Linux Audio

Check our new training course

Loading...
v3.1
   1#include "builtin.h"
   2#include "perf.h"
   3
 
 
   4#include "util/util.h"
   5#include "util/cache.h"
   6#include "util/symbol.h"
   7#include "util/thread.h"
   8#include "util/header.h"
   9
  10#include "util/parse-options.h"
  11#include "util/trace-event.h"
  12
  13#include "util/debug.h"
  14#include "util/session.h"
 
 
  15
  16#include <sys/types.h>
  17#include <sys/prctl.h>
  18#include <semaphore.h>
  19#include <pthread.h>
  20#include <math.h>
  21#include <limits.h>
  22
  23#include <linux/list.h>
  24#include <linux/hash.h>
  25
  26static struct perf_session *session;
  27
  28/* based on kernel/lockdep.c */
  29#define LOCKHASH_BITS		12
  30#define LOCKHASH_SIZE		(1UL << LOCKHASH_BITS)
  31
  32static struct list_head lockhash_table[LOCKHASH_SIZE];
  33
  34#define __lockhashfn(key)	hash_long((unsigned long)key, LOCKHASH_BITS)
  35#define lockhashentry(key)	(lockhash_table + __lockhashfn((key)))
  36
  37struct lock_stat {
  38	struct list_head	hash_entry;
  39	struct rb_node		rb;		/* used for sorting */
  40
  41	/*
  42	 * FIXME: raw_field_value() returns unsigned long long,
  43	 * so address of lockdep_map should be dealed as 64bit.
  44	 * Is there more better solution?
  45	 */
  46	void			*addr;		/* address of lockdep_map, used as ID */
  47	char			*name;		/* for strcpy(), we cannot use const */
  48
  49	unsigned int		nr_acquire;
  50	unsigned int		nr_acquired;
  51	unsigned int		nr_contended;
  52	unsigned int		nr_release;
  53
  54	unsigned int		nr_readlock;
  55	unsigned int		nr_trylock;
 
  56	/* these times are in nano sec. */
 
  57	u64			wait_time_total;
  58	u64			wait_time_min;
  59	u64			wait_time_max;
  60
  61	int			discard; /* flag of blacklist */
  62};
  63
  64/*
  65 * States of lock_seq_stat
  66 *
  67 * UNINITIALIZED is required for detecting first event of acquire.
  68 * As the nature of lock events, there is no guarantee
  69 * that the first event for the locks are acquire,
  70 * it can be acquired, contended or release.
  71 */
  72#define SEQ_STATE_UNINITIALIZED      0	       /* initial state */
  73#define SEQ_STATE_RELEASED	1
  74#define SEQ_STATE_ACQUIRING	2
  75#define SEQ_STATE_ACQUIRED	3
  76#define SEQ_STATE_READ_ACQUIRED	4
  77#define SEQ_STATE_CONTENDED	5
  78
  79/*
  80 * MAX_LOCK_DEPTH
  81 * Imported from include/linux/sched.h.
  82 * Should this be synchronized?
  83 */
  84#define MAX_LOCK_DEPTH 48
  85
  86/*
  87 * struct lock_seq_stat:
  88 * Place to put on state of one lock sequence
  89 * 1) acquire -> acquired -> release
  90 * 2) acquire -> contended -> acquired -> release
  91 * 3) acquire (with read or try) -> release
  92 * 4) Are there other patterns?
  93 */
  94struct lock_seq_stat {
  95	struct list_head        list;
  96	int			state;
  97	u64			prev_event_time;
  98	void                    *addr;
  99
 100	int                     read_count;
 101};
 102
 103struct thread_stat {
 104	struct rb_node		rb;
 105
 106	u32                     tid;
 107	struct list_head        seq_list;
 108};
 109
 110static struct rb_root		thread_stats;
 111
 112static struct thread_stat *thread_stat_find(u32 tid)
 113{
 114	struct rb_node *node;
 115	struct thread_stat *st;
 116
 117	node = thread_stats.rb_node;
 118	while (node) {
 119		st = container_of(node, struct thread_stat, rb);
 120		if (st->tid == tid)
 121			return st;
 122		else if (tid < st->tid)
 123			node = node->rb_left;
 124		else
 125			node = node->rb_right;
 126	}
 127
 128	return NULL;
 129}
 130
 131static void thread_stat_insert(struct thread_stat *new)
 132{
 133	struct rb_node **rb = &thread_stats.rb_node;
 134	struct rb_node *parent = NULL;
 135	struct thread_stat *p;
 136
 137	while (*rb) {
 138		p = container_of(*rb, struct thread_stat, rb);
 139		parent = *rb;
 140
 141		if (new->tid < p->tid)
 142			rb = &(*rb)->rb_left;
 143		else if (new->tid > p->tid)
 144			rb = &(*rb)->rb_right;
 145		else
 146			BUG_ON("inserting invalid thread_stat\n");
 147	}
 148
 149	rb_link_node(&new->rb, parent, rb);
 150	rb_insert_color(&new->rb, &thread_stats);
 151}
 152
 153static struct thread_stat *thread_stat_findnew_after_first(u32 tid)
 154{
 155	struct thread_stat *st;
 156
 157	st = thread_stat_find(tid);
 158	if (st)
 159		return st;
 160
 161	st = zalloc(sizeof(struct thread_stat));
 162	if (!st)
 163		die("memory allocation failed\n");
 
 
 164
 165	st->tid = tid;
 166	INIT_LIST_HEAD(&st->seq_list);
 167
 168	thread_stat_insert(st);
 169
 170	return st;
 171}
 172
 173static struct thread_stat *thread_stat_findnew_first(u32 tid);
 174static struct thread_stat *(*thread_stat_findnew)(u32 tid) =
 175	thread_stat_findnew_first;
 176
 177static struct thread_stat *thread_stat_findnew_first(u32 tid)
 178{
 179	struct thread_stat *st;
 180
 181	st = zalloc(sizeof(struct thread_stat));
 182	if (!st)
 183		die("memory allocation failed\n");
 
 
 184	st->tid = tid;
 185	INIT_LIST_HEAD(&st->seq_list);
 186
 187	rb_link_node(&st->rb, NULL, &thread_stats.rb_node);
 188	rb_insert_color(&st->rb, &thread_stats);
 189
 190	thread_stat_findnew = thread_stat_findnew_after_first;
 191	return st;
 192}
 193
 194/* build simple key function one is bigger than two */
 195#define SINGLE_KEY(member)						\
 196	static int lock_stat_key_ ## member(struct lock_stat *one,	\
 197					 struct lock_stat *two)		\
 198	{								\
 199		return one->member > two->member;			\
 200	}
 201
 202SINGLE_KEY(nr_acquired)
 203SINGLE_KEY(nr_contended)
 
 204SINGLE_KEY(wait_time_total)
 205SINGLE_KEY(wait_time_max)
 206
 207static int lock_stat_key_wait_time_min(struct lock_stat *one,
 208					struct lock_stat *two)
 209{
 210	u64 s1 = one->wait_time_min;
 211	u64 s2 = two->wait_time_min;
 212	if (s1 == ULLONG_MAX)
 213		s1 = 0;
 214	if (s2 == ULLONG_MAX)
 215		s2 = 0;
 216	return s1 > s2;
 217}
 218
 219struct lock_key {
 220	/*
 221	 * name: the value for specify by user
 222	 * this should be simpler than raw name of member
 223	 * e.g. nr_acquired -> acquired, wait_time_total -> wait_total
 224	 */
 225	const char		*name;
 226	int			(*key)(struct lock_stat*, struct lock_stat*);
 227};
 228
 229static const char		*sort_key = "acquired";
 230
 231static int			(*compare)(struct lock_stat *, struct lock_stat *);
 232
 233static struct rb_root		result;	/* place to store sorted data */
 234
 235#define DEF_KEY_LOCK(name, fn_suffix)	\
 236	{ #name, lock_stat_key_ ## fn_suffix }
 237struct lock_key keys[] = {
 238	DEF_KEY_LOCK(acquired, nr_acquired),
 239	DEF_KEY_LOCK(contended, nr_contended),
 
 240	DEF_KEY_LOCK(wait_total, wait_time_total),
 241	DEF_KEY_LOCK(wait_min, wait_time_min),
 242	DEF_KEY_LOCK(wait_max, wait_time_max),
 243
 244	/* extra comparisons much complicated should be here */
 245
 246	{ NULL, NULL }
 247};
 248
 249static void select_key(void)
 250{
 251	int i;
 252
 253	for (i = 0; keys[i].name; i++) {
 254		if (!strcmp(keys[i].name, sort_key)) {
 255			compare = keys[i].key;
 256			return;
 257		}
 258	}
 259
 260	die("Unknown compare key:%s\n", sort_key);
 
 
 261}
 262
 263static void insert_to_result(struct lock_stat *st,
 264			     int (*bigger)(struct lock_stat *, struct lock_stat *))
 265{
 266	struct rb_node **rb = &result.rb_node;
 267	struct rb_node *parent = NULL;
 268	struct lock_stat *p;
 269
 270	while (*rb) {
 271		p = container_of(*rb, struct lock_stat, rb);
 272		parent = *rb;
 273
 274		if (bigger(st, p))
 275			rb = &(*rb)->rb_left;
 276		else
 277			rb = &(*rb)->rb_right;
 278	}
 279
 280	rb_link_node(&st->rb, parent, rb);
 281	rb_insert_color(&st->rb, &result);
 282}
 283
 284/* returns left most element of result, and erase it */
 285static struct lock_stat *pop_from_result(void)
 286{
 287	struct rb_node *node = result.rb_node;
 288
 289	if (!node)
 290		return NULL;
 291
 292	while (node->rb_left)
 293		node = node->rb_left;
 294
 295	rb_erase(node, &result);
 296	return container_of(node, struct lock_stat, rb);
 297}
 298
 299static struct lock_stat *lock_stat_findnew(void *addr, const char *name)
 300{
 301	struct list_head *entry = lockhashentry(addr);
 302	struct lock_stat *ret, *new;
 303
 304	list_for_each_entry(ret, entry, hash_entry) {
 305		if (ret->addr == addr)
 306			return ret;
 307	}
 308
 309	new = zalloc(sizeof(struct lock_stat));
 310	if (!new)
 311		goto alloc_failed;
 312
 313	new->addr = addr;
 314	new->name = zalloc(sizeof(char) * strlen(name) + 1);
 315	if (!new->name)
 
 316		goto alloc_failed;
 317	strcpy(new->name, name);
 318
 
 319	new->wait_time_min = ULLONG_MAX;
 320
 321	list_add(&new->hash_entry, entry);
 322	return new;
 323
 324alloc_failed:
 325	die("memory allocation failed\n");
 
 326}
 327
 328static char			const *input_name = "perf.data";
 
 
 329
 330struct raw_event_sample {
 331	u32			size;
 332	char			data[0];
 333};
 334
 335struct trace_acquire_event {
 336	void			*addr;
 337	const char		*name;
 338	int			flag;
 339};
 340
 341struct trace_acquired_event {
 342	void			*addr;
 343	const char		*name;
 344};
 345
 346struct trace_contended_event {
 347	void			*addr;
 348	const char		*name;
 349};
 350
 351struct trace_release_event {
 352	void			*addr;
 353	const char		*name;
 354};
 355
 356struct trace_lock_handler {
 357	void (*acquire_event)(struct trace_acquire_event *,
 358			      struct event *,
 359			      int cpu,
 360			      u64 timestamp,
 361			      struct thread *thread);
 362
 363	void (*acquired_event)(struct trace_acquired_event *,
 364			       struct event *,
 365			       int cpu,
 366			       u64 timestamp,
 367			       struct thread *thread);
 368
 369	void (*contended_event)(struct trace_contended_event *,
 370				struct event *,
 371				int cpu,
 372				u64 timestamp,
 373				struct thread *thread);
 374
 375	void (*release_event)(struct trace_release_event *,
 376			      struct event *,
 377			      int cpu,
 378			      u64 timestamp,
 379			      struct thread *thread);
 380};
 381
 382static struct lock_seq_stat *get_seq(struct thread_stat *ts, void *addr)
 383{
 384	struct lock_seq_stat *seq;
 385
 386	list_for_each_entry(seq, &ts->seq_list, list) {
 387		if (seq->addr == addr)
 388			return seq;
 389	}
 390
 391	seq = zalloc(sizeof(struct lock_seq_stat));
 392	if (!seq)
 393		die("Not enough memory\n");
 
 
 394	seq->state = SEQ_STATE_UNINITIALIZED;
 395	seq->addr = addr;
 396
 397	list_add(&seq->list, &ts->seq_list);
 398	return seq;
 399}
 400
 401enum broken_state {
 402	BROKEN_ACQUIRE,
 403	BROKEN_ACQUIRED,
 404	BROKEN_CONTENDED,
 405	BROKEN_RELEASE,
 406	BROKEN_MAX,
 407};
 408
 409static int bad_hist[BROKEN_MAX];
 410
 411enum acquire_flags {
 412	TRY_LOCK = 1,
 413	READ_LOCK = 2,
 414};
 415
 416static void
 417report_lock_acquire_event(struct trace_acquire_event *acquire_event,
 418			struct event *__event __used,
 419			int cpu __used,
 420			u64 timestamp __used,
 421			struct thread *thread __used)
 422{
 
 423	struct lock_stat *ls;
 424	struct thread_stat *ts;
 425	struct lock_seq_stat *seq;
 
 
 
 
 
 426
 427	ls = lock_stat_findnew(acquire_event->addr, acquire_event->name);
 
 
 428	if (ls->discard)
 429		return;
 
 
 
 
 430
 431	ts = thread_stat_findnew(thread->pid);
 432	seq = get_seq(ts, acquire_event->addr);
 
 433
 434	switch (seq->state) {
 435	case SEQ_STATE_UNINITIALIZED:
 436	case SEQ_STATE_RELEASED:
 437		if (!acquire_event->flag) {
 438			seq->state = SEQ_STATE_ACQUIRING;
 439		} else {
 440			if (acquire_event->flag & TRY_LOCK)
 441				ls->nr_trylock++;
 442			if (acquire_event->flag & READ_LOCK)
 443				ls->nr_readlock++;
 444			seq->state = SEQ_STATE_READ_ACQUIRED;
 445			seq->read_count = 1;
 446			ls->nr_acquired++;
 447		}
 448		break;
 449	case SEQ_STATE_READ_ACQUIRED:
 450		if (acquire_event->flag & READ_LOCK) {
 451			seq->read_count++;
 452			ls->nr_acquired++;
 453			goto end;
 454		} else {
 455			goto broken;
 456		}
 457		break;
 458	case SEQ_STATE_ACQUIRED:
 459	case SEQ_STATE_ACQUIRING:
 460	case SEQ_STATE_CONTENDED:
 461broken:
 462		/* broken lock sequence, discard it */
 463		ls->discard = 1;
 464		bad_hist[BROKEN_ACQUIRE]++;
 465		list_del(&seq->list);
 466		free(seq);
 467		goto end;
 468		break;
 469	default:
 470		BUG_ON("Unknown state of lock sequence found!\n");
 471		break;
 472	}
 473
 474	ls->nr_acquire++;
 475	seq->prev_event_time = timestamp;
 476end:
 477	return;
 478}
 479
 480static void
 481report_lock_acquired_event(struct trace_acquired_event *acquired_event,
 482			 struct event *__event __used,
 483			 int cpu __used,
 484			 u64 timestamp __used,
 485			 struct thread *thread __used)
 486{
 
 487	struct lock_stat *ls;
 488	struct thread_stat *ts;
 489	struct lock_seq_stat *seq;
 490	u64 contended_term;
 
 
 491
 492	ls = lock_stat_findnew(acquired_event->addr, acquired_event->name);
 
 
 
 
 493	if (ls->discard)
 494		return;
 495
 496	ts = thread_stat_findnew(thread->pid);
 497	seq = get_seq(ts, acquired_event->addr);
 
 
 
 
 
 498
 499	switch (seq->state) {
 500	case SEQ_STATE_UNINITIALIZED:
 501		/* orphan event, do nothing */
 502		return;
 503	case SEQ_STATE_ACQUIRING:
 504		break;
 505	case SEQ_STATE_CONTENDED:
 506		contended_term = timestamp - seq->prev_event_time;
 507		ls->wait_time_total += contended_term;
 508		if (contended_term < ls->wait_time_min)
 509			ls->wait_time_min = contended_term;
 510		if (ls->wait_time_max < contended_term)
 511			ls->wait_time_max = contended_term;
 512		break;
 513	case SEQ_STATE_RELEASED:
 514	case SEQ_STATE_ACQUIRED:
 515	case SEQ_STATE_READ_ACQUIRED:
 516		/* broken lock sequence, discard it */
 517		ls->discard = 1;
 518		bad_hist[BROKEN_ACQUIRED]++;
 519		list_del(&seq->list);
 520		free(seq);
 521		goto end;
 522		break;
 523
 524	default:
 525		BUG_ON("Unknown state of lock sequence found!\n");
 526		break;
 527	}
 528
 529	seq->state = SEQ_STATE_ACQUIRED;
 530	ls->nr_acquired++;
 531	seq->prev_event_time = timestamp;
 
 532end:
 533	return;
 534}
 535
 536static void
 537report_lock_contended_event(struct trace_contended_event *contended_event,
 538			  struct event *__event __used,
 539			  int cpu __used,
 540			  u64 timestamp __used,
 541			  struct thread *thread __used)
 542{
 
 543	struct lock_stat *ls;
 544	struct thread_stat *ts;
 545	struct lock_seq_stat *seq;
 
 
 546
 547	ls = lock_stat_findnew(contended_event->addr, contended_event->name);
 
 
 
 
 548	if (ls->discard)
 549		return;
 550
 551	ts = thread_stat_findnew(thread->pid);
 552	seq = get_seq(ts, contended_event->addr);
 
 
 
 
 
 553
 554	switch (seq->state) {
 555	case SEQ_STATE_UNINITIALIZED:
 556		/* orphan event, do nothing */
 557		return;
 558	case SEQ_STATE_ACQUIRING:
 559		break;
 560	case SEQ_STATE_RELEASED:
 561	case SEQ_STATE_ACQUIRED:
 562	case SEQ_STATE_READ_ACQUIRED:
 563	case SEQ_STATE_CONTENDED:
 564		/* broken lock sequence, discard it */
 565		ls->discard = 1;
 566		bad_hist[BROKEN_CONTENDED]++;
 567		list_del(&seq->list);
 568		free(seq);
 569		goto end;
 570		break;
 571	default:
 572		BUG_ON("Unknown state of lock sequence found!\n");
 573		break;
 574	}
 575
 576	seq->state = SEQ_STATE_CONTENDED;
 577	ls->nr_contended++;
 578	seq->prev_event_time = timestamp;
 
 579end:
 580	return;
 581}
 582
 583static void
 584report_lock_release_event(struct trace_release_event *release_event,
 585			struct event *__event __used,
 586			int cpu __used,
 587			u64 timestamp __used,
 588			struct thread *thread __used)
 589{
 
 590	struct lock_stat *ls;
 591	struct thread_stat *ts;
 592	struct lock_seq_stat *seq;
 
 
 
 
 593
 594	ls = lock_stat_findnew(release_event->addr, release_event->name);
 
 
 595	if (ls->discard)
 596		return;
 
 
 
 
 597
 598	ts = thread_stat_findnew(thread->pid);
 599	seq = get_seq(ts, release_event->addr);
 
 600
 601	switch (seq->state) {
 602	case SEQ_STATE_UNINITIALIZED:
 603		goto end;
 604		break;
 605	case SEQ_STATE_ACQUIRED:
 606		break;
 607	case SEQ_STATE_READ_ACQUIRED:
 608		seq->read_count--;
 609		BUG_ON(seq->read_count < 0);
 610		if (!seq->read_count) {
 611			ls->nr_release++;
 612			goto end;
 613		}
 614		break;
 615	case SEQ_STATE_ACQUIRING:
 616	case SEQ_STATE_CONTENDED:
 617	case SEQ_STATE_RELEASED:
 618		/* broken lock sequence, discard it */
 619		ls->discard = 1;
 620		bad_hist[BROKEN_RELEASE]++;
 621		goto free_seq;
 622		break;
 623	default:
 624		BUG_ON("Unknown state of lock sequence found!\n");
 625		break;
 626	}
 627
 628	ls->nr_release++;
 629free_seq:
 630	list_del(&seq->list);
 631	free(seq);
 632end:
 633	return;
 634}
 635
 636/* lock oriented handlers */
 637/* TODO: handlers for CPU oriented, thread oriented */
 638static struct trace_lock_handler report_lock_ops  = {
 639	.acquire_event		= report_lock_acquire_event,
 640	.acquired_event		= report_lock_acquired_event,
 641	.contended_event	= report_lock_contended_event,
 642	.release_event		= report_lock_release_event,
 643};
 644
 645static struct trace_lock_handler *trace_handler;
 646
 647static void
 648process_lock_acquire_event(void *data,
 649			   struct event *event __used,
 650			   int cpu __used,
 651			   u64 timestamp __used,
 652			   struct thread *thread __used)
 653{
 654	struct trace_acquire_event acquire_event;
 655	u64 tmp;		/* this is required for casting... */
 656
 657	tmp = raw_field_value(event, "lockdep_addr", data);
 658	memcpy(&acquire_event.addr, &tmp, sizeof(void *));
 659	acquire_event.name = (char *)raw_field_ptr(event, "name", data);
 660	acquire_event.flag = (int)raw_field_value(event, "flag", data);
 661
 662	if (trace_handler->acquire_event)
 663		trace_handler->acquire_event(&acquire_event, event, cpu, timestamp, thread);
 664}
 665
 666static void
 667process_lock_acquired_event(void *data,
 668			    struct event *event __used,
 669			    int cpu __used,
 670			    u64 timestamp __used,
 671			    struct thread *thread __used)
 672{
 673	struct trace_acquired_event acquired_event;
 674	u64 tmp;		/* this is required for casting... */
 675
 676	tmp = raw_field_value(event, "lockdep_addr", data);
 677	memcpy(&acquired_event.addr, &tmp, sizeof(void *));
 678	acquired_event.name = (char *)raw_field_ptr(event, "name", data);
 679
 680	if (trace_handler->acquire_event)
 681		trace_handler->acquired_event(&acquired_event, event, cpu, timestamp, thread);
 
 682}
 683
 684static void
 685process_lock_contended_event(void *data,
 686			     struct event *event __used,
 687			     int cpu __used,
 688			     u64 timestamp __used,
 689			     struct thread *thread __used)
 690{
 691	struct trace_contended_event contended_event;
 692	u64 tmp;		/* this is required for casting... */
 693
 694	tmp = raw_field_value(event, "lockdep_addr", data);
 695	memcpy(&contended_event.addr, &tmp, sizeof(void *));
 696	contended_event.name = (char *)raw_field_ptr(event, "name", data);
 697
 698	if (trace_handler->acquire_event)
 699		trace_handler->contended_event(&contended_event, event, cpu, timestamp, thread);
 700}
 701
 702static void
 703process_lock_release_event(void *data,
 704			   struct event *event __used,
 705			   int cpu __used,
 706			   u64 timestamp __used,
 707			   struct thread *thread __used)
 708{
 709	struct trace_release_event release_event;
 710	u64 tmp;		/* this is required for casting... */
 711
 712	tmp = raw_field_value(event, "lockdep_addr", data);
 713	memcpy(&release_event.addr, &tmp, sizeof(void *));
 714	release_event.name = (char *)raw_field_ptr(event, "name", data);
 715
 716	if (trace_handler->acquire_event)
 717		trace_handler->release_event(&release_event, event, cpu, timestamp, thread);
 718}
 719
 720static void
 721process_raw_event(void *data, int cpu, u64 timestamp, struct thread *thread)
 722{
 723	struct event *event;
 724	int type;
 725
 726	type = trace_parse_common_type(data);
 727	event = trace_find_event(type);
 728
 729	if (!strcmp(event->name, "lock_acquire"))
 730		process_lock_acquire_event(data, event, cpu, timestamp, thread);
 731	if (!strcmp(event->name, "lock_acquired"))
 732		process_lock_acquired_event(data, event, cpu, timestamp, thread);
 733	if (!strcmp(event->name, "lock_contended"))
 734		process_lock_contended_event(data, event, cpu, timestamp, thread);
 735	if (!strcmp(event->name, "lock_release"))
 736		process_lock_release_event(data, event, cpu, timestamp, thread);
 737}
 738
 739static void print_bad_events(int bad, int total)
 740{
 741	/* Output for debug, this have to be removed */
 742	int i;
 743	const char *name[4] =
 744		{ "acquire", "acquired", "contended", "release" };
 745
 746	pr_info("\n=== output for debug===\n\n");
 747	pr_info("bad: %d, total: %d\n", bad, total);
 748	pr_info("bad rate: %f %%\n", (double)bad / (double)total * 100);
 749	pr_info("histogram of events caused bad sequence\n");
 750	for (i = 0; i < BROKEN_MAX; i++)
 751		pr_info(" %10s: %d\n", name[i], bad_hist[i]);
 752}
 753
 754/* TODO: various way to print, coloring, nano or milli sec */
 755static void print_result(void)
 756{
 757	struct lock_stat *st;
 758	char cut_name[20];
 759	int bad, total;
 760
 761	pr_info("%20s ", "Name");
 762	pr_info("%10s ", "acquired");
 763	pr_info("%10s ", "contended");
 764
 
 765	pr_info("%15s ", "total wait (ns)");
 766	pr_info("%15s ", "max wait (ns)");
 767	pr_info("%15s ", "min wait (ns)");
 768
 769	pr_info("\n\n");
 770
 771	bad = total = 0;
 772	while ((st = pop_from_result())) {
 773		total++;
 774		if (st->discard) {
 775			bad++;
 776			continue;
 777		}
 778		bzero(cut_name, 20);
 779
 780		if (strlen(st->name) < 16) {
 781			/* output raw name */
 782			pr_info("%20s ", st->name);
 783		} else {
 784			strncpy(cut_name, st->name, 16);
 785			cut_name[16] = '.';
 786			cut_name[17] = '.';
 787			cut_name[18] = '.';
 788			cut_name[19] = '\0';
 789			/* cut off name for saving output style */
 790			pr_info("%20s ", cut_name);
 791		}
 792
 793		pr_info("%10u ", st->nr_acquired);
 794		pr_info("%10u ", st->nr_contended);
 795
 
 796		pr_info("%15" PRIu64 " ", st->wait_time_total);
 797		pr_info("%15" PRIu64 " ", st->wait_time_max);
 798		pr_info("%15" PRIu64 " ", st->wait_time_min == ULLONG_MAX ?
 799		       0 : st->wait_time_min);
 800		pr_info("\n");
 801	}
 802
 803	print_bad_events(bad, total);
 804}
 805
 806static bool info_threads, info_map;
 807
 808static void dump_threads(void)
 809{
 810	struct thread_stat *st;
 811	struct rb_node *node;
 812	struct thread *t;
 813
 814	pr_info("%10s: comm\n", "Thread ID");
 815
 816	node = rb_first(&thread_stats);
 817	while (node) {
 818		st = container_of(node, struct thread_stat, rb);
 819		t = perf_session__findnew(session, st->tid);
 820		pr_info("%10d: %s\n", st->tid, t->comm);
 821		node = rb_next(node);
 
 822	};
 823}
 824
 825static void dump_map(void)
 826{
 827	unsigned int i;
 828	struct lock_stat *st;
 829
 830	pr_info("Address of instance: name of class\n");
 831	for (i = 0; i < LOCKHASH_SIZE; i++) {
 832		list_for_each_entry(st, &lockhash_table[i], hash_entry) {
 833			pr_info(" %p: %s\n", st->addr, st->name);
 834		}
 835	}
 836}
 837
 838static void dump_info(void)
 839{
 
 
 840	if (info_threads)
 841		dump_threads();
 842	else if (info_map)
 843		dump_map();
 844	else
 845		die("Unknown type of information\n");
 
 
 
 
 846}
 847
 848static int process_sample_event(union perf_event *event,
 
 
 
 
 849				struct perf_sample *sample,
 850				struct perf_evsel *evsel __used,
 851				struct perf_session *s)
 852{
 853	struct thread *thread = perf_session__findnew(s, sample->tid);
 
 
 854
 855	if (thread == NULL) {
 856		pr_debug("problem processing %d event, skipping it.\n",
 857			event->header.type);
 858		return -1;
 859	}
 860
 861	process_raw_event(sample->raw_data, sample->cpu, sample->time, thread);
 862
 863	return 0;
 864}
 865
 866static struct perf_event_ops eops = {
 867	.sample			= process_sample_event,
 868	.comm			= perf_event__process_comm,
 869	.ordered_samples	= true,
 870};
 871
 872static int read_events(void)
 873{
 874	session = perf_session__new(input_name, O_RDONLY, 0, false, &eops);
 875	if (!session)
 876		die("Initializing perf session failed\n");
 877
 878	return perf_session__process_events(session, &eops);
 879}
 880
 881static void sort_result(void)
 882{
 883	unsigned int i;
 884	struct lock_stat *st;
 885
 886	for (i = 0; i < LOCKHASH_SIZE; i++) {
 887		list_for_each_entry(st, &lockhash_table[i], hash_entry) {
 888			insert_to_result(st, compare);
 889		}
 890	}
 891}
 892
 893static void __cmd_report(void)
 894{
 895	setup_pager();
 896	select_key();
 897	read_events();
 898	sort_result();
 899	print_result();
 900}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 901
 902static const char * const report_usage[] = {
 903	"perf lock report [<options>]",
 904	NULL
 905};
 
 906
 907static const struct option report_options[] = {
 908	OPT_STRING('k', "key", &sort_key, "acquired",
 909		    "key for sorting (acquired / contended / wait_total / wait_max / wait_min)"),
 910	/* TODO: type */
 911	OPT_END()
 912};
 913
 914static const char * const info_usage[] = {
 915	"perf lock info [<options>]",
 916	NULL
 917};
 918
 919static const struct option info_options[] = {
 920	OPT_BOOLEAN('t', "threads", &info_threads,
 921		    "dump thread list in perf.data"),
 922	OPT_BOOLEAN('m', "map", &info_map,
 923		    "map of lock instances (name:address table)"),
 924	OPT_END()
 925};
 926
 927static const char * const lock_usage[] = {
 928	"perf lock [<options>] {record|trace|report}",
 929	NULL
 930};
 931
 932static const struct option lock_options[] = {
 933	OPT_STRING('i', "input", &input_name, "file", "input file name"),
 934	OPT_INCR('v', "verbose", &verbose, "be more verbose (show symbol address, etc)"),
 935	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, "dump raw trace in ASCII"),
 936	OPT_END()
 937};
 938
 939static const char *record_args[] = {
 940	"record",
 941	"-R",
 942	"-f",
 943	"-m", "1024",
 944	"-c", "1",
 945	"-e", "lock:lock_acquire",
 946	"-e", "lock:lock_acquired",
 947	"-e", "lock:lock_contended",
 948	"-e", "lock:lock_release",
 949};
 
 950
 951static int __cmd_record(int argc, const char **argv)
 952{
 953	unsigned int rec_argc, i, j;
 
 
 
 954	const char **rec_argv;
 955
 
 
 
 
 
 
 
 
 
 956	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
 957	rec_argv = calloc(rec_argc + 1, sizeof(char *));
 
 958
 959	if (rec_argv == NULL)
 
 960		return -ENOMEM;
 961
 962	for (i = 0; i < ARRAY_SIZE(record_args); i++)
 963		rec_argv[i] = strdup(record_args[i]);
 964
 
 
 
 
 
 965	for (j = 1; j < (unsigned int)argc; j++, i++)
 966		rec_argv[i] = argv[j];
 967
 968	BUG_ON(i != rec_argc);
 969
 970	return cmd_record(i, rec_argv, NULL);
 
 
 971}
 972
 973int cmd_lock(int argc, const char **argv, const char *prefix __used)
 974{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 975	unsigned int i;
 
 976
 977	symbol__init();
 978	for (i = 0; i < LOCKHASH_SIZE; i++)
 979		INIT_LIST_HEAD(lockhash_table + i);
 980
 981	argc = parse_options(argc, argv, lock_options, lock_usage,
 982			     PARSE_OPT_STOP_AT_NON_OPTION);
 983	if (!argc)
 984		usage_with_options(lock_usage, lock_options);
 985
 986	if (!strncmp(argv[0], "rec", 3)) {
 987		return __cmd_record(argc, argv);
 988	} else if (!strncmp(argv[0], "report", 6)) {
 989		trace_handler = &report_lock_ops;
 990		if (argc) {
 991			argc = parse_options(argc, argv,
 992					     report_options, report_usage, 0);
 993			if (argc)
 994				usage_with_options(report_usage, report_options);
 995		}
 996		__cmd_report();
 997	} else if (!strcmp(argv[0], "script")) {
 998		/* Aliased to 'perf script' */
 999		return cmd_script(argc, argv, prefix);
1000	} else if (!strcmp(argv[0], "info")) {
1001		if (argc) {
1002			argc = parse_options(argc, argv,
1003					     info_options, info_usage, 0);
1004			if (argc)
1005				usage_with_options(info_usage, info_options);
1006		}
1007		/* recycling report_lock_ops */
1008		trace_handler = &report_lock_ops;
1009		setup_pager();
1010		read_events();
1011		dump_info();
1012	} else {
1013		usage_with_options(lock_usage, lock_options);
1014	}
1015
1016	return 0;
1017}
v4.10.11
   1#include "builtin.h"
   2#include "perf.h"
   3
   4#include "util/evlist.h"
   5#include "util/evsel.h"
   6#include "util/util.h"
   7#include "util/cache.h"
   8#include "util/symbol.h"
   9#include "util/thread.h"
  10#include "util/header.h"
  11
  12#include <subcmd/parse-options.h>
  13#include "util/trace-event.h"
  14
  15#include "util/debug.h"
  16#include "util/session.h"
  17#include "util/tool.h"
  18#include "util/data.h"
  19
  20#include <sys/types.h>
  21#include <sys/prctl.h>
  22#include <semaphore.h>
  23#include <pthread.h>
  24#include <math.h>
  25#include <limits.h>
  26
  27#include <linux/list.h>
  28#include <linux/hash.h>
  29
  30static struct perf_session *session;
  31
  32/* based on kernel/lockdep.c */
  33#define LOCKHASH_BITS		12
  34#define LOCKHASH_SIZE		(1UL << LOCKHASH_BITS)
  35
  36static struct list_head lockhash_table[LOCKHASH_SIZE];
  37
  38#define __lockhashfn(key)	hash_long((unsigned long)key, LOCKHASH_BITS)
  39#define lockhashentry(key)	(lockhash_table + __lockhashfn((key)))
  40
  41struct lock_stat {
  42	struct list_head	hash_entry;
  43	struct rb_node		rb;		/* used for sorting */
  44
  45	/*
  46	 * FIXME: perf_evsel__intval() returns u64,
  47	 * so address of lockdep_map should be dealed as 64bit.
  48	 * Is there more better solution?
  49	 */
  50	void			*addr;		/* address of lockdep_map, used as ID */
  51	char			*name;		/* for strcpy(), we cannot use const */
  52
  53	unsigned int		nr_acquire;
  54	unsigned int		nr_acquired;
  55	unsigned int		nr_contended;
  56	unsigned int		nr_release;
  57
  58	unsigned int		nr_readlock;
  59	unsigned int		nr_trylock;
  60
  61	/* these times are in nano sec. */
  62	u64                     avg_wait_time;
  63	u64			wait_time_total;
  64	u64			wait_time_min;
  65	u64			wait_time_max;
  66
  67	int			discard; /* flag of blacklist */
  68};
  69
  70/*
  71 * States of lock_seq_stat
  72 *
  73 * UNINITIALIZED is required for detecting first event of acquire.
  74 * As the nature of lock events, there is no guarantee
  75 * that the first event for the locks are acquire,
  76 * it can be acquired, contended or release.
  77 */
  78#define SEQ_STATE_UNINITIALIZED      0	       /* initial state */
  79#define SEQ_STATE_RELEASED	1
  80#define SEQ_STATE_ACQUIRING	2
  81#define SEQ_STATE_ACQUIRED	3
  82#define SEQ_STATE_READ_ACQUIRED	4
  83#define SEQ_STATE_CONTENDED	5
  84
  85/*
  86 * MAX_LOCK_DEPTH
  87 * Imported from include/linux/sched.h.
  88 * Should this be synchronized?
  89 */
  90#define MAX_LOCK_DEPTH 48
  91
  92/*
  93 * struct lock_seq_stat:
  94 * Place to put on state of one lock sequence
  95 * 1) acquire -> acquired -> release
  96 * 2) acquire -> contended -> acquired -> release
  97 * 3) acquire (with read or try) -> release
  98 * 4) Are there other patterns?
  99 */
 100struct lock_seq_stat {
 101	struct list_head        list;
 102	int			state;
 103	u64			prev_event_time;
 104	void                    *addr;
 105
 106	int                     read_count;
 107};
 108
 109struct thread_stat {
 110	struct rb_node		rb;
 111
 112	u32                     tid;
 113	struct list_head        seq_list;
 114};
 115
 116static struct rb_root		thread_stats;
 117
 118static struct thread_stat *thread_stat_find(u32 tid)
 119{
 120	struct rb_node *node;
 121	struct thread_stat *st;
 122
 123	node = thread_stats.rb_node;
 124	while (node) {
 125		st = container_of(node, struct thread_stat, rb);
 126		if (st->tid == tid)
 127			return st;
 128		else if (tid < st->tid)
 129			node = node->rb_left;
 130		else
 131			node = node->rb_right;
 132	}
 133
 134	return NULL;
 135}
 136
 137static void thread_stat_insert(struct thread_stat *new)
 138{
 139	struct rb_node **rb = &thread_stats.rb_node;
 140	struct rb_node *parent = NULL;
 141	struct thread_stat *p;
 142
 143	while (*rb) {
 144		p = container_of(*rb, struct thread_stat, rb);
 145		parent = *rb;
 146
 147		if (new->tid < p->tid)
 148			rb = &(*rb)->rb_left;
 149		else if (new->tid > p->tid)
 150			rb = &(*rb)->rb_right;
 151		else
 152			BUG_ON("inserting invalid thread_stat\n");
 153	}
 154
 155	rb_link_node(&new->rb, parent, rb);
 156	rb_insert_color(&new->rb, &thread_stats);
 157}
 158
 159static struct thread_stat *thread_stat_findnew_after_first(u32 tid)
 160{
 161	struct thread_stat *st;
 162
 163	st = thread_stat_find(tid);
 164	if (st)
 165		return st;
 166
 167	st = zalloc(sizeof(struct thread_stat));
 168	if (!st) {
 169		pr_err("memory allocation failed\n");
 170		return NULL;
 171	}
 172
 173	st->tid = tid;
 174	INIT_LIST_HEAD(&st->seq_list);
 175
 176	thread_stat_insert(st);
 177
 178	return st;
 179}
 180
 181static struct thread_stat *thread_stat_findnew_first(u32 tid);
 182static struct thread_stat *(*thread_stat_findnew)(u32 tid) =
 183	thread_stat_findnew_first;
 184
 185static struct thread_stat *thread_stat_findnew_first(u32 tid)
 186{
 187	struct thread_stat *st;
 188
 189	st = zalloc(sizeof(struct thread_stat));
 190	if (!st) {
 191		pr_err("memory allocation failed\n");
 192		return NULL;
 193	}
 194	st->tid = tid;
 195	INIT_LIST_HEAD(&st->seq_list);
 196
 197	rb_link_node(&st->rb, NULL, &thread_stats.rb_node);
 198	rb_insert_color(&st->rb, &thread_stats);
 199
 200	thread_stat_findnew = thread_stat_findnew_after_first;
 201	return st;
 202}
 203
 204/* build simple key function one is bigger than two */
 205#define SINGLE_KEY(member)						\
 206	static int lock_stat_key_ ## member(struct lock_stat *one,	\
 207					 struct lock_stat *two)		\
 208	{								\
 209		return one->member > two->member;			\
 210	}
 211
 212SINGLE_KEY(nr_acquired)
 213SINGLE_KEY(nr_contended)
 214SINGLE_KEY(avg_wait_time)
 215SINGLE_KEY(wait_time_total)
 216SINGLE_KEY(wait_time_max)
 217
 218static int lock_stat_key_wait_time_min(struct lock_stat *one,
 219					struct lock_stat *two)
 220{
 221	u64 s1 = one->wait_time_min;
 222	u64 s2 = two->wait_time_min;
 223	if (s1 == ULLONG_MAX)
 224		s1 = 0;
 225	if (s2 == ULLONG_MAX)
 226		s2 = 0;
 227	return s1 > s2;
 228}
 229
 230struct lock_key {
 231	/*
 232	 * name: the value for specify by user
 233	 * this should be simpler than raw name of member
 234	 * e.g. nr_acquired -> acquired, wait_time_total -> wait_total
 235	 */
 236	const char		*name;
 237	int			(*key)(struct lock_stat*, struct lock_stat*);
 238};
 239
 240static const char		*sort_key = "acquired";
 241
 242static int			(*compare)(struct lock_stat *, struct lock_stat *);
 243
 244static struct rb_root		result;	/* place to store sorted data */
 245
 246#define DEF_KEY_LOCK(name, fn_suffix)	\
 247	{ #name, lock_stat_key_ ## fn_suffix }
 248struct lock_key keys[] = {
 249	DEF_KEY_LOCK(acquired, nr_acquired),
 250	DEF_KEY_LOCK(contended, nr_contended),
 251	DEF_KEY_LOCK(avg_wait, avg_wait_time),
 252	DEF_KEY_LOCK(wait_total, wait_time_total),
 253	DEF_KEY_LOCK(wait_min, wait_time_min),
 254	DEF_KEY_LOCK(wait_max, wait_time_max),
 255
 256	/* extra comparisons much complicated should be here */
 257
 258	{ NULL, NULL }
 259};
 260
 261static int select_key(void)
 262{
 263	int i;
 264
 265	for (i = 0; keys[i].name; i++) {
 266		if (!strcmp(keys[i].name, sort_key)) {
 267			compare = keys[i].key;
 268			return 0;
 269		}
 270	}
 271
 272	pr_err("Unknown compare key: %s\n", sort_key);
 273
 274	return -1;
 275}
 276
 277static void insert_to_result(struct lock_stat *st,
 278			     int (*bigger)(struct lock_stat *, struct lock_stat *))
 279{
 280	struct rb_node **rb = &result.rb_node;
 281	struct rb_node *parent = NULL;
 282	struct lock_stat *p;
 283
 284	while (*rb) {
 285		p = container_of(*rb, struct lock_stat, rb);
 286		parent = *rb;
 287
 288		if (bigger(st, p))
 289			rb = &(*rb)->rb_left;
 290		else
 291			rb = &(*rb)->rb_right;
 292	}
 293
 294	rb_link_node(&st->rb, parent, rb);
 295	rb_insert_color(&st->rb, &result);
 296}
 297
 298/* returns left most element of result, and erase it */
 299static struct lock_stat *pop_from_result(void)
 300{
 301	struct rb_node *node = result.rb_node;
 302
 303	if (!node)
 304		return NULL;
 305
 306	while (node->rb_left)
 307		node = node->rb_left;
 308
 309	rb_erase(node, &result);
 310	return container_of(node, struct lock_stat, rb);
 311}
 312
 313static struct lock_stat *lock_stat_findnew(void *addr, const char *name)
 314{
 315	struct list_head *entry = lockhashentry(addr);
 316	struct lock_stat *ret, *new;
 317
 318	list_for_each_entry(ret, entry, hash_entry) {
 319		if (ret->addr == addr)
 320			return ret;
 321	}
 322
 323	new = zalloc(sizeof(struct lock_stat));
 324	if (!new)
 325		goto alloc_failed;
 326
 327	new->addr = addr;
 328	new->name = zalloc(sizeof(char) * strlen(name) + 1);
 329	if (!new->name) {
 330		free(new);
 331		goto alloc_failed;
 332	}
 333
 334	strcpy(new->name, name);
 335	new->wait_time_min = ULLONG_MAX;
 336
 337	list_add(&new->hash_entry, entry);
 338	return new;
 339
 340alloc_failed:
 341	pr_err("memory allocation failed\n");
 342	return NULL;
 343}
 344
 345struct trace_lock_handler {
 346	int (*acquire_event)(struct perf_evsel *evsel,
 347			     struct perf_sample *sample);
 348
 349	int (*acquired_event)(struct perf_evsel *evsel,
 350			      struct perf_sample *sample);
 
 
 351
 352	int (*contended_event)(struct perf_evsel *evsel,
 353			       struct perf_sample *sample);
 
 
 
 354
 355	int (*release_event)(struct perf_evsel *evsel,
 356			     struct perf_sample *sample);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 357};
 358
 359static struct lock_seq_stat *get_seq(struct thread_stat *ts, void *addr)
 360{
 361	struct lock_seq_stat *seq;
 362
 363	list_for_each_entry(seq, &ts->seq_list, list) {
 364		if (seq->addr == addr)
 365			return seq;
 366	}
 367
 368	seq = zalloc(sizeof(struct lock_seq_stat));
 369	if (!seq) {
 370		pr_err("memory allocation failed\n");
 371		return NULL;
 372	}
 373	seq->state = SEQ_STATE_UNINITIALIZED;
 374	seq->addr = addr;
 375
 376	list_add(&seq->list, &ts->seq_list);
 377	return seq;
 378}
 379
 380enum broken_state {
 381	BROKEN_ACQUIRE,
 382	BROKEN_ACQUIRED,
 383	BROKEN_CONTENDED,
 384	BROKEN_RELEASE,
 385	BROKEN_MAX,
 386};
 387
 388static int bad_hist[BROKEN_MAX];
 389
 390enum acquire_flags {
 391	TRY_LOCK = 1,
 392	READ_LOCK = 2,
 393};
 394
 395static int report_lock_acquire_event(struct perf_evsel *evsel,
 396				     struct perf_sample *sample)
 
 
 
 
 397{
 398	void *addr;
 399	struct lock_stat *ls;
 400	struct thread_stat *ts;
 401	struct lock_seq_stat *seq;
 402	const char *name = perf_evsel__strval(evsel, sample, "name");
 403	u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
 404	int flag = perf_evsel__intval(evsel, sample, "flag");
 405
 406	memcpy(&addr, &tmp, sizeof(void *));
 407
 408	ls = lock_stat_findnew(addr, name);
 409	if (!ls)
 410		return -ENOMEM;
 411	if (ls->discard)
 412		return 0;
 413
 414	ts = thread_stat_findnew(sample->tid);
 415	if (!ts)
 416		return -ENOMEM;
 417
 418	seq = get_seq(ts, addr);
 419	if (!seq)
 420		return -ENOMEM;
 421
 422	switch (seq->state) {
 423	case SEQ_STATE_UNINITIALIZED:
 424	case SEQ_STATE_RELEASED:
 425		if (!flag) {
 426			seq->state = SEQ_STATE_ACQUIRING;
 427		} else {
 428			if (flag & TRY_LOCK)
 429				ls->nr_trylock++;
 430			if (flag & READ_LOCK)
 431				ls->nr_readlock++;
 432			seq->state = SEQ_STATE_READ_ACQUIRED;
 433			seq->read_count = 1;
 434			ls->nr_acquired++;
 435		}
 436		break;
 437	case SEQ_STATE_READ_ACQUIRED:
 438		if (flag & READ_LOCK) {
 439			seq->read_count++;
 440			ls->nr_acquired++;
 441			goto end;
 442		} else {
 443			goto broken;
 444		}
 445		break;
 446	case SEQ_STATE_ACQUIRED:
 447	case SEQ_STATE_ACQUIRING:
 448	case SEQ_STATE_CONTENDED:
 449broken:
 450		/* broken lock sequence, discard it */
 451		ls->discard = 1;
 452		bad_hist[BROKEN_ACQUIRE]++;
 453		list_del(&seq->list);
 454		free(seq);
 455		goto end;
 
 456	default:
 457		BUG_ON("Unknown state of lock sequence found!\n");
 458		break;
 459	}
 460
 461	ls->nr_acquire++;
 462	seq->prev_event_time = sample->time;
 463end:
 464	return 0;
 465}
 466
 467static int report_lock_acquired_event(struct perf_evsel *evsel,
 468				      struct perf_sample *sample)
 
 
 
 
 469{
 470	void *addr;
 471	struct lock_stat *ls;
 472	struct thread_stat *ts;
 473	struct lock_seq_stat *seq;
 474	u64 contended_term;
 475	const char *name = perf_evsel__strval(evsel, sample, "name");
 476	u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
 477
 478	memcpy(&addr, &tmp, sizeof(void *));
 479
 480	ls = lock_stat_findnew(addr, name);
 481	if (!ls)
 482		return -ENOMEM;
 483	if (ls->discard)
 484		return 0;
 485
 486	ts = thread_stat_findnew(sample->tid);
 487	if (!ts)
 488		return -ENOMEM;
 489
 490	seq = get_seq(ts, addr);
 491	if (!seq)
 492		return -ENOMEM;
 493
 494	switch (seq->state) {
 495	case SEQ_STATE_UNINITIALIZED:
 496		/* orphan event, do nothing */
 497		return 0;
 498	case SEQ_STATE_ACQUIRING:
 499		break;
 500	case SEQ_STATE_CONTENDED:
 501		contended_term = sample->time - seq->prev_event_time;
 502		ls->wait_time_total += contended_term;
 503		if (contended_term < ls->wait_time_min)
 504			ls->wait_time_min = contended_term;
 505		if (ls->wait_time_max < contended_term)
 506			ls->wait_time_max = contended_term;
 507		break;
 508	case SEQ_STATE_RELEASED:
 509	case SEQ_STATE_ACQUIRED:
 510	case SEQ_STATE_READ_ACQUIRED:
 511		/* broken lock sequence, discard it */
 512		ls->discard = 1;
 513		bad_hist[BROKEN_ACQUIRED]++;
 514		list_del(&seq->list);
 515		free(seq);
 516		goto end;
 
 
 517	default:
 518		BUG_ON("Unknown state of lock sequence found!\n");
 519		break;
 520	}
 521
 522	seq->state = SEQ_STATE_ACQUIRED;
 523	ls->nr_acquired++;
 524	ls->avg_wait_time = ls->nr_contended ? ls->wait_time_total/ls->nr_contended : 0;
 525	seq->prev_event_time = sample->time;
 526end:
 527	return 0;
 528}
 529
 530static int report_lock_contended_event(struct perf_evsel *evsel,
 531				       struct perf_sample *sample)
 
 
 
 
 532{
 533	void *addr;
 534	struct lock_stat *ls;
 535	struct thread_stat *ts;
 536	struct lock_seq_stat *seq;
 537	const char *name = perf_evsel__strval(evsel, sample, "name");
 538	u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
 539
 540	memcpy(&addr, &tmp, sizeof(void *));
 541
 542	ls = lock_stat_findnew(addr, name);
 543	if (!ls)
 544		return -ENOMEM;
 545	if (ls->discard)
 546		return 0;
 547
 548	ts = thread_stat_findnew(sample->tid);
 549	if (!ts)
 550		return -ENOMEM;
 551
 552	seq = get_seq(ts, addr);
 553	if (!seq)
 554		return -ENOMEM;
 555
 556	switch (seq->state) {
 557	case SEQ_STATE_UNINITIALIZED:
 558		/* orphan event, do nothing */
 559		return 0;
 560	case SEQ_STATE_ACQUIRING:
 561		break;
 562	case SEQ_STATE_RELEASED:
 563	case SEQ_STATE_ACQUIRED:
 564	case SEQ_STATE_READ_ACQUIRED:
 565	case SEQ_STATE_CONTENDED:
 566		/* broken lock sequence, discard it */
 567		ls->discard = 1;
 568		bad_hist[BROKEN_CONTENDED]++;
 569		list_del(&seq->list);
 570		free(seq);
 571		goto end;
 
 572	default:
 573		BUG_ON("Unknown state of lock sequence found!\n");
 574		break;
 575	}
 576
 577	seq->state = SEQ_STATE_CONTENDED;
 578	ls->nr_contended++;
 579	ls->avg_wait_time = ls->wait_time_total/ls->nr_contended;
 580	seq->prev_event_time = sample->time;
 581end:
 582	return 0;
 583}
 584
 585static int report_lock_release_event(struct perf_evsel *evsel,
 586				     struct perf_sample *sample)
 
 
 
 
 587{
 588	void *addr;
 589	struct lock_stat *ls;
 590	struct thread_stat *ts;
 591	struct lock_seq_stat *seq;
 592	const char *name = perf_evsel__strval(evsel, sample, "name");
 593	u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
 594
 595	memcpy(&addr, &tmp, sizeof(void *));
 596
 597	ls = lock_stat_findnew(addr, name);
 598	if (!ls)
 599		return -ENOMEM;
 600	if (ls->discard)
 601		return 0;
 602
 603	ts = thread_stat_findnew(sample->tid);
 604	if (!ts)
 605		return -ENOMEM;
 606
 607	seq = get_seq(ts, addr);
 608	if (!seq)
 609		return -ENOMEM;
 610
 611	switch (seq->state) {
 612	case SEQ_STATE_UNINITIALIZED:
 613		goto end;
 
 614	case SEQ_STATE_ACQUIRED:
 615		break;
 616	case SEQ_STATE_READ_ACQUIRED:
 617		seq->read_count--;
 618		BUG_ON(seq->read_count < 0);
 619		if (!seq->read_count) {
 620			ls->nr_release++;
 621			goto end;
 622		}
 623		break;
 624	case SEQ_STATE_ACQUIRING:
 625	case SEQ_STATE_CONTENDED:
 626	case SEQ_STATE_RELEASED:
 627		/* broken lock sequence, discard it */
 628		ls->discard = 1;
 629		bad_hist[BROKEN_RELEASE]++;
 630		goto free_seq;
 
 631	default:
 632		BUG_ON("Unknown state of lock sequence found!\n");
 633		break;
 634	}
 635
 636	ls->nr_release++;
 637free_seq:
 638	list_del(&seq->list);
 639	free(seq);
 640end:
 641	return 0;
 642}
 643
 644/* lock oriented handlers */
 645/* TODO: handlers for CPU oriented, thread oriented */
 646static struct trace_lock_handler report_lock_ops  = {
 647	.acquire_event		= report_lock_acquire_event,
 648	.acquired_event		= report_lock_acquired_event,
 649	.contended_event	= report_lock_contended_event,
 650	.release_event		= report_lock_release_event,
 651};
 652
 653static struct trace_lock_handler *trace_handler;
 654
 655static int perf_evsel__process_lock_acquire(struct perf_evsel *evsel,
 656					     struct perf_sample *sample)
 657{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 658	if (trace_handler->acquire_event)
 659		return trace_handler->acquire_event(evsel, sample);
 660	return 0;
 661}
 662
 663static int perf_evsel__process_lock_acquired(struct perf_evsel *evsel,
 664					      struct perf_sample *sample)
 665{
 666	if (trace_handler->acquired_event)
 667		return trace_handler->acquired_event(evsel, sample);
 668	return 0;
 
 
 
 
 
 
 
 
 
 
 669}
 670
 671static int perf_evsel__process_lock_contended(struct perf_evsel *evsel,
 672					      struct perf_sample *sample)
 673{
 674	if (trace_handler->contended_event)
 675		return trace_handler->contended_event(evsel, sample);
 676	return 0;
 
 
 
 
 
 
 
 
 
 
 677}
 678
 679static int perf_evsel__process_lock_release(struct perf_evsel *evsel,
 680					    struct perf_sample *sample)
 681{
 682	if (trace_handler->release_event)
 683		return trace_handler->release_event(evsel, sample);
 684	return 0;
 
 
 
 
 
 
 
 
 
 
 
 685}
 686
 687static void print_bad_events(int bad, int total)
 688{
 689	/* Output for debug, this have to be removed */
 690	int i;
 691	const char *name[4] =
 692		{ "acquire", "acquired", "contended", "release" };
 693
 694	pr_info("\n=== output for debug===\n\n");
 695	pr_info("bad: %d, total: %d\n", bad, total);
 696	pr_info("bad rate: %.2f %%\n", (double)bad / (double)total * 100);
 697	pr_info("histogram of events caused bad sequence\n");
 698	for (i = 0; i < BROKEN_MAX; i++)
 699		pr_info(" %10s: %d\n", name[i], bad_hist[i]);
 700}
 701
 702/* TODO: various way to print, coloring, nano or milli sec */
 703static void print_result(void)
 704{
 705	struct lock_stat *st;
 706	char cut_name[20];
 707	int bad, total;
 708
 709	pr_info("%20s ", "Name");
 710	pr_info("%10s ", "acquired");
 711	pr_info("%10s ", "contended");
 712
 713	pr_info("%15s ", "avg wait (ns)");
 714	pr_info("%15s ", "total wait (ns)");
 715	pr_info("%15s ", "max wait (ns)");
 716	pr_info("%15s ", "min wait (ns)");
 717
 718	pr_info("\n\n");
 719
 720	bad = total = 0;
 721	while ((st = pop_from_result())) {
 722		total++;
 723		if (st->discard) {
 724			bad++;
 725			continue;
 726		}
 727		bzero(cut_name, 20);
 728
 729		if (strlen(st->name) < 16) {
 730			/* output raw name */
 731			pr_info("%20s ", st->name);
 732		} else {
 733			strncpy(cut_name, st->name, 16);
 734			cut_name[16] = '.';
 735			cut_name[17] = '.';
 736			cut_name[18] = '.';
 737			cut_name[19] = '\0';
 738			/* cut off name for saving output style */
 739			pr_info("%20s ", cut_name);
 740		}
 741
 742		pr_info("%10u ", st->nr_acquired);
 743		pr_info("%10u ", st->nr_contended);
 744
 745		pr_info("%15" PRIu64 " ", st->avg_wait_time);
 746		pr_info("%15" PRIu64 " ", st->wait_time_total);
 747		pr_info("%15" PRIu64 " ", st->wait_time_max);
 748		pr_info("%15" PRIu64 " ", st->wait_time_min == ULLONG_MAX ?
 749		       0 : st->wait_time_min);
 750		pr_info("\n");
 751	}
 752
 753	print_bad_events(bad, total);
 754}
 755
 756static bool info_threads, info_map;
 757
 758static void dump_threads(void)
 759{
 760	struct thread_stat *st;
 761	struct rb_node *node;
 762	struct thread *t;
 763
 764	pr_info("%10s: comm\n", "Thread ID");
 765
 766	node = rb_first(&thread_stats);
 767	while (node) {
 768		st = container_of(node, struct thread_stat, rb);
 769		t = perf_session__findnew(session, st->tid);
 770		pr_info("%10d: %s\n", st->tid, thread__comm_str(t));
 771		node = rb_next(node);
 772		thread__put(t);
 773	};
 774}
 775
 776static void dump_map(void)
 777{
 778	unsigned int i;
 779	struct lock_stat *st;
 780
 781	pr_info("Address of instance: name of class\n");
 782	for (i = 0; i < LOCKHASH_SIZE; i++) {
 783		list_for_each_entry(st, &lockhash_table[i], hash_entry) {
 784			pr_info(" %p: %s\n", st->addr, st->name);
 785		}
 786	}
 787}
 788
 789static int dump_info(void)
 790{
 791	int rc = 0;
 792
 793	if (info_threads)
 794		dump_threads();
 795	else if (info_map)
 796		dump_map();
 797	else {
 798		rc = -1;
 799		pr_err("Unknown type of information\n");
 800	}
 801
 802	return rc;
 803}
 804
 805typedef int (*tracepoint_handler)(struct perf_evsel *evsel,
 806				  struct perf_sample *sample);
 807
 808static int process_sample_event(struct perf_tool *tool __maybe_unused,
 809				union perf_event *event,
 810				struct perf_sample *sample,
 811				struct perf_evsel *evsel,
 812				struct machine *machine)
 813{
 814	int err = 0;
 815	struct thread *thread = machine__findnew_thread(machine, sample->pid,
 816							sample->tid);
 817
 818	if (thread == NULL) {
 819		pr_debug("problem processing %d event, skipping it.\n",
 820			event->header.type);
 821		return -1;
 822	}
 823
 824	if (evsel->handler != NULL) {
 825		tracepoint_handler f = evsel->handler;
 826		err = f(evsel, sample);
 827	}
 
 
 
 
 
 
 828
 829	thread__put(thread);
 
 
 
 
 830
 831	return err;
 832}
 833
 834static void sort_result(void)
 835{
 836	unsigned int i;
 837	struct lock_stat *st;
 838
 839	for (i = 0; i < LOCKHASH_SIZE; i++) {
 840		list_for_each_entry(st, &lockhash_table[i], hash_entry) {
 841			insert_to_result(st, compare);
 842		}
 843	}
 844}
 845
 846static const struct perf_evsel_str_handler lock_tracepoints[] = {
 847	{ "lock:lock_acquire",	 perf_evsel__process_lock_acquire,   }, /* CONFIG_LOCKDEP */
 848	{ "lock:lock_acquired",	 perf_evsel__process_lock_acquired,  }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
 849	{ "lock:lock_contended", perf_evsel__process_lock_contended, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
 850	{ "lock:lock_release",	 perf_evsel__process_lock_release,   }, /* CONFIG_LOCKDEP */
 851};
 852
 853static bool force;
 854
 855static int __cmd_report(bool display_info)
 856{
 857	int err = -EINVAL;
 858	struct perf_tool eops = {
 859		.sample		 = process_sample_event,
 860		.comm		 = perf_event__process_comm,
 861		.ordered_events	 = true,
 862	};
 863	struct perf_data_file file = {
 864		.path = input_name,
 865		.mode = PERF_DATA_MODE_READ,
 866		.force = force,
 867	};
 868
 869	session = perf_session__new(&file, false, &eops);
 870	if (!session) {
 871		pr_err("Initializing perf session failed\n");
 872		return -1;
 873	}
 874
 875	symbol__init(&session->header.env);
 
 
 
 
 
 876
 877	if (!perf_session__has_traces(session, "lock record"))
 878		goto out_delete;
 
 
 879
 880	if (perf_session__set_tracepoints_handlers(session, lock_tracepoints)) {
 881		pr_err("Initializing perf session tracepoint handlers failed\n");
 882		goto out_delete;
 883	}
 
 
 
 884
 885	if (select_key())
 886		goto out_delete;
 
 
 887
 888	err = perf_session__process_events(session);
 889	if (err)
 890		goto out_delete;
 
 
 
 891
 892	setup_pager();
 893	if (display_info) /* used for info subcommand */
 894		err = dump_info();
 895	else {
 896		sort_result();
 897		print_result();
 898	}
 899
 900out_delete:
 901	perf_session__delete(session);
 902	return err;
 903}
 904
 905static int __cmd_record(int argc, const char **argv)
 906{
 907	const char *record_args[] = {
 908		"record", "-R", "-m", "1024", "-c", "1",
 909	};
 910	unsigned int rec_argc, i, j, ret;
 911	const char **rec_argv;
 912
 913	for (i = 0; i < ARRAY_SIZE(lock_tracepoints); i++) {
 914		if (!is_valid_tracepoint(lock_tracepoints[i].name)) {
 915				pr_err("tracepoint %s is not enabled. "
 916				       "Are CONFIG_LOCKDEP and CONFIG_LOCK_STAT enabled?\n",
 917				       lock_tracepoints[i].name);
 918				return 1;
 919		}
 920	}
 921
 922	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
 923	/* factor of 2 is for -e in front of each tracepoint */
 924	rec_argc += 2 * ARRAY_SIZE(lock_tracepoints);
 925
 926	rec_argv = calloc(rec_argc + 1, sizeof(char *));
 927	if (!rec_argv)
 928		return -ENOMEM;
 929
 930	for (i = 0; i < ARRAY_SIZE(record_args); i++)
 931		rec_argv[i] = strdup(record_args[i]);
 932
 933	for (j = 0; j < ARRAY_SIZE(lock_tracepoints); j++) {
 934		rec_argv[i++] = "-e";
 935		rec_argv[i++] = strdup(lock_tracepoints[j].name);
 936	}
 937
 938	for (j = 1; j < (unsigned int)argc; j++, i++)
 939		rec_argv[i] = argv[j];
 940
 941	BUG_ON(i != rec_argc);
 942
 943	ret = cmd_record(i, rec_argv, NULL);
 944	free(rec_argv);
 945	return ret;
 946}
 947
 948int cmd_lock(int argc, const char **argv, const char *prefix __maybe_unused)
 949{
 950	const struct option info_options[] = {
 951	OPT_BOOLEAN('t', "threads", &info_threads,
 952		    "dump thread list in perf.data"),
 953	OPT_BOOLEAN('m', "map", &info_map,
 954		    "map of lock instances (address:name table)"),
 955	OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
 956	OPT_END()
 957	};
 958	const struct option lock_options[] = {
 959	OPT_STRING('i', "input", &input_name, "file", "input file name"),
 960	OPT_INCR('v', "verbose", &verbose, "be more verbose (show symbol address, etc)"),
 961	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, "dump raw trace in ASCII"),
 962	OPT_END()
 963	};
 964	const struct option report_options[] = {
 965	OPT_STRING('k', "key", &sort_key, "acquired",
 966		    "key for sorting (acquired / contended / avg_wait / wait_total / wait_max / wait_min)"),
 967	OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
 968	/* TODO: type */
 969	OPT_END()
 970	};
 971	const char * const info_usage[] = {
 972		"perf lock info [<options>]",
 973		NULL
 974	};
 975	const char *const lock_subcommands[] = { "record", "report", "script",
 976						 "info", NULL };
 977	const char *lock_usage[] = {
 978		NULL,
 979		NULL
 980	};
 981	const char * const report_usage[] = {
 982		"perf lock report [<options>]",
 983		NULL
 984	};
 985	unsigned int i;
 986	int rc = 0;
 987
 
 988	for (i = 0; i < LOCKHASH_SIZE; i++)
 989		INIT_LIST_HEAD(lockhash_table + i);
 990
 991	argc = parse_options_subcommand(argc, argv, lock_options, lock_subcommands,
 992					lock_usage, PARSE_OPT_STOP_AT_NON_OPTION);
 993	if (!argc)
 994		usage_with_options(lock_usage, lock_options);
 995
 996	if (!strncmp(argv[0], "rec", 3)) {
 997		return __cmd_record(argc, argv);
 998	} else if (!strncmp(argv[0], "report", 6)) {
 999		trace_handler = &report_lock_ops;
1000		if (argc) {
1001			argc = parse_options(argc, argv,
1002					     report_options, report_usage, 0);
1003			if (argc)
1004				usage_with_options(report_usage, report_options);
1005		}
1006		rc = __cmd_report(false);
1007	} else if (!strcmp(argv[0], "script")) {
1008		/* Aliased to 'perf script' */
1009		return cmd_script(argc, argv, prefix);
1010	} else if (!strcmp(argv[0], "info")) {
1011		if (argc) {
1012			argc = parse_options(argc, argv,
1013					     info_options, info_usage, 0);
1014			if (argc)
1015				usage_with_options(info_usage, info_options);
1016		}
1017		/* recycling report_lock_ops */
1018		trace_handler = &report_lock_ops;
1019		rc = __cmd_report(true);
 
 
1020	} else {
1021		usage_with_options(lock_usage, lock_options);
1022	}
1023
1024	return rc;
1025}