Home | History | Annotate | Download | only in perf
      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 <lk/debugfs.h>
     17 #include "util/tool.h"
     18 #include "util/stat.h"
     19 #include "util/top.h"
     20 
     21 #include <sys/prctl.h>
     22 #include <sys/timerfd.h>
     23 
     24 #include <termios.h>
     25 #include <semaphore.h>
     26 #include <pthread.h>
     27 #include <math.h>
     28 
     29 #if defined(__i386__) || defined(__x86_64__)
     30 #include <asm/svm.h>
     31 #include <asm/vmx.h>
     32 #include <asm/kvm.h>
     33 
     34 struct event_key {
     35 	#define INVALID_KEY     (~0ULL)
     36 	u64 key;
     37 	int info;
     38 };
     39 
     40 struct kvm_event_stats {
     41 	u64 time;
     42 	struct stats stats;
     43 };
     44 
     45 struct kvm_event {
     46 	struct list_head hash_entry;
     47 	struct rb_node rb;
     48 
     49 	struct event_key key;
     50 
     51 	struct kvm_event_stats total;
     52 
     53 	#define DEFAULT_VCPU_NUM 8
     54 	int max_vcpu;
     55 	struct kvm_event_stats *vcpu;
     56 };
     57 
     58 typedef int (*key_cmp_fun)(struct kvm_event*, struct kvm_event*, int);
     59 
     60 struct kvm_event_key {
     61 	const char *name;
     62 	key_cmp_fun key;
     63 };
     64 
     65 
     66 struct perf_kvm_stat;
     67 
     68 struct kvm_events_ops {
     69 	bool (*is_begin_event)(struct perf_evsel *evsel,
     70 			       struct perf_sample *sample,
     71 			       struct event_key *key);
     72 	bool (*is_end_event)(struct perf_evsel *evsel,
     73 			     struct perf_sample *sample, struct event_key *key);
     74 	void (*decode_key)(struct perf_kvm_stat *kvm, struct event_key *key,
     75 			   char decode[20]);
     76 	const char *name;
     77 };
     78 
     79 struct exit_reasons_table {
     80 	unsigned long exit_code;
     81 	const char *reason;
     82 };
     83 
     84 #define EVENTS_BITS		12
     85 #define EVENTS_CACHE_SIZE	(1UL << EVENTS_BITS)
     86 
     87 struct perf_kvm_stat {
     88 	struct perf_tool    tool;
     89 	struct perf_record_opts opts;
     90 	struct perf_evlist  *evlist;
     91 	struct perf_session *session;
     92 
     93 	const char *file_name;
     94 	const char *report_event;
     95 	const char *sort_key;
     96 	int trace_vcpu;
     97 
     98 	struct exit_reasons_table *exit_reasons;
     99 	int exit_reasons_size;
    100 	const char *exit_reasons_isa;
    101 
    102 	struct kvm_events_ops *events_ops;
    103 	key_cmp_fun compare;
    104 	struct list_head kvm_events_cache[EVENTS_CACHE_SIZE];
    105 
    106 	u64 total_time;
    107 	u64 total_count;
    108 	u64 lost_events;
    109 	u64 duration;
    110 
    111 	const char *pid_str;
    112 	struct intlist *pid_list;
    113 
    114 	struct rb_root result;
    115 
    116 	int timerfd;
    117 	unsigned int display_time;
    118 	bool live;
    119 };
    120 
    121 
    122 static void exit_event_get_key(struct perf_evsel *evsel,
    123 			       struct perf_sample *sample,
    124 			       struct event_key *key)
    125 {
    126 	key->info = 0;
    127 	key->key = perf_evsel__intval(evsel, sample, "exit_reason");
    128 }
    129 
    130 static bool kvm_exit_event(struct perf_evsel *evsel)
    131 {
    132 	return !strcmp(evsel->name, "kvm:kvm_exit");
    133 }
    134 
    135 static bool exit_event_begin(struct perf_evsel *evsel,
    136 			     struct perf_sample *sample, struct event_key *key)
    137 {
    138 	if (kvm_exit_event(evsel)) {
    139 		exit_event_get_key(evsel, sample, key);
    140 		return true;
    141 	}
    142 
    143 	return false;
    144 }
    145 
    146 static bool kvm_entry_event(struct perf_evsel *evsel)
    147 {
    148 	return !strcmp(evsel->name, "kvm:kvm_entry");
    149 }
    150 
    151 static bool exit_event_end(struct perf_evsel *evsel,
    152 			   struct perf_sample *sample __maybe_unused,
    153 			   struct event_key *key __maybe_unused)
    154 {
    155 	return kvm_entry_event(evsel);
    156 }
    157 
    158 static struct exit_reasons_table vmx_exit_reasons[] = {
    159 	VMX_EXIT_REASONS
    160 };
    161 
    162 static struct exit_reasons_table svm_exit_reasons[] = {
    163 	SVM_EXIT_REASONS
    164 };
    165 
    166 static const char *get_exit_reason(struct perf_kvm_stat *kvm, u64 exit_code)
    167 {
    168 	int i = kvm->exit_reasons_size;
    169 	struct exit_reasons_table *tbl = kvm->exit_reasons;
    170 
    171 	while (i--) {
    172 		if (tbl->exit_code == exit_code)
    173 			return tbl->reason;
    174 		tbl++;
    175 	}
    176 
    177 	pr_err("unknown kvm exit code:%lld on %s\n",
    178 		(unsigned long long)exit_code, kvm->exit_reasons_isa);
    179 	return "UNKNOWN";
    180 }
    181 
    182 static void exit_event_decode_key(struct perf_kvm_stat *kvm,
    183 				  struct event_key *key,
    184 				  char decode[20])
    185 {
    186 	const char *exit_reason = get_exit_reason(kvm, key->key);
    187 
    188 	scnprintf(decode, 20, "%s", exit_reason);
    189 }
    190 
    191 static struct kvm_events_ops exit_events = {
    192 	.is_begin_event = exit_event_begin,
    193 	.is_end_event = exit_event_end,
    194 	.decode_key = exit_event_decode_key,
    195 	.name = "VM-EXIT"
    196 };
    197 
    198 /*
    199  * For the mmio events, we treat:
    200  * the time of MMIO write: kvm_mmio(KVM_TRACE_MMIO_WRITE...) -> kvm_entry
    201  * the time of MMIO read: kvm_exit -> kvm_mmio(KVM_TRACE_MMIO_READ...).
    202  */
    203 static void mmio_event_get_key(struct perf_evsel *evsel, struct perf_sample *sample,
    204 			       struct event_key *key)
    205 {
    206 	key->key  = perf_evsel__intval(evsel, sample, "gpa");
    207 	key->info = perf_evsel__intval(evsel, sample, "type");
    208 }
    209 
    210 #define KVM_TRACE_MMIO_READ_UNSATISFIED 0
    211 #define KVM_TRACE_MMIO_READ 1
    212 #define KVM_TRACE_MMIO_WRITE 2
    213 
    214 static bool mmio_event_begin(struct perf_evsel *evsel,
    215 			     struct perf_sample *sample, struct event_key *key)
    216 {
    217 	/* MMIO read begin event in kernel. */
    218 	if (kvm_exit_event(evsel))
    219 		return true;
    220 
    221 	/* MMIO write begin event in kernel. */
    222 	if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
    223 	    perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_WRITE) {
    224 		mmio_event_get_key(evsel, sample, key);
    225 		return true;
    226 	}
    227 
    228 	return false;
    229 }
    230 
    231 static bool mmio_event_end(struct perf_evsel *evsel, struct perf_sample *sample,
    232 			   struct event_key *key)
    233 {
    234 	/* MMIO write end event in kernel. */
    235 	if (kvm_entry_event(evsel))
    236 		return true;
    237 
    238 	/* MMIO read end event in kernel.*/
    239 	if (!strcmp(evsel->name, "kvm:kvm_mmio") &&
    240 	    perf_evsel__intval(evsel, sample, "type") == KVM_TRACE_MMIO_READ) {
    241 		mmio_event_get_key(evsel, sample, key);
    242 		return true;
    243 	}
    244 
    245 	return false;
    246 }
    247 
    248 static void mmio_event_decode_key(struct perf_kvm_stat *kvm __maybe_unused,
    249 				  struct event_key *key,
    250 				  char decode[20])
    251 {
    252 	scnprintf(decode, 20, "%#lx:%s", (unsigned long)key->key,
    253 				key->info == KVM_TRACE_MMIO_WRITE ? "W" : "R");
    254 }
    255 
    256 static struct kvm_events_ops mmio_events = {
    257 	.is_begin_event = mmio_event_begin,
    258 	.is_end_event = mmio_event_end,
    259 	.decode_key = mmio_event_decode_key,
    260 	.name = "MMIO Access"
    261 };
    262 
    263  /* The time of emulation pio access is from kvm_pio to kvm_entry. */
    264 static void ioport_event_get_key(struct perf_evsel *evsel,
    265 				 struct perf_sample *sample,
    266 				 struct event_key *key)
    267 {
    268 	key->key  = perf_evsel__intval(evsel, sample, "port");
    269 	key->info = perf_evsel__intval(evsel, sample, "rw");
    270 }
    271 
    272 static bool ioport_event_begin(struct perf_evsel *evsel,
    273 			       struct perf_sample *sample,
    274 			       struct event_key *key)
    275 {
    276 	if (!strcmp(evsel->name, "kvm:kvm_pio")) {
    277 		ioport_event_get_key(evsel, sample, key);
    278 		return true;
    279 	}
    280 
    281 	return false;
    282 }
    283 
    284 static bool ioport_event_end(struct perf_evsel *evsel,
    285 			     struct perf_sample *sample __maybe_unused,
    286 			     struct event_key *key __maybe_unused)
    287 {
    288 	return kvm_entry_event(evsel);
    289 }
    290 
    291 static void ioport_event_decode_key(struct perf_kvm_stat *kvm __maybe_unused,
    292 				    struct event_key *key,
    293 				    char decode[20])
    294 {
    295 	scnprintf(decode, 20, "%#llx:%s", (unsigned long long)key->key,
    296 				key->info ? "POUT" : "PIN");
    297 }
    298 
    299 static struct kvm_events_ops ioport_events = {
    300 	.is_begin_event = ioport_event_begin,
    301 	.is_end_event = ioport_event_end,
    302 	.decode_key = ioport_event_decode_key,
    303 	.name = "IO Port Access"
    304 };
    305 
    306 static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
    307 {
    308 	bool ret = true;
    309 
    310 	if (!strcmp(kvm->report_event, "vmexit"))
    311 		kvm->events_ops = &exit_events;
    312 	else if (!strcmp(kvm->report_event, "mmio"))
    313 		kvm->events_ops = &mmio_events;
    314 	else if (!strcmp(kvm->report_event, "ioport"))
    315 		kvm->events_ops = &ioport_events;
    316 	else {
    317 		pr_err("Unknown report event:%s\n", kvm->report_event);
    318 		ret = false;
    319 	}
    320 
    321 	return ret;
    322 }
    323 
    324 struct vcpu_event_record {
    325 	int vcpu_id;
    326 	u64 start_time;
    327 	struct kvm_event *last_event;
    328 };
    329 
    330 
    331 static void init_kvm_event_record(struct perf_kvm_stat *kvm)
    332 {
    333 	unsigned int i;
    334 
    335 	for (i = 0; i < EVENTS_CACHE_SIZE; i++)
    336 		INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
    337 }
    338 
    339 static void clear_events_cache_stats(struct list_head *kvm_events_cache)
    340 {
    341 	struct list_head *head;
    342 	struct kvm_event *event;
    343 	unsigned int i;
    344 	int j;
    345 
    346 	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
    347 		head = &kvm_events_cache[i];
    348 		list_for_each_entry(event, head, hash_entry) {
    349 			/* reset stats for event */
    350 			event->total.time = 0;
    351 			init_stats(&event->total.stats);
    352 
    353 			for (j = 0; j < event->max_vcpu; ++j) {
    354 				event->vcpu[j].time = 0;
    355 				init_stats(&event->vcpu[j].stats);
    356 			}
    357 		}
    358 	}
    359 }
    360 
    361 static int kvm_events_hash_fn(u64 key)
    362 {
    363 	return key & (EVENTS_CACHE_SIZE - 1);
    364 }
    365 
    366 static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
    367 {
    368 	int old_max_vcpu = event->max_vcpu;
    369 	void *prev;
    370 
    371 	if (vcpu_id < event->max_vcpu)
    372 		return true;
    373 
    374 	while (event->max_vcpu <= vcpu_id)
    375 		event->max_vcpu += DEFAULT_VCPU_NUM;
    376 
    377 	prev = event->vcpu;
    378 	event->vcpu = realloc(event->vcpu,
    379 			      event->max_vcpu * sizeof(*event->vcpu));
    380 	if (!event->vcpu) {
    381 		free(prev);
    382 		pr_err("Not enough memory\n");
    383 		return false;
    384 	}
    385 
    386 	memset(event->vcpu + old_max_vcpu, 0,
    387 	       (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
    388 	return true;
    389 }
    390 
    391 static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
    392 {
    393 	struct kvm_event *event;
    394 
    395 	event = zalloc(sizeof(*event));
    396 	if (!event) {
    397 		pr_err("Not enough memory\n");
    398 		return NULL;
    399 	}
    400 
    401 	event->key = *key;
    402 	return event;
    403 }
    404 
    405 static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
    406 					       struct event_key *key)
    407 {
    408 	struct kvm_event *event;
    409 	struct list_head *head;
    410 
    411 	BUG_ON(key->key == INVALID_KEY);
    412 
    413 	head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
    414 	list_for_each_entry(event, head, hash_entry) {
    415 		if (event->key.key == key->key && event->key.info == key->info)
    416 			return event;
    417 	}
    418 
    419 	event = kvm_alloc_init_event(key);
    420 	if (!event)
    421 		return NULL;
    422 
    423 	list_add(&event->hash_entry, head);
    424 	return event;
    425 }
    426 
    427 static bool handle_begin_event(struct perf_kvm_stat *kvm,
    428 			       struct vcpu_event_record *vcpu_record,
    429 			       struct event_key *key, u64 timestamp)
    430 {
    431 	struct kvm_event *event = NULL;
    432 
    433 	if (key->key != INVALID_KEY)
    434 		event = find_create_kvm_event(kvm, key);
    435 
    436 	vcpu_record->last_event = event;
    437 	vcpu_record->start_time = timestamp;
    438 	return true;
    439 }
    440 
    441 static void
    442 kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
    443 {
    444 	kvm_stats->time += time_diff;
    445 	update_stats(&kvm_stats->stats, time_diff);
    446 }
    447 
    448 static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
    449 {
    450 	struct kvm_event_stats *kvm_stats = &event->total;
    451 
    452 	if (vcpu_id != -1)
    453 		kvm_stats = &event->vcpu[vcpu_id];
    454 
    455 	return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
    456 				avg_stats(&kvm_stats->stats));
    457 }
    458 
    459 static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
    460 			     u64 time_diff)
    461 {
    462 	if (vcpu_id == -1) {
    463 		kvm_update_event_stats(&event->total, time_diff);
    464 		return true;
    465 	}
    466 
    467 	if (!kvm_event_expand(event, vcpu_id))
    468 		return false;
    469 
    470 	kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
    471 	return true;
    472 }
    473 
    474 static bool handle_end_event(struct perf_kvm_stat *kvm,
    475 			     struct vcpu_event_record *vcpu_record,
    476 			     struct event_key *key,
    477 			     struct perf_sample *sample)
    478 {
    479 	struct kvm_event *event;
    480 	u64 time_begin, time_diff;
    481 	int vcpu;
    482 
    483 	if (kvm->trace_vcpu == -1)
    484 		vcpu = -1;
    485 	else
    486 		vcpu = vcpu_record->vcpu_id;
    487 
    488 	event = vcpu_record->last_event;
    489 	time_begin = vcpu_record->start_time;
    490 
    491 	/* The begin event is not caught. */
    492 	if (!time_begin)
    493 		return true;
    494 
    495 	/*
    496 	 * In some case, the 'begin event' only records the start timestamp,
    497 	 * the actual event is recognized in the 'end event' (e.g. mmio-event).
    498 	 */
    499 
    500 	/* Both begin and end events did not get the key. */
    501 	if (!event && key->key == INVALID_KEY)
    502 		return true;
    503 
    504 	if (!event)
    505 		event = find_create_kvm_event(kvm, key);
    506 
    507 	if (!event)
    508 		return false;
    509 
    510 	vcpu_record->last_event = NULL;
    511 	vcpu_record->start_time = 0;
    512 
    513 	/* seems to happen once in a while during live mode */
    514 	if (sample->time < time_begin) {
    515 		pr_debug("End time before begin time; skipping event.\n");
    516 		return true;
    517 	}
    518 
    519 	time_diff = sample->time - time_begin;
    520 
    521 	if (kvm->duration && time_diff > kvm->duration) {
    522 		char decode[32];
    523 
    524 		kvm->events_ops->decode_key(kvm, &event->key, decode);
    525 		if (strcmp(decode, "HLT")) {
    526 			pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
    527 				 sample->time, sample->pid, vcpu_record->vcpu_id,
    528 				 decode, time_diff/1000);
    529 		}
    530 	}
    531 
    532 	return update_kvm_event(event, vcpu, time_diff);
    533 }
    534 
    535 static
    536 struct vcpu_event_record *per_vcpu_record(struct thread *thread,
    537 					  struct perf_evsel *evsel,
    538 					  struct perf_sample *sample)
    539 {
    540 	/* Only kvm_entry records vcpu id. */
    541 	if (!thread->priv && kvm_entry_event(evsel)) {
    542 		struct vcpu_event_record *vcpu_record;
    543 
    544 		vcpu_record = zalloc(sizeof(*vcpu_record));
    545 		if (!vcpu_record) {
    546 			pr_err("%s: Not enough memory\n", __func__);
    547 			return NULL;
    548 		}
    549 
    550 		vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample, "vcpu_id");
    551 		thread->priv = vcpu_record;
    552 	}
    553 
    554 	return thread->priv;
    555 }
    556 
    557 static bool handle_kvm_event(struct perf_kvm_stat *kvm,
    558 			     struct thread *thread,
    559 			     struct perf_evsel *evsel,
    560 			     struct perf_sample *sample)
    561 {
    562 	struct vcpu_event_record *vcpu_record;
    563 	struct event_key key = {.key = INVALID_KEY};
    564 
    565 	vcpu_record = per_vcpu_record(thread, evsel, sample);
    566 	if (!vcpu_record)
    567 		return true;
    568 
    569 	/* only process events for vcpus user cares about */
    570 	if ((kvm->trace_vcpu != -1) &&
    571 	    (kvm->trace_vcpu != vcpu_record->vcpu_id))
    572 		return true;
    573 
    574 	if (kvm->events_ops->is_begin_event(evsel, sample, &key))
    575 		return handle_begin_event(kvm, vcpu_record, &key, sample->time);
    576 
    577 	if (kvm->events_ops->is_end_event(evsel, sample, &key))
    578 		return handle_end_event(kvm, vcpu_record, &key, sample);
    579 
    580 	return true;
    581 }
    582 
    583 #define GET_EVENT_KEY(func, field)					\
    584 static u64 get_event_ ##func(struct kvm_event *event, int vcpu)		\
    585 {									\
    586 	if (vcpu == -1)							\
    587 		return event->total.field;				\
    588 									\
    589 	if (vcpu >= event->max_vcpu)					\
    590 		return 0;						\
    591 									\
    592 	return event->vcpu[vcpu].field;					\
    593 }
    594 
    595 #define COMPARE_EVENT_KEY(func, field)					\
    596 GET_EVENT_KEY(func, field)						\
    597 static int compare_kvm_event_ ## func(struct kvm_event *one,		\
    598 					struct kvm_event *two, int vcpu)\
    599 {									\
    600 	return get_event_ ##func(one, vcpu) >				\
    601 				get_event_ ##func(two, vcpu);		\
    602 }
    603 
    604 GET_EVENT_KEY(time, time);
    605 COMPARE_EVENT_KEY(count, stats.n);
    606 COMPARE_EVENT_KEY(mean, stats.mean);
    607 GET_EVENT_KEY(max, stats.max);
    608 GET_EVENT_KEY(min, stats.min);
    609 
    610 #define DEF_SORT_NAME_KEY(name, compare_key)				\
    611 	{ #name, compare_kvm_event_ ## compare_key }
    612 
    613 static struct kvm_event_key keys[] = {
    614 	DEF_SORT_NAME_KEY(sample, count),
    615 	DEF_SORT_NAME_KEY(time, mean),
    616 	{ NULL, NULL }
    617 };
    618 
    619 static bool select_key(struct perf_kvm_stat *kvm)
    620 {
    621 	int i;
    622 
    623 	for (i = 0; keys[i].name; i++) {
    624 		if (!strcmp(keys[i].name, kvm->sort_key)) {
    625 			kvm->compare = keys[i].key;
    626 			return true;
    627 		}
    628 	}
    629 
    630 	pr_err("Unknown compare key:%s\n", kvm->sort_key);
    631 	return false;
    632 }
    633 
    634 static void insert_to_result(struct rb_root *result, struct kvm_event *event,
    635 			     key_cmp_fun bigger, int vcpu)
    636 {
    637 	struct rb_node **rb = &result->rb_node;
    638 	struct rb_node *parent = NULL;
    639 	struct kvm_event *p;
    640 
    641 	while (*rb) {
    642 		p = container_of(*rb, struct kvm_event, rb);
    643 		parent = *rb;
    644 
    645 		if (bigger(event, p, vcpu))
    646 			rb = &(*rb)->rb_left;
    647 		else
    648 			rb = &(*rb)->rb_right;
    649 	}
    650 
    651 	rb_link_node(&event->rb, parent, rb);
    652 	rb_insert_color(&event->rb, result);
    653 }
    654 
    655 static void
    656 update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
    657 {
    658 	int vcpu = kvm->trace_vcpu;
    659 
    660 	kvm->total_count += get_event_count(event, vcpu);
    661 	kvm->total_time += get_event_time(event, vcpu);
    662 }
    663 
    664 static bool event_is_valid(struct kvm_event *event, int vcpu)
    665 {
    666 	return !!get_event_count(event, vcpu);
    667 }
    668 
    669 static void sort_result(struct perf_kvm_stat *kvm)
    670 {
    671 	unsigned int i;
    672 	int vcpu = kvm->trace_vcpu;
    673 	struct kvm_event *event;
    674 
    675 	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
    676 		list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
    677 			if (event_is_valid(event, vcpu)) {
    678 				update_total_count(kvm, event);
    679 				insert_to_result(&kvm->result, event,
    680 						 kvm->compare, vcpu);
    681 			}
    682 		}
    683 	}
    684 }
    685 
    686 /* returns left most element of result, and erase it */
    687 static struct kvm_event *pop_from_result(struct rb_root *result)
    688 {
    689 	struct rb_node *node = rb_first(result);
    690 
    691 	if (!node)
    692 		return NULL;
    693 
    694 	rb_erase(node, result);
    695 	return container_of(node, struct kvm_event, rb);
    696 }
    697 
    698 static void print_vcpu_info(struct perf_kvm_stat *kvm)
    699 {
    700 	int vcpu = kvm->trace_vcpu;
    701 
    702 	pr_info("Analyze events for ");
    703 
    704 	if (kvm->live) {
    705 		if (kvm->opts.target.system_wide)
    706 			pr_info("all VMs, ");
    707 		else if (kvm->opts.target.pid)
    708 			pr_info("pid(s) %s, ", kvm->opts.target.pid);
    709 		else
    710 			pr_info("dazed and confused on what is monitored, ");
    711 	}
    712 
    713 	if (vcpu == -1)
    714 		pr_info("all VCPUs:\n\n");
    715 	else
    716 		pr_info("VCPU %d:\n\n", vcpu);
    717 }
    718 
    719 static void show_timeofday(void)
    720 {
    721 	char date[64];
    722 	struct timeval tv;
    723 	struct tm ltime;
    724 
    725 	gettimeofday(&tv, NULL);
    726 	if (localtime_r(&tv.tv_sec, &ltime)) {
    727 		strftime(date, sizeof(date), "%H:%M:%S", &ltime);
    728 		pr_info("%s.%06ld", date, tv.tv_usec);
    729 	} else
    730 		pr_info("00:00:00.000000");
    731 
    732 	return;
    733 }
    734 
    735 static void print_result(struct perf_kvm_stat *kvm)
    736 {
    737 	char decode[20];
    738 	struct kvm_event *event;
    739 	int vcpu = kvm->trace_vcpu;
    740 
    741 	if (kvm->live) {
    742 		puts(CONSOLE_CLEAR);
    743 		show_timeofday();
    744 	}
    745 
    746 	pr_info("\n\n");
    747 	print_vcpu_info(kvm);
    748 	pr_info("%20s ", kvm->events_ops->name);
    749 	pr_info("%10s ", "Samples");
    750 	pr_info("%9s ", "Samples%");
    751 
    752 	pr_info("%9s ", "Time%");
    753 	pr_info("%10s ", "Min Time");
    754 	pr_info("%10s ", "Max Time");
    755 	pr_info("%16s ", "Avg time");
    756 	pr_info("\n\n");
    757 
    758 	while ((event = pop_from_result(&kvm->result))) {
    759 		u64 ecount, etime, max, min;
    760 
    761 		ecount = get_event_count(event, vcpu);
    762 		etime = get_event_time(event, vcpu);
    763 		max = get_event_max(event, vcpu);
    764 		min = get_event_min(event, vcpu);
    765 
    766 		kvm->events_ops->decode_key(kvm, &event->key, decode);
    767 		pr_info("%20s ", decode);
    768 		pr_info("%10llu ", (unsigned long long)ecount);
    769 		pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
    770 		pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
    771 		pr_info("%8" PRIu64 "us ", min / 1000);
    772 		pr_info("%8" PRIu64 "us ", max / 1000);
    773 		pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount/1e3,
    774 			kvm_event_rel_stddev(vcpu, event));
    775 		pr_info("\n");
    776 	}
    777 
    778 	pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
    779 		kvm->total_count, kvm->total_time / 1e3);
    780 
    781 	if (kvm->lost_events)
    782 		pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
    783 }
    784 
    785 static int process_lost_event(struct perf_tool *tool,
    786 			      union perf_event *event __maybe_unused,
    787 			      struct perf_sample *sample __maybe_unused,
    788 			      struct machine *machine __maybe_unused)
    789 {
    790 	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
    791 
    792 	kvm->lost_events++;
    793 	return 0;
    794 }
    795 
    796 static bool skip_sample(struct perf_kvm_stat *kvm,
    797 			struct perf_sample *sample)
    798 {
    799 	if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
    800 		return true;
    801 
    802 	return false;
    803 }
    804 
    805 static int process_sample_event(struct perf_tool *tool,
    806 				union perf_event *event,
    807 				struct perf_sample *sample,
    808 				struct perf_evsel *evsel,
    809 				struct machine *machine)
    810 {
    811 	struct thread *thread;
    812 	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
    813 						 tool);
    814 
    815 	if (skip_sample(kvm, sample))
    816 		return 0;
    817 
    818 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
    819 	if (thread == NULL) {
    820 		pr_debug("problem processing %d event, skipping it.\n",
    821 			event->header.type);
    822 		return -1;
    823 	}
    824 
    825 	if (!handle_kvm_event(kvm, thread, evsel, sample))
    826 		return -1;
    827 
    828 	return 0;
    829 }
    830 
    831 static int cpu_isa_config(struct perf_kvm_stat *kvm)
    832 {
    833 	char buf[64], *cpuid;
    834 	int err, isa;
    835 
    836 	if (kvm->live) {
    837 		err = get_cpuid(buf, sizeof(buf));
    838 		if (err != 0) {
    839 			pr_err("Failed to look up CPU type (Intel or AMD)\n");
    840 			return err;
    841 		}
    842 		cpuid = buf;
    843 	} else
    844 		cpuid = kvm->session->header.env.cpuid;
    845 
    846 	if (strstr(cpuid, "Intel"))
    847 		isa = 1;
    848 	else if (strstr(cpuid, "AMD"))
    849 		isa = 0;
    850 	else {
    851 		pr_err("CPU %s is not supported.\n", cpuid);
    852 		return -ENOTSUP;
    853 	}
    854 
    855 	if (isa == 1) {
    856 		kvm->exit_reasons = vmx_exit_reasons;
    857 		kvm->exit_reasons_size = ARRAY_SIZE(vmx_exit_reasons);
    858 		kvm->exit_reasons_isa = "VMX";
    859 	}
    860 
    861 	return 0;
    862 }
    863 
    864 static bool verify_vcpu(int vcpu)
    865 {
    866 	if (vcpu != -1 && vcpu < 0) {
    867 		pr_err("Invalid vcpu:%d.\n", vcpu);
    868 		return false;
    869 	}
    870 
    871 	return true;
    872 }
    873 
    874 /* keeping the max events to a modest level to keep
    875  * the processing of samples per mmap smooth.
    876  */
    877 #define PERF_KVM__MAX_EVENTS_PER_MMAP  25
    878 
    879 static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
    880 				   u64 *mmap_time)
    881 {
    882 	union perf_event *event;
    883 	struct perf_sample sample;
    884 	s64 n = 0;
    885 	int err;
    886 
    887 	*mmap_time = ULLONG_MAX;
    888 	while ((event = perf_evlist__mmap_read(kvm->evlist, idx)) != NULL) {
    889 		err = perf_evlist__parse_sample(kvm->evlist, event, &sample);
    890 		if (err) {
    891 			perf_evlist__mmap_consume(kvm->evlist, idx);
    892 			pr_err("Failed to parse sample\n");
    893 			return -1;
    894 		}
    895 
    896 		err = perf_session_queue_event(kvm->session, event, &sample, 0);
    897 		/*
    898 		 * FIXME: Here we can't consume the event, as perf_session_queue_event will
    899 		 *        point to it, and it'll get possibly overwritten by the kernel.
    900 		 */
    901 		perf_evlist__mmap_consume(kvm->evlist, idx);
    902 
    903 		if (err) {
    904 			pr_err("Failed to enqueue sample: %d\n", err);
    905 			return -1;
    906 		}
    907 
    908 		/* save time stamp of our first sample for this mmap */
    909 		if (n == 0)
    910 			*mmap_time = sample.time;
    911 
    912 		/* limit events per mmap handled all at once */
    913 		n++;
    914 		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
    915 			break;
    916 	}
    917 
    918 	return n;
    919 }
    920 
    921 static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
    922 {
    923 	int i, err, throttled = 0;
    924 	s64 n, ntotal = 0;
    925 	u64 flush_time = ULLONG_MAX, mmap_time;
    926 
    927 	for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
    928 		n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
    929 		if (n < 0)
    930 			return -1;
    931 
    932 		/* flush time is going to be the minimum of all the individual
    933 		 * mmap times. Essentially, we flush all the samples queued up
    934 		 * from the last pass under our minimal start time -- that leaves
    935 		 * a very small race for samples to come in with a lower timestamp.
    936 		 * The ioctl to return the perf_clock timestamp should close the
    937 		 * race entirely.
    938 		 */
    939 		if (mmap_time < flush_time)
    940 			flush_time = mmap_time;
    941 
    942 		ntotal += n;
    943 		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
    944 			throttled = 1;
    945 	}
    946 
    947 	/* flush queue after each round in which we processed events */
    948 	if (ntotal) {
    949 		kvm->session->ordered_samples.next_flush = flush_time;
    950 		err = kvm->tool.finished_round(&kvm->tool, NULL, kvm->session);
    951 		if (err) {
    952 			if (kvm->lost_events)
    953 				pr_info("\nLost events: %" PRIu64 "\n\n",
    954 					kvm->lost_events);
    955 			return err;
    956 		}
    957 	}
    958 
    959 	return throttled;
    960 }
    961 
    962 static volatile int done;
    963 
    964 static void sig_handler(int sig __maybe_unused)
    965 {
    966 	done = 1;
    967 }
    968 
    969 static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
    970 {
    971 	struct itimerspec new_value;
    972 	int rc = -1;
    973 
    974 	kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
    975 	if (kvm->timerfd < 0) {
    976 		pr_err("timerfd_create failed\n");
    977 		goto out;
    978 	}
    979 
    980 	new_value.it_value.tv_sec = kvm->display_time;
    981 	new_value.it_value.tv_nsec = 0;
    982 	new_value.it_interval.tv_sec = kvm->display_time;
    983 	new_value.it_interval.tv_nsec = 0;
    984 
    985 	if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
    986 		pr_err("timerfd_settime failed: %d\n", errno);
    987 		close(kvm->timerfd);
    988 		goto out;
    989 	}
    990 
    991 	rc = 0;
    992 out:
    993 	return rc;
    994 }
    995 
    996 static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
    997 {
    998 	uint64_t c;
    999 	int rc;
   1000 
   1001 	rc = read(kvm->timerfd, &c, sizeof(uint64_t));
   1002 	if (rc < 0) {
   1003 		if (errno == EAGAIN)
   1004 			return 0;
   1005 
   1006 		pr_err("Failed to read timer fd: %d\n", errno);
   1007 		return -1;
   1008 	}
   1009 
   1010 	if (rc != sizeof(uint64_t)) {
   1011 		pr_err("Error reading timer fd - invalid size returned\n");
   1012 		return -1;
   1013 	}
   1014 
   1015 	if (c != 1)
   1016 		pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
   1017 
   1018 	/* update display */
   1019 	sort_result(kvm);
   1020 	print_result(kvm);
   1021 
   1022 	/* reset counts */
   1023 	clear_events_cache_stats(kvm->kvm_events_cache);
   1024 	kvm->total_count = 0;
   1025 	kvm->total_time = 0;
   1026 	kvm->lost_events = 0;
   1027 
   1028 	return 0;
   1029 }
   1030 
   1031 static int fd_set_nonblock(int fd)
   1032 {
   1033 	long arg = 0;
   1034 
   1035 	arg = fcntl(fd, F_GETFL);
   1036 	if (arg < 0) {
   1037 		pr_err("Failed to get current flags for fd %d\n", fd);
   1038 		return -1;
   1039 	}
   1040 
   1041 	if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
   1042 		pr_err("Failed to set non-block option on fd %d\n", fd);
   1043 		return -1;
   1044 	}
   1045 
   1046 	return 0;
   1047 }
   1048 
   1049 static
   1050 int perf_kvm__handle_stdin(struct termios *tc_now, struct termios *tc_save)
   1051 {
   1052 	int c;
   1053 
   1054 	tcsetattr(0, TCSANOW, tc_now);
   1055 	c = getc(stdin);
   1056 	tcsetattr(0, TCSAFLUSH, tc_save);
   1057 
   1058 	if (c == 'q')
   1059 		return 1;
   1060 
   1061 	return 0;
   1062 }
   1063 
   1064 static int kvm_events_live_report(struct perf_kvm_stat *kvm)
   1065 {
   1066 	struct pollfd *pollfds = NULL;
   1067 	int nr_fds, nr_stdin, ret, err = -EINVAL;
   1068 	struct termios tc, save;
   1069 
   1070 	/* live flag must be set first */
   1071 	kvm->live = true;
   1072 
   1073 	ret = cpu_isa_config(kvm);
   1074 	if (ret < 0)
   1075 		return ret;
   1076 
   1077 	if (!verify_vcpu(kvm->trace_vcpu) ||
   1078 	    !select_key(kvm) ||
   1079 	    !register_kvm_events_ops(kvm)) {
   1080 		goto out;
   1081 	}
   1082 
   1083 	init_kvm_event_record(kvm);
   1084 
   1085 	tcgetattr(0, &save);
   1086 	tc = save;
   1087 	tc.c_lflag &= ~(ICANON | ECHO);
   1088 	tc.c_cc[VMIN] = 0;
   1089 	tc.c_cc[VTIME] = 0;
   1090 
   1091 	signal(SIGINT, sig_handler);
   1092 	signal(SIGTERM, sig_handler);
   1093 
   1094 	/* copy pollfds -- need to add timerfd and stdin */
   1095 	nr_fds = kvm->evlist->nr_fds;
   1096 	pollfds = zalloc(sizeof(struct pollfd) * (nr_fds + 2));
   1097 	if (!pollfds) {
   1098 		err = -ENOMEM;
   1099 		goto out;
   1100 	}
   1101 	memcpy(pollfds, kvm->evlist->pollfd,
   1102 		sizeof(struct pollfd) * kvm->evlist->nr_fds);
   1103 
   1104 	/* add timer fd */
   1105 	if (perf_kvm__timerfd_create(kvm) < 0) {
   1106 		err = -1;
   1107 		goto out;
   1108 	}
   1109 
   1110 	pollfds[nr_fds].fd = kvm->timerfd;
   1111 	pollfds[nr_fds].events = POLLIN;
   1112 	nr_fds++;
   1113 
   1114 	pollfds[nr_fds].fd = fileno(stdin);
   1115 	pollfds[nr_fds].events = POLLIN;
   1116 	nr_stdin = nr_fds;
   1117 	nr_fds++;
   1118 	if (fd_set_nonblock(fileno(stdin)) != 0)
   1119 		goto out;
   1120 
   1121 	/* everything is good - enable the events and process */
   1122 	perf_evlist__enable(kvm->evlist);
   1123 
   1124 	while (!done) {
   1125 		int rc;
   1126 
   1127 		rc = perf_kvm__mmap_read(kvm);
   1128 		if (rc < 0)
   1129 			break;
   1130 
   1131 		err = perf_kvm__handle_timerfd(kvm);
   1132 		if (err)
   1133 			goto out;
   1134 
   1135 		if (pollfds[nr_stdin].revents & POLLIN)
   1136 			done = perf_kvm__handle_stdin(&tc, &save);
   1137 
   1138 		if (!rc && !done)
   1139 			err = poll(pollfds, nr_fds, 100);
   1140 	}
   1141 
   1142 	perf_evlist__disable(kvm->evlist);
   1143 
   1144 	if (err == 0) {
   1145 		sort_result(kvm);
   1146 		print_result(kvm);
   1147 	}
   1148 
   1149 out:
   1150 	if (kvm->timerfd >= 0)
   1151 		close(kvm->timerfd);
   1152 
   1153 	if (pollfds)
   1154 		free(pollfds);
   1155 
   1156 	return err;
   1157 }
   1158 
   1159 static int kvm_live_open_events(struct perf_kvm_stat *kvm)
   1160 {
   1161 	int err, rc = -1;
   1162 	struct perf_evsel *pos;
   1163 	struct perf_evlist *evlist = kvm->evlist;
   1164 
   1165 	perf_evlist__config(evlist, &kvm->opts);
   1166 
   1167 	/*
   1168 	 * Note: exclude_{guest,host} do not apply here.
   1169 	 *       This command processes KVM tracepoints from host only
   1170 	 */
   1171 	list_for_each_entry(pos, &evlist->entries, node) {
   1172 		struct perf_event_attr *attr = &pos->attr;
   1173 
   1174 		/* make sure these *are* set */
   1175 		perf_evsel__set_sample_bit(pos, TID);
   1176 		perf_evsel__set_sample_bit(pos, TIME);
   1177 		perf_evsel__set_sample_bit(pos, CPU);
   1178 		perf_evsel__set_sample_bit(pos, RAW);
   1179 		/* make sure these are *not*; want as small a sample as possible */
   1180 		perf_evsel__reset_sample_bit(pos, PERIOD);
   1181 		perf_evsel__reset_sample_bit(pos, IP);
   1182 		perf_evsel__reset_sample_bit(pos, CALLCHAIN);
   1183 		perf_evsel__reset_sample_bit(pos, ADDR);
   1184 		perf_evsel__reset_sample_bit(pos, READ);
   1185 		attr->mmap = 0;
   1186 		attr->comm = 0;
   1187 		attr->task = 0;
   1188 
   1189 		attr->sample_period = 1;
   1190 
   1191 		attr->watermark = 0;
   1192 		attr->wakeup_events = 1000;
   1193 
   1194 		/* will enable all once we are ready */
   1195 		attr->disabled = 1;
   1196 	}
   1197 
   1198 	err = perf_evlist__open(evlist);
   1199 	if (err < 0) {
   1200 		printf("Couldn't create the events: %s\n", strerror(errno));
   1201 		goto out;
   1202 	}
   1203 
   1204 	if (perf_evlist__mmap(evlist, kvm->opts.mmap_pages, false) < 0) {
   1205 		ui__error("Failed to mmap the events: %s\n", strerror(errno));
   1206 		perf_evlist__close(evlist);
   1207 		goto out;
   1208 	}
   1209 
   1210 	rc = 0;
   1211 
   1212 out:
   1213 	return rc;
   1214 }
   1215 
   1216 static int read_events(struct perf_kvm_stat *kvm)
   1217 {
   1218 	int ret;
   1219 
   1220 	struct perf_tool eops = {
   1221 		.sample			= process_sample_event,
   1222 		.comm			= perf_event__process_comm,
   1223 		.ordered_samples	= true,
   1224 	};
   1225 
   1226 	kvm->tool = eops;
   1227 	kvm->session = perf_session__new(kvm->file_name, O_RDONLY, 0, false,
   1228 					 &kvm->tool);
   1229 	if (!kvm->session) {
   1230 		pr_err("Initializing perf session failed\n");
   1231 		return -EINVAL;
   1232 	}
   1233 
   1234 	if (!perf_session__has_traces(kvm->session, "kvm record"))
   1235 		return -EINVAL;
   1236 
   1237 	/*
   1238 	 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
   1239 	 * traced in the old kernel.
   1240 	 */
   1241 	ret = cpu_isa_config(kvm);
   1242 	if (ret < 0)
   1243 		return ret;
   1244 
   1245 	return perf_session__process_events(kvm->session, &kvm->tool);
   1246 }
   1247 
   1248 static int parse_target_str(struct perf_kvm_stat *kvm)
   1249 {
   1250 	if (kvm->pid_str) {
   1251 		kvm->pid_list = intlist__new(kvm->pid_str);
   1252 		if (kvm->pid_list == NULL) {
   1253 			pr_err("Error parsing process id string\n");
   1254 			return -EINVAL;
   1255 		}
   1256 	}
   1257 
   1258 	return 0;
   1259 }
   1260 
   1261 static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
   1262 {
   1263 	int ret = -EINVAL;
   1264 	int vcpu = kvm->trace_vcpu;
   1265 
   1266 	if (parse_target_str(kvm) != 0)
   1267 		goto exit;
   1268 
   1269 	if (!verify_vcpu(vcpu))
   1270 		goto exit;
   1271 
   1272 	if (!select_key(kvm))
   1273 		goto exit;
   1274 
   1275 	if (!register_kvm_events_ops(kvm))
   1276 		goto exit;
   1277 
   1278 	init_kvm_event_record(kvm);
   1279 	setup_pager();
   1280 
   1281 	ret = read_events(kvm);
   1282 	if (ret)
   1283 		goto exit;
   1284 
   1285 	sort_result(kvm);
   1286 	print_result(kvm);
   1287 
   1288 exit:
   1289 	return ret;
   1290 }
   1291 
   1292 static const char * const kvm_events_tp[] = {
   1293 	"kvm:kvm_entry",
   1294 	"kvm:kvm_exit",
   1295 	"kvm:kvm_mmio",
   1296 	"kvm:kvm_pio",
   1297 };
   1298 
   1299 #define STRDUP_FAIL_EXIT(s)		\
   1300 	({	char *_p;		\
   1301 	_p = strdup(s);		\
   1302 		if (!_p)		\
   1303 			return -ENOMEM;	\
   1304 		_p;			\
   1305 	})
   1306 
   1307 static int
   1308 kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
   1309 {
   1310 	unsigned int rec_argc, i, j;
   1311 	const char **rec_argv;
   1312 	const char * const record_args[] = {
   1313 		"record",
   1314 		"-R",
   1315 		"-m", "1024",
   1316 		"-c", "1",
   1317 	};
   1318 
   1319 	rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
   1320 		   2 * ARRAY_SIZE(kvm_events_tp);
   1321 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
   1322 
   1323 	if (rec_argv == NULL)
   1324 		return -ENOMEM;
   1325 
   1326 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
   1327 		rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
   1328 
   1329 	for (j = 0; j < ARRAY_SIZE(kvm_events_tp); j++) {
   1330 		rec_argv[i++] = "-e";
   1331 		rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
   1332 	}
   1333 
   1334 	rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
   1335 	rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
   1336 
   1337 	for (j = 1; j < (unsigned int)argc; j++, i++)
   1338 		rec_argv[i] = argv[j];
   1339 
   1340 	return cmd_record(i, rec_argv, NULL);
   1341 }
   1342 
   1343 static int
   1344 kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
   1345 {
   1346 	const struct option kvm_events_report_options[] = {
   1347 		OPT_STRING(0, "event", &kvm->report_event, "report event",
   1348 			    "event for reporting: vmexit, mmio, ioport"),
   1349 		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
   1350 			    "vcpu id to report"),
   1351 		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
   1352 			    "key for sorting: sample(sort by samples number)"
   1353 			    " time (sort by avg time)"),
   1354 		OPT_STRING('p', "pid", &kvm->pid_str, "pid",
   1355 			   "analyze events only for given process id(s)"),
   1356 		OPT_END()
   1357 	};
   1358 
   1359 	const char * const kvm_events_report_usage[] = {
   1360 		"perf kvm stat report [<options>]",
   1361 		NULL
   1362 	};
   1363 
   1364 	symbol__init();
   1365 
   1366 	if (argc) {
   1367 		argc = parse_options(argc, argv,
   1368 				     kvm_events_report_options,
   1369 				     kvm_events_report_usage, 0);
   1370 		if (argc)
   1371 			usage_with_options(kvm_events_report_usage,
   1372 					   kvm_events_report_options);
   1373 	}
   1374 
   1375 	return kvm_events_report_vcpu(kvm);
   1376 }
   1377 
   1378 static struct perf_evlist *kvm_live_event_list(void)
   1379 {
   1380 	struct perf_evlist *evlist;
   1381 	char *tp, *name, *sys;
   1382 	unsigned int j;
   1383 	int err = -1;
   1384 
   1385 	evlist = perf_evlist__new();
   1386 	if (evlist == NULL)
   1387 		return NULL;
   1388 
   1389 	for (j = 0; j < ARRAY_SIZE(kvm_events_tp); j++) {
   1390 
   1391 		tp = strdup(kvm_events_tp[j]);
   1392 		if (tp == NULL)
   1393 			goto out;
   1394 
   1395 		/* split tracepoint into subsystem and name */
   1396 		sys = tp;
   1397 		name = strchr(tp, ':');
   1398 		if (name == NULL) {
   1399 			pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
   1400 				kvm_events_tp[j]);
   1401 			free(tp);
   1402 			goto out;
   1403 		}
   1404 		*name = '\0';
   1405 		name++;
   1406 
   1407 		if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
   1408 			pr_err("Failed to add %s tracepoint to the list\n", kvm_events_tp[j]);
   1409 			free(tp);
   1410 			goto out;
   1411 		}
   1412 
   1413 		free(tp);
   1414 	}
   1415 
   1416 	err = 0;
   1417 
   1418 out:
   1419 	if (err) {
   1420 		perf_evlist__delete(evlist);
   1421 		evlist = NULL;
   1422 	}
   1423 
   1424 	return evlist;
   1425 }
   1426 
   1427 static int kvm_events_live(struct perf_kvm_stat *kvm,
   1428 			   int argc, const char **argv)
   1429 {
   1430 	char errbuf[BUFSIZ];
   1431 	int err;
   1432 
   1433 	const struct option live_options[] = {
   1434 		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
   1435 			"record events on existing process id"),
   1436 		OPT_UINTEGER('m', "mmap-pages", &kvm->opts.mmap_pages,
   1437 			"number of mmap data pages"),
   1438 		OPT_INCR('v', "verbose", &verbose,
   1439 			"be more verbose (show counter open errors, etc)"),
   1440 		OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
   1441 			"system-wide collection from all CPUs"),
   1442 		OPT_UINTEGER('d', "display", &kvm->display_time,
   1443 			"time in seconds between display updates"),
   1444 		OPT_STRING(0, "event", &kvm->report_event, "report event",
   1445 			"event for reporting: vmexit, mmio, ioport"),
   1446 		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
   1447 			"vcpu id to report"),
   1448 		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
   1449 			"key for sorting: sample(sort by samples number)"
   1450 			" time (sort by avg time)"),
   1451 		OPT_U64(0, "duration", &kvm->duration,
   1452 		    "show events other than HALT that take longer than duration usecs"),
   1453 		OPT_END()
   1454 	};
   1455 	const char * const live_usage[] = {
   1456 		"perf kvm stat live [<options>]",
   1457 		NULL
   1458 	};
   1459 
   1460 
   1461 	/* event handling */
   1462 	kvm->tool.sample = process_sample_event;
   1463 	kvm->tool.comm   = perf_event__process_comm;
   1464 	kvm->tool.exit   = perf_event__process_exit;
   1465 	kvm->tool.fork   = perf_event__process_fork;
   1466 	kvm->tool.lost   = process_lost_event;
   1467 	kvm->tool.ordered_samples = true;
   1468 	perf_tool__fill_defaults(&kvm->tool);
   1469 
   1470 	/* set defaults */
   1471 	kvm->display_time = 1;
   1472 	kvm->opts.user_interval = 1;
   1473 	kvm->opts.mmap_pages = 512;
   1474 	kvm->opts.target.uses_mmap = false;
   1475 	kvm->opts.target.uid_str = NULL;
   1476 	kvm->opts.target.uid = UINT_MAX;
   1477 
   1478 	symbol__init();
   1479 	disable_buildid_cache();
   1480 
   1481 	use_browser = 0;
   1482 	setup_browser(false);
   1483 
   1484 	if (argc) {
   1485 		argc = parse_options(argc, argv, live_options,
   1486 				     live_usage, 0);
   1487 		if (argc)
   1488 			usage_with_options(live_usage, live_options);
   1489 	}
   1490 
   1491 	kvm->duration *= NSEC_PER_USEC;   /* convert usec to nsec */
   1492 
   1493 	/*
   1494 	 * target related setups
   1495 	 */
   1496 	err = perf_target__validate(&kvm->opts.target);
   1497 	if (err) {
   1498 		perf_target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
   1499 		ui__warning("%s", errbuf);
   1500 	}
   1501 
   1502 	if (perf_target__none(&kvm->opts.target))
   1503 		kvm->opts.target.system_wide = true;
   1504 
   1505 
   1506 	/*
   1507 	 * generate the event list
   1508 	 */
   1509 	kvm->evlist = kvm_live_event_list();
   1510 	if (kvm->evlist == NULL) {
   1511 		err = -1;
   1512 		goto out;
   1513 	}
   1514 
   1515 	symbol_conf.nr_events = kvm->evlist->nr_entries;
   1516 
   1517 	if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
   1518 		usage_with_options(live_usage, live_options);
   1519 
   1520 	/*
   1521 	 * perf session
   1522 	 */
   1523 	kvm->session = perf_session__new(NULL, O_WRONLY, false, false, &kvm->tool);
   1524 	if (kvm->session == NULL) {
   1525 		err = -ENOMEM;
   1526 		goto out;
   1527 	}
   1528 	kvm->session->evlist = kvm->evlist;
   1529 	perf_session__set_id_hdr_size(kvm->session);
   1530 
   1531 
   1532 	if (perf_target__has_task(&kvm->opts.target))
   1533 		perf_event__synthesize_thread_map(&kvm->tool,
   1534 						  kvm->evlist->threads,
   1535 						  perf_event__process,
   1536 						  &kvm->session->machines.host);
   1537 	else
   1538 		perf_event__synthesize_threads(&kvm->tool, perf_event__process,
   1539 					       &kvm->session->machines.host);
   1540 
   1541 
   1542 	err = kvm_live_open_events(kvm);
   1543 	if (err)
   1544 		goto out;
   1545 
   1546 	err = kvm_events_live_report(kvm);
   1547 
   1548 out:
   1549 	exit_browser(0);
   1550 
   1551 	if (kvm->session)
   1552 		perf_session__delete(kvm->session);
   1553 	kvm->session = NULL;
   1554 	if (kvm->evlist) {
   1555 		perf_evlist__delete_maps(kvm->evlist);
   1556 		perf_evlist__delete(kvm->evlist);
   1557 	}
   1558 
   1559 	return err;
   1560 }
   1561 
   1562 static void print_kvm_stat_usage(void)
   1563 {
   1564 	printf("Usage: perf kvm stat <command>\n\n");
   1565 
   1566 	printf("# Available commands:\n");
   1567 	printf("\trecord: record kvm events\n");
   1568 	printf("\treport: report statistical data of kvm events\n");
   1569 	printf("\tlive:   live reporting of statistical data of kvm events\n");
   1570 
   1571 	printf("\nOtherwise, it is the alias of 'perf stat':\n");
   1572 }
   1573 
   1574 static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
   1575 {
   1576 	struct perf_kvm_stat kvm = {
   1577 		.file_name = file_name,
   1578 
   1579 		.trace_vcpu	= -1,
   1580 		.report_event	= "vmexit",
   1581 		.sort_key	= "sample",
   1582 
   1583 		.exit_reasons = svm_exit_reasons,
   1584 		.exit_reasons_size = ARRAY_SIZE(svm_exit_reasons),
   1585 		.exit_reasons_isa = "SVM",
   1586 	};
   1587 
   1588 	if (argc == 1) {
   1589 		print_kvm_stat_usage();
   1590 		goto perf_stat;
   1591 	}
   1592 
   1593 	if (!strncmp(argv[1], "rec", 3))
   1594 		return kvm_events_record(&kvm, argc - 1, argv + 1);
   1595 
   1596 	if (!strncmp(argv[1], "rep", 3))
   1597 		return kvm_events_report(&kvm, argc - 1 , argv + 1);
   1598 
   1599 	if (!strncmp(argv[1], "live", 4))
   1600 		return kvm_events_live(&kvm, argc - 1 , argv + 1);
   1601 
   1602 perf_stat:
   1603 	return cmd_stat(argc, argv, NULL);
   1604 }
   1605 #endif
   1606 
   1607 static int __cmd_record(const char *file_name, int argc, const char **argv)
   1608 {
   1609 	int rec_argc, i = 0, j;
   1610 	const char **rec_argv;
   1611 
   1612 	rec_argc = argc + 2;
   1613 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
   1614 	rec_argv[i++] = strdup("record");
   1615 	rec_argv[i++] = strdup("-o");
   1616 	rec_argv[i++] = strdup(file_name);
   1617 	for (j = 1; j < argc; j++, i++)
   1618 		rec_argv[i] = argv[j];
   1619 
   1620 	BUG_ON(i != rec_argc);
   1621 
   1622 	return cmd_record(i, rec_argv, NULL);
   1623 }
   1624 
   1625 static int __cmd_report(const char *file_name, int argc, const char **argv)
   1626 {
   1627 	int rec_argc, i = 0, j;
   1628 	const char **rec_argv;
   1629 
   1630 	rec_argc = argc + 2;
   1631 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
   1632 	rec_argv[i++] = strdup("report");
   1633 	rec_argv[i++] = strdup("-i");
   1634 	rec_argv[i++] = strdup(file_name);
   1635 	for (j = 1; j < argc; j++, i++)
   1636 		rec_argv[i] = argv[j];
   1637 
   1638 	BUG_ON(i != rec_argc);
   1639 
   1640 	return cmd_report(i, rec_argv, NULL);
   1641 }
   1642 
   1643 static int
   1644 __cmd_buildid_list(const char *file_name, int argc, const char **argv)
   1645 {
   1646 	int rec_argc, i = 0, j;
   1647 	const char **rec_argv;
   1648 
   1649 	rec_argc = argc + 2;
   1650 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
   1651 	rec_argv[i++] = strdup("buildid-list");
   1652 	rec_argv[i++] = strdup("-i");
   1653 	rec_argv[i++] = strdup(file_name);
   1654 	for (j = 1; j < argc; j++, i++)
   1655 		rec_argv[i] = argv[j];
   1656 
   1657 	BUG_ON(i != rec_argc);
   1658 
   1659 	return cmd_buildid_list(i, rec_argv, NULL);
   1660 }
   1661 
   1662 int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
   1663 {
   1664 	const char *file_name = NULL;
   1665 	const struct option kvm_options[] = {
   1666 		OPT_STRING('i', "input", &file_name, "file",
   1667 			   "Input file name"),
   1668 		OPT_STRING('o', "output", &file_name, "file",
   1669 			   "Output file name"),
   1670 		OPT_BOOLEAN(0, "guest", &perf_guest,
   1671 			    "Collect guest os data"),
   1672 		OPT_BOOLEAN(0, "host", &perf_host,
   1673 			    "Collect host os data"),
   1674 		OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
   1675 			   "guest mount directory under which every guest os"
   1676 			   " instance has a subdir"),
   1677 		OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
   1678 			   "file", "file saving guest os vmlinux"),
   1679 		OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
   1680 			   "file", "file saving guest os /proc/kallsyms"),
   1681 		OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
   1682 			   "file", "file saving guest os /proc/modules"),
   1683 		OPT_END()
   1684 	};
   1685 
   1686 
   1687 	const char * const kvm_usage[] = {
   1688 		"perf kvm [<options>] {top|record|report|diff|buildid-list|stat}",
   1689 		NULL
   1690 	};
   1691 
   1692 	perf_host  = 0;
   1693 	perf_guest = 1;
   1694 
   1695 	argc = parse_options(argc, argv, kvm_options, kvm_usage,
   1696 			PARSE_OPT_STOP_AT_NON_OPTION);
   1697 	if (!argc)
   1698 		usage_with_options(kvm_usage, kvm_options);
   1699 
   1700 	if (!perf_host)
   1701 		perf_guest = 1;
   1702 
   1703 	if (!file_name) {
   1704 		if (perf_host && !perf_guest)
   1705 			file_name = strdup("perf.data.host");
   1706 		else if (!perf_host && perf_guest)
   1707 			file_name = strdup("perf.data.guest");
   1708 		else
   1709 			file_name = strdup("perf.data.kvm");
   1710 
   1711 		if (!file_name) {
   1712 			pr_err("Failed to allocate memory for filename\n");
   1713 			return -ENOMEM;
   1714 		}
   1715 	}
   1716 
   1717 	if (!strncmp(argv[0], "rec", 3))
   1718 		return __cmd_record(file_name, argc, argv);
   1719 	else if (!strncmp(argv[0], "rep", 3))
   1720 		return __cmd_report(file_name, argc, argv);
   1721 	else if (!strncmp(argv[0], "diff", 4))
   1722 		return cmd_diff(argc, argv, NULL);
   1723 	else if (!strncmp(argv[0], "top", 3))
   1724 		return cmd_top(argc, argv, NULL);
   1725 	else if (!strncmp(argv[0], "buildid-list", 12))
   1726 		return __cmd_buildid_list(file_name, argc, argv);
   1727 #if defined(__i386__) || defined(__x86_64__)
   1728 	else if (!strncmp(argv[0], "stat", 4))
   1729 		return kvm_cmd_stat(file_name, argc, argv);
   1730 #endif
   1731 	else
   1732 		usage_with_options(kvm_usage, kvm_options);
   1733 
   1734 	return 0;
   1735 }
   1736