xref: /openbmc/linux/tools/perf/builtin-kvm.c (revision 8b036556)
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(thread) && 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__set_priv(thread, vcpu_record);
390 	}
391 
392 	return thread__priv(thread);
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->opts.target.system_wide)
547 		pr_info("all VMs, ");
548 	else if (kvm->opts.target.pid)
549 		pr_info("pid(s) %s, ", kvm->opts.target.pid);
550 	else
551 		pr_info("dazed and confused on what is monitored, ");
552 
553 	if (vcpu == -1)
554 		pr_info("all VCPUs:\n\n");
555 	else
556 		pr_info("VCPU %d:\n\n", vcpu);
557 }
558 
559 static void show_timeofday(void)
560 {
561 	char date[64];
562 	struct timeval tv;
563 	struct tm ltime;
564 
565 	gettimeofday(&tv, NULL);
566 	if (localtime_r(&tv.tv_sec, &ltime)) {
567 		strftime(date, sizeof(date), "%H:%M:%S", &ltime);
568 		pr_info("%s.%06ld", date, tv.tv_usec);
569 	} else
570 		pr_info("00:00:00.000000");
571 
572 	return;
573 }
574 
575 static void print_result(struct perf_kvm_stat *kvm)
576 {
577 	char decode[DECODE_STR_LEN];
578 	struct kvm_event *event;
579 	int vcpu = kvm->trace_vcpu;
580 
581 	if (kvm->live) {
582 		puts(CONSOLE_CLEAR);
583 		show_timeofday();
584 	}
585 
586 	pr_info("\n\n");
587 	print_vcpu_info(kvm);
588 	pr_info("%*s ", DECODE_STR_LEN, kvm->events_ops->name);
589 	pr_info("%10s ", "Samples");
590 	pr_info("%9s ", "Samples%");
591 
592 	pr_info("%9s ", "Time%");
593 	pr_info("%11s ", "Min Time");
594 	pr_info("%11s ", "Max Time");
595 	pr_info("%16s ", "Avg time");
596 	pr_info("\n\n");
597 
598 	while ((event = pop_from_result(&kvm->result))) {
599 		u64 ecount, etime, max, min;
600 
601 		ecount = get_event_count(event, vcpu);
602 		etime = get_event_time(event, vcpu);
603 		max = get_event_max(event, vcpu);
604 		min = get_event_min(event, vcpu);
605 
606 		kvm->events_ops->decode_key(kvm, &event->key, decode);
607 		pr_info("%*s ", DECODE_STR_LEN, decode);
608 		pr_info("%10llu ", (unsigned long long)ecount);
609 		pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
610 		pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
611 		pr_info("%9.2fus ", (double)min / 1e3);
612 		pr_info("%9.2fus ", (double)max / 1e3);
613 		pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount/1e3,
614 			kvm_event_rel_stddev(vcpu, event));
615 		pr_info("\n");
616 	}
617 
618 	pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
619 		kvm->total_count, kvm->total_time / 1e3);
620 
621 	if (kvm->lost_events)
622 		pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
623 }
624 
625 #ifdef HAVE_TIMERFD_SUPPORT
626 static int process_lost_event(struct perf_tool *tool,
627 			      union perf_event *event __maybe_unused,
628 			      struct perf_sample *sample __maybe_unused,
629 			      struct machine *machine __maybe_unused)
630 {
631 	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
632 
633 	kvm->lost_events++;
634 	return 0;
635 }
636 #endif
637 
638 static bool skip_sample(struct perf_kvm_stat *kvm,
639 			struct perf_sample *sample)
640 {
641 	if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
642 		return true;
643 
644 	return false;
645 }
646 
647 static int process_sample_event(struct perf_tool *tool,
648 				union perf_event *event,
649 				struct perf_sample *sample,
650 				struct perf_evsel *evsel,
651 				struct machine *machine)
652 {
653 	struct thread *thread;
654 	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
655 						 tool);
656 
657 	if (skip_sample(kvm, sample))
658 		return 0;
659 
660 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
661 	if (thread == NULL) {
662 		pr_debug("problem processing %d event, skipping it.\n",
663 			event->header.type);
664 		return -1;
665 	}
666 
667 	if (!handle_kvm_event(kvm, thread, evsel, sample))
668 		return -1;
669 
670 	return 0;
671 }
672 
673 static int cpu_isa_config(struct perf_kvm_stat *kvm)
674 {
675 	char buf[64], *cpuid;
676 	int err;
677 
678 	if (kvm->live) {
679 		err = get_cpuid(buf, sizeof(buf));
680 		if (err != 0) {
681 			pr_err("Failed to look up CPU type\n");
682 			return err;
683 		}
684 		cpuid = buf;
685 	} else
686 		cpuid = kvm->session->header.env.cpuid;
687 
688 	if (!cpuid) {
689 		pr_err("Failed to look up CPU type\n");
690 		return -EINVAL;
691 	}
692 
693 	err = cpu_isa_init(kvm, cpuid);
694 	if (err == -ENOTSUP)
695 		pr_err("CPU %s is not supported.\n", cpuid);
696 
697 	return err;
698 }
699 
700 static bool verify_vcpu(int vcpu)
701 {
702 	if (vcpu != -1 && vcpu < 0) {
703 		pr_err("Invalid vcpu:%d.\n", vcpu);
704 		return false;
705 	}
706 
707 	return true;
708 }
709 
710 #ifdef HAVE_TIMERFD_SUPPORT
711 /* keeping the max events to a modest level to keep
712  * the processing of samples per mmap smooth.
713  */
714 #define PERF_KVM__MAX_EVENTS_PER_MMAP  25
715 
716 static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
717 				   u64 *mmap_time)
718 {
719 	union perf_event *event;
720 	struct perf_sample sample;
721 	s64 n = 0;
722 	int err;
723 
724 	*mmap_time = ULLONG_MAX;
725 	while ((event = perf_evlist__mmap_read(kvm->evlist, idx)) != NULL) {
726 		err = perf_evlist__parse_sample(kvm->evlist, event, &sample);
727 		if (err) {
728 			perf_evlist__mmap_consume(kvm->evlist, idx);
729 			pr_err("Failed to parse sample\n");
730 			return -1;
731 		}
732 
733 		err = perf_session_queue_event(kvm->session, event, &kvm->tool, &sample, 0);
734 		/*
735 		 * FIXME: Here we can't consume the event, as perf_session_queue_event will
736 		 *        point to it, and it'll get possibly overwritten by the kernel.
737 		 */
738 		perf_evlist__mmap_consume(kvm->evlist, idx);
739 
740 		if (err) {
741 			pr_err("Failed to enqueue sample: %d\n", err);
742 			return -1;
743 		}
744 
745 		/* save time stamp of our first sample for this mmap */
746 		if (n == 0)
747 			*mmap_time = sample.time;
748 
749 		/* limit events per mmap handled all at once */
750 		n++;
751 		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
752 			break;
753 	}
754 
755 	return n;
756 }
757 
758 static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
759 {
760 	int i, err, throttled = 0;
761 	s64 n, ntotal = 0;
762 	u64 flush_time = ULLONG_MAX, mmap_time;
763 
764 	for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
765 		n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
766 		if (n < 0)
767 			return -1;
768 
769 		/* flush time is going to be the minimum of all the individual
770 		 * mmap times. Essentially, we flush all the samples queued up
771 		 * from the last pass under our minimal start time -- that leaves
772 		 * a very small race for samples to come in with a lower timestamp.
773 		 * The ioctl to return the perf_clock timestamp should close the
774 		 * race entirely.
775 		 */
776 		if (mmap_time < flush_time)
777 			flush_time = mmap_time;
778 
779 		ntotal += n;
780 		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
781 			throttled = 1;
782 	}
783 
784 	/* flush queue after each round in which we processed events */
785 	if (ntotal) {
786 		kvm->session->ordered_events.next_flush = flush_time;
787 		err = kvm->tool.finished_round(&kvm->tool, NULL, kvm->session);
788 		if (err) {
789 			if (kvm->lost_events)
790 				pr_info("\nLost events: %" PRIu64 "\n\n",
791 					kvm->lost_events);
792 			return err;
793 		}
794 	}
795 
796 	return throttled;
797 }
798 
799 static volatile int done;
800 
801 static void sig_handler(int sig __maybe_unused)
802 {
803 	done = 1;
804 }
805 
806 static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
807 {
808 	struct itimerspec new_value;
809 	int rc = -1;
810 
811 	kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
812 	if (kvm->timerfd < 0) {
813 		pr_err("timerfd_create failed\n");
814 		goto out;
815 	}
816 
817 	new_value.it_value.tv_sec = kvm->display_time;
818 	new_value.it_value.tv_nsec = 0;
819 	new_value.it_interval.tv_sec = kvm->display_time;
820 	new_value.it_interval.tv_nsec = 0;
821 
822 	if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
823 		pr_err("timerfd_settime failed: %d\n", errno);
824 		close(kvm->timerfd);
825 		goto out;
826 	}
827 
828 	rc = 0;
829 out:
830 	return rc;
831 }
832 
833 static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
834 {
835 	uint64_t c;
836 	int rc;
837 
838 	rc = read(kvm->timerfd, &c, sizeof(uint64_t));
839 	if (rc < 0) {
840 		if (errno == EAGAIN)
841 			return 0;
842 
843 		pr_err("Failed to read timer fd: %d\n", errno);
844 		return -1;
845 	}
846 
847 	if (rc != sizeof(uint64_t)) {
848 		pr_err("Error reading timer fd - invalid size returned\n");
849 		return -1;
850 	}
851 
852 	if (c != 1)
853 		pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
854 
855 	/* update display */
856 	sort_result(kvm);
857 	print_result(kvm);
858 
859 	/* reset counts */
860 	clear_events_cache_stats(kvm->kvm_events_cache);
861 	kvm->total_count = 0;
862 	kvm->total_time = 0;
863 	kvm->lost_events = 0;
864 
865 	return 0;
866 }
867 
868 static int fd_set_nonblock(int fd)
869 {
870 	long arg = 0;
871 
872 	arg = fcntl(fd, F_GETFL);
873 	if (arg < 0) {
874 		pr_err("Failed to get current flags for fd %d\n", fd);
875 		return -1;
876 	}
877 
878 	if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
879 		pr_err("Failed to set non-block option on fd %d\n", fd);
880 		return -1;
881 	}
882 
883 	return 0;
884 }
885 
886 static int perf_kvm__handle_stdin(void)
887 {
888 	int c;
889 
890 	c = getc(stdin);
891 	if (c == 'q')
892 		return 1;
893 
894 	return 0;
895 }
896 
897 static int kvm_events_live_report(struct perf_kvm_stat *kvm)
898 {
899 	int nr_stdin, ret, err = -EINVAL;
900 	struct termios save;
901 
902 	/* live flag must be set first */
903 	kvm->live = true;
904 
905 	ret = cpu_isa_config(kvm);
906 	if (ret < 0)
907 		return ret;
908 
909 	if (!verify_vcpu(kvm->trace_vcpu) ||
910 	    !select_key(kvm) ||
911 	    !register_kvm_events_ops(kvm)) {
912 		goto out;
913 	}
914 
915 	set_term_quiet_input(&save);
916 	init_kvm_event_record(kvm);
917 
918 	signal(SIGINT, sig_handler);
919 	signal(SIGTERM, sig_handler);
920 
921 	/* add timer fd */
922 	if (perf_kvm__timerfd_create(kvm) < 0) {
923 		err = -1;
924 		goto out;
925 	}
926 
927 	if (perf_evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
928 		goto out;
929 
930 	nr_stdin = perf_evlist__add_pollfd(kvm->evlist, fileno(stdin));
931 	if (nr_stdin < 0)
932 		goto out;
933 
934 	if (fd_set_nonblock(fileno(stdin)) != 0)
935 		goto out;
936 
937 	/* everything is good - enable the events and process */
938 	perf_evlist__enable(kvm->evlist);
939 
940 	while (!done) {
941 		struct fdarray *fda = &kvm->evlist->pollfd;
942 		int rc;
943 
944 		rc = perf_kvm__mmap_read(kvm);
945 		if (rc < 0)
946 			break;
947 
948 		err = perf_kvm__handle_timerfd(kvm);
949 		if (err)
950 			goto out;
951 
952 		if (fda->entries[nr_stdin].revents & POLLIN)
953 			done = perf_kvm__handle_stdin();
954 
955 		if (!rc && !done)
956 			err = fdarray__poll(fda, 100);
957 	}
958 
959 	perf_evlist__disable(kvm->evlist);
960 
961 	if (err == 0) {
962 		sort_result(kvm);
963 		print_result(kvm);
964 	}
965 
966 out:
967 	if (kvm->timerfd >= 0)
968 		close(kvm->timerfd);
969 
970 	tcsetattr(0, TCSAFLUSH, &save);
971 	return err;
972 }
973 
974 static int kvm_live_open_events(struct perf_kvm_stat *kvm)
975 {
976 	int err, rc = -1;
977 	struct perf_evsel *pos;
978 	struct perf_evlist *evlist = kvm->evlist;
979 	char sbuf[STRERR_BUFSIZE];
980 
981 	perf_evlist__config(evlist, &kvm->opts);
982 
983 	/*
984 	 * Note: exclude_{guest,host} do not apply here.
985 	 *       This command processes KVM tracepoints from host only
986 	 */
987 	evlist__for_each(evlist, pos) {
988 		struct perf_event_attr *attr = &pos->attr;
989 
990 		/* make sure these *are* set */
991 		perf_evsel__set_sample_bit(pos, TID);
992 		perf_evsel__set_sample_bit(pos, TIME);
993 		perf_evsel__set_sample_bit(pos, CPU);
994 		perf_evsel__set_sample_bit(pos, RAW);
995 		/* make sure these are *not*; want as small a sample as possible */
996 		perf_evsel__reset_sample_bit(pos, PERIOD);
997 		perf_evsel__reset_sample_bit(pos, IP);
998 		perf_evsel__reset_sample_bit(pos, CALLCHAIN);
999 		perf_evsel__reset_sample_bit(pos, ADDR);
1000 		perf_evsel__reset_sample_bit(pos, READ);
1001 		attr->mmap = 0;
1002 		attr->comm = 0;
1003 		attr->task = 0;
1004 
1005 		attr->sample_period = 1;
1006 
1007 		attr->watermark = 0;
1008 		attr->wakeup_events = 1000;
1009 
1010 		/* will enable all once we are ready */
1011 		attr->disabled = 1;
1012 	}
1013 
1014 	err = perf_evlist__open(evlist);
1015 	if (err < 0) {
1016 		printf("Couldn't create the events: %s\n",
1017 		       strerror_r(errno, sbuf, sizeof(sbuf)));
1018 		goto out;
1019 	}
1020 
1021 	if (perf_evlist__mmap(evlist, kvm->opts.mmap_pages, false) < 0) {
1022 		ui__error("Failed to mmap the events: %s\n",
1023 			  strerror_r(errno, sbuf, sizeof(sbuf)));
1024 		perf_evlist__close(evlist);
1025 		goto out;
1026 	}
1027 
1028 	rc = 0;
1029 
1030 out:
1031 	return rc;
1032 }
1033 #endif
1034 
1035 static int read_events(struct perf_kvm_stat *kvm)
1036 {
1037 	int ret;
1038 
1039 	struct perf_tool eops = {
1040 		.sample			= process_sample_event,
1041 		.comm			= perf_event__process_comm,
1042 		.ordered_events		= true,
1043 	};
1044 	struct perf_data_file file = {
1045 		.path = kvm->file_name,
1046 		.mode = PERF_DATA_MODE_READ,
1047 	};
1048 
1049 	kvm->tool = eops;
1050 	kvm->session = perf_session__new(&file, false, &kvm->tool);
1051 	if (!kvm->session) {
1052 		pr_err("Initializing perf session failed\n");
1053 		return -1;
1054 	}
1055 
1056 	symbol__init(&kvm->session->header.env);
1057 
1058 	if (!perf_session__has_traces(kvm->session, "kvm record"))
1059 		return -EINVAL;
1060 
1061 	/*
1062 	 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1063 	 * traced in the old kernel.
1064 	 */
1065 	ret = cpu_isa_config(kvm);
1066 	if (ret < 0)
1067 		return ret;
1068 
1069 	return perf_session__process_events(kvm->session, &kvm->tool);
1070 }
1071 
1072 static int parse_target_str(struct perf_kvm_stat *kvm)
1073 {
1074 	if (kvm->opts.target.pid) {
1075 		kvm->pid_list = intlist__new(kvm->opts.target.pid);
1076 		if (kvm->pid_list == NULL) {
1077 			pr_err("Error parsing process id string\n");
1078 			return -EINVAL;
1079 		}
1080 	}
1081 
1082 	return 0;
1083 }
1084 
1085 static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1086 {
1087 	int ret = -EINVAL;
1088 	int vcpu = kvm->trace_vcpu;
1089 
1090 	if (parse_target_str(kvm) != 0)
1091 		goto exit;
1092 
1093 	if (!verify_vcpu(vcpu))
1094 		goto exit;
1095 
1096 	if (!select_key(kvm))
1097 		goto exit;
1098 
1099 	if (!register_kvm_events_ops(kvm))
1100 		goto exit;
1101 
1102 	init_kvm_event_record(kvm);
1103 	setup_pager();
1104 
1105 	ret = read_events(kvm);
1106 	if (ret)
1107 		goto exit;
1108 
1109 	sort_result(kvm);
1110 	print_result(kvm);
1111 
1112 exit:
1113 	return ret;
1114 }
1115 
1116 #define STRDUP_FAIL_EXIT(s)		\
1117 	({	char *_p;		\
1118 	_p = strdup(s);		\
1119 		if (!_p)		\
1120 			return -ENOMEM;	\
1121 		_p;			\
1122 	})
1123 
1124 static int
1125 kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1126 {
1127 	unsigned int rec_argc, i, j, events_tp_size;
1128 	const char **rec_argv;
1129 	const char * const record_args[] = {
1130 		"record",
1131 		"-R",
1132 		"-m", "1024",
1133 		"-c", "1",
1134 	};
1135 	const char * const kvm_stat_record_usage[] = {
1136 		"perf kvm stat record [<options>]",
1137 		NULL
1138 	};
1139 	const char * const *events_tp;
1140 	events_tp_size = 0;
1141 
1142 	for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1143 		events_tp_size++;
1144 
1145 	rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1146 		   2 * events_tp_size;
1147 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1148 
1149 	if (rec_argv == NULL)
1150 		return -ENOMEM;
1151 
1152 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1153 		rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1154 
1155 	for (j = 0; j < events_tp_size; j++) {
1156 		rec_argv[i++] = "-e";
1157 		rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1158 	}
1159 
1160 	rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1161 	rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1162 
1163 	for (j = 1; j < (unsigned int)argc; j++, i++)
1164 		rec_argv[i] = argv[j];
1165 
1166 	set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1167 	set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1168 	set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1169 
1170 	set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1171 	set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1172 	set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1173 	set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1174 	set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1175 	set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1176 	set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1177 	set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1178 	set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1179 	set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1180 	set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1181 	set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1182 	set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1183 	set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1184 	set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1185 
1186 	record_usage = kvm_stat_record_usage;
1187 	return cmd_record(i, rec_argv, NULL);
1188 }
1189 
1190 static int
1191 kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1192 {
1193 	const struct option kvm_events_report_options[] = {
1194 		OPT_STRING(0, "event", &kvm->report_event, "report event",
1195 			   "event for reporting: vmexit, "
1196 			   "mmio (x86 only), ioport (x86 only)"),
1197 		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1198 			    "vcpu id to report"),
1199 		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1200 			    "key for sorting: sample(sort by samples number)"
1201 			    " time (sort by avg time)"),
1202 		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1203 			   "analyze events only for given process id(s)"),
1204 		OPT_END()
1205 	};
1206 
1207 	const char * const kvm_events_report_usage[] = {
1208 		"perf kvm stat report [<options>]",
1209 		NULL
1210 	};
1211 
1212 	if (argc) {
1213 		argc = parse_options(argc, argv,
1214 				     kvm_events_report_options,
1215 				     kvm_events_report_usage, 0);
1216 		if (argc)
1217 			usage_with_options(kvm_events_report_usage,
1218 					   kvm_events_report_options);
1219 	}
1220 
1221 	if (!kvm->opts.target.pid)
1222 		kvm->opts.target.system_wide = true;
1223 
1224 	return kvm_events_report_vcpu(kvm);
1225 }
1226 
1227 #ifdef HAVE_TIMERFD_SUPPORT
1228 static struct perf_evlist *kvm_live_event_list(void)
1229 {
1230 	struct perf_evlist *evlist;
1231 	char *tp, *name, *sys;
1232 	int err = -1;
1233 	const char * const *events_tp;
1234 
1235 	evlist = perf_evlist__new();
1236 	if (evlist == NULL)
1237 		return NULL;
1238 
1239 	for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1240 
1241 		tp = strdup(*events_tp);
1242 		if (tp == NULL)
1243 			goto out;
1244 
1245 		/* split tracepoint into subsystem and name */
1246 		sys = tp;
1247 		name = strchr(tp, ':');
1248 		if (name == NULL) {
1249 			pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1250 			       *events_tp);
1251 			free(tp);
1252 			goto out;
1253 		}
1254 		*name = '\0';
1255 		name++;
1256 
1257 		if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1258 			pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1259 			free(tp);
1260 			goto out;
1261 		}
1262 
1263 		free(tp);
1264 	}
1265 
1266 	err = 0;
1267 
1268 out:
1269 	if (err) {
1270 		perf_evlist__delete(evlist);
1271 		evlist = NULL;
1272 	}
1273 
1274 	return evlist;
1275 }
1276 
1277 static int kvm_events_live(struct perf_kvm_stat *kvm,
1278 			   int argc, const char **argv)
1279 {
1280 	char errbuf[BUFSIZ];
1281 	int err;
1282 
1283 	const struct option live_options[] = {
1284 		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1285 			"record events on existing process id"),
1286 		OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1287 			"number of mmap data pages",
1288 			perf_evlist__parse_mmap_pages),
1289 		OPT_INCR('v', "verbose", &verbose,
1290 			"be more verbose (show counter open errors, etc)"),
1291 		OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1292 			"system-wide collection from all CPUs"),
1293 		OPT_UINTEGER('d', "display", &kvm->display_time,
1294 			"time in seconds between display updates"),
1295 		OPT_STRING(0, "event", &kvm->report_event, "report event",
1296 			"event for reporting: "
1297 			"vmexit, mmio (x86 only), ioport (x86 only)"),
1298 		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1299 			"vcpu id to report"),
1300 		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1301 			"key for sorting: sample(sort by samples number)"
1302 			" time (sort by avg time)"),
1303 		OPT_U64(0, "duration", &kvm->duration,
1304 			"show events other than"
1305 			" HLT (x86 only) or Wait state (s390 only)"
1306 			" that take longer than duration usecs"),
1307 		OPT_END()
1308 	};
1309 	const char * const live_usage[] = {
1310 		"perf kvm stat live [<options>]",
1311 		NULL
1312 	};
1313 	struct perf_data_file file = {
1314 		.mode = PERF_DATA_MODE_WRITE,
1315 	};
1316 
1317 
1318 	/* event handling */
1319 	kvm->tool.sample = process_sample_event;
1320 	kvm->tool.comm   = perf_event__process_comm;
1321 	kvm->tool.exit   = perf_event__process_exit;
1322 	kvm->tool.fork   = perf_event__process_fork;
1323 	kvm->tool.lost   = process_lost_event;
1324 	kvm->tool.ordered_events = true;
1325 	perf_tool__fill_defaults(&kvm->tool);
1326 
1327 	/* set defaults */
1328 	kvm->display_time = 1;
1329 	kvm->opts.user_interval = 1;
1330 	kvm->opts.mmap_pages = 512;
1331 	kvm->opts.target.uses_mmap = false;
1332 	kvm->opts.target.uid_str = NULL;
1333 	kvm->opts.target.uid = UINT_MAX;
1334 
1335 	symbol__init(NULL);
1336 	disable_buildid_cache();
1337 
1338 	use_browser = 0;
1339 	setup_browser(false);
1340 
1341 	if (argc) {
1342 		argc = parse_options(argc, argv, live_options,
1343 				     live_usage, 0);
1344 		if (argc)
1345 			usage_with_options(live_usage, live_options);
1346 	}
1347 
1348 	kvm->duration *= NSEC_PER_USEC;   /* convert usec to nsec */
1349 
1350 	/*
1351 	 * target related setups
1352 	 */
1353 	err = target__validate(&kvm->opts.target);
1354 	if (err) {
1355 		target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1356 		ui__warning("%s", errbuf);
1357 	}
1358 
1359 	if (target__none(&kvm->opts.target))
1360 		kvm->opts.target.system_wide = true;
1361 
1362 
1363 	/*
1364 	 * generate the event list
1365 	 */
1366 	kvm->evlist = kvm_live_event_list();
1367 	if (kvm->evlist == NULL) {
1368 		err = -1;
1369 		goto out;
1370 	}
1371 
1372 	symbol_conf.nr_events = kvm->evlist->nr_entries;
1373 
1374 	if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1375 		usage_with_options(live_usage, live_options);
1376 
1377 	/*
1378 	 * perf session
1379 	 */
1380 	kvm->session = perf_session__new(&file, false, &kvm->tool);
1381 	if (kvm->session == NULL) {
1382 		err = -1;
1383 		goto out;
1384 	}
1385 	kvm->session->evlist = kvm->evlist;
1386 	perf_session__set_id_hdr_size(kvm->session);
1387 	ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
1388 	machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1389 				    kvm->evlist->threads, false);
1390 	err = kvm_live_open_events(kvm);
1391 	if (err)
1392 		goto out;
1393 
1394 	err = kvm_events_live_report(kvm);
1395 
1396 out:
1397 	exit_browser(0);
1398 
1399 	if (kvm->session)
1400 		perf_session__delete(kvm->session);
1401 	kvm->session = NULL;
1402 	if (kvm->evlist)
1403 		perf_evlist__delete(kvm->evlist);
1404 
1405 	return err;
1406 }
1407 #endif
1408 
1409 static void print_kvm_stat_usage(void)
1410 {
1411 	printf("Usage: perf kvm stat <command>\n\n");
1412 
1413 	printf("# Available commands:\n");
1414 	printf("\trecord: record kvm events\n");
1415 	printf("\treport: report statistical data of kvm events\n");
1416 	printf("\tlive:   live reporting of statistical data of kvm events\n");
1417 
1418 	printf("\nOtherwise, it is the alias of 'perf stat':\n");
1419 }
1420 
1421 static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1422 {
1423 	struct perf_kvm_stat kvm = {
1424 		.file_name = file_name,
1425 
1426 		.trace_vcpu	= -1,
1427 		.report_event	= "vmexit",
1428 		.sort_key	= "sample",
1429 
1430 	};
1431 
1432 	if (argc == 1) {
1433 		print_kvm_stat_usage();
1434 		goto perf_stat;
1435 	}
1436 
1437 	if (!strncmp(argv[1], "rec", 3))
1438 		return kvm_events_record(&kvm, argc - 1, argv + 1);
1439 
1440 	if (!strncmp(argv[1], "rep", 3))
1441 		return kvm_events_report(&kvm, argc - 1 , argv + 1);
1442 
1443 #ifdef HAVE_TIMERFD_SUPPORT
1444 	if (!strncmp(argv[1], "live", 4))
1445 		return kvm_events_live(&kvm, argc - 1 , argv + 1);
1446 #endif
1447 
1448 perf_stat:
1449 	return cmd_stat(argc, argv, NULL);
1450 }
1451 #endif /* HAVE_KVM_STAT_SUPPORT */
1452 
1453 static int __cmd_record(const char *file_name, int argc, const char **argv)
1454 {
1455 	int rec_argc, i = 0, j;
1456 	const char **rec_argv;
1457 
1458 	rec_argc = argc + 2;
1459 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1460 	rec_argv[i++] = strdup("record");
1461 	rec_argv[i++] = strdup("-o");
1462 	rec_argv[i++] = strdup(file_name);
1463 	for (j = 1; j < argc; j++, i++)
1464 		rec_argv[i] = argv[j];
1465 
1466 	BUG_ON(i != rec_argc);
1467 
1468 	return cmd_record(i, rec_argv, NULL);
1469 }
1470 
1471 static int __cmd_report(const char *file_name, int argc, const char **argv)
1472 {
1473 	int rec_argc, i = 0, j;
1474 	const char **rec_argv;
1475 
1476 	rec_argc = argc + 2;
1477 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1478 	rec_argv[i++] = strdup("report");
1479 	rec_argv[i++] = strdup("-i");
1480 	rec_argv[i++] = strdup(file_name);
1481 	for (j = 1; j < argc; j++, i++)
1482 		rec_argv[i] = argv[j];
1483 
1484 	BUG_ON(i != rec_argc);
1485 
1486 	return cmd_report(i, rec_argv, NULL);
1487 }
1488 
1489 static int
1490 __cmd_buildid_list(const char *file_name, int argc, const char **argv)
1491 {
1492 	int rec_argc, i = 0, j;
1493 	const char **rec_argv;
1494 
1495 	rec_argc = argc + 2;
1496 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1497 	rec_argv[i++] = strdup("buildid-list");
1498 	rec_argv[i++] = strdup("-i");
1499 	rec_argv[i++] = strdup(file_name);
1500 	for (j = 1; j < argc; j++, i++)
1501 		rec_argv[i] = argv[j];
1502 
1503 	BUG_ON(i != rec_argc);
1504 
1505 	return cmd_buildid_list(i, rec_argv, NULL);
1506 }
1507 
1508 int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
1509 {
1510 	const char *file_name = NULL;
1511 	const struct option kvm_options[] = {
1512 		OPT_STRING('i', "input", &file_name, "file",
1513 			   "Input file name"),
1514 		OPT_STRING('o', "output", &file_name, "file",
1515 			   "Output file name"),
1516 		OPT_BOOLEAN(0, "guest", &perf_guest,
1517 			    "Collect guest os data"),
1518 		OPT_BOOLEAN(0, "host", &perf_host,
1519 			    "Collect host os data"),
1520 		OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1521 			   "guest mount directory under which every guest os"
1522 			   " instance has a subdir"),
1523 		OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1524 			   "file", "file saving guest os vmlinux"),
1525 		OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1526 			   "file", "file saving guest os /proc/kallsyms"),
1527 		OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1528 			   "file", "file saving guest os /proc/modules"),
1529 		OPT_INCR('v', "verbose", &verbose,
1530 			    "be more verbose (show counter open errors, etc)"),
1531 		OPT_END()
1532 	};
1533 
1534 	const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1535 						"buildid-list", "stat", NULL };
1536 	const char *kvm_usage[] = { NULL, NULL };
1537 
1538 	perf_host  = 0;
1539 	perf_guest = 1;
1540 
1541 	argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1542 					PARSE_OPT_STOP_AT_NON_OPTION);
1543 	if (!argc)
1544 		usage_with_options(kvm_usage, kvm_options);
1545 
1546 	if (!perf_host)
1547 		perf_guest = 1;
1548 
1549 	if (!file_name) {
1550 		file_name = get_filename_for_perf_kvm();
1551 
1552 		if (!file_name) {
1553 			pr_err("Failed to allocate memory for filename\n");
1554 			return -ENOMEM;
1555 		}
1556 	}
1557 
1558 	if (!strncmp(argv[0], "rec", 3))
1559 		return __cmd_record(file_name, argc, argv);
1560 	else if (!strncmp(argv[0], "rep", 3))
1561 		return __cmd_report(file_name, argc, argv);
1562 	else if (!strncmp(argv[0], "diff", 4))
1563 		return cmd_diff(argc, argv, NULL);
1564 	else if (!strncmp(argv[0], "top", 3))
1565 		return cmd_top(argc, argv, NULL);
1566 	else if (!strncmp(argv[0], "buildid-list", 12))
1567 		return __cmd_buildid_list(file_name, argc, argv);
1568 #ifdef HAVE_KVM_STAT_SUPPORT
1569 	else if (!strncmp(argv[0], "stat", 4))
1570 		return kvm_cmd_stat(file_name, argc, argv);
1571 #endif
1572 	else
1573 		usage_with_options(kvm_usage, kvm_options);
1574 
1575 	return 0;
1576 }
1577