xref: /openbmc/linux/tools/perf/builtin-kvm.c (revision 3821a065)
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 #include "util/ordered-events.h"
22 
23 #include <sys/prctl.h>
24 #ifdef HAVE_TIMERFD_SUPPORT
25 #include <sys/timerfd.h>
26 #endif
27 
28 #include <termios.h>
29 #include <semaphore.h>
30 #include <pthread.h>
31 #include <math.h>
32 
33 #ifdef HAVE_KVM_STAT_SUPPORT
34 #include <asm/kvm_perf.h>
35 #include "util/kvm-stat.h"
36 
37 void exit_event_get_key(struct perf_evsel *evsel,
38 			struct perf_sample *sample,
39 			struct event_key *key)
40 {
41 	key->info = 0;
42 	key->key = perf_evsel__intval(evsel, sample, KVM_EXIT_REASON);
43 }
44 
45 bool kvm_exit_event(struct perf_evsel *evsel)
46 {
47 	return !strcmp(evsel->name, KVM_EXIT_TRACE);
48 }
49 
50 bool exit_event_begin(struct perf_evsel *evsel,
51 		      struct perf_sample *sample, struct event_key *key)
52 {
53 	if (kvm_exit_event(evsel)) {
54 		exit_event_get_key(evsel, sample, key);
55 		return true;
56 	}
57 
58 	return false;
59 }
60 
61 bool kvm_entry_event(struct perf_evsel *evsel)
62 {
63 	return !strcmp(evsel->name, KVM_ENTRY_TRACE);
64 }
65 
66 bool exit_event_end(struct perf_evsel *evsel,
67 		    struct perf_sample *sample __maybe_unused,
68 		    struct event_key *key __maybe_unused)
69 {
70 	return kvm_entry_event(evsel);
71 }
72 
73 static const char *get_exit_reason(struct perf_kvm_stat *kvm,
74 				   struct exit_reasons_table *tbl,
75 				   u64 exit_code)
76 {
77 	while (tbl->reason != NULL) {
78 		if (tbl->exit_code == exit_code)
79 			return tbl->reason;
80 		tbl++;
81 	}
82 
83 	pr_err("unknown kvm exit code:%lld on %s\n",
84 		(unsigned long long)exit_code, kvm->exit_reasons_isa);
85 	return "UNKNOWN";
86 }
87 
88 void exit_event_decode_key(struct perf_kvm_stat *kvm,
89 			   struct event_key *key,
90 			   char *decode)
91 {
92 	const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
93 						  key->key);
94 
95 	scnprintf(decode, DECODE_STR_LEN, "%s", exit_reason);
96 }
97 
98 static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
99 {
100 	struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
101 
102 	for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
103 		if (!strcmp(events_ops->name, kvm->report_event)) {
104 			kvm->events_ops = events_ops->ops;
105 			return true;
106 		}
107 	}
108 
109 	return false;
110 }
111 
112 struct vcpu_event_record {
113 	int vcpu_id;
114 	u64 start_time;
115 	struct kvm_event *last_event;
116 };
117 
118 
119 static void init_kvm_event_record(struct perf_kvm_stat *kvm)
120 {
121 	unsigned int i;
122 
123 	for (i = 0; i < EVENTS_CACHE_SIZE; i++)
124 		INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
125 }
126 
127 #ifdef HAVE_TIMERFD_SUPPORT
128 static void clear_events_cache_stats(struct list_head *kvm_events_cache)
129 {
130 	struct list_head *head;
131 	struct kvm_event *event;
132 	unsigned int i;
133 	int j;
134 
135 	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
136 		head = &kvm_events_cache[i];
137 		list_for_each_entry(event, head, hash_entry) {
138 			/* reset stats for event */
139 			event->total.time = 0;
140 			init_stats(&event->total.stats);
141 
142 			for (j = 0; j < event->max_vcpu; ++j) {
143 				event->vcpu[j].time = 0;
144 				init_stats(&event->vcpu[j].stats);
145 			}
146 		}
147 	}
148 }
149 #endif
150 
151 static int kvm_events_hash_fn(u64 key)
152 {
153 	return key & (EVENTS_CACHE_SIZE - 1);
154 }
155 
156 static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
157 {
158 	int old_max_vcpu = event->max_vcpu;
159 	void *prev;
160 
161 	if (vcpu_id < event->max_vcpu)
162 		return true;
163 
164 	while (event->max_vcpu <= vcpu_id)
165 		event->max_vcpu += DEFAULT_VCPU_NUM;
166 
167 	prev = event->vcpu;
168 	event->vcpu = realloc(event->vcpu,
169 			      event->max_vcpu * sizeof(*event->vcpu));
170 	if (!event->vcpu) {
171 		free(prev);
172 		pr_err("Not enough memory\n");
173 		return false;
174 	}
175 
176 	memset(event->vcpu + old_max_vcpu, 0,
177 	       (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
178 	return true;
179 }
180 
181 static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
182 {
183 	struct kvm_event *event;
184 
185 	event = zalloc(sizeof(*event));
186 	if (!event) {
187 		pr_err("Not enough memory\n");
188 		return NULL;
189 	}
190 
191 	event->key = *key;
192 	init_stats(&event->total.stats);
193 	return event;
194 }
195 
196 static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
197 					       struct event_key *key)
198 {
199 	struct kvm_event *event;
200 	struct list_head *head;
201 
202 	BUG_ON(key->key == INVALID_KEY);
203 
204 	head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
205 	list_for_each_entry(event, head, hash_entry) {
206 		if (event->key.key == key->key && event->key.info == key->info)
207 			return event;
208 	}
209 
210 	event = kvm_alloc_init_event(key);
211 	if (!event)
212 		return NULL;
213 
214 	list_add(&event->hash_entry, head);
215 	return event;
216 }
217 
218 static bool handle_begin_event(struct perf_kvm_stat *kvm,
219 			       struct vcpu_event_record *vcpu_record,
220 			       struct event_key *key, u64 timestamp)
221 {
222 	struct kvm_event *event = NULL;
223 
224 	if (key->key != INVALID_KEY)
225 		event = find_create_kvm_event(kvm, key);
226 
227 	vcpu_record->last_event = event;
228 	vcpu_record->start_time = timestamp;
229 	return true;
230 }
231 
232 static void
233 kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
234 {
235 	kvm_stats->time += time_diff;
236 	update_stats(&kvm_stats->stats, time_diff);
237 }
238 
239 static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
240 {
241 	struct kvm_event_stats *kvm_stats = &event->total;
242 
243 	if (vcpu_id != -1)
244 		kvm_stats = &event->vcpu[vcpu_id];
245 
246 	return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
247 				avg_stats(&kvm_stats->stats));
248 }
249 
250 static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
251 			     u64 time_diff)
252 {
253 	if (vcpu_id == -1) {
254 		kvm_update_event_stats(&event->total, time_diff);
255 		return true;
256 	}
257 
258 	if (!kvm_event_expand(event, vcpu_id))
259 		return false;
260 
261 	kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
262 	return true;
263 }
264 
265 static bool is_child_event(struct perf_kvm_stat *kvm,
266 			   struct perf_evsel *evsel,
267 			   struct perf_sample *sample,
268 			   struct event_key *key)
269 {
270 	struct child_event_ops *child_ops;
271 
272 	child_ops = kvm->events_ops->child_ops;
273 
274 	if (!child_ops)
275 		return false;
276 
277 	for (; child_ops->name; child_ops++) {
278 		if (!strcmp(evsel->name, child_ops->name)) {
279 			child_ops->get_key(evsel, sample, key);
280 			return true;
281 		}
282 	}
283 
284 	return false;
285 }
286 
287 static bool handle_child_event(struct perf_kvm_stat *kvm,
288 			       struct vcpu_event_record *vcpu_record,
289 			       struct event_key *key,
290 			       struct perf_sample *sample __maybe_unused)
291 {
292 	struct kvm_event *event = NULL;
293 
294 	if (key->key != INVALID_KEY)
295 		event = find_create_kvm_event(kvm, key);
296 
297 	vcpu_record->last_event = event;
298 
299 	return true;
300 }
301 
302 static bool skip_event(const char *event)
303 {
304 	const char * const *skip_events;
305 
306 	for (skip_events = kvm_skip_events; *skip_events; skip_events++)
307 		if (!strcmp(event, *skip_events))
308 			return true;
309 
310 	return false;
311 }
312 
313 static bool handle_end_event(struct perf_kvm_stat *kvm,
314 			     struct vcpu_event_record *vcpu_record,
315 			     struct event_key *key,
316 			     struct perf_sample *sample)
317 {
318 	struct kvm_event *event;
319 	u64 time_begin, time_diff;
320 	int vcpu;
321 
322 	if (kvm->trace_vcpu == -1)
323 		vcpu = -1;
324 	else
325 		vcpu = vcpu_record->vcpu_id;
326 
327 	event = vcpu_record->last_event;
328 	time_begin = vcpu_record->start_time;
329 
330 	/* The begin event is not caught. */
331 	if (!time_begin)
332 		return true;
333 
334 	/*
335 	 * In some case, the 'begin event' only records the start timestamp,
336 	 * the actual event is recognized in the 'end event' (e.g. mmio-event).
337 	 */
338 
339 	/* Both begin and end events did not get the key. */
340 	if (!event && key->key == INVALID_KEY)
341 		return true;
342 
343 	if (!event)
344 		event = find_create_kvm_event(kvm, key);
345 
346 	if (!event)
347 		return false;
348 
349 	vcpu_record->last_event = NULL;
350 	vcpu_record->start_time = 0;
351 
352 	/* seems to happen once in a while during live mode */
353 	if (sample->time < time_begin) {
354 		pr_debug("End time before begin time; skipping event.\n");
355 		return true;
356 	}
357 
358 	time_diff = sample->time - time_begin;
359 
360 	if (kvm->duration && time_diff > kvm->duration) {
361 		char decode[DECODE_STR_LEN];
362 
363 		kvm->events_ops->decode_key(kvm, &event->key, decode);
364 		if (!skip_event(decode)) {
365 			pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
366 				 sample->time, sample->pid, vcpu_record->vcpu_id,
367 				 decode, time_diff/1000);
368 		}
369 	}
370 
371 	return update_kvm_event(event, vcpu, time_diff);
372 }
373 
374 static
375 struct vcpu_event_record *per_vcpu_record(struct thread *thread,
376 					  struct perf_evsel *evsel,
377 					  struct perf_sample *sample)
378 {
379 	/* Only kvm_entry records vcpu id. */
380 	if (!thread__priv(thread) && kvm_entry_event(evsel)) {
381 		struct vcpu_event_record *vcpu_record;
382 
383 		vcpu_record = zalloc(sizeof(*vcpu_record));
384 		if (!vcpu_record) {
385 			pr_err("%s: Not enough memory\n", __func__);
386 			return NULL;
387 		}
388 
389 		vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample, VCPU_ID);
390 		thread__set_priv(thread, vcpu_record);
391 	}
392 
393 	return thread__priv(thread);
394 }
395 
396 static bool handle_kvm_event(struct perf_kvm_stat *kvm,
397 			     struct thread *thread,
398 			     struct perf_evsel *evsel,
399 			     struct perf_sample *sample)
400 {
401 	struct vcpu_event_record *vcpu_record;
402 	struct event_key key = { .key = INVALID_KEY,
403 				 .exit_reasons = kvm->exit_reasons };
404 
405 	vcpu_record = per_vcpu_record(thread, evsel, sample);
406 	if (!vcpu_record)
407 		return true;
408 
409 	/* only process events for vcpus user cares about */
410 	if ((kvm->trace_vcpu != -1) &&
411 	    (kvm->trace_vcpu != vcpu_record->vcpu_id))
412 		return true;
413 
414 	if (kvm->events_ops->is_begin_event(evsel, sample, &key))
415 		return handle_begin_event(kvm, vcpu_record, &key, sample->time);
416 
417 	if (is_child_event(kvm, evsel, sample, &key))
418 		return handle_child_event(kvm, vcpu_record, &key, sample);
419 
420 	if (kvm->events_ops->is_end_event(evsel, sample, &key))
421 		return handle_end_event(kvm, vcpu_record, &key, sample);
422 
423 	return true;
424 }
425 
426 #define GET_EVENT_KEY(func, field)					\
427 static u64 get_event_ ##func(struct kvm_event *event, int vcpu)		\
428 {									\
429 	if (vcpu == -1)							\
430 		return event->total.field;				\
431 									\
432 	if (vcpu >= event->max_vcpu)					\
433 		return 0;						\
434 									\
435 	return event->vcpu[vcpu].field;					\
436 }
437 
438 #define COMPARE_EVENT_KEY(func, field)					\
439 GET_EVENT_KEY(func, field)						\
440 static int compare_kvm_event_ ## func(struct kvm_event *one,		\
441 					struct kvm_event *two, int vcpu)\
442 {									\
443 	return get_event_ ##func(one, vcpu) >				\
444 				get_event_ ##func(two, vcpu);		\
445 }
446 
447 GET_EVENT_KEY(time, time);
448 COMPARE_EVENT_KEY(count, stats.n);
449 COMPARE_EVENT_KEY(mean, stats.mean);
450 GET_EVENT_KEY(max, stats.max);
451 GET_EVENT_KEY(min, stats.min);
452 
453 #define DEF_SORT_NAME_KEY(name, compare_key)				\
454 	{ #name, compare_kvm_event_ ## compare_key }
455 
456 static struct kvm_event_key keys[] = {
457 	DEF_SORT_NAME_KEY(sample, count),
458 	DEF_SORT_NAME_KEY(time, mean),
459 	{ NULL, NULL }
460 };
461 
462 static bool select_key(struct perf_kvm_stat *kvm)
463 {
464 	int i;
465 
466 	for (i = 0; keys[i].name; i++) {
467 		if (!strcmp(keys[i].name, kvm->sort_key)) {
468 			kvm->compare = keys[i].key;
469 			return true;
470 		}
471 	}
472 
473 	pr_err("Unknown compare key:%s\n", kvm->sort_key);
474 	return false;
475 }
476 
477 static void insert_to_result(struct rb_root *result, struct kvm_event *event,
478 			     key_cmp_fun bigger, int vcpu)
479 {
480 	struct rb_node **rb = &result->rb_node;
481 	struct rb_node *parent = NULL;
482 	struct kvm_event *p;
483 
484 	while (*rb) {
485 		p = container_of(*rb, struct kvm_event, rb);
486 		parent = *rb;
487 
488 		if (bigger(event, p, vcpu))
489 			rb = &(*rb)->rb_left;
490 		else
491 			rb = &(*rb)->rb_right;
492 	}
493 
494 	rb_link_node(&event->rb, parent, rb);
495 	rb_insert_color(&event->rb, result);
496 }
497 
498 static void
499 update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
500 {
501 	int vcpu = kvm->trace_vcpu;
502 
503 	kvm->total_count += get_event_count(event, vcpu);
504 	kvm->total_time += get_event_time(event, vcpu);
505 }
506 
507 static bool event_is_valid(struct kvm_event *event, int vcpu)
508 {
509 	return !!get_event_count(event, vcpu);
510 }
511 
512 static void sort_result(struct perf_kvm_stat *kvm)
513 {
514 	unsigned int i;
515 	int vcpu = kvm->trace_vcpu;
516 	struct kvm_event *event;
517 
518 	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
519 		list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
520 			if (event_is_valid(event, vcpu)) {
521 				update_total_count(kvm, event);
522 				insert_to_result(&kvm->result, event,
523 						 kvm->compare, vcpu);
524 			}
525 		}
526 	}
527 }
528 
529 /* returns left most element of result, and erase it */
530 static struct kvm_event *pop_from_result(struct rb_root *result)
531 {
532 	struct rb_node *node = rb_first(result);
533 
534 	if (!node)
535 		return NULL;
536 
537 	rb_erase(node, result);
538 	return container_of(node, struct kvm_event, rb);
539 }
540 
541 static void print_vcpu_info(struct perf_kvm_stat *kvm)
542 {
543 	int vcpu = kvm->trace_vcpu;
544 
545 	pr_info("Analyze events for ");
546 
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 	if (vcpu == -1)
555 		pr_info("all VCPUs:\n\n");
556 	else
557 		pr_info("VCPU %d:\n\n", vcpu);
558 }
559 
560 static void show_timeofday(void)
561 {
562 	char date[64];
563 	struct timeval tv;
564 	struct tm ltime;
565 
566 	gettimeofday(&tv, NULL);
567 	if (localtime_r(&tv.tv_sec, &ltime)) {
568 		strftime(date, sizeof(date), "%H:%M:%S", &ltime);
569 		pr_info("%s.%06ld", date, tv.tv_usec);
570 	} else
571 		pr_info("00:00:00.000000");
572 
573 	return;
574 }
575 
576 static void print_result(struct perf_kvm_stat *kvm)
577 {
578 	char decode[DECODE_STR_LEN];
579 	struct kvm_event *event;
580 	int vcpu = kvm->trace_vcpu;
581 
582 	if (kvm->live) {
583 		puts(CONSOLE_CLEAR);
584 		show_timeofday();
585 	}
586 
587 	pr_info("\n\n");
588 	print_vcpu_info(kvm);
589 	pr_info("%*s ", DECODE_STR_LEN, kvm->events_ops->name);
590 	pr_info("%10s ", "Samples");
591 	pr_info("%9s ", "Samples%");
592 
593 	pr_info("%9s ", "Time%");
594 	pr_info("%11s ", "Min Time");
595 	pr_info("%11s ", "Max Time");
596 	pr_info("%16s ", "Avg time");
597 	pr_info("\n\n");
598 
599 	while ((event = pop_from_result(&kvm->result))) {
600 		u64 ecount, etime, max, min;
601 
602 		ecount = get_event_count(event, vcpu);
603 		etime = get_event_time(event, vcpu);
604 		max = get_event_max(event, vcpu);
605 		min = get_event_min(event, vcpu);
606 
607 		kvm->events_ops->decode_key(kvm, &event->key, decode);
608 		pr_info("%*s ", DECODE_STR_LEN, decode);
609 		pr_info("%10llu ", (unsigned long long)ecount);
610 		pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
611 		pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
612 		pr_info("%9.2fus ", (double)min / 1e3);
613 		pr_info("%9.2fus ", (double)max / 1e3);
614 		pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount/1e3,
615 			kvm_event_rel_stddev(vcpu, event));
616 		pr_info("\n");
617 	}
618 
619 	pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
620 		kvm->total_count, kvm->total_time / 1e3);
621 
622 	if (kvm->lost_events)
623 		pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
624 }
625 
626 #ifdef HAVE_TIMERFD_SUPPORT
627 static int process_lost_event(struct perf_tool *tool,
628 			      union perf_event *event __maybe_unused,
629 			      struct perf_sample *sample __maybe_unused,
630 			      struct machine *machine __maybe_unused)
631 {
632 	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
633 
634 	kvm->lost_events++;
635 	return 0;
636 }
637 #endif
638 
639 static bool skip_sample(struct perf_kvm_stat *kvm,
640 			struct perf_sample *sample)
641 {
642 	if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
643 		return true;
644 
645 	return false;
646 }
647 
648 static int process_sample_event(struct perf_tool *tool,
649 				union perf_event *event,
650 				struct perf_sample *sample,
651 				struct perf_evsel *evsel,
652 				struct machine *machine)
653 {
654 	int err = 0;
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 		err = -1;
671 
672 	thread__put(thread);
673 	return err;
674 }
675 
676 static int cpu_isa_config(struct perf_kvm_stat *kvm)
677 {
678 	char buf[64], *cpuid;
679 	int err;
680 
681 	if (kvm->live) {
682 		err = get_cpuid(buf, sizeof(buf));
683 		if (err != 0) {
684 			pr_err("Failed to look up CPU type\n");
685 			return err;
686 		}
687 		cpuid = buf;
688 	} else
689 		cpuid = kvm->session->header.env.cpuid;
690 
691 	if (!cpuid) {
692 		pr_err("Failed to look up CPU type\n");
693 		return -EINVAL;
694 	}
695 
696 	err = cpu_isa_init(kvm, cpuid);
697 	if (err == -ENOTSUP)
698 		pr_err("CPU %s is not supported.\n", cpuid);
699 
700 	return err;
701 }
702 
703 static bool verify_vcpu(int vcpu)
704 {
705 	if (vcpu != -1 && vcpu < 0) {
706 		pr_err("Invalid vcpu:%d.\n", vcpu);
707 		return false;
708 	}
709 
710 	return true;
711 }
712 
713 #ifdef HAVE_TIMERFD_SUPPORT
714 /* keeping the max events to a modest level to keep
715  * the processing of samples per mmap smooth.
716  */
717 #define PERF_KVM__MAX_EVENTS_PER_MMAP  25
718 
719 static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
720 				   u64 *mmap_time)
721 {
722 	union perf_event *event;
723 	struct perf_sample sample;
724 	s64 n = 0;
725 	int err;
726 
727 	*mmap_time = ULLONG_MAX;
728 	while ((event = perf_evlist__mmap_read(kvm->evlist, idx)) != NULL) {
729 		err = perf_evlist__parse_sample(kvm->evlist, event, &sample);
730 		if (err) {
731 			perf_evlist__mmap_consume(kvm->evlist, idx);
732 			pr_err("Failed to parse sample\n");
733 			return -1;
734 		}
735 
736 		err = perf_session__queue_event(kvm->session, event, &sample, 0);
737 		/*
738 		 * FIXME: Here we can't consume the event, as perf_session__queue_event will
739 		 *        point to it, and it'll get possibly overwritten by the kernel.
740 		 */
741 		perf_evlist__mmap_consume(kvm->evlist, idx);
742 
743 		if (err) {
744 			pr_err("Failed to enqueue sample: %d\n", err);
745 			return -1;
746 		}
747 
748 		/* save time stamp of our first sample for this mmap */
749 		if (n == 0)
750 			*mmap_time = sample.time;
751 
752 		/* limit events per mmap handled all at once */
753 		n++;
754 		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
755 			break;
756 	}
757 
758 	return n;
759 }
760 
761 static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
762 {
763 	int i, err, throttled = 0;
764 	s64 n, ntotal = 0;
765 	u64 flush_time = ULLONG_MAX, mmap_time;
766 
767 	for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
768 		n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
769 		if (n < 0)
770 			return -1;
771 
772 		/* flush time is going to be the minimum of all the individual
773 		 * mmap times. Essentially, we flush all the samples queued up
774 		 * from the last pass under our minimal start time -- that leaves
775 		 * a very small race for samples to come in with a lower timestamp.
776 		 * The ioctl to return the perf_clock timestamp should close the
777 		 * race entirely.
778 		 */
779 		if (mmap_time < flush_time)
780 			flush_time = mmap_time;
781 
782 		ntotal += n;
783 		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
784 			throttled = 1;
785 	}
786 
787 	/* flush queue after each round in which we processed events */
788 	if (ntotal) {
789 		struct ordered_events *oe = &kvm->session->ordered_events;
790 
791 		oe->next_flush = flush_time;
792 		err = ordered_events__flush(oe, OE_FLUSH__ROUND);
793 		if (err) {
794 			if (kvm->lost_events)
795 				pr_info("\nLost events: %" PRIu64 "\n\n",
796 					kvm->lost_events);
797 			return err;
798 		}
799 	}
800 
801 	return throttled;
802 }
803 
804 static volatile int done;
805 
806 static void sig_handler(int sig __maybe_unused)
807 {
808 	done = 1;
809 }
810 
811 static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
812 {
813 	struct itimerspec new_value;
814 	int rc = -1;
815 
816 	kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
817 	if (kvm->timerfd < 0) {
818 		pr_err("timerfd_create failed\n");
819 		goto out;
820 	}
821 
822 	new_value.it_value.tv_sec = kvm->display_time;
823 	new_value.it_value.tv_nsec = 0;
824 	new_value.it_interval.tv_sec = kvm->display_time;
825 	new_value.it_interval.tv_nsec = 0;
826 
827 	if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
828 		pr_err("timerfd_settime failed: %d\n", errno);
829 		close(kvm->timerfd);
830 		goto out;
831 	}
832 
833 	rc = 0;
834 out:
835 	return rc;
836 }
837 
838 static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
839 {
840 	uint64_t c;
841 	int rc;
842 
843 	rc = read(kvm->timerfd, &c, sizeof(uint64_t));
844 	if (rc < 0) {
845 		if (errno == EAGAIN)
846 			return 0;
847 
848 		pr_err("Failed to read timer fd: %d\n", errno);
849 		return -1;
850 	}
851 
852 	if (rc != sizeof(uint64_t)) {
853 		pr_err("Error reading timer fd - invalid size returned\n");
854 		return -1;
855 	}
856 
857 	if (c != 1)
858 		pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
859 
860 	/* update display */
861 	sort_result(kvm);
862 	print_result(kvm);
863 
864 	/* reset counts */
865 	clear_events_cache_stats(kvm->kvm_events_cache);
866 	kvm->total_count = 0;
867 	kvm->total_time = 0;
868 	kvm->lost_events = 0;
869 
870 	return 0;
871 }
872 
873 static int fd_set_nonblock(int fd)
874 {
875 	long arg = 0;
876 
877 	arg = fcntl(fd, F_GETFL);
878 	if (arg < 0) {
879 		pr_err("Failed to get current flags for fd %d\n", fd);
880 		return -1;
881 	}
882 
883 	if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
884 		pr_err("Failed to set non-block option on fd %d\n", fd);
885 		return -1;
886 	}
887 
888 	return 0;
889 }
890 
891 static int perf_kvm__handle_stdin(void)
892 {
893 	int c;
894 
895 	c = getc(stdin);
896 	if (c == 'q')
897 		return 1;
898 
899 	return 0;
900 }
901 
902 static int kvm_events_live_report(struct perf_kvm_stat *kvm)
903 {
904 	int nr_stdin, ret, err = -EINVAL;
905 	struct termios save;
906 
907 	/* live flag must be set first */
908 	kvm->live = true;
909 
910 	ret = cpu_isa_config(kvm);
911 	if (ret < 0)
912 		return ret;
913 
914 	if (!verify_vcpu(kvm->trace_vcpu) ||
915 	    !select_key(kvm) ||
916 	    !register_kvm_events_ops(kvm)) {
917 		goto out;
918 	}
919 
920 	set_term_quiet_input(&save);
921 	init_kvm_event_record(kvm);
922 
923 	signal(SIGINT, sig_handler);
924 	signal(SIGTERM, sig_handler);
925 
926 	/* add timer fd */
927 	if (perf_kvm__timerfd_create(kvm) < 0) {
928 		err = -1;
929 		goto out;
930 	}
931 
932 	if (perf_evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
933 		goto out;
934 
935 	nr_stdin = perf_evlist__add_pollfd(kvm->evlist, fileno(stdin));
936 	if (nr_stdin < 0)
937 		goto out;
938 
939 	if (fd_set_nonblock(fileno(stdin)) != 0)
940 		goto out;
941 
942 	/* everything is good - enable the events and process */
943 	perf_evlist__enable(kvm->evlist);
944 
945 	while (!done) {
946 		struct fdarray *fda = &kvm->evlist->pollfd;
947 		int rc;
948 
949 		rc = perf_kvm__mmap_read(kvm);
950 		if (rc < 0)
951 			break;
952 
953 		err = perf_kvm__handle_timerfd(kvm);
954 		if (err)
955 			goto out;
956 
957 		if (fda->entries[nr_stdin].revents & POLLIN)
958 			done = perf_kvm__handle_stdin();
959 
960 		if (!rc && !done)
961 			err = fdarray__poll(fda, 100);
962 	}
963 
964 	perf_evlist__disable(kvm->evlist);
965 
966 	if (err == 0) {
967 		sort_result(kvm);
968 		print_result(kvm);
969 	}
970 
971 out:
972 	if (kvm->timerfd >= 0)
973 		close(kvm->timerfd);
974 
975 	tcsetattr(0, TCSAFLUSH, &save);
976 	return err;
977 }
978 
979 static int kvm_live_open_events(struct perf_kvm_stat *kvm)
980 {
981 	int err, rc = -1;
982 	struct perf_evsel *pos;
983 	struct perf_evlist *evlist = kvm->evlist;
984 	char sbuf[STRERR_BUFSIZE];
985 
986 	perf_evlist__config(evlist, &kvm->opts);
987 
988 	/*
989 	 * Note: exclude_{guest,host} do not apply here.
990 	 *       This command processes KVM tracepoints from host only
991 	 */
992 	evlist__for_each(evlist, pos) {
993 		struct perf_event_attr *attr = &pos->attr;
994 
995 		/* make sure these *are* set */
996 		perf_evsel__set_sample_bit(pos, TID);
997 		perf_evsel__set_sample_bit(pos, TIME);
998 		perf_evsel__set_sample_bit(pos, CPU);
999 		perf_evsel__set_sample_bit(pos, RAW);
1000 		/* make sure these are *not*; want as small a sample as possible */
1001 		perf_evsel__reset_sample_bit(pos, PERIOD);
1002 		perf_evsel__reset_sample_bit(pos, IP);
1003 		perf_evsel__reset_sample_bit(pos, CALLCHAIN);
1004 		perf_evsel__reset_sample_bit(pos, ADDR);
1005 		perf_evsel__reset_sample_bit(pos, READ);
1006 		attr->mmap = 0;
1007 		attr->comm = 0;
1008 		attr->task = 0;
1009 
1010 		attr->sample_period = 1;
1011 
1012 		attr->watermark = 0;
1013 		attr->wakeup_events = 1000;
1014 
1015 		/* will enable all once we are ready */
1016 		attr->disabled = 1;
1017 	}
1018 
1019 	err = perf_evlist__open(evlist);
1020 	if (err < 0) {
1021 		printf("Couldn't create the events: %s\n",
1022 		       strerror_r(errno, sbuf, sizeof(sbuf)));
1023 		goto out;
1024 	}
1025 
1026 	if (perf_evlist__mmap(evlist, kvm->opts.mmap_pages, false) < 0) {
1027 		ui__error("Failed to mmap the events: %s\n",
1028 			  strerror_r(errno, sbuf, sizeof(sbuf)));
1029 		perf_evlist__close(evlist);
1030 		goto out;
1031 	}
1032 
1033 	rc = 0;
1034 
1035 out:
1036 	return rc;
1037 }
1038 #endif
1039 
1040 static int read_events(struct perf_kvm_stat *kvm)
1041 {
1042 	int ret;
1043 
1044 	struct perf_tool eops = {
1045 		.sample			= process_sample_event,
1046 		.comm			= perf_event__process_comm,
1047 		.ordered_events		= true,
1048 	};
1049 	struct perf_data_file file = {
1050 		.path = kvm->file_name,
1051 		.mode = PERF_DATA_MODE_READ,
1052 		.force = kvm->force,
1053 	};
1054 
1055 	kvm->tool = eops;
1056 	kvm->session = perf_session__new(&file, false, &kvm->tool);
1057 	if (!kvm->session) {
1058 		pr_err("Initializing perf session failed\n");
1059 		return -1;
1060 	}
1061 
1062 	symbol__init(&kvm->session->header.env);
1063 
1064 	if (!perf_session__has_traces(kvm->session, "kvm record")) {
1065 		ret = -EINVAL;
1066 		goto out_delete;
1067 	}
1068 
1069 	/*
1070 	 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1071 	 * traced in the old kernel.
1072 	 */
1073 	ret = cpu_isa_config(kvm);
1074 	if (ret < 0)
1075 		goto out_delete;
1076 
1077 	ret = perf_session__process_events(kvm->session);
1078 
1079 out_delete:
1080 	perf_session__delete(kvm->session);
1081 	return ret;
1082 }
1083 
1084 static int parse_target_str(struct perf_kvm_stat *kvm)
1085 {
1086 	if (kvm->opts.target.pid) {
1087 		kvm->pid_list = intlist__new(kvm->opts.target.pid);
1088 		if (kvm->pid_list == NULL) {
1089 			pr_err("Error parsing process id string\n");
1090 			return -EINVAL;
1091 		}
1092 	}
1093 
1094 	return 0;
1095 }
1096 
1097 static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1098 {
1099 	int ret = -EINVAL;
1100 	int vcpu = kvm->trace_vcpu;
1101 
1102 	if (parse_target_str(kvm) != 0)
1103 		goto exit;
1104 
1105 	if (!verify_vcpu(vcpu))
1106 		goto exit;
1107 
1108 	if (!select_key(kvm))
1109 		goto exit;
1110 
1111 	if (!register_kvm_events_ops(kvm))
1112 		goto exit;
1113 
1114 	init_kvm_event_record(kvm);
1115 	setup_pager();
1116 
1117 	ret = read_events(kvm);
1118 	if (ret)
1119 		goto exit;
1120 
1121 	sort_result(kvm);
1122 	print_result(kvm);
1123 
1124 exit:
1125 	return ret;
1126 }
1127 
1128 #define STRDUP_FAIL_EXIT(s)		\
1129 	({	char *_p;		\
1130 	_p = strdup(s);		\
1131 		if (!_p)		\
1132 			return -ENOMEM;	\
1133 		_p;			\
1134 	})
1135 
1136 static int
1137 kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1138 {
1139 	unsigned int rec_argc, i, j, events_tp_size;
1140 	const char **rec_argv;
1141 	const char * const record_args[] = {
1142 		"record",
1143 		"-R",
1144 		"-m", "1024",
1145 		"-c", "1",
1146 	};
1147 	const char * const kvm_stat_record_usage[] = {
1148 		"perf kvm stat record [<options>]",
1149 		NULL
1150 	};
1151 	const char * const *events_tp;
1152 	events_tp_size = 0;
1153 
1154 	for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1155 		events_tp_size++;
1156 
1157 	rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1158 		   2 * events_tp_size;
1159 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1160 
1161 	if (rec_argv == NULL)
1162 		return -ENOMEM;
1163 
1164 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1165 		rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1166 
1167 	for (j = 0; j < events_tp_size; j++) {
1168 		rec_argv[i++] = "-e";
1169 		rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1170 	}
1171 
1172 	rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1173 	rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1174 
1175 	for (j = 1; j < (unsigned int)argc; j++, i++)
1176 		rec_argv[i] = argv[j];
1177 
1178 	set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1179 	set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1180 	set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1181 
1182 	set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1183 	set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1184 	set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1185 	set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1186 	set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1187 	set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1188 	set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1189 	set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1190 	set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1191 	set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1192 	set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1193 	set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1194 	set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1195 	set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1196 	set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1197 
1198 	record_usage = kvm_stat_record_usage;
1199 	return cmd_record(i, rec_argv, NULL);
1200 }
1201 
1202 static int
1203 kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1204 {
1205 	const struct option kvm_events_report_options[] = {
1206 		OPT_STRING(0, "event", &kvm->report_event, "report event",
1207 			   "event for reporting: vmexit, "
1208 			   "mmio (x86 only), ioport (x86 only)"),
1209 		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1210 			    "vcpu id to report"),
1211 		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1212 			    "key for sorting: sample(sort by samples number)"
1213 			    " time (sort by avg time)"),
1214 		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1215 			   "analyze events only for given process id(s)"),
1216 		OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
1217 		OPT_END()
1218 	};
1219 
1220 	const char * const kvm_events_report_usage[] = {
1221 		"perf kvm stat report [<options>]",
1222 		NULL
1223 	};
1224 
1225 	if (argc) {
1226 		argc = parse_options(argc, argv,
1227 				     kvm_events_report_options,
1228 				     kvm_events_report_usage, 0);
1229 		if (argc)
1230 			usage_with_options(kvm_events_report_usage,
1231 					   kvm_events_report_options);
1232 	}
1233 
1234 	if (!kvm->opts.target.pid)
1235 		kvm->opts.target.system_wide = true;
1236 
1237 	return kvm_events_report_vcpu(kvm);
1238 }
1239 
1240 #ifdef HAVE_TIMERFD_SUPPORT
1241 static struct perf_evlist *kvm_live_event_list(void)
1242 {
1243 	struct perf_evlist *evlist;
1244 	char *tp, *name, *sys;
1245 	int err = -1;
1246 	const char * const *events_tp;
1247 
1248 	evlist = perf_evlist__new();
1249 	if (evlist == NULL)
1250 		return NULL;
1251 
1252 	for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1253 
1254 		tp = strdup(*events_tp);
1255 		if (tp == NULL)
1256 			goto out;
1257 
1258 		/* split tracepoint into subsystem and name */
1259 		sys = tp;
1260 		name = strchr(tp, ':');
1261 		if (name == NULL) {
1262 			pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1263 			       *events_tp);
1264 			free(tp);
1265 			goto out;
1266 		}
1267 		*name = '\0';
1268 		name++;
1269 
1270 		if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1271 			pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1272 			free(tp);
1273 			goto out;
1274 		}
1275 
1276 		free(tp);
1277 	}
1278 
1279 	err = 0;
1280 
1281 out:
1282 	if (err) {
1283 		perf_evlist__delete(evlist);
1284 		evlist = NULL;
1285 	}
1286 
1287 	return evlist;
1288 }
1289 
1290 static int kvm_events_live(struct perf_kvm_stat *kvm,
1291 			   int argc, const char **argv)
1292 {
1293 	char errbuf[BUFSIZ];
1294 	int err;
1295 
1296 	const struct option live_options[] = {
1297 		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1298 			"record events on existing process id"),
1299 		OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1300 			"number of mmap data pages",
1301 			perf_evlist__parse_mmap_pages),
1302 		OPT_INCR('v', "verbose", &verbose,
1303 			"be more verbose (show counter open errors, etc)"),
1304 		OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1305 			"system-wide collection from all CPUs"),
1306 		OPT_UINTEGER('d', "display", &kvm->display_time,
1307 			"time in seconds between display updates"),
1308 		OPT_STRING(0, "event", &kvm->report_event, "report event",
1309 			"event for reporting: "
1310 			"vmexit, mmio (x86 only), ioport (x86 only)"),
1311 		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1312 			"vcpu id to report"),
1313 		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1314 			"key for sorting: sample(sort by samples number)"
1315 			" time (sort by avg time)"),
1316 		OPT_U64(0, "duration", &kvm->duration,
1317 			"show events other than"
1318 			" HLT (x86 only) or Wait state (s390 only)"
1319 			" that take longer than duration usecs"),
1320 		OPT_UINTEGER(0, "proc-map-timeout", &kvm->opts.proc_map_timeout,
1321 				"per thread proc mmap processing timeout in ms"),
1322 		OPT_END()
1323 	};
1324 	const char * const live_usage[] = {
1325 		"perf kvm stat live [<options>]",
1326 		NULL
1327 	};
1328 	struct perf_data_file file = {
1329 		.mode = PERF_DATA_MODE_WRITE,
1330 	};
1331 
1332 
1333 	/* event handling */
1334 	kvm->tool.sample = process_sample_event;
1335 	kvm->tool.comm   = perf_event__process_comm;
1336 	kvm->tool.exit   = perf_event__process_exit;
1337 	kvm->tool.fork   = perf_event__process_fork;
1338 	kvm->tool.lost   = process_lost_event;
1339 	kvm->tool.ordered_events = true;
1340 	perf_tool__fill_defaults(&kvm->tool);
1341 
1342 	/* set defaults */
1343 	kvm->display_time = 1;
1344 	kvm->opts.user_interval = 1;
1345 	kvm->opts.mmap_pages = 512;
1346 	kvm->opts.target.uses_mmap = false;
1347 	kvm->opts.target.uid_str = NULL;
1348 	kvm->opts.target.uid = UINT_MAX;
1349 	kvm->opts.proc_map_timeout = 500;
1350 
1351 	symbol__init(NULL);
1352 	disable_buildid_cache();
1353 
1354 	use_browser = 0;
1355 	setup_browser(false);
1356 
1357 	if (argc) {
1358 		argc = parse_options(argc, argv, live_options,
1359 				     live_usage, 0);
1360 		if (argc)
1361 			usage_with_options(live_usage, live_options);
1362 	}
1363 
1364 	kvm->duration *= NSEC_PER_USEC;   /* convert usec to nsec */
1365 
1366 	/*
1367 	 * target related setups
1368 	 */
1369 	err = target__validate(&kvm->opts.target);
1370 	if (err) {
1371 		target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1372 		ui__warning("%s", errbuf);
1373 	}
1374 
1375 	if (target__none(&kvm->opts.target))
1376 		kvm->opts.target.system_wide = true;
1377 
1378 
1379 	/*
1380 	 * generate the event list
1381 	 */
1382 	kvm->evlist = kvm_live_event_list();
1383 	if (kvm->evlist == NULL) {
1384 		err = -1;
1385 		goto out;
1386 	}
1387 
1388 	symbol_conf.nr_events = kvm->evlist->nr_entries;
1389 
1390 	if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1391 		usage_with_options(live_usage, live_options);
1392 
1393 	/*
1394 	 * perf session
1395 	 */
1396 	kvm->session = perf_session__new(&file, false, &kvm->tool);
1397 	if (kvm->session == NULL) {
1398 		err = -1;
1399 		goto out;
1400 	}
1401 	kvm->session->evlist = kvm->evlist;
1402 	perf_session__set_id_hdr_size(kvm->session);
1403 	ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
1404 	machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1405 				    kvm->evlist->threads, false, kvm->opts.proc_map_timeout);
1406 	err = kvm_live_open_events(kvm);
1407 	if (err)
1408 		goto out;
1409 
1410 	err = kvm_events_live_report(kvm);
1411 
1412 out:
1413 	exit_browser(0);
1414 
1415 	if (kvm->session)
1416 		perf_session__delete(kvm->session);
1417 	kvm->session = NULL;
1418 	if (kvm->evlist)
1419 		perf_evlist__delete(kvm->evlist);
1420 
1421 	return err;
1422 }
1423 #endif
1424 
1425 static void print_kvm_stat_usage(void)
1426 {
1427 	printf("Usage: perf kvm stat <command>\n\n");
1428 
1429 	printf("# Available commands:\n");
1430 	printf("\trecord: record kvm events\n");
1431 	printf("\treport: report statistical data of kvm events\n");
1432 	printf("\tlive:   live reporting of statistical data of kvm events\n");
1433 
1434 	printf("\nOtherwise, it is the alias of 'perf stat':\n");
1435 }
1436 
1437 static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1438 {
1439 	struct perf_kvm_stat kvm = {
1440 		.file_name = file_name,
1441 
1442 		.trace_vcpu	= -1,
1443 		.report_event	= "vmexit",
1444 		.sort_key	= "sample",
1445 
1446 	};
1447 
1448 	if (argc == 1) {
1449 		print_kvm_stat_usage();
1450 		goto perf_stat;
1451 	}
1452 
1453 	if (!strncmp(argv[1], "rec", 3))
1454 		return kvm_events_record(&kvm, argc - 1, argv + 1);
1455 
1456 	if (!strncmp(argv[1], "rep", 3))
1457 		return kvm_events_report(&kvm, argc - 1 , argv + 1);
1458 
1459 #ifdef HAVE_TIMERFD_SUPPORT
1460 	if (!strncmp(argv[1], "live", 4))
1461 		return kvm_events_live(&kvm, argc - 1 , argv + 1);
1462 #endif
1463 
1464 perf_stat:
1465 	return cmd_stat(argc, argv, NULL);
1466 }
1467 #endif /* HAVE_KVM_STAT_SUPPORT */
1468 
1469 static int __cmd_record(const char *file_name, int argc, const char **argv)
1470 {
1471 	int rec_argc, i = 0, j;
1472 	const char **rec_argv;
1473 
1474 	rec_argc = argc + 2;
1475 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1476 	rec_argv[i++] = strdup("record");
1477 	rec_argv[i++] = strdup("-o");
1478 	rec_argv[i++] = strdup(file_name);
1479 	for (j = 1; j < argc; j++, i++)
1480 		rec_argv[i] = argv[j];
1481 
1482 	BUG_ON(i != rec_argc);
1483 
1484 	return cmd_record(i, rec_argv, NULL);
1485 }
1486 
1487 static int __cmd_report(const char *file_name, int argc, const char **argv)
1488 {
1489 	int rec_argc, i = 0, j;
1490 	const char **rec_argv;
1491 
1492 	rec_argc = argc + 2;
1493 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1494 	rec_argv[i++] = strdup("report");
1495 	rec_argv[i++] = strdup("-i");
1496 	rec_argv[i++] = strdup(file_name);
1497 	for (j = 1; j < argc; j++, i++)
1498 		rec_argv[i] = argv[j];
1499 
1500 	BUG_ON(i != rec_argc);
1501 
1502 	return cmd_report(i, rec_argv, NULL);
1503 }
1504 
1505 static int
1506 __cmd_buildid_list(const char *file_name, int argc, const char **argv)
1507 {
1508 	int rec_argc, i = 0, j;
1509 	const char **rec_argv;
1510 
1511 	rec_argc = argc + 2;
1512 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1513 	rec_argv[i++] = strdup("buildid-list");
1514 	rec_argv[i++] = strdup("-i");
1515 	rec_argv[i++] = strdup(file_name);
1516 	for (j = 1; j < argc; j++, i++)
1517 		rec_argv[i] = argv[j];
1518 
1519 	BUG_ON(i != rec_argc);
1520 
1521 	return cmd_buildid_list(i, rec_argv, NULL);
1522 }
1523 
1524 int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
1525 {
1526 	const char *file_name = NULL;
1527 	const struct option kvm_options[] = {
1528 		OPT_STRING('i', "input", &file_name, "file",
1529 			   "Input file name"),
1530 		OPT_STRING('o', "output", &file_name, "file",
1531 			   "Output file name"),
1532 		OPT_BOOLEAN(0, "guest", &perf_guest,
1533 			    "Collect guest os data"),
1534 		OPT_BOOLEAN(0, "host", &perf_host,
1535 			    "Collect host os data"),
1536 		OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1537 			   "guest mount directory under which every guest os"
1538 			   " instance has a subdir"),
1539 		OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1540 			   "file", "file saving guest os vmlinux"),
1541 		OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1542 			   "file", "file saving guest os /proc/kallsyms"),
1543 		OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1544 			   "file", "file saving guest os /proc/modules"),
1545 		OPT_INCR('v', "verbose", &verbose,
1546 			    "be more verbose (show counter open errors, etc)"),
1547 		OPT_END()
1548 	};
1549 
1550 	const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1551 						"buildid-list", "stat", NULL };
1552 	const char *kvm_usage[] = { NULL, NULL };
1553 
1554 	perf_host  = 0;
1555 	perf_guest = 1;
1556 
1557 	argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1558 					PARSE_OPT_STOP_AT_NON_OPTION);
1559 	if (!argc)
1560 		usage_with_options(kvm_usage, kvm_options);
1561 
1562 	if (!perf_host)
1563 		perf_guest = 1;
1564 
1565 	if (!file_name) {
1566 		file_name = get_filename_for_perf_kvm();
1567 
1568 		if (!file_name) {
1569 			pr_err("Failed to allocate memory for filename\n");
1570 			return -ENOMEM;
1571 		}
1572 	}
1573 
1574 	if (!strncmp(argv[0], "rec", 3))
1575 		return __cmd_record(file_name, argc, argv);
1576 	else if (!strncmp(argv[0], "rep", 3))
1577 		return __cmd_report(file_name, argc, argv);
1578 	else if (!strncmp(argv[0], "diff", 4))
1579 		return cmd_diff(argc, argv, NULL);
1580 	else if (!strncmp(argv[0], "top", 3))
1581 		return cmd_top(argc, argv, NULL);
1582 	else if (!strncmp(argv[0], "buildid-list", 12))
1583 		return __cmd_buildid_list(file_name, argc, argv);
1584 #ifdef HAVE_KVM_STAT_SUPPORT
1585 	else if (!strncmp(argv[0], "stat", 4))
1586 		return kvm_cmd_stat(file_name, argc, argv);
1587 #endif
1588 	else
1589 		usage_with_options(kvm_usage, kvm_options);
1590 
1591 	return 0;
1592 }
1593