Linux Audio

Check our new training course

Loading...
v6.2
   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#include "util/string2.h"
  28
  29#include <sys/prctl.h>
  30#ifdef HAVE_TIMERFD_SUPPORT
  31#include <sys/timerfd.h>
  32#endif
  33#include <sys/time.h>
  34#include <sys/types.h>
  35#include <sys/stat.h>
  36#include <fcntl.h>
  37
  38#include <linux/err.h>
  39#include <linux/kernel.h>
  40#include <linux/string.h>
  41#include <linux/time64.h>
  42#include <linux/zalloc.h>
  43#include <errno.h>
  44#include <inttypes.h>
  45#include <poll.h>
  46#include <termios.h>
  47#include <semaphore.h>
  48#include <signal.h>
  49#include <math.h>
  50#include <perf/mmap.h>
  51
  52static const char *get_filename_for_perf_kvm(void)
  53{
  54	const char *filename;
  55
  56	if (perf_host && !perf_guest)
  57		filename = strdup("perf.data.host");
  58	else if (!perf_host && perf_guest)
  59		filename = strdup("perf.data.guest");
  60	else
  61		filename = strdup("perf.data.kvm");
  62
  63	return filename;
  64}
  65
  66#if defined(HAVE_KVM_STAT_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
 
  67
  68void exit_event_get_key(struct evsel *evsel,
  69			struct perf_sample *sample,
  70			struct event_key *key)
  71{
  72	key->info = 0;
  73	key->key  = evsel__intval(evsel, sample, kvm_exit_reason);
  74}
  75
  76bool kvm_exit_event(struct evsel *evsel)
  77{
  78	return !strcmp(evsel->name, kvm_exit_trace);
  79}
  80
  81bool exit_event_begin(struct evsel *evsel,
  82		      struct perf_sample *sample, struct event_key *key)
  83{
  84	if (kvm_exit_event(evsel)) {
  85		exit_event_get_key(evsel, sample, key);
  86		return true;
  87	}
  88
  89	return false;
  90}
  91
  92bool kvm_entry_event(struct evsel *evsel)
  93{
  94	return !strcmp(evsel->name, kvm_entry_trace);
  95}
  96
  97bool exit_event_end(struct evsel *evsel,
  98		    struct perf_sample *sample __maybe_unused,
  99		    struct event_key *key __maybe_unused)
 100{
 101	return kvm_entry_event(evsel);
 102}
 103
 104static const char *get_exit_reason(struct perf_kvm_stat *kvm,
 105				   struct exit_reasons_table *tbl,
 106				   u64 exit_code)
 107{
 108	while (tbl->reason != NULL) {
 109		if (tbl->exit_code == exit_code)
 110			return tbl->reason;
 111		tbl++;
 112	}
 113
 114	pr_err("unknown kvm exit code:%lld on %s\n",
 115		(unsigned long long)exit_code, kvm->exit_reasons_isa);
 116	return "UNKNOWN";
 117}
 118
 119void exit_event_decode_key(struct perf_kvm_stat *kvm,
 120			   struct event_key *key,
 121			   char *decode)
 122{
 123	const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
 124						  key->key);
 125
 126	scnprintf(decode, decode_str_len, "%s", exit_reason);
 127}
 128
 129static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
 130{
 131	struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
 132
 133	for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
 134		if (!strcmp(events_ops->name, kvm->report_event)) {
 135			kvm->events_ops = events_ops->ops;
 136			return true;
 137		}
 138	}
 139
 140	return false;
 141}
 142
 143struct vcpu_event_record {
 144	int vcpu_id;
 145	u64 start_time;
 146	struct kvm_event *last_event;
 147};
 148
 149
 150static void init_kvm_event_record(struct perf_kvm_stat *kvm)
 151{
 152	unsigned int i;
 153
 154	for (i = 0; i < EVENTS_CACHE_SIZE; i++)
 155		INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
 156}
 157
 158#ifdef HAVE_TIMERFD_SUPPORT
 159static void clear_events_cache_stats(struct list_head *kvm_events_cache)
 160{
 161	struct list_head *head;
 162	struct kvm_event *event;
 163	unsigned int i;
 164	int j;
 165
 166	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
 167		head = &kvm_events_cache[i];
 168		list_for_each_entry(event, head, hash_entry) {
 169			/* reset stats for event */
 170			event->total.time = 0;
 171			init_stats(&event->total.stats);
 172
 173			for (j = 0; j < event->max_vcpu; ++j) {
 174				event->vcpu[j].time = 0;
 175				init_stats(&event->vcpu[j].stats);
 176			}
 177		}
 178	}
 179}
 180#endif
 181
 182static int kvm_events_hash_fn(u64 key)
 183{
 184	return key & (EVENTS_CACHE_SIZE - 1);
 185}
 186
 187static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
 188{
 189	int old_max_vcpu = event->max_vcpu;
 190	void *prev;
 191
 192	if (vcpu_id < event->max_vcpu)
 193		return true;
 194
 195	while (event->max_vcpu <= vcpu_id)
 196		event->max_vcpu += DEFAULT_VCPU_NUM;
 197
 198	prev = event->vcpu;
 199	event->vcpu = realloc(event->vcpu,
 200			      event->max_vcpu * sizeof(*event->vcpu));
 201	if (!event->vcpu) {
 202		free(prev);
 203		pr_err("Not enough memory\n");
 204		return false;
 205	}
 206
 207	memset(event->vcpu + old_max_vcpu, 0,
 208	       (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
 209	return true;
 210}
 211
 212static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
 213{
 214	struct kvm_event *event;
 215
 216	event = zalloc(sizeof(*event));
 217	if (!event) {
 218		pr_err("Not enough memory\n");
 219		return NULL;
 220	}
 221
 222	event->key = *key;
 223	init_stats(&event->total.stats);
 224	return event;
 225}
 226
 227static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
 228					       struct event_key *key)
 229{
 230	struct kvm_event *event;
 231	struct list_head *head;
 232
 233	BUG_ON(key->key == INVALID_KEY);
 234
 235	head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
 236	list_for_each_entry(event, head, hash_entry) {
 237		if (event->key.key == key->key && event->key.info == key->info)
 238			return event;
 239	}
 240
 241	event = kvm_alloc_init_event(key);
 242	if (!event)
 243		return NULL;
 244
 245	list_add(&event->hash_entry, head);
 246	return event;
 247}
 248
 249static bool handle_begin_event(struct perf_kvm_stat *kvm,
 250			       struct vcpu_event_record *vcpu_record,
 251			       struct event_key *key, u64 timestamp)
 252{
 253	struct kvm_event *event = NULL;
 254
 255	if (key->key != INVALID_KEY)
 256		event = find_create_kvm_event(kvm, key);
 257
 258	vcpu_record->last_event = event;
 259	vcpu_record->start_time = timestamp;
 260	return true;
 261}
 262
 263static void
 264kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
 265{
 266	kvm_stats->time += time_diff;
 267	update_stats(&kvm_stats->stats, time_diff);
 268}
 269
 270static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
 271{
 272	struct kvm_event_stats *kvm_stats = &event->total;
 273
 274	if (vcpu_id != -1)
 275		kvm_stats = &event->vcpu[vcpu_id];
 276
 277	return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
 278				avg_stats(&kvm_stats->stats));
 279}
 280
 281static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
 282			     u64 time_diff)
 283{
 284	if (vcpu_id == -1) {
 285		kvm_update_event_stats(&event->total, time_diff);
 286		return true;
 287	}
 288
 289	if (!kvm_event_expand(event, vcpu_id))
 290		return false;
 291
 292	kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
 293	return true;
 294}
 295
 296static bool is_child_event(struct perf_kvm_stat *kvm,
 297			   struct evsel *evsel,
 298			   struct perf_sample *sample,
 299			   struct event_key *key)
 300{
 301	struct child_event_ops *child_ops;
 302
 303	child_ops = kvm->events_ops->child_ops;
 304
 305	if (!child_ops)
 306		return false;
 307
 308	for (; child_ops->name; child_ops++) {
 309		if (!strcmp(evsel->name, child_ops->name)) {
 310			child_ops->get_key(evsel, sample, key);
 311			return true;
 312		}
 313	}
 314
 315	return false;
 316}
 317
 318static bool handle_child_event(struct perf_kvm_stat *kvm,
 319			       struct vcpu_event_record *vcpu_record,
 320			       struct event_key *key,
 321			       struct perf_sample *sample __maybe_unused)
 322{
 323	struct kvm_event *event = NULL;
 324
 325	if (key->key != INVALID_KEY)
 326		event = find_create_kvm_event(kvm, key);
 327
 328	vcpu_record->last_event = event;
 329
 330	return true;
 331}
 332
 333static bool skip_event(const char *event)
 334{
 335	const char * const *skip_events;
 336
 337	for (skip_events = kvm_skip_events; *skip_events; skip_events++)
 338		if (!strcmp(event, *skip_events))
 339			return true;
 340
 341	return false;
 342}
 343
 344static bool handle_end_event(struct perf_kvm_stat *kvm,
 345			     struct vcpu_event_record *vcpu_record,
 346			     struct event_key *key,
 347			     struct perf_sample *sample)
 348{
 349	struct kvm_event *event;
 350	u64 time_begin, time_diff;
 351	int vcpu;
 352
 353	if (kvm->trace_vcpu == -1)
 354		vcpu = -1;
 355	else
 356		vcpu = vcpu_record->vcpu_id;
 357
 358	event = vcpu_record->last_event;
 359	time_begin = vcpu_record->start_time;
 360
 361	/* The begin event is not caught. */
 362	if (!time_begin)
 363		return true;
 364
 365	/*
 366	 * In some case, the 'begin event' only records the start timestamp,
 367	 * the actual event is recognized in the 'end event' (e.g. mmio-event).
 368	 */
 369
 370	/* Both begin and end events did not get the key. */
 371	if (!event && key->key == INVALID_KEY)
 372		return true;
 373
 374	if (!event)
 375		event = find_create_kvm_event(kvm, key);
 376
 377	if (!event)
 378		return false;
 379
 380	vcpu_record->last_event = NULL;
 381	vcpu_record->start_time = 0;
 382
 383	/* seems to happen once in a while during live mode */
 384	if (sample->time < time_begin) {
 385		pr_debug("End time before begin time; skipping event.\n");
 386		return true;
 387	}
 388
 389	time_diff = sample->time - time_begin;
 390
 391	if (kvm->duration && time_diff > kvm->duration) {
 392		char decode[decode_str_len];
 393
 394		kvm->events_ops->decode_key(kvm, &event->key, decode);
 395		if (!skip_event(decode)) {
 396			pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
 397				 sample->time, sample->pid, vcpu_record->vcpu_id,
 398				 decode, time_diff / NSEC_PER_USEC);
 399		}
 400	}
 401
 402	return update_kvm_event(event, vcpu, time_diff);
 403}
 404
 405static
 406struct vcpu_event_record *per_vcpu_record(struct thread *thread,
 407					  struct evsel *evsel,
 408					  struct perf_sample *sample)
 409{
 410	/* Only kvm_entry records vcpu id. */
 411	if (!thread__priv(thread) && kvm_entry_event(evsel)) {
 412		struct vcpu_event_record *vcpu_record;
 413
 414		vcpu_record = zalloc(sizeof(*vcpu_record));
 415		if (!vcpu_record) {
 416			pr_err("%s: Not enough memory\n", __func__);
 417			return NULL;
 418		}
 419
 420		vcpu_record->vcpu_id = evsel__intval(evsel, sample, vcpu_id_str);
 
 421		thread__set_priv(thread, vcpu_record);
 422	}
 423
 424	return thread__priv(thread);
 425}
 426
 427static bool handle_kvm_event(struct perf_kvm_stat *kvm,
 428			     struct thread *thread,
 429			     struct evsel *evsel,
 430			     struct perf_sample *sample)
 431{
 432	struct vcpu_event_record *vcpu_record;
 433	struct event_key key = { .key = INVALID_KEY,
 434				 .exit_reasons = kvm->exit_reasons };
 435
 436	vcpu_record = per_vcpu_record(thread, evsel, sample);
 437	if (!vcpu_record)
 438		return true;
 439
 440	/* only process events for vcpus user cares about */
 441	if ((kvm->trace_vcpu != -1) &&
 442	    (kvm->trace_vcpu != vcpu_record->vcpu_id))
 443		return true;
 444
 445	if (kvm->events_ops->is_begin_event(evsel, sample, &key))
 446		return handle_begin_event(kvm, vcpu_record, &key, sample->time);
 447
 448	if (is_child_event(kvm, evsel, sample, &key))
 449		return handle_child_event(kvm, vcpu_record, &key, sample);
 450
 451	if (kvm->events_ops->is_end_event(evsel, sample, &key))
 452		return handle_end_event(kvm, vcpu_record, &key, sample);
 453
 454	return true;
 455}
 456
 457#define GET_EVENT_KEY(func, field)					\
 458static u64 get_event_ ##func(struct kvm_event *event, int vcpu)		\
 459{									\
 460	if (vcpu == -1)							\
 461		return event->total.field;				\
 462									\
 463	if (vcpu >= event->max_vcpu)					\
 464		return 0;						\
 465									\
 466	return event->vcpu[vcpu].field;					\
 467}
 468
 469#define COMPARE_EVENT_KEY(func, field)					\
 470GET_EVENT_KEY(func, field)						\
 471static int compare_kvm_event_ ## func(struct kvm_event *one,		\
 472					struct kvm_event *two, int vcpu)\
 473{									\
 474	return get_event_ ##func(one, vcpu) >				\
 475				get_event_ ##func(two, vcpu);		\
 476}
 477
 478GET_EVENT_KEY(time, time);
 479COMPARE_EVENT_KEY(count, stats.n);
 480COMPARE_EVENT_KEY(mean, stats.mean);
 481GET_EVENT_KEY(max, stats.max);
 482GET_EVENT_KEY(min, stats.min);
 483
 484#define DEF_SORT_NAME_KEY(name, compare_key)				\
 485	{ #name, compare_kvm_event_ ## compare_key }
 486
 487static struct kvm_event_key keys[] = {
 488	DEF_SORT_NAME_KEY(sample, count),
 489	DEF_SORT_NAME_KEY(time, mean),
 490	{ NULL, NULL }
 491};
 492
 493static bool select_key(struct perf_kvm_stat *kvm)
 494{
 495	int i;
 496
 497	for (i = 0; keys[i].name; i++) {
 498		if (!strcmp(keys[i].name, kvm->sort_key)) {
 499			kvm->compare = keys[i].key;
 500			return true;
 501		}
 502	}
 503
 504	pr_err("Unknown compare key:%s\n", kvm->sort_key);
 505	return false;
 506}
 507
 508static void insert_to_result(struct rb_root *result, struct kvm_event *event,
 509			     key_cmp_fun bigger, int vcpu)
 510{
 511	struct rb_node **rb = &result->rb_node;
 512	struct rb_node *parent = NULL;
 513	struct kvm_event *p;
 514
 515	while (*rb) {
 516		p = container_of(*rb, struct kvm_event, rb);
 517		parent = *rb;
 518
 519		if (bigger(event, p, vcpu))
 520			rb = &(*rb)->rb_left;
 521		else
 522			rb = &(*rb)->rb_right;
 523	}
 524
 525	rb_link_node(&event->rb, parent, rb);
 526	rb_insert_color(&event->rb, result);
 527}
 528
 529static void
 530update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
 531{
 532	int vcpu = kvm->trace_vcpu;
 533
 534	kvm->total_count += get_event_count(event, vcpu);
 535	kvm->total_time += get_event_time(event, vcpu);
 536}
 537
 538static bool event_is_valid(struct kvm_event *event, int vcpu)
 539{
 540	return !!get_event_count(event, vcpu);
 541}
 542
 543static void sort_result(struct perf_kvm_stat *kvm)
 544{
 545	unsigned int i;
 546	int vcpu = kvm->trace_vcpu;
 547	struct kvm_event *event;
 548
 549	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
 550		list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
 551			if (event_is_valid(event, vcpu)) {
 552				update_total_count(kvm, event);
 553				insert_to_result(&kvm->result, event,
 554						 kvm->compare, vcpu);
 555			}
 556		}
 557	}
 558}
 559
 560/* returns left most element of result, and erase it */
 561static struct kvm_event *pop_from_result(struct rb_root *result)
 562{
 563	struct rb_node *node = rb_first(result);
 564
 565	if (!node)
 566		return NULL;
 567
 568	rb_erase(node, result);
 569	return container_of(node, struct kvm_event, rb);
 570}
 571
 572static void print_vcpu_info(struct perf_kvm_stat *kvm)
 573{
 574	int vcpu = kvm->trace_vcpu;
 575
 576	pr_info("Analyze events for ");
 577
 578	if (kvm->opts.target.system_wide)
 579		pr_info("all VMs, ");
 580	else if (kvm->opts.target.pid)
 581		pr_info("pid(s) %s, ", kvm->opts.target.pid);
 582	else
 583		pr_info("dazed and confused on what is monitored, ");
 584
 585	if (vcpu == -1)
 586		pr_info("all VCPUs:\n\n");
 587	else
 588		pr_info("VCPU %d:\n\n", vcpu);
 589}
 590
 591static void show_timeofday(void)
 592{
 593	char date[64];
 594	struct timeval tv;
 595	struct tm ltime;
 596
 597	gettimeofday(&tv, NULL);
 598	if (localtime_r(&tv.tv_sec, &ltime)) {
 599		strftime(date, sizeof(date), "%H:%M:%S", &ltime);
 600		pr_info("%s.%06ld", date, tv.tv_usec);
 601	} else
 602		pr_info("00:00:00.000000");
 603
 604	return;
 605}
 606
 607static void print_result(struct perf_kvm_stat *kvm)
 608{
 609	char decode[decode_str_len];
 610	struct kvm_event *event;
 611	int vcpu = kvm->trace_vcpu;
 612
 613	if (kvm->live) {
 614		puts(CONSOLE_CLEAR);
 615		show_timeofday();
 616	}
 617
 618	pr_info("\n\n");
 619	print_vcpu_info(kvm);
 620	pr_info("%*s ", decode_str_len, kvm->events_ops->name);
 621	pr_info("%10s ", "Samples");
 622	pr_info("%9s ", "Samples%");
 623
 624	pr_info("%9s ", "Time%");
 625	pr_info("%11s ", "Min Time");
 626	pr_info("%11s ", "Max Time");
 627	pr_info("%16s ", "Avg time");
 628	pr_info("\n\n");
 629
 630	while ((event = pop_from_result(&kvm->result))) {
 631		u64 ecount, etime, max, min;
 632
 633		ecount = get_event_count(event, vcpu);
 634		etime = get_event_time(event, vcpu);
 635		max = get_event_max(event, vcpu);
 636		min = get_event_min(event, vcpu);
 637
 638		kvm->events_ops->decode_key(kvm, &event->key, decode);
 639		pr_info("%*s ", decode_str_len, decode);
 640		pr_info("%10llu ", (unsigned long long)ecount);
 641		pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
 642		pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
 643		pr_info("%9.2fus ", (double)min / NSEC_PER_USEC);
 644		pr_info("%9.2fus ", (double)max / NSEC_PER_USEC);
 645		pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount / NSEC_PER_USEC,
 646			kvm_event_rel_stddev(vcpu, event));
 647		pr_info("\n");
 648	}
 649
 650	pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
 651		kvm->total_count, kvm->total_time / (double)NSEC_PER_USEC);
 652
 653	if (kvm->lost_events)
 654		pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
 655}
 656
 657#if defined(HAVE_TIMERFD_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
 658static int process_lost_event(struct perf_tool *tool,
 659			      union perf_event *event __maybe_unused,
 660			      struct perf_sample *sample __maybe_unused,
 661			      struct machine *machine __maybe_unused)
 662{
 663	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
 664
 665	kvm->lost_events++;
 666	return 0;
 667}
 668#endif
 669
 670static bool skip_sample(struct perf_kvm_stat *kvm,
 671			struct perf_sample *sample)
 672{
 673	if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
 674		return true;
 675
 676	return false;
 677}
 678
 679static int process_sample_event(struct perf_tool *tool,
 680				union perf_event *event,
 681				struct perf_sample *sample,
 682				struct evsel *evsel,
 683				struct machine *machine)
 684{
 685	int err = 0;
 686	struct thread *thread;
 687	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
 688						 tool);
 689
 690	if (skip_sample(kvm, sample))
 691		return 0;
 692
 693	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
 694	if (thread == NULL) {
 695		pr_debug("problem processing %d event, skipping it.\n",
 696			event->header.type);
 697		return -1;
 698	}
 699
 700	if (!handle_kvm_event(kvm, thread, evsel, sample))
 701		err = -1;
 702
 703	thread__put(thread);
 704	return err;
 705}
 706
 707static int cpu_isa_config(struct perf_kvm_stat *kvm)
 708{
 709	char buf[128], *cpuid;
 710	int err;
 711
 712	if (kvm->live) {
 713		err = get_cpuid(buf, sizeof(buf));
 714		if (err != 0) {
 715			pr_err("Failed to look up CPU type: %s\n",
 716			       str_error_r(err, buf, sizeof(buf)));
 717			return -err;
 718		}
 719		cpuid = buf;
 720	} else
 721		cpuid = kvm->session->header.env.cpuid;
 722
 723	if (!cpuid) {
 724		pr_err("Failed to look up CPU type\n");
 725		return -EINVAL;
 726	}
 727
 728	err = cpu_isa_init(kvm, cpuid);
 729	if (err == -ENOTSUP)
 730		pr_err("CPU %s is not supported.\n", cpuid);
 731
 732	return err;
 733}
 734
 735static bool verify_vcpu(int vcpu)
 736{
 737	if (vcpu != -1 && vcpu < 0) {
 738		pr_err("Invalid vcpu:%d.\n", vcpu);
 739		return false;
 740	}
 741
 742	return true;
 743}
 744
 745#if defined(HAVE_TIMERFD_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
 746/* keeping the max events to a modest level to keep
 747 * the processing of samples per mmap smooth.
 748 */
 749#define PERF_KVM__MAX_EVENTS_PER_MMAP  25
 750
 751static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
 752				   u64 *mmap_time)
 753{
 754	struct evlist *evlist = kvm->evlist;
 755	union perf_event *event;
 756	struct mmap *md;
 757	u64 timestamp;
 758	s64 n = 0;
 759	int err;
 760
 761	*mmap_time = ULLONG_MAX;
 762	md = &evlist->mmap[idx];
 763	err = perf_mmap__read_init(&md->core);
 764	if (err < 0)
 765		return (err == -EAGAIN) ? 0 : -1;
 766
 767	while ((event = perf_mmap__read_event(&md->core)) != NULL) {
 768		err = evlist__parse_sample_timestamp(evlist, event, &timestamp);
 769		if (err) {
 770			perf_mmap__consume(&md->core);
 771			pr_err("Failed to parse sample\n");
 772			return -1;
 773		}
 774
 775		err = perf_session__queue_event(kvm->session, event, timestamp, 0, NULL);
 776		/*
 777		 * FIXME: Here we can't consume the event, as perf_session__queue_event will
 778		 *        point to it, and it'll get possibly overwritten by the kernel.
 779		 */
 780		perf_mmap__consume(&md->core);
 781
 782		if (err) {
 783			pr_err("Failed to enqueue sample: %d\n", err);
 784			return -1;
 785		}
 786
 787		/* save time stamp of our first sample for this mmap */
 788		if (n == 0)
 789			*mmap_time = timestamp;
 790
 791		/* limit events per mmap handled all at once */
 792		n++;
 793		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
 794			break;
 795	}
 796
 797	perf_mmap__read_done(&md->core);
 798	return n;
 799}
 800
 801static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
 802{
 803	int i, err, throttled = 0;
 804	s64 n, ntotal = 0;
 805	u64 flush_time = ULLONG_MAX, mmap_time;
 806
 807	for (i = 0; i < kvm->evlist->core.nr_mmaps; i++) {
 808		n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
 809		if (n < 0)
 810			return -1;
 811
 812		/* flush time is going to be the minimum of all the individual
 813		 * mmap times. Essentially, we flush all the samples queued up
 814		 * from the last pass under our minimal start time -- that leaves
 815		 * a very small race for samples to come in with a lower timestamp.
 816		 * The ioctl to return the perf_clock timestamp should close the
 817		 * race entirely.
 818		 */
 819		if (mmap_time < flush_time)
 820			flush_time = mmap_time;
 821
 822		ntotal += n;
 823		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
 824			throttled = 1;
 825	}
 826
 827	/* flush queue after each round in which we processed events */
 828	if (ntotal) {
 829		struct ordered_events *oe = &kvm->session->ordered_events;
 830
 831		oe->next_flush = flush_time;
 832		err = ordered_events__flush(oe, OE_FLUSH__ROUND);
 833		if (err) {
 834			if (kvm->lost_events)
 835				pr_info("\nLost events: %" PRIu64 "\n\n",
 836					kvm->lost_events);
 837			return err;
 838		}
 839	}
 840
 841	return throttled;
 842}
 843
 844static volatile int done;
 845
 846static void sig_handler(int sig __maybe_unused)
 847{
 848	done = 1;
 849}
 850
 851static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
 852{
 853	struct itimerspec new_value;
 854	int rc = -1;
 855
 856	kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
 857	if (kvm->timerfd < 0) {
 858		pr_err("timerfd_create failed\n");
 859		goto out;
 860	}
 861
 862	new_value.it_value.tv_sec = kvm->display_time;
 863	new_value.it_value.tv_nsec = 0;
 864	new_value.it_interval.tv_sec = kvm->display_time;
 865	new_value.it_interval.tv_nsec = 0;
 866
 867	if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
 868		pr_err("timerfd_settime failed: %d\n", errno);
 869		close(kvm->timerfd);
 870		goto out;
 871	}
 872
 873	rc = 0;
 874out:
 875	return rc;
 876}
 877
 878static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
 879{
 880	uint64_t c;
 881	int rc;
 882
 883	rc = read(kvm->timerfd, &c, sizeof(uint64_t));
 884	if (rc < 0) {
 885		if (errno == EAGAIN)
 886			return 0;
 887
 888		pr_err("Failed to read timer fd: %d\n", errno);
 889		return -1;
 890	}
 891
 892	if (rc != sizeof(uint64_t)) {
 893		pr_err("Error reading timer fd - invalid size returned\n");
 894		return -1;
 895	}
 896
 897	if (c != 1)
 898		pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
 899
 900	/* update display */
 901	sort_result(kvm);
 902	print_result(kvm);
 903
 904	/* reset counts */
 905	clear_events_cache_stats(kvm->kvm_events_cache);
 906	kvm->total_count = 0;
 907	kvm->total_time = 0;
 908	kvm->lost_events = 0;
 909
 910	return 0;
 911}
 912
 913static int fd_set_nonblock(int fd)
 914{
 915	long arg = 0;
 916
 917	arg = fcntl(fd, F_GETFL);
 918	if (arg < 0) {
 919		pr_err("Failed to get current flags for fd %d\n", fd);
 920		return -1;
 921	}
 922
 923	if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
 924		pr_err("Failed to set non-block option on fd %d\n", fd);
 925		return -1;
 926	}
 927
 928	return 0;
 929}
 930
 931static int perf_kvm__handle_stdin(void)
 932{
 933	int c;
 934
 935	c = getc(stdin);
 936	if (c == 'q')
 937		return 1;
 938
 939	return 0;
 940}
 941
 942static int kvm_events_live_report(struct perf_kvm_stat *kvm)
 943{
 944	int nr_stdin, ret, err = -EINVAL;
 945	struct termios save;
 946
 947	/* live flag must be set first */
 948	kvm->live = true;
 949
 950	ret = cpu_isa_config(kvm);
 951	if (ret < 0)
 952		return ret;
 953
 954	if (!verify_vcpu(kvm->trace_vcpu) ||
 955	    !select_key(kvm) ||
 956	    !register_kvm_events_ops(kvm)) {
 957		goto out;
 958	}
 959
 960	set_term_quiet_input(&save);
 961	init_kvm_event_record(kvm);
 962
 963	signal(SIGINT, sig_handler);
 964	signal(SIGTERM, sig_handler);
 965
 966	/* add timer fd */
 967	if (perf_kvm__timerfd_create(kvm) < 0) {
 968		err = -1;
 969		goto out;
 970	}
 971
 972	if (evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
 973		goto out;
 974
 975	nr_stdin = evlist__add_pollfd(kvm->evlist, fileno(stdin));
 976	if (nr_stdin < 0)
 977		goto out;
 978
 979	if (fd_set_nonblock(fileno(stdin)) != 0)
 980		goto out;
 981
 982	/* everything is good - enable the events and process */
 983	evlist__enable(kvm->evlist);
 984
 985	while (!done) {
 986		struct fdarray *fda = &kvm->evlist->core.pollfd;
 987		int rc;
 988
 989		rc = perf_kvm__mmap_read(kvm);
 990		if (rc < 0)
 991			break;
 992
 993		err = perf_kvm__handle_timerfd(kvm);
 994		if (err)
 995			goto out;
 996
 997		if (fda->entries[nr_stdin].revents & POLLIN)
 998			done = perf_kvm__handle_stdin();
 999
1000		if (!rc && !done)
1001			err = evlist__poll(kvm->evlist, 100);
1002	}
1003
1004	evlist__disable(kvm->evlist);
1005
1006	if (err == 0) {
1007		sort_result(kvm);
1008		print_result(kvm);
1009	}
1010
1011out:
1012	if (kvm->timerfd >= 0)
1013		close(kvm->timerfd);
1014
1015	tcsetattr(0, TCSAFLUSH, &save);
1016	return err;
1017}
1018
1019static int kvm_live_open_events(struct perf_kvm_stat *kvm)
1020{
1021	int err, rc = -1;
1022	struct evsel *pos;
1023	struct evlist *evlist = kvm->evlist;
1024	char sbuf[STRERR_BUFSIZE];
1025
1026	evlist__config(evlist, &kvm->opts, NULL);
1027
1028	/*
1029	 * Note: exclude_{guest,host} do not apply here.
1030	 *       This command processes KVM tracepoints from host only
1031	 */
1032	evlist__for_each_entry(evlist, pos) {
1033		struct perf_event_attr *attr = &pos->core.attr;
1034
1035		/* make sure these *are* set */
1036		evsel__set_sample_bit(pos, TID);
1037		evsel__set_sample_bit(pos, TIME);
1038		evsel__set_sample_bit(pos, CPU);
1039		evsel__set_sample_bit(pos, RAW);
1040		/* make sure these are *not*; want as small a sample as possible */
1041		evsel__reset_sample_bit(pos, PERIOD);
1042		evsel__reset_sample_bit(pos, IP);
1043		evsel__reset_sample_bit(pos, CALLCHAIN);
1044		evsel__reset_sample_bit(pos, ADDR);
1045		evsel__reset_sample_bit(pos, READ);
1046		attr->mmap = 0;
1047		attr->comm = 0;
1048		attr->task = 0;
1049
1050		attr->sample_period = 1;
1051
1052		attr->watermark = 0;
1053		attr->wakeup_events = 1000;
1054
1055		/* will enable all once we are ready */
1056		attr->disabled = 1;
1057	}
1058
1059	err = evlist__open(evlist);
1060	if (err < 0) {
1061		printf("Couldn't create the events: %s\n",
1062		       str_error_r(errno, sbuf, sizeof(sbuf)));
1063		goto out;
1064	}
1065
1066	if (evlist__mmap(evlist, kvm->opts.mmap_pages) < 0) {
1067		ui__error("Failed to mmap the events: %s\n",
1068			  str_error_r(errno, sbuf, sizeof(sbuf)));
1069		evlist__close(evlist);
1070		goto out;
1071	}
1072
1073	rc = 0;
1074
1075out:
1076	return rc;
1077}
1078#endif
1079
1080static int read_events(struct perf_kvm_stat *kvm)
1081{
1082	int ret;
1083
1084	struct perf_tool eops = {
1085		.sample			= process_sample_event,
1086		.comm			= perf_event__process_comm,
1087		.namespaces		= perf_event__process_namespaces,
1088		.ordered_events		= true,
1089	};
1090	struct perf_data file = {
1091		.path  = kvm->file_name,
1092		.mode  = PERF_DATA_MODE_READ,
1093		.force = kvm->force,
 
 
1094	};
1095
1096	kvm->tool = eops;
1097	kvm->session = perf_session__new(&file, &kvm->tool);
1098	if (IS_ERR(kvm->session)) {
1099		pr_err("Initializing perf session failed\n");
1100		return PTR_ERR(kvm->session);
1101	}
1102
1103	symbol__init(&kvm->session->header.env);
1104
1105	if (!perf_session__has_traces(kvm->session, "kvm record")) {
1106		ret = -EINVAL;
1107		goto out_delete;
1108	}
1109
1110	/*
1111	 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1112	 * traced in the old kernel.
1113	 */
1114	ret = cpu_isa_config(kvm);
1115	if (ret < 0)
1116		goto out_delete;
1117
1118	ret = perf_session__process_events(kvm->session);
1119
1120out_delete:
1121	perf_session__delete(kvm->session);
1122	return ret;
1123}
1124
1125static int parse_target_str(struct perf_kvm_stat *kvm)
1126{
1127	if (kvm->opts.target.pid) {
1128		kvm->pid_list = intlist__new(kvm->opts.target.pid);
1129		if (kvm->pid_list == NULL) {
1130			pr_err("Error parsing process id string\n");
1131			return -EINVAL;
1132		}
1133	}
1134
1135	return 0;
1136}
1137
1138static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1139{
1140	int ret = -EINVAL;
1141	int vcpu = kvm->trace_vcpu;
1142
1143	if (parse_target_str(kvm) != 0)
1144		goto exit;
1145
1146	if (!verify_vcpu(vcpu))
1147		goto exit;
1148
1149	if (!select_key(kvm))
1150		goto exit;
1151
1152	if (!register_kvm_events_ops(kvm))
1153		goto exit;
1154
1155	init_kvm_event_record(kvm);
1156	setup_pager();
1157
1158	ret = read_events(kvm);
1159	if (ret)
1160		goto exit;
1161
1162	sort_result(kvm);
1163	print_result(kvm);
1164
1165exit:
1166	return ret;
1167}
1168
1169#define STRDUP_FAIL_EXIT(s)		\
1170	({	char *_p;		\
1171	_p = strdup(s);		\
1172		if (!_p)		\
1173			return -ENOMEM;	\
1174		_p;			\
1175	})
1176
1177int __weak setup_kvm_events_tp(struct perf_kvm_stat *kvm __maybe_unused)
1178{
1179	return 0;
1180}
1181
1182static int
1183kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1184{
1185	unsigned int rec_argc, i, j, events_tp_size;
1186	const char **rec_argv;
1187	const char * const record_args[] = {
1188		"record",
1189		"-R",
1190		"-m", "1024",
1191		"-c", "1",
1192	};
1193	const char * const kvm_stat_record_usage[] = {
1194		"perf kvm stat record [<options>]",
1195		NULL
1196	};
1197	const char * const *events_tp;
1198	int ret;
1199
1200	events_tp_size = 0;
1201	ret = setup_kvm_events_tp(kvm);
1202	if (ret < 0) {
1203		pr_err("Unable to setup the kvm tracepoints\n");
1204		return ret;
1205	}
1206
1207	for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1208		events_tp_size++;
1209
1210	rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1211		   2 * events_tp_size;
1212	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1213
1214	if (rec_argv == NULL)
1215		return -ENOMEM;
1216
1217	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1218		rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1219
1220	for (j = 0; j < events_tp_size; j++) {
1221		rec_argv[i++] = "-e";
1222		rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1223	}
1224
1225	rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1226	rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1227
1228	for (j = 1; j < (unsigned int)argc; j++, i++)
1229		rec_argv[i] = argv[j];
1230
1231	set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1232	set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1233	set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1234
1235	set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1236	set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1237	set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1238	set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1239	set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1240	set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1241	set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1242	set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1243	set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1244	set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1245	set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1246	set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1247	set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1248	set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1249	set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1250
1251	record_usage = kvm_stat_record_usage;
1252	return cmd_record(i, rec_argv);
1253}
1254
1255static int
1256kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1257{
1258	const struct option kvm_events_report_options[] = {
1259		OPT_STRING(0, "event", &kvm->report_event, "report event",
1260			   "event for reporting: vmexit, "
1261			   "mmio (x86 only), ioport (x86 only)"),
1262		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1263			    "vcpu id to report"),
1264		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1265			    "key for sorting: sample(sort by samples number)"
1266			    " time (sort by avg time)"),
1267		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1268			   "analyze events only for given process id(s)"),
1269		OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
1270		OPT_END()
1271	};
1272
1273	const char * const kvm_events_report_usage[] = {
1274		"perf kvm stat report [<options>]",
1275		NULL
1276	};
1277
1278	if (argc) {
1279		argc = parse_options(argc, argv,
1280				     kvm_events_report_options,
1281				     kvm_events_report_usage, 0);
1282		if (argc)
1283			usage_with_options(kvm_events_report_usage,
1284					   kvm_events_report_options);
1285	}
1286
1287	if (!kvm->opts.target.pid)
1288		kvm->opts.target.system_wide = true;
1289
1290	return kvm_events_report_vcpu(kvm);
1291}
1292
1293#if defined(HAVE_TIMERFD_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
1294static struct evlist *kvm_live_event_list(void)
1295{
1296	struct evlist *evlist;
1297	char *tp, *name, *sys;
1298	int err = -1;
1299	const char * const *events_tp;
1300
1301	evlist = evlist__new();
1302	if (evlist == NULL)
1303		return NULL;
1304
1305	for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1306
1307		tp = strdup(*events_tp);
1308		if (tp == NULL)
1309			goto out;
1310
1311		/* split tracepoint into subsystem and name */
1312		sys = tp;
1313		name = strchr(tp, ':');
1314		if (name == NULL) {
1315			pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1316			       *events_tp);
1317			free(tp);
1318			goto out;
1319		}
1320		*name = '\0';
1321		name++;
1322
1323		if (evlist__add_newtp(evlist, sys, name, NULL)) {
1324			pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1325			free(tp);
1326			goto out;
1327		}
1328
1329		free(tp);
1330	}
1331
1332	err = 0;
1333
1334out:
1335	if (err) {
1336		evlist__delete(evlist);
1337		evlist = NULL;
1338	}
1339
1340	return evlist;
1341}
1342
1343static int kvm_events_live(struct perf_kvm_stat *kvm,
1344			   int argc, const char **argv)
1345{
1346	char errbuf[BUFSIZ];
1347	int err;
1348
1349	const struct option live_options[] = {
1350		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1351			"record events on existing process id"),
1352		OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1353			"number of mmap data pages", 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 (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, &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, true, 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 (strlen(argv[1]) > 2 && strstarts("record", argv[1]))
1505		return kvm_events_record(&kvm, argc - 1, argv + 1);
1506
1507	if (strlen(argv[1]) > 2 && strstarts("report", argv[1]))
1508		return kvm_events_report(&kvm, argc - 1 , argv + 1);
1509
1510#if defined(HAVE_TIMERFD_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
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_BOOLEAN(0, "guest-code", &symbol_conf.guest_code,
1607			    "Guest code can be found in hypervisor process"),
1608		OPT_INCR('v', "verbose", &verbose,
1609			    "be more verbose (show counter open errors, etc)"),
1610		OPT_END()
1611	};
1612
1613	const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1614						"buildid-list", "stat", NULL };
1615	const char *kvm_usage[] = { NULL, NULL };
1616
1617	perf_host  = 0;
1618	perf_guest = 1;
1619
1620	argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1621					PARSE_OPT_STOP_AT_NON_OPTION);
1622	if (!argc)
1623		usage_with_options(kvm_usage, kvm_options);
1624
1625	if (!perf_host)
1626		perf_guest = 1;
1627
1628	if (!file_name) {
1629		file_name = get_filename_for_perf_kvm();
1630
1631		if (!file_name) {
1632			pr_err("Failed to allocate memory for filename\n");
1633			return -ENOMEM;
1634		}
1635	}
1636
1637	if (strlen(argv[0]) > 2 && strstarts("record", argv[0]))
1638		return __cmd_record(file_name, argc, argv);
1639	else if (strlen(argv[0]) > 2 && strstarts("report", argv[0]))
1640		return __cmd_report(file_name, argc, argv);
1641	else if (strlen(argv[0]) > 2 && strstarts("diff", argv[0]))
1642		return cmd_diff(argc, argv);
1643	else if (!strcmp(argv[0], "top"))
1644		return cmd_top(argc, argv);
1645	else if (strlen(argv[0]) > 2 && strstarts("buildid-list", argv[0]))
1646		return __cmd_buildid_list(file_name, argc, argv);
1647#if defined(HAVE_KVM_STAT_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
1648	else if (strlen(argv[0]) > 2 && strstarts("stat", argv[0]))
1649		return kvm_cmd_stat(file_name, argc, argv);
1650#endif
1651	else
1652		usage_with_options(kvm_usage, kvm_options);
1653
1654	return 0;
1655}
v4.17
   1// SPDX-License-Identifier: GPL-2.0
   2#include "builtin.h"
   3#include "perf.h"
   4
 
   5#include "util/evsel.h"
   6#include "util/evlist.h"
 
   7#include "util/term.h"
   8#include "util/util.h"
   9#include "util/cache.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/parse-options.h>
  16#include "util/trace-event.h"
  17#include "util/debug.h"
  18#include "util/tool.h"
  19#include "util/stat.h"
 
  20#include "util/top.h"
  21#include "util/data.h"
  22#include "util/ordered-events.h"
 
 
 
  23
  24#include <sys/prctl.h>
  25#ifdef HAVE_TIMERFD_SUPPORT
  26#include <sys/timerfd.h>
  27#endif
  28#include <sys/time.h>
  29#include <sys/types.h>
  30#include <sys/stat.h>
  31#include <fcntl.h>
  32
 
  33#include <linux/kernel.h>
 
  34#include <linux/time64.h>
 
  35#include <errno.h>
  36#include <inttypes.h>
  37#include <poll.h>
  38#include <termios.h>
  39#include <semaphore.h>
  40#include <signal.h>
  41#include <math.h>
 
  42
  43static const char *get_filename_for_perf_kvm(void)
  44{
  45	const char *filename;
  46
  47	if (perf_host && !perf_guest)
  48		filename = strdup("perf.data.host");
  49	else if (!perf_host && perf_guest)
  50		filename = strdup("perf.data.guest");
  51	else
  52		filename = strdup("perf.data.kvm");
  53
  54	return filename;
  55}
  56
  57#ifdef HAVE_KVM_STAT_SUPPORT
  58#include "util/kvm-stat.h"
  59
  60void exit_event_get_key(struct perf_evsel *evsel,
  61			struct perf_sample *sample,
  62			struct event_key *key)
  63{
  64	key->info = 0;
  65	key->key = perf_evsel__intval(evsel, sample, kvm_exit_reason);
  66}
  67
  68bool kvm_exit_event(struct perf_evsel *evsel)
  69{
  70	return !strcmp(evsel->name, kvm_exit_trace);
  71}
  72
  73bool exit_event_begin(struct perf_evsel *evsel,
  74		      struct perf_sample *sample, struct event_key *key)
  75{
  76	if (kvm_exit_event(evsel)) {
  77		exit_event_get_key(evsel, sample, key);
  78		return true;
  79	}
  80
  81	return false;
  82}
  83
  84bool kvm_entry_event(struct perf_evsel *evsel)
  85{
  86	return !strcmp(evsel->name, kvm_entry_trace);
  87}
  88
  89bool exit_event_end(struct perf_evsel *evsel,
  90		    struct perf_sample *sample __maybe_unused,
  91		    struct event_key *key __maybe_unused)
  92{
  93	return kvm_entry_event(evsel);
  94}
  95
  96static const char *get_exit_reason(struct perf_kvm_stat *kvm,
  97				   struct exit_reasons_table *tbl,
  98				   u64 exit_code)
  99{
 100	while (tbl->reason != NULL) {
 101		if (tbl->exit_code == exit_code)
 102			return tbl->reason;
 103		tbl++;
 104	}
 105
 106	pr_err("unknown kvm exit code:%lld on %s\n",
 107		(unsigned long long)exit_code, kvm->exit_reasons_isa);
 108	return "UNKNOWN";
 109}
 110
 111void exit_event_decode_key(struct perf_kvm_stat *kvm,
 112			   struct event_key *key,
 113			   char *decode)
 114{
 115	const char *exit_reason = get_exit_reason(kvm, key->exit_reasons,
 116						  key->key);
 117
 118	scnprintf(decode, decode_str_len, "%s", exit_reason);
 119}
 120
 121static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
 122{
 123	struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
 124
 125	for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
 126		if (!strcmp(events_ops->name, kvm->report_event)) {
 127			kvm->events_ops = events_ops->ops;
 128			return true;
 129		}
 130	}
 131
 132	return false;
 133}
 134
 135struct vcpu_event_record {
 136	int vcpu_id;
 137	u64 start_time;
 138	struct kvm_event *last_event;
 139};
 140
 141
 142static void init_kvm_event_record(struct perf_kvm_stat *kvm)
 143{
 144	unsigned int i;
 145
 146	for (i = 0; i < EVENTS_CACHE_SIZE; i++)
 147		INIT_LIST_HEAD(&kvm->kvm_events_cache[i]);
 148}
 149
 150#ifdef HAVE_TIMERFD_SUPPORT
 151static void clear_events_cache_stats(struct list_head *kvm_events_cache)
 152{
 153	struct list_head *head;
 154	struct kvm_event *event;
 155	unsigned int i;
 156	int j;
 157
 158	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
 159		head = &kvm_events_cache[i];
 160		list_for_each_entry(event, head, hash_entry) {
 161			/* reset stats for event */
 162			event->total.time = 0;
 163			init_stats(&event->total.stats);
 164
 165			for (j = 0; j < event->max_vcpu; ++j) {
 166				event->vcpu[j].time = 0;
 167				init_stats(&event->vcpu[j].stats);
 168			}
 169		}
 170	}
 171}
 172#endif
 173
 174static int kvm_events_hash_fn(u64 key)
 175{
 176	return key & (EVENTS_CACHE_SIZE - 1);
 177}
 178
 179static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
 180{
 181	int old_max_vcpu = event->max_vcpu;
 182	void *prev;
 183
 184	if (vcpu_id < event->max_vcpu)
 185		return true;
 186
 187	while (event->max_vcpu <= vcpu_id)
 188		event->max_vcpu += DEFAULT_VCPU_NUM;
 189
 190	prev = event->vcpu;
 191	event->vcpu = realloc(event->vcpu,
 192			      event->max_vcpu * sizeof(*event->vcpu));
 193	if (!event->vcpu) {
 194		free(prev);
 195		pr_err("Not enough memory\n");
 196		return false;
 197	}
 198
 199	memset(event->vcpu + old_max_vcpu, 0,
 200	       (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
 201	return true;
 202}
 203
 204static struct kvm_event *kvm_alloc_init_event(struct event_key *key)
 205{
 206	struct kvm_event *event;
 207
 208	event = zalloc(sizeof(*event));
 209	if (!event) {
 210		pr_err("Not enough memory\n");
 211		return NULL;
 212	}
 213
 214	event->key = *key;
 215	init_stats(&event->total.stats);
 216	return event;
 217}
 218
 219static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
 220					       struct event_key *key)
 221{
 222	struct kvm_event *event;
 223	struct list_head *head;
 224
 225	BUG_ON(key->key == INVALID_KEY);
 226
 227	head = &kvm->kvm_events_cache[kvm_events_hash_fn(key->key)];
 228	list_for_each_entry(event, head, hash_entry) {
 229		if (event->key.key == key->key && event->key.info == key->info)
 230			return event;
 231	}
 232
 233	event = kvm_alloc_init_event(key);
 234	if (!event)
 235		return NULL;
 236
 237	list_add(&event->hash_entry, head);
 238	return event;
 239}
 240
 241static bool handle_begin_event(struct perf_kvm_stat *kvm,
 242			       struct vcpu_event_record *vcpu_record,
 243			       struct event_key *key, u64 timestamp)
 244{
 245	struct kvm_event *event = NULL;
 246
 247	if (key->key != INVALID_KEY)
 248		event = find_create_kvm_event(kvm, key);
 249
 250	vcpu_record->last_event = event;
 251	vcpu_record->start_time = timestamp;
 252	return true;
 253}
 254
 255static void
 256kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
 257{
 258	kvm_stats->time += time_diff;
 259	update_stats(&kvm_stats->stats, time_diff);
 260}
 261
 262static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
 263{
 264	struct kvm_event_stats *kvm_stats = &event->total;
 265
 266	if (vcpu_id != -1)
 267		kvm_stats = &event->vcpu[vcpu_id];
 268
 269	return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
 270				avg_stats(&kvm_stats->stats));
 271}
 272
 273static bool update_kvm_event(struct kvm_event *event, int vcpu_id,
 274			     u64 time_diff)
 275{
 276	if (vcpu_id == -1) {
 277		kvm_update_event_stats(&event->total, time_diff);
 278		return true;
 279	}
 280
 281	if (!kvm_event_expand(event, vcpu_id))
 282		return false;
 283
 284	kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
 285	return true;
 286}
 287
 288static bool is_child_event(struct perf_kvm_stat *kvm,
 289			   struct perf_evsel *evsel,
 290			   struct perf_sample *sample,
 291			   struct event_key *key)
 292{
 293	struct child_event_ops *child_ops;
 294
 295	child_ops = kvm->events_ops->child_ops;
 296
 297	if (!child_ops)
 298		return false;
 299
 300	for (; child_ops->name; child_ops++) {
 301		if (!strcmp(evsel->name, child_ops->name)) {
 302			child_ops->get_key(evsel, sample, key);
 303			return true;
 304		}
 305	}
 306
 307	return false;
 308}
 309
 310static bool handle_child_event(struct perf_kvm_stat *kvm,
 311			       struct vcpu_event_record *vcpu_record,
 312			       struct event_key *key,
 313			       struct perf_sample *sample __maybe_unused)
 314{
 315	struct kvm_event *event = NULL;
 316
 317	if (key->key != INVALID_KEY)
 318		event = find_create_kvm_event(kvm, key);
 319
 320	vcpu_record->last_event = event;
 321
 322	return true;
 323}
 324
 325static bool skip_event(const char *event)
 326{
 327	const char * const *skip_events;
 328
 329	for (skip_events = kvm_skip_events; *skip_events; skip_events++)
 330		if (!strcmp(event, *skip_events))
 331			return true;
 332
 333	return false;
 334}
 335
 336static bool handle_end_event(struct perf_kvm_stat *kvm,
 337			     struct vcpu_event_record *vcpu_record,
 338			     struct event_key *key,
 339			     struct perf_sample *sample)
 340{
 341	struct kvm_event *event;
 342	u64 time_begin, time_diff;
 343	int vcpu;
 344
 345	if (kvm->trace_vcpu == -1)
 346		vcpu = -1;
 347	else
 348		vcpu = vcpu_record->vcpu_id;
 349
 350	event = vcpu_record->last_event;
 351	time_begin = vcpu_record->start_time;
 352
 353	/* The begin event is not caught. */
 354	if (!time_begin)
 355		return true;
 356
 357	/*
 358	 * In some case, the 'begin event' only records the start timestamp,
 359	 * the actual event is recognized in the 'end event' (e.g. mmio-event).
 360	 */
 361
 362	/* Both begin and end events did not get the key. */
 363	if (!event && key->key == INVALID_KEY)
 364		return true;
 365
 366	if (!event)
 367		event = find_create_kvm_event(kvm, key);
 368
 369	if (!event)
 370		return false;
 371
 372	vcpu_record->last_event = NULL;
 373	vcpu_record->start_time = 0;
 374
 375	/* seems to happen once in a while during live mode */
 376	if (sample->time < time_begin) {
 377		pr_debug("End time before begin time; skipping event.\n");
 378		return true;
 379	}
 380
 381	time_diff = sample->time - time_begin;
 382
 383	if (kvm->duration && time_diff > kvm->duration) {
 384		char decode[decode_str_len];
 385
 386		kvm->events_ops->decode_key(kvm, &event->key, decode);
 387		if (!skip_event(decode)) {
 388			pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
 389				 sample->time, sample->pid, vcpu_record->vcpu_id,
 390				 decode, time_diff / NSEC_PER_USEC);
 391		}
 392	}
 393
 394	return update_kvm_event(event, vcpu, time_diff);
 395}
 396
 397static
 398struct vcpu_event_record *per_vcpu_record(struct thread *thread,
 399					  struct perf_evsel *evsel,
 400					  struct perf_sample *sample)
 401{
 402	/* Only kvm_entry records vcpu id. */
 403	if (!thread__priv(thread) && kvm_entry_event(evsel)) {
 404		struct vcpu_event_record *vcpu_record;
 405
 406		vcpu_record = zalloc(sizeof(*vcpu_record));
 407		if (!vcpu_record) {
 408			pr_err("%s: Not enough memory\n", __func__);
 409			return NULL;
 410		}
 411
 412		vcpu_record->vcpu_id = perf_evsel__intval(evsel, sample,
 413							  vcpu_id_str);
 414		thread__set_priv(thread, vcpu_record);
 415	}
 416
 417	return thread__priv(thread);
 418}
 419
 420static bool handle_kvm_event(struct perf_kvm_stat *kvm,
 421			     struct thread *thread,
 422			     struct perf_evsel *evsel,
 423			     struct perf_sample *sample)
 424{
 425	struct vcpu_event_record *vcpu_record;
 426	struct event_key key = { .key = INVALID_KEY,
 427				 .exit_reasons = kvm->exit_reasons };
 428
 429	vcpu_record = per_vcpu_record(thread, evsel, sample);
 430	if (!vcpu_record)
 431		return true;
 432
 433	/* only process events for vcpus user cares about */
 434	if ((kvm->trace_vcpu != -1) &&
 435	    (kvm->trace_vcpu != vcpu_record->vcpu_id))
 436		return true;
 437
 438	if (kvm->events_ops->is_begin_event(evsel, sample, &key))
 439		return handle_begin_event(kvm, vcpu_record, &key, sample->time);
 440
 441	if (is_child_event(kvm, evsel, sample, &key))
 442		return handle_child_event(kvm, vcpu_record, &key, sample);
 443
 444	if (kvm->events_ops->is_end_event(evsel, sample, &key))
 445		return handle_end_event(kvm, vcpu_record, &key, sample);
 446
 447	return true;
 448}
 449
 450#define GET_EVENT_KEY(func, field)					\
 451static u64 get_event_ ##func(struct kvm_event *event, int vcpu)		\
 452{									\
 453	if (vcpu == -1)							\
 454		return event->total.field;				\
 455									\
 456	if (vcpu >= event->max_vcpu)					\
 457		return 0;						\
 458									\
 459	return event->vcpu[vcpu].field;					\
 460}
 461
 462#define COMPARE_EVENT_KEY(func, field)					\
 463GET_EVENT_KEY(func, field)						\
 464static int compare_kvm_event_ ## func(struct kvm_event *one,		\
 465					struct kvm_event *two, int vcpu)\
 466{									\
 467	return get_event_ ##func(one, vcpu) >				\
 468				get_event_ ##func(two, vcpu);		\
 469}
 470
 471GET_EVENT_KEY(time, time);
 472COMPARE_EVENT_KEY(count, stats.n);
 473COMPARE_EVENT_KEY(mean, stats.mean);
 474GET_EVENT_KEY(max, stats.max);
 475GET_EVENT_KEY(min, stats.min);
 476
 477#define DEF_SORT_NAME_KEY(name, compare_key)				\
 478	{ #name, compare_kvm_event_ ## compare_key }
 479
 480static struct kvm_event_key keys[] = {
 481	DEF_SORT_NAME_KEY(sample, count),
 482	DEF_SORT_NAME_KEY(time, mean),
 483	{ NULL, NULL }
 484};
 485
 486static bool select_key(struct perf_kvm_stat *kvm)
 487{
 488	int i;
 489
 490	for (i = 0; keys[i].name; i++) {
 491		if (!strcmp(keys[i].name, kvm->sort_key)) {
 492			kvm->compare = keys[i].key;
 493			return true;
 494		}
 495	}
 496
 497	pr_err("Unknown compare key:%s\n", kvm->sort_key);
 498	return false;
 499}
 500
 501static void insert_to_result(struct rb_root *result, struct kvm_event *event,
 502			     key_cmp_fun bigger, int vcpu)
 503{
 504	struct rb_node **rb = &result->rb_node;
 505	struct rb_node *parent = NULL;
 506	struct kvm_event *p;
 507
 508	while (*rb) {
 509		p = container_of(*rb, struct kvm_event, rb);
 510		parent = *rb;
 511
 512		if (bigger(event, p, vcpu))
 513			rb = &(*rb)->rb_left;
 514		else
 515			rb = &(*rb)->rb_right;
 516	}
 517
 518	rb_link_node(&event->rb, parent, rb);
 519	rb_insert_color(&event->rb, result);
 520}
 521
 522static void
 523update_total_count(struct perf_kvm_stat *kvm, struct kvm_event *event)
 524{
 525	int vcpu = kvm->trace_vcpu;
 526
 527	kvm->total_count += get_event_count(event, vcpu);
 528	kvm->total_time += get_event_time(event, vcpu);
 529}
 530
 531static bool event_is_valid(struct kvm_event *event, int vcpu)
 532{
 533	return !!get_event_count(event, vcpu);
 534}
 535
 536static void sort_result(struct perf_kvm_stat *kvm)
 537{
 538	unsigned int i;
 539	int vcpu = kvm->trace_vcpu;
 540	struct kvm_event *event;
 541
 542	for (i = 0; i < EVENTS_CACHE_SIZE; i++) {
 543		list_for_each_entry(event, &kvm->kvm_events_cache[i], hash_entry) {
 544			if (event_is_valid(event, vcpu)) {
 545				update_total_count(kvm, event);
 546				insert_to_result(&kvm->result, event,
 547						 kvm->compare, vcpu);
 548			}
 549		}
 550	}
 551}
 552
 553/* returns left most element of result, and erase it */
 554static struct kvm_event *pop_from_result(struct rb_root *result)
 555{
 556	struct rb_node *node = rb_first(result);
 557
 558	if (!node)
 559		return NULL;
 560
 561	rb_erase(node, result);
 562	return container_of(node, struct kvm_event, rb);
 563}
 564
 565static void print_vcpu_info(struct perf_kvm_stat *kvm)
 566{
 567	int vcpu = kvm->trace_vcpu;
 568
 569	pr_info("Analyze events for ");
 570
 571	if (kvm->opts.target.system_wide)
 572		pr_info("all VMs, ");
 573	else if (kvm->opts.target.pid)
 574		pr_info("pid(s) %s, ", kvm->opts.target.pid);
 575	else
 576		pr_info("dazed and confused on what is monitored, ");
 577
 578	if (vcpu == -1)
 579		pr_info("all VCPUs:\n\n");
 580	else
 581		pr_info("VCPU %d:\n\n", vcpu);
 582}
 583
 584static void show_timeofday(void)
 585{
 586	char date[64];
 587	struct timeval tv;
 588	struct tm ltime;
 589
 590	gettimeofday(&tv, NULL);
 591	if (localtime_r(&tv.tv_sec, &ltime)) {
 592		strftime(date, sizeof(date), "%H:%M:%S", &ltime);
 593		pr_info("%s.%06ld", date, tv.tv_usec);
 594	} else
 595		pr_info("00:00:00.000000");
 596
 597	return;
 598}
 599
 600static void print_result(struct perf_kvm_stat *kvm)
 601{
 602	char decode[decode_str_len];
 603	struct kvm_event *event;
 604	int vcpu = kvm->trace_vcpu;
 605
 606	if (kvm->live) {
 607		puts(CONSOLE_CLEAR);
 608		show_timeofday();
 609	}
 610
 611	pr_info("\n\n");
 612	print_vcpu_info(kvm);
 613	pr_info("%*s ", decode_str_len, kvm->events_ops->name);
 614	pr_info("%10s ", "Samples");
 615	pr_info("%9s ", "Samples%");
 616
 617	pr_info("%9s ", "Time%");
 618	pr_info("%11s ", "Min Time");
 619	pr_info("%11s ", "Max Time");
 620	pr_info("%16s ", "Avg time");
 621	pr_info("\n\n");
 622
 623	while ((event = pop_from_result(&kvm->result))) {
 624		u64 ecount, etime, max, min;
 625
 626		ecount = get_event_count(event, vcpu);
 627		etime = get_event_time(event, vcpu);
 628		max = get_event_max(event, vcpu);
 629		min = get_event_min(event, vcpu);
 630
 631		kvm->events_ops->decode_key(kvm, &event->key, decode);
 632		pr_info("%*s ", decode_str_len, decode);
 633		pr_info("%10llu ", (unsigned long long)ecount);
 634		pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
 635		pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
 636		pr_info("%9.2fus ", (double)min / NSEC_PER_USEC);
 637		pr_info("%9.2fus ", (double)max / NSEC_PER_USEC);
 638		pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount / NSEC_PER_USEC,
 639			kvm_event_rel_stddev(vcpu, event));
 640		pr_info("\n");
 641	}
 642
 643	pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
 644		kvm->total_count, kvm->total_time / (double)NSEC_PER_USEC);
 645
 646	if (kvm->lost_events)
 647		pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
 648}
 649
 650#ifdef HAVE_TIMERFD_SUPPORT
 651static int process_lost_event(struct perf_tool *tool,
 652			      union perf_event *event __maybe_unused,
 653			      struct perf_sample *sample __maybe_unused,
 654			      struct machine *machine __maybe_unused)
 655{
 656	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
 657
 658	kvm->lost_events++;
 659	return 0;
 660}
 661#endif
 662
 663static bool skip_sample(struct perf_kvm_stat *kvm,
 664			struct perf_sample *sample)
 665{
 666	if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
 667		return true;
 668
 669	return false;
 670}
 671
 672static int process_sample_event(struct perf_tool *tool,
 673				union perf_event *event,
 674				struct perf_sample *sample,
 675				struct perf_evsel *evsel,
 676				struct machine *machine)
 677{
 678	int err = 0;
 679	struct thread *thread;
 680	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
 681						 tool);
 682
 683	if (skip_sample(kvm, sample))
 684		return 0;
 685
 686	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
 687	if (thread == NULL) {
 688		pr_debug("problem processing %d event, skipping it.\n",
 689			event->header.type);
 690		return -1;
 691	}
 692
 693	if (!handle_kvm_event(kvm, thread, evsel, sample))
 694		err = -1;
 695
 696	thread__put(thread);
 697	return err;
 698}
 699
 700static int cpu_isa_config(struct perf_kvm_stat *kvm)
 701{
 702	char buf[64], *cpuid;
 703	int err;
 704
 705	if (kvm->live) {
 706		err = get_cpuid(buf, sizeof(buf));
 707		if (err != 0) {
 708			pr_err("Failed to look up CPU type\n");
 709			return err;
 
 710		}
 711		cpuid = buf;
 712	} else
 713		cpuid = kvm->session->header.env.cpuid;
 714
 715	if (!cpuid) {
 716		pr_err("Failed to look up CPU type\n");
 717		return -EINVAL;
 718	}
 719
 720	err = cpu_isa_init(kvm, cpuid);
 721	if (err == -ENOTSUP)
 722		pr_err("CPU %s is not supported.\n", cpuid);
 723
 724	return err;
 725}
 726
 727static bool verify_vcpu(int vcpu)
 728{
 729	if (vcpu != -1 && vcpu < 0) {
 730		pr_err("Invalid vcpu:%d.\n", vcpu);
 731		return false;
 732	}
 733
 734	return true;
 735}
 736
 737#ifdef HAVE_TIMERFD_SUPPORT
 738/* keeping the max events to a modest level to keep
 739 * the processing of samples per mmap smooth.
 740 */
 741#define PERF_KVM__MAX_EVENTS_PER_MMAP  25
 742
 743static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
 744				   u64 *mmap_time)
 745{
 746	struct perf_evlist *evlist = kvm->evlist;
 747	union perf_event *event;
 748	struct perf_mmap *md;
 749	u64 timestamp;
 750	s64 n = 0;
 751	int err;
 752
 753	*mmap_time = ULLONG_MAX;
 754	md = &evlist->mmap[idx];
 755	err = perf_mmap__read_init(md);
 756	if (err < 0)
 757		return (err == -EAGAIN) ? 0 : -1;
 758
 759	while ((event = perf_mmap__read_event(md)) != NULL) {
 760		err = perf_evlist__parse_sample_timestamp(evlist, event, &timestamp);
 761		if (err) {
 762			perf_mmap__consume(md);
 763			pr_err("Failed to parse sample\n");
 764			return -1;
 765		}
 766
 767		err = perf_session__queue_event(kvm->session, event, timestamp, 0);
 768		/*
 769		 * FIXME: Here we can't consume the event, as perf_session__queue_event will
 770		 *        point to it, and it'll get possibly overwritten by the kernel.
 771		 */
 772		perf_mmap__consume(md);
 773
 774		if (err) {
 775			pr_err("Failed to enqueue sample: %d\n", err);
 776			return -1;
 777		}
 778
 779		/* save time stamp of our first sample for this mmap */
 780		if (n == 0)
 781			*mmap_time = timestamp;
 782
 783		/* limit events per mmap handled all at once */
 784		n++;
 785		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
 786			break;
 787	}
 788
 789	perf_mmap__read_done(md);
 790	return n;
 791}
 792
 793static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
 794{
 795	int i, err, throttled = 0;
 796	s64 n, ntotal = 0;
 797	u64 flush_time = ULLONG_MAX, mmap_time;
 798
 799	for (i = 0; i < kvm->evlist->nr_mmaps; i++) {
 800		n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
 801		if (n < 0)
 802			return -1;
 803
 804		/* flush time is going to be the minimum of all the individual
 805		 * mmap times. Essentially, we flush all the samples queued up
 806		 * from the last pass under our minimal start time -- that leaves
 807		 * a very small race for samples to come in with a lower timestamp.
 808		 * The ioctl to return the perf_clock timestamp should close the
 809		 * race entirely.
 810		 */
 811		if (mmap_time < flush_time)
 812			flush_time = mmap_time;
 813
 814		ntotal += n;
 815		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
 816			throttled = 1;
 817	}
 818
 819	/* flush queue after each round in which we processed events */
 820	if (ntotal) {
 821		struct ordered_events *oe = &kvm->session->ordered_events;
 822
 823		oe->next_flush = flush_time;
 824		err = ordered_events__flush(oe, OE_FLUSH__ROUND);
 825		if (err) {
 826			if (kvm->lost_events)
 827				pr_info("\nLost events: %" PRIu64 "\n\n",
 828					kvm->lost_events);
 829			return err;
 830		}
 831	}
 832
 833	return throttled;
 834}
 835
 836static volatile int done;
 837
 838static void sig_handler(int sig __maybe_unused)
 839{
 840	done = 1;
 841}
 842
 843static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
 844{
 845	struct itimerspec new_value;
 846	int rc = -1;
 847
 848	kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
 849	if (kvm->timerfd < 0) {
 850		pr_err("timerfd_create failed\n");
 851		goto out;
 852	}
 853
 854	new_value.it_value.tv_sec = kvm->display_time;
 855	new_value.it_value.tv_nsec = 0;
 856	new_value.it_interval.tv_sec = kvm->display_time;
 857	new_value.it_interval.tv_nsec = 0;
 858
 859	if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
 860		pr_err("timerfd_settime failed: %d\n", errno);
 861		close(kvm->timerfd);
 862		goto out;
 863	}
 864
 865	rc = 0;
 866out:
 867	return rc;
 868}
 869
 870static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
 871{
 872	uint64_t c;
 873	int rc;
 874
 875	rc = read(kvm->timerfd, &c, sizeof(uint64_t));
 876	if (rc < 0) {
 877		if (errno == EAGAIN)
 878			return 0;
 879
 880		pr_err("Failed to read timer fd: %d\n", errno);
 881		return -1;
 882	}
 883
 884	if (rc != sizeof(uint64_t)) {
 885		pr_err("Error reading timer fd - invalid size returned\n");
 886		return -1;
 887	}
 888
 889	if (c != 1)
 890		pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
 891
 892	/* update display */
 893	sort_result(kvm);
 894	print_result(kvm);
 895
 896	/* reset counts */
 897	clear_events_cache_stats(kvm->kvm_events_cache);
 898	kvm->total_count = 0;
 899	kvm->total_time = 0;
 900	kvm->lost_events = 0;
 901
 902	return 0;
 903}
 904
 905static int fd_set_nonblock(int fd)
 906{
 907	long arg = 0;
 908
 909	arg = fcntl(fd, F_GETFL);
 910	if (arg < 0) {
 911		pr_err("Failed to get current flags for fd %d\n", fd);
 912		return -1;
 913	}
 914
 915	if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
 916		pr_err("Failed to set non-block option on fd %d\n", fd);
 917		return -1;
 918	}
 919
 920	return 0;
 921}
 922
 923static int perf_kvm__handle_stdin(void)
 924{
 925	int c;
 926
 927	c = getc(stdin);
 928	if (c == 'q')
 929		return 1;
 930
 931	return 0;
 932}
 933
 934static int kvm_events_live_report(struct perf_kvm_stat *kvm)
 935{
 936	int nr_stdin, ret, err = -EINVAL;
 937	struct termios save;
 938
 939	/* live flag must be set first */
 940	kvm->live = true;
 941
 942	ret = cpu_isa_config(kvm);
 943	if (ret < 0)
 944		return ret;
 945
 946	if (!verify_vcpu(kvm->trace_vcpu) ||
 947	    !select_key(kvm) ||
 948	    !register_kvm_events_ops(kvm)) {
 949		goto out;
 950	}
 951
 952	set_term_quiet_input(&save);
 953	init_kvm_event_record(kvm);
 954
 955	signal(SIGINT, sig_handler);
 956	signal(SIGTERM, sig_handler);
 957
 958	/* add timer fd */
 959	if (perf_kvm__timerfd_create(kvm) < 0) {
 960		err = -1;
 961		goto out;
 962	}
 963
 964	if (perf_evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
 965		goto out;
 966
 967	nr_stdin = perf_evlist__add_pollfd(kvm->evlist, fileno(stdin));
 968	if (nr_stdin < 0)
 969		goto out;
 970
 971	if (fd_set_nonblock(fileno(stdin)) != 0)
 972		goto out;
 973
 974	/* everything is good - enable the events and process */
 975	perf_evlist__enable(kvm->evlist);
 976
 977	while (!done) {
 978		struct fdarray *fda = &kvm->evlist->pollfd;
 979		int rc;
 980
 981		rc = perf_kvm__mmap_read(kvm);
 982		if (rc < 0)
 983			break;
 984
 985		err = perf_kvm__handle_timerfd(kvm);
 986		if (err)
 987			goto out;
 988
 989		if (fda->entries[nr_stdin].revents & POLLIN)
 990			done = perf_kvm__handle_stdin();
 991
 992		if (!rc && !done)
 993			err = fdarray__poll(fda, 100);
 994	}
 995
 996	perf_evlist__disable(kvm->evlist);
 997
 998	if (err == 0) {
 999		sort_result(kvm);
1000		print_result(kvm);
1001	}
1002
1003out:
1004	if (kvm->timerfd >= 0)
1005		close(kvm->timerfd);
1006
1007	tcsetattr(0, TCSAFLUSH, &save);
1008	return err;
1009}
1010
1011static int kvm_live_open_events(struct perf_kvm_stat *kvm)
1012{
1013	int err, rc = -1;
1014	struct perf_evsel *pos;
1015	struct perf_evlist *evlist = kvm->evlist;
1016	char sbuf[STRERR_BUFSIZE];
1017
1018	perf_evlist__config(evlist, &kvm->opts, NULL);
1019
1020	/*
1021	 * Note: exclude_{guest,host} do not apply here.
1022	 *       This command processes KVM tracepoints from host only
1023	 */
1024	evlist__for_each_entry(evlist, pos) {
1025		struct perf_event_attr *attr = &pos->attr;
1026
1027		/* make sure these *are* set */
1028		perf_evsel__set_sample_bit(pos, TID);
1029		perf_evsel__set_sample_bit(pos, TIME);
1030		perf_evsel__set_sample_bit(pos, CPU);
1031		perf_evsel__set_sample_bit(pos, RAW);
1032		/* make sure these are *not*; want as small a sample as possible */
1033		perf_evsel__reset_sample_bit(pos, PERIOD);
1034		perf_evsel__reset_sample_bit(pos, IP);
1035		perf_evsel__reset_sample_bit(pos, CALLCHAIN);
1036		perf_evsel__reset_sample_bit(pos, ADDR);
1037		perf_evsel__reset_sample_bit(pos, READ);
1038		attr->mmap = 0;
1039		attr->comm = 0;
1040		attr->task = 0;
1041
1042		attr->sample_period = 1;
1043
1044		attr->watermark = 0;
1045		attr->wakeup_events = 1000;
1046
1047		/* will enable all once we are ready */
1048		attr->disabled = 1;
1049	}
1050
1051	err = perf_evlist__open(evlist);
1052	if (err < 0) {
1053		printf("Couldn't create the events: %s\n",
1054		       str_error_r(errno, sbuf, sizeof(sbuf)));
1055		goto out;
1056	}
1057
1058	if (perf_evlist__mmap(evlist, kvm->opts.mmap_pages) < 0) {
1059		ui__error("Failed to mmap the events: %s\n",
1060			  str_error_r(errno, sbuf, sizeof(sbuf)));
1061		perf_evlist__close(evlist);
1062		goto out;
1063	}
1064
1065	rc = 0;
1066
1067out:
1068	return rc;
1069}
1070#endif
1071
1072static int read_events(struct perf_kvm_stat *kvm)
1073{
1074	int ret;
1075
1076	struct perf_tool eops = {
1077		.sample			= process_sample_event,
1078		.comm			= perf_event__process_comm,
1079		.namespaces		= perf_event__process_namespaces,
1080		.ordered_events		= true,
1081	};
1082	struct perf_data file = {
1083		.file      = {
1084			.path = kvm->file_name,
1085		},
1086		.mode      = PERF_DATA_MODE_READ,
1087		.force     = kvm->force,
1088	};
1089
1090	kvm->tool = eops;
1091	kvm->session = perf_session__new(&file, false, &kvm->tool);
1092	if (!kvm->session) {
1093		pr_err("Initializing perf session failed\n");
1094		return -1;
1095	}
1096
1097	symbol__init(&kvm->session->header.env);
1098
1099	if (!perf_session__has_traces(kvm->session, "kvm record")) {
1100		ret = -EINVAL;
1101		goto out_delete;
1102	}
1103
1104	/*
1105	 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1106	 * traced in the old kernel.
1107	 */
1108	ret = cpu_isa_config(kvm);
1109	if (ret < 0)
1110		goto out_delete;
1111
1112	ret = perf_session__process_events(kvm->session);
1113
1114out_delete:
1115	perf_session__delete(kvm->session);
1116	return ret;
1117}
1118
1119static int parse_target_str(struct perf_kvm_stat *kvm)
1120{
1121	if (kvm->opts.target.pid) {
1122		kvm->pid_list = intlist__new(kvm->opts.target.pid);
1123		if (kvm->pid_list == NULL) {
1124			pr_err("Error parsing process id string\n");
1125			return -EINVAL;
1126		}
1127	}
1128
1129	return 0;
1130}
1131
1132static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1133{
1134	int ret = -EINVAL;
1135	int vcpu = kvm->trace_vcpu;
1136
1137	if (parse_target_str(kvm) != 0)
1138		goto exit;
1139
1140	if (!verify_vcpu(vcpu))
1141		goto exit;
1142
1143	if (!select_key(kvm))
1144		goto exit;
1145
1146	if (!register_kvm_events_ops(kvm))
1147		goto exit;
1148
1149	init_kvm_event_record(kvm);
1150	setup_pager();
1151
1152	ret = read_events(kvm);
1153	if (ret)
1154		goto exit;
1155
1156	sort_result(kvm);
1157	print_result(kvm);
1158
1159exit:
1160	return ret;
1161}
1162
1163#define STRDUP_FAIL_EXIT(s)		\
1164	({	char *_p;		\
1165	_p = strdup(s);		\
1166		if (!_p)		\
1167			return -ENOMEM;	\
1168		_p;			\
1169	})
1170
1171int __weak setup_kvm_events_tp(struct perf_kvm_stat *kvm __maybe_unused)
1172{
1173	return 0;
1174}
1175
1176static int
1177kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1178{
1179	unsigned int rec_argc, i, j, events_tp_size;
1180	const char **rec_argv;
1181	const char * const record_args[] = {
1182		"record",
1183		"-R",
1184		"-m", "1024",
1185		"-c", "1",
1186	};
1187	const char * const kvm_stat_record_usage[] = {
1188		"perf kvm stat record [<options>]",
1189		NULL
1190	};
1191	const char * const *events_tp;
1192	int ret;
1193
1194	events_tp_size = 0;
1195	ret = setup_kvm_events_tp(kvm);
1196	if (ret < 0) {
1197		pr_err("Unable to setup the kvm tracepoints\n");
1198		return ret;
1199	}
1200
1201	for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1202		events_tp_size++;
1203
1204	rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1205		   2 * events_tp_size;
1206	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1207
1208	if (rec_argv == NULL)
1209		return -ENOMEM;
1210
1211	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1212		rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1213
1214	for (j = 0; j < events_tp_size; j++) {
1215		rec_argv[i++] = "-e";
1216		rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1217	}
1218
1219	rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1220	rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1221
1222	for (j = 1; j < (unsigned int)argc; j++, i++)
1223		rec_argv[i] = argv[j];
1224
1225	set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1226	set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1227	set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1228
1229	set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1230	set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1231	set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1232	set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1233	set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1234	set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1235	set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1236	set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1237	set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1238	set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1239	set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1240	set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1241	set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1242	set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1243	set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1244
1245	record_usage = kvm_stat_record_usage;
1246	return cmd_record(i, rec_argv);
1247}
1248
1249static int
1250kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1251{
1252	const struct option kvm_events_report_options[] = {
1253		OPT_STRING(0, "event", &kvm->report_event, "report event",
1254			   "event for reporting: vmexit, "
1255			   "mmio (x86 only), ioport (x86 only)"),
1256		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1257			    "vcpu id to report"),
1258		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1259			    "key for sorting: sample(sort by samples number)"
1260			    " time (sort by avg time)"),
1261		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1262			   "analyze events only for given process id(s)"),
1263		OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
1264		OPT_END()
1265	};
1266
1267	const char * const kvm_events_report_usage[] = {
1268		"perf kvm stat report [<options>]",
1269		NULL
1270	};
1271
1272	if (argc) {
1273		argc = parse_options(argc, argv,
1274				     kvm_events_report_options,
1275				     kvm_events_report_usage, 0);
1276		if (argc)
1277			usage_with_options(kvm_events_report_usage,
1278					   kvm_events_report_options);
1279	}
1280
1281	if (!kvm->opts.target.pid)
1282		kvm->opts.target.system_wide = true;
1283
1284	return kvm_events_report_vcpu(kvm);
1285}
1286
1287#ifdef HAVE_TIMERFD_SUPPORT
1288static struct perf_evlist *kvm_live_event_list(void)
1289{
1290	struct perf_evlist *evlist;
1291	char *tp, *name, *sys;
1292	int err = -1;
1293	const char * const *events_tp;
1294
1295	evlist = perf_evlist__new();
1296	if (evlist == NULL)
1297		return NULL;
1298
1299	for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1300
1301		tp = strdup(*events_tp);
1302		if (tp == NULL)
1303			goto out;
1304
1305		/* split tracepoint into subsystem and name */
1306		sys = tp;
1307		name = strchr(tp, ':');
1308		if (name == NULL) {
1309			pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1310			       *events_tp);
1311			free(tp);
1312			goto out;
1313		}
1314		*name = '\0';
1315		name++;
1316
1317		if (perf_evlist__add_newtp(evlist, sys, name, NULL)) {
1318			pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1319			free(tp);
1320			goto out;
1321		}
1322
1323		free(tp);
1324	}
1325
1326	err = 0;
1327
1328out:
1329	if (err) {
1330		perf_evlist__delete(evlist);
1331		evlist = NULL;
1332	}
1333
1334	return evlist;
1335}
1336
1337static int kvm_events_live(struct perf_kvm_stat *kvm,
1338			   int argc, const char **argv)
1339{
1340	char errbuf[BUFSIZ];
1341	int err;
1342
1343	const struct option live_options[] = {
1344		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1345			"record events on existing process id"),
1346		OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1347			"number of mmap data pages",
1348			perf_evlist__parse_mmap_pages),
1349		OPT_INCR('v', "verbose", &verbose,
1350			"be more verbose (show counter open errors, etc)"),
1351		OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1352			"system-wide collection from all CPUs"),
1353		OPT_UINTEGER('d', "display", &kvm->display_time,
1354			"time in seconds between display updates"),
1355		OPT_STRING(0, "event", &kvm->report_event, "report event",
1356			"event for reporting: "
1357			"vmexit, mmio (x86 only), ioport (x86 only)"),
1358		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1359			"vcpu id to report"),
1360		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1361			"key for sorting: sample(sort by samples number)"
1362			" time (sort by avg time)"),
1363		OPT_U64(0, "duration", &kvm->duration,
1364			"show events other than"
1365			" HLT (x86 only) or Wait state (s390 only)"
1366			" that take longer than duration usecs"),
1367		OPT_UINTEGER(0, "proc-map-timeout", &kvm->opts.proc_map_timeout,
1368				"per thread proc mmap processing timeout in ms"),
1369		OPT_END()
1370	};
1371	const char * const live_usage[] = {
1372		"perf kvm stat live [<options>]",
1373		NULL
1374	};
1375	struct perf_data data = {
1376		.mode = PERF_DATA_MODE_WRITE,
1377	};
1378
1379
1380	/* event handling */
1381	kvm->tool.sample = process_sample_event;
1382	kvm->tool.comm   = perf_event__process_comm;
1383	kvm->tool.exit   = perf_event__process_exit;
1384	kvm->tool.fork   = perf_event__process_fork;
1385	kvm->tool.lost   = process_lost_event;
1386	kvm->tool.namespaces  = perf_event__process_namespaces;
1387	kvm->tool.ordered_events = true;
1388	perf_tool__fill_defaults(&kvm->tool);
1389
1390	/* set defaults */
1391	kvm->display_time = 1;
1392	kvm->opts.user_interval = 1;
1393	kvm->opts.mmap_pages = 512;
1394	kvm->opts.target.uses_mmap = false;
1395	kvm->opts.target.uid_str = NULL;
1396	kvm->opts.target.uid = UINT_MAX;
1397	kvm->opts.proc_map_timeout = 500;
1398
1399	symbol__init(NULL);
1400	disable_buildid_cache();
1401
1402	use_browser = 0;
1403
1404	if (argc) {
1405		argc = parse_options(argc, argv, live_options,
1406				     live_usage, 0);
1407		if (argc)
1408			usage_with_options(live_usage, live_options);
1409	}
1410
1411	kvm->duration *= NSEC_PER_USEC;   /* convert usec to nsec */
1412
1413	/*
1414	 * target related setups
1415	 */
1416	err = target__validate(&kvm->opts.target);
1417	if (err) {
1418		target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1419		ui__warning("%s", errbuf);
1420	}
1421
1422	if (target__none(&kvm->opts.target))
1423		kvm->opts.target.system_wide = true;
1424
1425
1426	/*
1427	 * generate the event list
1428	 */
1429	err = setup_kvm_events_tp(kvm);
1430	if (err < 0) {
1431		pr_err("Unable to setup the kvm tracepoints\n");
1432		return err;
1433	}
1434
1435	kvm->evlist = kvm_live_event_list();
1436	if (kvm->evlist == NULL) {
1437		err = -1;
1438		goto out;
1439	}
1440
1441	symbol_conf.nr_events = kvm->evlist->nr_entries;
1442
1443	if (perf_evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1444		usage_with_options(live_usage, live_options);
1445
1446	/*
1447	 * perf session
1448	 */
1449	kvm->session = perf_session__new(&data, false, &kvm->tool);
1450	if (kvm->session == NULL) {
1451		err = -1;
1452		goto out;
1453	}
1454	kvm->session->evlist = kvm->evlist;
1455	perf_session__set_id_hdr_size(kvm->session);
1456	ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
1457	machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1458				    kvm->evlist->threads, false,
1459				    kvm->opts.proc_map_timeout, 1);
1460	err = kvm_live_open_events(kvm);
1461	if (err)
1462		goto out;
1463
1464	err = kvm_events_live_report(kvm);
1465
1466out:
1467	perf_session__delete(kvm->session);
1468	kvm->session = NULL;
1469	perf_evlist__delete(kvm->evlist);
1470
1471	return err;
1472}
1473#endif
1474
1475static void print_kvm_stat_usage(void)
1476{
1477	printf("Usage: perf kvm stat <command>\n\n");
1478
1479	printf("# Available commands:\n");
1480	printf("\trecord: record kvm events\n");
1481	printf("\treport: report statistical data of kvm events\n");
1482	printf("\tlive:   live reporting of statistical data of kvm events\n");
1483
1484	printf("\nOtherwise, it is the alias of 'perf stat':\n");
1485}
1486
1487static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1488{
1489	struct perf_kvm_stat kvm = {
1490		.file_name = file_name,
1491
1492		.trace_vcpu	= -1,
1493		.report_event	= "vmexit",
1494		.sort_key	= "sample",
1495
1496	};
1497
1498	if (argc == 1) {
1499		print_kvm_stat_usage();
1500		goto perf_stat;
1501	}
1502
1503	if (!strncmp(argv[1], "rec", 3))
1504		return kvm_events_record(&kvm, argc - 1, argv + 1);
1505
1506	if (!strncmp(argv[1], "rep", 3))
1507		return kvm_events_report(&kvm, argc - 1 , argv + 1);
1508
1509#ifdef HAVE_TIMERFD_SUPPORT
1510	if (!strncmp(argv[1], "live", 4))
1511		return kvm_events_live(&kvm, argc - 1 , argv + 1);
1512#endif
1513
1514perf_stat:
1515	return cmd_stat(argc, argv);
1516}
1517#endif /* HAVE_KVM_STAT_SUPPORT */
1518
 
 
 
 
 
 
1519static int __cmd_record(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("record");
1527	rec_argv[i++] = strdup("-o");
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_record(i, rec_argv);
1535}
1536
1537static int __cmd_report(const char *file_name, int argc, const char **argv)
1538{
1539	int rec_argc, i = 0, j;
1540	const char **rec_argv;
1541
1542	rec_argc = argc + 2;
1543	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1544	rec_argv[i++] = strdup("report");
1545	rec_argv[i++] = strdup("-i");
1546	rec_argv[i++] = strdup(file_name);
1547	for (j = 1; j < argc; j++, i++)
1548		rec_argv[i] = argv[j];
1549
1550	BUG_ON(i != rec_argc);
1551
1552	return cmd_report(i, rec_argv);
1553}
1554
1555static int
1556__cmd_buildid_list(const char *file_name, int argc, const char **argv)
1557{
1558	int rec_argc, i = 0, j;
1559	const char **rec_argv;
1560
1561	rec_argc = argc + 2;
1562	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1563	rec_argv[i++] = strdup("buildid-list");
1564	rec_argv[i++] = strdup("-i");
1565	rec_argv[i++] = strdup(file_name);
1566	for (j = 1; j < argc; j++, i++)
1567		rec_argv[i] = argv[j];
1568
1569	BUG_ON(i != rec_argc);
1570
1571	return cmd_buildid_list(i, rec_argv);
1572}
1573
1574int cmd_kvm(int argc, const char **argv)
1575{
1576	const char *file_name = NULL;
1577	const struct option kvm_options[] = {
1578		OPT_STRING('i', "input", &file_name, "file",
1579			   "Input file name"),
1580		OPT_STRING('o', "output", &file_name, "file",
1581			   "Output file name"),
1582		OPT_BOOLEAN(0, "guest", &perf_guest,
1583			    "Collect guest os data"),
1584		OPT_BOOLEAN(0, "host", &perf_host,
1585			    "Collect host os data"),
1586		OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
1587			   "guest mount directory under which every guest os"
1588			   " instance has a subdir"),
1589		OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
1590			   "file", "file saving guest os vmlinux"),
1591		OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
1592			   "file", "file saving guest os /proc/kallsyms"),
1593		OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
1594			   "file", "file saving guest os /proc/modules"),
 
 
1595		OPT_INCR('v', "verbose", &verbose,
1596			    "be more verbose (show counter open errors, etc)"),
1597		OPT_END()
1598	};
1599
1600	const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
1601						"buildid-list", "stat", NULL };
1602	const char *kvm_usage[] = { NULL, NULL };
1603
1604	perf_host  = 0;
1605	perf_guest = 1;
1606
1607	argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
1608					PARSE_OPT_STOP_AT_NON_OPTION);
1609	if (!argc)
1610		usage_with_options(kvm_usage, kvm_options);
1611
1612	if (!perf_host)
1613		perf_guest = 1;
1614
1615	if (!file_name) {
1616		file_name = get_filename_for_perf_kvm();
1617
1618		if (!file_name) {
1619			pr_err("Failed to allocate memory for filename\n");
1620			return -ENOMEM;
1621		}
1622	}
1623
1624	if (!strncmp(argv[0], "rec", 3))
1625		return __cmd_record(file_name, argc, argv);
1626	else if (!strncmp(argv[0], "rep", 3))
1627		return __cmd_report(file_name, argc, argv);
1628	else if (!strncmp(argv[0], "diff", 4))
1629		return cmd_diff(argc, argv);
1630	else if (!strncmp(argv[0], "top", 3))
1631		return cmd_top(argc, argv);
1632	else if (!strncmp(argv[0], "buildid-list", 12))
1633		return __cmd_buildid_list(file_name, argc, argv);
1634#ifdef HAVE_KVM_STAT_SUPPORT
1635	else if (!strncmp(argv[0], "stat", 4))
1636		return kvm_cmd_stat(file_name, argc, argv);
1637#endif
1638	else
1639		usage_with_options(kvm_usage, kvm_options);
1640
1641	return 0;
1642}