xref: /linux/tools/perf/builtin-kvm.c (revision fd639726bf15fca8ee1a00dce8e0096d0ad9bd18)
1 // SPDX-License-Identifier: GPL-2.0
2 #include "builtin.h"
3 #include "perf.h"
4 
5 #include "util/evsel.h"
6 #include "util/evlist.h"
7 #include "util/term.h"
8 #include "util/util.h"
9 #include "util/cache.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/parse-options.h>
16 #include "util/trace-event.h"
17 #include "util/debug.h"
18 #include "util/tool.h"
19 #include "util/stat.h"
20 #include "util/top.h"
21 #include "util/data.h"
22 #include "util/ordered-events.h"
23 
24 #include <sys/prctl.h>
25 #ifdef HAVE_TIMERFD_SUPPORT
26 #include <sys/timerfd.h>
27 #endif
28 #include <sys/time.h>
29 
30 #include <linux/kernel.h>
31 #include <linux/time64.h>
32 #include <errno.h>
33 #include <inttypes.h>
34 #include <poll.h>
35 #include <termios.h>
36 #include <semaphore.h>
37 #include <signal.h>
38 #include <math.h>
39 
40 static const char *get_filename_for_perf_kvm(void)
41 {
42 	const char *filename;
43 
44 	if (perf_host && !perf_guest)
45 		filename = strdup("perf.data.host");
46 	else if (!perf_host && perf_guest)
47 		filename = strdup("perf.data.guest");
48 	else
49 		filename = strdup("perf.data.kvm");
50 
51 	return filename;
52 }
53 
54 #ifdef HAVE_KVM_STAT_SUPPORT
55 #include "util/kvm-stat.h"
56 
57 void exit_event_get_key(struct perf_evsel *evsel,
58 			struct perf_sample *sample,
59 			struct event_key *key)
60 {
61 	key->info = 0;
62 	key->key = perf_evsel__intval(evsel, sample, kvm_exit_reason);
63 }
64 
65 bool kvm_exit_event(struct perf_evsel *evsel)
66 {
67 	return !strcmp(evsel->name, kvm_exit_trace);
68 }
69 
70 bool exit_event_begin(struct perf_evsel *evsel,
71 		      struct perf_sample *sample, struct event_key *key)
72 {
73 	if (kvm_exit_event(evsel)) {
74 		exit_event_get_key(evsel, sample, key);
75 		return true;
76 	}
77 
78 	return false;
79 }
80 
81 bool kvm_entry_event(struct perf_evsel *evsel)
82 {
83 	return !strcmp(evsel->name, kvm_entry_trace);
84 }
85 
86 bool exit_event_end(struct perf_evsel *evsel,
87 		    struct perf_sample *sample __maybe_unused,
88 		    struct event_key *key __maybe_unused)
89 {
90 	return kvm_entry_event(evsel);
91 }
92 
93 static const char *get_exit_reason(struct perf_kvm_stat *kvm,
94 				   struct exit_reasons_table *tbl,
95 				   u64 exit_code)
96 {
97 	while (tbl->reason != NULL) {
98 		if (tbl->exit_code == exit_code)
99 			return tbl->reason;
100 		tbl++;
101 	}
102 
103 	pr_err("unknown kvm exit code:%lld on %s\n",
104 		(unsigned long long)exit_code, kvm->exit_reasons_isa);
105 	return "UNKNOWN";
106 }
107 
108 void exit_event_decode_key(struct perf_kvm_stat *kvm,
109 			   struct event_key *key,
110 			   char *decode)
111 {
112 	const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
113 						  key->key);
114 
115 	scnprintf(decode, decode_str_len, "%s", exit_reason);
116 }
117 
118 static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
119 {
120 	struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
121 
122 	for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
123 		if (!strcmp(events_ops->name, kvm->report_event)) {
124 			kvm->events_ops = events_ops->ops;
125 			return true;
126 		}
127 	}
128 
129 	return false;
130 }
131 
132 struct vcpu_event_record {
133 	int vcpu_id;
134 	u64 start_time;
135 	struct kvm_event *last_event;
136 };
137 
138 
139 static void init_kvm_event_record(struct perf_kvm_stat *kvm)
140 {
141 	unsigned int i;
142 
143 	for (i = 0; i < EVENTS_CACHE_SIZE; i++)
144 		INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
145 }
146 
147 #ifdef HAVE_TIMERFD_SUPPORT
148 static void clear_events_cache_stats(struct list_head *kvm_events_cache)
149 {
150 	struct list_head *head;
151 	struct kvm_event *event;
152 	unsigned int i;
153 	int j;
154 
155 	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
156 		head = &kvm_events_cache[i];
157 		list_for_each_entry(event, head, hash_entry) {
158 			/* reset stats for event */
159 			event->total.time = 0;
160 			init_stats(&event->total.stats);
161 
162 			for (j = 0; j < event->max_vcpu; ++j) {
163 				event->vcpu[j].time = 0;
164 				init_stats(&event->vcpu[j].stats);
165 			}
166 		}
167 	}
168 }
169 #endif
170 
171 static int kvm_events_hash_fn(u64 key)
172 {
173 	return key & (EVENTS_CACHE_SIZE - 1);
174 }
175 
176 static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
177 {
178 	int old_max_vcpu = event->max_vcpu;
179 	void *prev;
180 
181 	if (vcpu_id < event->max_vcpu)
182 		return true;
183 
184 	while (event->max_vcpu <= vcpu_id)
185 		event->max_vcpu += DEFAULT_VCPU_NUM;
186 
187 	prev = event->vcpu;
188 	event->vcpu = realloc(event->vcpu,
189 			      event->max_vcpu * sizeof(*event->vcpu));
190 	if (!event->vcpu) {
191 		free(prev);
192 		pr_err("Not enough memory\n");
193 		return false;
194 	}
195 
196 	memset(event->vcpu + old_max_vcpu, 0,
197 	       (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
198 	return true;
199 }
200 
201 static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
202 {
203 	struct kvm_event *event;
204 
205 	event = zalloc(sizeof(*event));
206 	if (!event) {
207 		pr_err("Not enough memory\n");
208 		return NULL;
209 	}
210 
211 	event->key = *key;
212 	init_stats(&event->total.stats);
213 	return event;
214 }
215 
216 static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
217 					       struct event_key *key)
218 {
219 	struct kvm_event *event;
220 	struct list_head *head;
221 
222 	BUG_ON(key->key == INVALID_KEY);
223 
224 	head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
225 	list_for_each_entry(event, head, hash_entry) {
226 		if (event->key.key == key->key && event->key.info == key->info)
227 			return event;
228 	}
229 
230 	event = kvm_alloc_init_event(key);
231 	if (!event)
232 		return NULL;
233 
234 	list_add(&event->hash_entry, head);
235 	return event;
236 }
237 
238 static bool handle_begin_event(struct perf_kvm_stat *kvm,
239 			       struct vcpu_event_record *vcpu_record,
240 			       struct event_key *key, u64 timestamp)
241 {
242 	struct kvm_event *event = NULL;
243 
244 	if (key->key != INVALID_KEY)
245 		event = find_create_kvm_event(kvm, key);
246 
247 	vcpu_record->last_event = event;
248 	vcpu_record->start_time = timestamp;
249 	return true;
250 }
251 
252 static void
253 kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
254 {
255 	kvm_stats->time += time_diff;
256 	update_stats(&kvm_stats->stats, time_diff);
257 }
258 
259 static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
260 {
261 	struct kvm_event_stats *kvm_stats = &event->total;
262 
263 	if (vcpu_id != -1)
264 		kvm_stats = &event->vcpu[vcpu_id];
265 
266 	return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
267 				avg_stats(&kvm_stats->stats));
268 }
269 
270 static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
271 			     u64 time_diff)
272 {
273 	if (vcpu_id == -1) {
274 		kvm_update_event_stats(&event->total, time_diff);
275 		return true;
276 	}
277 
278 	if (!kvm_event_expand(event, vcpu_id))
279 		return false;
280 
281 	kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
282 	return true;
283 }
284 
285 static bool is_child_event(struct perf_kvm_stat *kvm,
286 			   struct perf_evsel *evsel,
287 			   struct perf_sample *sample,
288 			   struct event_key *key)
289 {
290 	struct child_event_ops *child_ops;
291 
292 	child_ops = kvm->events_ops->child_ops;
293 
294 	if (!child_ops)
295 		return false;
296 
297 	for (; child_ops->name; child_ops++) {
298 		if (!strcmp(evsel->name, child_ops->name)) {
299 			child_ops->get_key(evsel, sample, key);
300 			return true;
301 		}
302 	}
303 
304 	return false;
305 }
306 
307 static bool handle_child_event(struct perf_kvm_stat *kvm,
308 			       struct vcpu_event_record *vcpu_record,
309 			       struct event_key *key,
310 			       struct perf_sample *sample __maybe_unused)
311 {
312 	struct kvm_event *event = NULL;
313 
314 	if (key->key != INVALID_KEY)
315 		event = find_create_kvm_event(kvm, key);
316 
317 	vcpu_record->last_event = event;
318 
319 	return true;
320 }
321 
322 static bool skip_event(const char *event)
323 {
324 	const char * const *skip_events;
325 
326 	for (skip_events = kvm_skip_events; *skip_events; skip_events++)
327 		if (!strcmp(event, *skip_events))
328 			return true;
329 
330 	return false;
331 }
332 
333 static bool handle_end_event(struct perf_kvm_stat *kvm,
334 			     struct vcpu_event_record *vcpu_record,
335 			     struct event_key *key,
336 			     struct perf_sample *sample)
337 {
338 	struct kvm_event *event;
339 	u64 time_begin, time_diff;
340 	int vcpu;
341 
342 	if (kvm->trace_vcpu == -1)
343 		vcpu = -1;
344 	else
345 		vcpu = vcpu_record->vcpu_id;
346 
347 	event = vcpu_record->last_event;
348 	time_begin = vcpu_record->start_time;
349 
350 	/* The begin event is not caught. */
351 	if (!time_begin)
352 		return true;
353 
354 	/*
355 	 * In some case, the 'begin event' only records the start timestamp,
356 	 * the actual event is recognized in the 'end event' (e.g. mmio-event).
357 	 */
358 
359 	/* Both begin and end events did not get the key. */
360 	if (!event && key->key == INVALID_KEY)
361 		return true;
362 
363 	if (!event)
364 		event = find_create_kvm_event(kvm, key);
365 
366 	if (!event)
367 		return false;
368 
369 	vcpu_record->last_event = NULL;
370 	vcpu_record->start_time = 0;
371 
372 	/* seems to happen once in a while during live mode */
373 	if (sample->time < time_begin) {
374 		pr_debug("End time before begin time; skipping event.\n");
375 		return true;
376 	}
377 
378 	time_diff = sample->time - time_begin;
379 
380 	if (kvm->duration && time_diff > kvm->duration) {
381 		char decode[decode_str_len];
382 
383 		kvm->events_ops->decode_key(kvm, &event->key, decode);
384 		if (!skip_event(decode)) {
385 			pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
386 				 sample->time, sample->pid, vcpu_record->vcpu_id,
387 				 decode, time_diff / NSEC_PER_USEC);
388 		}
389 	}
390 
391 	return update_kvm_event(event, vcpu, time_diff);
392 }
393 
394 static
395 struct vcpu_event_record *per_vcpu_record(struct thread *thread,
396 					  struct perf_evsel *evsel,
397 					  struct perf_sample *sample)
398 {
399 	/* Only kvm_entry records vcpu id. */
400 	if (!thread__priv(thread) && kvm_entry_event(evsel)) {
401 		struct vcpu_event_record *vcpu_record;
402 
403 		vcpu_record = zalloc(sizeof(*vcpu_record));
404 		if (!vcpu_record) {
405 			pr_err("%s: Not enough memory\n", __func__);
406 			return NULL;
407 		}
408 
409 		vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample,
410 							  vcpu_id_str);
411 		thread__set_priv(thread, vcpu_record);
412 	}
413 
414 	return thread__priv(thread);
415 }
416 
417 static bool handle_kvm_event(struct perf_kvm_stat *kvm,
418 			     struct thread *thread,
419 			     struct perf_evsel *evsel,
420 			     struct perf_sample *sample)
421 {
422 	struct vcpu_event_record *vcpu_record;
423 	struct event_key key = { .key = INVALID_KEY,
424 				 .exit_reasons = kvm->exit_reasons };
425 
426 	vcpu_record = per_vcpu_record(thread, evsel, sample);
427 	if (!vcpu_record)
428 		return true;
429 
430 	/* only process events for vcpus user cares about */
431 	if ((kvm->trace_vcpu != -1) &&
432 	    (kvm->trace_vcpu != vcpu_record->vcpu_id))
433 		return true;
434 
435 	if (kvm->events_ops->is_begin_event(evsel, sample, &key))
436 		return handle_begin_event(kvm, vcpu_record, &key, sample->time);
437 
438 	if (is_child_event(kvm, evsel, sample, &key))
439 		return handle_child_event(kvm, vcpu_record, &key, sample);
440 
441 	if (kvm->events_ops->is_end_event(evsel, sample, &key))
442 		return handle_end_event(kvm, vcpu_record, &key, sample);
443 
444 	return true;
445 }
446 
447 #define GET_EVENT_KEY(func, field)					\
448 static u64 get_event_ ##func(struct kvm_event *event, int vcpu)		\
449 {									\
450 	if (vcpu == -1)							\
451 		return event->total.field;				\
452 									\
453 	if (vcpu >= event->max_vcpu)					\
454 		return 0;						\
455 									\
456 	return event->vcpu[vcpu].field;					\
457 }
458 
459 #define COMPARE_EVENT_KEY(func, field)					\
460 GET_EVENT_KEY(func, field)						\
461 static int compare_kvm_event_ ## func(struct kvm_event *one,		\
462 					struct kvm_event *two, int vcpu)\
463 {									\
464 	return get_event_ ##func(one, vcpu) >				\
465 				get_event_ ##func(two, vcpu);		\
466 }
467 
468 GET_EVENT_KEY(time, time);
469 COMPARE_EVENT_KEY(count, stats.n);
470 COMPARE_EVENT_KEY(mean, stats.mean);
471 GET_EVENT_KEY(max, stats.max);
472 GET_EVENT_KEY(min, stats.min);
473 
474 #define DEF_SORT_NAME_KEY(name, compare_key)				\
475 	{ #name, compare_kvm_event_ ## compare_key }
476 
477 static struct kvm_event_key keys[] = {
478 	DEF_SORT_NAME_KEY(sample, count),
479 	DEF_SORT_NAME_KEY(time, mean),
480 	{ NULL, NULL }
481 };
482 
483 static bool select_key(struct perf_kvm_stat *kvm)
484 {
485 	int i;
486 
487 	for (i = 0; keys[i].name; i++) {
488 		if (!strcmp(keys[i].name, kvm->sort_key)) {
489 			kvm->compare = keys[i].key;
490 			return true;
491 		}
492 	}
493 
494 	pr_err("Unknown compare key:%s\n", kvm->sort_key);
495 	return false;
496 }
497 
498 static void insert_to_result(struct rb_root *result, struct kvm_event *event,
499 			     key_cmp_fun bigger, int vcpu)
500 {
501 	struct rb_node **rb = &result->rb_node;
502 	struct rb_node *parent = NULL;
503 	struct kvm_event *p;
504 
505 	while (*rb) {
506 		p = container_of(*rb, struct kvm_event, rb);
507 		parent = *rb;
508 
509 		if (bigger(event, p, vcpu))
510 			rb = &(*rb)->rb_left;
511 		else
512 			rb = &(*rb)->rb_right;
513 	}
514 
515 	rb_link_node(&event->rb, parent, rb);
516 	rb_insert_color(&event->rb, result);
517 }
518 
519 static void
520 update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
521 {
522 	int vcpu = kvm->trace_vcpu;
523 
524 	kvm->total_count += get_event_count(event, vcpu);
525 	kvm->total_time += get_event_time(event, vcpu);
526 }
527 
528 static bool event_is_valid(struct kvm_event *event, int vcpu)
529 {
530 	return !!get_event_count(event, vcpu);
531 }
532 
533 static void sort_result(struct perf_kvm_stat *kvm)
534 {
535 	unsigned int i;
536 	int vcpu = kvm->trace_vcpu;
537 	struct kvm_event *event;
538 
539 	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
540 		list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
541 			if (event_is_valid(event, vcpu)) {
542 				update_total_count(kvm, event);
543 				insert_to_result(&kvm->result, event,
544 						 kvm->compare, vcpu);
545 			}
546 		}
547 	}
548 }
549 
550 /* returns left most element of result, and erase it */
551 static struct kvm_event *pop_from_result(struct rb_root *result)
552 {
553 	struct rb_node *node = rb_first(result);
554 
555 	if (!node)
556 		return NULL;
557 
558 	rb_erase(node, result);
559 	return container_of(node, struct kvm_event, rb);
560 }
561 
562 static void print_vcpu_info(struct perf_kvm_stat *kvm)
563 {
564 	int vcpu = kvm->trace_vcpu;
565 
566 	pr_info("Analyze events for ");
567 
568 	if (kvm->opts.target.system_wide)
569 		pr_info("all VMs, ");
570 	else if (kvm->opts.target.pid)
571 		pr_info("pid(s) %s, ", kvm->opts.target.pid);
572 	else
573 		pr_info("dazed and confused on what is monitored, ");
574 
575 	if (vcpu == -1)
576 		pr_info("all VCPUs:\n\n");
577 	else
578 		pr_info("VCPU %d:\n\n", vcpu);
579 }
580 
581 static void show_timeofday(void)
582 {
583 	char date[64];
584 	struct timeval tv;
585 	struct tm ltime;
586 
587 	gettimeofday(&tv, NULL);
588 	if (localtime_r(&tv.tv_sec, &ltime)) {
589 		strftime(date, sizeof(date), "%H:%M:%S", &ltime);
590 		pr_info("%s.%06ld", date, tv.tv_usec);
591 	} else
592 		pr_info("00:00:00.000000");
593 
594 	return;
595 }
596 
597 static void print_result(struct perf_kvm_stat *kvm)
598 {
599 	char decode[decode_str_len];
600 	struct kvm_event *event;
601 	int vcpu = kvm->trace_vcpu;
602 
603 	if (kvm->live) {
604 		puts(CONSOLE_CLEAR);
605 		show_timeofday();
606 	}
607 
608 	pr_info("\n\n");
609 	print_vcpu_info(kvm);
610 	pr_info("%*s ", decode_str_len, kvm->events_ops->name);
611 	pr_info("%10s ", "Samples");
612 	pr_info("%9s ", "Samples%");
613 
614 	pr_info("%9s ", "Time%");
615 	pr_info("%11s ", "Min Time");
616 	pr_info("%11s ", "Max Time");
617 	pr_info("%16s ", "Avg time");
618 	pr_info("\n\n");
619 
620 	while ((event = pop_from_result(&kvm->result))) {
621 		u64 ecount, etime, max, min;
622 
623 		ecount = get_event_count(event, vcpu);
624 		etime = get_event_time(event, vcpu);
625 		max = get_event_max(event, vcpu);
626 		min = get_event_min(event, vcpu);
627 
628 		kvm->events_ops->decode_key(kvm, &event->key, decode);
629 		pr_info("%*s ", decode_str_len, decode);
630 		pr_info("%10llu ", (unsigned long long)ecount);
631 		pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
632 		pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
633 		pr_info("%9.2fus ", (double)min / NSEC_PER_USEC);
634 		pr_info("%9.2fus ", (double)max / NSEC_PER_USEC);
635 		pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount / NSEC_PER_USEC,
636 			kvm_event_rel_stddev(vcpu, event));
637 		pr_info("\n");
638 	}
639 
640 	pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
641 		kvm->total_count, kvm->total_time / (double)NSEC_PER_USEC);
642 
643 	if (kvm->lost_events)
644 		pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
645 }
646 
647 #ifdef HAVE_TIMERFD_SUPPORT
648 static int process_lost_event(struct perf_tool *tool,
649 			      union perf_event *event __maybe_unused,
650 			      struct perf_sample *sample __maybe_unused,
651 			      struct machine *machine __maybe_unused)
652 {
653 	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
654 
655 	kvm->lost_events++;
656 	return 0;
657 }
658 #endif
659 
660 static bool skip_sample(struct perf_kvm_stat *kvm,
661 			struct perf_sample *sample)
662 {
663 	if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
664 		return true;
665 
666 	return false;
667 }
668 
669 static int process_sample_event(struct perf_tool *tool,
670 				union perf_event *event,
671 				struct perf_sample *sample,
672 				struct perf_evsel *evsel,
673 				struct machine *machine)
674 {
675 	int err = 0;
676 	struct thread *thread;
677 	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
678 						 tool);
679 
680 	if (skip_sample(kvm, sample))
681 		return 0;
682 
683 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
684 	if (thread == NULL) {
685 		pr_debug("problem processing %d event, skipping it.\n",
686 			event->header.type);
687 		return -1;
688 	}
689 
690 	if (!handle_kvm_event(kvm, thread, evsel, sample))
691 		err = -1;
692 
693 	thread__put(thread);
694 	return err;
695 }
696 
697 static int cpu_isa_config(struct perf_kvm_stat *kvm)
698 {
699 	char buf[64], *cpuid;
700 	int err;
701 
702 	if (kvm->live) {
703 		err = get_cpuid(buf, sizeof(buf));
704 		if (err != 0) {
705 			pr_err("Failed to look up CPU type\n");
706 			return err;
707 		}
708 		cpuid = buf;
709 	} else
710 		cpuid = kvm->session->header.env.cpuid;
711 
712 	if (!cpuid) {
713 		pr_err("Failed to look up CPU type\n");
714 		return -EINVAL;
715 	}
716 
717 	err = cpu_isa_init(kvm, cpuid);
718 	if (err == -ENOTSUP)
719 		pr_err("CPU %s is not supported.\n", cpuid);
720 
721 	return err;
722 }
723 
724 static bool verify_vcpu(int vcpu)
725 {
726 	if (vcpu != -1 && vcpu < 0) {
727 		pr_err("Invalid vcpu:%d.\n", vcpu);
728 		return false;
729 	}
730 
731 	return true;
732 }
733 
734 #ifdef HAVE_TIMERFD_SUPPORT
735 /* keeping the max events to a modest level to keep
736  * the processing of samples per mmap smooth.
737  */
738 #define PERF_KVM__MAX_EVENTS_PER_MMAP  25
739 
740 static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
741 				   u64 *mmap_time)
742 {
743 	union perf_event *event;
744 	struct perf_sample sample;
745 	s64 n = 0;
746 	int err;
747 
748 	*mmap_time = ULLONG_MAX;
749 	while ((event = perf_evlist__mmap_read(kvm->evlist, idx)) != NULL) {
750 		err = perf_evlist__parse_sample(kvm->evlist, event, &sample);
751 		if (err) {
752 			perf_evlist__mmap_consume(kvm->evlist, idx);
753 			pr_err("Failed to parse sample\n");
754 			return -1;
755 		}
756 
757 		err = perf_session__queue_event(kvm->session, event, &sample, 0);
758 		/*
759 		 * FIXME: Here we can't consume the event, as perf_session__queue_event will
760 		 *        point to it, and it'll get possibly overwritten by the kernel.
761 		 */
762 		perf_evlist__mmap_consume(kvm->evlist, idx);
763 
764 		if (err) {
765 			pr_err("Failed to enqueue sample: %d\n", err);
766 			return -1;
767 		}
768 
769 		/* save time stamp of our first sample for this mmap */
770 		if (n == 0)
771 			*mmap_time = sample.time;
772 
773 		/* limit events per mmap handled all at once */
774 		n++;
775 		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
776 			break;
777 	}
778 
779 	return n;
780 }
781 
782 static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
783 {
784 	int i, err, throttled = 0;
785 	s64 n, ntotal = 0;
786 	u64 flush_time = ULLONG_MAX, mmap_time;
787 
788 	for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
789 		n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
790 		if (n < 0)
791 			return -1;
792 
793 		/* flush time is going to be the minimum of all the individual
794 		 * mmap times. Essentially, we flush all the samples queued up
795 		 * from the last pass under our minimal start time -- that leaves
796 		 * a very small race for samples to come in with a lower timestamp.
797 		 * The ioctl to return the perf_clock timestamp should close the
798 		 * race entirely.
799 		 */
800 		if (mmap_time < flush_time)
801 			flush_time = mmap_time;
802 
803 		ntotal += n;
804 		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
805 			throttled = 1;
806 	}
807 
808 	/* flush queue after each round in which we processed events */
809 	if (ntotal) {
810 		struct ordered_events *oe = &kvm->session->ordered_events;
811 
812 		oe->next_flush = flush_time;
813 		err = ordered_events__flush(oe, OE_FLUSH__ROUND);
814 		if (err) {
815 			if (kvm->lost_events)
816 				pr_info("\nLost events: %" PRIu64 "\n\n",
817 					kvm->lost_events);
818 			return err;
819 		}
820 	}
821 
822 	return throttled;
823 }
824 
825 static volatile int done;
826 
827 static void sig_handler(int sig __maybe_unused)
828 {
829 	done = 1;
830 }
831 
832 static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
833 {
834 	struct itimerspec new_value;
835 	int rc = -1;
836 
837 	kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
838 	if (kvm->timerfd < 0) {
839 		pr_err("timerfd_create failed\n");
840 		goto out;
841 	}
842 
843 	new_value.it_value.tv_sec = kvm->display_time;
844 	new_value.it_value.tv_nsec = 0;
845 	new_value.it_interval.tv_sec = kvm->display_time;
846 	new_value.it_interval.tv_nsec = 0;
847 
848 	if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
849 		pr_err("timerfd_settime failed: %d\n", errno);
850 		close(kvm->timerfd);
851 		goto out;
852 	}
853 
854 	rc = 0;
855 out:
856 	return rc;
857 }
858 
859 static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
860 {
861 	uint64_t c;
862 	int rc;
863 
864 	rc = read(kvm->timerfd, &c, sizeof(uint64_t));
865 	if (rc < 0) {
866 		if (errno == EAGAIN)
867 			return 0;
868 
869 		pr_err("Failed to read timer fd: %d\n", errno);
870 		return -1;
871 	}
872 
873 	if (rc != sizeof(uint64_t)) {
874 		pr_err("Error reading timer fd - invalid size returned\n");
875 		return -1;
876 	}
877 
878 	if (c != 1)
879 		pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
880 
881 	/* update display */
882 	sort_result(kvm);
883 	print_result(kvm);
884 
885 	/* reset counts */
886 	clear_events_cache_stats(kvm->kvm_events_cache);
887 	kvm->total_count = 0;
888 	kvm->total_time = 0;
889 	kvm->lost_events = 0;
890 
891 	return 0;
892 }
893 
894 static int fd_set_nonblock(int fd)
895 {
896 	long arg = 0;
897 
898 	arg = fcntl(fd, F_GETFL);
899 	if (arg < 0) {
900 		pr_err("Failed to get current flags for fd %d\n", fd);
901 		return -1;
902 	}
903 
904 	if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
905 		pr_err("Failed to set non-block option on fd %d\n", fd);
906 		return -1;
907 	}
908 
909 	return 0;
910 }
911 
912 static int perf_kvm__handle_stdin(void)
913 {
914 	int c;
915 
916 	c = getc(stdin);
917 	if (c == 'q')
918 		return 1;
919 
920 	return 0;
921 }
922 
923 static int kvm_events_live_report(struct perf_kvm_stat *kvm)
924 {
925 	int nr_stdin, ret, err = -EINVAL;
926 	struct termios save;
927 
928 	/* live flag must be set first */
929 	kvm->live = true;
930 
931 	ret = cpu_isa_config(kvm);
932 	if (ret < 0)
933 		return ret;
934 
935 	if (!verify_vcpu(kvm->trace_vcpu) ||
936 	    !select_key(kvm) ||
937 	    !register_kvm_events_ops(kvm)) {
938 		goto out;
939 	}
940 
941 	set_term_quiet_input(&save);
942 	init_kvm_event_record(kvm);
943 
944 	signal(SIGINT, sig_handler);
945 	signal(SIGTERM, sig_handler);
946 
947 	/* add timer fd */
948 	if (perf_kvm__timerfd_create(kvm) < 0) {
949 		err = -1;
950 		goto out;
951 	}
952 
953 	if (perf_evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
954 		goto out;
955 
956 	nr_stdin = perf_evlist__add_pollfd(kvm->evlist, fileno(stdin));
957 	if (nr_stdin < 0)
958 		goto out;
959 
960 	if (fd_set_nonblock(fileno(stdin)) != 0)
961 		goto out;
962 
963 	/* everything is good - enable the events and process */
964 	perf_evlist__enable(kvm->evlist);
965 
966 	while (!done) {
967 		struct fdarray *fda = &kvm->evlist->pollfd;
968 		int rc;
969 
970 		rc = perf_kvm__mmap_read(kvm);
971 		if (rc < 0)
972 			break;
973 
974 		err = perf_kvm__handle_timerfd(kvm);
975 		if (err)
976 			goto out;
977 
978 		if (fda->entries[nr_stdin].revents & POLLIN)
979 			done = perf_kvm__handle_stdin();
980 
981 		if (!rc && !done)
982 			err = fdarray__poll(fda, 100);
983 	}
984 
985 	perf_evlist__disable(kvm->evlist);
986 
987 	if (err == 0) {
988 		sort_result(kvm);
989 		print_result(kvm);
990 	}
991 
992 out:
993 	if (kvm->timerfd >= 0)
994 		close(kvm->timerfd);
995 
996 	tcsetattr(0, TCSAFLUSH, &save);
997 	return err;
998 }
999 
1000 static int kvm_live_open_events(struct perf_kvm_stat *kvm)
1001 {
1002 	int err, rc = -1;
1003 	struct perf_evsel *pos;
1004 	struct perf_evlist *evlist = kvm->evlist;
1005 	char sbuf[STRERR_BUFSIZE];
1006 
1007 	perf_evlist__config(evlist, &kvm->opts, NULL);
1008 
1009 	/*
1010 	 * Note: exclude_{guest,host} do not apply here.
1011 	 *       This command processes KVM tracepoints from host only
1012 	 */
1013 	evlist__for_each_entry(evlist, pos) {
1014 		struct perf_event_attr *attr = &pos->attr;
1015 
1016 		/* make sure these *are* set */
1017 		perf_evsel__set_sample_bit(pos, TID);
1018 		perf_evsel__set_sample_bit(pos, TIME);
1019 		perf_evsel__set_sample_bit(pos, CPU);
1020 		perf_evsel__set_sample_bit(pos, RAW);
1021 		/* make sure these are *not*; want as small a sample as possible */
1022 		perf_evsel__reset_sample_bit(pos, PERIOD);
1023 		perf_evsel__reset_sample_bit(pos, IP);
1024 		perf_evsel__reset_sample_bit(pos, CALLCHAIN);
1025 		perf_evsel__reset_sample_bit(pos, ADDR);
1026 		perf_evsel__reset_sample_bit(pos, READ);
1027 		attr->mmap = 0;
1028 		attr->comm = 0;
1029 		attr->task = 0;
1030 
1031 		attr->sample_period = 1;
1032 
1033 		attr->watermark = 0;
1034 		attr->wakeup_events = 1000;
1035 
1036 		/* will enable all once we are ready */
1037 		attr->disabled = 1;
1038 	}
1039 
1040 	err = perf_evlist__open(evlist);
1041 	if (err < 0) {
1042 		printf("Couldn't create the events: %s\n",
1043 		       str_error_r(errno, sbuf, sizeof(sbuf)));
1044 		goto out;
1045 	}
1046 
1047 	if (perf_evlist__mmap(evlist, kvm->opts.mmap_pages, false) < 0) {
1048 		ui__error("Failed to mmap the events: %s\n",
1049 			  str_error_r(errno, sbuf, sizeof(sbuf)));
1050 		perf_evlist__close(evlist);
1051 		goto out;
1052 	}
1053 
1054 	rc = 0;
1055 
1056 out:
1057 	return rc;
1058 }
1059 #endif
1060 
1061 static int read_events(struct perf_kvm_stat *kvm)
1062 {
1063 	int ret;
1064 
1065 	struct perf_tool eops = {
1066 		.sample			= process_sample_event,
1067 		.comm			= perf_event__process_comm,
1068 		.namespaces		= perf_event__process_namespaces,
1069 		.ordered_events		= true,
1070 	};
1071 	struct perf_data file = {
1072 		.file      = {
1073 			.path = kvm->file_name,
1074 		},
1075 		.mode      = PERF_DATA_MODE_READ,
1076 		.force     = kvm->force,
1077 	};
1078 
1079 	kvm->tool = eops;
1080 	kvm->session = perf_session__new(&file, false, &kvm->tool);
1081 	if (!kvm->session) {
1082 		pr_err("Initializing perf session failed\n");
1083 		return -1;
1084 	}
1085 
1086 	symbol__init(&kvm->session->header.env);
1087 
1088 	if (!perf_session__has_traces(kvm->session, "kvm record")) {
1089 		ret = -EINVAL;
1090 		goto out_delete;
1091 	}
1092 
1093 	/*
1094 	 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1095 	 * traced in the old kernel.
1096 	 */
1097 	ret = cpu_isa_config(kvm);
1098 	if (ret < 0)
1099 		goto out_delete;
1100 
1101 	ret = perf_session__process_events(kvm->session);
1102 
1103 out_delete:
1104 	perf_session__delete(kvm->session);
1105 	return ret;
1106 }
1107 
1108 static int parse_target_str(struct perf_kvm_stat *kvm)
1109 {
1110 	if (kvm->opts.target.pid) {
1111 		kvm->pid_list = intlist__new(kvm->opts.target.pid);
1112 		if (kvm->pid_list == NULL) {
1113 			pr_err("Error parsing process id string\n");
1114 			return -EINVAL;
1115 		}
1116 	}
1117 
1118 	return 0;
1119 }
1120 
1121 static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1122 {
1123 	int ret = -EINVAL;
1124 	int vcpu = kvm->trace_vcpu;
1125 
1126 	if (parse_target_str(kvm) != 0)
1127 		goto exit;
1128 
1129 	if (!verify_vcpu(vcpu))
1130 		goto exit;
1131 
1132 	if (!select_key(kvm))
1133 		goto exit;
1134 
1135 	if (!register_kvm_events_ops(kvm))
1136 		goto exit;
1137 
1138 	init_kvm_event_record(kvm);
1139 	setup_pager();
1140 
1141 	ret = read_events(kvm);
1142 	if (ret)
1143 		goto exit;
1144 
1145 	sort_result(kvm);
1146 	print_result(kvm);
1147 
1148 exit:
1149 	return ret;
1150 }
1151 
1152 #define STRDUP_FAIL_EXIT(s)		\
1153 	({	char *_p;		\
1154 	_p = strdup(s);		\
1155 		if (!_p)		\
1156 			return -ENOMEM;	\
1157 		_p;			\
1158 	})
1159 
1160 int __weak setup_kvm_events_tp(struct perf_kvm_stat *kvm __maybe_unused)
1161 {
1162 	return 0;
1163 }
1164 
1165 static int
1166 kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1167 {
1168 	unsigned int rec_argc, i, j, events_tp_size;
1169 	const char **rec_argv;
1170 	const char * const record_args[] = {
1171 		"record",
1172 		"-R",
1173 		"-m", "1024",
1174 		"-c", "1",
1175 	};
1176 	const char * const kvm_stat_record_usage[] = {
1177 		"perf kvm stat record [<options>]",
1178 		NULL
1179 	};
1180 	const char * const *events_tp;
1181 	int ret;
1182 
1183 	events_tp_size = 0;
1184 	ret = setup_kvm_events_tp(kvm);
1185 	if (ret < 0) {
1186 		pr_err("Unable to setup the kvm tracepoints\n");
1187 		return ret;
1188 	}
1189 
1190 	for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1191 		events_tp_size++;
1192 
1193 	rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1194 		   2 * events_tp_size;
1195 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1196 
1197 	if (rec_argv == NULL)
1198 		return -ENOMEM;
1199 
1200 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1201 		rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1202 
1203 	for (j = 0; j < events_tp_size; j++) {
1204 		rec_argv[i++] = "-e";
1205 		rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1206 	}
1207 
1208 	rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1209 	rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1210 
1211 	for (j = 1; j < (unsigned int)argc; j++, i++)
1212 		rec_argv[i] = argv[j];
1213 
1214 	set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1215 	set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1216 	set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1217 
1218 	set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1219 	set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1220 	set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1221 	set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1222 	set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1223 	set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1224 	set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1225 	set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1226 	set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1227 	set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1228 	set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1229 	set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1230 	set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1231 	set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1232 	set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1233 
1234 	record_usage = kvm_stat_record_usage;
1235 	return cmd_record(i, rec_argv);
1236 }
1237 
1238 static int
1239 kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1240 {
1241 	const struct option kvm_events_report_options[] = {
1242 		OPT_STRING(0, "event", &kvm->report_event, "report event",
1243 			   "event for reporting: vmexit, "
1244 			   "mmio (x86 only), ioport (x86 only)"),
1245 		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1246 			    "vcpu id to report"),
1247 		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1248 			    "key for sorting: sample(sort by samples number)"
1249 			    " time (sort by avg time)"),
1250 		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1251 			   "analyze events only for given process id(s)"),
1252 		OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
1253 		OPT_END()
1254 	};
1255 
1256 	const char * const kvm_events_report_usage[] = {
1257 		"perf kvm stat report [<options>]",
1258 		NULL
1259 	};
1260 
1261 	if (argc) {
1262 		argc = parse_options(argc, argv,
1263 				     kvm_events_report_options,
1264 				     kvm_events_report_usage, 0);
1265 		if (argc)
1266 			usage_with_options(kvm_events_report_usage,
1267 					   kvm_events_report_options);
1268 	}
1269 
1270 	if (!kvm->opts.target.pid)
1271 		kvm->opts.target.system_wide = true;
1272 
1273 	return kvm_events_report_vcpu(kvm);
1274 }
1275 
1276 #ifdef HAVE_TIMERFD_SUPPORT
1277 static struct perf_evlist *kvm_live_event_list(void)
1278 {
1279 	struct perf_evlist *evlist;
1280 	char *tp, *name, *sys;
1281 	int err = -1;
1282 	const char * const *events_tp;
1283 
1284 	evlist = perf_evlist__new();
1285 	if (evlist == NULL)
1286 		return NULL;
1287 
1288 	for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1289 
1290 		tp = strdup(*events_tp);
1291 		if (tp == NULL)
1292 			goto out;
1293 
1294 		/* split tracepoint into subsystem and name */
1295 		sys = tp;
1296 		name = strchr(tp, ':');
1297 		if (name == NULL) {
1298 			pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1299 			       *events_tp);
1300 			free(tp);
1301 			goto out;
1302 		}
1303 		*name = '\0';
1304 		name++;
1305 
1306 		if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1307 			pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1308 			free(tp);
1309 			goto out;
1310 		}
1311 
1312 		free(tp);
1313 	}
1314 
1315 	err = 0;
1316 
1317 out:
1318 	if (err) {
1319 		perf_evlist__delete(evlist);
1320 		evlist = NULL;
1321 	}
1322 
1323 	return evlist;
1324 }
1325 
1326 static int kvm_events_live(struct perf_kvm_stat *kvm,
1327 			   int argc, const char **argv)
1328 {
1329 	char errbuf[BUFSIZ];
1330 	int err;
1331 
1332 	const struct option live_options[] = {
1333 		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1334 			"record events on existing process id"),
1335 		OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1336 			"number of mmap data pages",
1337 			perf_evlist__parse_mmap_pages),
1338 		OPT_INCR('v', "verbose", &verbose,
1339 			"be more verbose (show counter open errors, etc)"),
1340 		OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1341 			"system-wide collection from all CPUs"),
1342 		OPT_UINTEGER('d', "display", &kvm->display_time,
1343 			"time in seconds between display updates"),
1344 		OPT_STRING(0, "event", &kvm->report_event, "report event",
1345 			"event for reporting: "
1346 			"vmexit, mmio (x86 only), ioport (x86 only)"),
1347 		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1348 			"vcpu id to report"),
1349 		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1350 			"key for sorting: sample(sort by samples number)"
1351 			" time (sort by avg time)"),
1352 		OPT_U64(0, "duration", &kvm->duration,
1353 			"show events other than"
1354 			" HLT (x86 only) or Wait state (s390 only)"
1355 			" that take longer than duration usecs"),
1356 		OPT_UINTEGER(0, "proc-map-timeout", &kvm->opts.proc_map_timeout,
1357 				"per thread proc mmap processing timeout in ms"),
1358 		OPT_END()
1359 	};
1360 	const char * const live_usage[] = {
1361 		"perf kvm stat live [<options>]",
1362 		NULL
1363 	};
1364 	struct perf_data data = {
1365 		.mode = PERF_DATA_MODE_WRITE,
1366 	};
1367 
1368 
1369 	/* event handling */
1370 	kvm->tool.sample = process_sample_event;
1371 	kvm->tool.comm   = perf_event__process_comm;
1372 	kvm->tool.exit   = perf_event__process_exit;
1373 	kvm->tool.fork   = perf_event__process_fork;
1374 	kvm->tool.lost   = process_lost_event;
1375 	kvm->tool.namespaces  = perf_event__process_namespaces;
1376 	kvm->tool.ordered_events = true;
1377 	perf_tool__fill_defaults(&kvm->tool);
1378 
1379 	/* set defaults */
1380 	kvm->display_time = 1;
1381 	kvm->opts.user_interval = 1;
1382 	kvm->opts.mmap_pages = 512;
1383 	kvm->opts.target.uses_mmap = false;
1384 	kvm->opts.target.uid_str = NULL;
1385 	kvm->opts.target.uid = UINT_MAX;
1386 	kvm->opts.proc_map_timeout = 500;
1387 
1388 	symbol__init(NULL);
1389 	disable_buildid_cache();
1390 
1391 	use_browser = 0;
1392 
1393 	if (argc) {
1394 		argc = parse_options(argc, argv, live_options,
1395 				     live_usage, 0);
1396 		if (argc)
1397 			usage_with_options(live_usage, live_options);
1398 	}
1399 
1400 	kvm->duration *= NSEC_PER_USEC;   /* convert usec to nsec */
1401 
1402 	/*
1403 	 * target related setups
1404 	 */
1405 	err = target__validate(&kvm->opts.target);
1406 	if (err) {
1407 		target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1408 		ui__warning("%s", errbuf);
1409 	}
1410 
1411 	if (target__none(&kvm->opts.target))
1412 		kvm->opts.target.system_wide = true;
1413 
1414 
1415 	/*
1416 	 * generate the event list
1417 	 */
1418 	err = setup_kvm_events_tp(kvm);
1419 	if (err < 0) {
1420 		pr_err("Unable to setup the kvm tracepoints\n");
1421 		return err;
1422 	}
1423 
1424 	kvm->evlist = kvm_live_event_list();
1425 	if (kvm->evlist == NULL) {
1426 		err = -1;
1427 		goto out;
1428 	}
1429 
1430 	symbol_conf.nr_events = kvm->evlist->nr_entries;
1431 
1432 	if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1433 		usage_with_options(live_usage, live_options);
1434 
1435 	/*
1436 	 * perf session
1437 	 */
1438 	kvm->session = perf_session__new(&data, false, &kvm->tool);
1439 	if (kvm->session == NULL) {
1440 		err = -1;
1441 		goto out;
1442 	}
1443 	kvm->session->evlist = kvm->evlist;
1444 	perf_session__set_id_hdr_size(kvm->session);
1445 	ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
1446 	machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1447 				    kvm->evlist->threads, false,
1448 				    kvm->opts.proc_map_timeout, 1);
1449 	err = kvm_live_open_events(kvm);
1450 	if (err)
1451 		goto out;
1452 
1453 	err = kvm_events_live_report(kvm);
1454 
1455 out:
1456 	perf_session__delete(kvm->session);
1457 	kvm->session = NULL;
1458 	perf_evlist__delete(kvm->evlist);
1459 
1460 	return err;
1461 }
1462 #endif
1463 
1464 static void print_kvm_stat_usage(void)
1465 {
1466 	printf("Usage: perf kvm stat <command>\n\n");
1467 
1468 	printf("# Available commands:\n");
1469 	printf("\trecord: record kvm events\n");
1470 	printf("\treport: report statistical data of kvm events\n");
1471 	printf("\tlive:   live reporting of statistical data of kvm events\n");
1472 
1473 	printf("\nOtherwise, it is the alias of 'perf stat':\n");
1474 }
1475 
1476 static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1477 {
1478 	struct perf_kvm_stat kvm = {
1479 		.file_name = file_name,
1480 
1481 		.trace_vcpu	= -1,
1482 		.report_event	= "vmexit",
1483 		.sort_key	= "sample",
1484 
1485 	};
1486 
1487 	if (argc == 1) {
1488 		print_kvm_stat_usage();
1489 		goto perf_stat;
1490 	}
1491 
1492 	if (!strncmp(argv[1], "rec", 3))
1493 		return kvm_events_record(&kvm, argc - 1, argv + 1);
1494 
1495 	if (!strncmp(argv[1], "rep", 3))
1496 		return kvm_events_report(&kvm, argc - 1 , argv + 1);
1497 
1498 #ifdef HAVE_TIMERFD_SUPPORT
1499 	if (!strncmp(argv[1], "live", 4))
1500 		return kvm_events_live(&kvm, argc - 1 , argv + 1);
1501 #endif
1502 
1503 perf_stat:
1504 	return cmd_stat(argc, argv);
1505 }
1506 #endif /* HAVE_KVM_STAT_SUPPORT */
1507 
1508 static int __cmd_record(const char *file_name, int argc, const char **argv)
1509 {
1510 	int rec_argc, i = 0, j;
1511 	const char **rec_argv;
1512 
1513 	rec_argc = argc + 2;
1514 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1515 	rec_argv[i++] = strdup("record");
1516 	rec_argv[i++] = strdup("-o");
1517 	rec_argv[i++] = strdup(file_name);
1518 	for (j = 1; j < argc; j++, i++)
1519 		rec_argv[i] = argv[j];
1520 
1521 	BUG_ON(i != rec_argc);
1522 
1523 	return cmd_record(i, rec_argv);
1524 }
1525 
1526 static int __cmd_report(const char *file_name, int argc, const char **argv)
1527 {
1528 	int rec_argc, i = 0, j;
1529 	const char **rec_argv;
1530 
1531 	rec_argc = argc + 2;
1532 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1533 	rec_argv[i++] = strdup("report");
1534 	rec_argv[i++] = strdup("-i");
1535 	rec_argv[i++] = strdup(file_name);
1536 	for (j = 1; j < argc; j++, i++)
1537 		rec_argv[i] = argv[j];
1538 
1539 	BUG_ON(i != rec_argc);
1540 
1541 	return cmd_report(i, rec_argv);
1542 }
1543 
1544 static int
1545 __cmd_buildid_list(const char *file_name, int argc, const char **argv)
1546 {
1547 	int rec_argc, i = 0, j;
1548 	const char **rec_argv;
1549 
1550 	rec_argc = argc + 2;
1551 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1552 	rec_argv[i++] = strdup("buildid-list");
1553 	rec_argv[i++] = strdup("-i");
1554 	rec_argv[i++] = strdup(file_name);
1555 	for (j = 1; j < argc; j++, i++)
1556 		rec_argv[i] = argv[j];
1557 
1558 	BUG_ON(i != rec_argc);
1559 
1560 	return cmd_buildid_list(i, rec_argv);
1561 }
1562 
1563 int cmd_kvm(int argc, const char **argv)
1564 {
1565 	const char *file_name = NULL;
1566 	const struct option kvm_options[] = {
1567 		OPT_STRING('i', "input", &file_name, "file",
1568 			   "Input file name"),
1569 		OPT_STRING('o', "output", &file_name, "file",
1570 			   "Output file name"),
1571 		OPT_BOOLEAN(0, "guest", &perf_guest,
1572 			    "Collect guest os data"),
1573 		OPT_BOOLEAN(0, "host", &perf_host,
1574 			    "Collect host os data"),
1575 		OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1576 			   "guest mount directory under which every guest os"
1577 			   " instance has a subdir"),
1578 		OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1579 			   "file", "file saving guest os vmlinux"),
1580 		OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1581 			   "file", "file saving guest os /proc/kallsyms"),
1582 		OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1583 			   "file", "file saving guest os /proc/modules"),
1584 		OPT_INCR('v', "verbose", &verbose,
1585 			    "be more verbose (show counter open errors, etc)"),
1586 		OPT_END()
1587 	};
1588 
1589 	const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1590 						"buildid-list", "stat", NULL };
1591 	const char *kvm_usage[] = { NULL, NULL };
1592 
1593 	perf_host  = 0;
1594 	perf_guest = 1;
1595 
1596 	argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1597 					PARSE_OPT_STOP_AT_NON_OPTION);
1598 	if (!argc)
1599 		usage_with_options(kvm_usage, kvm_options);
1600 
1601 	if (!perf_host)
1602 		perf_guest = 1;
1603 
1604 	if (!file_name) {
1605 		file_name = get_filename_for_perf_kvm();
1606 
1607 		if (!file_name) {
1608 			pr_err("Failed to allocate memory for filename\n");
1609 			return -ENOMEM;
1610 		}
1611 	}
1612 
1613 	if (!strncmp(argv[0], "rec", 3))
1614 		return __cmd_record(file_name, argc, argv);
1615 	else if (!strncmp(argv[0], "rep", 3))
1616 		return __cmd_report(file_name, argc, argv);
1617 	else if (!strncmp(argv[0], "diff", 4))
1618 		return cmd_diff(argc, argv);
1619 	else if (!strncmp(argv[0], "top", 3))
1620 		return cmd_top(argc, argv);
1621 	else if (!strncmp(argv[0], "buildid-list", 12))
1622 		return __cmd_buildid_list(file_name, argc, argv);
1623 #ifdef HAVE_KVM_STAT_SUPPORT
1624 	else if (!strncmp(argv[0], "stat", 4))
1625 		return kvm_cmd_stat(file_name, argc, argv);
1626 #endif
1627 	else
1628 		usage_with_options(kvm_usage, kvm_options);
1629 
1630 	return 0;
1631 }
1632