Linux Audio

Check our new training course

Loading...
v3.5.6
   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#include "util/tool.h"
 
  16
  17#include <sys/types.h>
  18#include <sys/prctl.h>
  19#include <semaphore.h>
  20#include <pthread.h>
  21#include <math.h>
  22#include <limits.h>
  23
  24#include <linux/list.h>
  25#include <linux/hash.h>
  26
  27static struct perf_session *session;
  28
  29/* based on kernel/lockdep.c */
  30#define LOCKHASH_BITS		12
  31#define LOCKHASH_SIZE		(1UL << LOCKHASH_BITS)
  32
  33static struct list_head lockhash_table[LOCKHASH_SIZE];
  34
  35#define __lockhashfn(key)	hash_long((unsigned long)key, LOCKHASH_BITS)
  36#define lockhashentry(key)	(lockhash_table + __lockhashfn((key)))
  37
  38struct lock_stat {
  39	struct list_head	hash_entry;
  40	struct rb_node		rb;		/* used for sorting */
  41
  42	/*
  43	 * FIXME: raw_field_value() returns unsigned long long,
  44	 * so address of lockdep_map should be dealed as 64bit.
  45	 * Is there more better solution?
  46	 */
  47	void			*addr;		/* address of lockdep_map, used as ID */
  48	char			*name;		/* for strcpy(), we cannot use const */
  49
  50	unsigned int		nr_acquire;
  51	unsigned int		nr_acquired;
  52	unsigned int		nr_contended;
  53	unsigned int		nr_release;
  54
  55	unsigned int		nr_readlock;
  56	unsigned int		nr_trylock;
 
  57	/* these times are in nano sec. */
 
  58	u64			wait_time_total;
  59	u64			wait_time_min;
  60	u64			wait_time_max;
  61
  62	int			discard; /* flag of blacklist */
  63};
  64
  65/*
  66 * States of lock_seq_stat
  67 *
  68 * UNINITIALIZED is required for detecting first event of acquire.
  69 * As the nature of lock events, there is no guarantee
  70 * that the first event for the locks are acquire,
  71 * it can be acquired, contended or release.
  72 */
  73#define SEQ_STATE_UNINITIALIZED      0	       /* initial state */
  74#define SEQ_STATE_RELEASED	1
  75#define SEQ_STATE_ACQUIRING	2
  76#define SEQ_STATE_ACQUIRED	3
  77#define SEQ_STATE_READ_ACQUIRED	4
  78#define SEQ_STATE_CONTENDED	5
  79
  80/*
  81 * MAX_LOCK_DEPTH
  82 * Imported from include/linux/sched.h.
  83 * Should this be synchronized?
  84 */
  85#define MAX_LOCK_DEPTH 48
  86
  87/*
  88 * struct lock_seq_stat:
  89 * Place to put on state of one lock sequence
  90 * 1) acquire -> acquired -> release
  91 * 2) acquire -> contended -> acquired -> release
  92 * 3) acquire (with read or try) -> release
  93 * 4) Are there other patterns?
  94 */
  95struct lock_seq_stat {
  96	struct list_head        list;
  97	int			state;
  98	u64			prev_event_time;
  99	void                    *addr;
 100
 101	int                     read_count;
 102};
 103
 104struct thread_stat {
 105	struct rb_node		rb;
 106
 107	u32                     tid;
 108	struct list_head        seq_list;
 109};
 110
 111static struct rb_root		thread_stats;
 112
 113static struct thread_stat *thread_stat_find(u32 tid)
 114{
 115	struct rb_node *node;
 116	struct thread_stat *st;
 117
 118	node = thread_stats.rb_node;
 119	while (node) {
 120		st = container_of(node, struct thread_stat, rb);
 121		if (st->tid == tid)
 122			return st;
 123		else if (tid < st->tid)
 124			node = node->rb_left;
 125		else
 126			node = node->rb_right;
 127	}
 128
 129	return NULL;
 130}
 131
 132static void thread_stat_insert(struct thread_stat *new)
 133{
 134	struct rb_node **rb = &thread_stats.rb_node;
 135	struct rb_node *parent = NULL;
 136	struct thread_stat *p;
 137
 138	while (*rb) {
 139		p = container_of(*rb, struct thread_stat, rb);
 140		parent = *rb;
 141
 142		if (new->tid < p->tid)
 143			rb = &(*rb)->rb_left;
 144		else if (new->tid > p->tid)
 145			rb = &(*rb)->rb_right;
 146		else
 147			BUG_ON("inserting invalid thread_stat\n");
 148	}
 149
 150	rb_link_node(&new->rb, parent, rb);
 151	rb_insert_color(&new->rb, &thread_stats);
 152}
 153
 154static struct thread_stat *thread_stat_findnew_after_first(u32 tid)
 155{
 156	struct thread_stat *st;
 157
 158	st = thread_stat_find(tid);
 159	if (st)
 160		return st;
 161
 162	st = zalloc(sizeof(struct thread_stat));
 163	if (!st)
 164		die("memory allocation failed\n");
 
 
 165
 166	st->tid = tid;
 167	INIT_LIST_HEAD(&st->seq_list);
 168
 169	thread_stat_insert(st);
 170
 171	return st;
 172}
 173
 174static struct thread_stat *thread_stat_findnew_first(u32 tid);
 175static struct thread_stat *(*thread_stat_findnew)(u32 tid) =
 176	thread_stat_findnew_first;
 177
 178static struct thread_stat *thread_stat_findnew_first(u32 tid)
 179{
 180	struct thread_stat *st;
 181
 182	st = zalloc(sizeof(struct thread_stat));
 183	if (!st)
 184		die("memory allocation failed\n");
 
 
 185	st->tid = tid;
 186	INIT_LIST_HEAD(&st->seq_list);
 187
 188	rb_link_node(&st->rb, NULL, &thread_stats.rb_node);
 189	rb_insert_color(&st->rb, &thread_stats);
 190
 191	thread_stat_findnew = thread_stat_findnew_after_first;
 192	return st;
 193}
 194
 195/* build simple key function one is bigger than two */
 196#define SINGLE_KEY(member)						\
 197	static int lock_stat_key_ ## member(struct lock_stat *one,	\
 198					 struct lock_stat *two)		\
 199	{								\
 200		return one->member > two->member;			\
 201	}
 202
 203SINGLE_KEY(nr_acquired)
 204SINGLE_KEY(nr_contended)
 
 205SINGLE_KEY(wait_time_total)
 206SINGLE_KEY(wait_time_max)
 207
 208static int lock_stat_key_wait_time_min(struct lock_stat *one,
 209					struct lock_stat *two)
 210{
 211	u64 s1 = one->wait_time_min;
 212	u64 s2 = two->wait_time_min;
 213	if (s1 == ULLONG_MAX)
 214		s1 = 0;
 215	if (s2 == ULLONG_MAX)
 216		s2 = 0;
 217	return s1 > s2;
 218}
 219
 220struct lock_key {
 221	/*
 222	 * name: the value for specify by user
 223	 * this should be simpler than raw name of member
 224	 * e.g. nr_acquired -> acquired, wait_time_total -> wait_total
 225	 */
 226	const char		*name;
 227	int			(*key)(struct lock_stat*, struct lock_stat*);
 228};
 229
 230static const char		*sort_key = "acquired";
 231
 232static int			(*compare)(struct lock_stat *, struct lock_stat *);
 233
 234static struct rb_root		result;	/* place to store sorted data */
 235
 236#define DEF_KEY_LOCK(name, fn_suffix)	\
 237	{ #name, lock_stat_key_ ## fn_suffix }
 238struct lock_key keys[] = {
 239	DEF_KEY_LOCK(acquired, nr_acquired),
 240	DEF_KEY_LOCK(contended, nr_contended),
 
 241	DEF_KEY_LOCK(wait_total, wait_time_total),
 242	DEF_KEY_LOCK(wait_min, wait_time_min),
 243	DEF_KEY_LOCK(wait_max, wait_time_max),
 244
 245	/* extra comparisons much complicated should be here */
 246
 247	{ NULL, NULL }
 248};
 249
 250static void select_key(void)
 251{
 252	int i;
 253
 254	for (i = 0; keys[i].name; i++) {
 255		if (!strcmp(keys[i].name, sort_key)) {
 256			compare = keys[i].key;
 257			return;
 258		}
 259	}
 260
 261	die("Unknown compare key:%s\n", sort_key);
 
 
 262}
 263
 264static void insert_to_result(struct lock_stat *st,
 265			     int (*bigger)(struct lock_stat *, struct lock_stat *))
 266{
 267	struct rb_node **rb = &result.rb_node;
 268	struct rb_node *parent = NULL;
 269	struct lock_stat *p;
 270
 271	while (*rb) {
 272		p = container_of(*rb, struct lock_stat, rb);
 273		parent = *rb;
 274
 275		if (bigger(st, p))
 276			rb = &(*rb)->rb_left;
 277		else
 278			rb = &(*rb)->rb_right;
 279	}
 280
 281	rb_link_node(&st->rb, parent, rb);
 282	rb_insert_color(&st->rb, &result);
 283}
 284
 285/* returns left most element of result, and erase it */
 286static struct lock_stat *pop_from_result(void)
 287{
 288	struct rb_node *node = result.rb_node;
 289
 290	if (!node)
 291		return NULL;
 292
 293	while (node->rb_left)
 294		node = node->rb_left;
 295
 296	rb_erase(node, &result);
 297	return container_of(node, struct lock_stat, rb);
 298}
 299
 300static struct lock_stat *lock_stat_findnew(void *addr, const char *name)
 301{
 302	struct list_head *entry = lockhashentry(addr);
 303	struct lock_stat *ret, *new;
 304
 305	list_for_each_entry(ret, entry, hash_entry) {
 306		if (ret->addr == addr)
 307			return ret;
 308	}
 309
 310	new = zalloc(sizeof(struct lock_stat));
 311	if (!new)
 312		goto alloc_failed;
 313
 314	new->addr = addr;
 315	new->name = zalloc(sizeof(char) * strlen(name) + 1);
 316	if (!new->name)
 
 317		goto alloc_failed;
 318	strcpy(new->name, name);
 319
 
 320	new->wait_time_min = ULLONG_MAX;
 321
 322	list_add(&new->hash_entry, entry);
 323	return new;
 324
 325alloc_failed:
 326	die("memory allocation failed\n");
 
 327}
 328
 329static const char *input_name;
 
 
 330
 331struct raw_event_sample {
 332	u32			size;
 333	char			data[0];
 334};
 335
 336struct trace_acquire_event {
 337	void			*addr;
 338	const char		*name;
 339	int			flag;
 340};
 341
 342struct trace_acquired_event {
 343	void			*addr;
 344	const char		*name;
 345};
 346
 347struct trace_contended_event {
 348	void			*addr;
 349	const char		*name;
 350};
 351
 352struct trace_release_event {
 353	void			*addr;
 354	const char		*name;
 355};
 356
 357struct trace_lock_handler {
 358	void (*acquire_event)(struct trace_acquire_event *,
 359			      struct event_format *,
 360			      int cpu,
 361			      u64 timestamp,
 362			      struct thread *thread);
 363
 364	void (*acquired_event)(struct trace_acquired_event *,
 365			       struct event_format *,
 366			       int cpu,
 367			       u64 timestamp,
 368			       struct thread *thread);
 369
 370	void (*contended_event)(struct trace_contended_event *,
 371				struct event_format *,
 372				int cpu,
 373				u64 timestamp,
 374				struct thread *thread);
 375
 376	void (*release_event)(struct trace_release_event *,
 377			      struct event_format *,
 378			      int cpu,
 379			      u64 timestamp,
 380			      struct thread *thread);
 381};
 382
 383static struct lock_seq_stat *get_seq(struct thread_stat *ts, void *addr)
 384{
 385	struct lock_seq_stat *seq;
 386
 387	list_for_each_entry(seq, &ts->seq_list, list) {
 388		if (seq->addr == addr)
 389			return seq;
 390	}
 391
 392	seq = zalloc(sizeof(struct lock_seq_stat));
 393	if (!seq)
 394		die("Not enough memory\n");
 
 
 395	seq->state = SEQ_STATE_UNINITIALIZED;
 396	seq->addr = addr;
 397
 398	list_add(&seq->list, &ts->seq_list);
 399	return seq;
 400}
 401
 402enum broken_state {
 403	BROKEN_ACQUIRE,
 404	BROKEN_ACQUIRED,
 405	BROKEN_CONTENDED,
 406	BROKEN_RELEASE,
 407	BROKEN_MAX,
 408};
 409
 410static int bad_hist[BROKEN_MAX];
 411
 412enum acquire_flags {
 413	TRY_LOCK = 1,
 414	READ_LOCK = 2,
 415};
 416
 417static void
 418report_lock_acquire_event(struct trace_acquire_event *acquire_event,
 419			struct event_format *__event __used,
 420			int cpu __used,
 421			u64 timestamp __used,
 422			struct thread *thread __used)
 423{
 
 424	struct lock_stat *ls;
 425	struct thread_stat *ts;
 426	struct lock_seq_stat *seq;
 
 
 
 
 
 427
 428	ls = lock_stat_findnew(acquire_event->addr, acquire_event->name);
 
 
 429	if (ls->discard)
 430		return;
 
 
 
 
 431
 432	ts = thread_stat_findnew(thread->pid);
 433	seq = get_seq(ts, acquire_event->addr);
 
 434
 435	switch (seq->state) {
 436	case SEQ_STATE_UNINITIALIZED:
 437	case SEQ_STATE_RELEASED:
 438		if (!acquire_event->flag) {
 439			seq->state = SEQ_STATE_ACQUIRING;
 440		} else {
 441			if (acquire_event->flag & TRY_LOCK)
 442				ls->nr_trylock++;
 443			if (acquire_event->flag & READ_LOCK)
 444				ls->nr_readlock++;
 445			seq->state = SEQ_STATE_READ_ACQUIRED;
 446			seq->read_count = 1;
 447			ls->nr_acquired++;
 448		}
 449		break;
 450	case SEQ_STATE_READ_ACQUIRED:
 451		if (acquire_event->flag & READ_LOCK) {
 452			seq->read_count++;
 453			ls->nr_acquired++;
 454			goto end;
 455		} else {
 456			goto broken;
 457		}
 458		break;
 459	case SEQ_STATE_ACQUIRED:
 460	case SEQ_STATE_ACQUIRING:
 461	case SEQ_STATE_CONTENDED:
 462broken:
 463		/* broken lock sequence, discard it */
 464		ls->discard = 1;
 465		bad_hist[BROKEN_ACQUIRE]++;
 466		list_del(&seq->list);
 467		free(seq);
 468		goto end;
 469		break;
 470	default:
 471		BUG_ON("Unknown state of lock sequence found!\n");
 472		break;
 473	}
 474
 475	ls->nr_acquire++;
 476	seq->prev_event_time = timestamp;
 477end:
 478	return;
 479}
 480
 481static void
 482report_lock_acquired_event(struct trace_acquired_event *acquired_event,
 483			 struct event_format *__event __used,
 484			 int cpu __used,
 485			 u64 timestamp __used,
 486			 struct thread *thread __used)
 487{
 
 488	struct lock_stat *ls;
 489	struct thread_stat *ts;
 490	struct lock_seq_stat *seq;
 491	u64 contended_term;
 
 
 492
 493	ls = lock_stat_findnew(acquired_event->addr, acquired_event->name);
 
 
 
 
 494	if (ls->discard)
 495		return;
 496
 497	ts = thread_stat_findnew(thread->pid);
 498	seq = get_seq(ts, acquired_event->addr);
 
 
 
 
 
 499
 500	switch (seq->state) {
 501	case SEQ_STATE_UNINITIALIZED:
 502		/* orphan event, do nothing */
 503		return;
 504	case SEQ_STATE_ACQUIRING:
 505		break;
 506	case SEQ_STATE_CONTENDED:
 507		contended_term = timestamp - seq->prev_event_time;
 508		ls->wait_time_total += contended_term;
 509		if (contended_term < ls->wait_time_min)
 510			ls->wait_time_min = contended_term;
 511		if (ls->wait_time_max < contended_term)
 512			ls->wait_time_max = contended_term;
 513		break;
 514	case SEQ_STATE_RELEASED:
 515	case SEQ_STATE_ACQUIRED:
 516	case SEQ_STATE_READ_ACQUIRED:
 517		/* broken lock sequence, discard it */
 518		ls->discard = 1;
 519		bad_hist[BROKEN_ACQUIRED]++;
 520		list_del(&seq->list);
 521		free(seq);
 522		goto end;
 523		break;
 524
 525	default:
 526		BUG_ON("Unknown state of lock sequence found!\n");
 527		break;
 528	}
 529
 530	seq->state = SEQ_STATE_ACQUIRED;
 531	ls->nr_acquired++;
 532	seq->prev_event_time = timestamp;
 
 533end:
 534	return;
 535}
 536
 537static void
 538report_lock_contended_event(struct trace_contended_event *contended_event,
 539			  struct event_format *__event __used,
 540			  int cpu __used,
 541			  u64 timestamp __used,
 542			  struct thread *thread __used)
 543{
 
 544	struct lock_stat *ls;
 545	struct thread_stat *ts;
 546	struct lock_seq_stat *seq;
 
 
 
 
 547
 548	ls = lock_stat_findnew(contended_event->addr, contended_event->name);
 
 
 549	if (ls->discard)
 550		return;
 551
 552	ts = thread_stat_findnew(thread->pid);
 553	seq = get_seq(ts, contended_event->addr);
 
 
 
 
 
 554
 555	switch (seq->state) {
 556	case SEQ_STATE_UNINITIALIZED:
 557		/* orphan event, do nothing */
 558		return;
 559	case SEQ_STATE_ACQUIRING:
 560		break;
 561	case SEQ_STATE_RELEASED:
 562	case SEQ_STATE_ACQUIRED:
 563	case SEQ_STATE_READ_ACQUIRED:
 564	case SEQ_STATE_CONTENDED:
 565		/* broken lock sequence, discard it */
 566		ls->discard = 1;
 567		bad_hist[BROKEN_CONTENDED]++;
 568		list_del(&seq->list);
 569		free(seq);
 570		goto end;
 571		break;
 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	seq->prev_event_time = timestamp;
 
 580end:
 581	return;
 582}
 583
 584static void
 585report_lock_release_event(struct trace_release_event *release_event,
 586			struct event_format *__event __used,
 587			int cpu __used,
 588			u64 timestamp __used,
 589			struct thread *thread __used)
 590{
 
 591	struct lock_stat *ls;
 592	struct thread_stat *ts;
 593	struct lock_seq_stat *seq;
 
 
 594
 595	ls = lock_stat_findnew(release_event->addr, release_event->name);
 
 
 
 
 596	if (ls->discard)
 597		return;
 598
 599	ts = thread_stat_findnew(thread->pid);
 600	seq = get_seq(ts, release_event->addr);
 
 
 
 
 
 601
 602	switch (seq->state) {
 603	case SEQ_STATE_UNINITIALIZED:
 604		goto end;
 605		break;
 606	case SEQ_STATE_ACQUIRED:
 607		break;
 608	case SEQ_STATE_READ_ACQUIRED:
 609		seq->read_count--;
 610		BUG_ON(seq->read_count < 0);
 611		if (!seq->read_count) {
 612			ls->nr_release++;
 613			goto end;
 614		}
 615		break;
 616	case SEQ_STATE_ACQUIRING:
 617	case SEQ_STATE_CONTENDED:
 618	case SEQ_STATE_RELEASED:
 619		/* broken lock sequence, discard it */
 620		ls->discard = 1;
 621		bad_hist[BROKEN_RELEASE]++;
 622		goto free_seq;
 623		break;
 624	default:
 625		BUG_ON("Unknown state of lock sequence found!\n");
 626		break;
 627	}
 628
 629	ls->nr_release++;
 630free_seq:
 631	list_del(&seq->list);
 632	free(seq);
 633end:
 634	return;
 635}
 636
 637/* lock oriented handlers */
 638/* TODO: handlers for CPU oriented, thread oriented */
 639static struct trace_lock_handler report_lock_ops  = {
 640	.acquire_event		= report_lock_acquire_event,
 641	.acquired_event		= report_lock_acquired_event,
 642	.contended_event	= report_lock_contended_event,
 643	.release_event		= report_lock_release_event,
 644};
 645
 646static struct trace_lock_handler *trace_handler;
 647
 648static void
 649process_lock_acquire_event(void *data,
 650			   struct event_format *event __used,
 651			   int cpu __used,
 652			   u64 timestamp __used,
 653			   struct thread *thread __used)
 654{
 655	struct trace_acquire_event acquire_event;
 656	u64 tmp;		/* this is required for casting... */
 657
 658	tmp = raw_field_value(event, "lockdep_addr", data);
 659	memcpy(&acquire_event.addr, &tmp, sizeof(void *));
 660	acquire_event.name = (char *)raw_field_ptr(event, "name", data);
 661	acquire_event.flag = (int)raw_field_value(event, "flag", data);
 662
 663	if (trace_handler->acquire_event)
 664		trace_handler->acquire_event(&acquire_event, event, cpu, timestamp, thread);
 665}
 666
 667static void
 668process_lock_acquired_event(void *data,
 669			    struct event_format *event __used,
 670			    int cpu __used,
 671			    u64 timestamp __used,
 672			    struct thread *thread __used)
 673{
 674	struct trace_acquired_event acquired_event;
 675	u64 tmp;		/* this is required for casting... */
 676
 677	tmp = raw_field_value(event, "lockdep_addr", data);
 678	memcpy(&acquired_event.addr, &tmp, sizeof(void *));
 679	acquired_event.name = (char *)raw_field_ptr(event, "name", data);
 680
 681	if (trace_handler->acquire_event)
 682		trace_handler->acquired_event(&acquired_event, event, cpu, timestamp, thread);
 
 683}
 684
 685static void
 686process_lock_contended_event(void *data,
 687			     struct event_format *event __used,
 688			     int cpu __used,
 689			     u64 timestamp __used,
 690			     struct thread *thread __used)
 691{
 692	struct trace_contended_event contended_event;
 693	u64 tmp;		/* this is required for casting... */
 694
 695	tmp = raw_field_value(event, "lockdep_addr", data);
 696	memcpy(&contended_event.addr, &tmp, sizeof(void *));
 697	contended_event.name = (char *)raw_field_ptr(event, "name", data);
 698
 699	if (trace_handler->acquire_event)
 700		trace_handler->contended_event(&contended_event, event, cpu, timestamp, thread);
 701}
 702
 703static void
 704process_lock_release_event(void *data,
 705			   struct event_format *event __used,
 706			   int cpu __used,
 707			   u64 timestamp __used,
 708			   struct thread *thread __used)
 709{
 710	struct trace_release_event release_event;
 711	u64 tmp;		/* this is required for casting... */
 712
 713	tmp = raw_field_value(event, "lockdep_addr", data);
 714	memcpy(&release_event.addr, &tmp, sizeof(void *));
 715	release_event.name = (char *)raw_field_ptr(event, "name", data);
 716
 717	if (trace_handler->acquire_event)
 718		trace_handler->release_event(&release_event, event, cpu, timestamp, thread);
 719}
 720
 721static void
 722process_raw_event(void *data, int cpu, u64 timestamp, struct thread *thread)
 723{
 724	struct event_format *event;
 725	int type;
 726
 727	type = trace_parse_common_type(data);
 728	event = trace_find_event(type);
 729
 730	if (!strcmp(event->name, "lock_acquire"))
 731		process_lock_acquire_event(data, event, cpu, timestamp, thread);
 732	if (!strcmp(event->name, "lock_acquired"))
 733		process_lock_acquired_event(data, event, cpu, timestamp, thread);
 734	if (!strcmp(event->name, "lock_contended"))
 735		process_lock_contended_event(data, event, cpu, timestamp, thread);
 736	if (!strcmp(event->name, "lock_release"))
 737		process_lock_release_event(data, event, cpu, timestamp, thread);
 738}
 739
 740static void print_bad_events(int bad, int total)
 741{
 742	/* Output for debug, this have to be removed */
 743	int i;
 744	const char *name[4] =
 745		{ "acquire", "acquired", "contended", "release" };
 746
 747	pr_info("\n=== output for debug===\n\n");
 748	pr_info("bad: %d, total: %d\n", bad, total);
 749	pr_info("bad rate: %f %%\n", (double)bad / (double)total * 100);
 750	pr_info("histogram of events caused bad sequence\n");
 751	for (i = 0; i < BROKEN_MAX; i++)
 752		pr_info(" %10s: %d\n", name[i], bad_hist[i]);
 753}
 754
 755/* TODO: various way to print, coloring, nano or milli sec */
 756static void print_result(void)
 757{
 758	struct lock_stat *st;
 759	char cut_name[20];
 760	int bad, total;
 761
 762	pr_info("%20s ", "Name");
 763	pr_info("%10s ", "acquired");
 764	pr_info("%10s ", "contended");
 765
 
 766	pr_info("%15s ", "total wait (ns)");
 767	pr_info("%15s ", "max wait (ns)");
 768	pr_info("%15s ", "min wait (ns)");
 769
 770	pr_info("\n\n");
 771
 772	bad = total = 0;
 773	while ((st = pop_from_result())) {
 774		total++;
 775		if (st->discard) {
 776			bad++;
 777			continue;
 778		}
 779		bzero(cut_name, 20);
 780
 781		if (strlen(st->name) < 16) {
 782			/* output raw name */
 783			pr_info("%20s ", st->name);
 784		} else {
 785			strncpy(cut_name, st->name, 16);
 786			cut_name[16] = '.';
 787			cut_name[17] = '.';
 788			cut_name[18] = '.';
 789			cut_name[19] = '\0';
 790			/* cut off name for saving output style */
 791			pr_info("%20s ", cut_name);
 792		}
 793
 794		pr_info("%10u ", st->nr_acquired);
 795		pr_info("%10u ", st->nr_contended);
 796
 
 797		pr_info("%15" PRIu64 " ", st->wait_time_total);
 798		pr_info("%15" PRIu64 " ", st->wait_time_max);
 799		pr_info("%15" PRIu64 " ", st->wait_time_min == ULLONG_MAX ?
 800		       0 : st->wait_time_min);
 801		pr_info("\n");
 802	}
 803
 804	print_bad_events(bad, total);
 805}
 806
 807static bool info_threads, info_map;
 808
 809static void dump_threads(void)
 810{
 811	struct thread_stat *st;
 812	struct rb_node *node;
 813	struct thread *t;
 814
 815	pr_info("%10s: comm\n", "Thread ID");
 816
 817	node = rb_first(&thread_stats);
 818	while (node) {
 819		st = container_of(node, struct thread_stat, rb);
 820		t = perf_session__findnew(session, st->tid);
 821		pr_info("%10d: %s\n", st->tid, t->comm);
 822		node = rb_next(node);
 
 823	};
 824}
 825
 826static void dump_map(void)
 827{
 828	unsigned int i;
 829	struct lock_stat *st;
 830
 831	pr_info("Address of instance: name of class\n");
 832	for (i = 0; i < LOCKHASH_SIZE; i++) {
 833		list_for_each_entry(st, &lockhash_table[i], hash_entry) {
 834			pr_info(" %p: %s\n", st->addr, st->name);
 835		}
 836	}
 837}
 838
 839static void dump_info(void)
 840{
 
 
 841	if (info_threads)
 842		dump_threads();
 843	else if (info_map)
 844		dump_map();
 845	else
 846		die("Unknown type of information\n");
 
 
 
 
 847}
 848
 849static int process_sample_event(struct perf_tool *tool __used,
 
 
 
 850				union perf_event *event,
 851				struct perf_sample *sample,
 852				struct perf_evsel *evsel __used,
 853				struct machine *machine)
 854{
 855	struct thread *thread = machine__findnew_thread(machine, sample->tid);
 
 
 856
 857	if (thread == NULL) {
 858		pr_debug("problem processing %d event, skipping it.\n",
 859			event->header.type);
 860		return -1;
 861	}
 862
 863	process_raw_event(sample->raw_data, sample->cpu, sample->time, thread);
 864
 865	return 0;
 866}
 867
 868static struct perf_tool eops = {
 869	.sample			= process_sample_event,
 870	.comm			= perf_event__process_comm,
 871	.ordered_samples	= true,
 872};
 873
 874static int read_events(void)
 875{
 876	session = perf_session__new(input_name, O_RDONLY, 0, false, &eops);
 877	if (!session)
 878		die("Initializing perf session failed\n");
 879
 880	return perf_session__process_events(session, &eops);
 881}
 882
 883static void sort_result(void)
 884{
 885	unsigned int i;
 886	struct lock_stat *st;
 887
 888	for (i = 0; i < LOCKHASH_SIZE; i++) {
 889		list_for_each_entry(st, &lockhash_table[i], hash_entry) {
 890			insert_to_result(st, compare);
 891		}
 892	}
 893}
 894
 895static void __cmd_report(void)
 896{
 897	setup_pager();
 898	select_key();
 899	read_events();
 900	sort_result();
 901	print_result();
 902}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 903
 904static const char * const report_usage[] = {
 905	"perf lock report [<options>]",
 906	NULL
 907};
 
 908
 909static const struct option report_options[] = {
 910	OPT_STRING('k', "key", &sort_key, "acquired",
 911		    "key for sorting (acquired / contended / wait_total / wait_max / wait_min)"),
 912	/* TODO: type */
 913	OPT_END()
 914};
 915
 916static const char * const info_usage[] = {
 917	"perf lock info [<options>]",
 918	NULL
 919};
 920
 921static const struct option info_options[] = {
 922	OPT_BOOLEAN('t', "threads", &info_threads,
 923		    "dump thread list in perf.data"),
 924	OPT_BOOLEAN('m', "map", &info_map,
 925		    "map of lock instances (address:name table)"),
 926	OPT_END()
 927};
 928
 929static const char * const lock_usage[] = {
 930	"perf lock [<options>] {record|report|script|info}",
 931	NULL
 932};
 933
 934static const struct option lock_options[] = {
 935	OPT_STRING('i', "input", &input_name, "file", "input file name"),
 936	OPT_INCR('v', "verbose", &verbose, "be more verbose (show symbol address, etc)"),
 937	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, "dump raw trace in ASCII"),
 938	OPT_END()
 939};
 940
 941static const char *record_args[] = {
 942	"record",
 943	"-R",
 944	"-f",
 945	"-m", "1024",
 946	"-c", "1",
 947	"-e", "lock:lock_acquire",
 948	"-e", "lock:lock_acquired",
 949	"-e", "lock:lock_contended",
 950	"-e", "lock:lock_release",
 951};
 
 952
 953static int __cmd_record(int argc, const char **argv)
 954{
 955	unsigned int rec_argc, i, j;
 
 
 
 956	const char **rec_argv;
 957
 
 
 
 
 
 
 
 
 
 958	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
 959	rec_argv = calloc(rec_argc + 1, sizeof(char *));
 
 960
 961	if (rec_argv == NULL)
 
 962		return -ENOMEM;
 963
 964	for (i = 0; i < ARRAY_SIZE(record_args); i++)
 965		rec_argv[i] = strdup(record_args[i]);
 966
 
 
 
 
 
 967	for (j = 1; j < (unsigned int)argc; j++, i++)
 968		rec_argv[i] = argv[j];
 969
 970	BUG_ON(i != rec_argc);
 971
 972	return cmd_record(i, rec_argv, NULL);
 
 
 973}
 974
 975int cmd_lock(int argc, const char **argv, const char *prefix __used)
 976{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 977	unsigned int i;
 
 978
 979	symbol__init();
 980	for (i = 0; i < LOCKHASH_SIZE; i++)
 981		INIT_LIST_HEAD(lockhash_table + i);
 982
 983	argc = parse_options(argc, argv, lock_options, lock_usage,
 984			     PARSE_OPT_STOP_AT_NON_OPTION);
 985	if (!argc)
 986		usage_with_options(lock_usage, lock_options);
 987
 988	if (!strncmp(argv[0], "rec", 3)) {
 989		return __cmd_record(argc, argv);
 990	} else if (!strncmp(argv[0], "report", 6)) {
 991		trace_handler = &report_lock_ops;
 992		if (argc) {
 993			argc = parse_options(argc, argv,
 994					     report_options, report_usage, 0);
 995			if (argc)
 996				usage_with_options(report_usage, report_options);
 997		}
 998		__cmd_report();
 999	} else if (!strcmp(argv[0], "script")) {
1000		/* Aliased to 'perf script' */
1001		return cmd_script(argc, argv, prefix);
1002	} else if (!strcmp(argv[0], "info")) {
1003		if (argc) {
1004			argc = parse_options(argc, argv,
1005					     info_options, info_usage, 0);
1006			if (argc)
1007				usage_with_options(info_usage, info_options);
1008		}
1009		/* recycling report_lock_ops */
1010		trace_handler = &report_lock_ops;
1011		setup_pager();
1012		read_events();
1013		dump_info();
1014	} else {
1015		usage_with_options(lock_usage, lock_options);
1016	}
1017
1018	return 0;
1019}
v4.6
   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}