Linux Audio

Check our new training course

Yocto / OpenEmbedded training

Mar 24-27, 2025, special US time zones
Register
Loading...
v4.6
   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 <subcmd/parse-options.h>
  14#include "util/trace-event.h"
  15#include "util/debug.h"
  16#include "util/tool.h"
  17#include "util/stat.h"
  18#include "util/top.h"
  19#include "util/data.h"
  20#include "util/ordered-events.h"
  21
  22#include <sys/prctl.h>
  23#ifdef HAVE_TIMERFD_SUPPORT
  24#include <sys/timerfd.h>
  25#endif
  26
 
  27#include <termios.h>
  28#include <semaphore.h>
  29#include <pthread.h>
  30#include <math.h>
  31
  32#ifdef HAVE_KVM_STAT_SUPPORT
  33#include "util/kvm-stat.h"
  34
  35void exit_event_get_key(struct perf_evsel *evsel,
  36			struct perf_sample *sample,
  37			struct event_key *key)
  38{
  39	key->info = 0;
  40	key->key = perf_evsel__intval(evsel, sample, kvm_exit_reason);
  41}
  42
  43bool kvm_exit_event(struct perf_evsel *evsel)
  44{
  45	return !strcmp(evsel->name, kvm_exit_trace);
  46}
  47
  48bool exit_event_begin(struct perf_evsel *evsel,
  49		      struct perf_sample *sample, struct event_key *key)
  50{
  51	if (kvm_exit_event(evsel)) {
  52		exit_event_get_key(evsel, sample, key);
  53		return true;
  54	}
  55
  56	return false;
  57}
  58
  59bool kvm_entry_event(struct perf_evsel *evsel)
  60{
  61	return !strcmp(evsel->name, kvm_entry_trace);
  62}
  63
  64bool exit_event_end(struct perf_evsel *evsel,
  65		    struct perf_sample *sample __maybe_unused,
  66		    struct event_key *key __maybe_unused)
  67{
  68	return kvm_entry_event(evsel);
  69}
  70
  71static const char *get_exit_reason(struct perf_kvm_stat *kvm,
  72				   struct exit_reasons_table *tbl,
  73				   u64 exit_code)
  74{
  75	while (tbl->reason != NULL) {
  76		if (tbl->exit_code == exit_code)
  77			return tbl->reason;
  78		tbl++;
  79	}
  80
  81	pr_err("unknown kvm exit code:%lld on %s\n",
  82		(unsigned long long)exit_code, kvm->exit_reasons_isa);
  83	return "UNKNOWN";
  84}
  85
  86void exit_event_decode_key(struct perf_kvm_stat *kvm,
  87			   struct event_key *key,
  88			   char *decode)
  89{
  90	const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
  91						  key->key);
  92
  93	scnprintf(decode, decode_str_len, "%s", exit_reason);
  94}
  95
  96static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
  97{
  98	struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
  99
 100	for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
 101		if (!strcmp(events_ops->name, kvm->report_event)) {
 102			kvm->events_ops = events_ops->ops;
 103			return true;
 104		}
 105	}
 106
 107	return false;
 108}
 109
 110struct vcpu_event_record {
 111	int vcpu_id;
 112	u64 start_time;
 113	struct kvm_event *last_event;
 114};
 115
 116
 117static void init_kvm_event_record(struct perf_kvm_stat *kvm)
 118{
 119	unsigned int i;
 120
 121	for (i = 0; i < EVENTS_CACHE_SIZE; i++)
 122		INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
 123}
 124
 125#ifdef HAVE_TIMERFD_SUPPORT
 126static void clear_events_cache_stats(struct list_head *kvm_events_cache)
 127{
 128	struct list_head *head;
 129	struct kvm_event *event;
 130	unsigned int i;
 131	int j;
 132
 133	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
 134		head = &kvm_events_cache[i];
 135		list_for_each_entry(event, head, hash_entry) {
 136			/* reset stats for event */
 137			event->total.time = 0;
 138			init_stats(&event->total.stats);
 139
 140			for (j = 0; j < event->max_vcpu; ++j) {
 141				event->vcpu[j].time = 0;
 142				init_stats(&event->vcpu[j].stats);
 143			}
 144		}
 145	}
 146}
 147#endif
 148
 149static int kvm_events_hash_fn(u64 key)
 150{
 151	return key & (EVENTS_CACHE_SIZE - 1);
 152}
 153
 154static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
 155{
 156	int old_max_vcpu = event->max_vcpu;
 157	void *prev;
 158
 159	if (vcpu_id < event->max_vcpu)
 160		return true;
 161
 162	while (event->max_vcpu <= vcpu_id)
 163		event->max_vcpu += DEFAULT_VCPU_NUM;
 164
 165	prev = event->vcpu;
 166	event->vcpu = realloc(event->vcpu,
 167			      event->max_vcpu * sizeof(*event->vcpu));
 168	if (!event->vcpu) {
 169		free(prev);
 170		pr_err("Not enough memory\n");
 171		return false;
 172	}
 173
 174	memset(event->vcpu + old_max_vcpu, 0,
 175	       (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
 176	return true;
 177}
 178
 179static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
 180{
 181	struct kvm_event *event;
 182
 183	event = zalloc(sizeof(*event));
 184	if (!event) {
 185		pr_err("Not enough memory\n");
 186		return NULL;
 187	}
 188
 189	event->key = *key;
 190	init_stats(&event->total.stats);
 191	return event;
 192}
 193
 194static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
 195					       struct event_key *key)
 196{
 197	struct kvm_event *event;
 198	struct list_head *head;
 199
 200	BUG_ON(key->key == INVALID_KEY);
 201
 202	head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
 203	list_for_each_entry(event, head, hash_entry) {
 204		if (event->key.key == key->key && event->key.info == key->info)
 205			return event;
 206	}
 207
 208	event = kvm_alloc_init_event(key);
 209	if (!event)
 210		return NULL;
 211
 212	list_add(&event->hash_entry, head);
 213	return event;
 214}
 215
 216static bool handle_begin_event(struct perf_kvm_stat *kvm,
 217			       struct vcpu_event_record *vcpu_record,
 218			       struct event_key *key, u64 timestamp)
 219{
 220	struct kvm_event *event = NULL;
 221
 222	if (key->key != INVALID_KEY)
 223		event = find_create_kvm_event(kvm, key);
 224
 225	vcpu_record->last_event = event;
 226	vcpu_record->start_time = timestamp;
 227	return true;
 228}
 229
 230static void
 231kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
 232{
 233	kvm_stats->time += time_diff;
 234	update_stats(&kvm_stats->stats, time_diff);
 235}
 236
 237static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
 238{
 239	struct kvm_event_stats *kvm_stats = &event->total;
 240
 241	if (vcpu_id != -1)
 242		kvm_stats = &event->vcpu[vcpu_id];
 243
 244	return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
 245				avg_stats(&kvm_stats->stats));
 246}
 247
 248static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
 249			     u64 time_diff)
 250{
 251	if (vcpu_id == -1) {
 252		kvm_update_event_stats(&event->total, time_diff);
 253		return true;
 254	}
 255
 256	if (!kvm_event_expand(event, vcpu_id))
 257		return false;
 258
 259	kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
 260	return true;
 261}
 262
 263static bool is_child_event(struct perf_kvm_stat *kvm,
 264			   struct perf_evsel *evsel,
 265			   struct perf_sample *sample,
 266			   struct event_key *key)
 267{
 268	struct child_event_ops *child_ops;
 269
 270	child_ops = kvm->events_ops->child_ops;
 271
 272	if (!child_ops)
 273		return false;
 274
 275	for (; child_ops->name; child_ops++) {
 276		if (!strcmp(evsel->name, child_ops->name)) {
 277			child_ops->get_key(evsel, sample, key);
 278			return true;
 279		}
 280	}
 281
 282	return false;
 283}
 284
 285static bool handle_child_event(struct perf_kvm_stat *kvm,
 286			       struct vcpu_event_record *vcpu_record,
 287			       struct event_key *key,
 288			       struct perf_sample *sample __maybe_unused)
 289{
 290	struct kvm_event *event = NULL;
 291
 292	if (key->key != INVALID_KEY)
 293		event = find_create_kvm_event(kvm, key);
 294
 295	vcpu_record->last_event = event;
 296
 297	return true;
 298}
 299
 300static bool skip_event(const char *event)
 301{
 302	const char * const *skip_events;
 303
 304	for (skip_events = kvm_skip_events; *skip_events; skip_events++)
 305		if (!strcmp(event, *skip_events))
 306			return true;
 307
 308	return false;
 309}
 310
 311static bool handle_end_event(struct perf_kvm_stat *kvm,
 312			     struct vcpu_event_record *vcpu_record,
 313			     struct event_key *key,
 314			     struct perf_sample *sample)
 315{
 316	struct kvm_event *event;
 317	u64 time_begin, time_diff;
 318	int vcpu;
 319
 320	if (kvm->trace_vcpu == -1)
 321		vcpu = -1;
 322	else
 323		vcpu = vcpu_record->vcpu_id;
 324
 325	event = vcpu_record->last_event;
 326	time_begin = vcpu_record->start_time;
 327
 328	/* The begin event is not caught. */
 329	if (!time_begin)
 330		return true;
 331
 332	/*
 333	 * In some case, the 'begin event' only records the start timestamp,
 334	 * the actual event is recognized in the 'end event' (e.g. mmio-event).
 335	 */
 336
 337	/* Both begin and end events did not get the key. */
 338	if (!event && key->key == INVALID_KEY)
 339		return true;
 340
 341	if (!event)
 342		event = find_create_kvm_event(kvm, key);
 343
 344	if (!event)
 345		return false;
 346
 347	vcpu_record->last_event = NULL;
 348	vcpu_record->start_time = 0;
 349
 350	/* seems to happen once in a while during live mode */
 351	if (sample->time < time_begin) {
 352		pr_debug("End time before begin time; skipping event.\n");
 353		return true;
 354	}
 355
 356	time_diff = sample->time - time_begin;
 357
 358	if (kvm->duration && time_diff > kvm->duration) {
 359		char decode[decode_str_len];
 360
 361		kvm->events_ops->decode_key(kvm, &event->key, decode);
 362		if (!skip_event(decode)) {
 363			pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
 364				 sample->time, sample->pid, vcpu_record->vcpu_id,
 365				 decode, time_diff/1000);
 366		}
 367	}
 368
 369	return update_kvm_event(event, vcpu, time_diff);
 370}
 371
 372static
 373struct vcpu_event_record *per_vcpu_record(struct thread *thread,
 374					  struct perf_evsel *evsel,
 375					  struct perf_sample *sample)
 376{
 377	/* Only kvm_entry records vcpu id. */
 378	if (!thread__priv(thread) && kvm_entry_event(evsel)) {
 379		struct vcpu_event_record *vcpu_record;
 380
 381		vcpu_record = zalloc(sizeof(*vcpu_record));
 382		if (!vcpu_record) {
 383			pr_err("%s: Not enough memory\n", __func__);
 384			return NULL;
 385		}
 386
 387		vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample,
 388							  vcpu_id_str);
 389		thread__set_priv(thread, vcpu_record);
 390	}
 391
 392	return thread__priv(thread);
 393}
 394
 395static bool handle_kvm_event(struct perf_kvm_stat *kvm,
 396			     struct thread *thread,
 397			     struct perf_evsel *evsel,
 398			     struct perf_sample *sample)
 399{
 400	struct vcpu_event_record *vcpu_record;
 401	struct event_key key = { .key = INVALID_KEY,
 402				 .exit_reasons = kvm->exit_reasons };
 403
 404	vcpu_record = per_vcpu_record(thread, evsel, sample);
 405	if (!vcpu_record)
 406		return true;
 407
 408	/* only process events for vcpus user cares about */
 409	if ((kvm->trace_vcpu != -1) &&
 410	    (kvm->trace_vcpu != vcpu_record->vcpu_id))
 411		return true;
 412
 413	if (kvm->events_ops->is_begin_event(evsel, sample, &key))
 414		return handle_begin_event(kvm, vcpu_record, &key, sample->time);
 415
 416	if (is_child_event(kvm, evsel, sample, &key))
 417		return handle_child_event(kvm, vcpu_record, &key, sample);
 418
 419	if (kvm->events_ops->is_end_event(evsel, sample, &key))
 420		return handle_end_event(kvm, vcpu_record, &key, sample);
 421
 422	return true;
 423}
 424
 425#define GET_EVENT_KEY(func, field)					\
 426static u64 get_event_ ##func(struct kvm_event *event, int vcpu)		\
 427{									\
 428	if (vcpu == -1)							\
 429		return event->total.field;				\
 430									\
 431	if (vcpu >= event->max_vcpu)					\
 432		return 0;						\
 433									\
 434	return event->vcpu[vcpu].field;					\
 435}
 436
 437#define COMPARE_EVENT_KEY(func, field)					\
 438GET_EVENT_KEY(func, field)						\
 439static int compare_kvm_event_ ## func(struct kvm_event *one,		\
 440					struct kvm_event *two, int vcpu)\
 441{									\
 442	return get_event_ ##func(one, vcpu) >				\
 443				get_event_ ##func(two, vcpu);		\
 444}
 445
 446GET_EVENT_KEY(time, time);
 447COMPARE_EVENT_KEY(count, stats.n);
 448COMPARE_EVENT_KEY(mean, stats.mean);
 449GET_EVENT_KEY(max, stats.max);
 450GET_EVENT_KEY(min, stats.min);
 451
 452#define DEF_SORT_NAME_KEY(name, compare_key)				\
 453	{ #name, compare_kvm_event_ ## compare_key }
 454
 455static struct kvm_event_key keys[] = {
 456	DEF_SORT_NAME_KEY(sample, count),
 457	DEF_SORT_NAME_KEY(time, mean),
 458	{ NULL, NULL }
 459};
 460
 461static bool select_key(struct perf_kvm_stat *kvm)
 462{
 463	int i;
 464
 465	for (i = 0; keys[i].name; i++) {
 466		if (!strcmp(keys[i].name, kvm->sort_key)) {
 467			kvm->compare = keys[i].key;
 468			return true;
 469		}
 470	}
 471
 472	pr_err("Unknown compare key:%s\n", kvm->sort_key);
 473	return false;
 474}
 475
 476static void insert_to_result(struct rb_root *result, struct kvm_event *event,
 477			     key_cmp_fun bigger, int vcpu)
 478{
 479	struct rb_node **rb = &result->rb_node;
 480	struct rb_node *parent = NULL;
 481	struct kvm_event *p;
 482
 483	while (*rb) {
 484		p = container_of(*rb, struct kvm_event, rb);
 485		parent = *rb;
 486
 487		if (bigger(event, p, vcpu))
 488			rb = &(*rb)->rb_left;
 489		else
 490			rb = &(*rb)->rb_right;
 491	}
 492
 493	rb_link_node(&event->rb, parent, rb);
 494	rb_insert_color(&event->rb, result);
 495}
 496
 497static void
 498update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
 499{
 500	int vcpu = kvm->trace_vcpu;
 501
 502	kvm->total_count += get_event_count(event, vcpu);
 503	kvm->total_time += get_event_time(event, vcpu);
 504}
 505
 506static bool event_is_valid(struct kvm_event *event, int vcpu)
 507{
 508	return !!get_event_count(event, vcpu);
 509}
 510
 511static void sort_result(struct perf_kvm_stat *kvm)
 512{
 513	unsigned int i;
 514	int vcpu = kvm->trace_vcpu;
 515	struct kvm_event *event;
 516
 517	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
 518		list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
 519			if (event_is_valid(event, vcpu)) {
 520				update_total_count(kvm, event);
 521				insert_to_result(&kvm->result, event,
 522						 kvm->compare, vcpu);
 523			}
 524		}
 525	}
 526}
 527
 528/* returns left most element of result, and erase it */
 529static struct kvm_event *pop_from_result(struct rb_root *result)
 530{
 531	struct rb_node *node = rb_first(result);
 532
 533	if (!node)
 534		return NULL;
 535
 536	rb_erase(node, result);
 537	return container_of(node, struct kvm_event, rb);
 538}
 539
 540static void print_vcpu_info(struct perf_kvm_stat *kvm)
 541{
 542	int vcpu = kvm->trace_vcpu;
 543
 544	pr_info("Analyze events for ");
 545
 546	if (kvm->opts.target.system_wide)
 547		pr_info("all VMs, ");
 548	else if (kvm->opts.target.pid)
 549		pr_info("pid(s) %s, ", kvm->opts.target.pid);
 550	else
 551		pr_info("dazed and confused on what is monitored, ");
 552
 553	if (vcpu == -1)
 554		pr_info("all VCPUs:\n\n");
 555	else
 556		pr_info("VCPU %d:\n\n", vcpu);
 557}
 558
 559static void show_timeofday(void)
 560{
 561	char date[64];
 562	struct timeval tv;
 563	struct tm ltime;
 564
 565	gettimeofday(&tv, NULL);
 566	if (localtime_r(&tv.tv_sec, &ltime)) {
 567		strftime(date, sizeof(date), "%H:%M:%S", &ltime);
 568		pr_info("%s.%06ld", date, tv.tv_usec);
 569	} else
 570		pr_info("00:00:00.000000");
 571
 572	return;
 573}
 574
 575static void print_result(struct perf_kvm_stat *kvm)
 576{
 577	char decode[decode_str_len];
 578	struct kvm_event *event;
 579	int vcpu = kvm->trace_vcpu;
 580
 581	if (kvm->live) {
 582		puts(CONSOLE_CLEAR);
 583		show_timeofday();
 584	}
 585
 586	pr_info("\n\n");
 587	print_vcpu_info(kvm);
 588	pr_info("%*s ", decode_str_len, kvm->events_ops->name);
 589	pr_info("%10s ", "Samples");
 590	pr_info("%9s ", "Samples%");
 591
 592	pr_info("%9s ", "Time%");
 593	pr_info("%11s ", "Min Time");
 594	pr_info("%11s ", "Max Time");
 595	pr_info("%16s ", "Avg time");
 596	pr_info("\n\n");
 597
 598	while ((event = pop_from_result(&kvm->result))) {
 599		u64 ecount, etime, max, min;
 600
 601		ecount = get_event_count(event, vcpu);
 602		etime = get_event_time(event, vcpu);
 603		max = get_event_max(event, vcpu);
 604		min = get_event_min(event, vcpu);
 605
 606		kvm->events_ops->decode_key(kvm, &event->key, decode);
 607		pr_info("%*s ", decode_str_len, decode);
 608		pr_info("%10llu ", (unsigned long long)ecount);
 609		pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
 610		pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
 611		pr_info("%9.2fus ", (double)min / 1e3);
 612		pr_info("%9.2fus ", (double)max / 1e3);
 613		pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount/1e3,
 614			kvm_event_rel_stddev(vcpu, event));
 615		pr_info("\n");
 616	}
 617
 618	pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
 619		kvm->total_count, kvm->total_time / 1e3);
 620
 621	if (kvm->lost_events)
 622		pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
 623}
 624
 625#ifdef HAVE_TIMERFD_SUPPORT
 626static int process_lost_event(struct perf_tool *tool,
 627			      union perf_event *event __maybe_unused,
 628			      struct perf_sample *sample __maybe_unused,
 629			      struct machine *machine __maybe_unused)
 630{
 631	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
 632
 633	kvm->lost_events++;
 634	return 0;
 635}
 636#endif
 637
 638static bool skip_sample(struct perf_kvm_stat *kvm,
 639			struct perf_sample *sample)
 640{
 641	if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
 642		return true;
 643
 644	return false;
 645}
 646
 647static int process_sample_event(struct perf_tool *tool,
 648				union perf_event *event,
 649				struct perf_sample *sample,
 650				struct perf_evsel *evsel,
 651				struct machine *machine)
 652{
 653	int err = 0;
 654	struct thread *thread;
 655	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
 656						 tool);
 657
 658	if (skip_sample(kvm, sample))
 659		return 0;
 660
 661	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
 662	if (thread == NULL) {
 663		pr_debug("problem processing %d event, skipping it.\n",
 664			event->header.type);
 665		return -1;
 666	}
 667
 668	if (!handle_kvm_event(kvm, thread, evsel, sample))
 669		err = -1;
 670
 671	thread__put(thread);
 672	return err;
 673}
 674
 675static int cpu_isa_config(struct perf_kvm_stat *kvm)
 676{
 677	char buf[64], *cpuid;
 678	int err;
 679
 680	if (kvm->live) {
 681		err = get_cpuid(buf, sizeof(buf));
 682		if (err != 0) {
 683			pr_err("Failed to look up CPU type\n");
 684			return err;
 685		}
 686		cpuid = buf;
 687	} else
 688		cpuid = kvm->session->header.env.cpuid;
 689
 690	if (!cpuid) {
 691		pr_err("Failed to look up CPU type\n");
 692		return -EINVAL;
 693	}
 694
 695	err = cpu_isa_init(kvm, cpuid);
 696	if (err == -ENOTSUP)
 697		pr_err("CPU %s is not supported.\n", cpuid);
 698
 699	return err;
 700}
 701
 702static bool verify_vcpu(int vcpu)
 703{
 704	if (vcpu != -1 && vcpu < 0) {
 705		pr_err("Invalid vcpu:%d.\n", vcpu);
 706		return false;
 707	}
 708
 709	return true;
 710}
 711
 712#ifdef HAVE_TIMERFD_SUPPORT
 713/* keeping the max events to a modest level to keep
 714 * the processing of samples per mmap smooth.
 715 */
 716#define PERF_KVM__MAX_EVENTS_PER_MMAP  25
 717
 718static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
 719				   u64 *mmap_time)
 720{
 721	union perf_event *event;
 722	struct perf_sample sample;
 723	s64 n = 0;
 724	int err;
 725
 726	*mmap_time = ULLONG_MAX;
 727	while ((event = perf_evlist__mmap_read(kvm->evlist, idx)) != NULL) {
 728		err = perf_evlist__parse_sample(kvm->evlist, event, &sample);
 729		if (err) {
 730			perf_evlist__mmap_consume(kvm->evlist, idx);
 731			pr_err("Failed to parse sample\n");
 732			return -1;
 733		}
 734
 735		err = perf_session__queue_event(kvm->session, event, &sample, 0);
 736		/*
 737		 * FIXME: Here we can't consume the event, as perf_session__queue_event will
 738		 *        point to it, and it'll get possibly overwritten by the kernel.
 739		 */
 740		perf_evlist__mmap_consume(kvm->evlist, idx);
 741
 742		if (err) {
 743			pr_err("Failed to enqueue sample: %d\n", err);
 744			return -1;
 745		}
 746
 747		/* save time stamp of our first sample for this mmap */
 748		if (n == 0)
 749			*mmap_time = sample.time;
 750
 751		/* limit events per mmap handled all at once */
 752		n++;
 753		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
 754			break;
 755	}
 756
 757	return n;
 758}
 759
 760static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
 761{
 762	int i, err, throttled = 0;
 763	s64 n, ntotal = 0;
 764	u64 flush_time = ULLONG_MAX, mmap_time;
 765
 766	for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
 767		n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
 768		if (n < 0)
 769			return -1;
 770
 771		/* flush time is going to be the minimum of all the individual
 772		 * mmap times. Essentially, we flush all the samples queued up
 773		 * from the last pass under our minimal start time -- that leaves
 774		 * a very small race for samples to come in with a lower timestamp.
 775		 * The ioctl to return the perf_clock timestamp should close the
 776		 * race entirely.
 777		 */
 778		if (mmap_time < flush_time)
 779			flush_time = mmap_time;
 780
 781		ntotal += n;
 782		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
 783			throttled = 1;
 784	}
 785
 786	/* flush queue after each round in which we processed events */
 787	if (ntotal) {
 788		struct ordered_events *oe = &kvm->session->ordered_events;
 789
 790		oe->next_flush = flush_time;
 791		err = ordered_events__flush(oe, OE_FLUSH__ROUND);
 792		if (err) {
 793			if (kvm->lost_events)
 794				pr_info("\nLost events: %" PRIu64 "\n\n",
 795					kvm->lost_events);
 796			return err;
 797		}
 798	}
 799
 800	return throttled;
 801}
 802
 803static volatile int done;
 804
 805static void sig_handler(int sig __maybe_unused)
 806{
 807	done = 1;
 808}
 809
 810static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
 811{
 812	struct itimerspec new_value;
 813	int rc = -1;
 814
 815	kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
 816	if (kvm->timerfd < 0) {
 817		pr_err("timerfd_create failed\n");
 818		goto out;
 819	}
 820
 821	new_value.it_value.tv_sec = kvm->display_time;
 822	new_value.it_value.tv_nsec = 0;
 823	new_value.it_interval.tv_sec = kvm->display_time;
 824	new_value.it_interval.tv_nsec = 0;
 825
 826	if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
 827		pr_err("timerfd_settime failed: %d\n", errno);
 828		close(kvm->timerfd);
 829		goto out;
 830	}
 831
 832	rc = 0;
 833out:
 834	return rc;
 835}
 836
 837static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
 838{
 839	uint64_t c;
 840	int rc;
 841
 842	rc = read(kvm->timerfd, &c, sizeof(uint64_t));
 843	if (rc < 0) {
 844		if (errno == EAGAIN)
 845			return 0;
 846
 847		pr_err("Failed to read timer fd: %d\n", errno);
 848		return -1;
 849	}
 850
 851	if (rc != sizeof(uint64_t)) {
 852		pr_err("Error reading timer fd - invalid size returned\n");
 853		return -1;
 854	}
 855
 856	if (c != 1)
 857		pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
 858
 859	/* update display */
 860	sort_result(kvm);
 861	print_result(kvm);
 862
 863	/* reset counts */
 864	clear_events_cache_stats(kvm->kvm_events_cache);
 865	kvm->total_count = 0;
 866	kvm->total_time = 0;
 867	kvm->lost_events = 0;
 868
 869	return 0;
 870}
 871
 872static int fd_set_nonblock(int fd)
 873{
 874	long arg = 0;
 875
 876	arg = fcntl(fd, F_GETFL);
 877	if (arg < 0) {
 878		pr_err("Failed to get current flags for fd %d\n", fd);
 879		return -1;
 880	}
 881
 882	if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
 883		pr_err("Failed to set non-block option on fd %d\n", fd);
 884		return -1;
 885	}
 886
 887	return 0;
 888}
 889
 890static int perf_kvm__handle_stdin(void)
 891{
 892	int c;
 893
 894	c = getc(stdin);
 895	if (c == 'q')
 896		return 1;
 897
 898	return 0;
 899}
 900
 901static int kvm_events_live_report(struct perf_kvm_stat *kvm)
 902{
 903	int nr_stdin, ret, err = -EINVAL;
 904	struct termios save;
 905
 906	/* live flag must be set first */
 907	kvm->live = true;
 908
 909	ret = cpu_isa_config(kvm);
 910	if (ret < 0)
 911		return ret;
 912
 913	if (!verify_vcpu(kvm->trace_vcpu) ||
 914	    !select_key(kvm) ||
 915	    !register_kvm_events_ops(kvm)) {
 916		goto out;
 917	}
 918
 919	set_term_quiet_input(&save);
 920	init_kvm_event_record(kvm);
 921
 922	signal(SIGINT, sig_handler);
 923	signal(SIGTERM, sig_handler);
 924
 925	/* add timer fd */
 926	if (perf_kvm__timerfd_create(kvm) < 0) {
 927		err = -1;
 928		goto out;
 929	}
 930
 931	if (perf_evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
 932		goto out;
 933
 934	nr_stdin = perf_evlist__add_pollfd(kvm->evlist, fileno(stdin));
 935	if (nr_stdin < 0)
 936		goto out;
 937
 938	if (fd_set_nonblock(fileno(stdin)) != 0)
 939		goto out;
 940
 941	/* everything is good - enable the events and process */
 942	perf_evlist__enable(kvm->evlist);
 943
 944	while (!done) {
 945		struct fdarray *fda = &kvm->evlist->pollfd;
 946		int rc;
 947
 948		rc = perf_kvm__mmap_read(kvm);
 949		if (rc < 0)
 950			break;
 951
 952		err = perf_kvm__handle_timerfd(kvm);
 953		if (err)
 954			goto out;
 955
 956		if (fda->entries[nr_stdin].revents & POLLIN)
 957			done = perf_kvm__handle_stdin();
 958
 959		if (!rc && !done)
 960			err = fdarray__poll(fda, 100);
 961	}
 962
 963	perf_evlist__disable(kvm->evlist);
 964
 965	if (err == 0) {
 966		sort_result(kvm);
 967		print_result(kvm);
 968	}
 969
 970out:
 971	if (kvm->timerfd >= 0)
 972		close(kvm->timerfd);
 973
 974	tcsetattr(0, TCSAFLUSH, &save);
 975	return err;
 976}
 977
 978static int kvm_live_open_events(struct perf_kvm_stat *kvm)
 979{
 980	int err, rc = -1;
 981	struct perf_evsel *pos;
 982	struct perf_evlist *evlist = kvm->evlist;
 983	char sbuf[STRERR_BUFSIZE];
 984
 985	perf_evlist__config(evlist, &kvm->opts);
 986
 987	/*
 988	 * Note: exclude_{guest,host} do not apply here.
 989	 *       This command processes KVM tracepoints from host only
 990	 */
 991	evlist__for_each(evlist, pos) {
 992		struct perf_event_attr *attr = &pos->attr;
 993
 994		/* make sure these *are* set */
 995		perf_evsel__set_sample_bit(pos, TID);
 996		perf_evsel__set_sample_bit(pos, TIME);
 997		perf_evsel__set_sample_bit(pos, CPU);
 998		perf_evsel__set_sample_bit(pos, RAW);
 999		/* make sure these are *not*; want as small a sample as possible */
1000		perf_evsel__reset_sample_bit(pos, PERIOD);
1001		perf_evsel__reset_sample_bit(pos, IP);
1002		perf_evsel__reset_sample_bit(pos, CALLCHAIN);
1003		perf_evsel__reset_sample_bit(pos, ADDR);
1004		perf_evsel__reset_sample_bit(pos, READ);
1005		attr->mmap = 0;
1006		attr->comm = 0;
1007		attr->task = 0;
1008
1009		attr->sample_period = 1;
1010
1011		attr->watermark = 0;
1012		attr->wakeup_events = 1000;
1013
1014		/* will enable all once we are ready */
1015		attr->disabled = 1;
1016	}
1017
1018	err = perf_evlist__open(evlist);
1019	if (err < 0) {
1020		printf("Couldn't create the events: %s\n",
1021		       strerror_r(errno, sbuf, sizeof(sbuf)));
1022		goto out;
1023	}
1024
1025	if (perf_evlist__mmap(evlist, kvm->opts.mmap_pages, false) < 0) {
1026		ui__error("Failed to mmap the events: %s\n",
1027			  strerror_r(errno, sbuf, sizeof(sbuf)));
1028		perf_evlist__close(evlist);
1029		goto out;
1030	}
1031
1032	rc = 0;
1033
1034out:
1035	return rc;
1036}
1037#endif
1038
1039static int read_events(struct perf_kvm_stat *kvm)
1040{
1041	int ret;
1042
1043	struct perf_tool eops = {
1044		.sample			= process_sample_event,
1045		.comm			= perf_event__process_comm,
1046		.ordered_events		= true,
1047	};
1048	struct perf_data_file file = {
1049		.path = kvm->file_name,
1050		.mode = PERF_DATA_MODE_READ,
1051		.force = kvm->force,
1052	};
1053
1054	kvm->tool = eops;
1055	kvm->session = perf_session__new(&file, false, &kvm->tool);
1056	if (!kvm->session) {
1057		pr_err("Initializing perf session failed\n");
1058		return -1;
1059	}
1060
1061	symbol__init(&kvm->session->header.env);
1062
1063	if (!perf_session__has_traces(kvm->session, "kvm record")) {
1064		ret = -EINVAL;
1065		goto out_delete;
1066	}
1067
1068	/*
1069	 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1070	 * traced in the old kernel.
1071	 */
1072	ret = cpu_isa_config(kvm);
1073	if (ret < 0)
1074		goto out_delete;
1075
1076	ret = perf_session__process_events(kvm->session);
1077
1078out_delete:
1079	perf_session__delete(kvm->session);
1080	return ret;
1081}
1082
1083static int parse_target_str(struct perf_kvm_stat *kvm)
1084{
1085	if (kvm->opts.target.pid) {
1086		kvm->pid_list = intlist__new(kvm->opts.target.pid);
1087		if (kvm->pid_list == NULL) {
1088			pr_err("Error parsing process id string\n");
1089			return -EINVAL;
1090		}
1091	}
1092
1093	return 0;
1094}
1095
1096static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1097{
1098	int ret = -EINVAL;
1099	int vcpu = kvm->trace_vcpu;
1100
1101	if (parse_target_str(kvm) != 0)
1102		goto exit;
1103
1104	if (!verify_vcpu(vcpu))
1105		goto exit;
1106
1107	if (!select_key(kvm))
1108		goto exit;
1109
1110	if (!register_kvm_events_ops(kvm))
1111		goto exit;
1112
1113	init_kvm_event_record(kvm);
1114	setup_pager();
1115
1116	ret = read_events(kvm);
1117	if (ret)
1118		goto exit;
1119
1120	sort_result(kvm);
1121	print_result(kvm);
1122
1123exit:
1124	return ret;
1125}
1126
1127#define STRDUP_FAIL_EXIT(s)		\
1128	({	char *_p;		\
1129	_p = strdup(s);		\
1130		if (!_p)		\
1131			return -ENOMEM;	\
1132		_p;			\
1133	})
1134
1135int __weak setup_kvm_events_tp(struct perf_kvm_stat *kvm __maybe_unused)
1136{
1137	return 0;
1138}
1139
1140static int
1141kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1142{
1143	unsigned int rec_argc, i, j, events_tp_size;
1144	const char **rec_argv;
1145	const char * const record_args[] = {
1146		"record",
1147		"-R",
1148		"-m", "1024",
1149		"-c", "1",
1150	};
1151	const char * const kvm_stat_record_usage[] = {
1152		"perf kvm stat record [<options>]",
1153		NULL
1154	};
1155	const char * const *events_tp;
1156	int ret;
1157
1158	events_tp_size = 0;
1159	ret = setup_kvm_events_tp(kvm);
1160	if (ret < 0) {
1161		pr_err("Unable to setup the kvm tracepoints\n");
1162		return ret;
1163	}
1164
1165	for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1166		events_tp_size++;
1167
1168	rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1169		   2 * events_tp_size;
1170	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1171
1172	if (rec_argv == NULL)
1173		return -ENOMEM;
1174
1175	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1176		rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1177
1178	for (j = 0; j < events_tp_size; j++) {
1179		rec_argv[i++] = "-e";
1180		rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1181	}
1182
1183	rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1184	rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1185
1186	for (j = 1; j < (unsigned int)argc; j++, i++)
1187		rec_argv[i] = argv[j];
1188
1189	set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1190	set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1191	set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1192
1193	set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1194	set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1195	set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1196	set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1197	set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1198	set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1199	set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1200	set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1201	set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1202	set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1203	set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1204	set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1205	set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1206	set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1207	set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1208
1209	record_usage = kvm_stat_record_usage;
1210	return cmd_record(i, rec_argv, NULL);
1211}
1212
1213static int
1214kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1215{
1216	const struct option kvm_events_report_options[] = {
1217		OPT_STRING(0, "event", &kvm->report_event, "report event",
1218			   "event for reporting: vmexit, "
1219			   "mmio (x86 only), ioport (x86 only)"),
1220		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1221			    "vcpu id to report"),
1222		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1223			    "key for sorting: sample(sort by samples number)"
1224			    " time (sort by avg time)"),
1225		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1226			   "analyze events only for given process id(s)"),
1227		OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
1228		OPT_END()
1229	};
1230
1231	const char * const kvm_events_report_usage[] = {
1232		"perf kvm stat report [<options>]",
1233		NULL
1234	};
1235
1236	if (argc) {
1237		argc = parse_options(argc, argv,
1238				     kvm_events_report_options,
1239				     kvm_events_report_usage, 0);
1240		if (argc)
1241			usage_with_options(kvm_events_report_usage,
1242					   kvm_events_report_options);
1243	}
1244
1245	if (!kvm->opts.target.pid)
1246		kvm->opts.target.system_wide = true;
1247
1248	return kvm_events_report_vcpu(kvm);
1249}
1250
1251#ifdef HAVE_TIMERFD_SUPPORT
1252static struct perf_evlist *kvm_live_event_list(void)
1253{
1254	struct perf_evlist *evlist;
1255	char *tp, *name, *sys;
1256	int err = -1;
1257	const char * const *events_tp;
1258
1259	evlist = perf_evlist__new();
1260	if (evlist == NULL)
1261		return NULL;
1262
1263	for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1264
1265		tp = strdup(*events_tp);
1266		if (tp == NULL)
1267			goto out;
1268
1269		/* split tracepoint into subsystem and name */
1270		sys = tp;
1271		name = strchr(tp, ':');
1272		if (name == NULL) {
1273			pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1274			       *events_tp);
1275			free(tp);
1276			goto out;
1277		}
1278		*name = '\0';
1279		name++;
1280
1281		if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1282			pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1283			free(tp);
1284			goto out;
1285		}
1286
1287		free(tp);
1288	}
1289
1290	err = 0;
1291
1292out:
1293	if (err) {
1294		perf_evlist__delete(evlist);
1295		evlist = NULL;
1296	}
1297
1298	return evlist;
1299}
1300
1301static int kvm_events_live(struct perf_kvm_stat *kvm,
1302			   int argc, const char **argv)
1303{
1304	char errbuf[BUFSIZ];
1305	int err;
1306
1307	const struct option live_options[] = {
1308		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1309			"record events on existing process id"),
1310		OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1311			"number of mmap data pages",
1312			perf_evlist__parse_mmap_pages),
1313		OPT_INCR('v', "verbose", &verbose,
1314			"be more verbose (show counter open errors, etc)"),
1315		OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1316			"system-wide collection from all CPUs"),
1317		OPT_UINTEGER('d', "display", &kvm->display_time,
1318			"time in seconds between display updates"),
1319		OPT_STRING(0, "event", &kvm->report_event, "report event",
1320			"event for reporting: "
1321			"vmexit, mmio (x86 only), ioport (x86 only)"),
1322		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1323			"vcpu id to report"),
1324		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1325			"key for sorting: sample(sort by samples number)"
1326			" time (sort by avg time)"),
1327		OPT_U64(0, "duration", &kvm->duration,
1328			"show events other than"
1329			" HLT (x86 only) or Wait state (s390 only)"
1330			" that take longer than duration usecs"),
1331		OPT_UINTEGER(0, "proc-map-timeout", &kvm->opts.proc_map_timeout,
1332				"per thread proc mmap processing timeout in ms"),
1333		OPT_END()
1334	};
1335	const char * const live_usage[] = {
1336		"perf kvm stat live [<options>]",
1337		NULL
1338	};
1339	struct perf_data_file file = {
1340		.mode = PERF_DATA_MODE_WRITE,
1341	};
1342
1343
1344	/* event handling */
1345	kvm->tool.sample = process_sample_event;
1346	kvm->tool.comm   = perf_event__process_comm;
1347	kvm->tool.exit   = perf_event__process_exit;
1348	kvm->tool.fork   = perf_event__process_fork;
1349	kvm->tool.lost   = process_lost_event;
1350	kvm->tool.ordered_events = true;
1351	perf_tool__fill_defaults(&kvm->tool);
1352
1353	/* set defaults */
1354	kvm->display_time = 1;
1355	kvm->opts.user_interval = 1;
1356	kvm->opts.mmap_pages = 512;
1357	kvm->opts.target.uses_mmap = false;
1358	kvm->opts.target.uid_str = NULL;
1359	kvm->opts.target.uid = UINT_MAX;
1360	kvm->opts.proc_map_timeout = 500;
1361
1362	symbol__init(NULL);
1363	disable_buildid_cache();
1364
1365	use_browser = 0;
1366
1367	if (argc) {
1368		argc = parse_options(argc, argv, live_options,
1369				     live_usage, 0);
1370		if (argc)
1371			usage_with_options(live_usage, live_options);
1372	}
1373
1374	kvm->duration *= NSEC_PER_USEC;   /* convert usec to nsec */
1375
1376	/*
1377	 * target related setups
1378	 */
1379	err = target__validate(&kvm->opts.target);
1380	if (err) {
1381		target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1382		ui__warning("%s", errbuf);
1383	}
1384
1385	if (target__none(&kvm->opts.target))
1386		kvm->opts.target.system_wide = true;
1387
1388
1389	/*
1390	 * generate the event list
1391	 */
1392	err = setup_kvm_events_tp(kvm);
1393	if (err < 0) {
1394		pr_err("Unable to setup the kvm tracepoints\n");
1395		return err;
1396	}
1397
1398	kvm->evlist = kvm_live_event_list();
1399	if (kvm->evlist == NULL) {
1400		err = -1;
1401		goto out;
1402	}
1403
1404	symbol_conf.nr_events = kvm->evlist->nr_entries;
1405
1406	if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1407		usage_with_options(live_usage, live_options);
1408
1409	/*
1410	 * perf session
1411	 */
1412	kvm->session = perf_session__new(&file, false, &kvm->tool);
1413	if (kvm->session == NULL) {
1414		err = -1;
1415		goto out;
1416	}
1417	kvm->session->evlist = kvm->evlist;
1418	perf_session__set_id_hdr_size(kvm->session);
1419	ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
1420	machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1421				    kvm->evlist->threads, false, kvm->opts.proc_map_timeout);
1422	err = kvm_live_open_events(kvm);
1423	if (err)
1424		goto out;
1425
1426	err = kvm_events_live_report(kvm);
1427
1428out:
1429	if (kvm->session)
1430		perf_session__delete(kvm->session);
1431	kvm->session = NULL;
1432	if (kvm->evlist)
1433		perf_evlist__delete(kvm->evlist);
1434
1435	return err;
1436}
1437#endif
1438
1439static void print_kvm_stat_usage(void)
1440{
1441	printf("Usage: perf kvm stat <command>\n\n");
1442
1443	printf("# Available commands:\n");
1444	printf("\trecord: record kvm events\n");
1445	printf("\treport: report statistical data of kvm events\n");
1446	printf("\tlive:   live reporting of statistical data of kvm events\n");
1447
1448	printf("\nOtherwise, it is the alias of 'perf stat':\n");
1449}
1450
1451static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1452{
1453	struct perf_kvm_stat kvm = {
1454		.file_name = file_name,
1455
1456		.trace_vcpu	= -1,
1457		.report_event	= "vmexit",
1458		.sort_key	= "sample",
1459
1460	};
1461
1462	if (argc == 1) {
1463		print_kvm_stat_usage();
1464		goto perf_stat;
1465	}
1466
1467	if (!strncmp(argv[1], "rec", 3))
1468		return kvm_events_record(&kvm, argc - 1, argv + 1);
1469
1470	if (!strncmp(argv[1], "rep", 3))
1471		return kvm_events_report(&kvm, argc - 1 , argv + 1);
1472
1473#ifdef HAVE_TIMERFD_SUPPORT
1474	if (!strncmp(argv[1], "live", 4))
1475		return kvm_events_live(&kvm, argc - 1 , argv + 1);
1476#endif
1477
1478perf_stat:
1479	return cmd_stat(argc, argv, NULL);
1480}
1481#endif /* HAVE_KVM_STAT_SUPPORT */
1482
1483static int __cmd_record(const char *file_name, int argc, const char **argv)
1484{
1485	int rec_argc, i = 0, j;
1486	const char **rec_argv;
1487
1488	rec_argc = argc + 2;
1489	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1490	rec_argv[i++] = strdup("record");
1491	rec_argv[i++] = strdup("-o");
1492	rec_argv[i++] = strdup(file_name);
1493	for (j = 1; j < argc; j++, i++)
1494		rec_argv[i] = argv[j];
1495
1496	BUG_ON(i != rec_argc);
1497
1498	return cmd_record(i, rec_argv, NULL);
1499}
1500
1501static int __cmd_report(const char *file_name, int argc, const char **argv)
1502{
1503	int rec_argc, i = 0, j;
1504	const char **rec_argv;
1505
1506	rec_argc = argc + 2;
1507	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1508	rec_argv[i++] = strdup("report");
1509	rec_argv[i++] = strdup("-i");
1510	rec_argv[i++] = strdup(file_name);
1511	for (j = 1; j < argc; j++, i++)
1512		rec_argv[i] = argv[j];
1513
1514	BUG_ON(i != rec_argc);
1515
1516	return cmd_report(i, rec_argv, NULL);
1517}
1518
1519static int
1520__cmd_buildid_list(const char *file_name, int argc, const char **argv)
1521{
1522	int rec_argc, i = 0, j;
1523	const char **rec_argv;
1524
1525	rec_argc = argc + 2;
1526	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1527	rec_argv[i++] = strdup("buildid-list");
1528	rec_argv[i++] = strdup("-i");
1529	rec_argv[i++] = strdup(file_name);
1530	for (j = 1; j < argc; j++, i++)
1531		rec_argv[i] = argv[j];
1532
1533	BUG_ON(i != rec_argc);
1534
1535	return cmd_buildid_list(i, rec_argv, NULL);
1536}
1537
1538int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
1539{
1540	const char *file_name = NULL;
1541	const struct option kvm_options[] = {
1542		OPT_STRING('i', "input", &file_name, "file",
1543			   "Input file name"),
1544		OPT_STRING('o', "output", &file_name, "file",
1545			   "Output file name"),
1546		OPT_BOOLEAN(0, "guest", &perf_guest,
1547			    "Collect guest os data"),
1548		OPT_BOOLEAN(0, "host", &perf_host,
1549			    "Collect host os data"),
1550		OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1551			   "guest mount directory under which every guest os"
1552			   " instance has a subdir"),
1553		OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1554			   "file", "file saving guest os vmlinux"),
1555		OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1556			   "file", "file saving guest os /proc/kallsyms"),
1557		OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1558			   "file", "file saving guest os /proc/modules"),
1559		OPT_INCR('v', "verbose", &verbose,
1560			    "be more verbose (show counter open errors, etc)"),
1561		OPT_END()
1562	};
1563
1564	const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1565						"buildid-list", "stat", NULL };
1566	const char *kvm_usage[] = { NULL, NULL };
1567
1568	perf_host  = 0;
1569	perf_guest = 1;
1570
1571	argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1572					PARSE_OPT_STOP_AT_NON_OPTION);
1573	if (!argc)
1574		usage_with_options(kvm_usage, kvm_options);
1575
1576	if (!perf_host)
1577		perf_guest = 1;
1578
1579	if (!file_name) {
1580		file_name = get_filename_for_perf_kvm();
1581
1582		if (!file_name) {
1583			pr_err("Failed to allocate memory for filename\n");
1584			return -ENOMEM;
1585		}
1586	}
1587
1588	if (!strncmp(argv[0], "rec", 3))
1589		return __cmd_record(file_name, argc, argv);
1590	else if (!strncmp(argv[0], "rep", 3))
1591		return __cmd_report(file_name, argc, argv);
1592	else if (!strncmp(argv[0], "diff", 4))
1593		return cmd_diff(argc, argv, NULL);
1594	else if (!strncmp(argv[0], "top", 3))
1595		return cmd_top(argc, argv, NULL);
1596	else if (!strncmp(argv[0], "buildid-list", 12))
1597		return __cmd_buildid_list(file_name, argc, argv);
1598#ifdef HAVE_KVM_STAT_SUPPORT
1599	else if (!strncmp(argv[0], "stat", 4))
1600		return kvm_cmd_stat(file_name, argc, argv);
1601#endif
1602	else
1603		usage_with_options(kvm_usage, kvm_options);
1604
1605	return 0;
1606}
v4.10.11
   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 <subcmd/parse-options.h>
  14#include "util/trace-event.h"
  15#include "util/debug.h"
  16#include "util/tool.h"
  17#include "util/stat.h"
  18#include "util/top.h"
  19#include "util/data.h"
  20#include "util/ordered-events.h"
  21
  22#include <sys/prctl.h>
  23#ifdef HAVE_TIMERFD_SUPPORT
  24#include <sys/timerfd.h>
  25#endif
  26
  27#include <linux/time64.h>
  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 "util/kvm-stat.h"
  35
  36void exit_event_get_key(struct perf_evsel *evsel,
  37			struct perf_sample *sample,
  38			struct event_key *key)
  39{
  40	key->info = 0;
  41	key->key = perf_evsel__intval(evsel, sample, kvm_exit_reason);
  42}
  43
  44bool kvm_exit_event(struct perf_evsel *evsel)
  45{
  46	return !strcmp(evsel->name, kvm_exit_trace);
  47}
  48
  49bool exit_event_begin(struct perf_evsel *evsel,
  50		      struct perf_sample *sample, struct event_key *key)
  51{
  52	if (kvm_exit_event(evsel)) {
  53		exit_event_get_key(evsel, sample, key);
  54		return true;
  55	}
  56
  57	return false;
  58}
  59
  60bool kvm_entry_event(struct perf_evsel *evsel)
  61{
  62	return !strcmp(evsel->name, kvm_entry_trace);
  63}
  64
  65bool exit_event_end(struct perf_evsel *evsel,
  66		    struct perf_sample *sample __maybe_unused,
  67		    struct event_key *key __maybe_unused)
  68{
  69	return kvm_entry_event(evsel);
  70}
  71
  72static const char *get_exit_reason(struct perf_kvm_stat *kvm,
  73				   struct exit_reasons_table *tbl,
  74				   u64 exit_code)
  75{
  76	while (tbl->reason != NULL) {
  77		if (tbl->exit_code == exit_code)
  78			return tbl->reason;
  79		tbl++;
  80	}
  81
  82	pr_err("unknown kvm exit code:%lld on %s\n",
  83		(unsigned long long)exit_code, kvm->exit_reasons_isa);
  84	return "UNKNOWN";
  85}
  86
  87void exit_event_decode_key(struct perf_kvm_stat *kvm,
  88			   struct event_key *key,
  89			   char *decode)
  90{
  91	const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
  92						  key->key);
  93
  94	scnprintf(decode, decode_str_len, "%s", exit_reason);
  95}
  96
  97static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
  98{
  99	struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
 100
 101	for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
 102		if (!strcmp(events_ops->name, kvm->report_event)) {
 103			kvm->events_ops = events_ops->ops;
 104			return true;
 105		}
 106	}
 107
 108	return false;
 109}
 110
 111struct vcpu_event_record {
 112	int vcpu_id;
 113	u64 start_time;
 114	struct kvm_event *last_event;
 115};
 116
 117
 118static void init_kvm_event_record(struct perf_kvm_stat *kvm)
 119{
 120	unsigned int i;
 121
 122	for (i = 0; i < EVENTS_CACHE_SIZE; i++)
 123		INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
 124}
 125
 126#ifdef HAVE_TIMERFD_SUPPORT
 127static void clear_events_cache_stats(struct list_head *kvm_events_cache)
 128{
 129	struct list_head *head;
 130	struct kvm_event *event;
 131	unsigned int i;
 132	int j;
 133
 134	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
 135		head = &kvm_events_cache[i];
 136		list_for_each_entry(event, head, hash_entry) {
 137			/* reset stats for event */
 138			event->total.time = 0;
 139			init_stats(&event->total.stats);
 140
 141			for (j = 0; j < event->max_vcpu; ++j) {
 142				event->vcpu[j].time = 0;
 143				init_stats(&event->vcpu[j].stats);
 144			}
 145		}
 146	}
 147}
 148#endif
 149
 150static int kvm_events_hash_fn(u64 key)
 151{
 152	return key & (EVENTS_CACHE_SIZE - 1);
 153}
 154
 155static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
 156{
 157	int old_max_vcpu = event->max_vcpu;
 158	void *prev;
 159
 160	if (vcpu_id < event->max_vcpu)
 161		return true;
 162
 163	while (event->max_vcpu <= vcpu_id)
 164		event->max_vcpu += DEFAULT_VCPU_NUM;
 165
 166	prev = event->vcpu;
 167	event->vcpu = realloc(event->vcpu,
 168			      event->max_vcpu * sizeof(*event->vcpu));
 169	if (!event->vcpu) {
 170		free(prev);
 171		pr_err("Not enough memory\n");
 172		return false;
 173	}
 174
 175	memset(event->vcpu + old_max_vcpu, 0,
 176	       (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
 177	return true;
 178}
 179
 180static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
 181{
 182	struct kvm_event *event;
 183
 184	event = zalloc(sizeof(*event));
 185	if (!event) {
 186		pr_err("Not enough memory\n");
 187		return NULL;
 188	}
 189
 190	event->key = *key;
 191	init_stats(&event->total.stats);
 192	return event;
 193}
 194
 195static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
 196					       struct event_key *key)
 197{
 198	struct kvm_event *event;
 199	struct list_head *head;
 200
 201	BUG_ON(key->key == INVALID_KEY);
 202
 203	head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
 204	list_for_each_entry(event, head, hash_entry) {
 205		if (event->key.key == key->key && event->key.info == key->info)
 206			return event;
 207	}
 208
 209	event = kvm_alloc_init_event(key);
 210	if (!event)
 211		return NULL;
 212
 213	list_add(&event->hash_entry, head);
 214	return event;
 215}
 216
 217static bool handle_begin_event(struct perf_kvm_stat *kvm,
 218			       struct vcpu_event_record *vcpu_record,
 219			       struct event_key *key, u64 timestamp)
 220{
 221	struct kvm_event *event = NULL;
 222
 223	if (key->key != INVALID_KEY)
 224		event = find_create_kvm_event(kvm, key);
 225
 226	vcpu_record->last_event = event;
 227	vcpu_record->start_time = timestamp;
 228	return true;
 229}
 230
 231static void
 232kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
 233{
 234	kvm_stats->time += time_diff;
 235	update_stats(&kvm_stats->stats, time_diff);
 236}
 237
 238static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
 239{
 240	struct kvm_event_stats *kvm_stats = &event->total;
 241
 242	if (vcpu_id != -1)
 243		kvm_stats = &event->vcpu[vcpu_id];
 244
 245	return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
 246				avg_stats(&kvm_stats->stats));
 247}
 248
 249static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
 250			     u64 time_diff)
 251{
 252	if (vcpu_id == -1) {
 253		kvm_update_event_stats(&event->total, time_diff);
 254		return true;
 255	}
 256
 257	if (!kvm_event_expand(event, vcpu_id))
 258		return false;
 259
 260	kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
 261	return true;
 262}
 263
 264static bool is_child_event(struct perf_kvm_stat *kvm,
 265			   struct perf_evsel *evsel,
 266			   struct perf_sample *sample,
 267			   struct event_key *key)
 268{
 269	struct child_event_ops *child_ops;
 270
 271	child_ops = kvm->events_ops->child_ops;
 272
 273	if (!child_ops)
 274		return false;
 275
 276	for (; child_ops->name; child_ops++) {
 277		if (!strcmp(evsel->name, child_ops->name)) {
 278			child_ops->get_key(evsel, sample, key);
 279			return true;
 280		}
 281	}
 282
 283	return false;
 284}
 285
 286static bool handle_child_event(struct perf_kvm_stat *kvm,
 287			       struct vcpu_event_record *vcpu_record,
 288			       struct event_key *key,
 289			       struct perf_sample *sample __maybe_unused)
 290{
 291	struct kvm_event *event = NULL;
 292
 293	if (key->key != INVALID_KEY)
 294		event = find_create_kvm_event(kvm, key);
 295
 296	vcpu_record->last_event = event;
 297
 298	return true;
 299}
 300
 301static bool skip_event(const char *event)
 302{
 303	const char * const *skip_events;
 304
 305	for (skip_events = kvm_skip_events; *skip_events; skip_events++)
 306		if (!strcmp(event, *skip_events))
 307			return true;
 308
 309	return false;
 310}
 311
 312static bool handle_end_event(struct perf_kvm_stat *kvm,
 313			     struct vcpu_event_record *vcpu_record,
 314			     struct event_key *key,
 315			     struct perf_sample *sample)
 316{
 317	struct kvm_event *event;
 318	u64 time_begin, time_diff;
 319	int vcpu;
 320
 321	if (kvm->trace_vcpu == -1)
 322		vcpu = -1;
 323	else
 324		vcpu = vcpu_record->vcpu_id;
 325
 326	event = vcpu_record->last_event;
 327	time_begin = vcpu_record->start_time;
 328
 329	/* The begin event is not caught. */
 330	if (!time_begin)
 331		return true;
 332
 333	/*
 334	 * In some case, the 'begin event' only records the start timestamp,
 335	 * the actual event is recognized in the 'end event' (e.g. mmio-event).
 336	 */
 337
 338	/* Both begin and end events did not get the key. */
 339	if (!event && key->key == INVALID_KEY)
 340		return true;
 341
 342	if (!event)
 343		event = find_create_kvm_event(kvm, key);
 344
 345	if (!event)
 346		return false;
 347
 348	vcpu_record->last_event = NULL;
 349	vcpu_record->start_time = 0;
 350
 351	/* seems to happen once in a while during live mode */
 352	if (sample->time < time_begin) {
 353		pr_debug("End time before begin time; skipping event.\n");
 354		return true;
 355	}
 356
 357	time_diff = sample->time - time_begin;
 358
 359	if (kvm->duration && time_diff > kvm->duration) {
 360		char decode[decode_str_len];
 361
 362		kvm->events_ops->decode_key(kvm, &event->key, decode);
 363		if (!skip_event(decode)) {
 364			pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
 365				 sample->time, sample->pid, vcpu_record->vcpu_id,
 366				 decode, time_diff / NSEC_PER_USEC);
 367		}
 368	}
 369
 370	return update_kvm_event(event, vcpu, time_diff);
 371}
 372
 373static
 374struct vcpu_event_record *per_vcpu_record(struct thread *thread,
 375					  struct perf_evsel *evsel,
 376					  struct perf_sample *sample)
 377{
 378	/* Only kvm_entry records vcpu id. */
 379	if (!thread__priv(thread) && kvm_entry_event(evsel)) {
 380		struct vcpu_event_record *vcpu_record;
 381
 382		vcpu_record = zalloc(sizeof(*vcpu_record));
 383		if (!vcpu_record) {
 384			pr_err("%s: Not enough memory\n", __func__);
 385			return NULL;
 386		}
 387
 388		vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample,
 389							  vcpu_id_str);
 390		thread__set_priv(thread, vcpu_record);
 391	}
 392
 393	return thread__priv(thread);
 394}
 395
 396static 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)					\
 427static 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)					\
 439GET_EVENT_KEY(func, field)						\
 440static 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
 447GET_EVENT_KEY(time, time);
 448COMPARE_EVENT_KEY(count, stats.n);
 449COMPARE_EVENT_KEY(mean, stats.mean);
 450GET_EVENT_KEY(max, stats.max);
 451GET_EVENT_KEY(min, stats.min);
 452
 453#define DEF_SORT_NAME_KEY(name, compare_key)				\
 454	{ #name, compare_kvm_event_ ## compare_key }
 455
 456static struct kvm_event_key keys[] = {
 457	DEF_SORT_NAME_KEY(sample, count),
 458	DEF_SORT_NAME_KEY(time, mean),
 459	{ NULL, NULL }
 460};
 461
 462static 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
 477static 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
 498static void
 499update_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
 507static bool event_is_valid(struct kvm_event *event, int vcpu)
 508{
 509	return !!get_event_count(event, vcpu);
 510}
 511
 512static 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 */
 530static 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
 541static 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
 560static 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
 576static 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 / NSEC_PER_USEC);
 613		pr_info("%9.2fus ", (double)max / NSEC_PER_USEC);
 614		pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount / NSEC_PER_USEC,
 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 / (double)NSEC_PER_USEC);
 621
 622	if (kvm->lost_events)
 623		pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
 624}
 625
 626#ifdef HAVE_TIMERFD_SUPPORT
 627static 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
 639static 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
 648static 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
 676static 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
 703static 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
 719static 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
 761static 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
 804static volatile int done;
 805
 806static void sig_handler(int sig __maybe_unused)
 807{
 808	done = 1;
 809}
 810
 811static 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;
 834out:
 835	return rc;
 836}
 837
 838static 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
 873static 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
 891static 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
 902static 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
 971out:
 972	if (kvm->timerfd >= 0)
 973		close(kvm->timerfd);
 974
 975	tcsetattr(0, TCSAFLUSH, &save);
 976	return err;
 977}
 978
 979static 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, NULL);
 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_entry(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		       str_error_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			  str_error_r(errno, sbuf, sizeof(sbuf)));
1029		perf_evlist__close(evlist);
1030		goto out;
1031	}
1032
1033	rc = 0;
1034
1035out:
1036	return rc;
1037}
1038#endif
1039
1040static 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
1079out_delete:
1080	perf_session__delete(kvm->session);
1081	return ret;
1082}
1083
1084static 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
1097static 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
1124exit:
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
1136int __weak setup_kvm_events_tp(struct perf_kvm_stat *kvm __maybe_unused)
1137{
1138	return 0;
1139}
1140
1141static int
1142kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1143{
1144	unsigned int rec_argc, i, j, events_tp_size;
1145	const char **rec_argv;
1146	const char * const record_args[] = {
1147		"record",
1148		"-R",
1149		"-m", "1024",
1150		"-c", "1",
1151	};
1152	const char * const kvm_stat_record_usage[] = {
1153		"perf kvm stat record [<options>]",
1154		NULL
1155	};
1156	const char * const *events_tp;
1157	int ret;
1158
1159	events_tp_size = 0;
1160	ret = setup_kvm_events_tp(kvm);
1161	if (ret < 0) {
1162		pr_err("Unable to setup the kvm tracepoints\n");
1163		return ret;
1164	}
1165
1166	for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1167		events_tp_size++;
1168
1169	rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1170		   2 * events_tp_size;
1171	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1172
1173	if (rec_argv == NULL)
1174		return -ENOMEM;
1175
1176	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1177		rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1178
1179	for (j = 0; j < events_tp_size; j++) {
1180		rec_argv[i++] = "-e";
1181		rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1182	}
1183
1184	rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1185	rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1186
1187	for (j = 1; j < (unsigned int)argc; j++, i++)
1188		rec_argv[i] = argv[j];
1189
1190	set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1191	set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1192	set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1193
1194	set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1195	set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1196	set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1197	set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1198	set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1199	set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1200	set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1201	set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1202	set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1203	set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1204	set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1205	set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1206	set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1207	set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1208	set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1209
1210	record_usage = kvm_stat_record_usage;
1211	return cmd_record(i, rec_argv, NULL);
1212}
1213
1214static int
1215kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1216{
1217	const struct option kvm_events_report_options[] = {
1218		OPT_STRING(0, "event", &kvm->report_event, "report event",
1219			   "event for reporting: vmexit, "
1220			   "mmio (x86 only), ioport (x86 only)"),
1221		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1222			    "vcpu id to report"),
1223		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1224			    "key for sorting: sample(sort by samples number)"
1225			    " time (sort by avg time)"),
1226		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1227			   "analyze events only for given process id(s)"),
1228		OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
1229		OPT_END()
1230	};
1231
1232	const char * const kvm_events_report_usage[] = {
1233		"perf kvm stat report [<options>]",
1234		NULL
1235	};
1236
1237	if (argc) {
1238		argc = parse_options(argc, argv,
1239				     kvm_events_report_options,
1240				     kvm_events_report_usage, 0);
1241		if (argc)
1242			usage_with_options(kvm_events_report_usage,
1243					   kvm_events_report_options);
1244	}
1245
1246	if (!kvm->opts.target.pid)
1247		kvm->opts.target.system_wide = true;
1248
1249	return kvm_events_report_vcpu(kvm);
1250}
1251
1252#ifdef HAVE_TIMERFD_SUPPORT
1253static struct perf_evlist *kvm_live_event_list(void)
1254{
1255	struct perf_evlist *evlist;
1256	char *tp, *name, *sys;
1257	int err = -1;
1258	const char * const *events_tp;
1259
1260	evlist = perf_evlist__new();
1261	if (evlist == NULL)
1262		return NULL;
1263
1264	for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1265
1266		tp = strdup(*events_tp);
1267		if (tp == NULL)
1268			goto out;
1269
1270		/* split tracepoint into subsystem and name */
1271		sys = tp;
1272		name = strchr(tp, ':');
1273		if (name == NULL) {
1274			pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1275			       *events_tp);
1276			free(tp);
1277			goto out;
1278		}
1279		*name = '\0';
1280		name++;
1281
1282		if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1283			pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1284			free(tp);
1285			goto out;
1286		}
1287
1288		free(tp);
1289	}
1290
1291	err = 0;
1292
1293out:
1294	if (err) {
1295		perf_evlist__delete(evlist);
1296		evlist = NULL;
1297	}
1298
1299	return evlist;
1300}
1301
1302static int kvm_events_live(struct perf_kvm_stat *kvm,
1303			   int argc, const char **argv)
1304{
1305	char errbuf[BUFSIZ];
1306	int err;
1307
1308	const struct option live_options[] = {
1309		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1310			"record events on existing process id"),
1311		OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1312			"number of mmap data pages",
1313			perf_evlist__parse_mmap_pages),
1314		OPT_INCR('v', "verbose", &verbose,
1315			"be more verbose (show counter open errors, etc)"),
1316		OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1317			"system-wide collection from all CPUs"),
1318		OPT_UINTEGER('d', "display", &kvm->display_time,
1319			"time in seconds between display updates"),
1320		OPT_STRING(0, "event", &kvm->report_event, "report event",
1321			"event for reporting: "
1322			"vmexit, mmio (x86 only), ioport (x86 only)"),
1323		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1324			"vcpu id to report"),
1325		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1326			"key for sorting: sample(sort by samples number)"
1327			" time (sort by avg time)"),
1328		OPT_U64(0, "duration", &kvm->duration,
1329			"show events other than"
1330			" HLT (x86 only) or Wait state (s390 only)"
1331			" that take longer than duration usecs"),
1332		OPT_UINTEGER(0, "proc-map-timeout", &kvm->opts.proc_map_timeout,
1333				"per thread proc mmap processing timeout in ms"),
1334		OPT_END()
1335	};
1336	const char * const live_usage[] = {
1337		"perf kvm stat live [<options>]",
1338		NULL
1339	};
1340	struct perf_data_file file = {
1341		.mode = PERF_DATA_MODE_WRITE,
1342	};
1343
1344
1345	/* event handling */
1346	kvm->tool.sample = process_sample_event;
1347	kvm->tool.comm   = perf_event__process_comm;
1348	kvm->tool.exit   = perf_event__process_exit;
1349	kvm->tool.fork   = perf_event__process_fork;
1350	kvm->tool.lost   = process_lost_event;
1351	kvm->tool.ordered_events = true;
1352	perf_tool__fill_defaults(&kvm->tool);
1353
1354	/* set defaults */
1355	kvm->display_time = 1;
1356	kvm->opts.user_interval = 1;
1357	kvm->opts.mmap_pages = 512;
1358	kvm->opts.target.uses_mmap = false;
1359	kvm->opts.target.uid_str = NULL;
1360	kvm->opts.target.uid = UINT_MAX;
1361	kvm->opts.proc_map_timeout = 500;
1362
1363	symbol__init(NULL);
1364	disable_buildid_cache();
1365
1366	use_browser = 0;
1367
1368	if (argc) {
1369		argc = parse_options(argc, argv, live_options,
1370				     live_usage, 0);
1371		if (argc)
1372			usage_with_options(live_usage, live_options);
1373	}
1374
1375	kvm->duration *= NSEC_PER_USEC;   /* convert usec to nsec */
1376
1377	/*
1378	 * target related setups
1379	 */
1380	err = target__validate(&kvm->opts.target);
1381	if (err) {
1382		target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1383		ui__warning("%s", errbuf);
1384	}
1385
1386	if (target__none(&kvm->opts.target))
1387		kvm->opts.target.system_wide = true;
1388
1389
1390	/*
1391	 * generate the event list
1392	 */
1393	err = setup_kvm_events_tp(kvm);
1394	if (err < 0) {
1395		pr_err("Unable to setup the kvm tracepoints\n");
1396		return err;
1397	}
1398
1399	kvm->evlist = kvm_live_event_list();
1400	if (kvm->evlist == NULL) {
1401		err = -1;
1402		goto out;
1403	}
1404
1405	symbol_conf.nr_events = kvm->evlist->nr_entries;
1406
1407	if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1408		usage_with_options(live_usage, live_options);
1409
1410	/*
1411	 * perf session
1412	 */
1413	kvm->session = perf_session__new(&file, false, &kvm->tool);
1414	if (kvm->session == NULL) {
1415		err = -1;
1416		goto out;
1417	}
1418	kvm->session->evlist = kvm->evlist;
1419	perf_session__set_id_hdr_size(kvm->session);
1420	ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
1421	machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1422				    kvm->evlist->threads, false, kvm->opts.proc_map_timeout);
1423	err = kvm_live_open_events(kvm);
1424	if (err)
1425		goto out;
1426
1427	err = kvm_events_live_report(kvm);
1428
1429out:
1430	perf_session__delete(kvm->session);
 
1431	kvm->session = NULL;
1432	perf_evlist__delete(kvm->evlist);
 
1433
1434	return err;
1435}
1436#endif
1437
1438static void print_kvm_stat_usage(void)
1439{
1440	printf("Usage: perf kvm stat <command>\n\n");
1441
1442	printf("# Available commands:\n");
1443	printf("\trecord: record kvm events\n");
1444	printf("\treport: report statistical data of kvm events\n");
1445	printf("\tlive:   live reporting of statistical data of kvm events\n");
1446
1447	printf("\nOtherwise, it is the alias of 'perf stat':\n");
1448}
1449
1450static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1451{
1452	struct perf_kvm_stat kvm = {
1453		.file_name = file_name,
1454
1455		.trace_vcpu	= -1,
1456		.report_event	= "vmexit",
1457		.sort_key	= "sample",
1458
1459	};
1460
1461	if (argc == 1) {
1462		print_kvm_stat_usage();
1463		goto perf_stat;
1464	}
1465
1466	if (!strncmp(argv[1], "rec", 3))
1467		return kvm_events_record(&kvm, argc - 1, argv + 1);
1468
1469	if (!strncmp(argv[1], "rep", 3))
1470		return kvm_events_report(&kvm, argc - 1 , argv + 1);
1471
1472#ifdef HAVE_TIMERFD_SUPPORT
1473	if (!strncmp(argv[1], "live", 4))
1474		return kvm_events_live(&kvm, argc - 1 , argv + 1);
1475#endif
1476
1477perf_stat:
1478	return cmd_stat(argc, argv, NULL);
1479}
1480#endif /* HAVE_KVM_STAT_SUPPORT */
1481
1482static int __cmd_record(const char *file_name, int argc, const char **argv)
1483{
1484	int rec_argc, i = 0, j;
1485	const char **rec_argv;
1486
1487	rec_argc = argc + 2;
1488	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1489	rec_argv[i++] = strdup("record");
1490	rec_argv[i++] = strdup("-o");
1491	rec_argv[i++] = strdup(file_name);
1492	for (j = 1; j < argc; j++, i++)
1493		rec_argv[i] = argv[j];
1494
1495	BUG_ON(i != rec_argc);
1496
1497	return cmd_record(i, rec_argv, NULL);
1498}
1499
1500static int __cmd_report(const char *file_name, int argc, const char **argv)
1501{
1502	int rec_argc, i = 0, j;
1503	const char **rec_argv;
1504
1505	rec_argc = argc + 2;
1506	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1507	rec_argv[i++] = strdup("report");
1508	rec_argv[i++] = strdup("-i");
1509	rec_argv[i++] = strdup(file_name);
1510	for (j = 1; j < argc; j++, i++)
1511		rec_argv[i] = argv[j];
1512
1513	BUG_ON(i != rec_argc);
1514
1515	return cmd_report(i, rec_argv, NULL);
1516}
1517
1518static int
1519__cmd_buildid_list(const char *file_name, int argc, const char **argv)
1520{
1521	int rec_argc, i = 0, j;
1522	const char **rec_argv;
1523
1524	rec_argc = argc + 2;
1525	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1526	rec_argv[i++] = strdup("buildid-list");
1527	rec_argv[i++] = strdup("-i");
1528	rec_argv[i++] = strdup(file_name);
1529	for (j = 1; j < argc; j++, i++)
1530		rec_argv[i] = argv[j];
1531
1532	BUG_ON(i != rec_argc);
1533
1534	return cmd_buildid_list(i, rec_argv, NULL);
1535}
1536
1537int cmd_kvm(int argc, const char **argv, const char *prefix __maybe_unused)
1538{
1539	const char *file_name = NULL;
1540	const struct option kvm_options[] = {
1541		OPT_STRING('i', "input", &file_name, "file",
1542			   "Input file name"),
1543		OPT_STRING('o', "output", &file_name, "file",
1544			   "Output file name"),
1545		OPT_BOOLEAN(0, "guest", &perf_guest,
1546			    "Collect guest os data"),
1547		OPT_BOOLEAN(0, "host", &perf_host,
1548			    "Collect host os data"),
1549		OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1550			   "guest mount directory under which every guest os"
1551			   " instance has a subdir"),
1552		OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1553			   "file", "file saving guest os vmlinux"),
1554		OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1555			   "file", "file saving guest os /proc/kallsyms"),
1556		OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1557			   "file", "file saving guest os /proc/modules"),
1558		OPT_INCR('v', "verbose", &verbose,
1559			    "be more verbose (show counter open errors, etc)"),
1560		OPT_END()
1561	};
1562
1563	const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1564						"buildid-list", "stat", NULL };
1565	const char *kvm_usage[] = { NULL, NULL };
1566
1567	perf_host  = 0;
1568	perf_guest = 1;
1569
1570	argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1571					PARSE_OPT_STOP_AT_NON_OPTION);
1572	if (!argc)
1573		usage_with_options(kvm_usage, kvm_options);
1574
1575	if (!perf_host)
1576		perf_guest = 1;
1577
1578	if (!file_name) {
1579		file_name = get_filename_for_perf_kvm();
1580
1581		if (!file_name) {
1582			pr_err("Failed to allocate memory for filename\n");
1583			return -ENOMEM;
1584		}
1585	}
1586
1587	if (!strncmp(argv[0], "rec", 3))
1588		return __cmd_record(file_name, argc, argv);
1589	else if (!strncmp(argv[0], "rep", 3))
1590		return __cmd_report(file_name, argc, argv);
1591	else if (!strncmp(argv[0], "diff", 4))
1592		return cmd_diff(argc, argv, NULL);
1593	else if (!strncmp(argv[0], "top", 3))
1594		return cmd_top(argc, argv, NULL);
1595	else if (!strncmp(argv[0], "buildid-list", 12))
1596		return __cmd_buildid_list(file_name, argc, argv);
1597#ifdef HAVE_KVM_STAT_SUPPORT
1598	else if (!strncmp(argv[0], "stat", 4))
1599		return kvm_cmd_stat(file_name, argc, argv);
1600#endif
1601	else
1602		usage_with_options(kvm_usage, kvm_options);
1603
1604	return 0;
1605}