 {
833 	kvm_stats->time += time_diff;
834 	update_stats(&kvm_stats->stats, time_diff);
835 }
836 
837 static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
838 {
839 	struct kvm_event_stats *kvm_stats = &event->total;
840 
841 	if (vcpu_id != -1)
842 		kvm_stats = &event->vcpu[vcpu_id];
843 
844 	return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
845 				avg_stats(&kvm_stats->stats));
846 }
847 
848 static bool update_kvm_event(struct perf_kvm_stat *kvm,
849 			     struct kvm_event *event, int vcpu_id,
850 			     u64 time_diff)
851 {
852 	/* Update overall statistics */
853 	kvm->total_count++;
854 	kvm->total_time += time_diff;
855 
856 	if (vcpu_id == -1) {
857 		kvm_update_event_stats(&event->total, time_diff);
858 		return true;
859 	}
860 
861 	if (!kvm_event_expand(event, vcpu_id))
862 		return false;
863 
864 	kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
865 	return true;
866 }
867 
868 static bool is_child_event(struct perf_kvm_stat *kvm,
869 			   struct evsel *evsel,
870 			   struct perf_sample *sample,
871 			   struct event_key *key)
872 {
873 	struct child_event_ops *child_ops;
874 
875 	child_ops = kvm->events_ops->child_ops;
876 
877 	if (!child_ops)
878 		return false;
879 
880 	for (; child_ops->name; child_ops++) {
881 		if (evsel__name_is(evsel, child_ops->name)) {
882 			child_ops->get_key(evsel, sample, key);
883 			return true;
884 		}
885 	}
886 
887 	return false;
888 }
889 
890 static bool handle_child_event(struct perf_kvm_stat *kvm,
891 			       struct vcpu_event_record *vcpu_record,
892 			       struct event_key *key,
893 			       struct perf_sample *sample)
894 {
895 	struct kvm_event *event = NULL;
896 
897 	if (key->key != INVALID_KEY)
898 		event = find_create_kvm_event(kvm, key, sample);
899 
900 	vcpu_record->last_event = event;
901 
902 	return true;
903 }
904 
905 static bool skip_event(const char *event)
906 {
907 	const char * const *skip_events;
908 
909 	for (skip_events = kvm_skip_events; *skip_events; skip_events++)
910 		if (!strcmp(event, *skip_events))
911 			return true;
912 
913 	return false;
914 }
915 
916 static bool handle_end_event(struct perf_kvm_stat *kvm,
917 			     struct vcpu_event_record *vcpu_record,
918 			     struct event_key *key,
919 			     struct perf_sample *sample)
920 {
921 	struct kvm_event *event;
922 	u64 time_begin, time_diff;
923 	int vcpu;
924 
925 	if (kvm->trace_vcpu == -1)
926 		vcpu = -1;
927 	else
928 		vcpu = vcpu_record->vcpu_id;
929 
930 	event = vcpu_record->last_event;
931 	time_begin = vcpu_record->start_time;
932 
933 	/* The begin event is not caught. */
934 	if (!time_begin)
935 		return true;
936 
937 	/*
938 	 * In some case, the 'begin event' only records the start timestamp,
939 	 * the actual event is recognized in the 'end event' (e.g. mmio-event).
940 	 */
941 
942 	/* Both begin and end events did not get the key. */
943 	if (!event && key->key == INVALID_KEY)
944 		return true;
945 
946 	if (!event)
947 		event = find_create_kvm_event(kvm, key, sample);
948 
949 	if (!event)
950 		return false;
951 
952 	vcpu_record->last_event = NULL;
953 	vcpu_record->start_time = 0;
954 
955 	/* seems to happen once in a while during live mode */
956 	if (sample->time < time_begin) {
957 		pr_debug("End time before begin time; skipping event.\n");
958 		return true;
959 	}
960 
961 	time_diff = sample->time - time_begin;
962 
963 	if (kvm->duration && time_diff > kvm->duration) {
964 		char decode[KVM_EVENT_NAME_LEN];
965 
966 		kvm->events_ops->decode_key(kvm, &event->key, decode);
967 		if (!skip_event(decode)) {
968 			pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
969 				 sample->time, sample->pid, vcpu_record->vcpu_id,
970 				 decode, time_diff / NSEC_PER_USEC);
971 		}
972 	}
973 
974 	return update_kvm_event(kvm, event, vcpu, time_diff);
975 }
976 
977 static
978 struct vcpu_event_record *per_vcpu_record(struct thread *thread,
979 					  struct evsel *evsel,
980 					  struct perf_sample *sample)
981 {
982 	/* Only kvm_entry records vcpu id. */
983 	if (!thread__priv(thread) && kvm_entry_event(evsel)) {
984 		struct vcpu_event_record *vcpu_record;
985 
986 		vcpu_record = zalloc(sizeof(*vcpu_record));
987 		if (!vcpu_record) {
988 			pr_err("%s: Not enough memory\n", __func__);
989 			return NULL;
990 		}
991 
992 		vcpu_record->vcpu_id = evsel__intval(evsel, sample, vcpu_id_str);
993 		thread__set_priv(thread, vcpu_record);
994 	}
995 
996 	return thread__priv(thread);
997 }
998 
999 static bool handle_kvm_event(struct perf_kvm_stat *kvm,
1000 			     struct thread *thread,
1001 			     struct evsel *evsel,
1002 			     struct perf_sample *sample)
1003 {
1004 	struct vcpu_event_record *vcpu_record;
1005 	struct event_key key = { .key = INVALID_KEY,
1006 				 .exit_reasons = kvm->exit_reasons };
1007 
1008 	vcpu_record = per_vcpu_record(thread, evsel, sample);
1009 	if (!vcpu_record)
1010 		return true;
1011 
1012 	/* only process events for vcpus user cares about */
1013 	if ((kvm->trace_vcpu != -1) &&
1014 	    (kvm->trace_vcpu != vcpu_record->vcpu_id))
1015 		return true;
1016 
1017 	if (kvm->events_ops->is_begin_event(evsel, sample, &key))
1018 		return handle_begin_event(kvm, vcpu_record, &key, sample);
1019 
1020 	if (is_child_event(kvm, evsel, sample, &key))
1021 		return handle_child_event(kvm, vcpu_record, &key, sample);
1022 
1023 	if (kvm->events_ops->is_end_event(evsel, sample, &key))
1024 		return handle_end_event(kvm, vcpu_record, &key, sample);
1025 
1026 	return true;
1027 }
1028 
1029 static bool is_valid_key(struct perf_kvm_stat *kvm)
1030 {
1031 	static const char *key_array[] = {
1032 		"ev_name", "sample", "time", "max_t", "min_t", "mean_t",
1033 	};
1034 	unsigned int i;
1035 
1036 	for (i = 0; i < ARRAY_SIZE(key_array); i++)
1037 		if (!strcmp(key_array[i], kvm->sort_key))
1038 			return true;
1039 
1040 	pr_err("Unsupported sort key: %s\n", kvm->sort_key);
1041 	return false;
1042 }
1043 
1044 static bool event_is_valid(struct kvm_event *event, int vcpu)
1045 {
1046 	return !!get_event_count(event, vcpu);
1047 }
1048 
1049 static int filter_cb(struct hist_entry *he, void *arg __maybe_unused)
1050 {
1051 	struct kvm_event *event;
1052 	struct perf_kvm_stat *perf_kvm;
1053 
1054 	event = container_of(he, struct kvm_event, he);
1055 	perf_kvm = event->perf_kvm;
1056 	if (!event_is_valid(event, perf_kvm->trace_vcpu))
1057 		he->filtered = 1;
1058 	else
1059 		he->filtered = 0;
1060 	return 0;
1061 }
1062 
1063 static void sort_result(struct perf_kvm_stat *kvm)
1064 {
1065 	struct ui_progress prog;
1066 	const char *output_columns = "ev_name,sample,percent_sample,"
1067 				     "time,percent_time,max_t,min_t,mean_t";
1068 
1069 	kvm_hists__reinit(output_columns, kvm->sort_key);
1070 	ui_progress__init(&prog, kvm_hists.hists.nr_entries, "Sorting...");
1071 	hists__collapse_resort(&kvm_hists.hists, NULL);
1072 	hists__output_resort_cb(&kvm_hists.hists, NULL, filter_cb);
1073 	ui_progress__finish();
1074 }
1075 
1076 static void print_vcpu_info(struct perf_kvm_stat *kvm)
1077 {
1078 	int vcpu = kvm->trace_vcpu;
1079 
1080 	pr_info("Analyze events for ");
1081 
1082 	if (kvm->opts.target.system_wide)
1083 		pr_info("all VMs, ");
1084 	else if (kvm->opts.target.pid)
1085 		pr_info("pid(s) %s, ", kvm->opts.target.pid);
1086 	else
1087 		pr_info("dazed and confused on what is monitored, ");
1088 
1089 	if (vcpu == -1)
1090 		pr_info("all VCPUs:\n\n");
1091 	else
1092 		pr_info("VCPU %d:\n\n", vcpu);
1093 }
1094 
1095 static void show_timeofday(void)
1096 {
1097 	char date[64];
1098 	struct timeval tv;
1099 	struct tm ltime;
1100 
1101 	gettimeofday(&tv, NULL);
1102 	if (localtime_r(&tv.tv_sec, &ltime)) {
1103 		strftime(date, sizeof(date), "%H:%M:%S", &ltime);
1104 		pr_info("%s.%06ld", date, tv.tv_usec);
1105 	} else
1106 		pr_info("00:00:00.000000");
1107 
1108 	return;
1109 }
1110 
1111 static void print_result(struct perf_kvm_stat *kvm)
1112 {
1113 	char decode[KVM_EVENT_NAME_LEN];
1114 	struct kvm_event *event;
1115 	int vcpu = kvm->trace_vcpu;
1116 	struct rb_node *nd;
1117 
1118 	if (kvm->live) {
1119 		puts(CONSOLE_CLEAR);
1120 		show_timeofday();
1121 	}
1122 
1123 	pr_info("\n\n");
1124 	print_vcpu_info(kvm);
1125 	pr_info("%*s ", KVM_EVENT_NAME_LEN, kvm->events_ops->name);
1126 	pr_info("%10s ", "Samples");
1127 	pr_info("%9s ", "Samples%");
1128 
1129 	pr_info("%9s ", "Time%");
1130 	pr_info("%11s ", "Min Time");
1131 	pr_info("%11s ", "Max Time");
1132 	pr_info("%16s ", "Avg time");
1133 	pr_info("\n\n");
1134 
1135 	for (nd = rb_first_cached(&kvm_hists.hists.entries); nd; nd = rb_next(nd)) {
1136 		struct hist_entry *he;
1137 		u64 ecount, etime, max, min;
1138 
1139 		he = rb_entry(nd, struct hist_entry, rb_node);
1140 		if (he->filtered)
1141 			continue;
1142 
1143 		event = container_of(he, struct kvm_event, he);
1144 		ecount = get_event_count(event, vcpu);
1145 		etime = get_event_time(event, vcpu);
1146 		max = get_event_max(event, vcpu);
1147 		min = get_event_min(event, vcpu);
1148 
1149 		kvm->events_ops->decode_key(kvm, &event->key, decode);
1150 		pr_info("%*s ", KVM_EVENT_NAME_LEN, decode);
1151 		pr_info("%10llu ", (unsigned long long)ecount);
1152 		pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
1153 		pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
1154 		pr_info("%9.2fus ", (double)min / NSEC_PER_USEC);
1155 		pr_info("%9.2fus ", (double)max / NSEC_PER_USEC);
1156 		pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount / NSEC_PER_USEC,
1157 			kvm_event_rel_stddev(vcpu, event));
1158 		pr_info("\n");
1159 	}
1160 
1161 	pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
1162 		kvm->total_count, kvm->total_time / (double)NSEC_PER_USEC);
1163 
1164 	if (kvm->lost_events)
1165 		pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
1166 }
1167 
1168 #if defined(HAVE_TIMERFD_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
1169 static int process_lost_event(const struct perf_tool *tool,
1170 			      union perf_event *event __maybe_unused,
1171 			      struct perf_sample *sample __maybe_unused,
1172 			      struct machine *machine __maybe_unused)
1173 {
1174 	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
1175 
1176 	kvm->lost_events++;
1177 	return 0;
1178 }
1179 #endif
1180 
1181 static bool skip_sample(struct perf_kvm_stat *kvm,
1182 			struct perf_sample *sample)
1183 {
1184 	if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
1185 		return true;
1186 
1187 	return false;
1188 }
1189 
1190 static int process_sample_event(const struct perf_tool *tool,
1191 				union perf_event *event,
1192 				struct perf_sample *sample,
1193 				struct evsel *evsel,
1194 				struct machine *machine)
1195 {
1196 	int err = 0;
1197 	struct thread *thread;
1198 	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
1199 						 tool);
1200 
1201 	if (skip_sample(kvm, sample))
1202 		return 0;
1203 
1204 	if (machine__resolve(machine, &kvm->al, sample) < 0) {
1205 		pr_warning("Fail to resolve address location, skip sample.\n");
1206 		return 0;
1207 	}
1208 
1209 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
1210 	if (thread == NULL) {
1211 		pr_debug("problem processing %d event, skipping it.\n",
1212 			event->header.type);
1213 		return -1;
1214 	}
1215 
1216 	if (!handle_kvm_event(kvm, thread, evsel, sample))
1217 		err = -1;
1218 
1219 	thread__put(thread);
1220 	return err;
1221 }
1222 
1223 static int cpu_isa_config(struct perf_kvm_stat *kvm)
1224 {
1225 	char buf[128], *cpuid;
1226 	int err;
1227 
1228 	if (kvm->live) {
1229 		struct perf_cpu cpu = {-1};
1230 
1231 		err = get_cpuid(buf, sizeof(buf), cpu);
1232 		if (err != 0) {
1233 			pr_err("Failed to look up CPU type: %s\n",
1234 			       str_error_r(err, buf, sizeof(buf)));
1235 			return -err;
1236 		}
1237 		cpuid = buf;
1238 	} else
1239 		cpuid = kvm->session->header.env.cpuid;
1240 
1241 	if (!cpuid) {
1242 		pr_err("Failed to look up CPU type\n");
1243 		return -EINVAL;
1244 	}
1245 
1246 	err = cpu_isa_init(kvm, cpuid);
1247 	if (err == -ENOTSUP)
1248 		pr_err("CPU %s is not supported.\n", cpuid);
1249 
1250 	return err;
1251 }
1252 
1253 static bool verify_vcpu(int vcpu)
1254 {
1255 	if (vcpu != -1 && vcpu < 0) {
1256 		pr_err("Invalid vcpu:%d.\n", vcpu);
1257 		return false;
1258 	}
1259 
1260 	return true;
1261 }
1262 
1263 #if defined(HAVE_TIMERFD_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
1264 /* keeping the max events to a modest level to keep
1265  * the processing of samples per mmap smooth.
1266  */
1267 #define PERF_KVM__MAX_EVENTS_PER_MMAP  25
1268 
1269 static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
1270 				   u64 *mmap_time)
1271 {
1272 	struct evlist *evlist = kvm->evlist;
1273 	union perf_event *event;
1274 	struct mmap *md;
1275 	u64 timestamp;
1276 	s64 n = 0;
1277 	int err;
1278 
1279 	*mmap_time = ULLONG_MAX;
1280 	md = &evlist->mmap[idx];
1281 	err = perf_mmap__read_init(&md->core);
1282 	if (err < 0)
1283 		return (err == -EAGAIN) ? 0 : -1;
1284 
1285 	while ((event = perf_mmap__read_event(&md->core)) != NULL) {
1286 		err = evlist__parse_sample_timestamp(evlist, event, &timestamp);
1287 		if (err) {
1288 			perf_mmap__consume(&md->core);
1289 			pr_err("Failed to parse sample\n");
1290 			return -1;
1291 		}
1292 
1293 		err = perf_session__queue_event(kvm->session, event, timestamp, 0, NULL);
1294 		/*
1295 		 * FIXME: Here we can't consume the event, as perf_session__queue_event will
1296 		 *        point to it, and it'll get possibly overwritten by the kernel.
1297 		 */
1298 		perf_mmap__consume(&md->core);
1299 
1300 		if (err) {
1301 			pr_err("Failed to enqueue sample: %d\n", err);
1302 			return -1;
1303 		}
1304 
1305 		/* save time stamp of our first sample for this mmap */
1306 		if (n == 0)
1307 			*mmap_time = timestamp;
1308 
1309 		/* limit events per mmap handled all at once */
1310 		n++;
1311 		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
1312 			break;
1313 	}
1314 
1315 	perf_mmap__read_done(&md->core);
1316 	return n;
1317 }
1318 
1319 static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
1320 {
1321 	int i, err, throttled = 0;
1322 	s64 n, ntotal = 0;
1323 	u64 flush_time = ULLONG_MAX, mmap_time;
1324 
1325 	for (i = 0; i < kvm->evlist->core.nr_mmaps; i++) {
1326 		n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
1327 		if (n < 0)
1328 			return -1;
1329 
1330 		/* flush time is going to be the minimum of all the individual
1331 		 * mmap times. Essentially, we flush all the samples queued up
1332 		 * from the last pass under our minimal start time -- that leaves
1333 		 * a very small race for samples to come in with a lower timestamp.
1334 		 * The ioctl to return the perf_clock timestamp should close the
1335 		 * race entirely.
1336 		 */
1337 		if (mmap_time < flush_time)
1338 			flush_time = mmap_time;
1339 
1340 		ntotal += n;
1341 		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
1342 			throttled = 1;
1343 	}
1344 
1345 	/* flush queue after each round in which we processed events */
1346 	if (ntotal) {
1347 		struct ordered_events *oe = &kvm->session->ordered_events;
1348 
1349 		oe->next_flush = flush_time;
1350 		err = ordered_events__flush(oe, OE_FLUSH__ROUND);
1351 		if (err) {
1352 			if (kvm->lost_events)
1353 				pr_info("\nLost events: %" PRIu64 "\n\n",
1354 					kvm->lost_events);
1355 			return err;
1356 		}
1357 	}
1358 
1359 	return throttled;
1360 }
1361 
1362 static volatile int done;
1363 
1364 static void sig_handler(int sig __maybe_unused)
1365 {
1366 	done = 1;
1367 }
1368 
1369 static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
1370 {
1371 	struct itimerspec new_value;
1372 	int rc = -1;
1373 
1374 	kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
1375 	if (kvm->timerfd < 0) {
1376 		pr_err("timerfd_create failed\n");
1377 		goto out;
1378 	}
1379 
1380 	new_value.it_value.tv_sec = kvm->display_time;
1381 	new_value.it_value.tv_nsec = 0;
1382 	new_value.it_interval.tv_sec = kvm->display_time;
1383 	new_value.it_interval.tv_nsec = 0;
1384 
1385 	if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
1386 		pr_err("timerfd_settime failed: %d\n", errno);
1387 		close(kvm->timerfd);
1388 		goto out;
1389 	}
1390 
1391 	rc = 0;
1392 out:
1393 	return rc;
1394 }
1395 
1396 static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
1397 {
1398 	uint64_t c;
1399 	int rc;
1400 
1401 	rc = read(kvm->timerfd, &c, sizeof(uint64_t));
1402 	if (rc < 0) {
1403 		if (errno == EAGAIN)
1404 			return 0;
1405 
1406 		pr_err("Failed to read timer fd: %d\n", errno);
1407 		return -1;
1408 	}
1409 
1410 	if (rc != sizeof(uint64_t)) {
1411 		pr_err("Error reading timer fd - invalid size returned\n");
1412 		return -1;
1413 	}
1414 
1415 	if (c != 1)
1416 		pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
1417 
1418 	/* update display */
1419 	sort_result(kvm);
1420 	print_result(kvm);
1421 
1422 	/* Reset sort list to "ev_name" */
1423 	kvm_hists__reinit(NULL, "ev_name");
1424 
1425 	/* reset counts */
1426 	clear_events_cache_stats();
1427 	kvm->total_count = 0;
1428 	kvm->total_time = 0;
1429 	kvm->lost_events = 0;
1430 
1431 	return 0;
1432 }
1433 
1434 static int fd_set_nonblock(int fd)
1435 {
1436 	long arg = 0;
1437 
1438 	arg = fcntl(fd, F_GETFL);
1439 	if (arg < 0) {
1440 		pr_err("Failed to get current flags for fd %d\n", fd);
1441 		return -1;
1442 	}
1443 
1444 	if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
1445 		pr_err("Failed to set non-block option on fd %d\n", fd);
1446 		return -1;
1447 	}
1448 
1449 	return 0;
1450 }
1451 
1452 static int perf_kvm__handle_stdin(void)
1453 {
1454 	int c;
1455 
1456 	c = getc(stdin);
1457 	if (c == 'q')
1458 		return 1;
1459 
1460 	return 0;
1461 }
1462 
1463 static int kvm_events_live_report(struct perf_kvm_stat *kvm)
1464 {
1465 	int nr_stdin, ret, err = -EINVAL;
1466 	struct termios save;
1467 
1468 	/* live flag must be set first */
1469 	kvm->live = true;
1470 
1471 	ret = cpu_isa_config(kvm);
1472 	if (ret < 0)
1473 		return ret;
1474 
1475 	if (!verify_vcpu(kvm->trace_vcpu) ||
1476 	    !is_valid_key(kvm) ||
1477 	    !register_kvm_events_ops(kvm)) {
1478 		goto out;
1479 	}
1480 
1481 	set_term_quiet_input(&save);
1482 
1483 	kvm_hists__init();
1484 
1485 	signal(SIGINT, sig_handler);
1486 	signal(SIGTERM, sig_handler);
1487 
1488 	/* add timer fd */
1489 	if (perf_kvm__timerfd_create(kvm) < 0) {
1490 		err = -1;
1491 		goto out;
1492 	}
1493 
1494 	if (evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
1495 		goto out;
1496 
1497 	nr_stdin = evlist__add_pollfd(kvm->evlist, fileno(stdin));
1498 	if (nr_stdin < 0)
1499 		goto out;
1500 
1501 	if (fd_set_nonblock(fileno(stdin)) != 0)
1502 		goto out;
1503 
1504 	/* everything is good - enable the events and process */
1505 	evlist__enable(kvm->evlist);
1506 
1507 	while (!done) {
1508 		struct fdarray *fda = &kvm->evlist->core.pollfd;
1509 		int rc;
1510 
1511 		rc = perf_kvm__mmap_read(kvm);
1512 		if (rc < 0)
1513 			break;
1514 
1515 		err = perf_kvm__handle_timerfd(kvm);
1516 		if (err)
1517 			goto out;
1518 
1519 		if (fda->entries[nr_stdin].revents & POLLIN)
1520 			done = perf_kvm__handle_stdin();
1521 
1522 		if (!rc && !done)
1523 			err = evlist__poll(kvm->evlist, 100);
1524 	}
1525 
1526 	evlist__disable(kvm->evlist);
1527 
1528 	if (err == 0) {
1529 		sort_result(kvm);
1530 		print_result(kvm);
1531 	}
1532 
1533 out:
1534 	hists__delete_entries(&kvm_hists.hists);
1535 
1536 	if (kvm->timerfd >= 0)
1537 		close(kvm->timerfd);
1538 
1539 	tcsetattr(0, TCSAFLUSH, &save);
1540 	return err;
1541 }
1542 
1543 static int kvm_live_open_events(struct perf_kvm_stat *kvm)
1544 {
1545 	int err, rc = -1;
1546 	struct evsel *pos;
1547 	struct evlist *evlist = kvm->evlist;
1548 	char sbuf[STRERR_BUFSIZE];
1549 
1550 	evlist__config(evlist, &kvm->opts, NULL);
1551 
1552 	/*
1553 	 * Note: exclude_{guest,host} do not apply here.
1554 	 *       This command processes KVM tracepoints from host only
1555 	 */
1556 	evlist__for_each_entry(evlist, pos) {
1557 		struct perf_event_attr *attr = &pos->core.attr;
1558 
1559 		/* make sure these *are* set */
1560 		evsel__set_sample_bit(pos, TID);
1561 		evsel__set_sample_bit(pos, TIME);
1562 		evsel__set_sample_bit(pos, CPU);
1563 		evsel__set_sample_bit(pos, RAW);
1564 		/* make sure these are *not*; want as small a sample as possible */
1565 		evsel__reset_sample_bit(pos, PERIOD);
1566 		evsel__reset_sample_bit(pos, IP);
1567 		evsel__reset_sample_bit(pos, CALLCHAIN);
1568 		evsel__reset_sample_bit(pos, ADDR);
1569 		evsel__reset_sample_bit(pos, READ);
1570 		attr->mmap = 0;
1571 		attr->comm = 0;
1572 		attr->task = 0;
1573 
1574 		attr->sample_period = 1;
1575 
1576 		attr->watermark = 0;
1577 		attr->wakeup_events = 1000;
1578 
1579 		/* will enable all once we are ready */
1580 		attr->disabled = 1;
1581 	}
1582 
1583 	err = evlist__open(evlist);
1584 	if (err < 0) {
1585 		printf("Couldn't create the events: %s\n",
1586 		       str_error_r(errno, sbuf, sizeof(sbuf)));
1587 		goto out;
1588 	}
1589 
1590 	if (evlist__mmap(evlist, kvm->opts.mmap_pages) < 0) {
1591 		ui__error("Failed to mmap the events: %s\n",
1592 			  str_error_r(errno, sbuf, sizeof(sbuf)));
1593 		evlist__close(evlist);
1594 		goto out;
1595 	}
1596 
1597 	rc = 0;
1598 
1599 out:
1600 	return rc;
1601 }
1602 #endif
1603 
1604 static int read_events(struct perf_kvm_stat *kvm)
1605 {
1606 	int ret;
1607 
1608 	struct perf_data file = {
1609 		.path  = kvm->file_name,
1610 		.mode  = PERF_DATA_MODE_READ,
1611 		.force = kvm->force,
1612 	};
1613 
1614 	perf_tool__init(&kvm->tool, /*ordered_events=*/true);
1615 	kvm->tool.sample	= process_sample_event;
1616 	kvm->tool.comm		= perf_event__process_comm;
1617 	kvm->tool.namespaces	= perf_event__process_namespaces;
1618 
1619 	kvm->session = perf_session__new(&file, &kvm->tool);
1620 	if (IS_ERR(kvm->session)) {
1621 		pr_err("Initializing perf session failed\n");
1622 		return PTR_ERR(kvm->session);
1623 	}
1624 
1625 	symbol__init(&kvm->session->header.env);
1626 
1627 	if (!perf_session__has_traces(kvm->session, "kvm record")) {
1628 		ret = -EINVAL;
1629 		goto out_delete;
1630 	}
1631 
1632 	/*
1633 	 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1634 	 * traced in the old kernel.
1635 	 */
1636 	ret = cpu_isa_config(kvm);
1637 	if (ret < 0)
1638 		goto out_delete;
1639 
1640 	ret = perf_session__process_events(kvm->session);
1641 
1642 out_delete:
1643 	perf_session__delete(kvm->session);
1644 	return ret;
1645 }
1646 
1647 static int parse_target_str(struct perf_kvm_stat *kvm)
1648 {
1649 	if (kvm->opts.target.pid) {
1650 		kvm->pid_list = intlist__new(kvm->opts.target.pid);
1651 		if (kvm->pid_list == NULL) {
1652 			pr_err("Error parsing process id string\n");
1653 			return -EINVAL;
1654 		}
1655 	}
1656 
1657 	return 0;
1658 }
1659 
1660 static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1661 {
1662 	int ret = -EINVAL;
1663 	int vcpu = kvm->trace_vcpu;
1664 
1665 	if (parse_target_str(kvm) != 0)
1666 		goto exit;
1667 
1668 	if (!verify_vcpu(vcpu))
1669 		goto exit;
1670 
1671 	if (!is_valid_key(kvm))
1672 		goto exit;
1673 
1674 	if (!register_kvm_events_ops(kvm))
1675 		goto exit;
1676 
1677 	if (kvm->use_stdio) {
1678 		use_browser = 0;
1679 		setup_pager();
1680 	} else {
1681 		use_browser = 1;
1682 	}
1683 
1684 	setup_browser(false);
1685 
1686 	kvm_hists__init();
1687 
1688 	ret = read_events(kvm);
1689 	if (ret)
1690 		goto exit;
1691 
1692 	sort_result(kvm);
1693 	kvm_display(kvm);
1694 
1695 exit:
1696 	hists__delete_entries(&kvm_hists.hists);
1697 	return ret;
1698 }
1699 
1700 #define STRDUP_FAIL_EXIT(s)		\
1701 	({	char *_p;		\
1702 	_p = strdup(s);		\
1703 		if (!_p)		\
1704 			return -ENOMEM;	\
1705 		_p;			\
1706 	})
1707 
1708 int __weak setup_kvm_events_tp(struct perf_kvm_stat *kvm __maybe_unused)
1709 {
1710 	return 0;
1711 }
1712 
1713 static int
1714 kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1715 {
1716 	unsigned int rec_argc, i, j, events_tp_size;
1717 	const char **rec_argv;
1718 	const char * const record_args[] = {
1719 		"record",
1720 		"-R",
1721 		"-m", "1024",
1722 		"-c", "1",
1723 	};
1724 	const char * const kvm_stat_record_usage[] = {
1725 		"perf kvm stat record [<options>]",
1726 		NULL
1727 	};
1728 	const char * const *events_tp;
1729 	int ret;
1730 
1731 	events_tp_size = 0;
1732 	ret = setup_kvm_events_tp(kvm);
1733 	if (ret < 0) {
1734 		pr_err("Unable to setup the kvm tracepoints\n");
1735 		return ret;
1736 	}
1737 
1738 	for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1739 		events_tp_size++;
1740 
1741 	rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1742 		   2 * events_tp_size;
1743 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1744 
1745 	if (rec_argv == NULL)
1746 		return -ENOMEM;
1747 
1748 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1749 		rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1750 
1751 	for (j = 0; j < events_tp_size; j++) {
1752 		rec_argv[i++] = "-e";
1753 		rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1754 	}
1755 
1756 	rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1757 	rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1758 
1759 	for (j = 1; j < (unsigned int)argc; j++, i++)
1760 		rec_argv[i] = argv[j];
1761 
1762 	set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1763 	set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1764 	set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1765 
1766 	set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1767 	set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1768 	set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1769 	set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1770 	set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1771 	set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1772 	set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1773 	set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1774 	set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1775 	set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1776 	set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1777 	set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1778 	set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1779 	set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1780 	set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1781 
1782 	record_usage = kvm_stat_record_usage;
1783 	return cmd_record(i, rec_argv);
1784 }
1785 
1786 static int
1787 kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1788 {
1789 	const struct option kvm_events_report_options[] = {
1790 		OPT_STRING(0, "event", &kvm->report_event, "report event",
1791 			   "event for reporting: vmexit, "
1792 			   "mmio (x86 only), ioport (x86 only)"),
1793 		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1794 			    "vcpu id to report"),
1795 		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1796 			    "key for sorting: sample(sort by samples number)"
1797 			    " time (sort by avg time)"),
1798 		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1799 			   "analyze events only for given process id(s)"),
1800 		OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
1801 		OPT_BOOLEAN(0, "stdio", &kvm->use_stdio, "use the stdio interface"),
1802 		OPT_END()
1803 	};
1804 
1805 	const char * const kvm_events_report_usage[] = {
1806 		"perf kvm stat report [<options>]",
1807 		NULL
1808 	};
1809 
1810 	if (argc) {
1811 		argc = parse_options(argc, argv,
1812 				     kvm_events_report_options,
1813 				     kvm_events_report_usage, 0);
1814 		if (argc)
1815 			usage_with_options(kvm_events_report_usage,
1816 					   kvm_events_report_options);
1817 	}
1818 
1819 #ifndef HAVE_SLANG_SUPPORT
1820 	kvm->use_stdio = true;
1821 #endif
1822 
1823 	if (!kvm->opts.target.pid)
1824 		kvm->opts.target.system_wide = true;
1825 
1826 	return kvm_events_report_vcpu(kvm);
1827 }
1828 
1829 #if defined(HAVE_TIMERFD_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
1830 static struct evlist *kvm_live_event_list(void)
1831 {
1832 	struct evlist *evlist;
1833 	char *tp, *name, *sys;
1834 	int err = -1;
1835 	const char * const *events_tp;
1836 
1837 	evlist = evlist__new();
1838 	if (evlist == NULL)
1839 		return NULL;
1840 
1841 	for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1842 
1843 		tp = strdup(*events_tp);
1844 		if (tp == NULL)
1845 			goto out;
1846 
1847 		/* split tracepoint into subsystem and name */
1848 		sys = tp;
1849 		name = strchr(tp, ':');
1850 		if (name == NULL) {
1851 			pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1852 			       *events_tp);
1853 			free(tp);
1854 			goto out;
1855 		}
1856 		*name = '\0';
1857 		name++;
1858 
1859 		if (evlist__add_newtp(evlist, sys, name, NULL)) {
1860 			pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1861 			free(tp);
1862 			goto out;
1863 		}
1864 
1865 		free(tp);
1866 	}
1867 
1868 	err = 0;
1869 
1870 out:
1871 	if (err) {
1872 		evlist__delete(evlist);
1873 		evlist = NULL;
1874 	}
1875 
1876 	return evlist;
1877 }
1878 
1879 static int kvm_events_live(struct perf_kvm_stat *kvm,
1880 			   int argc, const char **argv)
1881 {
1882 	char errbuf[BUFSIZ];
1883 	int err;
1884 
1885 	const struct option live_options[] = {
1886 		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1887 			"record events on existing process id"),
1888 		OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1889 			"number of mmap data pages", evlist__parse_mmap_pages),
1890 		OPT_INCR('v', "verbose", &verbose,
1891 			"be more verbose (show counter open errors, etc)"),
1892 		OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1893 			"system-wide collection from all CPUs"),
1894 		OPT_UINTEGER('d', "display", &kvm->display_time,
1895 			"time in seconds between display updates"),
1896 		OPT_STRING(0, "event", &kvm->report_event, "report event",
1897 			"event for reporting: "
1898 			"vmexit, mmio (x86 only), ioport (x86 only)"),
1899 		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1900 			"vcpu id to report"),
1901 		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1902 			"key for sorting: sample(sort by samples number)"
1903 			" time (sort by avg time)"),
1904 		OPT_U64(0, "duration", &kvm->duration,
1905 			"show events other than"
1906 			" HLT (x86 only) or Wait state (s390 only)"
1907 			" that take longer than duration usecs"),
1908 		OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1909 				"per thread proc mmap processing timeout in ms"),
1910 		OPT_END()
1911 	};
1912 	const char * const live_usage[] = {
1913 		"perf kvm stat live [<options>]",
1914 		NULL
1915 	};
1916 	struct perf_data data = {
1917 		.mode = PERF_DATA_MODE_WRITE,
1918 	};
1919 
1920 
1921 	/* event handling */
1922 	perf_tool__init(&kvm->tool, /*ordered_events=*/true);
1923 	kvm->tool.sample = process_sample_event;
1924 	kvm->tool.comm   = perf_event__process_comm;
1925 	kvm->tool.exit   = perf_event__process_exit;
1926 	kvm->tool.fork   = perf_event__process_fork;
1927 	kvm->tool.lost   = process_lost_event;
1928 	kvm->tool.namespaces  = perf_event__process_namespaces;
1929 
1930 	/* set defaults */
1931 	kvm->display_time = 1;
1932 	kvm->opts.user_interval = 1;
1933 	kvm->opts.mmap_pages = 512;
1934 	kvm->opts.target.uses_mmap = false;
1935 	kvm->opts.target.uid_str = NULL;
1936 	kvm->opts.target.uid = UINT_MAX;
1937 
1938 	symbol__init(NULL);
1939 	disable_buildid_cache();
1940 
1941 	use_browser = 0;
1942 
1943 	if (argc) {
1944 		argc = parse_options(argc, argv, live_options,
1945 				     live_usage, 0);
1946 		if (argc)
1947 			usage_with_options(live_usage, live_options);
1948 	}
1949 
1950 	kvm->duration *= NSEC_PER_USEC;   /* convert usec to nsec */
1951 
1952 	/*
1953 	 * target related setups
1954 	 */
1955 	err = target__validate(&kvm->opts.target);
1956 	if (err) {
1957 		target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1958 		ui__warning("%s", errbuf);
1959 	}
1960 
1961 	if (target__none(&kvm->opts.target))
1962 		kvm->opts.target.system_wide = true;
1963 
1964 
1965 	/*
1966 	 * generate the event list
1967 	 */
1968 	err = setup_kvm_events_tp(kvm);
1969 	if (err < 0) {
1970 		pr_err("Unable to setup the kvm tracepoints\n");
1971 		return err;
1972 	}
1973 
1974 	kvm->evlist = kvm_live_event_list();
1975 	if (kvm->evlist == NULL) {
1976 		err = -1;
1977 		goto out;
1978 	}
1979 
1980 	if (evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1981 		usage_with_options(live_usage, live_options);
1982 
1983 	/*
1984 	 * perf session
1985 	 */
1986 	kvm->session = perf_session__new(&data, &kvm->tool);
1987 	if (IS_ERR(kvm->session)) {
1988 		err = PTR_ERR(kvm->session);
1989 		goto out;
1990 	}
1991 	kvm->session->evlist = kvm->evlist;
1992 	perf_session__set_id_hdr_size(kvm->session);
1993 	ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
1994 	machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1995 				    kvm->evlist->core.threads, true, false, 1);
1996 	err = kvm_live_open_events(kvm);
1997 	if (err)
1998 		goto out;
1999 
2000 	err = kvm_events_live_report(kvm);
2001 
2002 out:
2003 	perf_session__delete(kvm->session);
2004 	kvm->session = NULL;
2005 	evlist__delete(kvm->evlist);
2006 
2007 	return err;
2008 }
2009 #endif
2010 
2011 static void print_kvm_stat_usage(void)
2012 {
2013 	printf("Usage: perf kvm stat <command>\n\n");
2014 
2015 	printf("# Available commands:\n");
2016 	printf("\trecord: record kvm events\n");
2017 	printf("\treport: report statistical data of kvm events\n");
2018 	printf("\tlive:   live reporting of statistical data of kvm events\n");
2019 
2020 	printf("\nOtherwise, it is the alias of 'perf stat':\n");
2021 }
2022 
2023 static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
2024 {
2025 	struct perf_kvm_stat kvm = {
2026 		.file_name = file_name,
2027 
2028 		.trace_vcpu	= -1,
2029 		.report_event	= "vmexit",
2030 		.sort_key	= "sample",
2031 
2032 	};
2033 
2034 	if (argc == 1) {
2035 		print_kvm_stat_usage();
2036 		goto perf_stat;
2037 	}
2038 
2039 	if (strlen(argv[1]) > 2 && strstarts("record", argv[1]))
2040 		return kvm_events_record(&kvm, argc - 1, argv + 1);
2041 
2042 	if (strlen(argv[1]) > 2 && strstarts("report", argv[1]))
2043 		return kvm_events_report(&kvm, argc - 1 , argv + 1);
2044 
2045 #if defined(HAVE_TIMERFD_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
2046 	if (!strncmp(argv[1], "live", 4))
2047 		return kvm_events_live(&kvm, argc - 1 , argv + 1);
2048 #endif
2049 
2050 perf_stat:
2051 	return cmd_stat(argc, argv);
2052 }
2053 #endif /* HAVE_KVM_STAT_SUPPORT */
2054 
2055 int __weak kvm_add_default_arch_event(int *argc __maybe_unused,
2056 					const char **argv __maybe_unused)
2057 {
2058 	return 0;
2059 }
2060 
2061 static int __cmd_record(const char *file_name, int argc, const char **argv)
2062 {
2063 	int rec_argc, i = 0, j, ret;
2064 	const char **rec_argv;
2065 
2066 	ret = kvm_add_default_arch_event(&argc, argv);
2067 	if (ret)
2068 		return -EINVAL;
2069 
2070 	rec_argc = argc + 2;
2071 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
2072 	rec_argv[i++] = strdup("record");
2073 	rec_argv[i++] = strdup("-o");
2074 	rec_argv[i++] = strdup(file_name);
2075 	for (j = 1; j < argc; j++, i++)
2076 		rec_argv[i] = argv[j];
2077 
2078 	BUG_ON(i != rec_argc);
2079 
2080 	return cmd_record(i, rec_argv);
2081 }
2082 
2083 static int __cmd_report(const char *file_name, int argc, const char **argv)
2084 {
2085 	int rec_argc, i = 0, j;
2086 	const char **rec_argv;
2087 
2088 	rec_argc = argc + 2;
2089 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
2090 	rec_argv[i++] = strdup("report");
2091 	rec_argv[i++] = strdup("-i");
2092 	rec_argv[i++] = strdup(file_name);
2093 	for (j = 1; j < argc; j++, i++)
2094 		rec_argv[i] = argv[j];
2095 
2096 	BUG_ON(i != rec_argc);
2097 
2098 	return cmd_report(i, rec_argv);
2099 }
2100 
2101 static int
2102 __cmd_buildid_list(const char *file_name, int argc, const char **argv)
2103 {
2104 	int rec_argc, i = 0, j;
2105 	const char **rec_argv;
2106 
2107 	rec_argc = argc + 2;
2108 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
2109 	rec_argv[i++] = strdup("buildid-list");
2110 	rec_argv[i++] = strdup("-i");
2111 	rec_argv[i++] = strdup(file_name);
2112 	for (j = 1; j < argc; j++, i++)
2113 		rec_argv[i] = argv[j];
2114 
2115 	BUG_ON(i != rec_argc);
2116 
2117 	return cmd_buildid_list(i, rec_argv);
2118 }
2119 
2120 int cmd_kvm(int argc, const char **argv)
2121 {
2122 	const char *file_name = NULL;
2123 	const struct option kvm_options[] = {
2124 		OPT_STRING('i', "input", &file_name, "file",
2125 			   "Input file name"),
2126 		OPT_STRING('o', "output", &file_name, "file",
2127 			   "Output file name"),
2128 		OPT_BOOLEAN(0, "guest", &perf_guest,
2129 			    "Collect guest os data"),
2130 		OPT_BOOLEAN(0, "host", &perf_host,
2131 			    "Collect host os data"),
2132 		OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
2133 			   "guest mount directory under which every guest os"
2134 			   " instance has a subdir"),
2135 		OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
2136 			   "file", "file saving guest os vmlinux"),
2137 		OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
2138 			   "file", "file saving guest os /proc/kallsyms"),
2139 		OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
2140 			   "file", "file saving guest os /proc/modules"),
2141 		OPT_BOOLEAN(0, "guest-code", &symbol_conf.guest_code,
2142 			    "Guest code can be found in hypervisor process"),
2143 		OPT_INCR('v', "verbose", &verbose,
2144 			    "be more verbose (show counter open errors, etc)"),
2145 		OPT_END()
2146 	};
2147 
2148 	const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
2149 						"buildid-list", "stat", NULL };
2150 	const char *kvm_usage[] = { NULL, NULL };
2151 
2152 	exclude_GH_default = true;
2153 	perf_host  = 0;
2154 	perf_guest = 1;
2155 
2156 	argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
2157 					PARSE_OPT_STOP_AT_NON_OPTION);
2158 	if (!argc)
2159 		usage_with_options(kvm_usage, kvm_options);
2160 
2161 	if (!perf_host)
2162 		perf_guest = 1;
2163 
2164 	if (!file_name) {
2165 		file_name = get_filename_for_perf_kvm();
2166 
2167 		if (!file_name) {
2168 			pr_err("Failed to allocate memory for filename\n");
2169 			return -ENOMEM;
2170 		}
2171 	}
2172 
2173 	if (strlen(argv[0]) > 2 && strstarts("record", argv[0]))
2174 		return __cmd_record(file_name, argc, argv);
2175 	else if (strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2176 		return __cmd_report(file_name, argc, argv);
2177 	else if (strlen(argv[0]) > 2 && strstarts("diff", argv[0]))
2178 		return cmd_diff(argc, argv);
2179 	else if (!strcmp(argv[0], "top"))
2180 		return cmd_top(argc, argv);
2181 	else if (strlen(argv[0]) > 2 && strstarts("buildid-list", argv[0]))
2182 		return __cmd_buildid_list(file_name, argc, argv);
2183 #if defined(HAVE_KVM_STAT_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
2184 	else if (strlen(argv[0]) > 2 && strstarts("stat", argv[0]))
2185 		return kvm_cmd_stat(file_name, argc, argv);
2186 #endif
2187 	else
2188 		usage_with_options(kvm_usage, kvm_options);
2189 
2190 	/* free usage string allocated by parse_options_subcommand */
2191 	free((void *)kvm_usage[0]);
2192 
2193 	return 0;
2194 }
2195