Home | History | Annotate | Line # | Download | only in kern
subr_lockdebug.c revision 1.41.4.1
      1  1.41.4.1    rmind /*	$NetBSD: subr_lockdebug.c,v 1.41.4.1 2011/03/05 20:55:18 rmind Exp $	*/
      2       1.2       ad 
      3       1.2       ad /*-
      4      1.28       ad  * Copyright (c) 2006, 2007, 2008 The NetBSD Foundation, Inc.
      5       1.2       ad  * All rights reserved.
      6       1.2       ad  *
      7       1.2       ad  * This code is derived from software contributed to The NetBSD Foundation
      8       1.2       ad  * by Andrew Doran.
      9       1.2       ad  *
     10       1.2       ad  * Redistribution and use in source and binary forms, with or without
     11       1.2       ad  * modification, are permitted provided that the following conditions
     12       1.2       ad  * are met:
     13       1.2       ad  * 1. Redistributions of source code must retain the above copyright
     14       1.2       ad  *    notice, this list of conditions and the following disclaimer.
     15       1.2       ad  * 2. Redistributions in binary form must reproduce the above copyright
     16       1.2       ad  *    notice, this list of conditions and the following disclaimer in the
     17       1.2       ad  *    documentation and/or other materials provided with the distribution.
     18       1.2       ad  *
     19       1.2       ad  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     20       1.2       ad  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     21       1.2       ad  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     22       1.2       ad  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     23       1.2       ad  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     24       1.2       ad  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     25       1.2       ad  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     26       1.2       ad  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     27       1.2       ad  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     28       1.2       ad  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     29       1.2       ad  * POSSIBILITY OF SUCH DAMAGE.
     30       1.2       ad  */
     31       1.2       ad 
     32       1.2       ad /*
     33      1.11       ad  * Basic lock debugging code shared among lock primitives.
     34       1.2       ad  */
     35       1.2       ad 
     36       1.9      dsl #include <sys/cdefs.h>
     37  1.41.4.1    rmind __KERNEL_RCSID(0, "$NetBSD: subr_lockdebug.c,v 1.41.4.1 2011/03/05 20:55:18 rmind Exp $");
     38       1.9      dsl 
     39       1.2       ad #include "opt_ddb.h"
     40       1.2       ad 
     41       1.2       ad #include <sys/param.h>
     42       1.2       ad #include <sys/proc.h>
     43       1.2       ad #include <sys/systm.h>
     44      1.10       ad #include <sys/kernel.h>
     45       1.2       ad #include <sys/kmem.h>
     46       1.2       ad #include <sys/lockdebug.h>
     47       1.2       ad #include <sys/sleepq.h>
     48      1.10       ad #include <sys/cpu.h>
     49      1.22       ad #include <sys/atomic.h>
     50      1.26       ad #include <sys/lock.h>
     51  1.41.4.1    rmind #include <sys/rbtree.h>
     52      1.16     yamt 
     53      1.25       ad #include <machine/lock.h>
     54      1.25       ad 
     55      1.28       ad unsigned int		ld_panic;
     56      1.28       ad 
     57       1.2       ad #ifdef LOCKDEBUG
     58       1.2       ad 
     59       1.2       ad #define	LD_BATCH_SHIFT	9
     60       1.2       ad #define	LD_BATCH	(1 << LD_BATCH_SHIFT)
     61       1.2       ad #define	LD_BATCH_MASK	(LD_BATCH - 1)
     62       1.2       ad #define	LD_MAX_LOCKS	1048576
     63       1.2       ad #define	LD_SLOP		16
     64       1.2       ad 
     65       1.2       ad #define	LD_LOCKED	0x01
     66       1.2       ad #define	LD_SLEEPER	0x02
     67       1.2       ad 
     68      1.23       ad #define	LD_WRITE_LOCK	0x80000000
     69      1.23       ad 
     70       1.2       ad typedef struct lockdebug {
     71  1.41.4.1    rmind 	struct rb_node	ld_rb_node;
     72      1.34       ad 	__cpu_simple_lock_t ld_spinlock;
     73       1.2       ad 	_TAILQ_ENTRY(struct lockdebug, volatile) ld_chain;
     74       1.2       ad 	_TAILQ_ENTRY(struct lockdebug, volatile) ld_achain;
     75       1.2       ad 	volatile void	*ld_lock;
     76       1.2       ad 	lockops_t	*ld_lockops;
     77       1.2       ad 	struct lwp	*ld_lwp;
     78       1.2       ad 	uintptr_t	ld_locked;
     79       1.2       ad 	uintptr_t	ld_unlocked;
     80      1.10       ad 	uintptr_t	ld_initaddr;
     81       1.2       ad 	uint16_t	ld_shares;
     82       1.2       ad 	uint16_t	ld_cpu;
     83       1.2       ad 	uint8_t		ld_flags;
     84       1.2       ad 	uint8_t		ld_shwant;	/* advisory */
     85       1.2       ad 	uint8_t		ld_exwant;	/* advisory */
     86       1.2       ad 	uint8_t		ld_unused;
     87       1.2       ad } volatile lockdebug_t;
     88       1.2       ad 
     89       1.2       ad typedef _TAILQ_HEAD(lockdebuglist, struct lockdebug, volatile) lockdebuglist_t;
     90       1.2       ad 
     91      1.34       ad __cpu_simple_lock_t	ld_mod_lk;
     92      1.13     matt lockdebuglist_t		ld_free = TAILQ_HEAD_INITIALIZER(ld_free);
     93      1.13     matt lockdebuglist_t		ld_all = TAILQ_HEAD_INITIALIZER(ld_all);
     94       1.2       ad int			ld_nfree;
     95       1.2       ad int			ld_freeptr;
     96       1.2       ad int			ld_recurse;
     97       1.5       ad bool			ld_nomore;
     98       1.2       ad lockdebug_t		ld_prime[LD_BATCH];
     99       1.2       ad 
    100      1.34       ad static void	lockdebug_abort1(lockdebug_t *, int, const char *,
    101      1.34       ad 				 const char *, bool);
    102      1.34       ad static int	lockdebug_more(int);
    103       1.5       ad static void	lockdebug_init(void);
    104       1.2       ad 
    105      1.16     yamt static signed int
    106  1.41.4.1    rmind ld_rbto_compare_nodes(void *ctx, const void *n1, const void *n2)
    107      1.16     yamt {
    108  1.41.4.1    rmind 	const lockdebug_t *ld1 = n1;
    109  1.41.4.1    rmind 	const lockdebug_t *ld2 = n2;
    110      1.20     yamt 	const uintptr_t a = (uintptr_t)ld1->ld_lock;
    111      1.20     yamt 	const uintptr_t b = (uintptr_t)ld2->ld_lock;
    112      1.20     yamt 
    113      1.20     yamt 	if (a < b)
    114      1.16     yamt 		return -1;
    115  1.41.4.1    rmind 	if (a > b)
    116  1.41.4.1    rmind 		return 1;
    117      1.16     yamt 	return 0;
    118      1.16     yamt }
    119      1.16     yamt 
    120      1.16     yamt static signed int
    121  1.41.4.1    rmind ld_rbto_compare_key(void *ctx, const void *n, const void *key)
    122      1.16     yamt {
    123  1.41.4.1    rmind 	const lockdebug_t *ld = n;
    124      1.20     yamt 	const uintptr_t a = (uintptr_t)ld->ld_lock;
    125      1.20     yamt 	const uintptr_t b = (uintptr_t)key;
    126      1.20     yamt 
    127      1.20     yamt 	if (a < b)
    128      1.16     yamt 		return -1;
    129  1.41.4.1    rmind 	if (a > b)
    130  1.41.4.1    rmind 		return 1;
    131      1.16     yamt 	return 0;
    132      1.16     yamt }
    133      1.16     yamt 
    134  1.41.4.1    rmind static rb_tree_t ld_rb_tree;
    135      1.16     yamt 
    136  1.41.4.1    rmind static const rb_tree_ops_t ld_rb_tree_ops = {
    137      1.37     matt 	.rbto_compare_nodes = ld_rbto_compare_nodes,
    138      1.37     matt 	.rbto_compare_key = ld_rbto_compare_key,
    139  1.41.4.1    rmind 	.rbto_node_offset = offsetof(lockdebug_t, ld_rb_node),
    140  1.41.4.1    rmind 	.rbto_context = NULL
    141      1.16     yamt };
    142      1.16     yamt 
    143      1.34       ad static inline lockdebug_t *
    144      1.34       ad lockdebug_lookup1(volatile void *lock)
    145      1.23       ad {
    146      1.34       ad 	lockdebug_t *ld;
    147      1.34       ad 	struct cpu_info *ci;
    148      1.23       ad 
    149      1.34       ad 	ci = curcpu();
    150      1.34       ad 	__cpu_simple_lock(&ci->ci_data.cpu_ld_lock);
    151      1.34       ad 	ld = (lockdebug_t *)rb_tree_find_node(&ld_rb_tree, __UNVOLATILE(lock));
    152      1.34       ad 	__cpu_simple_unlock(&ci->ci_data.cpu_ld_lock);
    153      1.34       ad 	if (ld == NULL) {
    154      1.34       ad 		return NULL;
    155      1.34       ad 	}
    156      1.34       ad 	__cpu_simple_lock(&ld->ld_spinlock);
    157      1.23       ad 
    158      1.34       ad 	return ld;
    159       1.2       ad }
    160       1.2       ad 
    161      1.23       ad static void
    162      1.34       ad lockdebug_lock_cpus(void)
    163       1.2       ad {
    164      1.34       ad 	CPU_INFO_ITERATOR cii;
    165      1.34       ad 	struct cpu_info *ci;
    166       1.2       ad 
    167      1.34       ad 	for (CPU_INFO_FOREACH(cii, ci)) {
    168      1.34       ad 		__cpu_simple_lock(&ci->ci_data.cpu_ld_lock);
    169      1.34       ad 	}
    170      1.23       ad }
    171      1.23       ad 
    172      1.23       ad static void
    173      1.34       ad lockdebug_unlock_cpus(void)
    174      1.23       ad {
    175      1.34       ad 	CPU_INFO_ITERATOR cii;
    176      1.34       ad 	struct cpu_info *ci;
    177      1.23       ad 
    178      1.34       ad 	for (CPU_INFO_FOREACH(cii, ci)) {
    179      1.34       ad 		__cpu_simple_unlock(&ci->ci_data.cpu_ld_lock);
    180      1.34       ad 	}
    181       1.2       ad }
    182       1.2       ad 
    183       1.2       ad /*
    184      1.19     yamt  * lockdebug_lookup:
    185      1.19     yamt  *
    186      1.19     yamt  *	Find a lockdebug structure by a pointer to a lock and return it locked.
    187      1.19     yamt  */
    188      1.19     yamt static inline lockdebug_t *
    189      1.38    rafal lockdebug_lookup(volatile void *lock, uintptr_t where)
    190      1.19     yamt {
    191      1.19     yamt 	lockdebug_t *ld;
    192      1.19     yamt 
    193      1.34       ad 	ld = lockdebug_lookup1(lock);
    194  1.41.4.1    rmind 	if (ld == NULL) {
    195  1.41.4.1    rmind 		panic("lockdebug_lookup: uninitialized lock "
    196  1.41.4.1    rmind 		    "(lock=%p, from=%08"PRIxPTR")", lock, where);
    197  1.41.4.1    rmind 	}
    198      1.19     yamt 	return ld;
    199      1.19     yamt }
    200      1.19     yamt 
    201      1.19     yamt /*
    202       1.2       ad  * lockdebug_init:
    203       1.2       ad  *
    204       1.2       ad  *	Initialize the lockdebug system.  Allocate an initial pool of
    205       1.2       ad  *	lockdebug structures before the VM system is up and running.
    206       1.2       ad  */
    207       1.5       ad static void
    208       1.2       ad lockdebug_init(void)
    209       1.2       ad {
    210       1.2       ad 	lockdebug_t *ld;
    211       1.2       ad 	int i;
    212       1.2       ad 
    213      1.34       ad 	TAILQ_INIT(&curcpu()->ci_data.cpu_ld_locks);
    214      1.34       ad 	TAILQ_INIT(&curlwp->l_ld_locks);
    215      1.34       ad 	__cpu_simple_lock_init(&curcpu()->ci_data.cpu_ld_lock);
    216      1.34       ad 	__cpu_simple_lock_init(&ld_mod_lk);
    217      1.15     matt 
    218      1.16     yamt 	rb_tree_init(&ld_rb_tree, &ld_rb_tree_ops);
    219      1.16     yamt 
    220       1.2       ad 	ld = ld_prime;
    221       1.2       ad 	for (i = 1, ld++; i < LD_BATCH; i++, ld++) {
    222      1.34       ad 		__cpu_simple_lock_init(&ld->ld_spinlock);
    223       1.2       ad 		TAILQ_INSERT_TAIL(&ld_free, ld, ld_chain);
    224       1.2       ad 		TAILQ_INSERT_TAIL(&ld_all, ld, ld_achain);
    225       1.2       ad 	}
    226       1.2       ad 	ld_freeptr = 1;
    227       1.2       ad 	ld_nfree = LD_BATCH - 1;
    228       1.2       ad }
    229       1.2       ad 
    230       1.2       ad /*
    231       1.2       ad  * lockdebug_alloc:
    232       1.2       ad  *
    233       1.2       ad  *	A lock is being initialized, so allocate an associated debug
    234       1.2       ad  *	structure.
    235       1.2       ad  */
    236      1.16     yamt bool
    237      1.10       ad lockdebug_alloc(volatile void *lock, lockops_t *lo, uintptr_t initaddr)
    238       1.2       ad {
    239       1.2       ad 	struct cpu_info *ci;
    240       1.2       ad 	lockdebug_t *ld;
    241      1.34       ad 	int s;
    242       1.2       ad 
    243      1.27       ad 	if (lo == NULL || panicstr != NULL || ld_panic)
    244      1.16     yamt 		return false;
    245       1.5       ad 	if (ld_freeptr == 0)
    246       1.5       ad 		lockdebug_init();
    247       1.2       ad 
    248      1.34       ad 	s = splhigh();
    249      1.34       ad 	__cpu_simple_lock(&ld_mod_lk);
    250      1.34       ad 	if ((ld = lockdebug_lookup1(lock)) != NULL) {
    251      1.34       ad 		__cpu_simple_unlock(&ld_mod_lk);
    252      1.34       ad 		lockdebug_abort1(ld, s, __func__, "already initialized", true);
    253      1.27       ad 		return false;
    254      1.19     yamt 	}
    255      1.19     yamt 
    256       1.2       ad 	/*
    257       1.2       ad 	 * Pinch a new debug structure.  We may recurse because we call
    258       1.2       ad 	 * kmem_alloc(), which may need to initialize new locks somewhere
    259       1.7    skrll 	 * down the path.  If not recursing, we try to maintain at least
    260       1.2       ad 	 * LD_SLOP structures free, which should hopefully be enough to
    261       1.2       ad 	 * satisfy kmem_alloc().  If we can't provide a structure, not to
    262       1.2       ad 	 * worry: we'll just mark the lock as not having an ID.
    263       1.2       ad 	 */
    264      1.23       ad 	ci = curcpu();
    265       1.2       ad 	ci->ci_lkdebug_recurse++;
    266       1.2       ad 	if (TAILQ_EMPTY(&ld_free)) {
    267       1.5       ad 		if (ci->ci_lkdebug_recurse > 1 || ld_nomore) {
    268       1.2       ad 			ci->ci_lkdebug_recurse--;
    269      1.34       ad 			__cpu_simple_unlock(&ld_mod_lk);
    270      1.34       ad 			splx(s);
    271      1.16     yamt 			return false;
    272       1.2       ad 		}
    273      1.34       ad 		s = lockdebug_more(s);
    274      1.34       ad 	} else if (ci->ci_lkdebug_recurse == 1 && ld_nfree < LD_SLOP) {
    275      1.34       ad 		s = lockdebug_more(s);
    276      1.34       ad 	}
    277       1.2       ad 	if ((ld = TAILQ_FIRST(&ld_free)) == NULL) {
    278      1.34       ad 		__cpu_simple_unlock(&ld_mod_lk);
    279      1.34       ad 		splx(s);
    280      1.16     yamt 		return false;
    281       1.2       ad 	}
    282       1.2       ad 	TAILQ_REMOVE(&ld_free, ld, ld_chain);
    283       1.2       ad 	ld_nfree--;
    284       1.2       ad 	ci->ci_lkdebug_recurse--;
    285       1.2       ad 
    286      1.34       ad 	if (ld->ld_lock != NULL) {
    287       1.2       ad 		panic("lockdebug_alloc: corrupt table");
    288      1.34       ad 	}
    289       1.2       ad 
    290       1.2       ad 	/* Initialise the structure. */
    291       1.2       ad 	ld->ld_lock = lock;
    292       1.2       ad 	ld->ld_lockops = lo;
    293       1.2       ad 	ld->ld_locked = 0;
    294       1.2       ad 	ld->ld_unlocked = 0;
    295       1.2       ad 	ld->ld_lwp = NULL;
    296      1.10       ad 	ld->ld_initaddr = initaddr;
    297      1.35       ad 	ld->ld_flags = (lo->lo_type == LOCKOPS_SLEEP ? LD_SLEEPER : 0);
    298      1.34       ad 	lockdebug_lock_cpus();
    299  1.41.4.1    rmind 	(void)rb_tree_insert_node(&ld_rb_tree, __UNVOLATILE(ld));
    300      1.34       ad 	lockdebug_unlock_cpus();
    301      1.34       ad 	__cpu_simple_unlock(&ld_mod_lk);
    302       1.2       ad 
    303      1.34       ad 	splx(s);
    304      1.16     yamt 	return true;
    305       1.2       ad }
    306       1.2       ad 
    307       1.2       ad /*
    308       1.2       ad  * lockdebug_free:
    309       1.2       ad  *
    310       1.2       ad  *	A lock is being destroyed, so release debugging resources.
    311       1.2       ad  */
    312       1.2       ad void
    313      1.16     yamt lockdebug_free(volatile void *lock)
    314       1.2       ad {
    315       1.2       ad 	lockdebug_t *ld;
    316      1.34       ad 	int s;
    317       1.2       ad 
    318      1.27       ad 	if (panicstr != NULL || ld_panic)
    319       1.2       ad 		return;
    320       1.2       ad 
    321      1.34       ad 	s = splhigh();
    322      1.34       ad 	__cpu_simple_lock(&ld_mod_lk);
    323      1.38    rafal 	ld = lockdebug_lookup(lock, (uintptr_t) __builtin_return_address(0));
    324      1.16     yamt 	if (ld == NULL) {
    325      1.34       ad 		__cpu_simple_unlock(&ld_mod_lk);
    326      1.35       ad 		panic("lockdebug_free: destroying uninitialized object %p"
    327      1.16     yamt 		    "(ld_lock=%p)", lock, ld->ld_lock);
    328      1.35       ad 		lockdebug_abort1(ld, s, __func__, "record follows", true);
    329      1.27       ad 		return;
    330       1.2       ad 	}
    331      1.27       ad 	if ((ld->ld_flags & LD_LOCKED) != 0 || ld->ld_shares != 0) {
    332      1.34       ad 		__cpu_simple_unlock(&ld_mod_lk);
    333      1.35       ad 		lockdebug_abort1(ld, s, __func__, "is locked or in use", true);
    334      1.27       ad 		return;
    335      1.27       ad 	}
    336      1.34       ad 	lockdebug_lock_cpus();
    337  1.41.4.1    rmind 	rb_tree_remove_node(&ld_rb_tree, __UNVOLATILE(ld));
    338      1.34       ad 	lockdebug_unlock_cpus();
    339       1.2       ad 	ld->ld_lock = NULL;
    340       1.2       ad 	TAILQ_INSERT_TAIL(&ld_free, ld, ld_chain);
    341       1.2       ad 	ld_nfree++;
    342      1.34       ad 	__cpu_simple_unlock(&ld->ld_spinlock);
    343      1.34       ad 	__cpu_simple_unlock(&ld_mod_lk);
    344      1.34       ad 	splx(s);
    345       1.2       ad }
    346       1.2       ad 
    347       1.2       ad /*
    348       1.2       ad  * lockdebug_more:
    349       1.2       ad  *
    350       1.2       ad  *	Allocate a batch of debug structures and add to the free list.
    351      1.34       ad  *	Must be called with ld_mod_lk held.
    352       1.2       ad  */
    353      1.34       ad static int
    354      1.34       ad lockdebug_more(int s)
    355       1.2       ad {
    356       1.2       ad 	lockdebug_t *ld;
    357       1.2       ad 	void *block;
    358       1.5       ad 	int i, base, m;
    359       1.2       ad 
    360      1.35       ad 	/*
    361      1.35       ad 	 * Can't call kmem_alloc() if in interrupt context.  XXX We could
    362      1.35       ad 	 * deadlock, because we don't know which locks the caller holds.
    363      1.35       ad 	 */
    364      1.35       ad 	if (cpu_intr_p() || (curlwp->l_pflag & LP_INTR) != 0) {
    365      1.35       ad 		return s;
    366      1.35       ad 	}
    367      1.35       ad 
    368       1.2       ad 	while (ld_nfree < LD_SLOP) {
    369      1.34       ad 		__cpu_simple_unlock(&ld_mod_lk);
    370      1.34       ad 		splx(s);
    371       1.2       ad 		block = kmem_zalloc(LD_BATCH * sizeof(lockdebug_t), KM_SLEEP);
    372      1.34       ad 		s = splhigh();
    373      1.34       ad 		__cpu_simple_lock(&ld_mod_lk);
    374       1.2       ad 
    375       1.2       ad 		if (block == NULL)
    376      1.34       ad 			return s;
    377       1.2       ad 
    378       1.2       ad 		if (ld_nfree > LD_SLOP) {
    379       1.2       ad 			/* Somebody beat us to it. */
    380      1.34       ad 			__cpu_simple_unlock(&ld_mod_lk);
    381      1.34       ad 			splx(s);
    382       1.2       ad 			kmem_free(block, LD_BATCH * sizeof(lockdebug_t));
    383      1.34       ad 			s = splhigh();
    384      1.34       ad 			__cpu_simple_lock(&ld_mod_lk);
    385       1.2       ad 			continue;
    386       1.2       ad 		}
    387       1.2       ad 
    388       1.2       ad 		base = ld_freeptr;
    389       1.2       ad 		ld_nfree += LD_BATCH;
    390       1.2       ad 		ld = block;
    391       1.2       ad 		base <<= LD_BATCH_SHIFT;
    392       1.5       ad 		m = min(LD_MAX_LOCKS, base + LD_BATCH);
    393       1.5       ad 
    394       1.5       ad 		if (m == LD_MAX_LOCKS)
    395       1.5       ad 			ld_nomore = true;
    396       1.2       ad 
    397       1.5       ad 		for (i = base; i < m; i++, ld++) {
    398      1.34       ad 			__cpu_simple_lock_init(&ld->ld_spinlock);
    399       1.2       ad 			TAILQ_INSERT_TAIL(&ld_free, ld, ld_chain);
    400       1.2       ad 			TAILQ_INSERT_TAIL(&ld_all, ld, ld_achain);
    401       1.2       ad 		}
    402       1.2       ad 
    403      1.22       ad 		membar_producer();
    404       1.2       ad 	}
    405      1.34       ad 
    406      1.34       ad 	return s;
    407       1.2       ad }
    408       1.2       ad 
    409       1.2       ad /*
    410       1.2       ad  * lockdebug_wantlock:
    411       1.2       ad  *
    412       1.2       ad  *	Process the preamble to a lock acquire.
    413       1.2       ad  */
    414       1.2       ad void
    415      1.33       ad lockdebug_wantlock(volatile void *lock, uintptr_t where, bool shared,
    416      1.33       ad 		   bool trylock)
    417       1.2       ad {
    418       1.2       ad 	struct lwp *l = curlwp;
    419       1.2       ad 	lockdebug_t *ld;
    420       1.3  thorpej 	bool recurse;
    421      1.34       ad 	int s;
    422       1.2       ad 
    423       1.2       ad 	(void)shared;
    424       1.4  thorpej 	recurse = false;
    425       1.2       ad 
    426      1.27       ad 	if (panicstr != NULL || ld_panic)
    427       1.2       ad 		return;
    428       1.2       ad 
    429      1.34       ad 	s = splhigh();
    430      1.38    rafal 	if ((ld = lockdebug_lookup(lock, where)) == NULL) {
    431      1.34       ad 		splx(s);
    432       1.2       ad 		return;
    433      1.34       ad 	}
    434      1.32     yamt 	if ((ld->ld_flags & LD_LOCKED) != 0 || ld->ld_shares != 0) {
    435       1.2       ad 		if ((ld->ld_flags & LD_SLEEPER) != 0) {
    436      1.33       ad 			if (ld->ld_lwp == l && !(shared && trylock))
    437       1.4  thorpej 				recurse = true;
    438      1.40    rmind 		} else if (ld->ld_cpu == (uint16_t)cpu_index(curcpu()))
    439       1.4  thorpej 			recurse = true;
    440       1.2       ad 	}
    441      1.10       ad 	if (cpu_intr_p()) {
    442      1.27       ad 		if ((ld->ld_flags & LD_SLEEPER) != 0) {
    443      1.34       ad 			lockdebug_abort1(ld, s, __func__,
    444      1.10       ad 			    "acquiring sleep lock from interrupt context",
    445      1.10       ad 			    true);
    446      1.27       ad 			return;
    447      1.27       ad 		}
    448      1.10       ad 	}
    449       1.2       ad 	if (shared)
    450       1.2       ad 		ld->ld_shwant++;
    451       1.2       ad 	else
    452       1.2       ad 		ld->ld_exwant++;
    453      1.27       ad 	if (recurse) {
    454      1.34       ad 		lockdebug_abort1(ld, s, __func__, "locking against myself",
    455      1.10       ad 		    true);
    456      1.27       ad 		return;
    457      1.27       ad 	}
    458      1.34       ad 	__cpu_simple_unlock(&ld->ld_spinlock);
    459      1.34       ad 	splx(s);
    460       1.2       ad }
    461       1.2       ad 
    462       1.2       ad /*
    463       1.2       ad  * lockdebug_locked:
    464       1.2       ad  *
    465       1.2       ad  *	Process a lock acquire operation.
    466       1.2       ad  */
    467       1.2       ad void
    468      1.35       ad lockdebug_locked(volatile void *lock, void *cvlock, uintptr_t where,
    469      1.35       ad 		 int shared)
    470       1.2       ad {
    471       1.2       ad 	struct lwp *l = curlwp;
    472       1.2       ad 	lockdebug_t *ld;
    473      1.34       ad 	int s;
    474       1.2       ad 
    475      1.27       ad 	if (panicstr != NULL || ld_panic)
    476       1.2       ad 		return;
    477       1.2       ad 
    478      1.34       ad 	s = splhigh();
    479      1.38    rafal 	if ((ld = lockdebug_lookup(lock, where)) == NULL) {
    480      1.34       ad 		splx(s);
    481       1.2       ad 		return;
    482      1.34       ad 	}
    483      1.35       ad 	if (cvlock) {
    484      1.35       ad 		KASSERT(ld->ld_lockops->lo_type == LOCKOPS_CV);
    485      1.35       ad 		if (lock == (void *)&lbolt) {
    486      1.35       ad 			/* nothing */
    487      1.35       ad 		} else if (ld->ld_shares++ == 0) {
    488      1.35       ad 			ld->ld_locked = (uintptr_t)cvlock;
    489      1.35       ad 		} else if (cvlock != (void *)ld->ld_locked) {
    490      1.35       ad 			lockdebug_abort1(ld, s, __func__, "multiple locks used"
    491      1.35       ad 			    " with condition variable", true);
    492      1.35       ad 			return;
    493      1.35       ad 		}
    494      1.35       ad 	} else if (shared) {
    495       1.2       ad 		l->l_shlocks++;
    496       1.2       ad 		ld->ld_shares++;
    497       1.2       ad 		ld->ld_shwant--;
    498       1.2       ad 	} else {
    499      1.27       ad 		if ((ld->ld_flags & LD_LOCKED) != 0) {
    500      1.34       ad 			lockdebug_abort1(ld, s, __func__, "already locked",
    501      1.34       ad 			    true);
    502      1.27       ad 			return;
    503      1.27       ad 		}
    504       1.2       ad 		ld->ld_flags |= LD_LOCKED;
    505       1.2       ad 		ld->ld_locked = where;
    506       1.2       ad 		ld->ld_exwant--;
    507       1.2       ad 		if ((ld->ld_flags & LD_SLEEPER) != 0) {
    508      1.34       ad 			TAILQ_INSERT_TAIL(&l->l_ld_locks, ld, ld_chain);
    509       1.2       ad 		} else {
    510      1.34       ad 			TAILQ_INSERT_TAIL(&curcpu()->ci_data.cpu_ld_locks,
    511      1.34       ad 			    ld, ld_chain);
    512       1.2       ad 		}
    513       1.2       ad 	}
    514      1.40    rmind 	ld->ld_cpu = (uint16_t)cpu_index(curcpu());
    515      1.32     yamt 	ld->ld_lwp = l;
    516      1.34       ad 	__cpu_simple_unlock(&ld->ld_spinlock);
    517      1.34       ad 	splx(s);
    518       1.2       ad }
    519       1.2       ad 
    520       1.2       ad /*
    521       1.2       ad  * lockdebug_unlocked:
    522       1.2       ad  *
    523       1.2       ad  *	Process a lock release operation.
    524       1.2       ad  */
    525       1.2       ad void
    526      1.16     yamt lockdebug_unlocked(volatile void *lock, uintptr_t where, int shared)
    527       1.2       ad {
    528       1.2       ad 	struct lwp *l = curlwp;
    529       1.2       ad 	lockdebug_t *ld;
    530      1.34       ad 	int s;
    531       1.2       ad 
    532      1.27       ad 	if (panicstr != NULL || ld_panic)
    533       1.2       ad 		return;
    534       1.2       ad 
    535      1.34       ad 	s = splhigh();
    536      1.38    rafal 	if ((ld = lockdebug_lookup(lock, where)) == NULL) {
    537      1.34       ad 		splx(s);
    538       1.2       ad 		return;
    539      1.34       ad 	}
    540      1.35       ad 	if (ld->ld_lockops->lo_type == LOCKOPS_CV) {
    541      1.35       ad 		if (lock == (void *)&lbolt) {
    542      1.35       ad 			/* nothing */
    543      1.35       ad 		} else {
    544      1.35       ad 			ld->ld_shares--;
    545      1.35       ad 		}
    546      1.35       ad 	} else if (shared) {
    547      1.27       ad 		if (l->l_shlocks == 0) {
    548      1.34       ad 			lockdebug_abort1(ld, s, __func__,
    549      1.10       ad 			    "no shared locks held by LWP", true);
    550      1.27       ad 			return;
    551      1.27       ad 		}
    552      1.27       ad 		if (ld->ld_shares == 0) {
    553      1.34       ad 			lockdebug_abort1(ld, s, __func__,
    554      1.10       ad 			    "no shared holds on this lock", true);
    555      1.27       ad 			return;
    556      1.27       ad 		}
    557       1.2       ad 		l->l_shlocks--;
    558       1.2       ad 		ld->ld_shares--;
    559      1.32     yamt 		if (ld->ld_lwp == l)
    560      1.32     yamt 			ld->ld_lwp = NULL;
    561      1.40    rmind 		if (ld->ld_cpu == (uint16_t)cpu_index(curcpu()))
    562      1.32     yamt 			ld->ld_cpu = (uint16_t)-1;
    563       1.2       ad 	} else {
    564      1.27       ad 		if ((ld->ld_flags & LD_LOCKED) == 0) {
    565      1.34       ad 			lockdebug_abort1(ld, s, __func__, "not locked", true);
    566      1.27       ad 			return;
    567      1.27       ad 		}
    568       1.2       ad 
    569       1.2       ad 		if ((ld->ld_flags & LD_SLEEPER) != 0) {
    570      1.27       ad 			if (ld->ld_lwp != curlwp) {
    571      1.34       ad 				lockdebug_abort1(ld, s, __func__,
    572      1.10       ad 				    "not held by current LWP", true);
    573      1.27       ad 				return;
    574      1.27       ad 			}
    575       1.2       ad 			ld->ld_flags &= ~LD_LOCKED;
    576       1.2       ad 			ld->ld_unlocked = where;
    577       1.2       ad 			ld->ld_lwp = NULL;
    578      1.34       ad 			TAILQ_REMOVE(&l->l_ld_locks, ld, ld_chain);
    579       1.2       ad 		} else {
    580      1.40    rmind 			if (ld->ld_cpu != (uint16_t)cpu_index(curcpu())) {
    581      1.34       ad 				lockdebug_abort1(ld, s, __func__,
    582      1.10       ad 				    "not held by current CPU", true);
    583      1.27       ad 				return;
    584      1.27       ad 			}
    585       1.2       ad 			ld->ld_flags &= ~LD_LOCKED;
    586       1.2       ad 			ld->ld_unlocked = where;
    587       1.2       ad 			ld->ld_lwp = NULL;
    588      1.34       ad 			TAILQ_REMOVE(&curcpu()->ci_data.cpu_ld_locks, ld,
    589      1.34       ad 			    ld_chain);
    590       1.2       ad 		}
    591       1.2       ad 	}
    592      1.34       ad 	__cpu_simple_unlock(&ld->ld_spinlock);
    593      1.34       ad 	splx(s);
    594       1.2       ad }
    595       1.2       ad 
    596       1.2       ad /*
    597      1.35       ad  * lockdebug_wakeup:
    598      1.35       ad  *
    599      1.35       ad  *	Process a wakeup on a condition variable.
    600      1.35       ad  */
    601      1.35       ad void
    602      1.35       ad lockdebug_wakeup(volatile void *lock, uintptr_t where)
    603      1.35       ad {
    604      1.35       ad 	lockdebug_t *ld;
    605      1.35       ad 	int s;
    606      1.35       ad 
    607      1.35       ad 	if (panicstr != NULL || ld_panic || lock == (void *)&lbolt)
    608      1.35       ad 		return;
    609      1.35       ad 
    610      1.35       ad 	s = splhigh();
    611      1.35       ad 	/* Find the CV... */
    612      1.38    rafal 	if ((ld = lockdebug_lookup(lock, where)) == NULL) {
    613      1.35       ad 		splx(s);
    614      1.35       ad 		return;
    615      1.35       ad 	}
    616      1.35       ad 	/*
    617      1.35       ad 	 * If it has any waiters, ensure that they are using the
    618      1.35       ad 	 * same interlock.
    619      1.35       ad 	 */
    620      1.35       ad 	if (ld->ld_shares != 0 && !mutex_owned((kmutex_t *)ld->ld_locked)) {
    621      1.35       ad 		lockdebug_abort1(ld, s, __func__, "interlocking mutex not "
    622      1.35       ad 		    "held during wakeup", true);
    623      1.35       ad 		return;
    624      1.35       ad 	}
    625      1.35       ad 	__cpu_simple_unlock(&ld->ld_spinlock);
    626      1.35       ad 	splx(s);
    627      1.35       ad }
    628      1.35       ad 
    629      1.35       ad /*
    630       1.2       ad  * lockdebug_barrier:
    631       1.2       ad  *
    632       1.2       ad  *	Panic if we hold more than one specified spin lock, and optionally,
    633       1.2       ad  *	if we hold sleep locks.
    634       1.2       ad  */
    635       1.2       ad void
    636       1.2       ad lockdebug_barrier(volatile void *spinlock, int slplocks)
    637       1.2       ad {
    638       1.2       ad 	struct lwp *l = curlwp;
    639       1.2       ad 	lockdebug_t *ld;
    640      1.34       ad 	int s;
    641       1.2       ad 
    642      1.27       ad 	if (panicstr != NULL || ld_panic)
    643       1.2       ad 		return;
    644       1.2       ad 
    645      1.34       ad 	s = splhigh();
    646      1.34       ad 	if ((l->l_pflag & LP_INTR) == 0) {
    647      1.34       ad 		TAILQ_FOREACH(ld, &curcpu()->ci_data.cpu_ld_locks, ld_chain) {
    648       1.2       ad 			if (ld->ld_lock == spinlock) {
    649       1.2       ad 				continue;
    650       1.2       ad 			}
    651      1.34       ad 			__cpu_simple_lock(&ld->ld_spinlock);
    652      1.34       ad 			lockdebug_abort1(ld, s, __func__,
    653      1.34       ad 			    "spin lock held", true);
    654      1.34       ad 			return;
    655       1.2       ad 		}
    656       1.2       ad 	}
    657      1.34       ad 	if (slplocks) {
    658      1.34       ad 		splx(s);
    659      1.34       ad 		return;
    660      1.34       ad 	}
    661      1.34       ad 	if ((ld = TAILQ_FIRST(&l->l_ld_locks)) != NULL) {
    662      1.34       ad 		__cpu_simple_lock(&ld->ld_spinlock);
    663      1.34       ad 		lockdebug_abort1(ld, s, __func__, "sleep lock held", true);
    664      1.34       ad 		return;
    665      1.34       ad 	}
    666      1.34       ad 	splx(s);
    667      1.34       ad 	if (l->l_shlocks != 0) {
    668      1.34       ad 		panic("lockdebug_barrier: holding %d shared locks",
    669      1.34       ad 		    l->l_shlocks);
    670       1.2       ad 	}
    671       1.2       ad }
    672       1.2       ad 
    673       1.2       ad /*
    674      1.10       ad  * lockdebug_mem_check:
    675      1.10       ad  *
    676      1.10       ad  *	Check for in-use locks within a memory region that is
    677      1.16     yamt  *	being freed.
    678      1.10       ad  */
    679      1.10       ad void
    680      1.10       ad lockdebug_mem_check(const char *func, void *base, size_t sz)
    681      1.10       ad {
    682      1.16     yamt 	lockdebug_t *ld;
    683      1.34       ad 	struct cpu_info *ci;
    684      1.23       ad 	int s;
    685      1.10       ad 
    686      1.27       ad 	if (panicstr != NULL || ld_panic)
    687      1.24       ad 		return;
    688      1.24       ad 
    689      1.34       ad 	s = splhigh();
    690      1.34       ad 	ci = curcpu();
    691      1.34       ad 	__cpu_simple_lock(&ci->ci_data.cpu_ld_lock);
    692      1.16     yamt 	ld = (lockdebug_t *)rb_tree_find_node_geq(&ld_rb_tree, base);
    693      1.23       ad 	if (ld != NULL) {
    694      1.23       ad 		const uintptr_t lock = (uintptr_t)ld->ld_lock;
    695      1.23       ad 
    696      1.23       ad 		if ((uintptr_t)base > lock)
    697      1.23       ad 			panic("%s: corrupt tree ld=%p, base=%p, sz=%zu",
    698      1.23       ad 			    __func__, ld, base, sz);
    699      1.23       ad 		if (lock >= (uintptr_t)base + sz)
    700      1.23       ad 			ld = NULL;
    701      1.23       ad 	}
    702      1.34       ad 	__cpu_simple_unlock(&ci->ci_data.cpu_ld_lock);
    703      1.34       ad 	if (ld != NULL) {
    704      1.34       ad 		__cpu_simple_lock(&ld->ld_spinlock);
    705      1.34       ad 		lockdebug_abort1(ld, s, func,
    706      1.34       ad 		    "allocation contains active lock", !cold);
    707      1.16     yamt 		return;
    708      1.34       ad 	}
    709      1.34       ad 	splx(s);
    710      1.10       ad }
    711      1.10       ad 
    712      1.10       ad /*
    713       1.2       ad  * lockdebug_dump:
    714       1.2       ad  *
    715       1.2       ad  *	Dump information about a lock on panic, or for DDB.
    716       1.2       ad  */
    717       1.2       ad static void
    718       1.2       ad lockdebug_dump(lockdebug_t *ld, void (*pr)(const char *, ...))
    719       1.2       ad {
    720       1.2       ad 	int sleeper = (ld->ld_flags & LD_SLEEPER);
    721       1.2       ad 
    722       1.2       ad 	(*pr)(
    723       1.2       ad 	    "lock address : %#018lx type     : %18s\n"
    724      1.35       ad 	    "initialized  : %#018lx",
    725       1.2       ad 	    (long)ld->ld_lock, (sleeper ? "sleep/adaptive" : "spin"),
    726      1.10       ad 	    (long)ld->ld_initaddr);
    727       1.2       ad 
    728      1.35       ad 	if (ld->ld_lockops->lo_type == LOCKOPS_CV) {
    729      1.35       ad 		(*pr)(" interlock: %#018lx\n", ld->ld_locked);
    730      1.35       ad 	} else {
    731      1.35       ad 		(*pr)("\n"
    732      1.35       ad 		    "shared holds : %18u exclusive: %18u\n"
    733      1.35       ad 		    "shares wanted: %18u exclusive: %18u\n"
    734      1.35       ad 		    "current cpu  : %18u last held: %18u\n"
    735      1.35       ad 		    "current lwp  : %#018lx last held: %#018lx\n"
    736      1.35       ad 		    "last locked  : %#018lx unlocked : %#018lx\n",
    737      1.35       ad 		    (unsigned)ld->ld_shares, ((ld->ld_flags & LD_LOCKED) != 0),
    738      1.35       ad 		    (unsigned)ld->ld_shwant, (unsigned)ld->ld_exwant,
    739      1.40    rmind 		    (unsigned)cpu_index(curcpu()), (unsigned)ld->ld_cpu,
    740      1.35       ad 		    (long)curlwp, (long)ld->ld_lwp,
    741      1.35       ad 		    (long)ld->ld_locked, (long)ld->ld_unlocked);
    742      1.35       ad 	}
    743      1.35       ad 
    744       1.2       ad 	if (ld->ld_lockops->lo_dump != NULL)
    745       1.2       ad 		(*ld->ld_lockops->lo_dump)(ld->ld_lock);
    746       1.2       ad 
    747       1.2       ad 	if (sleeper) {
    748       1.2       ad 		(*pr)("\n");
    749       1.2       ad 		turnstile_print(ld->ld_lock, pr);
    750       1.2       ad 	}
    751       1.2       ad }
    752       1.2       ad 
    753       1.2       ad /*
    754      1.27       ad  * lockdebug_abort1:
    755       1.2       ad  *
    756      1.27       ad  *	An error has been trapped - dump lock info and panic.
    757       1.2       ad  */
    758       1.5       ad static void
    759      1.34       ad lockdebug_abort1(lockdebug_t *ld, int s, const char *func,
    760      1.10       ad 		 const char *msg, bool dopanic)
    761       1.2       ad {
    762       1.2       ad 
    763      1.27       ad 	/*
    764      1.27       ad 	 * Don't make the situation wose if the system is already going
    765      1.27       ad 	 * down in flames.  Once a panic is triggered, lockdebug state
    766      1.27       ad 	 * becomes stale and cannot be trusted.
    767      1.27       ad 	 */
    768      1.27       ad 	if (atomic_inc_uint_nv(&ld_panic) != 1) {
    769      1.34       ad 		__cpu_simple_unlock(&ld->ld_spinlock);
    770      1.34       ad 		splx(s);
    771      1.27       ad 		return;
    772      1.27       ad 	}
    773      1.27       ad 
    774       1.2       ad 	printf_nolog("%s error: %s: %s\n\n", ld->ld_lockops->lo_name,
    775       1.2       ad 	    func, msg);
    776       1.2       ad 	lockdebug_dump(ld, printf_nolog);
    777      1.34       ad 	__cpu_simple_unlock(&ld->ld_spinlock);
    778      1.34       ad 	splx(s);
    779       1.2       ad 	printf_nolog("\n");
    780      1.10       ad 	if (dopanic)
    781      1.10       ad 		panic("LOCKDEBUG");
    782       1.2       ad }
    783       1.2       ad 
    784       1.2       ad #endif	/* LOCKDEBUG */
    785       1.2       ad 
    786       1.2       ad /*
    787       1.2       ad  * lockdebug_lock_print:
    788       1.2       ad  *
    789       1.2       ad  *	Handle the DDB 'show lock' command.
    790       1.2       ad  */
    791       1.2       ad #ifdef DDB
    792       1.2       ad void
    793       1.2       ad lockdebug_lock_print(void *addr, void (*pr)(const char *, ...))
    794       1.2       ad {
    795       1.2       ad #ifdef LOCKDEBUG
    796       1.2       ad 	lockdebug_t *ld;
    797       1.2       ad 
    798       1.2       ad 	TAILQ_FOREACH(ld, &ld_all, ld_achain) {
    799      1.41   dyoung 		if (ld->ld_lock == NULL)
    800      1.41   dyoung 			continue;
    801      1.41   dyoung 		if (addr == NULL || ld->ld_lock == addr) {
    802       1.2       ad 			lockdebug_dump(ld, pr);
    803      1.41   dyoung 			if (addr != NULL)
    804      1.41   dyoung 				return;
    805       1.2       ad 		}
    806       1.2       ad 	}
    807      1.41   dyoung 	if (addr != NULL) {
    808      1.41   dyoung 		(*pr)("Sorry, no record of a lock with address %p found.\n",
    809      1.41   dyoung 		    addr);
    810      1.41   dyoung 	}
    811       1.2       ad #else
    812       1.2       ad 	(*pr)("Sorry, kernel not built with the LOCKDEBUG option.\n");
    813       1.2       ad #endif	/* LOCKDEBUG */
    814       1.2       ad }
    815       1.2       ad #endif	/* DDB */
    816       1.2       ad 
    817       1.2       ad /*
    818       1.2       ad  * lockdebug_abort:
    819       1.2       ad  *
    820       1.2       ad  *	An error has been trapped - dump lock info and call panic().
    821       1.2       ad  */
    822       1.2       ad void
    823      1.16     yamt lockdebug_abort(volatile void *lock, lockops_t *ops, const char *func,
    824      1.16     yamt 		const char *msg)
    825       1.2       ad {
    826       1.2       ad #ifdef LOCKDEBUG
    827       1.2       ad 	lockdebug_t *ld;
    828      1.34       ad 	int s;
    829       1.2       ad 
    830      1.34       ad 	s = splhigh();
    831      1.38    rafal 	if ((ld = lockdebug_lookup(lock,
    832      1.38    rafal 			(uintptr_t) __builtin_return_address(0))) != NULL) {
    833      1.34       ad 		lockdebug_abort1(ld, s, func, msg, true);
    834      1.34       ad 		return;
    835       1.2       ad 	}
    836      1.34       ad 	splx(s);
    837       1.2       ad #endif	/* LOCKDEBUG */
    838       1.2       ad 
    839      1.27       ad 	/*
    840      1.27       ad 	 * Complain first on the occurrance only.  Otherwise proceeed to
    841      1.27       ad 	 * panic where we will `rendezvous' with other CPUs if the machine
    842      1.27       ad 	 * is going down in flames.
    843      1.27       ad 	 */
    844      1.27       ad 	if (atomic_inc_uint_nv(&ld_panic) == 1) {
    845      1.27       ad 		printf_nolog("%s error: %s: %s\n\n"
    846      1.27       ad 		    "lock address : %#018lx\n"
    847      1.27       ad 		    "current cpu  : %18d\n"
    848      1.27       ad 		    "current lwp  : %#018lx\n",
    849      1.40    rmind 		    ops->lo_name, func, msg, (long)lock,
    850      1.40    rmind 		    (int)cpu_index(curcpu()), (long)curlwp);
    851      1.27       ad 		(*ops->lo_dump)(lock);
    852      1.27       ad 		printf_nolog("\n");
    853      1.27       ad 	}
    854       1.2       ad 
    855       1.2       ad 	panic("lock error");
    856       1.2       ad }
    857