builtin-kvm.c

   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 "util/util.h"
  27#include "ui/browsers/hists.h"
  28#include "ui/progress.h"
  29#include "ui/ui.h"
  30#include "util/string2.h"
  31
  32#include <sys/prctl.h>
  33#ifdef HAVE_TIMERFD_SUPPORT
  34#include <sys/timerfd.h>
  35#endif
  36#include <sys/time.h>
  37#include <sys/types.h>
  38#include <sys/stat.h>
  39#include <fcntl.h>
  40
  41#include <linux/err.h>
  42#include <linux/kernel.h>
  43#include <linux/string.h>
  44#include <linux/time64.h>
  45#include <linux/zalloc.h>
  46#include <errno.h>
  47#include <inttypes.h>
  48#include <poll.h>
  49#include <termios.h>
  50#include <semaphore.h>
  51#include <signal.h>
  52#include <math.h>
  53#include <perf/mmap.h>
  54
  55#if defined(HAVE_KVM_STAT_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
  56#define GET_EVENT_KEY(func, field)					\
  57static u64 get_event_ ##func(struct kvm_event *event, int vcpu)		\
  58{									\
  59	if (vcpu == -1)							\
  60		return event->total.field;				\
  61									\
  62	if (vcpu >= event->max_vcpu)					\
  63		return 0;						\
  64									\
  65	return event->vcpu[vcpu].field;					\
  66}
  67
  68#define COMPARE_EVENT_KEY(func, field)					\
  69GET_EVENT_KEY(func, field)						\
  70static int64_t cmp_event_ ## func(struct kvm_event *one,		\
  71			      struct kvm_event *two, int vcpu)		\
  72{									\
  73	return get_event_ ##func(one, vcpu) -				\
  74	       get_event_ ##func(two, vcpu);				\
  75}
  76
  77COMPARE_EVENT_KEY(time, time);
  78COMPARE_EVENT_KEY(max, stats.max);
  79COMPARE_EVENT_KEY(min, stats.min);
  80COMPARE_EVENT_KEY(count, stats.n);
  81COMPARE_EVENT_KEY(mean, stats.mean);
  82
  83struct kvm_hists {
  84	struct hists		hists;
  85	struct perf_hpp_list	list;
  86};
  87
  88struct kvm_dimension {
  89	const char *name;
  90	const char *header;
  91	int width;
  92	int64_t (*cmp)(struct perf_hpp_fmt *fmt, struct hist_entry *left,
  93		       struct hist_entry *right);
  94	int (*entry)(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
  95		     struct hist_entry *he);
  96};
  97
  98struct kvm_fmt {
  99	struct perf_hpp_fmt	fmt;
 100	struct kvm_dimension	*dim;
 101};
 102
 103static struct kvm_hists kvm_hists;
 104
 105static int64_t ev_name_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
 106			   struct hist_entry *left,
 107			   struct hist_entry *right)
 108{
 109	/* Return opposite number for sorting in alphabetical order */
 110	return -strcmp(left->kvm_info->name, right->kvm_info->name);
 111}
 112
 113static int fmt_width(struct perf_hpp_fmt *fmt,
 114		     struct perf_hpp *hpp __maybe_unused,
 115		     struct hists *hists __maybe_unused);
 116
 117static int ev_name_entry(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 118			 struct hist_entry *he)
 119{
 120	int width = fmt_width(fmt, hpp, he->hists);
 121
 122	return scnprintf(hpp->buf, hpp->size, "%*s", width, he->kvm_info->name);
 123}
 124
 125static struct kvm_dimension dim_event = {
 126	.header		= "Event name",
 127	.name		= "ev_name",
 128	.cmp		= ev_name_cmp,
 129	.entry		= ev_name_entry,
 130	.width		= 40,
 131};
 132
 133#define EV_METRIC_CMP(metric)						\
 134static int64_t ev_cmp_##metric(struct perf_hpp_fmt *fmt __maybe_unused,	\
 135			       struct hist_entry *left,			\
 136			       struct hist_entry *right)		\
 137{									\
 138	struct kvm_event *event_left;					\
 139	struct kvm_event *event_right;					\
 140	struct perf_kvm_stat *perf_kvm;					\
 141									\
 142	event_left  = container_of(left, struct kvm_event, he);		\
 143	event_right = container_of(right, struct kvm_event, he);	\
 144									\
 145	perf_kvm = event_left->perf_kvm;				\
 146	return cmp_event_##metric(event_left, event_right,		\
 147				  perf_kvm->trace_vcpu);		\
 148}
 149
 150EV_METRIC_CMP(time)
 151EV_METRIC_CMP(count)
 152EV_METRIC_CMP(max)
 153EV_METRIC_CMP(min)
 154EV_METRIC_CMP(mean)
 155
 156#define EV_METRIC_ENTRY(metric)						\
 157static int ev_entry_##metric(struct perf_hpp_fmt *fmt,			\
 158			     struct perf_hpp *hpp,			\
 159			     struct hist_entry *he)			\
 160{									\
 161	struct kvm_event *event;					\
 162	int width = fmt_width(fmt, hpp, he->hists);			\
 163	struct perf_kvm_stat *perf_kvm;					\
 164									\
 165	event = container_of(he, struct kvm_event, he);			\
 166	perf_kvm = event->perf_kvm;					\
 167	return scnprintf(hpp->buf, hpp->size, "%*lu", width,		\
 168		get_event_##metric(event, perf_kvm->trace_vcpu));	\
 169}
 170
 171EV_METRIC_ENTRY(time)
 172EV_METRIC_ENTRY(count)
 173EV_METRIC_ENTRY(max)
 174EV_METRIC_ENTRY(min)
 175
 176static struct kvm_dimension dim_time = {
 177	.header		= "Time (ns)",
 178	.name		= "time",
 179	.cmp		= ev_cmp_time,
 180	.entry		= ev_entry_time,
 181	.width		= 12,
 182};
 183
 184static struct kvm_dimension dim_count = {
 185	.header		= "Samples",
 186	.name		= "sample",
 187	.cmp		= ev_cmp_count,
 188	.entry		= ev_entry_count,
 189	.width		= 12,
 190};
 191
 192static struct kvm_dimension dim_max_time = {
 193	.header		= "Max Time (ns)",
 194	.name		= "max_t",
 195	.cmp		= ev_cmp_max,
 196	.entry		= ev_entry_max,
 197	.width		= 14,
 198};
 199
 200static struct kvm_dimension dim_min_time = {
 201	.header		= "Min Time (ns)",
 202	.name		= "min_t",
 203	.cmp		= ev_cmp_min,
 204	.entry		= ev_entry_min,
 205	.width		= 14,
 206};
 207
 208static int ev_entry_mean(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 209			 struct hist_entry *he)
 210{
 211	struct kvm_event *event;
 212	int width = fmt_width(fmt, hpp, he->hists);
 213	struct perf_kvm_stat *perf_kvm;
 214
 215	event = container_of(he, struct kvm_event, he);
 216	perf_kvm = event->perf_kvm;
 217	return scnprintf(hpp->buf, hpp->size, "%*lu", width,
 218			 get_event_mean(event, perf_kvm->trace_vcpu));
 219}
 220
 221static struct kvm_dimension dim_mean_time = {
 222	.header		= "Mean Time (ns)",
 223	.name		= "mean_t",
 224	.cmp		= ev_cmp_mean,
 225	.entry		= ev_entry_mean,
 226	.width		= 14,
 227};
 228
 229#define PERC_STR(__s, __v)				\
 230({							\
 231	scnprintf(__s, sizeof(__s), "%.2F%%", __v);	\
 232	__s;						\
 233})
 234
 235static double percent(u64 st, u64 tot)
 236{
 237	return tot ? 100. * (double) st / (double) tot : 0;
 238}
 239
 240#define EV_METRIC_PERCENT(metric)					\
 241static int ev_percent_##metric(struct hist_entry *he)			\
 242{									\
 243	struct kvm_event *event;					\
 244	struct perf_kvm_stat *perf_kvm;					\
 245									\
 246	event = container_of(he, struct kvm_event, he);			\
 247	perf_kvm = event->perf_kvm;					\
 248									\
 249	return percent(get_event_##metric(event, perf_kvm->trace_vcpu),	\
 250		       perf_kvm->total_##metric);			\
 251}
 252
 253EV_METRIC_PERCENT(time)
 254EV_METRIC_PERCENT(count)
 255
 256static int ev_entry_time_precent(struct perf_hpp_fmt *fmt,
 257				 struct perf_hpp *hpp,
 258				 struct hist_entry *he)
 259{
 260	int width = fmt_width(fmt, hpp, he->hists);
 261	double per;
 262	char buf[10];
 263
 264	per = ev_percent_time(he);
 265	return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
 266}
 267
 268static int64_t
 269ev_cmp_time_precent(struct perf_hpp_fmt *fmt __maybe_unused,
 270		    struct hist_entry *left, struct hist_entry *right)
 271{
 272	double per_left;
 273	double per_right;
 274
 275	per_left  = ev_percent_time(left);
 276	per_right = ev_percent_time(right);
 277
 278	return per_left - per_right;
 279}
 280
 281static struct kvm_dimension dim_time_percent = {
 282	.header		= "Time%",
 283	.name		= "percent_time",
 284	.cmp		= ev_cmp_time_precent,
 285	.entry		= ev_entry_time_precent,
 286	.width		= 12,
 287};
 288
 289static int ev_entry_count_precent(struct perf_hpp_fmt *fmt,
 290				  struct perf_hpp *hpp,
 291				  struct hist_entry *he)
 292{
 293	int width = fmt_width(fmt, hpp, he->hists);
 294	double per;
 295	char buf[10];
 296
 297	per = ev_percent_count(he);
 298	return scnprintf(hpp->buf, hpp->size, "%*s", width, PERC_STR(buf, per));
 299}
 300
 301static int64_t
 302ev_cmp_count_precent(struct perf_hpp_fmt *fmt __maybe_unused,
 303		     struct hist_entry *left, struct hist_entry *right)
 304{
 305	double per_left;
 306	double per_right;
 307
 308	per_left  = ev_percent_count(left);
 309	per_right = ev_percent_count(right);
 310
 311	return per_left - per_right;
 312}
 313
 314static struct kvm_dimension dim_count_percent = {
 315	.header		= "Sample%",
 316	.name		= "percent_sample",
 317	.cmp		= ev_cmp_count_precent,
 318	.entry		= ev_entry_count_precent,
 319	.width		= 12,
 320};
 321
 322static struct kvm_dimension *dimensions[] = {
 323	&dim_event,
 324	&dim_time,
 325	&dim_time_percent,
 326	&dim_count,
 327	&dim_count_percent,
 328	&dim_max_time,
 329	&dim_min_time,
 330	&dim_mean_time,
 331	NULL,
 332};
 333
 334static int fmt_width(struct perf_hpp_fmt *fmt,
 335		     struct perf_hpp *hpp __maybe_unused,
 336		     struct hists *hists __maybe_unused)
 337{
 338	struct kvm_fmt *kvm_fmt;
 339
 340	kvm_fmt = container_of(fmt, struct kvm_fmt, fmt);
 341	return kvm_fmt->dim->width;
 342}
 343
 344static int fmt_header(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
 345		      struct hists *hists, int line __maybe_unused,
 346		      int *span __maybe_unused)
 347{
 348	struct kvm_fmt *kvm_fmt;
 349	struct kvm_dimension *dim;
 350	int width = fmt_width(fmt, hpp, hists);
 351
 352	kvm_fmt = container_of(fmt, struct kvm_fmt, fmt);
 353	dim = kvm_fmt->dim;
 354
 355	return scnprintf(hpp->buf, hpp->size, "%*s", width, dim->header);
 356}
 357
 358static bool fmt_equal(struct perf_hpp_fmt *a, struct perf_hpp_fmt *b)
 359{
 360	struct kvm_fmt *kvm_fmt_a = container_of(a, struct kvm_fmt, fmt);
 361	struct kvm_fmt *kvm_fmt_b = container_of(b, struct kvm_fmt, fmt);
 362
 363	return kvm_fmt_a->dim == kvm_fmt_b->dim;
 364}
 365
 366static void fmt_free(struct perf_hpp_fmt *fmt)
 367{
 368	struct kvm_fmt *kvm_fmt;
 369
 370	kvm_fmt = container_of(fmt, struct kvm_fmt, fmt);
 371	free(kvm_fmt);
 372}
 373
 374static struct kvm_dimension *get_dimension(const char *name)
 375{
 376	unsigned int i;
 377
 378	for (i = 0; dimensions[i] != NULL; i++) {
 379		if (!strcmp(dimensions[i]->name, name))
 380			return dimensions[i];
 381	}
 382
 383	return NULL;
 384}
 385
 386static struct kvm_fmt *get_format(const char *name)
 387{
 388	struct kvm_dimension *dim = get_dimension(name);
 389	struct kvm_fmt *kvm_fmt;
 390	struct perf_hpp_fmt *fmt;
 391
 392	if (!dim)
 393		return NULL;
 394
 395	kvm_fmt = zalloc(sizeof(*kvm_fmt));
 396	if (!kvm_fmt)
 397		return NULL;
 398
 399	kvm_fmt->dim = dim;
 400
 401	fmt = &kvm_fmt->fmt;
 402	INIT_LIST_HEAD(&fmt->list);
 403	INIT_LIST_HEAD(&fmt->sort_list);
 404	fmt->cmp	= dim->cmp;
 405	fmt->sort	= dim->cmp;
 406	fmt->color	= NULL;
 407	fmt->entry	= dim->entry;
 408	fmt->header	= fmt_header;
 409	fmt->width	= fmt_width;
 410	fmt->collapse	= dim->cmp;
 411	fmt->equal	= fmt_equal;
 412	fmt->free	= fmt_free;
 413
 414	return kvm_fmt;
 415}
 416
 417static int kvm_hists__init_output(struct perf_hpp_list *hpp_list, char *name)
 418{
 419	struct kvm_fmt *kvm_fmt = get_format(name);
 420
 421	if (!kvm_fmt) {
 422		pr_warning("Fail to find format for output field %s.\n", name);
 423		return -EINVAL;
 424	}
 425
 426	perf_hpp_list__column_register(hpp_list, &kvm_fmt->fmt);
 427	return 0;
 428}
 429
 430static int kvm_hists__init_sort(struct perf_hpp_list *hpp_list, char *name)
 431{
 432	struct kvm_fmt *kvm_fmt = get_format(name);
 433
 434	if (!kvm_fmt) {
 435		pr_warning("Fail to find format for sorting %s.\n", name);
 436		return -EINVAL;
 437	}
 438
 439	perf_hpp_list__register_sort_field(hpp_list, &kvm_fmt->fmt);
 440	return 0;
 441}
 442
 443static int kvm_hpp_list__init(char *list,
 444			      struct perf_hpp_list *hpp_list,
 445			      int (*fn)(struct perf_hpp_list *hpp_list,
 446					char *name))
 447{
 448	char *tmp, *tok;
 449	int ret;
 450
 451	if (!list || !fn)
 452		return 0;
 453
 454	for (tok = strtok_r(list, ", ", &tmp); tok;
 455	     tok = strtok_r(NULL, ", ", &tmp)) {
 456		ret = fn(hpp_list, tok);
 457		if (!ret)
 458			continue;
 459
 460		/* Handle errors */
 461		if (ret == -EINVAL)
 462			pr_err("Invalid field key: '%s'", tok);
 463		else if (ret == -ESRCH)
 464			pr_err("Unknown field key: '%s'", tok);
 465		else
 466			pr_err("Fail to initialize for field key: '%s'", tok);
 467
 468		break;
 469	}
 470
 471	return ret;
 472}
 473
 474static int kvm_hpp_list__parse(struct perf_hpp_list *hpp_list,
 475			       const char *output_, const char *sort_)
 476{
 477	char *output = output_ ? strdup(output_) : NULL;
 478	char *sort = sort_ ? strdup(sort_) : NULL;
 479	int ret;
 480
 481	ret = kvm_hpp_list__init(output, hpp_list, kvm_hists__init_output);
 482	if (ret)
 483		goto out;
 484
 485	ret = kvm_hpp_list__init(sort, hpp_list, kvm_hists__init_sort);
 486	if (ret)
 487		goto out;
 488
 489	/* Copy sort keys to output fields */
 490	perf_hpp__setup_output_field(hpp_list);
 491
 492	/* and then copy output fields to sort keys */
 493	perf_hpp__append_sort_keys(hpp_list);
 494out:
 495	free(output);
 496	free(sort);
 497	return ret;
 498}
 499
 500static int kvm_hists__init(void)
 501{
 502	kvm_hists.list.nr_header_lines = 1;
 503	__hists__init(&kvm_hists.hists, &kvm_hists.list);
 504	perf_hpp_list__init(&kvm_hists.list);
 505	return kvm_hpp_list__parse(&kvm_hists.list, NULL, "ev_name");
 506}
 507
 508static int kvm_hists__reinit(const char *output, const char *sort)
 509{
 510	perf_hpp__reset_output_field(&kvm_hists.list);
 511	return kvm_hpp_list__parse(&kvm_hists.list, output, sort);
 512}
 513static void print_result(struct perf_kvm_stat *kvm);
 514
 515#ifdef HAVE_SLANG_SUPPORT
 516static void kvm_browser__update_nr_entries(struct hist_browser *hb)
 517{
 518	struct rb_node *nd = rb_first_cached(&hb->hists->entries);
 519	u64 nr_entries = 0;
 520
 521	for (; nd; nd = rb_next(nd)) {
 522		struct hist_entry *he = rb_entry(nd, struct hist_entry,
 523						 rb_node);
 524
 525		if (!he->filtered)
 526			nr_entries++;
 527	}
 528
 529	hb->nr_non_filtered_entries = nr_entries;
 530}
 531
 532static int kvm_browser__title(struct hist_browser *browser,
 533			      char *buf, size_t size)
 534{
 535	scnprintf(buf, size, "KVM event statistics (%lu entries)",
 536		  browser->nr_non_filtered_entries);
 537	return 0;
 538}
 539
 540static struct hist_browser*
 541perf_kvm_browser__new(struct hists *hists)
 542{
 543	struct hist_browser *browser = hist_browser__new(hists);
 544
 545	if (browser)
 546		browser->title = kvm_browser__title;
 547
 548	return browser;
 549}
 550
 551static int kvm__hists_browse(struct hists *hists)
 552{
 553	struct hist_browser *browser;
 554	int key = -1;
 555
 556	browser = perf_kvm_browser__new(hists);
 557	if (browser == NULL)
 558		return -1;
 559
 560	/* reset abort key so that it can get Ctrl-C as a key */
 561	SLang_reset_tty();
 562	SLang_init_tty(0, 0, 0);
 563
 564	kvm_browser__update_nr_entries(browser);
 565
 566	while (1) {
 567		key = hist_browser__run(browser, "? - help", true, 0);
 568
 569		switch (key) {
 570		case 'q':
 571			goto out;
 572		default:
 573			break;
 574		}
 575	}
 576
 577out:
 578	hist_browser__delete(browser);
 579	return 0;
 580}
 581
 582static void kvm_display(struct perf_kvm_stat *kvm)
 583{
 584	if (!use_browser)
 585		print_result(kvm);
 586	else
 587		kvm__hists_browse(&kvm_hists.hists);
 588}
 589
 590#else
 591
 592static void kvm_display(struct perf_kvm_stat *kvm)
 593{
 594	use_browser = 0;
 595	print_result(kvm);
 596}
 597
 598#endif /* HAVE_SLANG_SUPPORT */
 599
 600#endif // defined(HAVE_KVM_STAT_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
 601
 602static const char *get_filename_for_perf_kvm(void)
 603{
 604	const char *filename;
 605
 606	if (perf_host && !perf_guest)
 607		filename = strdup("perf.data.host");
 608	else if (!perf_host && perf_guest)
 609		filename = strdup("perf.data.guest");
 610	else
 611		filename = strdup("perf.data.kvm");
 612
 613	return filename;
 614}
 615
 616#if defined(HAVE_KVM_STAT_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
 617
 618static bool register_kvm_events_ops(struct perf_kvm_stat *kvm)
 619{
 620	struct kvm_reg_events_ops *events_ops = kvm_reg_events_ops;
 621
 622	for (events_ops = kvm_reg_events_ops; events_ops->name; events_ops++) {
 623		if (!strcmp(events_ops->name, kvm->report_event)) {
 624			kvm->events_ops = events_ops->ops;
 625			return true;
 626		}
 627	}
 628
 629	return false;
 630}
 631
 632struct vcpu_event_record {
 633	int vcpu_id;
 634	u64 start_time;
 635	struct kvm_event *last_event;
 636};
 637
 638#ifdef HAVE_TIMERFD_SUPPORT
 639static void clear_events_cache_stats(void)
 640{
 641	struct rb_root_cached *root;
 642	struct rb_node *nd;
 643	struct kvm_event *event;
 644	int i;
 645
 646	if (hists__has(&kvm_hists.hists, need_collapse))
 647		root = &kvm_hists.hists.entries_collapsed;
 648	else
 649		root = kvm_hists.hists.entries_in;
 650
 651	for (nd = rb_first_cached(root); nd; nd = rb_next(nd)) {
 652		struct hist_entry *he;
 653
 654		he = rb_entry(nd, struct hist_entry, rb_node_in);
 655		event = container_of(he, struct kvm_event, he);
 656
 657		/* reset stats for event */
 658		event->total.time = 0;
 659		init_stats(&event->total.stats);
 660
 661		for (i = 0; i < event->max_vcpu; ++i) {
 662			event->vcpu[i].time = 0;
 663			init_stats(&event->vcpu[i].stats);
 664		}
 665	}
 666}
 667#endif
 668
 669static bool kvm_event_expand(struct kvm_event *event, int vcpu_id)
 670{
 671	int old_max_vcpu = event->max_vcpu;
 672	void *prev;
 673
 674	if (vcpu_id < event->max_vcpu)
 675		return true;
 676
 677	while (event->max_vcpu <= vcpu_id)
 678		event->max_vcpu += DEFAULT_VCPU_NUM;
 679
 680	prev = event->vcpu;
 681	event->vcpu = realloc(event->vcpu,
 682			      event->max_vcpu * sizeof(*event->vcpu));
 683	if (!event->vcpu) {
 684		free(prev);
 685		pr_err("Not enough memory\n");
 686		return false;
 687	}
 688
 689	memset(event->vcpu + old_max_vcpu, 0,
 690	       (event->max_vcpu - old_max_vcpu) * sizeof(*event->vcpu));
 691	return true;
 692}
 693
 694static void *kvm_he_zalloc(size_t size)
 695{
 696	struct kvm_event *kvm_ev;
 697
 698	kvm_ev = zalloc(size + sizeof(*kvm_ev));
 699	if (!kvm_ev)
 700		return NULL;
 701
 702	init_stats(&kvm_ev->total.stats);
 703	hists__inc_nr_samples(&kvm_hists.hists, 0);
 704	return &kvm_ev->he;
 705}
 706
 707static void kvm_he_free(void *he)
 708{
 709	struct kvm_event *kvm_ev;
 710
 711	kvm_ev = container_of(he, struct kvm_event, he);
 712	free(kvm_ev);
 713}
 714
 715static struct hist_entry_ops kvm_ev_entry_ops = {
 716	.new	= kvm_he_zalloc,
 717	.free	= kvm_he_free,
 718};
 719
 720static struct kvm_event *find_create_kvm_event(struct perf_kvm_stat *kvm,
 721					       struct event_key *key,
 722					       struct perf_sample *sample)
 723{
 724	struct kvm_event *event;
 725	struct hist_entry *he;
 726	struct kvm_info *ki;
 727
 728	BUG_ON(key->key == INVALID_KEY);
 729
 730	ki = kvm_info__new();
 731	if (!ki) {
 732		pr_err("Failed to allocate kvm info\n");
 733		return NULL;
 734	}
 735
 736	kvm->events_ops->decode_key(kvm, key, ki->name);
 737	he = hists__add_entry_ops(&kvm_hists.hists, &kvm_ev_entry_ops,
 738				  &kvm->al, NULL, NULL, NULL, ki, sample, true);
 739	if (he == NULL) {
 740		pr_err("Failed to allocate hist entry\n");
 741		free(ki);
 742		return NULL;
 743	}
 744
 745	event = container_of(he, struct kvm_event, he);
 746	if (!event->perf_kvm) {
 747		event->perf_kvm = kvm;
 748		event->key = *key;
 749	}
 750
 751	return event;
 752}
 753
 754static bool handle_begin_event(struct perf_kvm_stat *kvm,
 755			       struct vcpu_event_record *vcpu_record,
 756			       struct event_key *key,
 757			       struct perf_sample *sample)
 758{
 759	struct kvm_event *event = NULL;
 760
 761	if (key->key != INVALID_KEY)
 762		event = find_create_kvm_event(kvm, key, sample);
 763
 764	vcpu_record->last_event = event;
 765	vcpu_record->start_time = sample->time;
 766	return true;
 767}
 768
 769static void
 770kvm_update_event_stats(struct kvm_event_stats *kvm_stats, u64 time_diff)
 771{
 772	kvm_stats->time += time_diff;
 773	update_stats(&kvm_stats->stats, time_diff);
 774}
 775
 776static double kvm_event_rel_stddev(int vcpu_id, struct kvm_event *event)
 777{
 778	struct kvm_event_stats *kvm_stats = &event->total;
 779
 780	if (vcpu_id != -1)
 781		kvm_stats = &event->vcpu[vcpu_id];
 782
 783	return rel_stddev_stats(stddev_stats(&kvm_stats->stats),
 784				avg_stats(&kvm_stats->stats));
 785}
 786
 787static bool update_kvm_event(struct perf_kvm_stat *kvm,
 788			     struct kvm_event *event, int vcpu_id,
 789			     u64 time_diff)
 790{
 791	/* Update overall statistics */
 792	kvm->total_count++;
 793	kvm->total_time += time_diff;
 794
 795	if (vcpu_id == -1) {
 796		kvm_update_event_stats(&event->total, time_diff);
 797		return true;
 798	}
 799
 800	if (!kvm_event_expand(event, vcpu_id))
 801		return false;
 802
 803	kvm_update_event_stats(&event->vcpu[vcpu_id], time_diff);
 804	return true;
 805}
 806
 807static bool is_child_event(struct perf_kvm_stat *kvm,
 808			   struct evsel *evsel,
 809			   struct perf_sample *sample,
 810			   struct event_key *key)
 811{
 812	struct child_event_ops *child_ops;
 813
 814	child_ops = kvm->events_ops->child_ops;
 815
 816	if (!child_ops)
 817		return false;
 818
 819	for (; child_ops->name; child_ops++) {
 820		if (evsel__name_is(evsel, child_ops->name)) {
 821			child_ops->get_key(evsel, sample, key);
 822			return true;
 823		}
 824	}
 825
 826	return false;
 827}
 828
 829static bool handle_child_event(struct perf_kvm_stat *kvm,
 830			       struct vcpu_event_record *vcpu_record,
 831			       struct event_key *key,
 832			       struct perf_sample *sample)
 833{
 834	struct kvm_event *event = NULL;
 835
 836	if (key->key != INVALID_KEY)
 837		event = find_create_kvm_event(kvm, key, sample);
 838
 839	vcpu_record->last_event = event;
 840
 841	return true;
 842}
 843
 844static bool skip_event(const char *event)
 845{
 846	const char * const *skip_events;
 847
 848	for (skip_events = kvm_skip_events; *skip_events; skip_events++)
 849		if (!strcmp(event, *skip_events))
 850			return true;
 851
 852	return false;
 853}
 854
 855static bool handle_end_event(struct perf_kvm_stat *kvm,
 856			     struct vcpu_event_record *vcpu_record,
 857			     struct event_key *key,
 858			     struct perf_sample *sample)
 859{
 860	struct kvm_event *event;
 861	u64 time_begin, time_diff;
 862	int vcpu;
 863
 864	if (kvm->trace_vcpu == -1)
 865		vcpu = -1;
 866	else
 867		vcpu = vcpu_record->vcpu_id;
 868
 869	event = vcpu_record->last_event;
 870	time_begin = vcpu_record->start_time;
 871
 872	/* The begin event is not caught. */
 873	if (!time_begin)
 874		return true;
 875
 876	/*
 877	 * In some case, the 'begin event' only records the start timestamp,
 878	 * the actual event is recognized in the 'end event' (e.g. mmio-event).
 879	 */
 880
 881	/* Both begin and end events did not get the key. */
 882	if (!event && key->key == INVALID_KEY)
 883		return true;
 884
 885	if (!event)
 886		event = find_create_kvm_event(kvm, key, sample);
 887
 888	if (!event)
 889		return false;
 890
 891	vcpu_record->last_event = NULL;
 892	vcpu_record->start_time = 0;
 893
 894	/* seems to happen once in a while during live mode */
 895	if (sample->time < time_begin) {
 896		pr_debug("End time before begin time; skipping event.\n");
 897		return true;
 898	}
 899
 900	time_diff = sample->time - time_begin;
 901
 902	if (kvm->duration && time_diff > kvm->duration) {
 903		char decode[KVM_EVENT_NAME_LEN];
 904
 905		kvm->events_ops->decode_key(kvm, &event->key, decode);
 906		if (!skip_event(decode)) {
 907			pr_info("%" PRIu64 " VM %d, vcpu %d: %s event took %" PRIu64 "usec\n",
 908				 sample->time, sample->pid, vcpu_record->vcpu_id,
 909				 decode, time_diff / NSEC_PER_USEC);
 910		}
 911	}
 912
 913	return update_kvm_event(kvm, event, vcpu, time_diff);
 914}
 915
 916static
 917struct vcpu_event_record *per_vcpu_record(struct thread *thread,
 918					  struct evsel *evsel,
 919					  struct perf_sample *sample)
 920{
 921	/* Only kvm_entry records vcpu id. */
 922	if (!thread__priv(thread) && kvm_entry_event(evsel)) {
 923		struct vcpu_event_record *vcpu_record;
 924
 925		vcpu_record = zalloc(sizeof(*vcpu_record));
 926		if (!vcpu_record) {
 927			pr_err("%s: Not enough memory\n", __func__);
 928			return NULL;
 929		}
 930
 931		vcpu_record->vcpu_id = evsel__intval(evsel, sample, vcpu_id_str);
 932		thread__set_priv(thread, vcpu_record);
 933	}
 934
 935	return thread__priv(thread);
 936}
 937
 938static bool handle_kvm_event(struct perf_kvm_stat *kvm,
 939			     struct thread *thread,
 940			     struct evsel *evsel,
 941			     struct perf_sample *sample)
 942{
 943	struct vcpu_event_record *vcpu_record;
 944	struct event_key key = { .key = INVALID_KEY,
 945				 .exit_reasons = kvm->exit_reasons };
 946
 947	vcpu_record = per_vcpu_record(thread, evsel, sample);
 948	if (!vcpu_record)
 949		return true;
 950
 951	/* only process events for vcpus user cares about */
 952	if ((kvm->trace_vcpu != -1) &&
 953	    (kvm->trace_vcpu != vcpu_record->vcpu_id))
 954		return true;
 955
 956	if (kvm->events_ops->is_begin_event(evsel, sample, &key))
 957		return handle_begin_event(kvm, vcpu_record, &key, sample);
 958
 959	if (is_child_event(kvm, evsel, sample, &key))
 960		return handle_child_event(kvm, vcpu_record, &key, sample);
 961
 962	if (kvm->events_ops->is_end_event(evsel, sample, &key))
 963		return handle_end_event(kvm, vcpu_record, &key, sample);
 964
 965	return true;
 966}
 967
 968static bool is_valid_key(struct perf_kvm_stat *kvm)
 969{
 970	static const char *key_array[] = {
 971		"ev_name", "sample", "time", "max_t", "min_t", "mean_t",
 972	};
 973	unsigned int i;
 974
 975	for (i = 0; i < ARRAY_SIZE(key_array); i++)
 976		if (!strcmp(key_array[i], kvm->sort_key))
 977			return true;
 978
 979	pr_err("Unsupported sort key: %s\n", kvm->sort_key);
 980	return false;
 981}
 982
 983static bool event_is_valid(struct kvm_event *event, int vcpu)
 984{
 985	return !!get_event_count(event, vcpu);
 986}
 987
 988static int filter_cb(struct hist_entry *he, void *arg __maybe_unused)
 989{
 990	struct kvm_event *event;
 991	struct perf_kvm_stat *perf_kvm;
 992
 993	event = container_of(he, struct kvm_event, he);
 994	perf_kvm = event->perf_kvm;
 995	if (!event_is_valid(event, perf_kvm->trace_vcpu))
 996		he->filtered = 1;
 997	else
 998		he->filtered = 0;
 999	return 0;
1000}
1001
1002static void sort_result(struct perf_kvm_stat *kvm)
1003{
1004	struct ui_progress prog;
1005	const char *output_columns = "ev_name,sample,percent_sample,"
1006				     "time,percent_time,max_t,min_t,mean_t";
1007
1008	kvm_hists__reinit(output_columns, kvm->sort_key);
1009	ui_progress__init(&prog, kvm_hists.hists.nr_entries, "Sorting...");
1010	hists__collapse_resort(&kvm_hists.hists, NULL);
1011	hists__output_resort_cb(&kvm_hists.hists, NULL, filter_cb);
1012	ui_progress__finish();
1013}
1014
1015static void print_vcpu_info(struct perf_kvm_stat *kvm)
1016{
1017	int vcpu = kvm->trace_vcpu;
1018
1019	pr_info("Analyze events for ");
1020
1021	if (kvm->opts.target.system_wide)
1022		pr_info("all VMs, ");
1023	else if (kvm->opts.target.pid)
1024		pr_info("pid(s) %s, ", kvm->opts.target.pid);
1025	else
1026		pr_info("dazed and confused on what is monitored, ");
1027
1028	if (vcpu == -1)
1029		pr_info("all VCPUs:\n\n");
1030	else
1031		pr_info("VCPU %d:\n\n", vcpu);
1032}
1033
1034static void show_timeofday(void)
1035{
1036	char date[64];
1037	struct timeval tv;
1038	struct tm ltime;
1039
1040	gettimeofday(&tv, NULL);
1041	if (localtime_r(&tv.tv_sec, &ltime)) {
1042		strftime(date, sizeof(date), "%H:%M:%S", &ltime);
1043		pr_info("%s.%06ld", date, tv.tv_usec);
1044	} else
1045		pr_info("00:00:00.000000");
1046
1047	return;
1048}
1049
1050static void print_result(struct perf_kvm_stat *kvm)
1051{
1052	char decode[KVM_EVENT_NAME_LEN];
1053	struct kvm_event *event;
1054	int vcpu = kvm->trace_vcpu;
1055	struct rb_node *nd;
1056
1057	if (kvm->live) {
1058		puts(CONSOLE_CLEAR);
1059		show_timeofday();
1060	}
1061
1062	pr_info("\n\n");
1063	print_vcpu_info(kvm);
1064	pr_info("%*s ", KVM_EVENT_NAME_LEN, kvm->events_ops->name);
1065	pr_info("%10s ", "Samples");
1066	pr_info("%9s ", "Samples%");
1067
1068	pr_info("%9s ", "Time%");
1069	pr_info("%11s ", "Min Time");
1070	pr_info("%11s ", "Max Time");
1071	pr_info("%16s ", "Avg time");
1072	pr_info("\n\n");
1073
1074	for (nd = rb_first_cached(&kvm_hists.hists.entries); nd; nd = rb_next(nd)) {
1075		struct hist_entry *he;
1076		u64 ecount, etime, max, min;
1077
1078		he = rb_entry(nd, struct hist_entry, rb_node);
1079		if (he->filtered)
1080			continue;
1081
1082		event = container_of(he, struct kvm_event, he);
1083		ecount = get_event_count(event, vcpu);
1084		etime = get_event_time(event, vcpu);
1085		max = get_event_max(event, vcpu);
1086		min = get_event_min(event, vcpu);
1087
1088		kvm->events_ops->decode_key(kvm, &event->key, decode);
1089		pr_info("%*s ", KVM_EVENT_NAME_LEN, decode);
1090		pr_info("%10llu ", (unsigned long long)ecount);
1091		pr_info("%8.2f%% ", (double)ecount / kvm->total_count * 100);
1092		pr_info("%8.2f%% ", (double)etime / kvm->total_time * 100);
1093		pr_info("%9.2fus ", (double)min / NSEC_PER_USEC);
1094		pr_info("%9.2fus ", (double)max / NSEC_PER_USEC);
1095		pr_info("%9.2fus ( +-%7.2f%% )", (double)etime / ecount / NSEC_PER_USEC,
1096			kvm_event_rel_stddev(vcpu, event));
1097		pr_info("\n");
1098	}
1099
1100	pr_info("\nTotal Samples:%" PRIu64 ", Total events handled time:%.2fus.\n\n",
1101		kvm->total_count, kvm->total_time / (double)NSEC_PER_USEC);
1102
1103	if (kvm->lost_events)
1104		pr_info("\nLost events: %" PRIu64 "\n\n", kvm->lost_events);
1105}
1106
1107#if defined(HAVE_TIMERFD_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
1108static int process_lost_event(const struct perf_tool *tool,
1109			      union perf_event *event __maybe_unused,
1110			      struct perf_sample *sample __maybe_unused,
1111			      struct machine *machine __maybe_unused)
1112{
1113	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat, tool);
1114
1115	kvm->lost_events++;
1116	return 0;
1117}
1118#endif
1119
1120static bool skip_sample(struct perf_kvm_stat *kvm,
1121			struct perf_sample *sample)
1122{
1123	if (kvm->pid_list && intlist__find(kvm->pid_list, sample->pid) == NULL)
1124		return true;
1125
1126	return false;
1127}
1128
1129static int process_sample_event(const struct perf_tool *tool,
1130				union perf_event *event,
1131				struct perf_sample *sample,
1132				struct evsel *evsel,
1133				struct machine *machine)
1134{
1135	int err = 0;
1136	struct thread *thread;
1137	struct perf_kvm_stat *kvm = container_of(tool, struct perf_kvm_stat,
1138						 tool);
1139
1140	if (skip_sample(kvm, sample))
1141		return 0;
1142
1143	if (machine__resolve(machine, &kvm->al, sample) < 0) {
1144		pr_warning("Fail to resolve address location, skip sample.\n");
1145		return 0;
1146	}
1147
1148	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
1149	if (thread == NULL) {
1150		pr_debug("problem processing %d event, skipping it.\n",
1151			event->header.type);
1152		return -1;
1153	}
1154
1155	if (!handle_kvm_event(kvm, thread, evsel, sample))
1156		err = -1;
1157
1158	thread__put(thread);
1159	return err;
1160}
1161
1162static int cpu_isa_config(struct perf_kvm_stat *kvm)
1163{
1164	char buf[128], *cpuid;
1165	int err;
1166
1167	if (kvm->live) {
1168		struct perf_cpu cpu = {-1};
1169
1170		err = get_cpuid(buf, sizeof(buf), cpu);
1171		if (err != 0) {
1172			pr_err("Failed to look up CPU type: %s\n",
1173			       str_error_r(err, buf, sizeof(buf)));
1174			return -err;
1175		}
1176		cpuid = buf;
1177	} else
1178		cpuid = perf_session__env(kvm->session)->cpuid;
1179
1180	if (!cpuid) {
1181		pr_err("Failed to look up CPU type\n");
1182		return -EINVAL;
1183	}
1184
1185	err = cpu_isa_init(kvm, cpuid);
1186	if (err == -ENOTSUP)
1187		pr_err("CPU %s is not supported.\n", cpuid);
1188
1189	return err;
1190}
1191
1192static bool verify_vcpu(int vcpu)
1193{
1194	if (vcpu != -1 && vcpu < 0) {
1195		pr_err("Invalid vcpu:%d.\n", vcpu);
1196		return false;
1197	}
1198
1199	return true;
1200}
1201
1202#if defined(HAVE_TIMERFD_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
1203/* keeping the max events to a modest level to keep
1204 * the processing of samples per mmap smooth.
1205 */
1206#define PERF_KVM__MAX_EVENTS_PER_MMAP  25
1207
1208static s64 perf_kvm__mmap_read_idx(struct perf_kvm_stat *kvm, int idx,
1209				   u64 *mmap_time)
1210{
1211	struct evlist *evlist = kvm->evlist;
1212	union perf_event *event;
1213	struct mmap *md;
1214	u64 timestamp;
1215	s64 n = 0;
1216	int err;
1217
1218	*mmap_time = ULLONG_MAX;
1219	md = &evlist->mmap[idx];
1220	err = perf_mmap__read_init(&md->core);
1221	if (err < 0)
1222		return (err == -EAGAIN) ? 0 : -1;
1223
1224	while ((event = perf_mmap__read_event(&md->core)) != NULL) {
1225		err = evlist__parse_sample_timestamp(evlist, event, &timestamp);
1226		if (err) {
1227			perf_mmap__consume(&md->core);
1228			pr_err("Failed to parse sample\n");
1229			return -1;
1230		}
1231
1232		err = perf_session__queue_event(kvm->session, event, timestamp, 0, NULL);
1233		/*
1234		 * FIXME: Here we can't consume the event, as perf_session__queue_event will
1235		 *        point to it, and it'll get possibly overwritten by the kernel.
1236		 */
1237		perf_mmap__consume(&md->core);
1238
1239		if (err) {
1240			pr_err("Failed to enqueue sample: %d\n", err);
1241			return -1;
1242		}
1243
1244		/* save time stamp of our first sample for this mmap */
1245		if (n == 0)
1246			*mmap_time = timestamp;
1247
1248		/* limit events per mmap handled all at once */
1249		n++;
1250		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
1251			break;
1252	}
1253
1254	perf_mmap__read_done(&md->core);
1255	return n;
1256}
1257
1258static int perf_kvm__mmap_read(struct perf_kvm_stat *kvm)
1259{
1260	int i, err, throttled = 0;
1261	s64 n, ntotal = 0;
1262	u64 flush_time = ULLONG_MAX, mmap_time;
1263
1264	for (i = 0; i < kvm->evlist->core.nr_mmaps; i++) {
1265		n = perf_kvm__mmap_read_idx(kvm, i, &mmap_time);
1266		if (n < 0)
1267			return -1;
1268
1269		/* flush time is going to be the minimum of all the individual
1270		 * mmap times. Essentially, we flush all the samples queued up
1271		 * from the last pass under our minimal start time -- that leaves
1272		 * a very small race for samples to come in with a lower timestamp.
1273		 * The ioctl to return the perf_clock timestamp should close the
1274		 * race entirely.
1275		 */
1276		if (mmap_time < flush_time)
1277			flush_time = mmap_time;
1278
1279		ntotal += n;
1280		if (n == PERF_KVM__MAX_EVENTS_PER_MMAP)
1281			throttled = 1;
1282	}
1283
1284	/* flush queue after each round in which we processed events */
1285	if (ntotal) {
1286		struct ordered_events *oe = &kvm->session->ordered_events;
1287
1288		oe->next_flush = flush_time;
1289		err = ordered_events__flush(oe, OE_FLUSH__ROUND);
1290		if (err) {
1291			if (kvm->lost_events)
1292				pr_info("\nLost events: %" PRIu64 "\n\n",
1293					kvm->lost_events);
1294			return err;
1295		}
1296	}
1297
1298	return throttled;
1299}
1300
1301static volatile int done;
1302
1303static void sig_handler(int sig __maybe_unused)
1304{
1305	done = 1;
1306}
1307
1308static int perf_kvm__timerfd_create(struct perf_kvm_stat *kvm)
1309{
1310	struct itimerspec new_value;
1311	int rc = -1;
1312
1313	kvm->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
1314	if (kvm->timerfd < 0) {
1315		pr_err("timerfd_create failed\n");
1316		goto out;
1317	}
1318
1319	new_value.it_value.tv_sec = kvm->display_time;
1320	new_value.it_value.tv_nsec = 0;
1321	new_value.it_interval.tv_sec = kvm->display_time;
1322	new_value.it_interval.tv_nsec = 0;
1323
1324	if (timerfd_settime(kvm->timerfd, 0, &new_value, NULL) != 0) {
1325		pr_err("timerfd_settime failed: %d\n", errno);
1326		close(kvm->timerfd);
1327		goto out;
1328	}
1329
1330	rc = 0;
1331out:
1332	return rc;
1333}
1334
1335static int perf_kvm__handle_timerfd(struct perf_kvm_stat *kvm)
1336{
1337	uint64_t c;
1338	int rc;
1339
1340	rc = read(kvm->timerfd, &c, sizeof(uint64_t));
1341	if (rc < 0) {
1342		if (errno == EAGAIN)
1343			return 0;
1344
1345		pr_err("Failed to read timer fd: %d\n", errno);
1346		return -1;
1347	}
1348
1349	if (rc != sizeof(uint64_t)) {
1350		pr_err("Error reading timer fd - invalid size returned\n");
1351		return -1;
1352	}
1353
1354	if (c != 1)
1355		pr_debug("Missed timer beats: %" PRIu64 "\n", c-1);
1356
1357	/* update display */
1358	sort_result(kvm);
1359	print_result(kvm);
1360
1361	/* Reset sort list to "ev_name" */
1362	kvm_hists__reinit(NULL, "ev_name");
1363
1364	/* reset counts */
1365	clear_events_cache_stats();
1366	kvm->total_count = 0;
1367	kvm->total_time = 0;
1368	kvm->lost_events = 0;
1369
1370	return 0;
1371}
1372
1373static int fd_set_nonblock(int fd)
1374{
1375	long arg = 0;
1376
1377	arg = fcntl(fd, F_GETFL);
1378	if (arg < 0) {
1379		pr_err("Failed to get current flags for fd %d\n", fd);
1380		return -1;
1381	}
1382
1383	if (fcntl(fd, F_SETFL, arg | O_NONBLOCK) < 0) {
1384		pr_err("Failed to set non-block option on fd %d\n", fd);
1385		return -1;
1386	}
1387
1388	return 0;
1389}
1390
1391static int perf_kvm__handle_stdin(void)
1392{
1393	int c;
1394
1395	c = getc(stdin);
1396	if (c == 'q')
1397		return 1;
1398
1399	return 0;
1400}
1401
1402static int kvm_events_live_report(struct perf_kvm_stat *kvm)
1403{
1404	int nr_stdin, ret, err = -EINVAL;
1405	struct termios save;
1406
1407	/* live flag must be set first */
1408	kvm->live = true;
1409
1410	ret = cpu_isa_config(kvm);
1411	if (ret < 0)
1412		return ret;
1413
1414	if (!verify_vcpu(kvm->trace_vcpu) ||
1415	    !is_valid_key(kvm) ||
1416	    !register_kvm_events_ops(kvm)) {
1417		goto out;
1418	}
1419
1420	set_term_quiet_input(&save);
1421
1422	kvm_hists__init();
1423
1424	signal(SIGINT, sig_handler);
1425	signal(SIGTERM, sig_handler);
1426
1427	/* add timer fd */
1428	if (perf_kvm__timerfd_create(kvm) < 0) {
1429		err = -1;
1430		goto out;
1431	}
1432
1433	if (evlist__add_pollfd(kvm->evlist, kvm->timerfd) < 0)
1434		goto out;
1435
1436	nr_stdin = evlist__add_pollfd(kvm->evlist, fileno(stdin));
1437	if (nr_stdin < 0)
1438		goto out;
1439
1440	if (fd_set_nonblock(fileno(stdin)) != 0)
1441		goto out;
1442
1443	/* everything is good - enable the events and process */
1444	evlist__enable(kvm->evlist);
1445
1446	while (!done) {
1447		struct fdarray *fda = &kvm->evlist->core.pollfd;
1448		int rc;
1449
1450		rc = perf_kvm__mmap_read(kvm);
1451		if (rc < 0)
1452			break;
1453
1454		err = perf_kvm__handle_timerfd(kvm);
1455		if (err)
1456			goto out;
1457
1458		if (fda->entries[nr_stdin].revents & POLLIN)
1459			done = perf_kvm__handle_stdin();
1460
1461		if (!rc && !done)
1462			err = evlist__poll(kvm->evlist, 100);
1463	}
1464
1465	evlist__disable(kvm->evlist);
1466
1467	if (err == 0) {
1468		sort_result(kvm);
1469		print_result(kvm);
1470	}
1471
1472out:
1473	hists__delete_entries(&kvm_hists.hists);
1474
1475	if (kvm->timerfd >= 0)
1476		close(kvm->timerfd);
1477
1478	tcsetattr(0, TCSAFLUSH, &save);
1479	return err;
1480}
1481
1482static int kvm_live_open_events(struct perf_kvm_stat *kvm)
1483{
1484	int err, rc = -1;
1485	struct evsel *pos;
1486	struct evlist *evlist = kvm->evlist;
1487	char sbuf[STRERR_BUFSIZE];
1488
1489	evlist__config(evlist, &kvm->opts, NULL);
1490
1491	/*
1492	 * Note: exclude_{guest,host} do not apply here.
1493	 *       This command processes KVM tracepoints from host only
1494	 */
1495	evlist__for_each_entry(evlist, pos) {
1496		struct perf_event_attr *attr = &pos->core.attr;
1497
1498		/* make sure these *are* set */
1499		evsel__set_sample_bit(pos, TID);
1500		evsel__set_sample_bit(pos, TIME);
1501		evsel__set_sample_bit(pos, CPU);
1502		evsel__set_sample_bit(pos, RAW);
1503		/* make sure these are *not*; want as small a sample as possible */
1504		evsel__reset_sample_bit(pos, PERIOD);
1505		evsel__reset_sample_bit(pos, IP);
1506		evsel__reset_sample_bit(pos, CALLCHAIN);
1507		evsel__reset_sample_bit(pos, ADDR);
1508		evsel__reset_sample_bit(pos, READ);
1509		attr->mmap = 0;
1510		attr->comm = 0;
1511		attr->task = 0;
1512
1513		attr->sample_period = 1;
1514
1515		attr->watermark = 0;
1516		attr->wakeup_events = 1000;
1517
1518		/* will enable all once we are ready */
1519		attr->disabled = 1;
1520	}
1521
1522	err = evlist__open(evlist);
1523	if (err < 0) {
1524		printf("Couldn't create the events: %s\n",
1525		       str_error_r(errno, sbuf, sizeof(sbuf)));
1526		goto out;
1527	}
1528
1529	if (evlist__mmap(evlist, kvm->opts.mmap_pages) < 0) {
1530		ui__error("Failed to mmap the events: %s\n",
1531			  str_error_r(errno, sbuf, sizeof(sbuf)));
1532		evlist__close(evlist);
1533		goto out;
1534	}
1535
1536	rc = 0;
1537
1538out:
1539	return rc;
1540}
1541#endif
1542
1543static int read_events(struct perf_kvm_stat *kvm)
1544{
1545	int ret;
1546
1547	struct perf_data file = {
1548		.path  = kvm->file_name,
1549		.mode  = PERF_DATA_MODE_READ,
1550		.force = kvm->force,
1551	};
1552
1553	perf_tool__init(&kvm->tool, /*ordered_events=*/true);
1554	kvm->tool.sample	= process_sample_event;
1555	kvm->tool.comm		= perf_event__process_comm;
1556	kvm->tool.namespaces	= perf_event__process_namespaces;
1557
1558	kvm->session = perf_session__new(&file, &kvm->tool);
1559	if (IS_ERR(kvm->session)) {
1560		pr_err("Initializing perf session failed\n");
1561		return PTR_ERR(kvm->session);
1562	}
1563
1564	symbol__init(perf_session__env(kvm->session));
1565
1566	if (!perf_session__has_traces(kvm->session, "kvm record")) {
1567		ret = -EINVAL;
1568		goto out_delete;
1569	}
1570
1571	/*
1572	 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1573	 * traced in the old kernel.
1574	 */
1575	ret = cpu_isa_config(kvm);
1576	if (ret < 0)
1577		goto out_delete;
1578
1579	ret = perf_session__process_events(kvm->session);
1580
1581out_delete:
1582	perf_session__delete(kvm->session);
1583	return ret;
1584}
1585
1586static int parse_target_str(struct perf_kvm_stat *kvm)
1587{
1588	if (kvm->opts.target.pid) {
1589		kvm->pid_list = intlist__new(kvm->opts.target.pid);
1590		if (kvm->pid_list == NULL) {
1591			pr_err("Error parsing process id string\n");
1592			return -EINVAL;
1593		}
1594	}
1595
1596	return 0;
1597}
1598
1599static int kvm_events_report_vcpu(struct perf_kvm_stat *kvm)
1600{
1601	int ret = -EINVAL;
1602	int vcpu = kvm->trace_vcpu;
1603
1604	if (parse_target_str(kvm) != 0)
1605		goto exit;
1606
1607	if (!verify_vcpu(vcpu))
1608		goto exit;
1609
1610	if (!is_valid_key(kvm))
1611		goto exit;
1612
1613	if (!register_kvm_events_ops(kvm))
1614		goto exit;
1615
1616	if (kvm->use_stdio) {
1617		use_browser = 0;
1618		setup_pager();
1619	} else {
1620		use_browser = 1;
1621	}
1622
1623	setup_browser(false);
1624
1625	kvm_hists__init();
1626
1627	ret = read_events(kvm);
1628	if (ret)
1629		goto exit;
1630
1631	sort_result(kvm);
1632	kvm_display(kvm);
1633
1634exit:
1635	hists__delete_entries(&kvm_hists.hists);
1636	return ret;
1637}
1638
1639int __weak setup_kvm_events_tp(struct perf_kvm_stat *kvm __maybe_unused)
1640{
1641	return 0;
1642}
1643
1644static int
1645kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
1646{
1647	unsigned int rec_argc, i, j, events_tp_size;
1648	const char **rec_argv;
1649	const char * const record_args[] = {
1650		"record",
1651		"-R",
1652		"-m", "1024",
1653		"-c", "1",
1654	};
1655	const char * const kvm_stat_record_usage[] = {
1656		"perf kvm stat record [<options>]",
1657		NULL
1658	};
1659	const char * const *events_tp;
1660	int ret;
1661
1662	events_tp_size = 0;
1663	ret = setup_kvm_events_tp(kvm);
1664	if (ret < 0) {
1665		pr_err("Unable to setup the kvm tracepoints\n");
1666		return ret;
1667	}
1668
1669	for (events_tp = kvm_events_tp; *events_tp; events_tp++)
1670		events_tp_size++;
1671
1672	rec_argc = ARRAY_SIZE(record_args) + argc + 2 +
1673		   2 * events_tp_size;
1674	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1675
1676	if (rec_argv == NULL)
1677		return -ENOMEM;
1678
1679	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1680		rec_argv[i] = STRDUP_FAIL_EXIT(record_args[i]);
1681
1682	for (j = 0; j < events_tp_size; j++) {
1683		rec_argv[i++] = STRDUP_FAIL_EXIT("-e");
1684		rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
1685	}
1686
1687	rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
1688	rec_argv[i++] = STRDUP_FAIL_EXIT(kvm->file_name);
1689
1690	for (j = 1; j < (unsigned int)argc; j++, i++)
1691		rec_argv[i] = STRDUP_FAIL_EXIT(argv[j]);
1692
1693	set_option_flag(record_options, 'e', "event", PARSE_OPT_HIDDEN);
1694	set_option_flag(record_options, 0, "filter", PARSE_OPT_HIDDEN);
1695	set_option_flag(record_options, 'R', "raw-samples", PARSE_OPT_HIDDEN);
1696
1697	set_option_flag(record_options, 'F', "freq", PARSE_OPT_DISABLED);
1698	set_option_flag(record_options, 0, "group", PARSE_OPT_DISABLED);
1699	set_option_flag(record_options, 'g', NULL, PARSE_OPT_DISABLED);
1700	set_option_flag(record_options, 0, "call-graph", PARSE_OPT_DISABLED);
1701	set_option_flag(record_options, 'd', "data", PARSE_OPT_DISABLED);
1702	set_option_flag(record_options, 'T', "timestamp", PARSE_OPT_DISABLED);
1703	set_option_flag(record_options, 'P', "period", PARSE_OPT_DISABLED);
1704	set_option_flag(record_options, 'n', "no-samples", PARSE_OPT_DISABLED);
1705	set_option_flag(record_options, 'N', "no-buildid-cache", PARSE_OPT_DISABLED);
1706	set_option_flag(record_options, 'B', "no-buildid", PARSE_OPT_DISABLED);
1707	set_option_flag(record_options, 'G', "cgroup", PARSE_OPT_DISABLED);
1708	set_option_flag(record_options, 'b', "branch-any", PARSE_OPT_DISABLED);
1709	set_option_flag(record_options, 'j', "branch-filter", PARSE_OPT_DISABLED);
1710	set_option_flag(record_options, 'W', "weight", PARSE_OPT_DISABLED);
1711	set_option_flag(record_options, 0, "transaction", PARSE_OPT_DISABLED);
1712
1713	record_usage = kvm_stat_record_usage;
1714	ret = cmd_record(i, rec_argv);
1715
1716EXIT:
1717	for (i = 0; i < rec_argc; i++)
1718		free((void *)rec_argv[i]);
1719	free(rec_argv);
1720	return ret;
1721}
1722
1723static int
1724kvm_events_report(struct perf_kvm_stat *kvm, int argc, const char **argv)
1725{
1726	const struct option kvm_events_report_options[] = {
1727		OPT_STRING(0, "event", &kvm->report_event, "report event",
1728			   "event for reporting: vmexit, "
1729			   "mmio (x86 only), ioport (x86 only)"),
1730		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1731			    "vcpu id to report"),
1732		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1733			    "key for sorting: sample(sort by samples number)"
1734			    " time (sort by avg time)"),
1735		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1736			   "analyze events only for given process id(s)"),
1737		OPT_BOOLEAN('f', "force", &kvm->force, "don't complain, do it"),
1738		OPT_BOOLEAN(0, "stdio", &kvm->use_stdio, "use the stdio interface"),
1739		OPT_END()
1740	};
1741
1742	const char * const kvm_events_report_usage[] = {
1743		"perf kvm stat report [<options>]",
1744		NULL
1745	};
1746
1747	if (argc) {
1748		argc = parse_options(argc, argv,
1749				     kvm_events_report_options,
1750				     kvm_events_report_usage, 0);
1751		if (argc)
1752			usage_with_options(kvm_events_report_usage,
1753					   kvm_events_report_options);
1754	}
1755
1756#ifndef HAVE_SLANG_SUPPORT
1757	kvm->use_stdio = true;
1758#endif
1759
1760	if (!kvm->opts.target.pid)
1761		kvm->opts.target.system_wide = true;
1762
1763	return kvm_events_report_vcpu(kvm);
1764}
1765
1766#if defined(HAVE_TIMERFD_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
1767static struct evlist *kvm_live_event_list(void)
1768{
1769	struct evlist *evlist;
1770	char *tp, *name, *sys;
1771	int err = -1;
1772	const char * const *events_tp;
1773
1774	evlist = evlist__new();
1775	if (evlist == NULL)
1776		return NULL;
1777
1778	for (events_tp = kvm_events_tp; *events_tp; events_tp++) {
1779
1780		tp = strdup(*events_tp);
1781		if (tp == NULL)
1782			goto out;
1783
1784		/* split tracepoint into subsystem and name */
1785		sys = tp;
1786		name = strchr(tp, ':');
1787		if (name == NULL) {
1788			pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1789			       *events_tp);
1790			free(tp);
1791			goto out;
1792		}
1793		*name = '\0';
1794		name++;
1795
1796		if (evlist__add_newtp(evlist, sys, name, NULL)) {
1797			pr_err("Failed to add %s tracepoint to the list\n", *events_tp);
1798			free(tp);
1799			goto out;
1800		}
1801
1802		free(tp);
1803	}
1804
1805	err = 0;
1806
1807out:
1808	if (err) {
1809		evlist__delete(evlist);
1810		evlist = NULL;
1811	}
1812
1813	return evlist;
1814}
1815
1816static int kvm_events_live(struct perf_kvm_stat *kvm,
1817			   int argc, const char **argv)
1818{
1819	char errbuf[BUFSIZ];
1820	int err;
1821
1822	const struct option live_options[] = {
1823		OPT_STRING('p', "pid", &kvm->opts.target.pid, "pid",
1824			"record events on existing process id"),
1825		OPT_CALLBACK('m', "mmap-pages", &kvm->opts.mmap_pages, "pages",
1826			"number of mmap data pages", evlist__parse_mmap_pages),
1827		OPT_INCR('v', "verbose", &verbose,
1828			"be more verbose (show counter open errors, etc)"),
1829		OPT_BOOLEAN('a', "all-cpus", &kvm->opts.target.system_wide,
1830			"system-wide collection from all CPUs"),
1831		OPT_UINTEGER('d', "display", &kvm->display_time,
1832			"time in seconds between display updates"),
1833		OPT_STRING(0, "event", &kvm->report_event, "report event",
1834			"event for reporting: "
1835			"vmexit, mmio (x86 only), ioport (x86 only)"),
1836		OPT_INTEGER(0, "vcpu", &kvm->trace_vcpu,
1837			"vcpu id to report"),
1838		OPT_STRING('k', "key", &kvm->sort_key, "sort-key",
1839			"key for sorting: sample(sort by samples number)"
1840			" time (sort by avg time)"),
1841		OPT_U64(0, "duration", &kvm->duration,
1842			"show events other than"
1843			" HLT (x86 only) or Wait state (s390 only)"
1844			" that take longer than duration usecs"),
1845		OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1846				"per thread proc mmap processing timeout in ms"),
1847		OPT_END()
1848	};
1849	const char * const live_usage[] = {
1850		"perf kvm stat live [<options>]",
1851		NULL
1852	};
1853	struct perf_data data = {
1854		.mode = PERF_DATA_MODE_WRITE,
1855	};
1856
1857
1858	/* event handling */
1859	perf_tool__init(&kvm->tool, /*ordered_events=*/true);
1860	kvm->tool.sample = process_sample_event;
1861	kvm->tool.comm   = perf_event__process_comm;
1862	kvm->tool.exit   = perf_event__process_exit;
1863	kvm->tool.fork   = perf_event__process_fork;
1864	kvm->tool.lost   = process_lost_event;
1865	kvm->tool.namespaces  = perf_event__process_namespaces;
1866
1867	/* set defaults */
1868	kvm->display_time = 1;
1869	kvm->opts.user_interval = 1;
1870	kvm->opts.mmap_pages = 512;
1871	kvm->opts.target.uses_mmap = false;
1872
1873	symbol__init(NULL);
1874	disable_buildid_cache();
1875
1876	use_browser = 0;
1877
1878	if (argc) {
1879		argc = parse_options(argc, argv, live_options,
1880				     live_usage, 0);
1881		if (argc)
1882			usage_with_options(live_usage, live_options);
1883	}
1884
1885	kvm->duration *= NSEC_PER_USEC;   /* convert usec to nsec */
1886
1887	/*
1888	 * target related setups
1889	 */
1890	err = target__validate(&kvm->opts.target);
1891	if (err) {
1892		target__strerror(&kvm->opts.target, err, errbuf, BUFSIZ);
1893		ui__warning("%s", errbuf);
1894	}
1895
1896	if (target__none(&kvm->opts.target))
1897		kvm->opts.target.system_wide = true;
1898
1899
1900	/*
1901	 * generate the event list
1902	 */
1903	err = setup_kvm_events_tp(kvm);
1904	if (err < 0) {
1905		pr_err("Unable to setup the kvm tracepoints\n");
1906		return err;
1907	}
1908
1909	kvm->evlist = kvm_live_event_list();
1910	if (kvm->evlist == NULL) {
1911		err = -1;
1912		goto out;
1913	}
1914
1915	if (evlist__create_maps(kvm->evlist, &kvm->opts.target) < 0)
1916		usage_with_options(live_usage, live_options);
1917
1918	/*
1919	 * perf session
1920	 */
1921	kvm->session = perf_session__new(&data, &kvm->tool);
1922	if (IS_ERR(kvm->session)) {
1923		err = PTR_ERR(kvm->session);
1924		goto out;
1925	}
1926	kvm->session->evlist = kvm->evlist;
1927	perf_session__set_id_hdr_size(kvm->session);
1928	ordered_events__set_copy_on_queue(&kvm->session->ordered_events, true);
1929	machine__synthesize_threads(&kvm->session->machines.host, &kvm->opts.target,
1930				    kvm->evlist->core.threads, true, false, 1);
1931	err = kvm_live_open_events(kvm);
1932	if (err)
1933		goto out;
1934
1935	err = kvm_events_live_report(kvm);
1936
1937out:
1938	perf_session__delete(kvm->session);
1939	kvm->session = NULL;
1940	evlist__delete(kvm->evlist);
1941
1942	return err;
1943}
1944#endif
1945
1946static void print_kvm_stat_usage(void)
1947{
1948	printf("Usage: perf kvm stat <command>\n\n");
1949
1950	printf("# Available commands:\n");
1951	printf("\trecord: record kvm events\n");
1952	printf("\treport: report statistical data of kvm events\n");
1953	printf("\tlive:   live reporting of statistical data of kvm events\n");
1954
1955	printf("\nOtherwise, it is the alias of 'perf stat':\n");
1956}
1957
1958static int kvm_cmd_stat(const char *file_name, int argc, const char **argv)
1959{
1960	struct perf_kvm_stat kvm = {
1961		.file_name = file_name,
1962
1963		.trace_vcpu	= -1,
1964		.report_event	= "vmexit",
1965		.sort_key	= "sample",
1966
1967	};
1968
1969	if (argc == 1) {
1970		print_kvm_stat_usage();
1971		goto perf_stat;
1972	}
1973
1974	if (strlen(argv[1]) > 2 && strstarts("record", argv[1]))
1975		return kvm_events_record(&kvm, argc - 1, argv + 1);
1976
1977	if (strlen(argv[1]) > 2 && strstarts("report", argv[1]))
1978		return kvm_events_report(&kvm, argc - 1 , argv + 1);
1979
1980#if defined(HAVE_TIMERFD_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
1981	if (!strncmp(argv[1], "live", 4))
1982		return kvm_events_live(&kvm, argc - 1 , argv + 1);
1983#endif
1984
1985perf_stat:
1986	return cmd_stat(argc, argv);
1987}
1988#endif /* HAVE_KVM_STAT_SUPPORT */
1989
1990int __weak kvm_add_default_arch_event(int *argc __maybe_unused,
1991					const char **argv __maybe_unused)
1992{
1993	return 0;
1994}
1995
1996static int __cmd_record(const char *file_name, int argc, const char **argv)
1997{
1998	int rec_argc, i = 0, j, ret;
1999	const char **rec_argv;
2000
2001	/*
2002	 * Besides the 2 more options "-o" and "filename",
2003	 * kvm_add_default_arch_event() may add 2 extra options,
2004	 * so allocate 4 more items.
2005	 */
2006	rec_argc = argc + 2 + 2;
2007	rec_argv = calloc(rec_argc + 1, sizeof(char *));
2008	if (!rec_argv)
2009		return -ENOMEM;
2010
2011	rec_argv[i++] = STRDUP_FAIL_EXIT("record");
2012	rec_argv[i++] = STRDUP_FAIL_EXIT("-o");
2013	rec_argv[i++] = STRDUP_FAIL_EXIT(file_name);
2014	for (j = 1; j < argc; j++, i++)
2015		rec_argv[i] = STRDUP_FAIL_EXIT(argv[j]);
2016
2017	BUG_ON(i + 2 != rec_argc);
2018
2019	ret = kvm_add_default_arch_event(&i, rec_argv);
2020	if (ret)
2021		goto EXIT;
2022
2023	ret = cmd_record(i, rec_argv);
2024
2025EXIT:
2026	for (i = 0; i < rec_argc; i++)
2027		free((void *)rec_argv[i]);
2028	free(rec_argv);
2029	return ret;
2030}
2031
2032static int __cmd_report(const char *file_name, int argc, const char **argv)
2033{
2034	int rec_argc, i = 0, j, ret;
2035	const char **rec_argv;
2036
2037	rec_argc = argc + 2;
2038	rec_argv = calloc(rec_argc + 1, sizeof(char *));
2039	if (!rec_argv)
2040		return -ENOMEM;
2041
2042	rec_argv[i++] = STRDUP_FAIL_EXIT("report");
2043	rec_argv[i++] = STRDUP_FAIL_EXIT("-i");
2044	rec_argv[i++] = STRDUP_FAIL_EXIT(file_name);
2045	for (j = 1; j < argc; j++, i++)
2046		rec_argv[i] = STRDUP_FAIL_EXIT(argv[j]);
2047
2048	BUG_ON(i != rec_argc);
2049
2050	ret = cmd_report(i, rec_argv);
2051
2052EXIT:
2053	for (i = 0; i < rec_argc; i++)
2054		free((void *)rec_argv[i]);
2055	free(rec_argv);
2056	return ret;
2057}
2058
2059static int
2060__cmd_buildid_list(const char *file_name, int argc, const char **argv)
2061{
2062	int rec_argc, i = 0, j, ret;
2063	const char **rec_argv;
2064
2065	rec_argc = argc + 2;
2066	rec_argv = calloc(rec_argc + 1, sizeof(char *));
2067	if (!rec_argv)
2068		return -ENOMEM;
2069
2070	rec_argv[i++] = STRDUP_FAIL_EXIT("buildid-list");
2071	rec_argv[i++] = STRDUP_FAIL_EXIT("-i");
2072	rec_argv[i++] = STRDUP_FAIL_EXIT(file_name);
2073	for (j = 1; j < argc; j++, i++)
2074		rec_argv[i] = STRDUP_FAIL_EXIT(argv[j]);
2075
2076	BUG_ON(i != rec_argc);
2077
2078	ret = cmd_buildid_list(i, rec_argv);
2079
2080EXIT:
2081	for (i = 0; i < rec_argc; i++)
2082		free((void *)rec_argv[i]);
2083	free(rec_argv);
2084	return ret;
2085}
2086
2087static int __cmd_top(int argc, const char **argv)
2088{
2089	int rec_argc, i = 0, ret;
2090	const char **rec_argv;
2091
2092	/*
2093	 * kvm_add_default_arch_event() may add 2 extra options, so
2094	 * allocate 2 more pointers in adavance.
2095	 */
2096	rec_argc = argc + 2;
2097	rec_argv = calloc(rec_argc + 1, sizeof(char *));
2098	if (!rec_argv)
2099		return -ENOMEM;
2100
2101	for (i = 0; i < argc; i++)
2102		rec_argv[i] = STRDUP_FAIL_EXIT(argv[i]);
2103
2104	BUG_ON(i != argc);
2105
2106	ret = kvm_add_default_arch_event(&i, rec_argv);
2107	if (ret)
2108		goto EXIT;
2109
2110	ret = cmd_top(i, rec_argv);
2111
2112EXIT:
2113	for (i = 0; i < rec_argc; i++)
2114		free((void *)rec_argv[i]);
2115	free(rec_argv);
2116	return ret;
2117}
2118
2119int cmd_kvm(int argc, const char **argv)
2120{
2121	const char *file_name = NULL;
2122	const struct option kvm_options[] = {
2123		OPT_STRING('i', "input", &file_name, "file",
2124			   "Input file name"),
2125		OPT_STRING('o', "output", &file_name, "file",
2126			   "Output file name"),
2127		OPT_BOOLEAN(0, "guest", &perf_guest,
2128			    "Collect guest os data"),
2129		OPT_BOOLEAN(0, "host", &perf_host,
2130			    "Collect host os data"),
2131		OPT_STRING(0, "guestmount", &symbol_conf.guestmount, "directory",
2132			   "guest mount directory under which every guest os"
2133			   " instance has a subdir"),
2134		OPT_STRING(0, "guestvmlinux", &symbol_conf.default_guest_vmlinux_name,
2135			   "file", "file saving guest os vmlinux"),
2136		OPT_STRING(0, "guestkallsyms", &symbol_conf.default_guest_kallsyms,
2137			   "file", "file saving guest os /proc/kallsyms"),
2138		OPT_STRING(0, "guestmodules", &symbol_conf.default_guest_modules,
2139			   "file", "file saving guest os /proc/modules"),
2140		OPT_BOOLEAN(0, "guest-code", &symbol_conf.guest_code,
2141			    "Guest code can be found in hypervisor process"),
2142		OPT_INCR('v', "verbose", &verbose,
2143			    "be more verbose (show counter open errors, etc)"),
2144		OPT_END()
2145	};
2146
2147	const char *const kvm_subcommands[] = { "top", "record", "report", "diff",
2148						"buildid-list", "stat", NULL };
2149	const char *kvm_usage[] = { NULL, NULL };
2150
2151	exclude_GH_default = true;
2152	perf_host  = 0;
2153	perf_guest = 1;
2154
2155	argc = parse_options_subcommand(argc, argv, kvm_options, kvm_subcommands, kvm_usage,
2156					PARSE_OPT_STOP_AT_NON_OPTION);
2157	if (!argc)
2158		usage_with_options(kvm_usage, kvm_options);
2159
2160	if (!perf_host)
2161		perf_guest = 1;
2162
2163	if (!file_name) {
2164		file_name = get_filename_for_perf_kvm();
2165
2166		if (!file_name) {
2167			pr_err("Failed to allocate memory for filename\n");
2168			return -ENOMEM;
2169		}
2170	}
2171
2172	if (strlen(argv[0]) > 2 && strstarts("record", argv[0]))
2173		return __cmd_record(file_name, argc, argv);
2174	else if (strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2175		return __cmd_report(file_name, argc, argv);
2176	else if (strlen(argv[0]) > 2 && strstarts("diff", argv[0]))
2177		return cmd_diff(argc, argv);
2178	else if (!strcmp(argv[0], "top"))
2179		return __cmd_top(argc, argv);
2180	else if (strlen(argv[0]) > 2 && strstarts("buildid-list", argv[0]))
2181		return __cmd_buildid_list(file_name, argc, argv);
2182#if defined(HAVE_KVM_STAT_SUPPORT) && defined(HAVE_LIBTRACEEVENT)
2183	else if (strlen(argv[0]) > 2 && strstarts("stat", argv[0]))
2184		return kvm_cmd_stat(file_name, argc, argv);
2185#endif
2186	else
2187		usage_with_options(kvm_usage, kvm_options);
2188
2189	/* free usage string allocated by parse_options_subcommand */
2190	free((void *)kvm_usage[0]);
2191
2192	return 0;
2193}