xref: /openbmc/linux/tools/perf/builtin-kvm.c (revision bc000245)
1 #include "builtin.h"
2 #include "perf.h"
3 
4 #include "util/evsel.h"
5 #include "util/evlist.h"
6 #include "util/util.h"
7 #include "util/cache.h"
8 #include "util/symbol.h"
9 #include "util/thread.h"
10 #include "util/header.h"
11 #include "util/session.h"
12 #include "util/intlist.h"
13 #include "util/parse-options.h"
14 #include "util/trace-event.h"
15 #include "util/debug.h"
16 #include <lk/debugfs.h>
17 #include "util/tool.h"
18 #include "util/stat.h"
19 #include "util/top.h"
20 #include "util/data.h"
21 
22 #include <sys/prctl.h>
23 #ifdef HAVE_TIMERFD_SUPPORT
24 #include <sys/timerfd.h>
25 #endif
26 
27 #include <termios.h>
28 #include <semaphore.h>
29 #include <pthread.h>
30 #include <math.h>
31 
32 #if defined(__i386__) || defined(__x86_64__)
33 #include <asm/svm.h>
34 #include <asm/vmx.h>
35 #include <asm/kvm.h>
36 
37 struct event_key {
38 	#define INVALID_KEY     (~0ULL)
39 	u64 key;
40 	int info;
41 };
42 
43 struct kvm_event_stats {
44 	u64 time;
45 	struct stats stats;
46 };
47 
48 struct kvm_event {
49 	struct list_head hash_entry;
50 	struct rb_node rb;
51 
52 	struct event_key key;
53 
54 	struct kvm_event_stats total;
55 
56 	#define DEFAULT_VCPU_NUM 8
57 	int max_vcpu;
58 	struct kvm_event_stats *vcpu;
59 };
60 
61 typedef int (*key_cmp_fun)(struct kvm_event*, struct kvm_event*, int);
62 
63 struct kvm_event_key {
64 	const char *name;
65 	key_cmp_fun key;
66 };
67 
68 
69 struct perf_kvm_stat;
70 
71 struct kvm_events_ops {
72 	bool (*is_begin_event)(struct perf_evsel *evsel,
73 			       struct perf_sample *sample,
74 			       struct event_key *key);
75 	bool (*is_end_event)(struct perf_evsel *evsel,
76 			     struct perf_sample *sample, struct event_key *key);
77 	void (*decode_key)(struct perf_kvm_stat *kvm, struct event_key *key,
78 			   char decode[20]);
79 	const char *name;
80 };
81 
82 struct exit_reasons_table {
83 	unsigned long exit_code;
84 	const char *reason;
85 };
86 
87 #define EVENTS_BITS		12
88 #define EVENTS_CACHE_SIZE	(1UL << EVENTS_BITS)
89 
90 struct perf_kvm_stat {
91 	struct perf_tool    tool;
92 	struct perf_record_opts opts;
93 	struct perf_evlist  *evlist;
94 	struct perf_session *session;
95 
96 	const char *file_name;
97 	const char *report_event;
98 	const char *sort_key;
99 	int trace_vcpu;
100 
101 	struct exit_reasons_table *exit_reasons;
102 	int exit_reasons_size;
103 	const char *exit_reasons_isa;
104 
105 	struct kvm_events_ops *events_ops;
106 	key_cmp_fun compare;
107 	struct list_head kvm_events_cache[EVENTS_CACHE_SIZE];
108 
109 	u64 total_time;
110 	u64 total_count;
111 	u64 lost_events;
112 	u64 duration;
113 
114 	const char *pid_str;
115 	struct intlist *pid_list;
116 
117 	struct rb_root result;
118 
119 	int timerfd;
120 	unsigned int display_time;
121 	bool live;
122 };
123 
124 
125 static void exit_event_get_key(struct perf_evsel *evsel,
126 			       struct perf_sample *sample,
127 			       struct event_key *key)
128 {
129 	key->info = 0;
130 	key->key = perf_evsel__intval(evsel, sample, "exit_reason");
131 }
132 
133 static bool kvm_exit_event(struct perf_evsel *evsel)
134 {
135 	return !strcmp(evsel->name, "kvm:kvm_exit");
136 }
137 
138 static bool exit_event_begin(struct perf_evsel *evsel,
139 			     struct perf_sample *sample, struct event_key *key)
140 {
141 	if (kvm_exit_event(evsel)) {
142 		exit_event_get_key(evsel, sample, key);
143 		return true;
144 	}
145 
146 	return false;
147 }
148 
149 static bool kvm_entry_event(struct perf_evsel *evsel)
150 {
151 	return !strcmp(evsel->name, "kvm:kvm_entry");
152 }
153 
154 static bool exit_event_end(struct perf_evsel *evsel,
155 			   struct perf_sample *sample __maybe_unused,
156 			   struct event_key *key __maybe_unused)
157 {
158 	return kvm_entry_event(evsel);
159 }
160 
161 static struct exit_reasons_table vmx_exit_reasons[] = {
162 	VMX_EXIT_REASONS
163 };
164 
165 static struct exit_reasons_table svm_exit_reasons[] = {
166 	SVM_EXIT_REASONS
167 };
168 
169 static const char *get_exit_reason(struct perf_kvm_stat *kvm, u64 exit_code)
170 {
171 	int i = kvm->exit_reasons_size;
172 	struct exit_reasons_table *tbl = kvm->exit_reasons;
173 
174 	while (i--) {
175 		if (tbl->exit_code == exit_code)
176 			return tbl->reason;
177 		tbl++;
178 	}
179 
180 	pr_err("unknown kvm exit code:%lld on %s\n",
181 		(unsigned long long)exit_code, kvm->exit_reasons_isa);
182 	return "UNKNOWN";
183 }
184 
185 static void exit_event_decode_key(struct perf_kvm_stat *kvm,
186 				  struct event_key *key,
187 				  char decode[20])
188 {
189 	const char *exit_reason = get_exit_reason(kvm, key->key);
190 
191 	scnprintf(decode, 20, "%s", exit_reason);
192 }
193 
194 static struct kvm_events_ops exit_events = {
195 	.is_begin_event = exit_event_begin,
196 	.is_end_event = exit_event_end,
197 	.decode_key = exit_event_decode_key,
198 	.name = "VM-EXIT"
199 };
200 
201 /*
202  * For the mmio events, we treat:
203  * the time of MMIO write: kvm_mmio(KVM_TRACE_MMIO_WRITE...) -> kvm_entry
204  * the time of MMIO read: kvm_exit -> kvm_mmio(KVM_TRACE_MMIO_READ...).
205  */
206 static void mmio_event_get_key(struct perf_evsel *evsel, struct perf_sample *sample,
207 			       struct event_key *key)
208 {
209 	key->key  = perf_evsel__intval(evsel, sample, "gpa");
210 	key->info = perf_evsel__intval(evsel, sample, "type");
211 }
212 
213 #define KVM_TRACE_MMIO_READ_UNSATISFIED 0
214 #define KVM_TRACE_MMIO_READ 1
215 #define KVM_TRACE_MMIO_WRITE 2
216 
217 static bool mmio_event_begin(struct perf_evsel *evsel,
218 			     struct perf_sample *sample, struct event_key *key)
219 {
220 	/* MMIO read begin event in kernel. */
221 	if (kvm_exit_event(evsel))
222 		return true;
223 
224 	/* MMIO write begin event in kernel. */
225 	if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
226 	    perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_WRITE) {
227 		mmio_event_get_key(evsel, sample, key);
228 		return true;
229 	}
230 
231 	return false;
232 }
233 
234 static bool mmio_event_end(struct perf_evsel *evsel, struct perf_sample *sample,
235 			   struct event_key *key)
236 {
237 	/* MMIO write end event in kernel. */
238 	if (kvm_entry_event(evsel))
239 		return true;
240 
241 	/* MMIO read end event in kernel.*/
242 	if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
243 	    perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_READ) {
244 		mmio_event_get_key(evsel, sample, key);
245 		return true;
246 	}
247 
248 	return false;
249 }
250 
251 static void mmio_event_decode_key(struct perf_kvm_stat *kvm __maybe_unused,
252 				  struct event_key *key,
253 				  char decode[20])
254 {
255 	scnprintf(decode, 20, "%#lx:%s", (unsigned long)key->key,
256 				key->info == KVM_TRACE_MMIO_WRITE ? "W" : "R");
257 }
258 
259 static struct kvm_events_ops mmio_events = {
260 	.is_begin_event = mmio_event_begin,
261 	.is_end_event = mmio_event_end,
262 	.decode_key = mmio_event_decode_key,
263 	.name = "MMIO Access"
264 };
265 
266  /* The time of emulation pio access is from kvm_pio to kvm_entry. */
267 static void ioport_event_get_key(struct perf_evsel *evsel,
268 				 struct perf_sample *sample,
269 				 struct event_key *key)
270 {
271 	key->key  = perf_evsel__intval(evsel, sample, "port");
272 	key->info = perf_evsel__intval(evsel, sample, "rw");
273 }
274 
275 static bool ioport_event_begin(struct perf_evsel *evsel,
276 			       struct perf_sample *sample,
277 			       struct event_key *key)
278 {
279 	if (!strcmp(evsel->name, "kvm:kvm_pio")) {
280 		ioport_event_get_key(evsel, sample, key);
281 		return true;
282 	}
283 
284 	return false;
285 }
286 
287 static bool ioport_event_end(struct perf_evsel *evsel,
288 			     struct perf_sample *sample __maybe_unused,
289 			     struct event_key *key __maybe_unused)
290 {
291 	return kvm_entry_event(evsel);
292 }
293 
294 static void ioport_event_decode_key(struct perf_kvm_stat *kvm __maybe_unused,
295 				    struct event_key *key,
296 				    char decode[20])
297 {
298 	scnprintf(decode, 20, "%#llx:%s", (unsigned long long)key->key,
299 				key->info ? "POUT" : "PIN");
300 }
301 
302 static struct kvm_events_ops ioport_events = {
303 	.is_begin_event = ioport_event_begin,
304 	.is_end_event = ioport_event_end,
305 	.decode_key = ioport_event_decode_key,
306 	.name = "IO Port Access"
307 };
308 
309 static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
310 {
311 	bool ret = true;
312 
313 	if (!strcmp(kvm->report_event, "vmexit"))
314 		kvm->events_ops = &exit_events;
315 	else if (!strcmp(kvm->report_event, "mmio"))
316 		kvm->events_ops = &mmio_events;
317 	else if (!strcmp(kvm->report_event, "ioport"))
318 		kvm->events_ops = &ioport_events;
319 	else {
320 		pr_err("Unknown report event:%s\n", kvm->report_event);
321 		ret = false;
322 	}
323 
324 	return ret;
325 }
326 
327 struct vcpu_event_record {
328 	int vcpu_id;
329 	u64 start_time;
330 	struct kvm_event *last_event;
331 };
332 
333 
334 static void init_kvm_event_record(struct perf_kvm_stat *kvm)
335 {
336 	unsigned int i;
337 
338 	for (i = 0; i < EVENTS_CACHE_SIZE; i++)
339 		INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
340 }
341 
342 #ifdef HAVE_TIMERFD_SUPPORT
343 static void clear_events_cache_stats(struct list_head *kvm_events_cache)
344 {
345 	struct list_head *head;
346 	struct kvm_event *event;
347 	unsigned int i;
348 	int j;
349 
350 	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
351 		head = &kvm_events_cache[i];
352 		list_for_each_entry(event, head, hash_entry) {
353 			/* reset stats for event */
354 			event->total.time = 0;
355 			init_stats(&event->total.stats);
356 
357 			for (j = 0; j < event->max_vcpu; ++j) {
358 				event->vcpu[j].time = 0;
359 				init_stats(&event->vcpu[j].stats);
360 			}
361 		}
362 	}
363 }
364 #endif
365 
366 static int kvm_events_hash_fn(u64 key)
367 {
368 	return key & (EVENTS_CACHE_SIZE - 1);
369 }
370 
371 static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
372 {
373 	int old_max_vcpu = event->max_vcpu;
374 	void *prev;
375 
376 	if (vcpu_id < event->max_vcpu)
377 		return true;
378 
379 	while (event->max_vcpu <= vcpu_id)
380 		event->max_vcpu += DEFAULT_VCPU_NUM;
381 
382 	prev = event->vcpu;
383 	event->vcpu = realloc(event->vcpu,
384 			      event->max_vcpu * sizeof(*event->vcpu));
385 	if (!event->vcpu) {
386 		free(prev);
387 		pr_err("Not enough memory\n");
388 		return false;
389 	}
390 
391 	memset(event->vcpu + old_max_vcpu, 0,
392 	       (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
393 	return true;
394 }
395 
396 static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
397 {
398 	struct kvm_event *event;
399 
400 	event = zalloc(sizeof(*event));
401 	if (!event) {
402 		pr_err("Not enough memory\n");
403 		return NULL;
404 	}
405 
406 	event->key = *key;
407 	return event;
408 }
409 
410 static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
411 					       struct event_key *key)
412 {
413 	struct kvm_event *event;
414 	struct list_head *head;
415 
416 	BUG_ON(key->key == INVALID_KEY);
417 
418 	head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
419 	list_for_each_entry(event, head, hash_entry) {
420 		if (event->key.key == key->key && event->key.info == key->info)
421 			return event;
422 	}
423 
424 	event = kvm_alloc_init_event(key);
425 	if (!event)
426 		return NULL;
427 
428 	list_add(&event->hash_entry, head);
429 	return event;
430 }
431 
432 static bool handle_begin_event(struct perf_kvm_stat *kvm,
433 			       struct vcpu_event_record *vcpu_record,
434 			       struct event_key *key, u64 timestamp)
435 {
436 	struct kvm_event *event = NULL;
437 
438 	if (key->key != INVALID_KEY)
439 		event = find_create_kvm_event(kvm, key);
440 
441 	vcpu_record->last_event = event;
442 	vcpu_record->start_time = timestamp;
443 	return true;
444 }
445 
446 static void
447 kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
448 {
449 	kvm_stats->time += time_diff;
450 	update_stats(&kvm_stats->stats, time_diff);
451 }
452 
453 static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
454 {
455 	struct kvm_event_stats *kvm_stats = &event->total;
456 
457 	if (vcpu_id != -1)
458 		kvm_stats = &event->vcpu[vcpu_id];
459 
460 	return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
461 				avg_stats(&kvm_stats->stats));
462 }
463 
464 static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
465 			     u64 time_diff)
466 {
467 	if (vcpu_id == -1) {
468 		kvm_update_event_stats(&event->total, time_diff);
469 		return true;
470 	}
471 
472 	if (!kvm_event_expand(event, vcpu_id))
473 		return false;
474 
475 	kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
476 	return true;
477 }
478 
479 static bool handle_end_event(struct perf_kvm_stat *kvm,
480 			     struct vcpu_event_record *vcpu_record,
481 			     struct event_key *key,
482 			     struct perf_sample *sample)
483 {
484 	struct kvm_event *event;
485 	u64 time_begin, time_diff;
486 	int vcpu;
487 
488 	if (kvm->trace_vcpu == -1)
489 		vcpu = -1;
490 	else
491 		vcpu = vcpu_record->vcpu_id;
492 
493 	event = vcpu_record->last_event;
494 	time_begin = vcpu_record->start_time;
495 
496 	/* The begin event is not caught. */
497 	if (!time_begin)
498 		return true;
499 
500 	/*
501 	 * In some case, the 'begin event' only records the start timestamp,
502 	 * the actual event is recognized in the 'end event' (e.g. mmio-event).
503 	 */
504 
505 	/* Both begin and end events did not get the key. */
506 	if (!event && key->key == INVALID_KEY)
507 		return true;
508 
509 	if (!event)
510 		event = find_create_kvm_event(kvm, key);
511 
512 	if (!event)
513 		return false;
514 
515 	vcpu_record->last_event = NULL;
516 	vcpu_record->start_time = 0;
517 
518 	/* seems to happen once in a while during live mode */
519 	if (sample->time < time_begin) {
520 		pr_debug("End time before begin time; skipping event.\n");
521 		return true;
522 	}
523 
524 	time_diff = sample->time - time_begin;
525 
526 	if (kvm->duration && time_diff > kvm->duration) {
527 		char decode[32];
528 
529 		kvm->events_ops->decode_key(kvm, &event->key, decode);
530 		if (strcmp(decode, "HLT")) {
531 			pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
532 				 sample->time, sample->pid, vcpu_record->vcpu_id,
533 				 decode, time_diff/1000);
534 		}
535 	}
536 
537 	return update_kvm_event(event, vcpu, time_diff);
538 }
539 
540 static
541 struct vcpu_event_record *per_vcpu_record(struct thread *thread,
542 					  struct perf_evsel *evsel,
543 					  struct perf_sample *sample)
544 {
545 	/* Only kvm_entry records vcpu id. */
546 	if (!thread->priv && kvm_entry_event(evsel)) {
547 		struct vcpu_event_record *vcpu_record;
548 
549 		vcpu_record = zalloc(sizeof(*vcpu_record));
550 		if (!vcpu_record) {
551 			pr_err("%s: Not enough memory\n", __func__);
552 			return NULL;
553 		}
554 
555 		vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample, "vcpu_id");
556 		thread->priv = vcpu_record;
557 	}
558 
559 	return thread->priv;
560 }
561 
562 static bool handle_kvm_event(struct perf_kvm_stat *kvm,
563 			     struct thread *thread,
564 			     struct perf_evsel *evsel,
565 			     struct perf_sample *sample)
566 {
567 	struct vcpu_event_record *vcpu_record;
568 	struct event_key key = {.key = INVALID_KEY};
569 
570 	vcpu_record = per_vcpu_record(thread, evsel, sample);
571 	if (!vcpu_record)
572 		return true;
573 
574 	/* only process events for vcpus user cares about */
575 	if ((kvm->trace_vcpu != -1) &&
576 	    (kvm->trace_vcpu != vcpu_record->vcpu_id))
577 		return true;
578 
579 	if (kvm->events_ops->is_begin_event(evsel, sample, &key))
580 		return handle_begin_event(kvm, vcpu_record, &key, sample->time);
581 
582 	if (kvm->events_ops->is_end_event(evsel, sample, &key))
583 		return handle_end_event(kvm, vcpu_record, &key, sample);
584 
585 	return true;
586 }
587 
588 #define GET_EVENT_KEY(func, field)					\
589 static u64 get_event_ ##func(struct kvm_event *event, int vcpu)		\
590 {									\
591 	if (vcpu == -1)							\
592 		return event->total.field;				\
593 									\
594 	if (vcpu >= event->max_vcpu)					\
595 		return 0;						\
596 									\
597 	return event->vcpu[vcpu].field;					\
598 }
599 
600 #define COMPARE_EVENT_KEY(func, field)					\
601 GET_EVENT_KEY(func, field)						\
602 static int compare_kvm_event_ ## func(struct kvm_event *one,		\
603 					struct kvm_event *two, int vcpu)\
604 {									\
605 	return get_event_ ##func(one, vcpu) >				\
606 				get_event_ ##func(two, vcpu);		\
607 }
608 
609 GET_EVENT_KEY(time, time);
610 COMPARE_EVENT_KEY(count, stats.n);
611 COMPARE_EVENT_KEY(mean, stats.mean);
612 GET_EVENT_KEY(max, stats.max);
613 GET_EVENT_KEY(min, stats.min);
614 
615 #define DEF_SORT_NAME_KEY(name, compare_key)				\
616 	{ #name, compare_kvm_event_ ## compare_key }
617 
618 static struct kvm_event_key keys[] = {
619 	DEF_SORT_NAME_KEY(sample, count),
620 	DEF_SORT_NAME_KEY(time, mean),
621 	{ NULL, NULL }
622 };
623 
624 static bool select_key(struct perf_kvm_stat *kvm)
625 {
626 	int i;
627 
628 	for (i = 0; keys[i].name; i++) {
629 		if (!strcmp(keys[i].name, kvm->sort_key)) {
630 			kvm->compare = keys[i].key;
631 			return true;
632 		}
633 	}
634 
635 	pr_err("Unknown compare key:%s\n", kvm->sort_key);
636 	return false;
637 }
638 
639 static void insert_to_result(struct rb_root *result, struct kvm_event *event,
640 			     key_cmp_fun bigger, int vcpu)
641 {
642 	struct rb_node **rb = &result->rb_node;
643 	struct rb_node *parent = NULL;
644 	struct kvm_event *p;
645 
646 	while (*rb) {
647 		p = container_of(*rb, struct kvm_event, rb);
648 		parent = *rb;
649 
650 		if (bigger(event, p, vcpu))
651 			rb = &(*rb)->rb_left;
652 		else
653 			rb = &(*rb)->rb_right;
654 	}
655 
656 	rb_link_node(&event->rb, parent, rb);
657 	rb_insert_color(&event->rb, result);
658 }
659 
660 static void
661 update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
662 {
663 	int vcpu = kvm->trace_vcpu;
664 
665 	kvm->total_count += get_event_count(event, vcpu);
666 	kvm->total_time += get_event_time(event, vcpu);
667 }
668 
669 static bool event_is_valid(struct kvm_event *event, int vcpu)
670 {
671 	return !!get_event_count(event, vcpu);
672 }
673 
674 static void sort_result(struct perf_kvm_stat *kvm)
675 {
676 	unsigned int i;
677 	int vcpu = kvm->trace_vcpu;
678 	struct kvm_event *event;
679 
680 	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
681 		list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
682 			if (event_is_valid(event, vcpu)) {
683 				update_total_count(kvm, event);
684 				insert_to_result(&kvm->result, event,
685 						 kvm->compare, vcpu);
686 			}
687 		}
688 	}
689 }
690 
691 /* returns left most element of result, and erase it */
692 static struct kvm_event *pop_from_result(struct rb_root *result)
693 {
694 	struct rb_node *node = rb_first(result);
695 
696 	if (!node)
697 		return NULL;
698 
699 	rb_erase(node, result);
700 	return container_of(node, struct kvm_event, rb);
701 }
702 
703 static void print_vcpu_info(struct perf_kvm_stat *kvm)
704 {
705 	int vcpu = kvm->trace_vcpu;
706 
707 	pr_info("Analyze events for ");
708 
709 	if (kvm->live) {
710 		if (kvm->opts.target.system_wide)
711 			pr_info("all VMs, ");
712 		else if (kvm->opts.target.pid)
713 			pr_info("pid(s) %s, ", kvm->opts.target.pid);
714 		else
715 			pr_info("dazed and confused on what is monitored, ");
716 	}
717 
718 	if (vcpu == -1)
719 		pr_info("all VCPUs:\n\n");
720 	else
721 		pr_info("VCPU %d:\n\n", vcpu);
722 }
723 
724 static void show_timeofday(void)
725 {
726 	char date[64];
727 	struct timeval tv;
728 	struct tm ltime;
729 
730 	gettimeofday(&tv, NULL);
731 	if (localtime_r(&tv.tv_sec, &ltime)) {
732 		strftime(date, sizeof(date), "%H:%M:%S", &ltime);
733 		pr_info("%s.%06ld", date, tv.tv_usec);
734 	} else
735 		pr_info("00:00:00.000000");
736 
737 	return;
738 }
739 
740 static void print_result(struct perf_kvm_stat *kvm)
741 {
742 	char decode[20];
743 	struct kvm_event *event;
744 	int vcpu = kvm->trace_vcpu;
745 
746 	if (kvm->live) {
747 		puts(CONSOLE_CLEAR);
748 		show_timeofday();
749 	}
750 
751 	pr_info("\n\n");
752 	print_vcpu_info(kvm);
753 	pr_info("%20s ", kvm->events_ops->name);
754 	pr_info("%10s ", "Samples");
755 	pr_info("%9s ", "Samples%");
756 
757 	pr_info("%9s ", "Time%");
758 	pr_info("%10s ", "Min Time");
759 	pr_info("%10s ", "Max Time");
760 	pr_info("%16s ", "Avg time");
761 	pr_info("\n\n");
762 
763 	while ((event = pop_from_result(&kvm->result))) {
764 		u64 ecount, etime, max, min;
765 
766 		ecount = get_event_count(event, vcpu);
767 		etime = get_event_time(event, vcpu);
768 		max = get_event_max(event, vcpu);
769 		min = get_event_min(event, vcpu);
770 
771 		kvm->events_ops->decode_key(kvm, &event->key, decode);
772 		pr_info("%20s ", decode);
773 		pr_info("%10llu ", (unsigned long long)ecount);
774 		pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
775 		pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
776 		pr_info("%8" PRIu64 "us ", min / 1000);
777 		pr_info("%8" PRIu64 "us ", max / 1000);
778 		pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount/1e3,
779 			kvm_event_rel_stddev(vcpu, event));
780 		pr_info("\n");
781 	}
782 
783 	pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
784 		kvm->total_count, kvm->total_time / 1e3);
785 
786 	if (kvm->lost_events)
787 		pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
788 }
789 
790 #ifdef HAVE_TIMERFD_SUPPORT
791 static int process_lost_event(struct perf_tool *tool,
792 			      union perf_event *event __maybe_unused,
793 			      struct perf_sample *sample __maybe_unused,
794 			      struct machine *machine __maybe_unused)
795 {
796 	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
797 
798 	kvm->lost_events++;
799 	return 0;
800 }
801 #endif
802 
803 static bool skip_sample(struct perf_kvm_stat *kvm,
804 			struct perf_sample *sample)
805 {
806 	if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
807 		return true;
808 
809 	return false;
810 }
811 
812 static int process_sample_event(struct perf_tool *tool,
813 				union perf_event *event,
814 				struct perf_sample *sample,
815 				struct perf_evsel *evsel,
816 				struct machine *machine)
817 {
818 	struct thread *thread;
819 	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
820 						 tool);
821 
822 	if (skip_sample(kvm, sample))
823 		return 0;
824 
825 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
826 	if (thread == NULL) {
827 		pr_debug("problem processing %d event, skipping it.\n",
828 			event->header.type);
829 		return -1;
830 	}
831 
832 	if (!handle_kvm_event(kvm, thread, evsel, sample))
833 		return -1;
834 
835 	return 0;
836 }
837 
838 static int cpu_isa_config(struct perf_kvm_stat *kvm)
839 {
840 	char buf[64], *cpuid;
841 	int err, isa;
842 
843 	if (kvm->live) {
844 		err = get_cpuid(buf, sizeof(buf));
845 		if (err != 0) {
846 			pr_err("Failed to look up CPU type (Intel or AMD)\n");
847 			return err;
848 		}
849 		cpuid = buf;
850 	} else
851 		cpuid = kvm->session->header.env.cpuid;
852 
853 	if (strstr(cpuid, "Intel"))
854 		isa = 1;
855 	else if (strstr(cpuid, "AMD"))
856 		isa = 0;
857 	else {
858 		pr_err("CPU %s is not supported.\n", cpuid);
859 		return -ENOTSUP;
860 	}
861 
862 	if (isa == 1) {
863 		kvm->exit_reasons = vmx_exit_reasons;
864 		kvm->exit_reasons_size = ARRAY_SIZE(vmx_exit_reasons);
865 		kvm->exit_reasons_isa = "VMX";
866 	}
867 
868 	return 0;
869 }
870 
871 static bool verify_vcpu(int vcpu)
872 {
873 	if (vcpu != -1 && vcpu < 0) {
874 		pr_err("Invalid vcpu:%d.\n", vcpu);
875 		return false;
876 	}
877 
878 	return true;
879 }
880 
881 #ifdef HAVE_TIMERFD_SUPPORT
882 /* keeping the max events to a modest level to keep
883  * the processing of samples per mmap smooth.
884  */
885 #define PERF_KVM__MAX_EVENTS_PER_MMAP  25
886 
887 static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
888 				   u64 *mmap_time)
889 {
890 	union perf_event *event;
891 	struct perf_sample sample;
892 	s64 n = 0;
893 	int err;
894 
895 	*mmap_time = ULLONG_MAX;
896 	while ((event = perf_evlist__mmap_read(kvm->evlist, idx)) != NULL) {
897 		err = perf_evlist__parse_sample(kvm->evlist, event, &sample);
898 		if (err) {
899 			perf_evlist__mmap_consume(kvm->evlist, idx);
900 			pr_err("Failed to parse sample\n");
901 			return -1;
902 		}
903 
904 		err = perf_session_queue_event(kvm->session, event, &sample, 0);
905 		/*
906 		 * FIXME: Here we can't consume the event, as perf_session_queue_event will
907 		 *        point to it, and it'll get possibly overwritten by the kernel.
908 		 */
909 		perf_evlist__mmap_consume(kvm->evlist, idx);
910 
911 		if (err) {
912 			pr_err("Failed to enqueue sample: %d\n", err);
913 			return -1;
914 		}
915 
916 		/* save time stamp of our first sample for this mmap */
917 		if (n == 0)
918 			*mmap_time = sample.time;
919 
920 		/* limit events per mmap handled all at once */
921 		n++;
922 		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
923 			break;
924 	}
925 
926 	return n;
927 }
928 
929 static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
930 {
931 	int i, err, throttled = 0;
932 	s64 n, ntotal = 0;
933 	u64 flush_time = ULLONG_MAX, mmap_time;
934 
935 	for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
936 		n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
937 		if (n < 0)
938 			return -1;
939 
940 		/* flush time is going to be the minimum of all the individual
941 		 * mmap times. Essentially, we flush all the samples queued up
942 		 * from the last pass under our minimal start time -- that leaves
943 		 * a very small race for samples to come in with a lower timestamp.
944 		 * The ioctl to return the perf_clock timestamp should close the
945 		 * race entirely.
946 		 */
947 		if (mmap_time < flush_time)
948 			flush_time = mmap_time;
949 
950 		ntotal += n;
951 		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
952 			throttled = 1;
953 	}
954 
955 	/* flush queue after each round in which we processed events */
956 	if (ntotal) {
957 		kvm->session->ordered_samples.next_flush = flush_time;
958 		err = kvm->tool.finished_round(&kvm->tool, NULL, kvm->session);
959 		if (err) {
960 			if (kvm->lost_events)
961 				pr_info("\nLost events: %" PRIu64 "\n\n",
962 					kvm->lost_events);
963 			return err;
964 		}
965 	}
966 
967 	return throttled;
968 }
969 
970 static volatile int done;
971 
972 static void sig_handler(int sig __maybe_unused)
973 {
974 	done = 1;
975 }
976 
977 static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
978 {
979 	struct itimerspec new_value;
980 	int rc = -1;
981 
982 	kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
983 	if (kvm->timerfd < 0) {
984 		pr_err("timerfd_create failed\n");
985 		goto out;
986 	}
987 
988 	new_value.it_value.tv_sec = kvm->display_time;
989 	new_value.it_value.tv_nsec = 0;
990 	new_value.it_interval.tv_sec = kvm->display_time;
991 	new_value.it_interval.tv_nsec = 0;
992 
993 	if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
994 		pr_err("timerfd_settime failed: %d\n", errno);
995 		close(kvm->timerfd);
996 		goto out;
997 	}
998 
999 	rc = 0;
1000 out:
1001 	return rc;
1002 }
1003 
1004 static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
1005 {
1006 	uint64_t c;
1007 	int rc;
1008 
1009 	rc = read(kvm->timerfd, &c, sizeof(uint64_t));
1010 	if (rc < 0) {
1011 		if (errno == EAGAIN)
1012 			return 0;
1013 
1014 		pr_err("Failed to read timer fd: %d\n", errno);
1015 		return -1;
1016 	}
1017 
1018 	if (rc != sizeof(uint64_t)) {
1019 		pr_err("Error reading timer fd - invalid size returned\n");
1020 		return -1;
1021 	}
1022 
1023 	if (c != 1)
1024 		pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
1025 
1026 	/* update display */
1027 	sort_result(kvm);
1028 	print_result(kvm);
1029 
1030 	/* reset counts */
1031 	clear_events_cache_stats(kvm->kvm_events_cache);
1032 	kvm->total_count = 0;
1033 	kvm->total_time = 0;
1034 	kvm->lost_events = 0;
1035 
1036 	return 0;
1037 }
1038 
1039 static int fd_set_nonblock(int fd)
1040 {
1041 	long arg = 0;
1042 
1043 	arg = fcntl(fd, F_GETFL);
1044 	if (arg < 0) {
1045 		pr_err("Failed to get current flags for fd %d\n", fd);
1046 		return -1;
1047 	}
1048 
1049 	if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
1050 		pr_err("Failed to set non-block option on fd %d\n", fd);
1051 		return -1;
1052 	}
1053 
1054 	return 0;
1055 }
1056 
1057 static
1058 int perf_kvm__handle_stdin(struct termios *tc_now, struct termios *tc_save)
1059 {
1060 	int c;
1061 
1062 	tcsetattr(0, TCSANOW, tc_now);
1063 	c = getc(stdin);
1064 	tcsetattr(0, TCSAFLUSH, tc_save);
1065 
1066 	if (c == 'q')
1067 		return 1;
1068 
1069 	return 0;
1070 }
1071 
1072 static int kvm_events_live_report(struct perf_kvm_stat *kvm)
1073 {
1074 	struct pollfd *pollfds = NULL;
1075 	int nr_fds, nr_stdin, ret, err = -EINVAL;
1076 	struct termios tc, save;
1077 
1078 	/* live flag must be set first */
1079 	kvm->live = true;
1080 
1081 	ret = cpu_isa_config(kvm);
1082 	if (ret < 0)
1083 		return ret;
1084 
1085 	if (!verify_vcpu(kvm->trace_vcpu) ||
1086 	    !select_key(kvm) ||
1087 	    !register_kvm_events_ops(kvm)) {
1088 		goto out;
1089 	}
1090 
1091 	init_kvm_event_record(kvm);
1092 
1093 	tcgetattr(0, &save);
1094 	tc = save;
1095 	tc.c_lflag &= ~(ICANON | ECHO);
1096 	tc.c_cc[VMIN] = 0;
1097 	tc.c_cc[VTIME] = 0;
1098 
1099 	signal(SIGINT, sig_handler);
1100 	signal(SIGTERM, sig_handler);
1101 
1102 	/* copy pollfds -- need to add timerfd and stdin */
1103 	nr_fds = kvm->evlist->nr_fds;
1104 	pollfds = zalloc(sizeof(struct pollfd) * (nr_fds + 2));
1105 	if (!pollfds) {
1106 		err = -ENOMEM;
1107 		goto out;
1108 	}
1109 	memcpy(pollfds, kvm->evlist->pollfd,
1110 		sizeof(struct pollfd) * kvm->evlist->nr_fds);
1111 
1112 	/* add timer fd */
1113 	if (perf_kvm__timerfd_create(kvm) < 0) {
1114 		err = -1;
1115 		goto out;
1116 	}
1117 
1118 	pollfds[nr_fds].fd = kvm->timerfd;
1119 	pollfds[nr_fds].events = POLLIN;
1120 	nr_fds++;
1121 
1122 	pollfds[nr_fds].fd = fileno(stdin);
1123 	pollfds[nr_fds].events = POLLIN;
1124 	nr_stdin = nr_fds;
1125 	nr_fds++;
1126 	if (fd_set_nonblock(fileno(stdin)) != 0)
1127 		goto out;
1128 
1129 	/* everything is good - enable the events and process */
1130 	perf_evlist__enable(kvm->evlist);
1131 
1132 	while (!done) {
1133 		int rc;
1134 
1135 		rc = perf_kvm__mmap_read(kvm);
1136 		if (rc < 0)
1137 			break;
1138 
1139 		err = perf_kvm__handle_timerfd(kvm);
1140 		if (err)
1141 			goto out;
1142 
1143 		if (pollfds[nr_stdin].revents & POLLIN)
1144 			done = perf_kvm__handle_stdin(&tc, &save);
1145 
1146 		if (!rc && !done)
1147 			err = poll(pollfds, nr_fds, 100);
1148 	}
1149 
1150 	perf_evlist__disable(kvm->evlist);
1151 
1152 	if (err == 0) {
1153 		sort_result(kvm);
1154 		print_result(kvm);
1155 	}
1156 
1157 out:
1158 	if (kvm->timerfd >= 0)
1159 		close(kvm->timerfd);
1160 
1161 	if (pollfds)
1162 		free(pollfds);
1163 
1164 	return err;
1165 }
1166 
1167 static int kvm_live_open_events(struct perf_kvm_stat *kvm)
1168 {
1169 	int err, rc = -1;
1170 	struct perf_evsel *pos;
1171 	struct perf_evlist *evlist = kvm->evlist;
1172 
1173 	perf_evlist__config(evlist, &kvm->opts);
1174 
1175 	/*
1176 	 * Note: exclude_{guest,host} do not apply here.
1177 	 *       This command processes KVM tracepoints from host only
1178 	 */
1179 	list_for_each_entry(pos, &evlist->entries, node) {
1180 		struct perf_event_attr *attr = &pos->attr;
1181 
1182 		/* make sure these *are* set */
1183 		perf_evsel__set_sample_bit(pos, TID);
1184 		perf_evsel__set_sample_bit(pos, TIME);
1185 		perf_evsel__set_sample_bit(pos, CPU);
1186 		perf_evsel__set_sample_bit(pos, RAW);
1187 		/* make sure these are *not*; want as small a sample as possible */
1188 		perf_evsel__reset_sample_bit(pos, PERIOD);
1189 		perf_evsel__reset_sample_bit(pos, IP);
1190 		perf_evsel__reset_sample_bit(pos, CALLCHAIN);
1191 		perf_evsel__reset_sample_bit(pos, ADDR);
1192 		perf_evsel__reset_sample_bit(pos, READ);
1193 		attr->mmap = 0;
1194 		attr->comm = 0;
1195 		attr->task = 0;
1196 
1197 		attr->sample_period = 1;
1198 
1199 		attr->watermark = 0;
1200 		attr->wakeup_events = 1000;
1201 
1202 		/* will enable all once we are ready */
1203 		attr->disabled = 1;
1204 	}
1205 
1206 	err = perf_evlist__open(evlist);
1207 	if (err < 0) {
1208 		printf("Couldn't create the events: %s\n", strerror(errno));
1209 		goto out;
1210 	}
1211 
1212 	if (perf_evlist__mmap(evlist, kvm->opts.mmap_pages, false) < 0) {
1213 		ui__error("Failed to mmap the events: %s\n", strerror(errno));
1214 		perf_evlist__close(evlist);
1215 		goto out;
1216 	}
1217 
1218 	rc = 0;
1219 
1220 out:
1221 	return rc;
1222 }
1223 #endif
1224 
1225 static int read_events(struct perf_kvm_stat *kvm)
1226 {
1227 	int ret;
1228 
1229 	struct perf_tool eops = {
1230 		.sample			= process_sample_event,
1231 		.comm			= perf_event__process_comm,
1232 		.ordered_samples	= true,
1233 	};
1234 	struct perf_data_file file = {
1235 		.path = input_name,
1236 		.mode = PERF_DATA_MODE_READ,
1237 	};
1238 
1239 	kvm->tool = eops;
1240 	kvm->session = perf_session__new(&file, false, &kvm->tool);
1241 	if (!kvm->session) {
1242 		pr_err("Initializing perf session failed\n");
1243 		return -EINVAL;
1244 	}
1245 
1246 	if (!perf_session__has_traces(kvm->session, "kvm record"))
1247 		return -EINVAL;
1248 
1249 	/*
1250 	 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1251 	 * traced in the old kernel.
1252 	 */
1253 	ret = cpu_isa_config(kvm);
1254 	if (ret < 0)
1255 		return ret;
1256 
1257 	return perf_session__process_events(kvm->session, &kvm->tool);
1258 }
1259 
1260 static int parse_target_str(struct perf_kvm_stat *kvm)
1261 {
1262 	if (kvm->pid_str) {
1263 		kvm->pid_list = intlist__new(kvm->pid_str);
1264 		if (kvm->pid_list == NULL) {
1265 			pr_err("Error parsing process id string\n");
1266 			return -EINVAL;
1267 		}
1268 	}
1269 
1270 	return 0;
1271 }
1272 
1273 static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1274 {
1275 	int ret = -EINVAL;
1276 	int vcpu = kvm->trace_vcpu;
1277 
1278 	if (parse_target_str(kvm) != 0)
1279 		goto exit;
1280 
1281 	if (!verify_vcpu(vcpu))
1282 		goto exit;
1283 
1284 	if (!select_key(kvm))
1285 		goto exit;
1286 
1287 	if (!register_kvm_events_ops(kvm))
1288 		goto exit;
1289 
1290 	init_kvm_event_record(kvm);
1291 	setup_pager();
1292 
1293 	ret = read_events(kvm);
1294 	if (ret)
1295 		goto exit;
1296 
1297 	sort_result(kvm);
1298 	print_result(kvm);
1299 
1300 exit:
1301 	return ret;
1302 }
1303 
1304 static const char * const kvm_events_tp[] = {
1305 	"kvm:kvm_entry",
1306 	"kvm:kvm_exit",
1307 	"kvm:kvm_mmio",
1308 	"kvm:kvm_pio",
1309 };
1310 
1311 #define STRDUP_FAIL_EXIT(s)		\
1312 	({	char *_p;		\
1313 	_p = strdup(s);		\
1314 		if (!_p)		\
1315 			return -ENOMEM;	\
1316 		_p;			\
1317 	})
1318 
1319 static int
1320 kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1321 {
1322 	unsigned int rec_argc, i, j;
1323 	const char **rec_argv;
1324 	const char * const record_args[] = {
1325 		"record",
1326 		"-R",
1327 		"-m", "1024",
1328 		"-c", "1",
1329 	};
1330 
1331 	rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1332 		   2 * ARRAY_SIZE(kvm_events_tp);
1333 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1334 
1335 	if (rec_argv == NULL)
1336 		return -ENOMEM;
1337 
1338 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1339 		rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1340 
1341 	for (j = 0; j < ARRAY_SIZE(kvm_events_tp); j++) {
1342 		rec_argv[i++] = "-e";
1343 		rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1344 	}
1345 
1346 	rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1347 	rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1348 
1349 	for (j = 1; j < (unsigned int)argc; j++, i++)
1350 		rec_argv[i] = argv[j];
1351 
1352 	return cmd_record(i, rec_argv, NULL);
1353 }
1354 
1355 static int
1356 kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1357 {
1358 	const struct option kvm_events_report_options[] = {
1359 		OPT_STRING(0, "event", &kvm->report_event, "report event",
1360 			    "event for reporting: vmexit, mmio, ioport"),
1361 		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1362 			    "vcpu id to report"),
1363 		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1364 			    "key for sorting: sample(sort by samples number)"
1365 			    " time (sort by avg time)"),
1366 		OPT_STRING('p', "pid", &kvm->pid_str, "pid",
1367 			   "analyze events only for given process id(s)"),
1368 		OPT_END()
1369 	};
1370 
1371 	const char * const kvm_events_report_usage[] = {
1372 		"perf kvm stat report [<options>]",
1373 		NULL
1374 	};
1375 
1376 	symbol__init();
1377 
1378 	if (argc) {
1379 		argc = parse_options(argc, argv,
1380 				     kvm_events_report_options,
1381 				     kvm_events_report_usage, 0);
1382 		if (argc)
1383 			usage_with_options(kvm_events_report_usage,
1384 					   kvm_events_report_options);
1385 	}
1386 
1387 	return kvm_events_report_vcpu(kvm);
1388 }
1389 
1390 #ifdef HAVE_TIMERFD_SUPPORT
1391 static struct perf_evlist *kvm_live_event_list(void)
1392 {
1393 	struct perf_evlist *evlist;
1394 	char *tp, *name, *sys;
1395 	unsigned int j;
1396 	int err = -1;
1397 
1398 	evlist = perf_evlist__new();
1399 	if (evlist == NULL)
1400 		return NULL;
1401 
1402 	for (j = 0; j < ARRAY_SIZE(kvm_events_tp); j++) {
1403 
1404 		tp = strdup(kvm_events_tp[j]);
1405 		if (tp == NULL)
1406 			goto out;
1407 
1408 		/* split tracepoint into subsystem and name */
1409 		sys = tp;
1410 		name = strchr(tp, ':');
1411 		if (name == NULL) {
1412 			pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1413 				kvm_events_tp[j]);
1414 			free(tp);
1415 			goto out;
1416 		}
1417 		*name = '\0';
1418 		name++;
1419 
1420 		if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1421 			pr_err("Failed to add %s tracepoint to the list\n", kvm_events_tp[j]);
1422 			free(tp);
1423 			goto out;
1424 		}
1425 
1426 		free(tp);
1427 	}
1428 
1429 	err = 0;
1430 
1431 out:
1432 	if (err) {
1433 		perf_evlist__delete(evlist);
1434 		evlist = NULL;
1435 	}
1436 
1437 	return evlist;
1438 }
1439 
1440 static int kvm_events_live(struct perf_kvm_stat *kvm,
1441 			   int argc, const char **argv)
1442 {
1443 	char errbuf[BUFSIZ];
1444 	int err;
1445 
1446 	const struct option live_options[] = {
1447 		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1448 			"record events on existing process id"),
1449 		OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1450 			"number of mmap data pages",
1451 			perf_evlist__parse_mmap_pages),
1452 		OPT_INCR('v', "verbose", &verbose,
1453 			"be more verbose (show counter open errors, etc)"),
1454 		OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1455 			"system-wide collection from all CPUs"),
1456 		OPT_UINTEGER('d', "display", &kvm->display_time,
1457 			"time in seconds between display updates"),
1458 		OPT_STRING(0, "event", &kvm->report_event, "report event",
1459 			"event for reporting: vmexit, mmio, ioport"),
1460 		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1461 			"vcpu id to report"),
1462 		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1463 			"key for sorting: sample(sort by samples number)"
1464 			" time (sort by avg time)"),
1465 		OPT_U64(0, "duration", &kvm->duration,
1466 		    "show events other than HALT that take longer than duration usecs"),
1467 		OPT_END()
1468 	};
1469 	const char * const live_usage[] = {
1470 		"perf kvm stat live [<options>]",
1471 		NULL
1472 	};
1473 	struct perf_data_file file = {
1474 		.mode = PERF_DATA_MODE_WRITE,
1475 	};
1476 
1477 
1478 	/* event handling */
1479 	kvm->tool.sample = process_sample_event;
1480 	kvm->tool.comm   = perf_event__process_comm;
1481 	kvm->tool.exit   = perf_event__process_exit;
1482 	kvm->tool.fork   = perf_event__process_fork;
1483 	kvm->tool.lost   = process_lost_event;
1484 	kvm->tool.ordered_samples = true;
1485 	perf_tool__fill_defaults(&kvm->tool);
1486 
1487 	/* set defaults */
1488 	kvm->display_time = 1;
1489 	kvm->opts.user_interval = 1;
1490 	kvm->opts.mmap_pages = 512;
1491 	kvm->opts.target.uses_mmap = false;
1492 	kvm->opts.target.uid_str = NULL;
1493 	kvm->opts.target.uid = UINT_MAX;
1494 
1495 	symbol__init();
1496 	disable_buildid_cache();
1497 
1498 	use_browser = 0;
1499 	setup_browser(false);
1500 
1501 	if (argc) {
1502 		argc = parse_options(argc, argv, live_options,
1503 				     live_usage, 0);
1504 		if (argc)
1505 			usage_with_options(live_usage, live_options);
1506 	}
1507 
1508 	kvm->duration *= NSEC_PER_USEC;   /* convert usec to nsec */
1509 
1510 	/*
1511 	 * target related setups
1512 	 */
1513 	err = target__validate(&kvm->opts.target);
1514 	if (err) {
1515 		target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1516 		ui__warning("%s", errbuf);
1517 	}
1518 
1519 	if (target__none(&kvm->opts.target))
1520 		kvm->opts.target.system_wide = true;
1521 
1522 
1523 	/*
1524 	 * generate the event list
1525 	 */
1526 	kvm->evlist = kvm_live_event_list();
1527 	if (kvm->evlist == NULL) {
1528 		err = -1;
1529 		goto out;
1530 	}
1531 
1532 	symbol_conf.nr_events = kvm->evlist->nr_entries;
1533 
1534 	if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1535 		usage_with_options(live_usage, live_options);
1536 
1537 	/*
1538 	 * perf session
1539 	 */
1540 	kvm->session = perf_session__new(&file, false, &kvm->tool);
1541 	if (kvm->session == NULL) {
1542 		err = -ENOMEM;
1543 		goto out;
1544 	}
1545 	kvm->session->evlist = kvm->evlist;
1546 	perf_session__set_id_hdr_size(kvm->session);
1547 	machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1548 				    kvm->evlist->threads, false);
1549 	err = kvm_live_open_events(kvm);
1550 	if (err)
1551 		goto out;
1552 
1553 	err = kvm_events_live_report(kvm);
1554 
1555 out:
1556 	exit_browser(0);
1557 
1558 	if (kvm->session)
1559 		perf_session__delete(kvm->session);
1560 	kvm->session = NULL;
1561 	if (kvm->evlist) {
1562 		perf_evlist__delete_maps(kvm->evlist);
1563 		perf_evlist__delete(kvm->evlist);
1564 	}
1565 
1566 	return err;
1567 }
1568 #endif
1569 
1570 static void print_kvm_stat_usage(void)
1571 {
1572 	printf("Usage: perf kvm stat <command>\n\n");
1573 
1574 	printf("# Available commands:\n");
1575 	printf("\trecord: record kvm events\n");
1576 	printf("\treport: report statistical data of kvm events\n");
1577 	printf("\tlive:   live reporting of statistical data of kvm events\n");
1578 
1579 	printf("\nOtherwise, it is the alias of 'perf stat':\n");
1580 }
1581 
1582 static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1583 {
1584 	struct perf_kvm_stat kvm = {
1585 		.file_name = file_name,
1586 
1587 		.trace_vcpu	= -1,
1588 		.report_event	= "vmexit",
1589 		.sort_key	= "sample",
1590 
1591 		.exit_reasons = svm_exit_reasons,
1592 		.exit_reasons_size = ARRAY_SIZE(svm_exit_reasons),
1593 		.exit_reasons_isa = "SVM",
1594 	};
1595 
1596 	if (argc == 1) {
1597 		print_kvm_stat_usage();
1598 		goto perf_stat;
1599 	}
1600 
1601 	if (!strncmp(argv[1], "rec", 3))
1602 		return kvm_events_record(&kvm, argc - 1, argv + 1);
1603 
1604 	if (!strncmp(argv[1], "rep", 3))
1605 		return kvm_events_report(&kvm, argc - 1 , argv + 1);
1606 
1607 #ifdef HAVE_TIMERFD_SUPPORT
1608 	if (!strncmp(argv[1], "live", 4))
1609 		return kvm_events_live(&kvm, argc - 1 , argv + 1);
1610 #endif
1611 
1612 perf_stat:
1613 	return cmd_stat(argc, argv, NULL);
1614 }
1615 #endif
1616 
1617 static int __cmd_record(const char *file_name, int argc, const char **argv)
1618 {
1619 	int rec_argc, i = 0, j;
1620 	const char **rec_argv;
1621 
1622 	rec_argc = argc + 2;
1623 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1624 	rec_argv[i++] = strdup("record");
1625 	rec_argv[i++] = strdup("-o");
1626 	rec_argv[i++] = strdup(file_name);
1627 	for (j = 1; j < argc; j++, i++)
1628 		rec_argv[i] = argv[j];
1629 
1630 	BUG_ON(i != rec_argc);
1631 
1632 	return cmd_record(i, rec_argv, NULL);
1633 }
1634 
1635 static int __cmd_report(const char *file_name, int argc, const char **argv)
1636 {
1637 	int rec_argc, i = 0, j;
1638 	const char **rec_argv;
1639 
1640 	rec_argc = argc + 2;
1641 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1642 	rec_argv[i++] = strdup("report");
1643 	rec_argv[i++] = strdup("-i");
1644 	rec_argv[i++] = strdup(file_name);
1645 	for (j = 1; j < argc; j++, i++)
1646 		rec_argv[i] = argv[j];
1647 
1648 	BUG_ON(i != rec_argc);
1649 
1650 	return cmd_report(i, rec_argv, NULL);
1651 }
1652 
1653 static int
1654 __cmd_buildid_list(const char *file_name, int argc, const char **argv)
1655 {
1656 	int rec_argc, i = 0, j;
1657 	const char **rec_argv;
1658 
1659 	rec_argc = argc + 2;
1660 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1661 	rec_argv[i++] = strdup("buildid-list");
1662 	rec_argv[i++] = strdup("-i");
1663 	rec_argv[i++] = strdup(file_name);
1664 	for (j = 1; j < argc; j++, i++)
1665 		rec_argv[i] = argv[j];
1666 
1667 	BUG_ON(i != rec_argc);
1668 
1669 	return cmd_buildid_list(i, rec_argv, NULL);
1670 }
1671 
1672 int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
1673 {
1674 	const char *file_name = NULL;
1675 	const struct option kvm_options[] = {
1676 		OPT_STRING('i', "input", &file_name, "file",
1677 			   "Input file name"),
1678 		OPT_STRING('o', "output", &file_name, "file",
1679 			   "Output file name"),
1680 		OPT_BOOLEAN(0, "guest", &perf_guest,
1681 			    "Collect guest os data"),
1682 		OPT_BOOLEAN(0, "host", &perf_host,
1683 			    "Collect host os data"),
1684 		OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1685 			   "guest mount directory under which every guest os"
1686 			   " instance has a subdir"),
1687 		OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1688 			   "file", "file saving guest os vmlinux"),
1689 		OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1690 			   "file", "file saving guest os /proc/kallsyms"),
1691 		OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1692 			   "file", "file saving guest os /proc/modules"),
1693 		OPT_END()
1694 	};
1695 
1696 
1697 	const char * const kvm_usage[] = {
1698 		"perf kvm [<options>] {top|record|report|diff|buildid-list|stat}",
1699 		NULL
1700 	};
1701 
1702 	perf_host  = 0;
1703 	perf_guest = 1;
1704 
1705 	argc = parse_options(argc, argv, kvm_options, kvm_usage,
1706 			PARSE_OPT_STOP_AT_NON_OPTION);
1707 	if (!argc)
1708 		usage_with_options(kvm_usage, kvm_options);
1709 
1710 	if (!perf_host)
1711 		perf_guest = 1;
1712 
1713 	if (!file_name) {
1714 		if (perf_host && !perf_guest)
1715 			file_name = strdup("perf.data.host");
1716 		else if (!perf_host && perf_guest)
1717 			file_name = strdup("perf.data.guest");
1718 		else
1719 			file_name = strdup("perf.data.kvm");
1720 
1721 		if (!file_name) {
1722 			pr_err("Failed to allocate memory for filename\n");
1723 			return -ENOMEM;
1724 		}
1725 	}
1726 
1727 	if (!strncmp(argv[0], "rec", 3))
1728 		return __cmd_record(file_name, argc, argv);
1729 	else if (!strncmp(argv[0], "rep", 3))
1730 		return __cmd_report(file_name, argc, argv);
1731 	else if (!strncmp(argv[0], "diff", 4))
1732 		return cmd_diff(argc, argv, NULL);
1733 	else if (!strncmp(argv[0], "top", 3))
1734 		return cmd_top(argc, argv, NULL);
1735 	else if (!strncmp(argv[0], "buildid-list", 12))
1736 		return __cmd_buildid_list(file_name, argc, argv);
1737 #if defined(__i386__) || defined(__x86_64__)
1738 	else if (!strncmp(argv[0], "stat", 4))
1739 		return kvm_cmd_stat(file_name, argc, argv);
1740 #endif
1741 	else
1742 		usage_with_options(kvm_usage, kvm_options);
1743 
1744 	return 0;
1745 }
1746