Linux Audio

Check our new training course

Linux kernel drivers training

Mar 31-Apr 9, 2025, special US time zones
Register
Loading...
v3.15
 
 
 
   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 "util/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	};
 773}
 774
 775static void dump_map(void)
 776{
 777	unsigned int i;
 778	struct lock_stat *st;
 779
 780	pr_info("Address of instance: name of class\n");
 781	for (i = 0; i < LOCKHASH_SIZE; i++) {
 782		list_for_each_entry(st, &lockhash_table[i], hash_entry) {
 783			pr_info(" %p: %s\n", st->addr, st->name);
 784		}
 785	}
 786}
 787
 788static int dump_info(void)
 789{
 790	int rc = 0;
 791
 792	if (info_threads)
 793		dump_threads();
 794	else if (info_map)
 795		dump_map();
 796	else {
 797		rc = -1;
 798		pr_err("Unknown type of information\n");
 799	}
 800
 801	return rc;
 802}
 803
 804typedef int (*tracepoint_handler)(struct perf_evsel *evsel,
 805				  struct perf_sample *sample);
 806
 807static int process_sample_event(struct perf_tool *tool __maybe_unused,
 808				union perf_event *event,
 809				struct perf_sample *sample,
 810				struct perf_evsel *evsel,
 811				struct machine *machine)
 812{
 
 813	struct thread *thread = machine__findnew_thread(machine, sample->pid,
 814							sample->tid);
 815
 816	if (thread == NULL) {
 817		pr_debug("problem processing %d event, skipping it.\n",
 818			event->header.type);
 819		return -1;
 820	}
 821
 822	if (evsel->handler != NULL) {
 823		tracepoint_handler f = evsel->handler;
 824		return f(evsel, sample);
 825	}
 826
 827	return 0;
 
 
 828}
 829
 830static void sort_result(void)
 831{
 832	unsigned int i;
 833	struct lock_stat *st;
 834
 835	for (i = 0; i < LOCKHASH_SIZE; i++) {
 836		list_for_each_entry(st, &lockhash_table[i], hash_entry) {
 837			insert_to_result(st, compare);
 838		}
 839	}
 840}
 841
 842static const struct perf_evsel_str_handler lock_tracepoints[] = {
 843	{ "lock:lock_acquire",	 perf_evsel__process_lock_acquire,   }, /* CONFIG_LOCKDEP */
 844	{ "lock:lock_acquired",	 perf_evsel__process_lock_acquired,  }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
 845	{ "lock:lock_contended", perf_evsel__process_lock_contended, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
 846	{ "lock:lock_release",	 perf_evsel__process_lock_release,   }, /* CONFIG_LOCKDEP */
 847};
 848
 
 
 849static int __cmd_report(bool display_info)
 850{
 851	int err = -EINVAL;
 852	struct perf_tool eops = {
 853		.sample		 = process_sample_event,
 854		.comm		 = perf_event__process_comm,
 855		.ordered_samples = true,
 
 856	};
 857	struct perf_data_file file = {
 858		.path = input_name,
 859		.mode = PERF_DATA_MODE_READ,
 
 
 
 860	};
 861
 862	session = perf_session__new(&file, false, &eops);
 863	if (!session) {
 864		pr_err("Initializing perf session failed\n");
 865		return -ENOMEM;
 866	}
 867
 
 
 868	if (!perf_session__has_traces(session, "lock record"))
 869		goto out_delete;
 870
 871	if (perf_session__set_tracepoints_handlers(session, lock_tracepoints)) {
 872		pr_err("Initializing perf session tracepoint handlers failed\n");
 873		goto out_delete;
 874	}
 875
 876	if (select_key())
 877		goto out_delete;
 878
 879	err = perf_session__process_events(session, &eops);
 880	if (err)
 881		goto out_delete;
 882
 883	setup_pager();
 884	if (display_info) /* used for info subcommand */
 885		err = dump_info();
 886	else {
 887		sort_result();
 888		print_result();
 889	}
 890
 891out_delete:
 892	perf_session__delete(session);
 893	return err;
 894}
 895
 896static int __cmd_record(int argc, const char **argv)
 897{
 898	const char *record_args[] = {
 899		"record", "-R", "-m", "1024", "-c", "1",
 900	};
 901	unsigned int rec_argc, i, j, ret;
 902	const char **rec_argv;
 903
 904	for (i = 0; i < ARRAY_SIZE(lock_tracepoints); i++) {
 905		if (!is_valid_tracepoint(lock_tracepoints[i].name)) {
 906				pr_err("tracepoint %s is not enabled. "
 907				       "Are CONFIG_LOCKDEP and CONFIG_LOCK_STAT enabled?\n",
 908				       lock_tracepoints[i].name);
 909				return 1;
 910		}
 911	}
 912
 913	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
 914	/* factor of 2 is for -e in front of each tracepoint */
 915	rec_argc += 2 * ARRAY_SIZE(lock_tracepoints);
 916
 917	rec_argv = calloc(rec_argc + 1, sizeof(char *));
 918	if (!rec_argv)
 919		return -ENOMEM;
 920
 921	for (i = 0; i < ARRAY_SIZE(record_args); i++)
 922		rec_argv[i] = strdup(record_args[i]);
 923
 924	for (j = 0; j < ARRAY_SIZE(lock_tracepoints); j++) {
 925		rec_argv[i++] = "-e";
 926		rec_argv[i++] = strdup(lock_tracepoints[j].name);
 927	}
 928
 929	for (j = 1; j < (unsigned int)argc; j++, i++)
 930		rec_argv[i] = argv[j];
 931
 932	BUG_ON(i != rec_argc);
 933
 934	ret = cmd_record(i, rec_argv, NULL);
 935	free(rec_argv);
 936	return ret;
 937}
 938
 939int cmd_lock(int argc, const char **argv, const char *prefix __maybe_unused)
 940{
 941	const struct option info_options[] = {
 942	OPT_BOOLEAN('t', "threads", &info_threads,
 943		    "dump thread list in perf.data"),
 944	OPT_BOOLEAN('m', "map", &info_map,
 945		    "map of lock instances (address:name table)"),
 946	OPT_END()
 947	};
 948	const struct option lock_options[] = {
 949	OPT_STRING('i', "input", &input_name, "file", "input file name"),
 950	OPT_INCR('v', "verbose", &verbose, "be more verbose (show symbol address, etc)"),
 951	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, "dump raw trace in ASCII"),
 
 952	OPT_END()
 953	};
 
 
 
 
 
 
 
 
 
 954	const struct option report_options[] = {
 955	OPT_STRING('k', "key", &sort_key, "acquired",
 956		    "key for sorting (acquired / contended / avg_wait / wait_total / wait_max / wait_min)"),
 957	/* TODO: type */
 958	OPT_END()
 959	};
 
 960	const char * const info_usage[] = {
 961		"perf lock info [<options>]",
 962		NULL
 963	};
 964	const char * const lock_usage[] = {
 965		"perf lock [<options>] {record|report|script|info}",
 
 
 966		NULL
 967	};
 968	const char * const report_usage[] = {
 969		"perf lock report [<options>]",
 970		NULL
 971	};
 972	unsigned int i;
 973	int rc = 0;
 974
 975	symbol__init();
 976	for (i = 0; i < LOCKHASH_SIZE; i++)
 977		INIT_LIST_HEAD(lockhash_table + i);
 978
 979	argc = parse_options(argc, argv, lock_options, lock_usage,
 980			     PARSE_OPT_STOP_AT_NON_OPTION);
 981	if (!argc)
 982		usage_with_options(lock_usage, lock_options);
 983
 984	if (!strncmp(argv[0], "rec", 3)) {
 985		return __cmd_record(argc, argv);
 986	} else if (!strncmp(argv[0], "report", 6)) {
 987		trace_handler = &report_lock_ops;
 988		if (argc) {
 989			argc = parse_options(argc, argv,
 990					     report_options, report_usage, 0);
 991			if (argc)
 992				usage_with_options(report_usage, report_options);
 993		}
 994		rc = __cmd_report(false);
 995	} else if (!strcmp(argv[0], "script")) {
 996		/* Aliased to 'perf script' */
 997		return cmd_script(argc, argv, prefix);
 998	} else if (!strcmp(argv[0], "info")) {
 999		if (argc) {
1000			argc = parse_options(argc, argv,
1001					     info_options, info_usage, 0);
1002			if (argc)
1003				usage_with_options(info_usage, info_options);
1004		}
1005		/* recycling report_lock_ops */
1006		trace_handler = &report_lock_ops;
1007		rc = __cmd_report(true);
1008	} else {
1009		usage_with_options(lock_usage, lock_options);
1010	}
1011
1012	return rc;
1013}
v4.17
   1// SPDX-License-Identifier: GPL-2.0
   2#include <errno.h>
   3#include <inttypes.h>
   4#include "builtin.h"
   5#include "perf.h"
   6
   7#include "util/evlist.h"
   8#include "util/evsel.h"
   9#include "util/util.h"
  10#include "util/cache.h"
  11#include "util/symbol.h"
  12#include "util/thread.h"
  13#include "util/header.h"
  14
  15#include <subcmd/parse-options.h>
  16#include "util/trace-event.h"
  17
  18#include "util/debug.h"
  19#include "util/session.h"
  20#include "util/tool.h"
  21#include "util/data.h"
  22
  23#include <sys/types.h>
  24#include <sys/prctl.h>
  25#include <semaphore.h>
  26#include <pthread.h>
  27#include <math.h>
  28#include <limits.h>
  29
  30#include <linux/list.h>
  31#include <linux/hash.h>
  32#include <linux/kernel.h>
  33
  34static struct perf_session *session;
  35
  36/* based on kernel/lockdep.c */
  37#define LOCKHASH_BITS		12
  38#define LOCKHASH_SIZE		(1UL << LOCKHASH_BITS)
  39
  40static struct list_head lockhash_table[LOCKHASH_SIZE];
  41
  42#define __lockhashfn(key)	hash_long((unsigned long)key, LOCKHASH_BITS)
  43#define lockhashentry(key)	(lockhash_table + __lockhashfn((key)))
  44
  45struct lock_stat {
  46	struct list_head	hash_entry;
  47	struct rb_node		rb;		/* used for sorting */
  48
  49	/*
  50	 * FIXME: perf_evsel__intval() returns u64,
  51	 * so address of lockdep_map should be dealed as 64bit.
  52	 * Is there more better solution?
  53	 */
  54	void			*addr;		/* address of lockdep_map, used as ID */
  55	char			*name;		/* for strcpy(), we cannot use const */
  56
  57	unsigned int		nr_acquire;
  58	unsigned int		nr_acquired;
  59	unsigned int		nr_contended;
  60	unsigned int		nr_release;
  61
  62	unsigned int		nr_readlock;
  63	unsigned int		nr_trylock;
  64
  65	/* these times are in nano sec. */
  66	u64                     avg_wait_time;
  67	u64			wait_time_total;
  68	u64			wait_time_min;
  69	u64			wait_time_max;
  70
  71	int			discard; /* flag of blacklist */
  72};
  73
  74/*
  75 * States of lock_seq_stat
  76 *
  77 * UNINITIALIZED is required for detecting first event of acquire.
  78 * As the nature of lock events, there is no guarantee
  79 * that the first event for the locks are acquire,
  80 * it can be acquired, contended or release.
  81 */
  82#define SEQ_STATE_UNINITIALIZED      0	       /* initial state */
  83#define SEQ_STATE_RELEASED	1
  84#define SEQ_STATE_ACQUIRING	2
  85#define SEQ_STATE_ACQUIRED	3
  86#define SEQ_STATE_READ_ACQUIRED	4
  87#define SEQ_STATE_CONTENDED	5
  88
  89/*
  90 * MAX_LOCK_DEPTH
  91 * Imported from include/linux/sched.h.
  92 * Should this be synchronized?
  93 */
  94#define MAX_LOCK_DEPTH 48
  95
  96/*
  97 * struct lock_seq_stat:
  98 * Place to put on state of one lock sequence
  99 * 1) acquire -> acquired -> release
 100 * 2) acquire -> contended -> acquired -> release
 101 * 3) acquire (with read or try) -> release
 102 * 4) Are there other patterns?
 103 */
 104struct lock_seq_stat {
 105	struct list_head        list;
 106	int			state;
 107	u64			prev_event_time;
 108	void                    *addr;
 109
 110	int                     read_count;
 111};
 112
 113struct thread_stat {
 114	struct rb_node		rb;
 115
 116	u32                     tid;
 117	struct list_head        seq_list;
 118};
 119
 120static struct rb_root		thread_stats;
 121
 122static struct thread_stat *thread_stat_find(u32 tid)
 123{
 124	struct rb_node *node;
 125	struct thread_stat *st;
 126
 127	node = thread_stats.rb_node;
 128	while (node) {
 129		st = container_of(node, struct thread_stat, rb);
 130		if (st->tid == tid)
 131			return st;
 132		else if (tid < st->tid)
 133			node = node->rb_left;
 134		else
 135			node = node->rb_right;
 136	}
 137
 138	return NULL;
 139}
 140
 141static void thread_stat_insert(struct thread_stat *new)
 142{
 143	struct rb_node **rb = &thread_stats.rb_node;
 144	struct rb_node *parent = NULL;
 145	struct thread_stat *p;
 146
 147	while (*rb) {
 148		p = container_of(*rb, struct thread_stat, rb);
 149		parent = *rb;
 150
 151		if (new->tid < p->tid)
 152			rb = &(*rb)->rb_left;
 153		else if (new->tid > p->tid)
 154			rb = &(*rb)->rb_right;
 155		else
 156			BUG_ON("inserting invalid thread_stat\n");
 157	}
 158
 159	rb_link_node(&new->rb, parent, rb);
 160	rb_insert_color(&new->rb, &thread_stats);
 161}
 162
 163static struct thread_stat *thread_stat_findnew_after_first(u32 tid)
 164{
 165	struct thread_stat *st;
 166
 167	st = thread_stat_find(tid);
 168	if (st)
 169		return st;
 170
 171	st = zalloc(sizeof(struct thread_stat));
 172	if (!st) {
 173		pr_err("memory allocation failed\n");
 174		return NULL;
 175	}
 176
 177	st->tid = tid;
 178	INIT_LIST_HEAD(&st->seq_list);
 179
 180	thread_stat_insert(st);
 181
 182	return st;
 183}
 184
 185static struct thread_stat *thread_stat_findnew_first(u32 tid);
 186static struct thread_stat *(*thread_stat_findnew)(u32 tid) =
 187	thread_stat_findnew_first;
 188
 189static struct thread_stat *thread_stat_findnew_first(u32 tid)
 190{
 191	struct thread_stat *st;
 192
 193	st = zalloc(sizeof(struct thread_stat));
 194	if (!st) {
 195		pr_err("memory allocation failed\n");
 196		return NULL;
 197	}
 198	st->tid = tid;
 199	INIT_LIST_HEAD(&st->seq_list);
 200
 201	rb_link_node(&st->rb, NULL, &thread_stats.rb_node);
 202	rb_insert_color(&st->rb, &thread_stats);
 203
 204	thread_stat_findnew = thread_stat_findnew_after_first;
 205	return st;
 206}
 207
 208/* build simple key function one is bigger than two */
 209#define SINGLE_KEY(member)						\
 210	static int lock_stat_key_ ## member(struct lock_stat *one,	\
 211					 struct lock_stat *two)		\
 212	{								\
 213		return one->member > two->member;			\
 214	}
 215
 216SINGLE_KEY(nr_acquired)
 217SINGLE_KEY(nr_contended)
 218SINGLE_KEY(avg_wait_time)
 219SINGLE_KEY(wait_time_total)
 220SINGLE_KEY(wait_time_max)
 221
 222static int lock_stat_key_wait_time_min(struct lock_stat *one,
 223					struct lock_stat *two)
 224{
 225	u64 s1 = one->wait_time_min;
 226	u64 s2 = two->wait_time_min;
 227	if (s1 == ULLONG_MAX)
 228		s1 = 0;
 229	if (s2 == ULLONG_MAX)
 230		s2 = 0;
 231	return s1 > s2;
 232}
 233
 234struct lock_key {
 235	/*
 236	 * name: the value for specify by user
 237	 * this should be simpler than raw name of member
 238	 * e.g. nr_acquired -> acquired, wait_time_total -> wait_total
 239	 */
 240	const char		*name;
 241	int			(*key)(struct lock_stat*, struct lock_stat*);
 242};
 243
 244static const char		*sort_key = "acquired";
 245
 246static int			(*compare)(struct lock_stat *, struct lock_stat *);
 247
 248static struct rb_root		result;	/* place to store sorted data */
 249
 250#define DEF_KEY_LOCK(name, fn_suffix)	\
 251	{ #name, lock_stat_key_ ## fn_suffix }
 252struct lock_key keys[] = {
 253	DEF_KEY_LOCK(acquired, nr_acquired),
 254	DEF_KEY_LOCK(contended, nr_contended),
 255	DEF_KEY_LOCK(avg_wait, avg_wait_time),
 256	DEF_KEY_LOCK(wait_total, wait_time_total),
 257	DEF_KEY_LOCK(wait_min, wait_time_min),
 258	DEF_KEY_LOCK(wait_max, wait_time_max),
 259
 260	/* extra comparisons much complicated should be here */
 261
 262	{ NULL, NULL }
 263};
 264
 265static int select_key(void)
 266{
 267	int i;
 268
 269	for (i = 0; keys[i].name; i++) {
 270		if (!strcmp(keys[i].name, sort_key)) {
 271			compare = keys[i].key;
 272			return 0;
 273		}
 274	}
 275
 276	pr_err("Unknown compare key: %s\n", sort_key);
 277
 278	return -1;
 279}
 280
 281static void insert_to_result(struct lock_stat *st,
 282			     int (*bigger)(struct lock_stat *, struct lock_stat *))
 283{
 284	struct rb_node **rb = &result.rb_node;
 285	struct rb_node *parent = NULL;
 286	struct lock_stat *p;
 287
 288	while (*rb) {
 289		p = container_of(*rb, struct lock_stat, rb);
 290		parent = *rb;
 291
 292		if (bigger(st, p))
 293			rb = &(*rb)->rb_left;
 294		else
 295			rb = &(*rb)->rb_right;
 296	}
 297
 298	rb_link_node(&st->rb, parent, rb);
 299	rb_insert_color(&st->rb, &result);
 300}
 301
 302/* returns left most element of result, and erase it */
 303static struct lock_stat *pop_from_result(void)
 304{
 305	struct rb_node *node = result.rb_node;
 306
 307	if (!node)
 308		return NULL;
 309
 310	while (node->rb_left)
 311		node = node->rb_left;
 312
 313	rb_erase(node, &result);
 314	return container_of(node, struct lock_stat, rb);
 315}
 316
 317static struct lock_stat *lock_stat_findnew(void *addr, const char *name)
 318{
 319	struct list_head *entry = lockhashentry(addr);
 320	struct lock_stat *ret, *new;
 321
 322	list_for_each_entry(ret, entry, hash_entry) {
 323		if (ret->addr == addr)
 324			return ret;
 325	}
 326
 327	new = zalloc(sizeof(struct lock_stat));
 328	if (!new)
 329		goto alloc_failed;
 330
 331	new->addr = addr;
 332	new->name = zalloc(sizeof(char) * strlen(name) + 1);
 333	if (!new->name) {
 334		free(new);
 335		goto alloc_failed;
 336	}
 337
 338	strcpy(new->name, name);
 339	new->wait_time_min = ULLONG_MAX;
 340
 341	list_add(&new->hash_entry, entry);
 342	return new;
 343
 344alloc_failed:
 345	pr_err("memory allocation failed\n");
 346	return NULL;
 347}
 348
 349struct trace_lock_handler {
 350	int (*acquire_event)(struct perf_evsel *evsel,
 351			     struct perf_sample *sample);
 352
 353	int (*acquired_event)(struct perf_evsel *evsel,
 354			      struct perf_sample *sample);
 355
 356	int (*contended_event)(struct perf_evsel *evsel,
 357			       struct perf_sample *sample);
 358
 359	int (*release_event)(struct perf_evsel *evsel,
 360			     struct perf_sample *sample);
 361};
 362
 363static struct lock_seq_stat *get_seq(struct thread_stat *ts, void *addr)
 364{
 365	struct lock_seq_stat *seq;
 366
 367	list_for_each_entry(seq, &ts->seq_list, list) {
 368		if (seq->addr == addr)
 369			return seq;
 370	}
 371
 372	seq = zalloc(sizeof(struct lock_seq_stat));
 373	if (!seq) {
 374		pr_err("memory allocation failed\n");
 375		return NULL;
 376	}
 377	seq->state = SEQ_STATE_UNINITIALIZED;
 378	seq->addr = addr;
 379
 380	list_add(&seq->list, &ts->seq_list);
 381	return seq;
 382}
 383
 384enum broken_state {
 385	BROKEN_ACQUIRE,
 386	BROKEN_ACQUIRED,
 387	BROKEN_CONTENDED,
 388	BROKEN_RELEASE,
 389	BROKEN_MAX,
 390};
 391
 392static int bad_hist[BROKEN_MAX];
 393
 394enum acquire_flags {
 395	TRY_LOCK = 1,
 396	READ_LOCK = 2,
 397};
 398
 399static int report_lock_acquire_event(struct perf_evsel *evsel,
 400				     struct perf_sample *sample)
 401{
 402	void *addr;
 403	struct lock_stat *ls;
 404	struct thread_stat *ts;
 405	struct lock_seq_stat *seq;
 406	const char *name = perf_evsel__strval(evsel, sample, "name");
 407	u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
 408	int flag = perf_evsel__intval(evsel, sample, "flag");
 409
 410	memcpy(&addr, &tmp, sizeof(void *));
 411
 412	ls = lock_stat_findnew(addr, name);
 413	if (!ls)
 414		return -ENOMEM;
 415	if (ls->discard)
 416		return 0;
 417
 418	ts = thread_stat_findnew(sample->tid);
 419	if (!ts)
 420		return -ENOMEM;
 421
 422	seq = get_seq(ts, addr);
 423	if (!seq)
 424		return -ENOMEM;
 425
 426	switch (seq->state) {
 427	case SEQ_STATE_UNINITIALIZED:
 428	case SEQ_STATE_RELEASED:
 429		if (!flag) {
 430			seq->state = SEQ_STATE_ACQUIRING;
 431		} else {
 432			if (flag & TRY_LOCK)
 433				ls->nr_trylock++;
 434			if (flag & READ_LOCK)
 435				ls->nr_readlock++;
 436			seq->state = SEQ_STATE_READ_ACQUIRED;
 437			seq->read_count = 1;
 438			ls->nr_acquired++;
 439		}
 440		break;
 441	case SEQ_STATE_READ_ACQUIRED:
 442		if (flag & READ_LOCK) {
 443			seq->read_count++;
 444			ls->nr_acquired++;
 445			goto end;
 446		} else {
 447			goto broken;
 448		}
 449		break;
 450	case SEQ_STATE_ACQUIRED:
 451	case SEQ_STATE_ACQUIRING:
 452	case SEQ_STATE_CONTENDED:
 453broken:
 454		/* broken lock sequence, discard it */
 455		ls->discard = 1;
 456		bad_hist[BROKEN_ACQUIRE]++;
 457		list_del(&seq->list);
 458		free(seq);
 459		goto end;
 460	default:
 461		BUG_ON("Unknown state of lock sequence found!\n");
 462		break;
 463	}
 464
 465	ls->nr_acquire++;
 466	seq->prev_event_time = sample->time;
 467end:
 468	return 0;
 469}
 470
 471static int report_lock_acquired_event(struct perf_evsel *evsel,
 472				      struct perf_sample *sample)
 473{
 474	void *addr;
 475	struct lock_stat *ls;
 476	struct thread_stat *ts;
 477	struct lock_seq_stat *seq;
 478	u64 contended_term;
 479	const char *name = perf_evsel__strval(evsel, sample, "name");
 480	u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
 481
 482	memcpy(&addr, &tmp, sizeof(void *));
 483
 484	ls = lock_stat_findnew(addr, name);
 485	if (!ls)
 486		return -ENOMEM;
 487	if (ls->discard)
 488		return 0;
 489
 490	ts = thread_stat_findnew(sample->tid);
 491	if (!ts)
 492		return -ENOMEM;
 493
 494	seq = get_seq(ts, addr);
 495	if (!seq)
 496		return -ENOMEM;
 497
 498	switch (seq->state) {
 499	case SEQ_STATE_UNINITIALIZED:
 500		/* orphan event, do nothing */
 501		return 0;
 502	case SEQ_STATE_ACQUIRING:
 503		break;
 504	case SEQ_STATE_CONTENDED:
 505		contended_term = sample->time - seq->prev_event_time;
 506		ls->wait_time_total += contended_term;
 507		if (contended_term < ls->wait_time_min)
 508			ls->wait_time_min = contended_term;
 509		if (ls->wait_time_max < contended_term)
 510			ls->wait_time_max = contended_term;
 511		break;
 512	case SEQ_STATE_RELEASED:
 513	case SEQ_STATE_ACQUIRED:
 514	case SEQ_STATE_READ_ACQUIRED:
 515		/* broken lock sequence, discard it */
 516		ls->discard = 1;
 517		bad_hist[BROKEN_ACQUIRED]++;
 518		list_del(&seq->list);
 519		free(seq);
 520		goto end;
 521	default:
 522		BUG_ON("Unknown state of lock sequence found!\n");
 523		break;
 524	}
 525
 526	seq->state = SEQ_STATE_ACQUIRED;
 527	ls->nr_acquired++;
 528	ls->avg_wait_time = ls->nr_contended ? ls->wait_time_total/ls->nr_contended : 0;
 529	seq->prev_event_time = sample->time;
 530end:
 531	return 0;
 532}
 533
 534static int report_lock_contended_event(struct perf_evsel *evsel,
 535				       struct perf_sample *sample)
 536{
 537	void *addr;
 538	struct lock_stat *ls;
 539	struct thread_stat *ts;
 540	struct lock_seq_stat *seq;
 541	const char *name = perf_evsel__strval(evsel, sample, "name");
 542	u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
 543
 544	memcpy(&addr, &tmp, sizeof(void *));
 545
 546	ls = lock_stat_findnew(addr, name);
 547	if (!ls)
 548		return -ENOMEM;
 549	if (ls->discard)
 550		return 0;
 551
 552	ts = thread_stat_findnew(sample->tid);
 553	if (!ts)
 554		return -ENOMEM;
 555
 556	seq = get_seq(ts, addr);
 557	if (!seq)
 558		return -ENOMEM;
 559
 560	switch (seq->state) {
 561	case SEQ_STATE_UNINITIALIZED:
 562		/* orphan event, do nothing */
 563		return 0;
 564	case SEQ_STATE_ACQUIRING:
 565		break;
 566	case SEQ_STATE_RELEASED:
 567	case SEQ_STATE_ACQUIRED:
 568	case SEQ_STATE_READ_ACQUIRED:
 569	case SEQ_STATE_CONTENDED:
 570		/* broken lock sequence, discard it */
 571		ls->discard = 1;
 572		bad_hist[BROKEN_CONTENDED]++;
 573		list_del(&seq->list);
 574		free(seq);
 575		goto end;
 576	default:
 577		BUG_ON("Unknown state of lock sequence found!\n");
 578		break;
 579	}
 580
 581	seq->state = SEQ_STATE_CONTENDED;
 582	ls->nr_contended++;
 583	ls->avg_wait_time = ls->wait_time_total/ls->nr_contended;
 584	seq->prev_event_time = sample->time;
 585end:
 586	return 0;
 587}
 588
 589static int report_lock_release_event(struct perf_evsel *evsel,
 590				     struct perf_sample *sample)
 591{
 592	void *addr;
 593	struct lock_stat *ls;
 594	struct thread_stat *ts;
 595	struct lock_seq_stat *seq;
 596	const char *name = perf_evsel__strval(evsel, sample, "name");
 597	u64 tmp = perf_evsel__intval(evsel, sample, "lockdep_addr");
 598
 599	memcpy(&addr, &tmp, sizeof(void *));
 600
 601	ls = lock_stat_findnew(addr, name);
 602	if (!ls)
 603		return -ENOMEM;
 604	if (ls->discard)
 605		return 0;
 606
 607	ts = thread_stat_findnew(sample->tid);
 608	if (!ts)
 609		return -ENOMEM;
 610
 611	seq = get_seq(ts, addr);
 612	if (!seq)
 613		return -ENOMEM;
 614
 615	switch (seq->state) {
 616	case SEQ_STATE_UNINITIALIZED:
 617		goto end;
 618	case SEQ_STATE_ACQUIRED:
 619		break;
 620	case SEQ_STATE_READ_ACQUIRED:
 621		seq->read_count--;
 622		BUG_ON(seq->read_count < 0);
 623		if (!seq->read_count) {
 624			ls->nr_release++;
 625			goto end;
 626		}
 627		break;
 628	case SEQ_STATE_ACQUIRING:
 629	case SEQ_STATE_CONTENDED:
 630	case SEQ_STATE_RELEASED:
 631		/* broken lock sequence, discard it */
 632		ls->discard = 1;
 633		bad_hist[BROKEN_RELEASE]++;
 634		goto free_seq;
 635	default:
 636		BUG_ON("Unknown state of lock sequence found!\n");
 637		break;
 638	}
 639
 640	ls->nr_release++;
 641free_seq:
 642	list_del(&seq->list);
 643	free(seq);
 644end:
 645	return 0;
 646}
 647
 648/* lock oriented handlers */
 649/* TODO: handlers for CPU oriented, thread oriented */
 650static struct trace_lock_handler report_lock_ops  = {
 651	.acquire_event		= report_lock_acquire_event,
 652	.acquired_event		= report_lock_acquired_event,
 653	.contended_event	= report_lock_contended_event,
 654	.release_event		= report_lock_release_event,
 655};
 656
 657static struct trace_lock_handler *trace_handler;
 658
 659static int perf_evsel__process_lock_acquire(struct perf_evsel *evsel,
 660					     struct perf_sample *sample)
 661{
 662	if (trace_handler->acquire_event)
 663		return trace_handler->acquire_event(evsel, sample);
 664	return 0;
 665}
 666
 667static int perf_evsel__process_lock_acquired(struct perf_evsel *evsel,
 668					      struct perf_sample *sample)
 669{
 670	if (trace_handler->acquired_event)
 671		return trace_handler->acquired_event(evsel, sample);
 672	return 0;
 673}
 674
 675static int perf_evsel__process_lock_contended(struct perf_evsel *evsel,
 676					      struct perf_sample *sample)
 677{
 678	if (trace_handler->contended_event)
 679		return trace_handler->contended_event(evsel, sample);
 680	return 0;
 681}
 682
 683static int perf_evsel__process_lock_release(struct perf_evsel *evsel,
 684					    struct perf_sample *sample)
 685{
 686	if (trace_handler->release_event)
 687		return trace_handler->release_event(evsel, sample);
 688	return 0;
 689}
 690
 691static void print_bad_events(int bad, int total)
 692{
 693	/* Output for debug, this have to be removed */
 694	int i;
 695	const char *name[4] =
 696		{ "acquire", "acquired", "contended", "release" };
 697
 698	pr_info("\n=== output for debug===\n\n");
 699	pr_info("bad: %d, total: %d\n", bad, total);
 700	pr_info("bad rate: %.2f %%\n", (double)bad / (double)total * 100);
 701	pr_info("histogram of events caused bad sequence\n");
 702	for (i = 0; i < BROKEN_MAX; i++)
 703		pr_info(" %10s: %d\n", name[i], bad_hist[i]);
 704}
 705
 706/* TODO: various way to print, coloring, nano or milli sec */
 707static void print_result(void)
 708{
 709	struct lock_stat *st;
 710	char cut_name[20];
 711	int bad, total;
 712
 713	pr_info("%20s ", "Name");
 714	pr_info("%10s ", "acquired");
 715	pr_info("%10s ", "contended");
 716
 717	pr_info("%15s ", "avg wait (ns)");
 718	pr_info("%15s ", "total wait (ns)");
 719	pr_info("%15s ", "max wait (ns)");
 720	pr_info("%15s ", "min wait (ns)");
 721
 722	pr_info("\n\n");
 723
 724	bad = total = 0;
 725	while ((st = pop_from_result())) {
 726		total++;
 727		if (st->discard) {
 728			bad++;
 729			continue;
 730		}
 731		bzero(cut_name, 20);
 732
 733		if (strlen(st->name) < 16) {
 734			/* output raw name */
 735			pr_info("%20s ", st->name);
 736		} else {
 737			strncpy(cut_name, st->name, 16);
 738			cut_name[16] = '.';
 739			cut_name[17] = '.';
 740			cut_name[18] = '.';
 741			cut_name[19] = '\0';
 742			/* cut off name for saving output style */
 743			pr_info("%20s ", cut_name);
 744		}
 745
 746		pr_info("%10u ", st->nr_acquired);
 747		pr_info("%10u ", st->nr_contended);
 748
 749		pr_info("%15" PRIu64 " ", st->avg_wait_time);
 750		pr_info("%15" PRIu64 " ", st->wait_time_total);
 751		pr_info("%15" PRIu64 " ", st->wait_time_max);
 752		pr_info("%15" PRIu64 " ", st->wait_time_min == ULLONG_MAX ?
 753		       0 : st->wait_time_min);
 754		pr_info("\n");
 755	}
 756
 757	print_bad_events(bad, total);
 758}
 759
 760static bool info_threads, info_map;
 761
 762static void dump_threads(void)
 763{
 764	struct thread_stat *st;
 765	struct rb_node *node;
 766	struct thread *t;
 767
 768	pr_info("%10s: comm\n", "Thread ID");
 769
 770	node = rb_first(&thread_stats);
 771	while (node) {
 772		st = container_of(node, struct thread_stat, rb);
 773		t = perf_session__findnew(session, st->tid);
 774		pr_info("%10d: %s\n", st->tid, thread__comm_str(t));
 775		node = rb_next(node);
 776		thread__put(t);
 777	};
 778}
 779
 780static void dump_map(void)
 781{
 782	unsigned int i;
 783	struct lock_stat *st;
 784
 785	pr_info("Address of instance: name of class\n");
 786	for (i = 0; i < LOCKHASH_SIZE; i++) {
 787		list_for_each_entry(st, &lockhash_table[i], hash_entry) {
 788			pr_info(" %p: %s\n", st->addr, st->name);
 789		}
 790	}
 791}
 792
 793static int dump_info(void)
 794{
 795	int rc = 0;
 796
 797	if (info_threads)
 798		dump_threads();
 799	else if (info_map)
 800		dump_map();
 801	else {
 802		rc = -1;
 803		pr_err("Unknown type of information\n");
 804	}
 805
 806	return rc;
 807}
 808
 809typedef int (*tracepoint_handler)(struct perf_evsel *evsel,
 810				  struct perf_sample *sample);
 811
 812static int process_sample_event(struct perf_tool *tool __maybe_unused,
 813				union perf_event *event,
 814				struct perf_sample *sample,
 815				struct perf_evsel *evsel,
 816				struct machine *machine)
 817{
 818	int err = 0;
 819	struct thread *thread = machine__findnew_thread(machine, sample->pid,
 820							sample->tid);
 821
 822	if (thread == NULL) {
 823		pr_debug("problem processing %d event, skipping it.\n",
 824			event->header.type);
 825		return -1;
 826	}
 827
 828	if (evsel->handler != NULL) {
 829		tracepoint_handler f = evsel->handler;
 830		err = f(evsel, sample);
 831	}
 832
 833	thread__put(thread);
 834
 835	return err;
 836}
 837
 838static void sort_result(void)
 839{
 840	unsigned int i;
 841	struct lock_stat *st;
 842
 843	for (i = 0; i < LOCKHASH_SIZE; i++) {
 844		list_for_each_entry(st, &lockhash_table[i], hash_entry) {
 845			insert_to_result(st, compare);
 846		}
 847	}
 848}
 849
 850static const struct perf_evsel_str_handler lock_tracepoints[] = {
 851	{ "lock:lock_acquire",	 perf_evsel__process_lock_acquire,   }, /* CONFIG_LOCKDEP */
 852	{ "lock:lock_acquired",	 perf_evsel__process_lock_acquired,  }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
 853	{ "lock:lock_contended", perf_evsel__process_lock_contended, }, /* CONFIG_LOCKDEP, CONFIG_LOCK_STAT */
 854	{ "lock:lock_release",	 perf_evsel__process_lock_release,   }, /* CONFIG_LOCKDEP */
 855};
 856
 857static bool force;
 858
 859static int __cmd_report(bool display_info)
 860{
 861	int err = -EINVAL;
 862	struct perf_tool eops = {
 863		.sample		 = process_sample_event,
 864		.comm		 = perf_event__process_comm,
 865		.namespaces	 = perf_event__process_namespaces,
 866		.ordered_events	 = true,
 867	};
 868	struct perf_data data = {
 869		.file      = {
 870			.path = input_name,
 871		},
 872		.mode      = PERF_DATA_MODE_READ,
 873		.force     = force,
 874	};
 875
 876	session = perf_session__new(&data, false, &eops);
 877	if (!session) {
 878		pr_err("Initializing perf session failed\n");
 879		return -1;
 880	}
 881
 882	symbol__init(&session->header.env);
 883
 884	if (!perf_session__has_traces(session, "lock record"))
 885		goto out_delete;
 886
 887	if (perf_session__set_tracepoints_handlers(session, lock_tracepoints)) {
 888		pr_err("Initializing perf session tracepoint handlers failed\n");
 889		goto out_delete;
 890	}
 891
 892	if (select_key())
 893		goto out_delete;
 894
 895	err = perf_session__process_events(session);
 896	if (err)
 897		goto out_delete;
 898
 899	setup_pager();
 900	if (display_info) /* used for info subcommand */
 901		err = dump_info();
 902	else {
 903		sort_result();
 904		print_result();
 905	}
 906
 907out_delete:
 908	perf_session__delete(session);
 909	return err;
 910}
 911
 912static int __cmd_record(int argc, const char **argv)
 913{
 914	const char *record_args[] = {
 915		"record", "-R", "-m", "1024", "-c", "1",
 916	};
 917	unsigned int rec_argc, i, j, ret;
 918	const char **rec_argv;
 919
 920	for (i = 0; i < ARRAY_SIZE(lock_tracepoints); i++) {
 921		if (!is_valid_tracepoint(lock_tracepoints[i].name)) {
 922				pr_err("tracepoint %s is not enabled. "
 923				       "Are CONFIG_LOCKDEP and CONFIG_LOCK_STAT enabled?\n",
 924				       lock_tracepoints[i].name);
 925				return 1;
 926		}
 927	}
 928
 929	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
 930	/* factor of 2 is for -e in front of each tracepoint */
 931	rec_argc += 2 * ARRAY_SIZE(lock_tracepoints);
 932
 933	rec_argv = calloc(rec_argc + 1, sizeof(char *));
 934	if (!rec_argv)
 935		return -ENOMEM;
 936
 937	for (i = 0; i < ARRAY_SIZE(record_args); i++)
 938		rec_argv[i] = strdup(record_args[i]);
 939
 940	for (j = 0; j < ARRAY_SIZE(lock_tracepoints); j++) {
 941		rec_argv[i++] = "-e";
 942		rec_argv[i++] = strdup(lock_tracepoints[j].name);
 943	}
 944
 945	for (j = 1; j < (unsigned int)argc; j++, i++)
 946		rec_argv[i] = argv[j];
 947
 948	BUG_ON(i != rec_argc);
 949
 950	ret = cmd_record(i, rec_argv);
 951	free(rec_argv);
 952	return ret;
 953}
 954
 955int cmd_lock(int argc, const char **argv)
 956{
 
 
 
 
 
 
 
 957	const struct option lock_options[] = {
 958	OPT_STRING('i', "input", &input_name, "file", "input file name"),
 959	OPT_INCR('v', "verbose", &verbose, "be more verbose (show symbol address, etc)"),
 960	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, "dump raw trace in ASCII"),
 961	OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
 962	OPT_END()
 963	};
 964
 965	const struct option info_options[] = {
 966	OPT_BOOLEAN('t', "threads", &info_threads,
 967		    "dump thread list in perf.data"),
 968	OPT_BOOLEAN('m', "map", &info_map,
 969		    "map of lock instances (address:name table)"),
 970	OPT_PARENT(lock_options)
 971	};
 972
 973	const struct option report_options[] = {
 974	OPT_STRING('k', "key", &sort_key, "acquired",
 975		    "key for sorting (acquired / contended / avg_wait / wait_total / wait_max / wait_min)"),
 976	/* TODO: type */
 977	OPT_PARENT(lock_options)
 978	};
 979
 980	const char * const info_usage[] = {
 981		"perf lock info [<options>]",
 982		NULL
 983	};
 984	const char *const lock_subcommands[] = { "record", "report", "script",
 985						 "info", NULL };
 986	const char *lock_usage[] = {
 987		NULL,
 988		NULL
 989	};
 990	const char * const report_usage[] = {
 991		"perf lock report [<options>]",
 992		NULL
 993	};
 994	unsigned int i;
 995	int rc = 0;
 996
 
 997	for (i = 0; i < LOCKHASH_SIZE; i++)
 998		INIT_LIST_HEAD(lockhash_table + i);
 999
1000	argc = parse_options_subcommand(argc, argv, lock_options, lock_subcommands,
1001					lock_usage, PARSE_OPT_STOP_AT_NON_OPTION);
1002	if (!argc)
1003		usage_with_options(lock_usage, lock_options);
1004
1005	if (!strncmp(argv[0], "rec", 3)) {
1006		return __cmd_record(argc, argv);
1007	} else if (!strncmp(argv[0], "report", 6)) {
1008		trace_handler = &report_lock_ops;
1009		if (argc) {
1010			argc = parse_options(argc, argv,
1011					     report_options, report_usage, 0);
1012			if (argc)
1013				usage_with_options(report_usage, report_options);
1014		}
1015		rc = __cmd_report(false);
1016	} else if (!strcmp(argv[0], "script")) {
1017		/* Aliased to 'perf script' */
1018		return cmd_script(argc, argv);
1019	} else if (!strcmp(argv[0], "info")) {
1020		if (argc) {
1021			argc = parse_options(argc, argv,
1022					     info_options, info_usage, 0);
1023			if (argc)
1024				usage_with_options(info_usage, info_options);
1025		}
1026		/* recycling report_lock_ops */
1027		trace_handler = &report_lock_ops;
1028		rc = __cmd_report(true);
1029	} else {
1030		usage_with_options(lock_usage, lock_options);
1031	}
1032
1033	return rc;
1034}