xref: /linux/tools/perf/builtin-kvm.c (revision 4413e16d9d21673bb5048a2e542f1aaa00015c2e)
1 #include "builtin.h"
2 #include "perf.h"
3 
4 #include "util/evsel.h"
5 #include "util/util.h"
6 #include "util/cache.h"
7 #include "util/symbol.h"
8 #include "util/thread.h"
9 #include "util/header.h"
10 #include "util/session.h"
11 
12 #include "util/parse-options.h"
13 #include "util/trace-event.h"
14 #include "util/debug.h"
15 #include "util/debugfs.h"
16 #include "util/tool.h"
17 #include "util/stat.h"
18 
19 #include <sys/prctl.h>
20 
21 #include <semaphore.h>
22 #include <pthread.h>
23 #include <math.h>
24 
25 #include "../../arch/x86/include/asm/svm.h"
26 #include "../../arch/x86/include/asm/vmx.h"
27 #include "../../arch/x86/include/asm/kvm.h"
28 
29 struct event_key {
30 	#define INVALID_KEY     (~0ULL)
31 	u64 key;
32 	int info;
33 };
34 
35 struct kvm_events_ops {
36 	bool (*is_begin_event)(struct perf_evsel *evsel,
37 			       struct perf_sample *sample,
38 			       struct event_key *key);
39 	bool (*is_end_event)(struct perf_evsel *evsel,
40 			     struct perf_sample *sample, struct event_key *key);
41 	void (*decode_key)(struct event_key *key, char decode[20]);
42 	const char *name;
43 };
44 
45 static void exit_event_get_key(struct perf_evsel *evsel,
46 			       struct perf_sample *sample,
47 			       struct event_key *key)
48 {
49 	key->info = 0;
50 	key->key = perf_evsel__intval(evsel, sample, "exit_reason");
51 }
52 
53 static bool kvm_exit_event(struct perf_evsel *evsel)
54 {
55 	return !strcmp(evsel->name, "kvm:kvm_exit");
56 }
57 
58 static bool exit_event_begin(struct perf_evsel *evsel,
59 			     struct perf_sample *sample, struct event_key *key)
60 {
61 	if (kvm_exit_event(evsel)) {
62 		exit_event_get_key(evsel, sample, key);
63 		return true;
64 	}
65 
66 	return false;
67 }
68 
69 static bool kvm_entry_event(struct perf_evsel *evsel)
70 {
71 	return !strcmp(evsel->name, "kvm:kvm_entry");
72 }
73 
74 static bool exit_event_end(struct perf_evsel *evsel,
75 			   struct perf_sample *sample __maybe_unused,
76 			   struct event_key *key __maybe_unused)
77 {
78 	return kvm_entry_event(evsel);
79 }
80 
81 struct exit_reasons_table {
82 	unsigned long exit_code;
83 	const char *reason;
84 };
85 
86 struct exit_reasons_table vmx_exit_reasons[] = {
87 	VMX_EXIT_REASONS
88 };
89 
90 struct exit_reasons_table svm_exit_reasons[] = {
91 	SVM_EXIT_REASONS
92 };
93 
94 static int cpu_isa;
95 
96 static const char *get_exit_reason(u64 exit_code)
97 {
98 	int table_size = ARRAY_SIZE(svm_exit_reasons);
99 	struct exit_reasons_table *table = svm_exit_reasons;
100 
101 	if (cpu_isa == 1) {
102 		table = vmx_exit_reasons;
103 		table_size = ARRAY_SIZE(vmx_exit_reasons);
104 	}
105 
106 	while (table_size--) {
107 		if (table->exit_code == exit_code)
108 			return table->reason;
109 		table++;
110 	}
111 
112 	pr_err("unknown kvm exit code:%lld on %s\n",
113 		(unsigned long long)exit_code, cpu_isa ? "VMX" : "SVM");
114 	return "UNKNOWN";
115 }
116 
117 static void exit_event_decode_key(struct event_key *key, char decode[20])
118 {
119 	const char *exit_reason = get_exit_reason(key->key);
120 
121 	scnprintf(decode, 20, "%s", exit_reason);
122 }
123 
124 static struct kvm_events_ops exit_events = {
125 	.is_begin_event = exit_event_begin,
126 	.is_end_event = exit_event_end,
127 	.decode_key = exit_event_decode_key,
128 	.name = "VM-EXIT"
129 };
130 
131     /*
132      * For the mmio events, we treat:
133      * the time of MMIO write: kvm_mmio(KVM_TRACE_MMIO_WRITE...) -> kvm_entry
134      * the time of MMIO read: kvm_exit -> kvm_mmio(KVM_TRACE_MMIO_READ...).
135      */
136 static void mmio_event_get_key(struct perf_evsel *evsel, struct perf_sample *sample,
137 			       struct event_key *key)
138 {
139 	key->key  = perf_evsel__intval(evsel, sample, "gpa");
140 	key->info = perf_evsel__intval(evsel, sample, "type");
141 }
142 
143 #define KVM_TRACE_MMIO_READ_UNSATISFIED 0
144 #define KVM_TRACE_MMIO_READ 1
145 #define KVM_TRACE_MMIO_WRITE 2
146 
147 static bool mmio_event_begin(struct perf_evsel *evsel,
148 			     struct perf_sample *sample, struct event_key *key)
149 {
150 	/* MMIO read begin event in kernel. */
151 	if (kvm_exit_event(evsel))
152 		return true;
153 
154 	/* MMIO write begin event in kernel. */
155 	if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
156 	    perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_WRITE) {
157 		mmio_event_get_key(evsel, sample, key);
158 		return true;
159 	}
160 
161 	return false;
162 }
163 
164 static bool mmio_event_end(struct perf_evsel *evsel, struct perf_sample *sample,
165 			   struct event_key *key)
166 {
167 	/* MMIO write end event in kernel. */
168 	if (kvm_entry_event(evsel))
169 		return true;
170 
171 	/* MMIO read end event in kernel.*/
172 	if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
173 	    perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_READ) {
174 		mmio_event_get_key(evsel, sample, key);
175 		return true;
176 	}
177 
178 	return false;
179 }
180 
181 static void mmio_event_decode_key(struct event_key *key, char decode[20])
182 {
183 	scnprintf(decode, 20, "%#lx:%s", (unsigned long)key->key,
184 				key->info == KVM_TRACE_MMIO_WRITE ? "W" : "R");
185 }
186 
187 static struct kvm_events_ops mmio_events = {
188 	.is_begin_event = mmio_event_begin,
189 	.is_end_event = mmio_event_end,
190 	.decode_key = mmio_event_decode_key,
191 	.name = "MMIO Access"
192 };
193 
194  /* The time of emulation pio access is from kvm_pio to kvm_entry. */
195 static void ioport_event_get_key(struct perf_evsel *evsel,
196 				 struct perf_sample *sample,
197 				 struct event_key *key)
198 {
199 	key->key  = perf_evsel__intval(evsel, sample, "port");
200 	key->info = perf_evsel__intval(evsel, sample, "rw");
201 }
202 
203 static bool ioport_event_begin(struct perf_evsel *evsel,
204 			       struct perf_sample *sample,
205 			       struct event_key *key)
206 {
207 	if (!strcmp(evsel->name, "kvm:kvm_pio")) {
208 		ioport_event_get_key(evsel, sample, key);
209 		return true;
210 	}
211 
212 	return false;
213 }
214 
215 static bool ioport_event_end(struct perf_evsel *evsel,
216 			     struct perf_sample *sample __maybe_unused,
217 			     struct event_key *key __maybe_unused)
218 {
219 	return kvm_entry_event(evsel);
220 }
221 
222 static void ioport_event_decode_key(struct event_key *key, char decode[20])
223 {
224 	scnprintf(decode, 20, "%#llx:%s", (unsigned long long)key->key,
225 				key->info ? "POUT" : "PIN");
226 }
227 
228 static struct kvm_events_ops ioport_events = {
229 	.is_begin_event = ioport_event_begin,
230 	.is_end_event = ioport_event_end,
231 	.decode_key = ioport_event_decode_key,
232 	.name = "IO Port Access"
233 };
234 
235 static const char *report_event = "vmexit";
236 struct kvm_events_ops *events_ops;
237 
238 static bool register_kvm_events_ops(void)
239 {
240 	bool ret = true;
241 
242 	if (!strcmp(report_event, "vmexit"))
243 		events_ops = &exit_events;
244 	else if (!strcmp(report_event, "mmio"))
245 		events_ops = &mmio_events;
246 	else if (!strcmp(report_event, "ioport"))
247 		events_ops = &ioport_events;
248 	else {
249 		pr_err("Unknown report event:%s\n", report_event);
250 		ret = false;
251 	}
252 
253 	return ret;
254 }
255 
256 struct kvm_event_stats {
257 	u64 time;
258 	struct stats stats;
259 };
260 
261 struct kvm_event {
262 	struct list_head hash_entry;
263 	struct rb_node rb;
264 
265 	struct event_key key;
266 
267 	struct kvm_event_stats total;
268 
269 	#define DEFAULT_VCPU_NUM 8
270 	int max_vcpu;
271 	struct kvm_event_stats *vcpu;
272 };
273 
274 struct vcpu_event_record {
275 	int vcpu_id;
276 	u64 start_time;
277 	struct kvm_event *last_event;
278 };
279 
280 #define EVENTS_BITS			12
281 #define EVENTS_CACHE_SIZE	(1UL << EVENTS_BITS)
282 
283 static u64 total_time;
284 static u64 total_count;
285 static struct list_head kvm_events_cache[EVENTS_CACHE_SIZE];
286 
287 static void init_kvm_event_record(void)
288 {
289 	int i;
290 
291 	for (i = 0; i < (int)EVENTS_CACHE_SIZE; i++)
292 		INIT_LIST_HEAD(&kvm_events_cache[i]);
293 }
294 
295 static int kvm_events_hash_fn(u64 key)
296 {
297 	return key & (EVENTS_CACHE_SIZE - 1);
298 }
299 
300 static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
301 {
302 	int old_max_vcpu = event->max_vcpu;
303 
304 	if (vcpu_id < event->max_vcpu)
305 		return true;
306 
307 	while (event->max_vcpu <= vcpu_id)
308 		event->max_vcpu += DEFAULT_VCPU_NUM;
309 
310 	event->vcpu = realloc(event->vcpu,
311 			      event->max_vcpu * sizeof(*event->vcpu));
312 	if (!event->vcpu) {
313 		pr_err("Not enough memory\n");
314 		return false;
315 	}
316 
317 	memset(event->vcpu + old_max_vcpu, 0,
318 	       (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
319 	return true;
320 }
321 
322 static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
323 {
324 	struct kvm_event *event;
325 
326 	event = zalloc(sizeof(*event));
327 	if (!event) {
328 		pr_err("Not enough memory\n");
329 		return NULL;
330 	}
331 
332 	event->key = *key;
333 	return event;
334 }
335 
336 static struct kvm_event *find_create_kvm_event(struct event_key *key)
337 {
338 	struct kvm_event *event;
339 	struct list_head *head;
340 
341 	BUG_ON(key->key == INVALID_KEY);
342 
343 	head = &kvm_events_cache[kvm_events_hash_fn(key->key)];
344 	list_for_each_entry(event, head, hash_entry)
345 		if (event->key.key == key->key && event->key.info == key->info)
346 			return event;
347 
348 	event = kvm_alloc_init_event(key);
349 	if (!event)
350 		return NULL;
351 
352 	list_add(&event->hash_entry, head);
353 	return event;
354 }
355 
356 static bool handle_begin_event(struct vcpu_event_record *vcpu_record,
357 			       struct event_key *key, u64 timestamp)
358 {
359 	struct kvm_event *event = NULL;
360 
361 	if (key->key != INVALID_KEY)
362 		event = find_create_kvm_event(key);
363 
364 	vcpu_record->last_event = event;
365 	vcpu_record->start_time = timestamp;
366 	return true;
367 }
368 
369 static void
370 kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
371 {
372 	kvm_stats->time += time_diff;
373 	update_stats(&kvm_stats->stats, time_diff);
374 }
375 
376 static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
377 {
378 	struct kvm_event_stats *kvm_stats = &event->total;
379 
380 	if (vcpu_id != -1)
381 		kvm_stats = &event->vcpu[vcpu_id];
382 
383 	return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
384 				avg_stats(&kvm_stats->stats));
385 }
386 
387 static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
388 			     u64 time_diff)
389 {
390 	kvm_update_event_stats(&event->total, time_diff);
391 
392 	if (!kvm_event_expand(event, vcpu_id))
393 		return false;
394 
395 	kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
396 	return true;
397 }
398 
399 static bool handle_end_event(struct vcpu_event_record *vcpu_record,
400 			     struct event_key *key, u64 timestamp)
401 {
402 	struct kvm_event *event;
403 	u64 time_begin, time_diff;
404 
405 	event = vcpu_record->last_event;
406 	time_begin = vcpu_record->start_time;
407 
408 	/* The begin event is not caught. */
409 	if (!time_begin)
410 		return true;
411 
412 	/*
413 	 * In some case, the 'begin event' only records the start timestamp,
414 	 * the actual event is recognized in the 'end event' (e.g. mmio-event).
415 	 */
416 
417 	/* Both begin and end events did not get the key. */
418 	if (!event && key->key == INVALID_KEY)
419 		return true;
420 
421 	if (!event)
422 		event = find_create_kvm_event(key);
423 
424 	if (!event)
425 		return false;
426 
427 	vcpu_record->last_event = NULL;
428 	vcpu_record->start_time = 0;
429 
430 	BUG_ON(timestamp < time_begin);
431 
432 	time_diff = timestamp - time_begin;
433 	return update_kvm_event(event, vcpu_record->vcpu_id, time_diff);
434 }
435 
436 static
437 struct vcpu_event_record *per_vcpu_record(struct thread *thread,
438 					  struct perf_evsel *evsel,
439 					  struct perf_sample *sample)
440 {
441 	/* Only kvm_entry records vcpu id. */
442 	if (!thread->priv && kvm_entry_event(evsel)) {
443 		struct vcpu_event_record *vcpu_record;
444 
445 		vcpu_record = zalloc(sizeof(*vcpu_record));
446 		if (!vcpu_record) {
447 			pr_err("%s: Not enough memory\n", __func__);
448 			return NULL;
449 		}
450 
451 		vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample, "vcpu_id");
452 		thread->priv = vcpu_record;
453 	}
454 
455 	return thread->priv;
456 }
457 
458 static bool handle_kvm_event(struct thread *thread, struct perf_evsel *evsel,
459 			     struct perf_sample *sample)
460 {
461 	struct vcpu_event_record *vcpu_record;
462 	struct event_key key = {.key = INVALID_KEY};
463 
464 	vcpu_record = per_vcpu_record(thread, evsel, sample);
465 	if (!vcpu_record)
466 		return true;
467 
468 	if (events_ops->is_begin_event(evsel, sample, &key))
469 		return handle_begin_event(vcpu_record, &key, sample->time);
470 
471 	if (events_ops->is_end_event(evsel, sample, &key))
472 		return handle_end_event(vcpu_record, &key, sample->time);
473 
474 	return true;
475 }
476 
477 typedef int (*key_cmp_fun)(struct kvm_event*, struct kvm_event*, int);
478 struct kvm_event_key {
479 	const char *name;
480 	key_cmp_fun key;
481 };
482 
483 static int trace_vcpu = -1;
484 #define GET_EVENT_KEY(func, field)					\
485 static u64 get_event_ ##func(struct kvm_event *event, int vcpu)		\
486 {									\
487 	if (vcpu == -1)							\
488 		return event->total.field;				\
489 									\
490 	if (vcpu >= event->max_vcpu)					\
491 		return 0;						\
492 									\
493 	return event->vcpu[vcpu].field;					\
494 }
495 
496 #define COMPARE_EVENT_KEY(func, field)					\
497 GET_EVENT_KEY(func, field)						\
498 static int compare_kvm_event_ ## func(struct kvm_event *one,		\
499 					struct kvm_event *two, int vcpu)\
500 {									\
501 	return get_event_ ##func(one, vcpu) >				\
502 				get_event_ ##func(two, vcpu);		\
503 }
504 
505 GET_EVENT_KEY(time, time);
506 COMPARE_EVENT_KEY(count, stats.n);
507 COMPARE_EVENT_KEY(mean, stats.mean);
508 
509 #define DEF_SORT_NAME_KEY(name, compare_key)				\
510 	{ #name, compare_kvm_event_ ## compare_key }
511 
512 static struct kvm_event_key keys[] = {
513 	DEF_SORT_NAME_KEY(sample, count),
514 	DEF_SORT_NAME_KEY(time, mean),
515 	{ NULL, NULL }
516 };
517 
518 static const char *sort_key = "sample";
519 static key_cmp_fun compare;
520 
521 static bool select_key(void)
522 {
523 	int i;
524 
525 	for (i = 0; keys[i].name; i++) {
526 		if (!strcmp(keys[i].name, sort_key)) {
527 			compare = keys[i].key;
528 			return true;
529 		}
530 	}
531 
532 	pr_err("Unknown compare key:%s\n", sort_key);
533 	return false;
534 }
535 
536 static struct rb_root result;
537 static void insert_to_result(struct kvm_event *event, key_cmp_fun bigger,
538 			     int vcpu)
539 {
540 	struct rb_node **rb = &result.rb_node;
541 	struct rb_node *parent = NULL;
542 	struct kvm_event *p;
543 
544 	while (*rb) {
545 		p = container_of(*rb, struct kvm_event, rb);
546 		parent = *rb;
547 
548 		if (bigger(event, p, vcpu))
549 			rb = &(*rb)->rb_left;
550 		else
551 			rb = &(*rb)->rb_right;
552 	}
553 
554 	rb_link_node(&event->rb, parent, rb);
555 	rb_insert_color(&event->rb, &result);
556 }
557 
558 static void update_total_count(struct kvm_event *event, int vcpu)
559 {
560 	total_count += get_event_count(event, vcpu);
561 	total_time += get_event_time(event, vcpu);
562 }
563 
564 static bool event_is_valid(struct kvm_event *event, int vcpu)
565 {
566 	return !!get_event_count(event, vcpu);
567 }
568 
569 static void sort_result(int vcpu)
570 {
571 	unsigned int i;
572 	struct kvm_event *event;
573 
574 	for (i = 0; i < EVENTS_CACHE_SIZE; i++)
575 		list_for_each_entry(event, &kvm_events_cache[i], hash_entry)
576 			if (event_is_valid(event, vcpu)) {
577 				update_total_count(event, vcpu);
578 				insert_to_result(event, compare, vcpu);
579 			}
580 }
581 
582 /* returns left most element of result, and erase it */
583 static struct kvm_event *pop_from_result(void)
584 {
585 	struct rb_node *node = rb_first(&result);
586 
587 	if (!node)
588 		return NULL;
589 
590 	rb_erase(node, &result);
591 	return container_of(node, struct kvm_event, rb);
592 }
593 
594 static void print_vcpu_info(int vcpu)
595 {
596 	pr_info("Analyze events for ");
597 
598 	if (vcpu == -1)
599 		pr_info("all VCPUs:\n\n");
600 	else
601 		pr_info("VCPU %d:\n\n", vcpu);
602 }
603 
604 static void print_result(int vcpu)
605 {
606 	char decode[20];
607 	struct kvm_event *event;
608 
609 	pr_info("\n\n");
610 	print_vcpu_info(vcpu);
611 	pr_info("%20s ", events_ops->name);
612 	pr_info("%10s ", "Samples");
613 	pr_info("%9s ", "Samples%");
614 
615 	pr_info("%9s ", "Time%");
616 	pr_info("%16s ", "Avg time");
617 	pr_info("\n\n");
618 
619 	while ((event = pop_from_result())) {
620 		u64 ecount, etime;
621 
622 		ecount = get_event_count(event, vcpu);
623 		etime = get_event_time(event, vcpu);
624 
625 		events_ops->decode_key(&event->key, decode);
626 		pr_info("%20s ", decode);
627 		pr_info("%10llu ", (unsigned long long)ecount);
628 		pr_info("%8.2f%% ", (double)ecount / total_count * 100);
629 		pr_info("%8.2f%% ", (double)etime / total_time * 100);
630 		pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount/1e3,
631 			kvm_event_rel_stddev(vcpu, event));
632 		pr_info("\n");
633 	}
634 
635 	pr_info("\nTotal Samples:%lld, Total events handled time:%.2fus.\n\n",
636 		(unsigned long long)total_count, total_time / 1e3);
637 }
638 
639 static int process_sample_event(struct perf_tool *tool __maybe_unused,
640 				union perf_event *event,
641 				struct perf_sample *sample,
642 				struct perf_evsel *evsel,
643 				struct machine *machine)
644 {
645 	struct thread *thread = machine__findnew_thread(machine, sample->tid);
646 
647 	if (thread == NULL) {
648 		pr_debug("problem processing %d event, skipping it.\n",
649 			event->header.type);
650 		return -1;
651 	}
652 
653 	if (!handle_kvm_event(thread, evsel, sample))
654 		return -1;
655 
656 	return 0;
657 }
658 
659 static struct perf_tool eops = {
660 	.sample			= process_sample_event,
661 	.comm			= perf_event__process_comm,
662 	.ordered_samples	= true,
663 };
664 
665 static int get_cpu_isa(struct perf_session *session)
666 {
667 	char *cpuid = session->header.env.cpuid;
668 	int isa;
669 
670 	if (strstr(cpuid, "Intel"))
671 		isa = 1;
672 	else if (strstr(cpuid, "AMD"))
673 		isa = 0;
674 	else {
675 		pr_err("CPU %s is not supported.\n", cpuid);
676 		isa = -ENOTSUP;
677 	}
678 
679 	return isa;
680 }
681 
682 static const char *file_name;
683 
684 static int read_events(void)
685 {
686 	struct perf_session *kvm_session;
687 	int ret;
688 
689 	kvm_session = perf_session__new(file_name, O_RDONLY, 0, false, &eops);
690 	if (!kvm_session) {
691 		pr_err("Initializing perf session failed\n");
692 		return -EINVAL;
693 	}
694 
695 	if (!perf_session__has_traces(kvm_session, "kvm record"))
696 		return -EINVAL;
697 
698 	/*
699 	 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
700 	 * traced in the old kernel.
701 	 */
702 	ret = get_cpu_isa(kvm_session);
703 
704 	if (ret < 0)
705 		return ret;
706 
707 	cpu_isa = ret;
708 
709 	return perf_session__process_events(kvm_session, &eops);
710 }
711 
712 static bool verify_vcpu(int vcpu)
713 {
714 	if (vcpu != -1 && vcpu < 0) {
715 		pr_err("Invalid vcpu:%d.\n", vcpu);
716 		return false;
717 	}
718 
719 	return true;
720 }
721 
722 static int kvm_events_report_vcpu(int vcpu)
723 {
724 	int ret = -EINVAL;
725 
726 	if (!verify_vcpu(vcpu))
727 		goto exit;
728 
729 	if (!select_key())
730 		goto exit;
731 
732 	if (!register_kvm_events_ops())
733 		goto exit;
734 
735 	init_kvm_event_record();
736 	setup_pager();
737 
738 	ret = read_events();
739 	if (ret)
740 		goto exit;
741 
742 	sort_result(vcpu);
743 	print_result(vcpu);
744 exit:
745 	return ret;
746 }
747 
748 static const char * const record_args[] = {
749 	"record",
750 	"-R",
751 	"-f",
752 	"-m", "1024",
753 	"-c", "1",
754 	"-e", "kvm:kvm_entry",
755 	"-e", "kvm:kvm_exit",
756 	"-e", "kvm:kvm_mmio",
757 	"-e", "kvm:kvm_pio",
758 };
759 
760 #define STRDUP_FAIL_EXIT(s)		\
761 	({	char *_p;		\
762 	_p = strdup(s);		\
763 		if (!_p)		\
764 			return -ENOMEM;	\
765 		_p;			\
766 	})
767 
768 static int kvm_events_record(int argc, const char **argv)
769 {
770 	unsigned int rec_argc, i, j;
771 	const char **rec_argv;
772 
773 	rec_argc = ARRAY_SIZE(record_args) + argc + 2;
774 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
775 
776 	if (rec_argv == NULL)
777 		return -ENOMEM;
778 
779 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
780 		rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
781 
782 	rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
783 	rec_argv[i++] = STRDUP_FAIL_EXIT(file_name);
784 
785 	for (j = 1; j < (unsigned int)argc; j++, i++)
786 		rec_argv[i] = argv[j];
787 
788 	return cmd_record(i, rec_argv, NULL);
789 }
790 
791 static const char * const kvm_events_report_usage[] = {
792 	"perf kvm stat report [<options>]",
793 	NULL
794 };
795 
796 static const struct option kvm_events_report_options[] = {
797 	OPT_STRING(0, "event", &report_event, "report event",
798 		    "event for reporting: vmexit, mmio, ioport"),
799 	OPT_INTEGER(0, "vcpu", &trace_vcpu,
800 		    "vcpu id to report"),
801 	OPT_STRING('k', "key", &sort_key, "sort-key",
802 		    "key for sorting: sample(sort by samples number)"
803 		    " time (sort by avg time)"),
804 	OPT_END()
805 };
806 
807 static int kvm_events_report(int argc, const char **argv)
808 {
809 	symbol__init();
810 
811 	if (argc) {
812 		argc = parse_options(argc, argv,
813 				     kvm_events_report_options,
814 				     kvm_events_report_usage, 0);
815 		if (argc)
816 			usage_with_options(kvm_events_report_usage,
817 					   kvm_events_report_options);
818 	}
819 
820 	return kvm_events_report_vcpu(trace_vcpu);
821 }
822 
823 static void print_kvm_stat_usage(void)
824 {
825 	printf("Usage: perf kvm stat <command>\n\n");
826 
827 	printf("# Available commands:\n");
828 	printf("\trecord: record kvm events\n");
829 	printf("\treport: report statistical data of kvm events\n");
830 
831 	printf("\nOtherwise, it is the alias of 'perf stat':\n");
832 }
833 
834 static int kvm_cmd_stat(int argc, const char **argv)
835 {
836 	if (argc == 1) {
837 		print_kvm_stat_usage();
838 		goto perf_stat;
839 	}
840 
841 	if (!strncmp(argv[1], "rec", 3))
842 		return kvm_events_record(argc - 1, argv + 1);
843 
844 	if (!strncmp(argv[1], "rep", 3))
845 		return kvm_events_report(argc - 1 , argv + 1);
846 
847 perf_stat:
848 	return cmd_stat(argc, argv, NULL);
849 }
850 
851 static char			name_buffer[256];
852 
853 static const char * const kvm_usage[] = {
854 	"perf kvm [<options>] {top|record|report|diff|buildid-list|stat}",
855 	NULL
856 };
857 
858 static const struct option kvm_options[] = {
859 	OPT_STRING('i', "input", &file_name, "file",
860 		   "Input file name"),
861 	OPT_STRING('o', "output", &file_name, "file",
862 		   "Output file name"),
863 	OPT_BOOLEAN(0, "guest", &perf_guest,
864 		    "Collect guest os data"),
865 	OPT_BOOLEAN(0, "host", &perf_host,
866 		    "Collect host os data"),
867 	OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
868 		   "guest mount directory under which every guest os"
869 		   " instance has a subdir"),
870 	OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
871 		   "file", "file saving guest os vmlinux"),
872 	OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
873 		   "file", "file saving guest os /proc/kallsyms"),
874 	OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
875 		   "file", "file saving guest os /proc/modules"),
876 	OPT_END()
877 };
878 
879 static int __cmd_record(int argc, const char **argv)
880 {
881 	int rec_argc, i = 0, j;
882 	const char **rec_argv;
883 
884 	rec_argc = argc + 2;
885 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
886 	rec_argv[i++] = strdup("record");
887 	rec_argv[i++] = strdup("-o");
888 	rec_argv[i++] = strdup(file_name);
889 	for (j = 1; j < argc; j++, i++)
890 		rec_argv[i] = argv[j];
891 
892 	BUG_ON(i != rec_argc);
893 
894 	return cmd_record(i, rec_argv, NULL);
895 }
896 
897 static int __cmd_report(int argc, const char **argv)
898 {
899 	int rec_argc, i = 0, j;
900 	const char **rec_argv;
901 
902 	rec_argc = argc + 2;
903 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
904 	rec_argv[i++] = strdup("report");
905 	rec_argv[i++] = strdup("-i");
906 	rec_argv[i++] = strdup(file_name);
907 	for (j = 1; j < argc; j++, i++)
908 		rec_argv[i] = argv[j];
909 
910 	BUG_ON(i != rec_argc);
911 
912 	return cmd_report(i, rec_argv, NULL);
913 }
914 
915 static int __cmd_buildid_list(int argc, const char **argv)
916 {
917 	int rec_argc, i = 0, j;
918 	const char **rec_argv;
919 
920 	rec_argc = argc + 2;
921 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
922 	rec_argv[i++] = strdup("buildid-list");
923 	rec_argv[i++] = strdup("-i");
924 	rec_argv[i++] = strdup(file_name);
925 	for (j = 1; j < argc; j++, i++)
926 		rec_argv[i] = argv[j];
927 
928 	BUG_ON(i != rec_argc);
929 
930 	return cmd_buildid_list(i, rec_argv, NULL);
931 }
932 
933 int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
934 {
935 	perf_host  = 0;
936 	perf_guest = 1;
937 
938 	argc = parse_options(argc, argv, kvm_options, kvm_usage,
939 			PARSE_OPT_STOP_AT_NON_OPTION);
940 	if (!argc)
941 		usage_with_options(kvm_usage, kvm_options);
942 
943 	if (!perf_host)
944 		perf_guest = 1;
945 
946 	if (!file_name) {
947 		if (perf_host && !perf_guest)
948 			sprintf(name_buffer, "perf.data.host");
949 		else if (!perf_host && perf_guest)
950 			sprintf(name_buffer, "perf.data.guest");
951 		else
952 			sprintf(name_buffer, "perf.data.kvm");
953 		file_name = name_buffer;
954 	}
955 
956 	if (!strncmp(argv[0], "rec", 3))
957 		return __cmd_record(argc, argv);
958 	else if (!strncmp(argv[0], "rep", 3))
959 		return __cmd_report(argc, argv);
960 	else if (!strncmp(argv[0], "diff", 4))
961 		return cmd_diff(argc, argv, NULL);
962 	else if (!strncmp(argv[0], "top", 3))
963 		return cmd_top(argc, argv, NULL);
964 	else if (!strncmp(argv[0], "buildid-list", 12))
965 		return __cmd_buildid_list(argc, argv);
966 	else if (!strncmp(argv[0], "stat", 4))
967 		return kvm_cmd_stat(argc, argv);
968 	else
969 		usage_with_options(kvm_usage, kvm_options);
970 
971 	return 0;
972 }
973