Linux Audio

Check our new training course

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}
v5.4
   1// SPDX-License-Identifier: GPL-2.0
   2#include "builtin.h"
   3#include "perf.h"
   4
   5#include "util/build-id.h"
   6#include "util/evsel.h"
   7#include "util/evlist.h"
   8#include "util/mmap.h"
   9#include "util/term.h"
  10#include "util/symbol.h"
  11#include "util/thread.h"
  12#include "util/header.h"
  13#include "util/session.h"
  14#include "util/intlist.h"
  15#include <subcmd/pager.h>
  16#include <subcmd/parse-options.h>
  17#include "util/trace-event.h"
  18#include "util/debug.h"
  19#include "util/tool.h"
  20#include "util/stat.h"
  21#include "util/synthetic-events.h"
  22#include "util/top.h"
  23#include "util/data.h"
  24#include "util/ordered-events.h"
  25#include "util/kvm-stat.h"
  26#include "ui/ui.h"
  27
  28#include <sys/prctl.h>
  29#ifdef HAVE_TIMERFD_SUPPORT
  30#include <sys/timerfd.h>
  31#endif
  32#include <sys/time.h>
  33#include <sys/types.h>
  34#include <sys/stat.h>
  35#include <fcntl.h>
  36
  37#include <linux/err.h>
  38#include <linux/kernel.h>
  39#include <linux/string.h>
  40#include <linux/time64.h>
  41#include <linux/zalloc.h>
  42#include <errno.h>
  43#include <inttypes.h>
  44#include <poll.h>
  45#include <termios.h>
  46#include <semaphore.h>
  47#include <signal.h>
  48#include <math.h>
  49
  50static const char *get_filename_for_perf_kvm(void)
  51{
  52	const char *filename;
  53
  54	if (perf_host && !perf_guest)
  55		filename = strdup("perf.data.host");
  56	else if (!perf_host && perf_guest)
  57		filename = strdup("perf.data.guest");
  58	else
  59		filename = strdup("perf.data.kvm");
  60
  61	return filename;
  62}
  63
  64#ifdef HAVE_KVM_STAT_SUPPORT
 
  65
  66void exit_event_get_key(struct evsel *evsel,
  67			struct perf_sample *sample,
  68			struct event_key *key)
  69{
  70	key->info = 0;
  71	key->key = perf_evsel__intval(evsel, sample, kvm_exit_reason);
  72}
  73
  74bool kvm_exit_event(struct evsel *evsel)
  75{
  76	return !strcmp(evsel->name, kvm_exit_trace);
  77}
  78
  79bool exit_event_begin(struct evsel *evsel,
  80		      struct perf_sample *sample, struct event_key *key)
  81{
  82	if (kvm_exit_event(evsel)) {
  83		exit_event_get_key(evsel, sample, key);
  84		return true;
  85	}
  86
  87	return false;
  88}
  89
  90bool kvm_entry_event(struct evsel *evsel)
  91{
  92	return !strcmp(evsel->name, kvm_entry_trace);
  93}
  94
  95bool exit_event_end(struct evsel *evsel,
  96		    struct perf_sample *sample __maybe_unused,
  97		    struct event_key *key __maybe_unused)
  98{
  99	return kvm_entry_event(evsel);
 100}
 101
 102static const char *get_exit_reason(struct perf_kvm_stat *kvm,
 103				   struct exit_reasons_table *tbl,
 104				   u64 exit_code)
 105{
 106	while (tbl->reason != NULL) {
 107		if (tbl->exit_code == exit_code)
 108			return tbl->reason;
 109		tbl++;
 110	}
 111
 112	pr_err("unknown kvm exit code:%lld on %s\n",
 113		(unsigned long long)exit_code, kvm->exit_reasons_isa);
 114	return "UNKNOWN";
 115}
 116
 117void exit_event_decode_key(struct perf_kvm_stat *kvm,
 118			   struct event_key *key,
 119			   char *decode)
 120{
 121	const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
 122						  key->key);
 123
 124	scnprintf(decode, decode_str_len, "%s", exit_reason);
 125}
 126
 127static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
 128{
 129	struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
 130
 131	for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
 132		if (!strcmp(events_ops->name, kvm->report_event)) {
 133			kvm->events_ops = events_ops->ops;
 134			return true;
 135		}
 136	}
 137
 138	return false;
 139}
 140
 141struct vcpu_event_record {
 142	int vcpu_id;
 143	u64 start_time;
 144	struct kvm_event *last_event;
 145};
 146
 147
 148static void init_kvm_event_record(struct perf_kvm_stat *kvm)
 149{
 150	unsigned int i;
 151
 152	for (i = 0; i < EVENTS_CACHE_SIZE; i++)
 153		INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
 154}
 155
 156#ifdef HAVE_TIMERFD_SUPPORT
 157static void clear_events_cache_stats(struct list_head *kvm_events_cache)
 158{
 159	struct list_head *head;
 160	struct kvm_event *event;
 161	unsigned int i;
 162	int j;
 163
 164	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
 165		head = &kvm_events_cache[i];
 166		list_for_each_entry(event, head, hash_entry) {
 167			/* reset stats for event */
 168			event->total.time = 0;
 169			init_stats(&event->total.stats);
 170
 171			for (j = 0; j < event->max_vcpu; ++j) {
 172				event->vcpu[j].time = 0;
 173				init_stats(&event->vcpu[j].stats);
 174			}
 175		}
 176	}
 177}
 178#endif
 179
 180static int kvm_events_hash_fn(u64 key)
 181{
 182	return key & (EVENTS_CACHE_SIZE - 1);
 183}
 184
 185static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
 186{
 187	int old_max_vcpu = event->max_vcpu;
 188	void *prev;
 189
 190	if (vcpu_id < event->max_vcpu)
 191		return true;
 192
 193	while (event->max_vcpu <= vcpu_id)
 194		event->max_vcpu += DEFAULT_VCPU_NUM;
 195
 196	prev = event->vcpu;
 197	event->vcpu = realloc(event->vcpu,
 198			      event->max_vcpu * sizeof(*event->vcpu));
 199	if (!event->vcpu) {
 200		free(prev);
 201		pr_err("Not enough memory\n");
 202		return false;
 203	}
 204
 205	memset(event->vcpu + old_max_vcpu, 0,
 206	       (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
 207	return true;
 208}
 209
 210static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
 211{
 212	struct kvm_event *event;
 213
 214	event = zalloc(sizeof(*event));
 215	if (!event) {
 216		pr_err("Not enough memory\n");
 217		return NULL;
 218	}
 219
 220	event->key = *key;
 221	init_stats(&event->total.stats);
 222	return event;
 223}
 224
 225static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
 226					       struct event_key *key)
 227{
 228	struct kvm_event *event;
 229	struct list_head *head;
 230
 231	BUG_ON(key->key == INVALID_KEY);
 232
 233	head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
 234	list_for_each_entry(event, head, hash_entry) {
 235		if (event->key.key == key->key && event->key.info == key->info)
 236			return event;
 237	}
 238
 239	event = kvm_alloc_init_event(key);
 240	if (!event)
 241		return NULL;
 242
 243	list_add(&event->hash_entry, head);
 244	return event;
 245}
 246
 247static bool handle_begin_event(struct perf_kvm_stat *kvm,
 248			       struct vcpu_event_record *vcpu_record,
 249			       struct event_key *key, u64 timestamp)
 250{
 251	struct kvm_event *event = NULL;
 252
 253	if (key->key != INVALID_KEY)
 254		event = find_create_kvm_event(kvm, key);
 255
 256	vcpu_record->last_event = event;
 257	vcpu_record->start_time = timestamp;
 258	return true;
 259}
 260
 261static void
 262kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
 263{
 264	kvm_stats->time += time_diff;
 265	update_stats(&kvm_stats->stats, time_diff);
 266}
 267
 268static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
 269{
 270	struct kvm_event_stats *kvm_stats = &event->total;
 271
 272	if (vcpu_id != -1)
 273		kvm_stats = &event->vcpu[vcpu_id];
 274
 275	return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
 276				avg_stats(&kvm_stats->stats));
 277}
 278
 279static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
 280			     u64 time_diff)
 281{
 282	if (vcpu_id == -1) {
 283		kvm_update_event_stats(&event->total, time_diff);
 284		return true;
 285	}
 286
 287	if (!kvm_event_expand(event, vcpu_id))
 288		return false;
 289
 290	kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
 291	return true;
 292}
 293
 294static bool is_child_event(struct perf_kvm_stat *kvm,
 295			   struct evsel *evsel,
 296			   struct perf_sample *sample,
 297			   struct event_key *key)
 298{
 299	struct child_event_ops *child_ops;
 300
 301	child_ops = kvm->events_ops->child_ops;
 302
 303	if (!child_ops)
 304		return false;
 305
 306	for (; child_ops->name; child_ops++) {
 307		if (!strcmp(evsel->name, child_ops->name)) {
 308			child_ops->get_key(evsel, sample, key);
 309			return true;
 310		}
 311	}
 312
 313	return false;
 314}
 315
 316static bool handle_child_event(struct perf_kvm_stat *kvm,
 317			       struct vcpu_event_record *vcpu_record,
 318			       struct event_key *key,
 319			       struct perf_sample *sample __maybe_unused)
 320{
 321	struct kvm_event *event = NULL;
 322
 323	if (key->key != INVALID_KEY)
 324		event = find_create_kvm_event(kvm, key);
 325
 326	vcpu_record->last_event = event;
 327
 328	return true;
 329}
 330
 331static bool skip_event(const char *event)
 332{
 333	const char * const *skip_events;
 334
 335	for (skip_events = kvm_skip_events; *skip_events; skip_events++)
 336		if (!strcmp(event, *skip_events))
 337			return true;
 338
 339	return false;
 340}
 341
 342static bool handle_end_event(struct perf_kvm_stat *kvm,
 343			     struct vcpu_event_record *vcpu_record,
 344			     struct event_key *key,
 345			     struct perf_sample *sample)
 346{
 347	struct kvm_event *event;
 348	u64 time_begin, time_diff;
 349	int vcpu;
 350
 351	if (kvm->trace_vcpu == -1)
 352		vcpu = -1;
 353	else
 354		vcpu = vcpu_record->vcpu_id;
 355
 356	event = vcpu_record->last_event;
 357	time_begin = vcpu_record->start_time;
 358
 359	/* The begin event is not caught. */
 360	if (!time_begin)
 361		return true;
 362
 363	/*
 364	 * In some case, the 'begin event' only records the start timestamp,
 365	 * the actual event is recognized in the 'end event' (e.g. mmio-event).
 366	 */
 367
 368	/* Both begin and end events did not get the key. */
 369	if (!event && key->key == INVALID_KEY)
 370		return true;
 371
 372	if (!event)
 373		event = find_create_kvm_event(kvm, key);
 374
 375	if (!event)
 376		return false;
 377
 378	vcpu_record->last_event = NULL;
 379	vcpu_record->start_time = 0;
 380
 381	/* seems to happen once in a while during live mode */
 382	if (sample->time < time_begin) {
 383		pr_debug("End time before begin time; skipping event.\n");
 384		return true;
 385	}
 386
 387	time_diff = sample->time - time_begin;
 388
 389	if (kvm->duration && time_diff > kvm->duration) {
 390		char decode[decode_str_len];
 391
 392		kvm->events_ops->decode_key(kvm, &event->key, decode);
 393		if (!skip_event(decode)) {
 394			pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
 395				 sample->time, sample->pid, vcpu_record->vcpu_id,
 396				 decode, time_diff / NSEC_PER_USEC);
 397		}
 398	}
 399
 400	return update_kvm_event(event, vcpu, time_diff);
 401}
 402
 403static
 404struct vcpu_event_record *per_vcpu_record(struct thread *thread,
 405					  struct evsel *evsel,
 406					  struct perf_sample *sample)
 407{
 408	/* Only kvm_entry records vcpu id. */
 409	if (!thread__priv(thread) && kvm_entry_event(evsel)) {
 410		struct vcpu_event_record *vcpu_record;
 411
 412		vcpu_record = zalloc(sizeof(*vcpu_record));
 413		if (!vcpu_record) {
 414			pr_err("%s: Not enough memory\n", __func__);
 415			return NULL;
 416		}
 417
 418		vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample,
 419							  vcpu_id_str);
 420		thread__set_priv(thread, vcpu_record);
 421	}
 422
 423	return thread__priv(thread);
 424}
 425
 426static bool handle_kvm_event(struct perf_kvm_stat *kvm,
 427			     struct thread *thread,
 428			     struct evsel *evsel,
 429			     struct perf_sample *sample)
 430{
 431	struct vcpu_event_record *vcpu_record;
 432	struct event_key key = { .key = INVALID_KEY,
 433				 .exit_reasons = kvm->exit_reasons };
 434
 435	vcpu_record = per_vcpu_record(thread, evsel, sample);
 436	if (!vcpu_record)
 437		return true;
 438
 439	/* only process events for vcpus user cares about */
 440	if ((kvm->trace_vcpu != -1) &&
 441	    (kvm->trace_vcpu != vcpu_record->vcpu_id))
 442		return true;
 443
 444	if (kvm->events_ops->is_begin_event(evsel, sample, &key))
 445		return handle_begin_event(kvm, vcpu_record, &key, sample->time);
 446
 447	if (is_child_event(kvm, evsel, sample, &key))
 448		return handle_child_event(kvm, vcpu_record, &key, sample);
 449
 450	if (kvm->events_ops->is_end_event(evsel, sample, &key))
 451		return handle_end_event(kvm, vcpu_record, &key, sample);
 452
 453	return true;
 454}
 455
 456#define GET_EVENT_KEY(func, field)					\
 457static u64 get_event_ ##func(struct kvm_event *event, int vcpu)		\
 458{									\
 459	if (vcpu == -1)							\
 460		return event->total.field;				\
 461									\
 462	if (vcpu >= event->max_vcpu)					\
 463		return 0;						\
 464									\
 465	return event->vcpu[vcpu].field;					\
 466}
 467
 468#define COMPARE_EVENT_KEY(func, field)					\
 469GET_EVENT_KEY(func, field)						\
 470static int compare_kvm_event_ ## func(struct kvm_event *one,		\
 471					struct kvm_event *two, int vcpu)\
 472{									\
 473	return get_event_ ##func(one, vcpu) >				\
 474				get_event_ ##func(two, vcpu);		\
 475}
 476
 477GET_EVENT_KEY(time, time);
 478COMPARE_EVENT_KEY(count, stats.n);
 479COMPARE_EVENT_KEY(mean, stats.mean);
 480GET_EVENT_KEY(max, stats.max);
 481GET_EVENT_KEY(min, stats.min);
 482
 483#define DEF_SORT_NAME_KEY(name, compare_key)				\
 484	{ #name, compare_kvm_event_ ## compare_key }
 485
 486static struct kvm_event_key keys[] = {
 487	DEF_SORT_NAME_KEY(sample, count),
 488	DEF_SORT_NAME_KEY(time, mean),
 489	{ NULL, NULL }
 490};
 491
 492static bool select_key(struct perf_kvm_stat *kvm)
 493{
 494	int i;
 495
 496	for (i = 0; keys[i].name; i++) {
 497		if (!strcmp(keys[i].name, kvm->sort_key)) {
 498			kvm->compare = keys[i].key;
 499			return true;
 500		}
 501	}
 502
 503	pr_err("Unknown compare key:%s\n", kvm->sort_key);
 504	return false;
 505}
 506
 507static void insert_to_result(struct rb_root *result, struct kvm_event *event,
 508			     key_cmp_fun bigger, int vcpu)
 509{
 510	struct rb_node **rb = &result->rb_node;
 511	struct rb_node *parent = NULL;
 512	struct kvm_event *p;
 513
 514	while (*rb) {
 515		p = container_of(*rb, struct kvm_event, rb);
 516		parent = *rb;
 517
 518		if (bigger(event, p, vcpu))
 519			rb = &(*rb)->rb_left;
 520		else
 521			rb = &(*rb)->rb_right;
 522	}
 523
 524	rb_link_node(&event->rb, parent, rb);
 525	rb_insert_color(&event->rb, result);
 526}
 527
 528static void
 529update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
 530{
 531	int vcpu = kvm->trace_vcpu;
 532
 533	kvm->total_count += get_event_count(event, vcpu);
 534	kvm->total_time += get_event_time(event, vcpu);
 535}
 536
 537static bool event_is_valid(struct kvm_event *event, int vcpu)
 538{
 539	return !!get_event_count(event, vcpu);
 540}
 541
 542static void sort_result(struct perf_kvm_stat *kvm)
 543{
 544	unsigned int i;
 545	int vcpu = kvm->trace_vcpu;
 546	struct kvm_event *event;
 547
 548	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
 549		list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
 550			if (event_is_valid(event, vcpu)) {
 551				update_total_count(kvm, event);
 552				insert_to_result(&kvm->result, event,
 553						 kvm->compare, vcpu);
 554			}
 555		}
 556	}
 557}
 558
 559/* returns left most element of result, and erase it */
 560static struct kvm_event *pop_from_result(struct rb_root *result)
 561{
 562	struct rb_node *node = rb_first(result);
 563
 564	if (!node)
 565		return NULL;
 566
 567	rb_erase(node, result);
 568	return container_of(node, struct kvm_event, rb);
 569}
 570
 571static void print_vcpu_info(struct perf_kvm_stat *kvm)
 572{
 573	int vcpu = kvm->trace_vcpu;
 574
 575	pr_info("Analyze events for ");
 576
 577	if (kvm->opts.target.system_wide)
 578		pr_info("all VMs, ");
 579	else if (kvm->opts.target.pid)
 580		pr_info("pid(s) %s, ", kvm->opts.target.pid);
 581	else
 582		pr_info("dazed and confused on what is monitored, ");
 583
 584	if (vcpu == -1)
 585		pr_info("all VCPUs:\n\n");
 586	else
 587		pr_info("VCPU %d:\n\n", vcpu);
 588}
 589
 590static void show_timeofday(void)
 591{
 592	char date[64];
 593	struct timeval tv;
 594	struct tm ltime;
 595
 596	gettimeofday(&tv, NULL);
 597	if (localtime_r(&tv.tv_sec, &ltime)) {
 598		strftime(date, sizeof(date), "%H:%M:%S", &ltime);
 599		pr_info("%s.%06ld", date, tv.tv_usec);
 600	} else
 601		pr_info("00:00:00.000000");
 602
 603	return;
 604}
 605
 606static void print_result(struct perf_kvm_stat *kvm)
 607{
 608	char decode[decode_str_len];
 609	struct kvm_event *event;
 610	int vcpu = kvm->trace_vcpu;
 611
 612	if (kvm->live) {
 613		puts(CONSOLE_CLEAR);
 614		show_timeofday();
 615	}
 616
 617	pr_info("\n\n");
 618	print_vcpu_info(kvm);
 619	pr_info("%*s ", decode_str_len, kvm->events_ops->name);
 620	pr_info("%10s ", "Samples");
 621	pr_info("%9s ", "Samples%");
 622
 623	pr_info("%9s ", "Time%");
 624	pr_info("%11s ", "Min Time");
 625	pr_info("%11s ", "Max Time");
 626	pr_info("%16s ", "Avg time");
 627	pr_info("\n\n");
 628
 629	while ((event = pop_from_result(&kvm->result))) {
 630		u64 ecount, etime, max, min;
 631
 632		ecount = get_event_count(event, vcpu);
 633		etime = get_event_time(event, vcpu);
 634		max = get_event_max(event, vcpu);
 635		min = get_event_min(event, vcpu);
 636
 637		kvm->events_ops->decode_key(kvm, &event->key, decode);
 638		pr_info("%*s ", decode_str_len, decode);
 639		pr_info("%10llu ", (unsigned long long)ecount);
 640		pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
 641		pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
 642		pr_info("%9.2fus ", (double)min / NSEC_PER_USEC);
 643		pr_info("%9.2fus ", (double)max / NSEC_PER_USEC);
 644		pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount / NSEC_PER_USEC,
 645			kvm_event_rel_stddev(vcpu, event));
 646		pr_info("\n");
 647	}
 648
 649	pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
 650		kvm->total_count, kvm->total_time / (double)NSEC_PER_USEC);
 651
 652	if (kvm->lost_events)
 653		pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
 654}
 655
 656#ifdef HAVE_TIMERFD_SUPPORT
 657static int process_lost_event(struct perf_tool *tool,
 658			      union perf_event *event __maybe_unused,
 659			      struct perf_sample *sample __maybe_unused,
 660			      struct machine *machine __maybe_unused)
 661{
 662	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
 663
 664	kvm->lost_events++;
 665	return 0;
 666}
 667#endif
 668
 669static bool skip_sample(struct perf_kvm_stat *kvm,
 670			struct perf_sample *sample)
 671{
 672	if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
 673		return true;
 674
 675	return false;
 676}
 677
 678static int process_sample_event(struct perf_tool *tool,
 679				union perf_event *event,
 680				struct perf_sample *sample,
 681				struct evsel *evsel,
 682				struct machine *machine)
 683{
 684	int err = 0;
 685	struct thread *thread;
 686	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
 687						 tool);
 688
 689	if (skip_sample(kvm, sample))
 690		return 0;
 691
 692	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
 693	if (thread == NULL) {
 694		pr_debug("problem processing %d event, skipping it.\n",
 695			event->header.type);
 696		return -1;
 697	}
 698
 699	if (!handle_kvm_event(kvm, thread, evsel, sample))
 700		err = -1;
 701
 702	thread__put(thread);
 703	return err;
 704}
 705
 706static int cpu_isa_config(struct perf_kvm_stat *kvm)
 707{
 708	char buf[128], *cpuid;
 709	int err;
 710
 711	if (kvm->live) {
 712		err = get_cpuid(buf, sizeof(buf));
 713		if (err != 0) {
 714			pr_err("Failed to look up CPU type: %s\n",
 715			       str_error_r(err, buf, sizeof(buf)));
 716			return -err;
 717		}
 718		cpuid = buf;
 719	} else
 720		cpuid = kvm->session->header.env.cpuid;
 721
 722	if (!cpuid) {
 723		pr_err("Failed to look up CPU type\n");
 724		return -EINVAL;
 725	}
 726
 727	err = cpu_isa_init(kvm, cpuid);
 728	if (err == -ENOTSUP)
 729		pr_err("CPU %s is not supported.\n", cpuid);
 730
 731	return err;
 732}
 733
 734static bool verify_vcpu(int vcpu)
 735{
 736	if (vcpu != -1 && vcpu < 0) {
 737		pr_err("Invalid vcpu:%d.\n", vcpu);
 738		return false;
 739	}
 740
 741	return true;
 742}
 743
 744#ifdef HAVE_TIMERFD_SUPPORT
 745/* keeping the max events to a modest level to keep
 746 * the processing of samples per mmap smooth.
 747 */
 748#define PERF_KVM__MAX_EVENTS_PER_MMAP  25
 749
 750static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
 751				   u64 *mmap_time)
 752{
 753	struct evlist *evlist = kvm->evlist;
 754	union perf_event *event;
 755	struct mmap *md;
 756	u64 timestamp;
 757	s64 n = 0;
 758	int err;
 759
 760	*mmap_time = ULLONG_MAX;
 761	md = &evlist->mmap[idx];
 762	err = perf_mmap__read_init(md);
 763	if (err < 0)
 764		return (err == -EAGAIN) ? 0 : -1;
 765
 766	while ((event = perf_mmap__read_event(md)) != NULL) {
 767		err = perf_evlist__parse_sample_timestamp(evlist, event, &timestamp);
 768		if (err) {
 769			perf_mmap__consume(md);
 770			pr_err("Failed to parse sample\n");
 771			return -1;
 772		}
 773
 774		err = perf_session__queue_event(kvm->session, event, timestamp, 0);
 775		/*
 776		 * FIXME: Here we can't consume the event, as perf_session__queue_event will
 777		 *        point to it, and it'll get possibly overwritten by the kernel.
 778		 */
 779		perf_mmap__consume(md);
 780
 781		if (err) {
 782			pr_err("Failed to enqueue sample: %d\n", err);
 783			return -1;
 784		}
 785
 786		/* save time stamp of our first sample for this mmap */
 787		if (n == 0)
 788			*mmap_time = timestamp;
 789
 790		/* limit events per mmap handled all at once */
 791		n++;
 792		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
 793			break;
 794	}
 795
 796	perf_mmap__read_done(md);
 797	return n;
 798}
 799
 800static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
 801{
 802	int i, err, throttled = 0;
 803	s64 n, ntotal = 0;
 804	u64 flush_time = ULLONG_MAX, mmap_time;
 805
 806	for (i = 0; i < kvm->evlist->core.nr_mmaps; i++) {
 807		n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
 808		if (n < 0)
 809			return -1;
 810
 811		/* flush time is going to be the minimum of all the individual
 812		 * mmap times. Essentially, we flush all the samples queued up
 813		 * from the last pass under our minimal start time -- that leaves
 814		 * a very small race for samples to come in with a lower timestamp.
 815		 * The ioctl to return the perf_clock timestamp should close the
 816		 * race entirely.
 817		 */
 818		if (mmap_time < flush_time)
 819			flush_time = mmap_time;
 820
 821		ntotal += n;
 822		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
 823			throttled = 1;
 824	}
 825
 826	/* flush queue after each round in which we processed events */
 827	if (ntotal) {
 828		struct ordered_events *oe = &kvm->session->ordered_events;
 829
 830		oe->next_flush = flush_time;
 831		err = ordered_events__flush(oe, OE_FLUSH__ROUND);
 832		if (err) {
 833			if (kvm->lost_events)
 834				pr_info("\nLost events: %" PRIu64 "\n\n",
 835					kvm->lost_events);
 836			return err;
 837		}
 838	}
 839
 840	return throttled;
 841}
 842
 843static volatile int done;
 844
 845static void sig_handler(int sig __maybe_unused)
 846{
 847	done = 1;
 848}
 849
 850static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
 851{
 852	struct itimerspec new_value;
 853	int rc = -1;
 854
 855	kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
 856	if (kvm->timerfd < 0) {
 857		pr_err("timerfd_create failed\n");
 858		goto out;
 859	}
 860
 861	new_value.it_value.tv_sec = kvm->display_time;
 862	new_value.it_value.tv_nsec = 0;
 863	new_value.it_interval.tv_sec = kvm->display_time;
 864	new_value.it_interval.tv_nsec = 0;
 865
 866	if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
 867		pr_err("timerfd_settime failed: %d\n", errno);
 868		close(kvm->timerfd);
 869		goto out;
 870	}
 871
 872	rc = 0;
 873out:
 874	return rc;
 875}
 876
 877static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
 878{
 879	uint64_t c;
 880	int rc;
 881
 882	rc = read(kvm->timerfd, &c, sizeof(uint64_t));
 883	if (rc < 0) {
 884		if (errno == EAGAIN)
 885			return 0;
 886
 887		pr_err("Failed to read timer fd: %d\n", errno);
 888		return -1;
 889	}
 890
 891	if (rc != sizeof(uint64_t)) {
 892		pr_err("Error reading timer fd - invalid size returned\n");
 893		return -1;
 894	}
 895
 896	if (c != 1)
 897		pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
 898
 899	/* update display */
 900	sort_result(kvm);
 901	print_result(kvm);
 902
 903	/* reset counts */
 904	clear_events_cache_stats(kvm->kvm_events_cache);
 905	kvm->total_count = 0;
 906	kvm->total_time = 0;
 907	kvm->lost_events = 0;
 908
 909	return 0;
 910}
 911
 912static int fd_set_nonblock(int fd)
 913{
 914	long arg = 0;
 915
 916	arg = fcntl(fd, F_GETFL);
 917	if (arg < 0) {
 918		pr_err("Failed to get current flags for fd %d\n", fd);
 919		return -1;
 920	}
 921
 922	if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
 923		pr_err("Failed to set non-block option on fd %d\n", fd);
 924		return -1;
 925	}
 926
 927	return 0;
 928}
 929
 930static int perf_kvm__handle_stdin(void)
 931{
 932	int c;
 933
 934	c = getc(stdin);
 935	if (c == 'q')
 936		return 1;
 937
 938	return 0;
 939}
 940
 941static int kvm_events_live_report(struct perf_kvm_stat *kvm)
 942{
 943	int nr_stdin, ret, err = -EINVAL;
 944	struct termios save;
 945
 946	/* live flag must be set first */
 947	kvm->live = true;
 948
 949	ret = cpu_isa_config(kvm);
 950	if (ret < 0)
 951		return ret;
 952
 953	if (!verify_vcpu(kvm->trace_vcpu) ||
 954	    !select_key(kvm) ||
 955	    !register_kvm_events_ops(kvm)) {
 956		goto out;
 957	}
 958
 959	set_term_quiet_input(&save);
 960	init_kvm_event_record(kvm);
 961
 962	signal(SIGINT, sig_handler);
 963	signal(SIGTERM, sig_handler);
 964
 965	/* add timer fd */
 966	if (perf_kvm__timerfd_create(kvm) < 0) {
 967		err = -1;
 968		goto out;
 969	}
 970
 971	if (evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
 972		goto out;
 973
 974	nr_stdin = evlist__add_pollfd(kvm->evlist, fileno(stdin));
 975	if (nr_stdin < 0)
 976		goto out;
 977
 978	if (fd_set_nonblock(fileno(stdin)) != 0)
 979		goto out;
 980
 981	/* everything is good - enable the events and process */
 982	evlist__enable(kvm->evlist);
 983
 984	while (!done) {
 985		struct fdarray *fda = &kvm->evlist->core.pollfd;
 986		int rc;
 987
 988		rc = perf_kvm__mmap_read(kvm);
 989		if (rc < 0)
 990			break;
 991
 992		err = perf_kvm__handle_timerfd(kvm);
 993		if (err)
 994			goto out;
 995
 996		if (fda->entries[nr_stdin].revents & POLLIN)
 997			done = perf_kvm__handle_stdin();
 998
 999		if (!rc && !done)
1000			err = fdarray__poll(fda, 100);
1001	}
1002
1003	evlist__disable(kvm->evlist);
1004
1005	if (err == 0) {
1006		sort_result(kvm);
1007		print_result(kvm);
1008	}
1009
1010out:
1011	if (kvm->timerfd >= 0)
1012		close(kvm->timerfd);
1013
1014	tcsetattr(0, TCSAFLUSH, &save);
1015	return err;
1016}
1017
1018static int kvm_live_open_events(struct perf_kvm_stat *kvm)
1019{
1020	int err, rc = -1;
1021	struct evsel *pos;
1022	struct evlist *evlist = kvm->evlist;
1023	char sbuf[STRERR_BUFSIZE];
1024
1025	perf_evlist__config(evlist, &kvm->opts, NULL);
1026
1027	/*
1028	 * Note: exclude_{guest,host} do not apply here.
1029	 *       This command processes KVM tracepoints from host only
1030	 */
1031	evlist__for_each_entry(evlist, pos) {
1032		struct perf_event_attr *attr = &pos->core.attr;
1033
1034		/* make sure these *are* set */
1035		perf_evsel__set_sample_bit(pos, TID);
1036		perf_evsel__set_sample_bit(pos, TIME);
1037		perf_evsel__set_sample_bit(pos, CPU);
1038		perf_evsel__set_sample_bit(pos, RAW);
1039		/* make sure these are *not*; want as small a sample as possible */
1040		perf_evsel__reset_sample_bit(pos, PERIOD);
1041		perf_evsel__reset_sample_bit(pos, IP);
1042		perf_evsel__reset_sample_bit(pos, CALLCHAIN);
1043		perf_evsel__reset_sample_bit(pos, ADDR);
1044		perf_evsel__reset_sample_bit(pos, READ);
1045		attr->mmap = 0;
1046		attr->comm = 0;
1047		attr->task = 0;
1048
1049		attr->sample_period = 1;
1050
1051		attr->watermark = 0;
1052		attr->wakeup_events = 1000;
1053
1054		/* will enable all once we are ready */
1055		attr->disabled = 1;
1056	}
1057
1058	err = evlist__open(evlist);
1059	if (err < 0) {
1060		printf("Couldn't create the events: %s\n",
1061		       str_error_r(errno, sbuf, sizeof(sbuf)));
1062		goto out;
1063	}
1064
1065	if (evlist__mmap(evlist, kvm->opts.mmap_pages) < 0) {
1066		ui__error("Failed to mmap the events: %s\n",
1067			  str_error_r(errno, sbuf, sizeof(sbuf)));
1068		evlist__close(evlist);
1069		goto out;
1070	}
1071
1072	rc = 0;
1073
1074out:
1075	return rc;
1076}
1077#endif
1078
1079static int read_events(struct perf_kvm_stat *kvm)
1080{
1081	int ret;
1082
1083	struct perf_tool eops = {
1084		.sample			= process_sample_event,
1085		.comm			= perf_event__process_comm,
1086		.namespaces		= perf_event__process_namespaces,
1087		.ordered_events		= true,
1088	};
1089	struct perf_data file = {
1090		.path  = kvm->file_name,
1091		.mode  = PERF_DATA_MODE_READ,
1092		.force = kvm->force,
1093	};
1094
1095	kvm->tool = eops;
1096	kvm->session = perf_session__new(&file, false, &kvm->tool);
1097	if (IS_ERR(kvm->session)) {
1098		pr_err("Initializing perf session failed\n");
1099		return PTR_ERR(kvm->session);
1100	}
1101
1102	symbol__init(&kvm->session->header.env);
1103
1104	if (!perf_session__has_traces(kvm->session, "kvm record")) {
1105		ret = -EINVAL;
1106		goto out_delete;
1107	}
1108
1109	/*
1110	 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1111	 * traced in the old kernel.
1112	 */
1113	ret = cpu_isa_config(kvm);
1114	if (ret < 0)
1115		goto out_delete;
1116
1117	ret = perf_session__process_events(kvm->session);
1118
1119out_delete:
1120	perf_session__delete(kvm->session);
1121	return ret;
1122}
1123
1124static int parse_target_str(struct perf_kvm_stat *kvm)
1125{
1126	if (kvm->opts.target.pid) {
1127		kvm->pid_list = intlist__new(kvm->opts.target.pid);
1128		if (kvm->pid_list == NULL) {
1129			pr_err("Error parsing process id string\n");
1130			return -EINVAL;
1131		}
1132	}
1133
1134	return 0;
1135}
1136
1137static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1138{
1139	int ret = -EINVAL;
1140	int vcpu = kvm->trace_vcpu;
1141
1142	if (parse_target_str(kvm) != 0)
1143		goto exit;
1144
1145	if (!verify_vcpu(vcpu))
1146		goto exit;
1147
1148	if (!select_key(kvm))
1149		goto exit;
1150
1151	if (!register_kvm_events_ops(kvm))
1152		goto exit;
1153
1154	init_kvm_event_record(kvm);
1155	setup_pager();
1156
1157	ret = read_events(kvm);
1158	if (ret)
1159		goto exit;
1160
1161	sort_result(kvm);
1162	print_result(kvm);
1163
1164exit:
1165	return ret;
1166}
1167
1168#define STRDUP_FAIL_EXIT(s)		\
1169	({	char *_p;		\
1170	_p = strdup(s);		\
1171		if (!_p)		\
1172			return -ENOMEM;	\
1173		_p;			\
1174	})
1175
1176int __weak setup_kvm_events_tp(struct perf_kvm_stat *kvm __maybe_unused)
1177{
1178	return 0;
1179}
1180
1181static int
1182kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1183{
1184	unsigned int rec_argc, i, j, events_tp_size;
1185	const char **rec_argv;
1186	const char * const record_args[] = {
1187		"record",
1188		"-R",
1189		"-m", "1024",
1190		"-c", "1",
1191	};
1192	const char * const kvm_stat_record_usage[] = {
1193		"perf kvm stat record [<options>]",
1194		NULL
1195	};
1196	const char * const *events_tp;
1197	int ret;
1198
1199	events_tp_size = 0;
1200	ret = setup_kvm_events_tp(kvm);
1201	if (ret < 0) {
1202		pr_err("Unable to setup the kvm tracepoints\n");
1203		return ret;
1204	}
1205
1206	for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1207		events_tp_size++;
1208
1209	rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1210		   2 * events_tp_size;
1211	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1212
1213	if (rec_argv == NULL)
1214		return -ENOMEM;
1215
1216	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1217		rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1218
1219	for (j = 0; j < events_tp_size; j++) {
1220		rec_argv[i++] = "-e";
1221		rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1222	}
1223
1224	rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1225	rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1226
1227	for (j = 1; j < (unsigned int)argc; j++, i++)
1228		rec_argv[i] = argv[j];
1229
1230	set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1231	set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1232	set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1233
1234	set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1235	set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1236	set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1237	set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1238	set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1239	set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1240	set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1241	set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1242	set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1243	set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1244	set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1245	set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1246	set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1247	set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1248	set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1249
1250	record_usage = kvm_stat_record_usage;
1251	return cmd_record(i, rec_argv);
1252}
1253
1254static int
1255kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1256{
1257	const struct option kvm_events_report_options[] = {
1258		OPT_STRING(0, "event", &kvm->report_event, "report event",
1259			   "event for reporting: vmexit, "
1260			   "mmio (x86 only), ioport (x86 only)"),
1261		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1262			    "vcpu id to report"),
1263		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1264			    "key for sorting: sample(sort by samples number)"
1265			    " time (sort by avg time)"),
1266		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1267			   "analyze events only for given process id(s)"),
1268		OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
1269		OPT_END()
1270	};
1271
1272	const char * const kvm_events_report_usage[] = {
1273		"perf kvm stat report [<options>]",
1274		NULL
1275	};
1276
1277	if (argc) {
1278		argc = parse_options(argc, argv,
1279				     kvm_events_report_options,
1280				     kvm_events_report_usage, 0);
1281		if (argc)
1282			usage_with_options(kvm_events_report_usage,
1283					   kvm_events_report_options);
1284	}
1285
1286	if (!kvm->opts.target.pid)
1287		kvm->opts.target.system_wide = true;
1288
1289	return kvm_events_report_vcpu(kvm);
1290}
1291
1292#ifdef HAVE_TIMERFD_SUPPORT
1293static struct evlist *kvm_live_event_list(void)
1294{
1295	struct evlist *evlist;
1296	char *tp, *name, *sys;
1297	int err = -1;
1298	const char * const *events_tp;
1299
1300	evlist = evlist__new();
1301	if (evlist == NULL)
1302		return NULL;
1303
1304	for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1305
1306		tp = strdup(*events_tp);
1307		if (tp == NULL)
1308			goto out;
1309
1310		/* split tracepoint into subsystem and name */
1311		sys = tp;
1312		name = strchr(tp, ':');
1313		if (name == NULL) {
1314			pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1315			       *events_tp);
1316			free(tp);
1317			goto out;
1318		}
1319		*name = '\0';
1320		name++;
1321
1322		if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1323			pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1324			free(tp);
1325			goto out;
1326		}
1327
1328		free(tp);
1329	}
1330
1331	err = 0;
1332
1333out:
1334	if (err) {
1335		evlist__delete(evlist);
1336		evlist = NULL;
1337	}
1338
1339	return evlist;
1340}
1341
1342static int kvm_events_live(struct perf_kvm_stat *kvm,
1343			   int argc, const char **argv)
1344{
1345	char errbuf[BUFSIZ];
1346	int err;
1347
1348	const struct option live_options[] = {
1349		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1350			"record events on existing process id"),
1351		OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1352			"number of mmap data pages",
1353			perf_evlist__parse_mmap_pages),
1354		OPT_INCR('v', "verbose", &verbose,
1355			"be more verbose (show counter open errors, etc)"),
1356		OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1357			"system-wide collection from all CPUs"),
1358		OPT_UINTEGER('d', "display", &kvm->display_time,
1359			"time in seconds between display updates"),
1360		OPT_STRING(0, "event", &kvm->report_event, "report event",
1361			"event for reporting: "
1362			"vmexit, mmio (x86 only), ioport (x86 only)"),
1363		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1364			"vcpu id to report"),
1365		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1366			"key for sorting: sample(sort by samples number)"
1367			" time (sort by avg time)"),
1368		OPT_U64(0, "duration", &kvm->duration,
1369			"show events other than"
1370			" HLT (x86 only) or Wait state (s390 only)"
1371			" that take longer than duration usecs"),
1372		OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1373				"per thread proc mmap processing timeout in ms"),
1374		OPT_END()
1375	};
1376	const char * const live_usage[] = {
1377		"perf kvm stat live [<options>]",
1378		NULL
1379	};
1380	struct perf_data data = {
1381		.mode = PERF_DATA_MODE_WRITE,
1382	};
1383
1384
1385	/* event handling */
1386	kvm->tool.sample = process_sample_event;
1387	kvm->tool.comm   = perf_event__process_comm;
1388	kvm->tool.exit   = perf_event__process_exit;
1389	kvm->tool.fork   = perf_event__process_fork;
1390	kvm->tool.lost   = process_lost_event;
1391	kvm->tool.namespaces  = perf_event__process_namespaces;
1392	kvm->tool.ordered_events = true;
1393	perf_tool__fill_defaults(&kvm->tool);
1394
1395	/* set defaults */
1396	kvm->display_time = 1;
1397	kvm->opts.user_interval = 1;
1398	kvm->opts.mmap_pages = 512;
1399	kvm->opts.target.uses_mmap = false;
1400	kvm->opts.target.uid_str = NULL;
1401	kvm->opts.target.uid = UINT_MAX;
 
1402
1403	symbol__init(NULL);
1404	disable_buildid_cache();
1405
1406	use_browser = 0;
1407
1408	if (argc) {
1409		argc = parse_options(argc, argv, live_options,
1410				     live_usage, 0);
1411		if (argc)
1412			usage_with_options(live_usage, live_options);
1413	}
1414
1415	kvm->duration *= NSEC_PER_USEC;   /* convert usec to nsec */
1416
1417	/*
1418	 * target related setups
1419	 */
1420	err = target__validate(&kvm->opts.target);
1421	if (err) {
1422		target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1423		ui__warning("%s", errbuf);
1424	}
1425
1426	if (target__none(&kvm->opts.target))
1427		kvm->opts.target.system_wide = true;
1428
1429
1430	/*
1431	 * generate the event list
1432	 */
1433	err = setup_kvm_events_tp(kvm);
1434	if (err < 0) {
1435		pr_err("Unable to setup the kvm tracepoints\n");
1436		return err;
1437	}
1438
1439	kvm->evlist = kvm_live_event_list();
1440	if (kvm->evlist == NULL) {
1441		err = -1;
1442		goto out;
1443	}
1444
 
 
1445	if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1446		usage_with_options(live_usage, live_options);
1447
1448	/*
1449	 * perf session
1450	 */
1451	kvm->session = perf_session__new(&data, false, &kvm->tool);
1452	if (IS_ERR(kvm->session)) {
1453		err = PTR_ERR(kvm->session);
1454		goto out;
1455	}
1456	kvm->session->evlist = kvm->evlist;
1457	perf_session__set_id_hdr_size(kvm->session);
1458	ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
1459	machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1460				    kvm->evlist->core.threads, false, 1);
1461	err = kvm_live_open_events(kvm);
1462	if (err)
1463		goto out;
1464
1465	err = kvm_events_live_report(kvm);
1466
1467out:
1468	perf_session__delete(kvm->session);
 
1469	kvm->session = NULL;
1470	evlist__delete(kvm->evlist);
 
1471
1472	return err;
1473}
1474#endif
1475
1476static void print_kvm_stat_usage(void)
1477{
1478	printf("Usage: perf kvm stat <command>\n\n");
1479
1480	printf("# Available commands:\n");
1481	printf("\trecord: record kvm events\n");
1482	printf("\treport: report statistical data of kvm events\n");
1483	printf("\tlive:   live reporting of statistical data of kvm events\n");
1484
1485	printf("\nOtherwise, it is the alias of 'perf stat':\n");
1486}
1487
1488static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1489{
1490	struct perf_kvm_stat kvm = {
1491		.file_name = file_name,
1492
1493		.trace_vcpu	= -1,
1494		.report_event	= "vmexit",
1495		.sort_key	= "sample",
1496
1497	};
1498
1499	if (argc == 1) {
1500		print_kvm_stat_usage();
1501		goto perf_stat;
1502	}
1503
1504	if (!strncmp(argv[1], "rec", 3))
1505		return kvm_events_record(&kvm, argc - 1, argv + 1);
1506
1507	if (!strncmp(argv[1], "rep", 3))
1508		return kvm_events_report(&kvm, argc - 1 , argv + 1);
1509
1510#ifdef HAVE_TIMERFD_SUPPORT
1511	if (!strncmp(argv[1], "live", 4))
1512		return kvm_events_live(&kvm, argc - 1 , argv + 1);
1513#endif
1514
1515perf_stat:
1516	return cmd_stat(argc, argv);
1517}
1518#endif /* HAVE_KVM_STAT_SUPPORT */
1519
1520int __weak kvm_add_default_arch_event(int *argc __maybe_unused,
1521					const char **argv __maybe_unused)
1522{
1523	return 0;
1524}
1525
1526static int __cmd_record(const char *file_name, int argc, const char **argv)
1527{
1528	int rec_argc, i = 0, j, ret;
1529	const char **rec_argv;
1530
1531	ret = kvm_add_default_arch_event(&argc, argv);
1532	if (ret)
1533		return -EINVAL;
1534
1535	rec_argc = argc + 2;
1536	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1537	rec_argv[i++] = strdup("record");
1538	rec_argv[i++] = strdup("-o");
1539	rec_argv[i++] = strdup(file_name);
1540	for (j = 1; j < argc; j++, i++)
1541		rec_argv[i] = argv[j];
1542
1543	BUG_ON(i != rec_argc);
1544
1545	return cmd_record(i, rec_argv);
1546}
1547
1548static int __cmd_report(const char *file_name, int argc, const char **argv)
1549{
1550	int rec_argc, i = 0, j;
1551	const char **rec_argv;
1552
1553	rec_argc = argc + 2;
1554	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1555	rec_argv[i++] = strdup("report");
1556	rec_argv[i++] = strdup("-i");
1557	rec_argv[i++] = strdup(file_name);
1558	for (j = 1; j < argc; j++, i++)
1559		rec_argv[i] = argv[j];
1560
1561	BUG_ON(i != rec_argc);
1562
1563	return cmd_report(i, rec_argv);
1564}
1565
1566static int
1567__cmd_buildid_list(const char *file_name, int argc, const char **argv)
1568{
1569	int rec_argc, i = 0, j;
1570	const char **rec_argv;
1571
1572	rec_argc = argc + 2;
1573	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1574	rec_argv[i++] = strdup("buildid-list");
1575	rec_argv[i++] = strdup("-i");
1576	rec_argv[i++] = strdup(file_name);
1577	for (j = 1; j < argc; j++, i++)
1578		rec_argv[i] = argv[j];
1579
1580	BUG_ON(i != rec_argc);
1581
1582	return cmd_buildid_list(i, rec_argv);
1583}
1584
1585int cmd_kvm(int argc, const char **argv)
1586{
1587	const char *file_name = NULL;
1588	const struct option kvm_options[] = {
1589		OPT_STRING('i', "input", &file_name, "file",
1590			   "Input file name"),
1591		OPT_STRING('o', "output", &file_name, "file",
1592			   "Output file name"),
1593		OPT_BOOLEAN(0, "guest", &perf_guest,
1594			    "Collect guest os data"),
1595		OPT_BOOLEAN(0, "host", &perf_host,
1596			    "Collect host os data"),
1597		OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1598			   "guest mount directory under which every guest os"
1599			   " instance has a subdir"),
1600		OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1601			   "file", "file saving guest os vmlinux"),
1602		OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1603			   "file", "file saving guest os /proc/kallsyms"),
1604		OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1605			   "file", "file saving guest os /proc/modules"),
1606		OPT_INCR('v', "verbose", &verbose,
1607			    "be more verbose (show counter open errors, etc)"),
1608		OPT_END()
1609	};
1610
1611	const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1612						"buildid-list", "stat", NULL };
1613	const char *kvm_usage[] = { NULL, NULL };
1614
1615	perf_host  = 0;
1616	perf_guest = 1;
1617
1618	argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1619					PARSE_OPT_STOP_AT_NON_OPTION);
1620	if (!argc)
1621		usage_with_options(kvm_usage, kvm_options);
1622
1623	if (!perf_host)
1624		perf_guest = 1;
1625
1626	if (!file_name) {
1627		file_name = get_filename_for_perf_kvm();
1628
1629		if (!file_name) {
1630			pr_err("Failed to allocate memory for filename\n");
1631			return -ENOMEM;
1632		}
1633	}
1634
1635	if (!strncmp(argv[0], "rec", 3))
1636		return __cmd_record(file_name, argc, argv);
1637	else if (!strncmp(argv[0], "rep", 3))
1638		return __cmd_report(file_name, argc, argv);
1639	else if (!strncmp(argv[0], "diff", 4))
1640		return cmd_diff(argc, argv);
1641	else if (!strncmp(argv[0], "top", 3))
1642		return cmd_top(argc, argv);
1643	else if (!strncmp(argv[0], "buildid-list", 12))
1644		return __cmd_buildid_list(file_name, argc, argv);
1645#ifdef HAVE_KVM_STAT_SUPPORT
1646	else if (!strncmp(argv[0], "stat", 4))
1647		return kvm_cmd_stat(file_name, argc, argv);
1648#endif
1649	else
1650		usage_with_options(kvm_usage, kvm_options);
1651
1652	return 0;
1653}