xref: /openbmc/linux/tools/perf/builtin-kvm.c (revision 05bcf503)
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_event_stats {
36 	u64 time;
37 	struct stats stats;
38 };
39 
40 struct kvm_event {
41 	struct list_head hash_entry;
42 	struct rb_node rb;
43 
44 	struct event_key key;
45 
46 	struct kvm_event_stats total;
47 
48 	#define DEFAULT_VCPU_NUM 8
49 	int max_vcpu;
50 	struct kvm_event_stats *vcpu;
51 };
52 
53 typedef int (*key_cmp_fun)(struct kvm_event*, struct kvm_event*, int);
54 
55 struct kvm_event_key {
56 	const char *name;
57 	key_cmp_fun key;
58 };
59 
60 
61 struct perf_kvm;
62 
63 struct kvm_events_ops {
64 	bool (*is_begin_event)(struct perf_evsel *evsel,
65 			       struct perf_sample *sample,
66 			       struct event_key *key);
67 	bool (*is_end_event)(struct perf_evsel *evsel,
68 			     struct perf_sample *sample, struct event_key *key);
69 	void (*decode_key)(struct perf_kvm *kvm, struct event_key *key,
70 			   char decode[20]);
71 	const char *name;
72 };
73 
74 struct exit_reasons_table {
75 	unsigned long exit_code;
76 	const char *reason;
77 };
78 
79 #define EVENTS_BITS		12
80 #define EVENTS_CACHE_SIZE	(1UL << EVENTS_BITS)
81 
82 struct perf_kvm {
83 	struct perf_tool    tool;
84 	struct perf_session *session;
85 
86 	const char *file_name;
87 	const char *report_event;
88 	const char *sort_key;
89 	int trace_vcpu;
90 
91 	struct exit_reasons_table *exit_reasons;
92 	int exit_reasons_size;
93 	const char *exit_reasons_isa;
94 
95 	struct kvm_events_ops *events_ops;
96 	key_cmp_fun compare;
97 	struct list_head kvm_events_cache[EVENTS_CACHE_SIZE];
98 	u64 total_time;
99 	u64 total_count;
100 
101 	struct rb_root result;
102 };
103 
104 
105 static void exit_event_get_key(struct perf_evsel *evsel,
106 			       struct perf_sample *sample,
107 			       struct event_key *key)
108 {
109 	key->info = 0;
110 	key->key = perf_evsel__intval(evsel, sample, "exit_reason");
111 }
112 
113 static bool kvm_exit_event(struct perf_evsel *evsel)
114 {
115 	return !strcmp(evsel->name, "kvm:kvm_exit");
116 }
117 
118 static bool exit_event_begin(struct perf_evsel *evsel,
119 			     struct perf_sample *sample, struct event_key *key)
120 {
121 	if (kvm_exit_event(evsel)) {
122 		exit_event_get_key(evsel, sample, key);
123 		return true;
124 	}
125 
126 	return false;
127 }
128 
129 static bool kvm_entry_event(struct perf_evsel *evsel)
130 {
131 	return !strcmp(evsel->name, "kvm:kvm_entry");
132 }
133 
134 static bool exit_event_end(struct perf_evsel *evsel,
135 			   struct perf_sample *sample __maybe_unused,
136 			   struct event_key *key __maybe_unused)
137 {
138 	return kvm_entry_event(evsel);
139 }
140 
141 static struct exit_reasons_table vmx_exit_reasons[] = {
142 	VMX_EXIT_REASONS
143 };
144 
145 static struct exit_reasons_table svm_exit_reasons[] = {
146 	SVM_EXIT_REASONS
147 };
148 
149 static const char *get_exit_reason(struct perf_kvm *kvm, u64 exit_code)
150 {
151 	int i = kvm->exit_reasons_size;
152 	struct exit_reasons_table *tbl = kvm->exit_reasons;
153 
154 	while (i--) {
155 		if (tbl->exit_code == exit_code)
156 			return tbl->reason;
157 		tbl++;
158 	}
159 
160 	pr_err("unknown kvm exit code:%lld on %s\n",
161 		(unsigned long long)exit_code, kvm->exit_reasons_isa);
162 	return "UNKNOWN";
163 }
164 
165 static void exit_event_decode_key(struct perf_kvm *kvm,
166 				  struct event_key *key,
167 				  char decode[20])
168 {
169 	const char *exit_reason = get_exit_reason(kvm, key->key);
170 
171 	scnprintf(decode, 20, "%s", exit_reason);
172 }
173 
174 static struct kvm_events_ops exit_events = {
175 	.is_begin_event = exit_event_begin,
176 	.is_end_event = exit_event_end,
177 	.decode_key = exit_event_decode_key,
178 	.name = "VM-EXIT"
179 };
180 
181 /*
182  * For the mmio events, we treat:
183  * the time of MMIO write: kvm_mmio(KVM_TRACE_MMIO_WRITE...) -> kvm_entry
184  * the time of MMIO read: kvm_exit -> kvm_mmio(KVM_TRACE_MMIO_READ...).
185  */
186 static void mmio_event_get_key(struct perf_evsel *evsel, struct perf_sample *sample,
187 			       struct event_key *key)
188 {
189 	key->key  = perf_evsel__intval(evsel, sample, "gpa");
190 	key->info = perf_evsel__intval(evsel, sample, "type");
191 }
192 
193 #define KVM_TRACE_MMIO_READ_UNSATISFIED 0
194 #define KVM_TRACE_MMIO_READ 1
195 #define KVM_TRACE_MMIO_WRITE 2
196 
197 static bool mmio_event_begin(struct perf_evsel *evsel,
198 			     struct perf_sample *sample, struct event_key *key)
199 {
200 	/* MMIO read begin event in kernel. */
201 	if (kvm_exit_event(evsel))
202 		return true;
203 
204 	/* MMIO write begin event in kernel. */
205 	if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
206 	    perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_WRITE) {
207 		mmio_event_get_key(evsel, sample, key);
208 		return true;
209 	}
210 
211 	return false;
212 }
213 
214 static bool mmio_event_end(struct perf_evsel *evsel, struct perf_sample *sample,
215 			   struct event_key *key)
216 {
217 	/* MMIO write end event in kernel. */
218 	if (kvm_entry_event(evsel))
219 		return true;
220 
221 	/* MMIO read end event in kernel.*/
222 	if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
223 	    perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_READ) {
224 		mmio_event_get_key(evsel, sample, key);
225 		return true;
226 	}
227 
228 	return false;
229 }
230 
231 static void mmio_event_decode_key(struct perf_kvm *kvm __maybe_unused,
232 				  struct event_key *key,
233 				  char decode[20])
234 {
235 	scnprintf(decode, 20, "%#lx:%s", (unsigned long)key->key,
236 				key->info == KVM_TRACE_MMIO_WRITE ? "W" : "R");
237 }
238 
239 static struct kvm_events_ops mmio_events = {
240 	.is_begin_event = mmio_event_begin,
241 	.is_end_event = mmio_event_end,
242 	.decode_key = mmio_event_decode_key,
243 	.name = "MMIO Access"
244 };
245 
246  /* The time of emulation pio access is from kvm_pio to kvm_entry. */
247 static void ioport_event_get_key(struct perf_evsel *evsel,
248 				 struct perf_sample *sample,
249 				 struct event_key *key)
250 {
251 	key->key  = perf_evsel__intval(evsel, sample, "port");
252 	key->info = perf_evsel__intval(evsel, sample, "rw");
253 }
254 
255 static bool ioport_event_begin(struct perf_evsel *evsel,
256 			       struct perf_sample *sample,
257 			       struct event_key *key)
258 {
259 	if (!strcmp(evsel->name, "kvm:kvm_pio")) {
260 		ioport_event_get_key(evsel, sample, key);
261 		return true;
262 	}
263 
264 	return false;
265 }
266 
267 static bool ioport_event_end(struct perf_evsel *evsel,
268 			     struct perf_sample *sample __maybe_unused,
269 			     struct event_key *key __maybe_unused)
270 {
271 	return kvm_entry_event(evsel);
272 }
273 
274 static void ioport_event_decode_key(struct perf_kvm *kvm __maybe_unused,
275 				    struct event_key *key,
276 				    char decode[20])
277 {
278 	scnprintf(decode, 20, "%#llx:%s", (unsigned long long)key->key,
279 				key->info ? "POUT" : "PIN");
280 }
281 
282 static struct kvm_events_ops ioport_events = {
283 	.is_begin_event = ioport_event_begin,
284 	.is_end_event = ioport_event_end,
285 	.decode_key = ioport_event_decode_key,
286 	.name = "IO Port Access"
287 };
288 
289 static bool register_kvm_events_ops(struct perf_kvm *kvm)
290 {
291 	bool ret = true;
292 
293 	if (!strcmp(kvm->report_event, "vmexit"))
294 		kvm->events_ops = &exit_events;
295 	else if (!strcmp(kvm->report_event, "mmio"))
296 		kvm->events_ops = &mmio_events;
297 	else if (!strcmp(kvm->report_event, "ioport"))
298 		kvm->events_ops = &ioport_events;
299 	else {
300 		pr_err("Unknown report event:%s\n", kvm->report_event);
301 		ret = false;
302 	}
303 
304 	return ret;
305 }
306 
307 struct vcpu_event_record {
308 	int vcpu_id;
309 	u64 start_time;
310 	struct kvm_event *last_event;
311 };
312 
313 
314 static void init_kvm_event_record(struct perf_kvm *kvm)
315 {
316 	int i;
317 
318 	for (i = 0; i < (int)EVENTS_CACHE_SIZE; i++)
319 		INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
320 }
321 
322 static int kvm_events_hash_fn(u64 key)
323 {
324 	return key & (EVENTS_CACHE_SIZE - 1);
325 }
326 
327 static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
328 {
329 	int old_max_vcpu = event->max_vcpu;
330 
331 	if (vcpu_id < event->max_vcpu)
332 		return true;
333 
334 	while (event->max_vcpu <= vcpu_id)
335 		event->max_vcpu += DEFAULT_VCPU_NUM;
336 
337 	event->vcpu = realloc(event->vcpu,
338 			      event->max_vcpu * sizeof(*event->vcpu));
339 	if (!event->vcpu) {
340 		pr_err("Not enough memory\n");
341 		return false;
342 	}
343 
344 	memset(event->vcpu + old_max_vcpu, 0,
345 	       (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
346 	return true;
347 }
348 
349 static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
350 {
351 	struct kvm_event *event;
352 
353 	event = zalloc(sizeof(*event));
354 	if (!event) {
355 		pr_err("Not enough memory\n");
356 		return NULL;
357 	}
358 
359 	event->key = *key;
360 	return event;
361 }
362 
363 static struct kvm_event *find_create_kvm_event(struct perf_kvm *kvm,
364 					       struct event_key *key)
365 {
366 	struct kvm_event *event;
367 	struct list_head *head;
368 
369 	BUG_ON(key->key == INVALID_KEY);
370 
371 	head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
372 	list_for_each_entry(event, head, hash_entry)
373 		if (event->key.key == key->key && event->key.info == key->info)
374 			return event;
375 
376 	event = kvm_alloc_init_event(key);
377 	if (!event)
378 		return NULL;
379 
380 	list_add(&event->hash_entry, head);
381 	return event;
382 }
383 
384 static bool handle_begin_event(struct perf_kvm *kvm,
385 			       struct vcpu_event_record *vcpu_record,
386 			       struct event_key *key, u64 timestamp)
387 {
388 	struct kvm_event *event = NULL;
389 
390 	if (key->key != INVALID_KEY)
391 		event = find_create_kvm_event(kvm, key);
392 
393 	vcpu_record->last_event = event;
394 	vcpu_record->start_time = timestamp;
395 	return true;
396 }
397 
398 static void
399 kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
400 {
401 	kvm_stats->time += time_diff;
402 	update_stats(&kvm_stats->stats, time_diff);
403 }
404 
405 static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
406 {
407 	struct kvm_event_stats *kvm_stats = &event->total;
408 
409 	if (vcpu_id != -1)
410 		kvm_stats = &event->vcpu[vcpu_id];
411 
412 	return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
413 				avg_stats(&kvm_stats->stats));
414 }
415 
416 static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
417 			     u64 time_diff)
418 {
419 	kvm_update_event_stats(&event->total, time_diff);
420 
421 	if (!kvm_event_expand(event, vcpu_id))
422 		return false;
423 
424 	kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
425 	return true;
426 }
427 
428 static bool handle_end_event(struct perf_kvm *kvm,
429 			     struct vcpu_event_record *vcpu_record,
430 			     struct event_key *key,
431 			     u64 timestamp)
432 {
433 	struct kvm_event *event;
434 	u64 time_begin, time_diff;
435 
436 	event = vcpu_record->last_event;
437 	time_begin = vcpu_record->start_time;
438 
439 	/* The begin event is not caught. */
440 	if (!time_begin)
441 		return true;
442 
443 	/*
444 	 * In some case, the 'begin event' only records the start timestamp,
445 	 * the actual event is recognized in the 'end event' (e.g. mmio-event).
446 	 */
447 
448 	/* Both begin and end events did not get the key. */
449 	if (!event && key->key == INVALID_KEY)
450 		return true;
451 
452 	if (!event)
453 		event = find_create_kvm_event(kvm, key);
454 
455 	if (!event)
456 		return false;
457 
458 	vcpu_record->last_event = NULL;
459 	vcpu_record->start_time = 0;
460 
461 	BUG_ON(timestamp < time_begin);
462 
463 	time_diff = timestamp - time_begin;
464 	return update_kvm_event(event, vcpu_record->vcpu_id, time_diff);
465 }
466 
467 static
468 struct vcpu_event_record *per_vcpu_record(struct thread *thread,
469 					  struct perf_evsel *evsel,
470 					  struct perf_sample *sample)
471 {
472 	/* Only kvm_entry records vcpu id. */
473 	if (!thread->priv && kvm_entry_event(evsel)) {
474 		struct vcpu_event_record *vcpu_record;
475 
476 		vcpu_record = zalloc(sizeof(*vcpu_record));
477 		if (!vcpu_record) {
478 			pr_err("%s: Not enough memory\n", __func__);
479 			return NULL;
480 		}
481 
482 		vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample, "vcpu_id");
483 		thread->priv = vcpu_record;
484 	}
485 
486 	return thread->priv;
487 }
488 
489 static bool handle_kvm_event(struct perf_kvm *kvm,
490 			     struct thread *thread,
491 			     struct perf_evsel *evsel,
492 			     struct perf_sample *sample)
493 {
494 	struct vcpu_event_record *vcpu_record;
495 	struct event_key key = {.key = INVALID_KEY};
496 
497 	vcpu_record = per_vcpu_record(thread, evsel, sample);
498 	if (!vcpu_record)
499 		return true;
500 
501 	if (kvm->events_ops->is_begin_event(evsel, sample, &key))
502 		return handle_begin_event(kvm, vcpu_record, &key, sample->time);
503 
504 	if (kvm->events_ops->is_end_event(evsel, sample, &key))
505 		return handle_end_event(kvm, vcpu_record, &key, sample->time);
506 
507 	return true;
508 }
509 
510 #define GET_EVENT_KEY(func, field)					\
511 static u64 get_event_ ##func(struct kvm_event *event, int vcpu)		\
512 {									\
513 	if (vcpu == -1)							\
514 		return event->total.field;				\
515 									\
516 	if (vcpu >= event->max_vcpu)					\
517 		return 0;						\
518 									\
519 	return event->vcpu[vcpu].field;					\
520 }
521 
522 #define COMPARE_EVENT_KEY(func, field)					\
523 GET_EVENT_KEY(func, field)						\
524 static int compare_kvm_event_ ## func(struct kvm_event *one,		\
525 					struct kvm_event *two, int vcpu)\
526 {									\
527 	return get_event_ ##func(one, vcpu) >				\
528 				get_event_ ##func(two, vcpu);		\
529 }
530 
531 GET_EVENT_KEY(time, time);
532 COMPARE_EVENT_KEY(count, stats.n);
533 COMPARE_EVENT_KEY(mean, stats.mean);
534 
535 #define DEF_SORT_NAME_KEY(name, compare_key)				\
536 	{ #name, compare_kvm_event_ ## compare_key }
537 
538 static struct kvm_event_key keys[] = {
539 	DEF_SORT_NAME_KEY(sample, count),
540 	DEF_SORT_NAME_KEY(time, mean),
541 	{ NULL, NULL }
542 };
543 
544 static bool select_key(struct perf_kvm *kvm)
545 {
546 	int i;
547 
548 	for (i = 0; keys[i].name; i++) {
549 		if (!strcmp(keys[i].name, kvm->sort_key)) {
550 			kvm->compare = keys[i].key;
551 			return true;
552 		}
553 	}
554 
555 	pr_err("Unknown compare key:%s\n", kvm->sort_key);
556 	return false;
557 }
558 
559 static void insert_to_result(struct rb_root *result, struct kvm_event *event,
560 			     key_cmp_fun bigger, int vcpu)
561 {
562 	struct rb_node **rb = &result->rb_node;
563 	struct rb_node *parent = NULL;
564 	struct kvm_event *p;
565 
566 	while (*rb) {
567 		p = container_of(*rb, struct kvm_event, rb);
568 		parent = *rb;
569 
570 		if (bigger(event, p, vcpu))
571 			rb = &(*rb)->rb_left;
572 		else
573 			rb = &(*rb)->rb_right;
574 	}
575 
576 	rb_link_node(&event->rb, parent, rb);
577 	rb_insert_color(&event->rb, result);
578 }
579 
580 static void update_total_count(struct perf_kvm *kvm, struct kvm_event *event)
581 {
582 	int vcpu = kvm->trace_vcpu;
583 
584 	kvm->total_count += get_event_count(event, vcpu);
585 	kvm->total_time += get_event_time(event, vcpu);
586 }
587 
588 static bool event_is_valid(struct kvm_event *event, int vcpu)
589 {
590 	return !!get_event_count(event, vcpu);
591 }
592 
593 static void sort_result(struct perf_kvm *kvm)
594 {
595 	unsigned int i;
596 	int vcpu = kvm->trace_vcpu;
597 	struct kvm_event *event;
598 
599 	for (i = 0; i < EVENTS_CACHE_SIZE; i++)
600 		list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry)
601 			if (event_is_valid(event, vcpu)) {
602 				update_total_count(kvm, event);
603 				insert_to_result(&kvm->result, event,
604 						 kvm->compare, vcpu);
605 			}
606 }
607 
608 /* returns left most element of result, and erase it */
609 static struct kvm_event *pop_from_result(struct rb_root *result)
610 {
611 	struct rb_node *node = rb_first(result);
612 
613 	if (!node)
614 		return NULL;
615 
616 	rb_erase(node, result);
617 	return container_of(node, struct kvm_event, rb);
618 }
619 
620 static void print_vcpu_info(int vcpu)
621 {
622 	pr_info("Analyze events for ");
623 
624 	if (vcpu == -1)
625 		pr_info("all VCPUs:\n\n");
626 	else
627 		pr_info("VCPU %d:\n\n", vcpu);
628 }
629 
630 static void print_result(struct perf_kvm *kvm)
631 {
632 	char decode[20];
633 	struct kvm_event *event;
634 	int vcpu = kvm->trace_vcpu;
635 
636 	pr_info("\n\n");
637 	print_vcpu_info(vcpu);
638 	pr_info("%20s ", kvm->events_ops->name);
639 	pr_info("%10s ", "Samples");
640 	pr_info("%9s ", "Samples%");
641 
642 	pr_info("%9s ", "Time%");
643 	pr_info("%16s ", "Avg time");
644 	pr_info("\n\n");
645 
646 	while ((event = pop_from_result(&kvm->result))) {
647 		u64 ecount, etime;
648 
649 		ecount = get_event_count(event, vcpu);
650 		etime = get_event_time(event, vcpu);
651 
652 		kvm->events_ops->decode_key(kvm, &event->key, decode);
653 		pr_info("%20s ", decode);
654 		pr_info("%10llu ", (unsigned long long)ecount);
655 		pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
656 		pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
657 		pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount/1e3,
658 			kvm_event_rel_stddev(vcpu, event));
659 		pr_info("\n");
660 	}
661 
662 	pr_info("\nTotal Samples:%lld, Total events handled time:%.2fus.\n\n",
663 		(unsigned long long)kvm->total_count, kvm->total_time / 1e3);
664 }
665 
666 static int process_sample_event(struct perf_tool *tool,
667 				union perf_event *event,
668 				struct perf_sample *sample,
669 				struct perf_evsel *evsel,
670 				struct machine *machine)
671 {
672 	struct thread *thread = machine__findnew_thread(machine, sample->tid);
673 	struct perf_kvm *kvm = container_of(tool, struct perf_kvm, tool);
674 
675 	if (thread == NULL) {
676 		pr_debug("problem processing %d event, skipping it.\n",
677 			event->header.type);
678 		return -1;
679 	}
680 
681 	if (!handle_kvm_event(kvm, thread, evsel, sample))
682 		return -1;
683 
684 	return 0;
685 }
686 
687 static int get_cpu_isa(struct perf_session *session)
688 {
689 	char *cpuid = session->header.env.cpuid;
690 	int isa;
691 
692 	if (strstr(cpuid, "Intel"))
693 		isa = 1;
694 	else if (strstr(cpuid, "AMD"))
695 		isa = 0;
696 	else {
697 		pr_err("CPU %s is not supported.\n", cpuid);
698 		isa = -ENOTSUP;
699 	}
700 
701 	return isa;
702 }
703 
704 static int read_events(struct perf_kvm *kvm)
705 {
706 	int ret;
707 
708 	struct perf_tool eops = {
709 		.sample			= process_sample_event,
710 		.comm			= perf_event__process_comm,
711 		.ordered_samples	= true,
712 	};
713 
714 	kvm->tool = eops;
715 	kvm->session = perf_session__new(kvm->file_name, O_RDONLY, 0, false,
716 					 &kvm->tool);
717 	if (!kvm->session) {
718 		pr_err("Initializing perf session failed\n");
719 		return -EINVAL;
720 	}
721 
722 	if (!perf_session__has_traces(kvm->session, "kvm record"))
723 		return -EINVAL;
724 
725 	/*
726 	 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
727 	 * traced in the old kernel.
728 	 */
729 	ret = get_cpu_isa(kvm->session);
730 
731 	if (ret < 0)
732 		return ret;
733 
734 	if (ret == 1) {
735 		kvm->exit_reasons = vmx_exit_reasons;
736 		kvm->exit_reasons_size = ARRAY_SIZE(vmx_exit_reasons);
737 		kvm->exit_reasons_isa = "VMX";
738 	}
739 
740 	return perf_session__process_events(kvm->session, &kvm->tool);
741 }
742 
743 static bool verify_vcpu(int vcpu)
744 {
745 	if (vcpu != -1 && vcpu < 0) {
746 		pr_err("Invalid vcpu:%d.\n", vcpu);
747 		return false;
748 	}
749 
750 	return true;
751 }
752 
753 static int kvm_events_report_vcpu(struct perf_kvm *kvm)
754 {
755 	int ret = -EINVAL;
756 	int vcpu = kvm->trace_vcpu;
757 
758 	if (!verify_vcpu(vcpu))
759 		goto exit;
760 
761 	if (!select_key(kvm))
762 		goto exit;
763 
764 	if (!register_kvm_events_ops(kvm))
765 		goto exit;
766 
767 	init_kvm_event_record(kvm);
768 	setup_pager();
769 
770 	ret = read_events(kvm);
771 	if (ret)
772 		goto exit;
773 
774 	sort_result(kvm);
775 	print_result(kvm);
776 
777 exit:
778 	return ret;
779 }
780 
781 static const char * const record_args[] = {
782 	"record",
783 	"-R",
784 	"-f",
785 	"-m", "1024",
786 	"-c", "1",
787 	"-e", "kvm:kvm_entry",
788 	"-e", "kvm:kvm_exit",
789 	"-e", "kvm:kvm_mmio",
790 	"-e", "kvm:kvm_pio",
791 };
792 
793 #define STRDUP_FAIL_EXIT(s)		\
794 	({	char *_p;		\
795 	_p = strdup(s);		\
796 		if (!_p)		\
797 			return -ENOMEM;	\
798 		_p;			\
799 	})
800 
801 static int kvm_events_record(struct perf_kvm *kvm, int argc, const char **argv)
802 {
803 	unsigned int rec_argc, i, j;
804 	const char **rec_argv;
805 
806 	rec_argc = ARRAY_SIZE(record_args) + argc + 2;
807 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
808 
809 	if (rec_argv == NULL)
810 		return -ENOMEM;
811 
812 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
813 		rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
814 
815 	rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
816 	rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
817 
818 	for (j = 1; j < (unsigned int)argc; j++, i++)
819 		rec_argv[i] = argv[j];
820 
821 	return cmd_record(i, rec_argv, NULL);
822 }
823 
824 static int kvm_events_report(struct perf_kvm *kvm, int argc, const char **argv)
825 {
826 	const struct option kvm_events_report_options[] = {
827 		OPT_STRING(0, "event", &kvm->report_event, "report event",
828 			    "event for reporting: vmexit, mmio, ioport"),
829 		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
830 			    "vcpu id to report"),
831 		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
832 			    "key for sorting: sample(sort by samples number)"
833 			    " time (sort by avg time)"),
834 		OPT_END()
835 	};
836 
837 	const char * const kvm_events_report_usage[] = {
838 		"perf kvm stat report [<options>]",
839 		NULL
840 	};
841 
842 	symbol__init();
843 
844 	if (argc) {
845 		argc = parse_options(argc, argv,
846 				     kvm_events_report_options,
847 				     kvm_events_report_usage, 0);
848 		if (argc)
849 			usage_with_options(kvm_events_report_usage,
850 					   kvm_events_report_options);
851 	}
852 
853 	return kvm_events_report_vcpu(kvm);
854 }
855 
856 static void print_kvm_stat_usage(void)
857 {
858 	printf("Usage: perf kvm stat <command>\n\n");
859 
860 	printf("# Available commands:\n");
861 	printf("\trecord: record kvm events\n");
862 	printf("\treport: report statistical data of kvm events\n");
863 
864 	printf("\nOtherwise, it is the alias of 'perf stat':\n");
865 }
866 
867 static int kvm_cmd_stat(struct perf_kvm *kvm, int argc, const char **argv)
868 {
869 	if (argc == 1) {
870 		print_kvm_stat_usage();
871 		goto perf_stat;
872 	}
873 
874 	if (!strncmp(argv[1], "rec", 3))
875 		return kvm_events_record(kvm, argc - 1, argv + 1);
876 
877 	if (!strncmp(argv[1], "rep", 3))
878 		return kvm_events_report(kvm, argc - 1 , argv + 1);
879 
880 perf_stat:
881 	return cmd_stat(argc, argv, NULL);
882 }
883 
884 static int __cmd_record(struct perf_kvm *kvm, int argc, const char **argv)
885 {
886 	int rec_argc, i = 0, j;
887 	const char **rec_argv;
888 
889 	rec_argc = argc + 2;
890 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
891 	rec_argv[i++] = strdup("record");
892 	rec_argv[i++] = strdup("-o");
893 	rec_argv[i++] = strdup(kvm->file_name);
894 	for (j = 1; j < argc; j++, i++)
895 		rec_argv[i] = argv[j];
896 
897 	BUG_ON(i != rec_argc);
898 
899 	return cmd_record(i, rec_argv, NULL);
900 }
901 
902 static int __cmd_report(struct perf_kvm *kvm, int argc, const char **argv)
903 {
904 	int rec_argc, i = 0, j;
905 	const char **rec_argv;
906 
907 	rec_argc = argc + 2;
908 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
909 	rec_argv[i++] = strdup("report");
910 	rec_argv[i++] = strdup("-i");
911 	rec_argv[i++] = strdup(kvm->file_name);
912 	for (j = 1; j < argc; j++, i++)
913 		rec_argv[i] = argv[j];
914 
915 	BUG_ON(i != rec_argc);
916 
917 	return cmd_report(i, rec_argv, NULL);
918 }
919 
920 static int __cmd_buildid_list(struct perf_kvm *kvm, int argc, const char **argv)
921 {
922 	int rec_argc, i = 0, j;
923 	const char **rec_argv;
924 
925 	rec_argc = argc + 2;
926 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
927 	rec_argv[i++] = strdup("buildid-list");
928 	rec_argv[i++] = strdup("-i");
929 	rec_argv[i++] = strdup(kvm->file_name);
930 	for (j = 1; j < argc; j++, i++)
931 		rec_argv[i] = argv[j];
932 
933 	BUG_ON(i != rec_argc);
934 
935 	return cmd_buildid_list(i, rec_argv, NULL);
936 }
937 
938 int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
939 {
940 	struct perf_kvm kvm = {
941 		.trace_vcpu	= -1,
942 		.report_event	= "vmexit",
943 		.sort_key	= "sample",
944 
945 		.exit_reasons = svm_exit_reasons,
946 		.exit_reasons_size = ARRAY_SIZE(svm_exit_reasons),
947 		.exit_reasons_isa = "SVM",
948 	};
949 
950 	const struct option kvm_options[] = {
951 		OPT_STRING('i', "input", &kvm.file_name, "file",
952 			   "Input file name"),
953 		OPT_STRING('o', "output", &kvm.file_name, "file",
954 			   "Output file name"),
955 		OPT_BOOLEAN(0, "guest", &perf_guest,
956 			    "Collect guest os data"),
957 		OPT_BOOLEAN(0, "host", &perf_host,
958 			    "Collect host os data"),
959 		OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
960 			   "guest mount directory under which every guest os"
961 			   " instance has a subdir"),
962 		OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
963 			   "file", "file saving guest os vmlinux"),
964 		OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
965 			   "file", "file saving guest os /proc/kallsyms"),
966 		OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
967 			   "file", "file saving guest os /proc/modules"),
968 		OPT_END()
969 	};
970 
971 
972 	const char * const kvm_usage[] = {
973 		"perf kvm [<options>] {top|record|report|diff|buildid-list|stat}",
974 		NULL
975 	};
976 
977 	perf_host  = 0;
978 	perf_guest = 1;
979 
980 	argc = parse_options(argc, argv, kvm_options, kvm_usage,
981 			PARSE_OPT_STOP_AT_NON_OPTION);
982 	if (!argc)
983 		usage_with_options(kvm_usage, kvm_options);
984 
985 	if (!perf_host)
986 		perf_guest = 1;
987 
988 	if (!kvm.file_name) {
989 		if (perf_host && !perf_guest)
990 			kvm.file_name = strdup("perf.data.host");
991 		else if (!perf_host && perf_guest)
992 			kvm.file_name = strdup("perf.data.guest");
993 		else
994 			kvm.file_name = strdup("perf.data.kvm");
995 
996 		if (!kvm.file_name) {
997 			pr_err("Failed to allocate memory for filename\n");
998 			return -ENOMEM;
999 		}
1000 	}
1001 
1002 	if (!strncmp(argv[0], "rec", 3))
1003 		return __cmd_record(&kvm, argc, argv);
1004 	else if (!strncmp(argv[0], "rep", 3))
1005 		return __cmd_report(&kvm, argc, argv);
1006 	else if (!strncmp(argv[0], "diff", 4))
1007 		return cmd_diff(argc, argv, NULL);
1008 	else if (!strncmp(argv[0], "top", 3))
1009 		return cmd_top(argc, argv, NULL);
1010 	else if (!strncmp(argv[0], "buildid-list", 12))
1011 		return __cmd_buildid_list(&kvm, argc, argv);
1012 	else if (!strncmp(argv[0], "stat", 4))
1013 		return kvm_cmd_stat(&kvm, argc, argv);
1014 	else
1015 		usage_with_options(kvm_usage, kvm_options);
1016 
1017 	return 0;
1018 }
1019