xref: /openbmc/linux/tools/perf/builtin-kvm.c (revision 3932b9ca)
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 <api/fs/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 #ifdef HAVE_KVM_STAT_SUPPORT
33 #include <asm/kvm_perf.h>
34 #include "util/kvm-stat.h"
35 
36 void exit_event_get_key(struct perf_evsel *evsel,
37 			struct perf_sample *sample,
38 			struct event_key *key)
39 {
40 	key->info = 0;
41 	key->key = perf_evsel__intval(evsel, sample, KVM_EXIT_REASON);
42 }
43 
44 bool kvm_exit_event(struct perf_evsel *evsel)
45 {
46 	return !strcmp(evsel->name, KVM_EXIT_TRACE);
47 }
48 
49 bool exit_event_begin(struct perf_evsel *evsel,
50 		      struct perf_sample *sample, struct event_key *key)
51 {
52 	if (kvm_exit_event(evsel)) {
53 		exit_event_get_key(evsel, sample, key);
54 		return true;
55 	}
56 
57 	return false;
58 }
59 
60 bool kvm_entry_event(struct perf_evsel *evsel)
61 {
62 	return !strcmp(evsel->name, KVM_ENTRY_TRACE);
63 }
64 
65 bool exit_event_end(struct perf_evsel *evsel,
66 		    struct perf_sample *sample __maybe_unused,
67 		    struct event_key *key __maybe_unused)
68 {
69 	return kvm_entry_event(evsel);
70 }
71 
72 static const char *get_exit_reason(struct perf_kvm_stat *kvm,
73 				   struct exit_reasons_table *tbl,
74 				   u64 exit_code)
75 {
76 	while (tbl->reason != NULL) {
77 		if (tbl->exit_code == exit_code)
78 			return tbl->reason;
79 		tbl++;
80 	}
81 
82 	pr_err("unknown kvm exit code:%lld on %s\n",
83 		(unsigned long long)exit_code, kvm->exit_reasons_isa);
84 	return "UNKNOWN";
85 }
86 
87 void exit_event_decode_key(struct perf_kvm_stat *kvm,
88 			   struct event_key *key,
89 			   char *decode)
90 {
91 	const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
92 						  key->key);
93 
94 	scnprintf(decode, DECODE_STR_LEN, "%s", exit_reason);
95 }
96 
97 static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
98 {
99 	struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
100 
101 	for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
102 		if (!strcmp(events_ops->name, kvm->report_event)) {
103 			kvm->events_ops = events_ops->ops;
104 			return true;
105 		}
106 	}
107 
108 	return false;
109 }
110 
111 struct vcpu_event_record {
112 	int vcpu_id;
113 	u64 start_time;
114 	struct kvm_event *last_event;
115 };
116 
117 
118 static void init_kvm_event_record(struct perf_kvm_stat *kvm)
119 {
120 	unsigned int i;
121 
122 	for (i = 0; i < EVENTS_CACHE_SIZE; i++)
123 		INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
124 }
125 
126 #ifdef HAVE_TIMERFD_SUPPORT
127 static void clear_events_cache_stats(struct list_head *kvm_events_cache)
128 {
129 	struct list_head *head;
130 	struct kvm_event *event;
131 	unsigned int i;
132 	int j;
133 
134 	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
135 		head = &kvm_events_cache[i];
136 		list_for_each_entry(event, head, hash_entry) {
137 			/* reset stats for event */
138 			event->total.time = 0;
139 			init_stats(&event->total.stats);
140 
141 			for (j = 0; j < event->max_vcpu; ++j) {
142 				event->vcpu[j].time = 0;
143 				init_stats(&event->vcpu[j].stats);
144 			}
145 		}
146 	}
147 }
148 #endif
149 
150 static int kvm_events_hash_fn(u64 key)
151 {
152 	return key & (EVENTS_CACHE_SIZE - 1);
153 }
154 
155 static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
156 {
157 	int old_max_vcpu = event->max_vcpu;
158 	void *prev;
159 
160 	if (vcpu_id < event->max_vcpu)
161 		return true;
162 
163 	while (event->max_vcpu <= vcpu_id)
164 		event->max_vcpu += DEFAULT_VCPU_NUM;
165 
166 	prev = event->vcpu;
167 	event->vcpu = realloc(event->vcpu,
168 			      event->max_vcpu * sizeof(*event->vcpu));
169 	if (!event->vcpu) {
170 		free(prev);
171 		pr_err("Not enough memory\n");
172 		return false;
173 	}
174 
175 	memset(event->vcpu + old_max_vcpu, 0,
176 	       (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
177 	return true;
178 }
179 
180 static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
181 {
182 	struct kvm_event *event;
183 
184 	event = zalloc(sizeof(*event));
185 	if (!event) {
186 		pr_err("Not enough memory\n");
187 		return NULL;
188 	}
189 
190 	event->key = *key;
191 	init_stats(&event->total.stats);
192 	return event;
193 }
194 
195 static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
196 					       struct event_key *key)
197 {
198 	struct kvm_event *event;
199 	struct list_head *head;
200 
201 	BUG_ON(key->key == INVALID_KEY);
202 
203 	head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
204 	list_for_each_entry(event, head, hash_entry) {
205 		if (event->key.key == key->key && event->key.info == key->info)
206 			return event;
207 	}
208 
209 	event = kvm_alloc_init_event(key);
210 	if (!event)
211 		return NULL;
212 
213 	list_add(&event->hash_entry, head);
214 	return event;
215 }
216 
217 static bool handle_begin_event(struct perf_kvm_stat *kvm,
218 			       struct vcpu_event_record *vcpu_record,
219 			       struct event_key *key, u64 timestamp)
220 {
221 	struct kvm_event *event = NULL;
222 
223 	if (key->key != INVALID_KEY)
224 		event = find_create_kvm_event(kvm, key);
225 
226 	vcpu_record->last_event = event;
227 	vcpu_record->start_time = timestamp;
228 	return true;
229 }
230 
231 static void
232 kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
233 {
234 	kvm_stats->time += time_diff;
235 	update_stats(&kvm_stats->stats, time_diff);
236 }
237 
238 static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
239 {
240 	struct kvm_event_stats *kvm_stats = &event->total;
241 
242 	if (vcpu_id != -1)
243 		kvm_stats = &event->vcpu[vcpu_id];
244 
245 	return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
246 				avg_stats(&kvm_stats->stats));
247 }
248 
249 static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
250 			     u64 time_diff)
251 {
252 	if (vcpu_id == -1) {
253 		kvm_update_event_stats(&event->total, time_diff);
254 		return true;
255 	}
256 
257 	if (!kvm_event_expand(event, vcpu_id))
258 		return false;
259 
260 	kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
261 	return true;
262 }
263 
264 static bool is_child_event(struct perf_kvm_stat *kvm,
265 			   struct perf_evsel *evsel,
266 			   struct perf_sample *sample,
267 			   struct event_key *key)
268 {
269 	struct child_event_ops *child_ops;
270 
271 	child_ops = kvm->events_ops->child_ops;
272 
273 	if (!child_ops)
274 		return false;
275 
276 	for (; child_ops->name; child_ops++) {
277 		if (!strcmp(evsel->name, child_ops->name)) {
278 			child_ops->get_key(evsel, sample, key);
279 			return true;
280 		}
281 	}
282 
283 	return false;
284 }
285 
286 static bool handle_child_event(struct perf_kvm_stat *kvm,
287 			       struct vcpu_event_record *vcpu_record,
288 			       struct event_key *key,
289 			       struct perf_sample *sample __maybe_unused)
290 {
291 	struct kvm_event *event = NULL;
292 
293 	if (key->key != INVALID_KEY)
294 		event = find_create_kvm_event(kvm, key);
295 
296 	vcpu_record->last_event = event;
297 
298 	return true;
299 }
300 
301 static bool skip_event(const char *event)
302 {
303 	const char * const *skip_events;
304 
305 	for (skip_events = kvm_skip_events; *skip_events; skip_events++)
306 		if (!strcmp(event, *skip_events))
307 			return true;
308 
309 	return false;
310 }
311 
312 static bool handle_end_event(struct perf_kvm_stat *kvm,
313 			     struct vcpu_event_record *vcpu_record,
314 			     struct event_key *key,
315 			     struct perf_sample *sample)
316 {
317 	struct kvm_event *event;
318 	u64 time_begin, time_diff;
319 	int vcpu;
320 
321 	if (kvm->trace_vcpu == -1)
322 		vcpu = -1;
323 	else
324 		vcpu = vcpu_record->vcpu_id;
325 
326 	event = vcpu_record->last_event;
327 	time_begin = vcpu_record->start_time;
328 
329 	/* The begin event is not caught. */
330 	if (!time_begin)
331 		return true;
332 
333 	/*
334 	 * In some case, the 'begin event' only records the start timestamp,
335 	 * the actual event is recognized in the 'end event' (e.g. mmio-event).
336 	 */
337 
338 	/* Both begin and end events did not get the key. */
339 	if (!event && key->key == INVALID_KEY)
340 		return true;
341 
342 	if (!event)
343 		event = find_create_kvm_event(kvm, key);
344 
345 	if (!event)
346 		return false;
347 
348 	vcpu_record->last_event = NULL;
349 	vcpu_record->start_time = 0;
350 
351 	/* seems to happen once in a while during live mode */
352 	if (sample->time < time_begin) {
353 		pr_debug("End time before begin time; skipping event.\n");
354 		return true;
355 	}
356 
357 	time_diff = sample->time - time_begin;
358 
359 	if (kvm->duration && time_diff > kvm->duration) {
360 		char decode[DECODE_STR_LEN];
361 
362 		kvm->events_ops->decode_key(kvm, &event->key, decode);
363 		if (!skip_event(decode)) {
364 			pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
365 				 sample->time, sample->pid, vcpu_record->vcpu_id,
366 				 decode, time_diff/1000);
367 		}
368 	}
369 
370 	return update_kvm_event(event, vcpu, time_diff);
371 }
372 
373 static
374 struct vcpu_event_record *per_vcpu_record(struct thread *thread,
375 					  struct perf_evsel *evsel,
376 					  struct perf_sample *sample)
377 {
378 	/* Only kvm_entry records vcpu id. */
379 	if (!thread->priv && kvm_entry_event(evsel)) {
380 		struct vcpu_event_record *vcpu_record;
381 
382 		vcpu_record = zalloc(sizeof(*vcpu_record));
383 		if (!vcpu_record) {
384 			pr_err("%s: Not enough memory\n", __func__);
385 			return NULL;
386 		}
387 
388 		vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample, VCPU_ID);
389 		thread->priv = vcpu_record;
390 	}
391 
392 	return thread->priv;
393 }
394 
395 static bool handle_kvm_event(struct perf_kvm_stat *kvm,
396 			     struct thread *thread,
397 			     struct perf_evsel *evsel,
398 			     struct perf_sample *sample)
399 {
400 	struct vcpu_event_record *vcpu_record;
401 	struct event_key key = { .key = INVALID_KEY,
402 				 .exit_reasons = kvm->exit_reasons };
403 
404 	vcpu_record = per_vcpu_record(thread, evsel, sample);
405 	if (!vcpu_record)
406 		return true;
407 
408 	/* only process events for vcpus user cares about */
409 	if ((kvm->trace_vcpu != -1) &&
410 	    (kvm->trace_vcpu != vcpu_record->vcpu_id))
411 		return true;
412 
413 	if (kvm->events_ops->is_begin_event(evsel, sample, &key))
414 		return handle_begin_event(kvm, vcpu_record, &key, sample->time);
415 
416 	if (is_child_event(kvm, evsel, sample, &key))
417 		return handle_child_event(kvm, vcpu_record, &key, sample);
418 
419 	if (kvm->events_ops->is_end_event(evsel, sample, &key))
420 		return handle_end_event(kvm, vcpu_record, &key, sample);
421 
422 	return true;
423 }
424 
425 #define GET_EVENT_KEY(func, field)					\
426 static u64 get_event_ ##func(struct kvm_event *event, int vcpu)		\
427 {									\
428 	if (vcpu == -1)							\
429 		return event->total.field;				\
430 									\
431 	if (vcpu >= event->max_vcpu)					\
432 		return 0;						\
433 									\
434 	return event->vcpu[vcpu].field;					\
435 }
436 
437 #define COMPARE_EVENT_KEY(func, field)					\
438 GET_EVENT_KEY(func, field)						\
439 static int compare_kvm_event_ ## func(struct kvm_event *one,		\
440 					struct kvm_event *two, int vcpu)\
441 {									\
442 	return get_event_ ##func(one, vcpu) >				\
443 				get_event_ ##func(two, vcpu);		\
444 }
445 
446 GET_EVENT_KEY(time, time);
447 COMPARE_EVENT_KEY(count, stats.n);
448 COMPARE_EVENT_KEY(mean, stats.mean);
449 GET_EVENT_KEY(max, stats.max);
450 GET_EVENT_KEY(min, stats.min);
451 
452 #define DEF_SORT_NAME_KEY(name, compare_key)				\
453 	{ #name, compare_kvm_event_ ## compare_key }
454 
455 static struct kvm_event_key keys[] = {
456 	DEF_SORT_NAME_KEY(sample, count),
457 	DEF_SORT_NAME_KEY(time, mean),
458 	{ NULL, NULL }
459 };
460 
461 static bool select_key(struct perf_kvm_stat *kvm)
462 {
463 	int i;
464 
465 	for (i = 0; keys[i].name; i++) {
466 		if (!strcmp(keys[i].name, kvm->sort_key)) {
467 			kvm->compare = keys[i].key;
468 			return true;
469 		}
470 	}
471 
472 	pr_err("Unknown compare key:%s\n", kvm->sort_key);
473 	return false;
474 }
475 
476 static void insert_to_result(struct rb_root *result, struct kvm_event *event,
477 			     key_cmp_fun bigger, int vcpu)
478 {
479 	struct rb_node **rb = &result->rb_node;
480 	struct rb_node *parent = NULL;
481 	struct kvm_event *p;
482 
483 	while (*rb) {
484 		p = container_of(*rb, struct kvm_event, rb);
485 		parent = *rb;
486 
487 		if (bigger(event, p, vcpu))
488 			rb = &(*rb)->rb_left;
489 		else
490 			rb = &(*rb)->rb_right;
491 	}
492 
493 	rb_link_node(&event->rb, parent, rb);
494 	rb_insert_color(&event->rb, result);
495 }
496 
497 static void
498 update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
499 {
500 	int vcpu = kvm->trace_vcpu;
501 
502 	kvm->total_count += get_event_count(event, vcpu);
503 	kvm->total_time += get_event_time(event, vcpu);
504 }
505 
506 static bool event_is_valid(struct kvm_event *event, int vcpu)
507 {
508 	return !!get_event_count(event, vcpu);
509 }
510 
511 static void sort_result(struct perf_kvm_stat *kvm)
512 {
513 	unsigned int i;
514 	int vcpu = kvm->trace_vcpu;
515 	struct kvm_event *event;
516 
517 	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
518 		list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
519 			if (event_is_valid(event, vcpu)) {
520 				update_total_count(kvm, event);
521 				insert_to_result(&kvm->result, event,
522 						 kvm->compare, vcpu);
523 			}
524 		}
525 	}
526 }
527 
528 /* returns left most element of result, and erase it */
529 static struct kvm_event *pop_from_result(struct rb_root *result)
530 {
531 	struct rb_node *node = rb_first(result);
532 
533 	if (!node)
534 		return NULL;
535 
536 	rb_erase(node, result);
537 	return container_of(node, struct kvm_event, rb);
538 }
539 
540 static void print_vcpu_info(struct perf_kvm_stat *kvm)
541 {
542 	int vcpu = kvm->trace_vcpu;
543 
544 	pr_info("Analyze events for ");
545 
546 	if (kvm->live) {
547 		if (kvm->opts.target.system_wide)
548 			pr_info("all VMs, ");
549 		else if (kvm->opts.target.pid)
550 			pr_info("pid(s) %s, ", kvm->opts.target.pid);
551 		else
552 			pr_info("dazed and confused on what is monitored, ");
553 	}
554 
555 	if (vcpu == -1)
556 		pr_info("all VCPUs:\n\n");
557 	else
558 		pr_info("VCPU %d:\n\n", vcpu);
559 }
560 
561 static void show_timeofday(void)
562 {
563 	char date[64];
564 	struct timeval tv;
565 	struct tm ltime;
566 
567 	gettimeofday(&tv, NULL);
568 	if (localtime_r(&tv.tv_sec, &ltime)) {
569 		strftime(date, sizeof(date), "%H:%M:%S", &ltime);
570 		pr_info("%s.%06ld", date, tv.tv_usec);
571 	} else
572 		pr_info("00:00:00.000000");
573 
574 	return;
575 }
576 
577 static void print_result(struct perf_kvm_stat *kvm)
578 {
579 	char decode[DECODE_STR_LEN];
580 	struct kvm_event *event;
581 	int vcpu = kvm->trace_vcpu;
582 
583 	if (kvm->live) {
584 		puts(CONSOLE_CLEAR);
585 		show_timeofday();
586 	}
587 
588 	pr_info("\n\n");
589 	print_vcpu_info(kvm);
590 	pr_info("%*s ", DECODE_STR_LEN, kvm->events_ops->name);
591 	pr_info("%10s ", "Samples");
592 	pr_info("%9s ", "Samples%");
593 
594 	pr_info("%9s ", "Time%");
595 	pr_info("%10s ", "Min Time");
596 	pr_info("%10s ", "Max Time");
597 	pr_info("%16s ", "Avg time");
598 	pr_info("\n\n");
599 
600 	while ((event = pop_from_result(&kvm->result))) {
601 		u64 ecount, etime, max, min;
602 
603 		ecount = get_event_count(event, vcpu);
604 		etime = get_event_time(event, vcpu);
605 		max = get_event_max(event, vcpu);
606 		min = get_event_min(event, vcpu);
607 
608 		kvm->events_ops->decode_key(kvm, &event->key, decode);
609 		pr_info("%*s ", DECODE_STR_LEN, decode);
610 		pr_info("%10llu ", (unsigned long long)ecount);
611 		pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
612 		pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
613 		pr_info("%8" PRIu64 "us ", min / 1000);
614 		pr_info("%8" PRIu64 "us ", max / 1000);
615 		pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount/1e3,
616 			kvm_event_rel_stddev(vcpu, event));
617 		pr_info("\n");
618 	}
619 
620 	pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
621 		kvm->total_count, kvm->total_time / 1e3);
622 
623 	if (kvm->lost_events)
624 		pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
625 }
626 
627 #ifdef HAVE_TIMERFD_SUPPORT
628 static int process_lost_event(struct perf_tool *tool,
629 			      union perf_event *event __maybe_unused,
630 			      struct perf_sample *sample __maybe_unused,
631 			      struct machine *machine __maybe_unused)
632 {
633 	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
634 
635 	kvm->lost_events++;
636 	return 0;
637 }
638 #endif
639 
640 static bool skip_sample(struct perf_kvm_stat *kvm,
641 			struct perf_sample *sample)
642 {
643 	if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
644 		return true;
645 
646 	return false;
647 }
648 
649 static int process_sample_event(struct perf_tool *tool,
650 				union perf_event *event,
651 				struct perf_sample *sample,
652 				struct perf_evsel *evsel,
653 				struct machine *machine)
654 {
655 	struct thread *thread;
656 	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
657 						 tool);
658 
659 	if (skip_sample(kvm, sample))
660 		return 0;
661 
662 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
663 	if (thread == NULL) {
664 		pr_debug("problem processing %d event, skipping it.\n",
665 			event->header.type);
666 		return -1;
667 	}
668 
669 	if (!handle_kvm_event(kvm, thread, evsel, sample))
670 		return -1;
671 
672 	return 0;
673 }
674 
675 static int cpu_isa_config(struct perf_kvm_stat *kvm)
676 {
677 	char buf[64], *cpuid;
678 	int err;
679 
680 	if (kvm->live) {
681 		err = get_cpuid(buf, sizeof(buf));
682 		if (err != 0) {
683 			pr_err("Failed to look up CPU type\n");
684 			return err;
685 		}
686 		cpuid = buf;
687 	} else
688 		cpuid = kvm->session->header.env.cpuid;
689 
690 	if (!cpuid) {
691 		pr_err("Failed to look up CPU type\n");
692 		return -EINVAL;
693 	}
694 
695 	err = cpu_isa_init(kvm, cpuid);
696 	if (err == -ENOTSUP)
697 		pr_err("CPU %s is not supported.\n", cpuid);
698 
699 	return err;
700 }
701 
702 static bool verify_vcpu(int vcpu)
703 {
704 	if (vcpu != -1 && vcpu < 0) {
705 		pr_err("Invalid vcpu:%d.\n", vcpu);
706 		return false;
707 	}
708 
709 	return true;
710 }
711 
712 #ifdef HAVE_TIMERFD_SUPPORT
713 /* keeping the max events to a modest level to keep
714  * the processing of samples per mmap smooth.
715  */
716 #define PERF_KVM__MAX_EVENTS_PER_MMAP  25
717 
718 static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
719 				   u64 *mmap_time)
720 {
721 	union perf_event *event;
722 	struct perf_sample sample;
723 	s64 n = 0;
724 	int err;
725 
726 	*mmap_time = ULLONG_MAX;
727 	while ((event = perf_evlist__mmap_read(kvm->evlist, idx)) != NULL) {
728 		err = perf_evlist__parse_sample(kvm->evlist, event, &sample);
729 		if (err) {
730 			perf_evlist__mmap_consume(kvm->evlist, idx);
731 			pr_err("Failed to parse sample\n");
732 			return -1;
733 		}
734 
735 		err = perf_session_queue_event(kvm->session, event, &sample, 0);
736 		/*
737 		 * FIXME: Here we can't consume the event, as perf_session_queue_event will
738 		 *        point to it, and it'll get possibly overwritten by the kernel.
739 		 */
740 		perf_evlist__mmap_consume(kvm->evlist, idx);
741 
742 		if (err) {
743 			pr_err("Failed to enqueue sample: %d\n", err);
744 			return -1;
745 		}
746 
747 		/* save time stamp of our first sample for this mmap */
748 		if (n == 0)
749 			*mmap_time = sample.time;
750 
751 		/* limit events per mmap handled all at once */
752 		n++;
753 		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
754 			break;
755 	}
756 
757 	return n;
758 }
759 
760 static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
761 {
762 	int i, err, throttled = 0;
763 	s64 n, ntotal = 0;
764 	u64 flush_time = ULLONG_MAX, mmap_time;
765 
766 	for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
767 		n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
768 		if (n < 0)
769 			return -1;
770 
771 		/* flush time is going to be the minimum of all the individual
772 		 * mmap times. Essentially, we flush all the samples queued up
773 		 * from the last pass under our minimal start time -- that leaves
774 		 * a very small race for samples to come in with a lower timestamp.
775 		 * The ioctl to return the perf_clock timestamp should close the
776 		 * race entirely.
777 		 */
778 		if (mmap_time < flush_time)
779 			flush_time = mmap_time;
780 
781 		ntotal += n;
782 		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
783 			throttled = 1;
784 	}
785 
786 	/* flush queue after each round in which we processed events */
787 	if (ntotal) {
788 		kvm->session->ordered_samples.next_flush = flush_time;
789 		err = kvm->tool.finished_round(&kvm->tool, NULL, kvm->session);
790 		if (err) {
791 			if (kvm->lost_events)
792 				pr_info("\nLost events: %" PRIu64 "\n\n",
793 					kvm->lost_events);
794 			return err;
795 		}
796 	}
797 
798 	return throttled;
799 }
800 
801 static volatile int done;
802 
803 static void sig_handler(int sig __maybe_unused)
804 {
805 	done = 1;
806 }
807 
808 static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
809 {
810 	struct itimerspec new_value;
811 	int rc = -1;
812 
813 	kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
814 	if (kvm->timerfd < 0) {
815 		pr_err("timerfd_create failed\n");
816 		goto out;
817 	}
818 
819 	new_value.it_value.tv_sec = kvm->display_time;
820 	new_value.it_value.tv_nsec = 0;
821 	new_value.it_interval.tv_sec = kvm->display_time;
822 	new_value.it_interval.tv_nsec = 0;
823 
824 	if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
825 		pr_err("timerfd_settime failed: %d\n", errno);
826 		close(kvm->timerfd);
827 		goto out;
828 	}
829 
830 	rc = 0;
831 out:
832 	return rc;
833 }
834 
835 static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
836 {
837 	uint64_t c;
838 	int rc;
839 
840 	rc = read(kvm->timerfd, &c, sizeof(uint64_t));
841 	if (rc < 0) {
842 		if (errno == EAGAIN)
843 			return 0;
844 
845 		pr_err("Failed to read timer fd: %d\n", errno);
846 		return -1;
847 	}
848 
849 	if (rc != sizeof(uint64_t)) {
850 		pr_err("Error reading timer fd - invalid size returned\n");
851 		return -1;
852 	}
853 
854 	if (c != 1)
855 		pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
856 
857 	/* update display */
858 	sort_result(kvm);
859 	print_result(kvm);
860 
861 	/* reset counts */
862 	clear_events_cache_stats(kvm->kvm_events_cache);
863 	kvm->total_count = 0;
864 	kvm->total_time = 0;
865 	kvm->lost_events = 0;
866 
867 	return 0;
868 }
869 
870 static int fd_set_nonblock(int fd)
871 {
872 	long arg = 0;
873 
874 	arg = fcntl(fd, F_GETFL);
875 	if (arg < 0) {
876 		pr_err("Failed to get current flags for fd %d\n", fd);
877 		return -1;
878 	}
879 
880 	if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
881 		pr_err("Failed to set non-block option on fd %d\n", fd);
882 		return -1;
883 	}
884 
885 	return 0;
886 }
887 
888 static
889 int perf_kvm__handle_stdin(struct termios *tc_now, struct termios *tc_save)
890 {
891 	int c;
892 
893 	tcsetattr(0, TCSANOW, tc_now);
894 	c = getc(stdin);
895 	tcsetattr(0, TCSAFLUSH, tc_save);
896 
897 	if (c == 'q')
898 		return 1;
899 
900 	return 0;
901 }
902 
903 static int kvm_events_live_report(struct perf_kvm_stat *kvm)
904 {
905 	struct pollfd *pollfds = NULL;
906 	int nr_fds, nr_stdin, ret, err = -EINVAL;
907 	struct termios tc, save;
908 
909 	/* live flag must be set first */
910 	kvm->live = true;
911 
912 	ret = cpu_isa_config(kvm);
913 	if (ret < 0)
914 		return ret;
915 
916 	if (!verify_vcpu(kvm->trace_vcpu) ||
917 	    !select_key(kvm) ||
918 	    !register_kvm_events_ops(kvm)) {
919 		goto out;
920 	}
921 
922 	init_kvm_event_record(kvm);
923 
924 	tcgetattr(0, &save);
925 	tc = save;
926 	tc.c_lflag &= ~(ICANON | ECHO);
927 	tc.c_cc[VMIN] = 0;
928 	tc.c_cc[VTIME] = 0;
929 
930 	signal(SIGINT, sig_handler);
931 	signal(SIGTERM, sig_handler);
932 
933 	/* copy pollfds -- need to add timerfd and stdin */
934 	nr_fds = kvm->evlist->nr_fds;
935 	pollfds = zalloc(sizeof(struct pollfd) * (nr_fds + 2));
936 	if (!pollfds) {
937 		err = -ENOMEM;
938 		goto out;
939 	}
940 	memcpy(pollfds, kvm->evlist->pollfd,
941 		sizeof(struct pollfd) * kvm->evlist->nr_fds);
942 
943 	/* add timer fd */
944 	if (perf_kvm__timerfd_create(kvm) < 0) {
945 		err = -1;
946 		goto out;
947 	}
948 
949 	pollfds[nr_fds].fd = kvm->timerfd;
950 	pollfds[nr_fds].events = POLLIN;
951 	nr_fds++;
952 
953 	pollfds[nr_fds].fd = fileno(stdin);
954 	pollfds[nr_fds].events = POLLIN;
955 	nr_stdin = nr_fds;
956 	nr_fds++;
957 	if (fd_set_nonblock(fileno(stdin)) != 0)
958 		goto out;
959 
960 	/* everything is good - enable the events and process */
961 	perf_evlist__enable(kvm->evlist);
962 
963 	while (!done) {
964 		int rc;
965 
966 		rc = perf_kvm__mmap_read(kvm);
967 		if (rc < 0)
968 			break;
969 
970 		err = perf_kvm__handle_timerfd(kvm);
971 		if (err)
972 			goto out;
973 
974 		if (pollfds[nr_stdin].revents & POLLIN)
975 			done = perf_kvm__handle_stdin(&tc, &save);
976 
977 		if (!rc && !done)
978 			err = poll(pollfds, nr_fds, 100);
979 	}
980 
981 	perf_evlist__disable(kvm->evlist);
982 
983 	if (err == 0) {
984 		sort_result(kvm);
985 		print_result(kvm);
986 	}
987 
988 out:
989 	if (kvm->timerfd >= 0)
990 		close(kvm->timerfd);
991 
992 	free(pollfds);
993 	return err;
994 }
995 
996 static int kvm_live_open_events(struct perf_kvm_stat *kvm)
997 {
998 	int err, rc = -1;
999 	struct perf_evsel *pos;
1000 	struct perf_evlist *evlist = kvm->evlist;
1001 
1002 	perf_evlist__config(evlist, &kvm->opts);
1003 
1004 	/*
1005 	 * Note: exclude_{guest,host} do not apply here.
1006 	 *       This command processes KVM tracepoints from host only
1007 	 */
1008 	evlist__for_each(evlist, pos) {
1009 		struct perf_event_attr *attr = &pos->attr;
1010 
1011 		/* make sure these *are* set */
1012 		perf_evsel__set_sample_bit(pos, TID);
1013 		perf_evsel__set_sample_bit(pos, TIME);
1014 		perf_evsel__set_sample_bit(pos, CPU);
1015 		perf_evsel__set_sample_bit(pos, RAW);
1016 		/* make sure these are *not*; want as small a sample as possible */
1017 		perf_evsel__reset_sample_bit(pos, PERIOD);
1018 		perf_evsel__reset_sample_bit(pos, IP);
1019 		perf_evsel__reset_sample_bit(pos, CALLCHAIN);
1020 		perf_evsel__reset_sample_bit(pos, ADDR);
1021 		perf_evsel__reset_sample_bit(pos, READ);
1022 		attr->mmap = 0;
1023 		attr->comm = 0;
1024 		attr->task = 0;
1025 
1026 		attr->sample_period = 1;
1027 
1028 		attr->watermark = 0;
1029 		attr->wakeup_events = 1000;
1030 
1031 		/* will enable all once we are ready */
1032 		attr->disabled = 1;
1033 	}
1034 
1035 	err = perf_evlist__open(evlist);
1036 	if (err < 0) {
1037 		printf("Couldn't create the events: %s\n", strerror(errno));
1038 		goto out;
1039 	}
1040 
1041 	if (perf_evlist__mmap(evlist, kvm->opts.mmap_pages, false) < 0) {
1042 		ui__error("Failed to mmap the events: %s\n", strerror(errno));
1043 		perf_evlist__close(evlist);
1044 		goto out;
1045 	}
1046 
1047 	rc = 0;
1048 
1049 out:
1050 	return rc;
1051 }
1052 #endif
1053 
1054 static int read_events(struct perf_kvm_stat *kvm)
1055 {
1056 	int ret;
1057 
1058 	struct perf_tool eops = {
1059 		.sample			= process_sample_event,
1060 		.comm			= perf_event__process_comm,
1061 		.ordered_samples	= true,
1062 	};
1063 	struct perf_data_file file = {
1064 		.path = kvm->file_name,
1065 		.mode = PERF_DATA_MODE_READ,
1066 	};
1067 
1068 	kvm->tool = eops;
1069 	kvm->session = perf_session__new(&file, false, &kvm->tool);
1070 	if (!kvm->session) {
1071 		pr_err("Initializing perf session failed\n");
1072 		return -EINVAL;
1073 	}
1074 
1075 	if (!perf_session__has_traces(kvm->session, "kvm record"))
1076 		return -EINVAL;
1077 
1078 	/*
1079 	 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1080 	 * traced in the old kernel.
1081 	 */
1082 	ret = cpu_isa_config(kvm);
1083 	if (ret < 0)
1084 		return ret;
1085 
1086 	return perf_session__process_events(kvm->session, &kvm->tool);
1087 }
1088 
1089 static int parse_target_str(struct perf_kvm_stat *kvm)
1090 {
1091 	if (kvm->pid_str) {
1092 		kvm->pid_list = intlist__new(kvm->pid_str);
1093 		if (kvm->pid_list == NULL) {
1094 			pr_err("Error parsing process id string\n");
1095 			return -EINVAL;
1096 		}
1097 	}
1098 
1099 	return 0;
1100 }
1101 
1102 static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1103 {
1104 	int ret = -EINVAL;
1105 	int vcpu = kvm->trace_vcpu;
1106 
1107 	if (parse_target_str(kvm) != 0)
1108 		goto exit;
1109 
1110 	if (!verify_vcpu(vcpu))
1111 		goto exit;
1112 
1113 	if (!select_key(kvm))
1114 		goto exit;
1115 
1116 	if (!register_kvm_events_ops(kvm))
1117 		goto exit;
1118 
1119 	init_kvm_event_record(kvm);
1120 	setup_pager();
1121 
1122 	ret = read_events(kvm);
1123 	if (ret)
1124 		goto exit;
1125 
1126 	sort_result(kvm);
1127 	print_result(kvm);
1128 
1129 exit:
1130 	return ret;
1131 }
1132 
1133 #define STRDUP_FAIL_EXIT(s)		\
1134 	({	char *_p;		\
1135 	_p = strdup(s);		\
1136 		if (!_p)		\
1137 			return -ENOMEM;	\
1138 		_p;			\
1139 	})
1140 
1141 static int
1142 kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1143 {
1144 	unsigned int rec_argc, i, j, events_tp_size;
1145 	const char **rec_argv;
1146 	const char * const record_args[] = {
1147 		"record",
1148 		"-R",
1149 		"-m", "1024",
1150 		"-c", "1",
1151 	};
1152 	const char * const *events_tp;
1153 	events_tp_size = 0;
1154 
1155 	for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1156 		events_tp_size++;
1157 
1158 	rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1159 		   2 * events_tp_size;
1160 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1161 
1162 	if (rec_argv == NULL)
1163 		return -ENOMEM;
1164 
1165 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1166 		rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1167 
1168 	for (j = 0; j < events_tp_size; j++) {
1169 		rec_argv[i++] = "-e";
1170 		rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1171 	}
1172 
1173 	rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1174 	rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1175 
1176 	for (j = 1; j < (unsigned int)argc; j++, i++)
1177 		rec_argv[i] = argv[j];
1178 
1179 	return cmd_record(i, rec_argv, NULL);
1180 }
1181 
1182 static int
1183 kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1184 {
1185 	const struct option kvm_events_report_options[] = {
1186 		OPT_STRING(0, "event", &kvm->report_event, "report event",
1187 			   "event for reporting: vmexit, "
1188 			   "mmio (x86 only), ioport (x86 only)"),
1189 		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1190 			    "vcpu id to report"),
1191 		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1192 			    "key for sorting: sample(sort by samples number)"
1193 			    " time (sort by avg time)"),
1194 		OPT_STRING('p', "pid", &kvm->pid_str, "pid",
1195 			   "analyze events only for given process id(s)"),
1196 		OPT_END()
1197 	};
1198 
1199 	const char * const kvm_events_report_usage[] = {
1200 		"perf kvm stat report [<options>]",
1201 		NULL
1202 	};
1203 
1204 	symbol__init();
1205 
1206 	if (argc) {
1207 		argc = parse_options(argc, argv,
1208 				     kvm_events_report_options,
1209 				     kvm_events_report_usage, 0);
1210 		if (argc)
1211 			usage_with_options(kvm_events_report_usage,
1212 					   kvm_events_report_options);
1213 	}
1214 
1215 	return kvm_events_report_vcpu(kvm);
1216 }
1217 
1218 #ifdef HAVE_TIMERFD_SUPPORT
1219 static struct perf_evlist *kvm_live_event_list(void)
1220 {
1221 	struct perf_evlist *evlist;
1222 	char *tp, *name, *sys;
1223 	int err = -1;
1224 	const char * const *events_tp;
1225 
1226 	evlist = perf_evlist__new();
1227 	if (evlist == NULL)
1228 		return NULL;
1229 
1230 	for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1231 
1232 		tp = strdup(*events_tp);
1233 		if (tp == NULL)
1234 			goto out;
1235 
1236 		/* split tracepoint into subsystem and name */
1237 		sys = tp;
1238 		name = strchr(tp, ':');
1239 		if (name == NULL) {
1240 			pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1241 			       *events_tp);
1242 			free(tp);
1243 			goto out;
1244 		}
1245 		*name = '\0';
1246 		name++;
1247 
1248 		if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1249 			pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1250 			free(tp);
1251 			goto out;
1252 		}
1253 
1254 		free(tp);
1255 	}
1256 
1257 	err = 0;
1258 
1259 out:
1260 	if (err) {
1261 		perf_evlist__delete(evlist);
1262 		evlist = NULL;
1263 	}
1264 
1265 	return evlist;
1266 }
1267 
1268 static int kvm_events_live(struct perf_kvm_stat *kvm,
1269 			   int argc, const char **argv)
1270 {
1271 	char errbuf[BUFSIZ];
1272 	int err;
1273 
1274 	const struct option live_options[] = {
1275 		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1276 			"record events on existing process id"),
1277 		OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1278 			"number of mmap data pages",
1279 			perf_evlist__parse_mmap_pages),
1280 		OPT_INCR('v', "verbose", &verbose,
1281 			"be more verbose (show counter open errors, etc)"),
1282 		OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1283 			"system-wide collection from all CPUs"),
1284 		OPT_UINTEGER('d', "display", &kvm->display_time,
1285 			"time in seconds between display updates"),
1286 		OPT_STRING(0, "event", &kvm->report_event, "report event",
1287 			"event for reporting: vmexit, mmio, ioport"),
1288 		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1289 			"vcpu id to report"),
1290 		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1291 			"key for sorting: sample(sort by samples number)"
1292 			" time (sort by avg time)"),
1293 		OPT_U64(0, "duration", &kvm->duration,
1294 			"show events other than"
1295 			" HLT (x86 only) or Wait state (s390 only)"
1296 			" that take longer than duration usecs"),
1297 		OPT_END()
1298 	};
1299 	const char * const live_usage[] = {
1300 		"perf kvm stat live [<options>]",
1301 		NULL
1302 	};
1303 	struct perf_data_file file = {
1304 		.mode = PERF_DATA_MODE_WRITE,
1305 	};
1306 
1307 
1308 	/* event handling */
1309 	kvm->tool.sample = process_sample_event;
1310 	kvm->tool.comm   = perf_event__process_comm;
1311 	kvm->tool.exit   = perf_event__process_exit;
1312 	kvm->tool.fork   = perf_event__process_fork;
1313 	kvm->tool.lost   = process_lost_event;
1314 	kvm->tool.ordered_samples = true;
1315 	perf_tool__fill_defaults(&kvm->tool);
1316 
1317 	/* set defaults */
1318 	kvm->display_time = 1;
1319 	kvm->opts.user_interval = 1;
1320 	kvm->opts.mmap_pages = 512;
1321 	kvm->opts.target.uses_mmap = false;
1322 	kvm->opts.target.uid_str = NULL;
1323 	kvm->opts.target.uid = UINT_MAX;
1324 
1325 	symbol__init();
1326 	disable_buildid_cache();
1327 
1328 	use_browser = 0;
1329 	setup_browser(false);
1330 
1331 	if (argc) {
1332 		argc = parse_options(argc, argv, live_options,
1333 				     live_usage, 0);
1334 		if (argc)
1335 			usage_with_options(live_usage, live_options);
1336 	}
1337 
1338 	kvm->duration *= NSEC_PER_USEC;   /* convert usec to nsec */
1339 
1340 	/*
1341 	 * target related setups
1342 	 */
1343 	err = target__validate(&kvm->opts.target);
1344 	if (err) {
1345 		target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1346 		ui__warning("%s", errbuf);
1347 	}
1348 
1349 	if (target__none(&kvm->opts.target))
1350 		kvm->opts.target.system_wide = true;
1351 
1352 
1353 	/*
1354 	 * generate the event list
1355 	 */
1356 	kvm->evlist = kvm_live_event_list();
1357 	if (kvm->evlist == NULL) {
1358 		err = -1;
1359 		goto out;
1360 	}
1361 
1362 	symbol_conf.nr_events = kvm->evlist->nr_entries;
1363 
1364 	if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1365 		usage_with_options(live_usage, live_options);
1366 
1367 	/*
1368 	 * perf session
1369 	 */
1370 	kvm->session = perf_session__new(&file, false, &kvm->tool);
1371 	if (kvm->session == NULL) {
1372 		err = -ENOMEM;
1373 		goto out;
1374 	}
1375 	kvm->session->evlist = kvm->evlist;
1376 	perf_session__set_id_hdr_size(kvm->session);
1377 	machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1378 				    kvm->evlist->threads, false);
1379 	err = kvm_live_open_events(kvm);
1380 	if (err)
1381 		goto out;
1382 
1383 	err = kvm_events_live_report(kvm);
1384 
1385 out:
1386 	exit_browser(0);
1387 
1388 	if (kvm->session)
1389 		perf_session__delete(kvm->session);
1390 	kvm->session = NULL;
1391 	if (kvm->evlist)
1392 		perf_evlist__delete(kvm->evlist);
1393 
1394 	return err;
1395 }
1396 #endif
1397 
1398 static void print_kvm_stat_usage(void)
1399 {
1400 	printf("Usage: perf kvm stat <command>\n\n");
1401 
1402 	printf("# Available commands:\n");
1403 	printf("\trecord: record kvm events\n");
1404 	printf("\treport: report statistical data of kvm events\n");
1405 	printf("\tlive:   live reporting of statistical data of kvm events\n");
1406 
1407 	printf("\nOtherwise, it is the alias of 'perf stat':\n");
1408 }
1409 
1410 static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1411 {
1412 	struct perf_kvm_stat kvm = {
1413 		.file_name = file_name,
1414 
1415 		.trace_vcpu	= -1,
1416 		.report_event	= "vmexit",
1417 		.sort_key	= "sample",
1418 
1419 	};
1420 
1421 	if (argc == 1) {
1422 		print_kvm_stat_usage();
1423 		goto perf_stat;
1424 	}
1425 
1426 	if (!strncmp(argv[1], "rec", 3))
1427 		return kvm_events_record(&kvm, argc - 1, argv + 1);
1428 
1429 	if (!strncmp(argv[1], "rep", 3))
1430 		return kvm_events_report(&kvm, argc - 1 , argv + 1);
1431 
1432 #ifdef HAVE_TIMERFD_SUPPORT
1433 	if (!strncmp(argv[1], "live", 4))
1434 		return kvm_events_live(&kvm, argc - 1 , argv + 1);
1435 #endif
1436 
1437 perf_stat:
1438 	return cmd_stat(argc, argv, NULL);
1439 }
1440 #endif /* HAVE_KVM_STAT_SUPPORT */
1441 
1442 static int __cmd_record(const char *file_name, int argc, const char **argv)
1443 {
1444 	int rec_argc, i = 0, j;
1445 	const char **rec_argv;
1446 
1447 	rec_argc = argc + 2;
1448 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1449 	rec_argv[i++] = strdup("record");
1450 	rec_argv[i++] = strdup("-o");
1451 	rec_argv[i++] = strdup(file_name);
1452 	for (j = 1; j < argc; j++, i++)
1453 		rec_argv[i] = argv[j];
1454 
1455 	BUG_ON(i != rec_argc);
1456 
1457 	return cmd_record(i, rec_argv, NULL);
1458 }
1459 
1460 static int __cmd_report(const char *file_name, int argc, const char **argv)
1461 {
1462 	int rec_argc, i = 0, j;
1463 	const char **rec_argv;
1464 
1465 	rec_argc = argc + 2;
1466 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1467 	rec_argv[i++] = strdup("report");
1468 	rec_argv[i++] = strdup("-i");
1469 	rec_argv[i++] = strdup(file_name);
1470 	for (j = 1; j < argc; j++, i++)
1471 		rec_argv[i] = argv[j];
1472 
1473 	BUG_ON(i != rec_argc);
1474 
1475 	return cmd_report(i, rec_argv, NULL);
1476 }
1477 
1478 static int
1479 __cmd_buildid_list(const char *file_name, int argc, const char **argv)
1480 {
1481 	int rec_argc, i = 0, j;
1482 	const char **rec_argv;
1483 
1484 	rec_argc = argc + 2;
1485 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1486 	rec_argv[i++] = strdup("buildid-list");
1487 	rec_argv[i++] = strdup("-i");
1488 	rec_argv[i++] = strdup(file_name);
1489 	for (j = 1; j < argc; j++, i++)
1490 		rec_argv[i] = argv[j];
1491 
1492 	BUG_ON(i != rec_argc);
1493 
1494 	return cmd_buildid_list(i, rec_argv, NULL);
1495 }
1496 
1497 int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
1498 {
1499 	const char *file_name = NULL;
1500 	const struct option kvm_options[] = {
1501 		OPT_STRING('i', "input", &file_name, "file",
1502 			   "Input file name"),
1503 		OPT_STRING('o', "output", &file_name, "file",
1504 			   "Output file name"),
1505 		OPT_BOOLEAN(0, "guest", &perf_guest,
1506 			    "Collect guest os data"),
1507 		OPT_BOOLEAN(0, "host", &perf_host,
1508 			    "Collect host os data"),
1509 		OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1510 			   "guest mount directory under which every guest os"
1511 			   " instance has a subdir"),
1512 		OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1513 			   "file", "file saving guest os vmlinux"),
1514 		OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1515 			   "file", "file saving guest os /proc/kallsyms"),
1516 		OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1517 			   "file", "file saving guest os /proc/modules"),
1518 		OPT_INCR('v', "verbose", &verbose,
1519 			    "be more verbose (show counter open errors, etc)"),
1520 		OPT_END()
1521 	};
1522 
1523 	const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1524 						"buildid-list", "stat", NULL };
1525 	const char *kvm_usage[] = { NULL, NULL };
1526 
1527 	perf_host  = 0;
1528 	perf_guest = 1;
1529 
1530 	argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1531 					PARSE_OPT_STOP_AT_NON_OPTION);
1532 	if (!argc)
1533 		usage_with_options(kvm_usage, kvm_options);
1534 
1535 	if (!perf_host)
1536 		perf_guest = 1;
1537 
1538 	if (!file_name) {
1539 		file_name = get_filename_for_perf_kvm();
1540 
1541 		if (!file_name) {
1542 			pr_err("Failed to allocate memory for filename\n");
1543 			return -ENOMEM;
1544 		}
1545 	}
1546 
1547 	if (!strncmp(argv[0], "rec", 3))
1548 		return __cmd_record(file_name, argc, argv);
1549 	else if (!strncmp(argv[0], "rep", 3))
1550 		return __cmd_report(file_name, argc, argv);
1551 	else if (!strncmp(argv[0], "diff", 4))
1552 		return cmd_diff(argc, argv, NULL);
1553 	else if (!strncmp(argv[0], "top", 3))
1554 		return cmd_top(argc, argv, NULL);
1555 	else if (!strncmp(argv[0], "buildid-list", 12))
1556 		return __cmd_buildid_list(file_name, argc, argv);
1557 #ifdef HAVE_KVM_STAT_SUPPORT
1558 	else if (!strncmp(argv[0], "stat", 4))
1559 		return kvm_cmd_stat(file_name, argc, argv);
1560 #endif
1561 	else
1562 		usage_with_options(kvm_usage, kvm_options);
1563 
1564 	return 0;
1565 }
1566