X-Git-Url: https://git.sur5r.net/?a=blobdiff_plain;f=servers%2Fslapd%2Fback-bdb%2Fcache.c;h=5c9dc31490405d4a4e79bcadc1cc97e1e9213e57;hb=822888eafcbeb0814cacf9155b8f6e0ae3822fda;hp=8e39d952dffd3a597e9f0dbd6b248a5ef8f14090;hpb=af23be36751dcb34555f7ac7092f6edf4c3a4171;p=openldap

diff --git a/servers/slapd/back-bdb/cache.c b/servers/slapd/back-bdb/cache.c
index 8e39d952df..5c9dc31490 100644
--- a/servers/slapd/back-bdb/cache.c
+++ b/servers/slapd/back-bdb/cache.c
@@ -2,7 +2,7 @@
 /* $OpenLDAP$ */
 /* This work is part of OpenLDAP Software <http://www.openldap.org/>.
  *
- * Copyright 2000-2005 The OpenLDAP Foundation.
+ * Copyright 2000-2006 The OpenLDAP Foundation.
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
@@ -26,9 +26,12 @@
 
 #include "back-bdb.h"
 
+#include "ldap_rq.h"
+
 #ifdef BDB_HIER
-#define bdb_cache_lru_add	hdb_cache_lru_add
+#define bdb_cache_lru_purge	hdb_cache_lru_purge
 #endif
+static void bdb_cache_lru_purge( struct bdb_info *bdb );
 
 static int	bdb_cache_delete_internal(Cache *cache, EntryInfo *e, int decr);
 #ifdef LDAP_DEBUG
@@ -37,41 +40,94 @@ static void	bdb_lru_print(Cache *cache);
 #endif
 #endif
 
+/* For concurrency experiments only! */
+#if 0
+#define	ldap_pvt_thread_rdwr_wlock(a)	0
+#define	ldap_pvt_thread_rdwr_wunlock(a)	0
+#define	ldap_pvt_thread_rdwr_rlock(a)	0
+#define	ldap_pvt_thread_rdwr_runlock(a)	0
+#endif
+
+#if 0
+#define ldap_pvt_thread_mutex_trylock(a) 0
+#endif
+
 static EntryInfo *
 bdb_cache_entryinfo_new( Cache *cache )
 {
 	EntryInfo *ei = NULL;
 
 	if ( cache->c_eifree ) {
-		ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
+		ldap_pvt_thread_mutex_lock( &cache->c_eifree_mutex );
 		if ( cache->c_eifree ) {
 			ei = cache->c_eifree;
 			cache->c_eifree = ei->bei_lrunext;
 		}
-		ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
+		ldap_pvt_thread_mutex_unlock( &cache->c_eifree_mutex );
 	}
-	if ( ei ) {
-		ei->bei_lrunext = NULL;
-		ei->bei_state = 0;
-	} else {
-		ei = ch_calloc(1, sizeof(struct bdb_entry_info));
+	if ( !ei ) {
+		ei = ch_calloc(1, sizeof(EntryInfo));
 		ldap_pvt_thread_mutex_init( &ei->bei_kids_mutex );
 	}
 
+	ei->bei_state = CACHE_ENTRY_REFERENCED;
+
 	return ei;
 }
 
+/* Note - we now use a Second-Chance / Clock algorithm instead of
+ * Least-Recently-Used. This tremendously improves concurrency
+ * because we no longer need to manipulate the lists every time an
+ * entry is touched. We only need to lock the lists when adding
+ * or deleting an entry. It's now a circular doubly-linked list.
+ * We always append to the tail, but the head traverses the circle
+ * during a purge operation.
+ */
+static void
+bdb_cache_lru_link( Cache *cache, EntryInfo *ei )
+{
+	/* Insert into circular LRU list */
+	ldap_pvt_thread_mutex_lock( &cache->lru_tail_mutex );
+	ei->bei_lruprev = cache->c_lrutail;
+	if ( cache->c_lrutail ) {
+		ei->bei_lrunext = cache->c_lrutail->bei_lrunext;
+		cache->c_lrutail->bei_lrunext = ei;
+		if ( ei->bei_lrunext )
+			ei->bei_lrunext->bei_lruprev = ei;
+	} else {
+		ei->bei_lrunext = ei->bei_lruprev = ei;
+		cache->c_lruhead = ei;
+	}
+	cache->c_lrutail = ei;
+	ldap_pvt_thread_mutex_unlock( &cache->lru_tail_mutex );
+}
+
+#ifdef NO_THREADS
+#define NO_DB_LOCK
+#endif
+
+/* #define NO_DB_LOCK 1 */
+/* Note: The BerkeleyDB locks are much slower than regular
+ * mutexes or rdwr locks. But the BDB implementation has the
+ * advantage of using a fixed size lock table, instead of
+ * allocating a lock object per entry in the DB. That's a
+ * key benefit for scaling. It also frees us from worrying
+ * about undetectable deadlocks between BDB activity and our
+ * own cache activity. It's still worth exploring faster
+ * alternatives though.
+ */
+
 /* Atomically release and reacquire a lock */
 int
 bdb_cache_entry_db_relock(
-	DB_ENV *env,
+	struct bdb_info *bdb,
 	u_int32_t locker,
 	EntryInfo *ei,
 	int rw,
 	int tryOnly,
 	DB_LOCK *lock )
 {
-#ifdef NO_THREADS
+#ifdef NO_DB_LOCK
 	return 0;
 #else
 	int	rc;
@@ -89,7 +145,7 @@ bdb_cache_entry_db_relock(
 	list[1].lock = *lock;
 	list[1].mode = rw ? DB_LOCK_WRITE : DB_LOCK_READ;
 	list[1].obj = &lockobj;
-	rc = env->lock_vec(env, locker, tryOnly ? DB_LOCK_NOWAIT : 0,
+	rc = bdb->bi_dbenv->lock_vec(bdb->bi_dbenv, locker, tryOnly ? DB_LOCK_NOWAIT : 0,
 		list, 2, NULL );
 
 	if (rc && !tryOnly) {
@@ -104,10 +160,10 @@ bdb_cache_entry_db_relock(
 }
 
 static int
-bdb_cache_entry_db_lock( DB_ENV *env, u_int32_t locker, EntryInfo *ei,
+bdb_cache_entry_db_lock( struct bdb_info *bdb, u_int32_t locker, EntryInfo *ei,
 	int rw, int tryOnly, DB_LOCK *lock )
 {
-#ifdef NO_THREADS
+#ifdef NO_DB_LOCK
 	return 0;
 #else
 	int       rc;
@@ -124,7 +180,7 @@ bdb_cache_entry_db_lock( DB_ENV *env, u_int32_t locker, EntryInfo *ei,
 	lockobj.data = &ei->bei_id;
 	lockobj.size = sizeof(ei->bei_id) + 1;
 
-	rc = LOCK_GET(env, locker, tryOnly ? DB_LOCK_NOWAIT : 0,
+	rc = LOCK_GET(bdb->bi_dbenv, locker, tryOnly ? DB_LOCK_NOWAIT : 0,
 					&lockobj, db_rw, lock);
 	if (rc && !tryOnly) {
 		Debug( LDAP_DEBUG_TRACE,
@@ -132,20 +188,20 @@ bdb_cache_entry_db_lock( DB_ENV *env, u_int32_t locker, EntryInfo *ei,
 			ei->bei_id, rw, rc );
 	}
 	return rc;
-#endif /* NO_THREADS */
+#endif /* NO_DB_LOCK */
 }
 
 int
-bdb_cache_entry_db_unlock ( DB_ENV *env, DB_LOCK *lock )
+bdb_cache_entry_db_unlock ( struct bdb_info *bdb, DB_LOCK *lock )
 {
-#ifdef NO_THREADS
+#ifdef NO_DB_LOCK
 	return 0;
 #else
 	int rc;
 
-	if ( !lock ) return 0;
+	if ( !lock || lock->mode == DB_LOCK_NG ) return 0;
 
-	rc = LOCK_PUT ( env, lock );
+	rc = LOCK_PUT ( bdb->bi_dbenv, lock );
 	return rc;
 #endif
 }
@@ -162,32 +218,6 @@ bdb_cache_entryinfo_destroy( EntryInfo *e )
 	return 0;
 }
 
-#define LRU_DELETE( cache, ei ) do { \
-	if ( (ei)->bei_lruprev != NULL ) { \
-		(ei)->bei_lruprev->bei_lrunext = (ei)->bei_lrunext; \
-	} else { \
-		(cache)->c_lruhead = (ei)->bei_lrunext; \
-	} \
-	if ( (ei)->bei_lrunext != NULL ) { \
-		(ei)->bei_lrunext->bei_lruprev = (ei)->bei_lruprev; \
-	} else { \
-		(cache)->c_lrutail = (ei)->bei_lruprev; \
-	} \
-	(ei)->bei_lrunext = (ei)->bei_lruprev = NULL; \
-} while(0)
-
-#define LRU_ADD( cache, ei ) do { \
-	(ei)->bei_lrunext = (cache)->c_lruhead; \
-	if ( (ei)->bei_lrunext != NULL ) { \
-		(ei)->bei_lrunext->bei_lruprev = (ei); \
-	} \
-	(cache)->c_lruhead = (ei); \
-	(ei)->bei_lruprev = NULL; \
-	if ( (cache)->c_lrutail == NULL ) { \
-		(cache)->c_lrutail = (ei); \
-	} \
-} while(0)
-
 /* Do a length-ordered sort on normalized RDNs */
 static int
 bdb_rdn_cmp( const void *v_e1, const void *v_e2 )
@@ -262,6 +292,7 @@ bdb_entryinfo_add_internal(
 		ei->bei_parent->bei_ckids++;
 #endif
 	}
+	bdb_cache_lru_link( &bdb->bi_cache, ei2 );
 
 	*res = ei2;
 	return 0;
@@ -310,6 +341,7 @@ bdb_cache_find_ndn(
 	}
 	
 	for ( bdb_cache_entryinfo_lock( eip ); eip; ) {
+		eip->bei_state |= CACHE_ENTRY_REFERENCED;
 		ei.bei_parent = eip;
 		ei2 = (EntryInfo *)avl_find( eip->bei_kids, &ei, bdb_rdn_cmp );
 		if ( !ei2 ) {
@@ -387,7 +419,6 @@ hdb_cache_find_parent(
 	struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
 	EntryInfo ei, eip, *ei2 = NULL, *ein = NULL, *eir = NULL;
 	int rc;
-	int addlru = 0;
 
 	ei.bei_id = id;
 	ei.bei_kids = NULL;
@@ -413,7 +444,7 @@ hdb_cache_find_parent(
 		ei.bei_ckids = 0;
 		
 		/* This node is not fully connected yet */
-		ein->bei_state = CACHE_ENTRY_NOT_LINKED;
+		ein->bei_state |= CACHE_ENTRY_NOT_LINKED;
 
 		/* Insert this node into the ID tree */
 		ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
@@ -435,15 +466,9 @@ hdb_cache_find_parent(
 				ein->bei_ckids++;
 				bdb_cache_entryinfo_unlock( ein );
 			}
-			addlru = 0;
-
-		}
-		if ( addlru ) {
-			ldap_pvt_thread_mutex_lock( &bdb->bi_cache.lru_mutex );
-			LRU_ADD( &bdb->bi_cache, ein );
-			ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_mutex );
+		} else {
+			bdb_cache_lru_link( &bdb->bi_cache, ein );
 		}
-		addlru = 1;
 
 		/* If this is the first time, save this node
 		 * to be returned later.
@@ -469,15 +494,18 @@ hdb_cache_find_parent(
 		if ( ei2 ) {
 			bdb_cache_entryinfo_lock( ei2 );
 			ein->bei_parent = ei2;
+
 			avl_insert( &ei2->bei_kids, (caddr_t)ein, bdb_rdn_cmp,
 				avl_dup_error);
 			ei2->bei_ckids++;
-			bdb_cache_entryinfo_unlock( ei2 );
-			bdb_cache_entryinfo_lock( eir );
 
 			/* Reset all the state info */
 			for (ein = eir; ein != ei2; ein=ein->bei_parent)
 				ein->bei_state &= ~CACHE_ENTRY_NOT_LINKED;
+
+			bdb_cache_entryinfo_unlock( ei2 );
+			bdb_cache_entryinfo_lock( eir );
+
 			*res = eir;
 			break;
 		}
@@ -526,94 +554,108 @@ int hdb_cache_load(
 }
 #endif
 
-/* caller must have lru_mutex locked. mutex
- * will be unlocked on return.
- */
 static void
-bdb_cache_lru_add(
-	struct bdb_info *bdb,
-	u_int32_t	locker,
-	EntryInfo *ei )
+bdb_cache_lru_purge( struct bdb_info *bdb )
 {
 	DB_LOCK		lock, *lockp;
+	EntryInfo *elru, *elnext;
+	int count, islocked;
+
+	/* Don't bother if we can't get the lock */
+	if ( ldap_pvt_thread_mutex_trylock( &bdb->bi_cache.lru_head_mutex ) )
+		return;
 
-	if ( locker ) {
+	if ( bdb->bi_cache.c_cursize <= bdb->bi_cache.c_maxsize ) {
+		ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_head_mutex );
+		return;
+	}
+
+	if ( bdb->bi_cache.c_locker ) {
 		lockp = &lock;
 	} else {
 		lockp = NULL;
 	}
 
-	/* See if we're above the cache size limit */
-	if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize ) {
-		EntryInfo *elru, *elprev;
-		int i = 0;
+	count = 0;
+	/* Look for an unused entry to remove */
+	for (elru = bdb->bi_cache.c_lruhead; elru; elru = elnext ) {
+		elnext = elru->bei_lrunext;
 
-		/* Look for an unused entry to remove */
-		for (elru = bdb->bi_cache.c_lrutail; elru; elru = elprev, i++ ) {
-			elprev = elru->bei_lruprev;
+		if ( ldap_pvt_thread_mutex_trylock( &elru->bei_kids_mutex ))
+			continue;
 
-			/* Too many probes, not enough idle, give up */
-			if (i > 10)
-				break;
+		/* This flag implements the clock replacement behavior */
+		if ( elru->bei_state & ( CACHE_ENTRY_REFERENCED )) {
+			elru->bei_state &= ~CACHE_ENTRY_REFERENCED;
+			bdb_cache_entryinfo_unlock( elru );
+			continue;
+		}
 
-			/* If we can successfully writelock it, then
-			 * the object is idle.
-			 */
-			if ( bdb_cache_entry_db_lock( bdb->bi_dbenv,
-					bdb->bi_cache.c_locker, elru, 1, 1, lockp ) == 0 ) {
+		/* If this node is in the process of linking into the cache,
+		 * or this node is being deleted, skip it.
+		 *
+		 * Also, if this node has no entry attached, skip it, there's
+		 * nothing to purge anyway.
+		 */
+		if (( elru->bei_state & ( CACHE_ENTRY_NOT_LINKED |
+			CACHE_ENTRY_DELETED | CACHE_ENTRY_LOADING )) ||
+			!elru->bei_e ) {
+			bdb_cache_entryinfo_unlock( elru );
+			continue;
+		}
 
-				int stop = 0, decr = 0;
+		/* entryinfo is locked */
+		islocked = 1;
 
-				/* If this node is in the process of linking into the cache,
-				 * or this node is being deleted, skip it.
-				 */
-				if ( elru->bei_state &
-					( CACHE_ENTRY_NOT_LINKED | CACHE_ENTRY_DELETED )) {
-					bdb_cache_entry_db_unlock( bdb->bi_dbenv, lockp );
-					continue;
-				}
-				/* Free entry for this node if it's present */
-				if ( elru->bei_e ) {
-					elru->bei_e->e_private = NULL;
+		/* If we can successfully writelock it, then
+		 * the object is idle.
+		 */
+		if ( bdb_cache_entry_db_lock( bdb,
+			bdb->bi_cache.c_locker, elru, 1, 1, lockp ) == 0 ) {
+
+			/* Free entry for this node if it's present */
+			if ( elru->bei_e ) {
+				elru->bei_e->e_private = NULL;
 #ifdef SLAP_ZONE_ALLOC
-					bdb_entry_return( bdb, elru->bei_e, elru->bei_zseq );
+				bdb_entry_return( bdb, elru->bei_e, elru->bei_zseq );
 #else
-					bdb_entry_return( elru->bei_e );
+				bdb_entry_return( elru->bei_e );
 #endif
-					elru->bei_e = NULL;
-					decr = 1;
-				}
-				/* ITS#4010 if we're in slapcat, and this node is a leaf
-				 * node, free it.
-				 *
-				 * FIXME: we need to do this for slapd as well, (which is
-				 * why we compute bi_cache.c_leaves now) but at the moment
-				 * we can't because it causes unresolvable deadlocks. 
-				 */
-				if ( slapMode & SLAP_TOOL_READONLY ) {
-					if ( !elru->bei_kids ) {
-						/* This does LRU_DELETE for us */
-						bdb_cache_delete_internal( &bdb->bi_cache, elru, 0 );
-						bdb_cache_delete_cleanup( &bdb->bi_cache, elru );
-					}
-					/* Leave node on LRU list for a future pass */
-				} else {
-					LRU_DELETE( &bdb->bi_cache, elru );
+				elru->bei_e = NULL;
+				count++;
+			}
+			bdb_cache_entry_db_unlock( bdb, lockp );
+
+			/* ITS#4010 if we're in slapcat, and this node is a leaf
+			 * node, free it.
+			 *
+			 * FIXME: we need to do this for slapd as well, (which is
+			 * why we compute bi_cache.c_leaves now) but at the moment
+			 * we can't because it causes unresolvable deadlocks. 
+			 */
+			if ( slapMode & SLAP_TOOL_READONLY ) {
+				if ( !elru->bei_kids ) {
+					bdb_cache_delete_internal( &bdb->bi_cache, elru, 0 );
+					bdb_cache_delete_cleanup( &bdb->bi_cache, elru );
+					islocked = 0;
 				}
-				bdb_cache_entry_db_unlock( bdb->bi_dbenv, lockp );
-
-				ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
-				if ( decr )
-					--bdb->bi_cache.c_cursize;
-				if (bdb->bi_cache.c_cursize <= bdb->bi_cache.c_maxsize)
-					stop = 1;
-				ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
-				if (stop) break;
+				/* Leave node on LRU list for a future pass */
 			}
 		}
+
+		if ( islocked )
+			bdb_cache_entryinfo_unlock( elru );
+
+		if ( count >= bdb->bi_cache.c_minfree ) {
+			ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
+			bdb->bi_cache.c_cursize -= count;
+			ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
+			break;
+		}
 	}
-	LRU_ADD( &bdb->bi_cache, ei );
-	ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_mutex );
+
+	bdb->bi_cache.c_lruhead = elnext;
+	ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_head_mutex );
 }
 
 EntryInfo *
@@ -702,6 +744,7 @@ again:	ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
 				&ep->e_nname, eip );
 			if ( *eip ) islocked = 1;
 			if ( rc ) {
+				ep->e_private = NULL;
 #ifdef SLAP_ZONE_ALLOC
 				bdb_entry_return( bdb, ep, (*eip)->bei_zseq );
 #else
@@ -712,12 +755,12 @@ again:	ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
 		}
 #else
 		rc = hdb_cache_find_parent(op, tid, locker, id, eip );
-		if ( rc == 0 && *eip ) islocked = 1;
+		if ( rc == 0 ) islocked = 1;
 #endif
 	}
 
 	/* Ok, we found the info, do we have the entry? */
-	if ( *eip && rc == 0 ) {
+	if ( rc == 0 ) {
 		if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
 			rc = DB_NOTFOUND;
 		} else {
@@ -738,19 +781,13 @@ load1:
 				bdb_cache_entryinfo_unlock( *eip );
 				islocked = 0;
 			}
-			rc = bdb_cache_entry_db_lock( bdb->bi_dbenv, locker, *eip, 0, 0, lock );
+			rc = bdb_cache_entry_db_lock( bdb, locker, *eip, load, 0, lock );
 			if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
 				rc = DB_NOTFOUND;
-				bdb_cache_entry_db_unlock( bdb->bi_dbenv, lock );
+				bdb_cache_entry_db_unlock( bdb, lock );
 			} else if ( rc == 0 ) {
 				if ( load ) {
-					/* Give up original read lock, obtain write lock
-					 */
-				    if ( rc == 0 ) {
-						rc = bdb_cache_entry_db_relock( bdb->bi_dbenv, locker,
-							*eip, 1, 0, lock );
-					}
-					if ( rc == 0 && !ep) {
+					if ( !ep) {
 						rc = bdb_id2entry( op->o_bd, tid, locker, id, &ep );
 					}
 					if ( rc == 0 ) {
@@ -764,21 +801,22 @@ load1:
 #endif
 						ep = NULL;
 					}
+					bdb_cache_entryinfo_lock( *eip );
 					(*eip)->bei_state ^= CACHE_ENTRY_LOADING;
+					bdb_cache_entryinfo_unlock( *eip );
 					if ( rc == 0 ) {
 						/* If we succeeded, downgrade back to a readlock. */
-						rc = bdb_cache_entry_db_relock( bdb->bi_dbenv, locker,
+						rc = bdb_cache_entry_db_relock( bdb, locker,
 							*eip, 0, 0, lock );
 					} else {
 						/* Otherwise, release the lock. */
-						bdb_cache_entry_db_unlock( bdb->bi_dbenv, lock );
+						bdb_cache_entry_db_unlock( bdb, lock );
 					}
 				} else if ( !(*eip)->bei_e ) {
 					/* Some other thread is trying to load the entry,
-					 * give it a chance to finish.
+					 * wait for it to finish.
 					 */
-					bdb_cache_entry_db_unlock( bdb->bi_dbenv, lock );
-					ldap_pvt_thread_yield();
+					bdb_cache_entry_db_unlock( bdb, lock );
 					bdb_cache_entryinfo_lock( *eip );
 					islocked = 1;
 					goto load1;
@@ -788,12 +826,12 @@ load1:
 					 */
 					rc = bdb_fix_dn( (*eip)->bei_e, 1 );
 					if ( rc ) {
-						bdb_cache_entry_db_relock( bdb->bi_dbenv,
+						bdb_cache_entry_db_relock( bdb,
 							locker, *eip, 1, 0, lock );
 						/* check again in case other modifier did it already */
 						if ( bdb_fix_dn( (*eip)->bei_e, 1 ) )
 							rc = bdb_fix_dn( (*eip)->bei_e, 2 );
-						bdb_cache_entry_db_relock( bdb->bi_dbenv,
+						bdb_cache_entry_db_relock( bdb,
 							locker, *eip, 0, 0, lock );
 					}
 #endif
@@ -806,6 +844,7 @@ load1:
 		bdb_cache_entryinfo_unlock( *eip );
 	}
 	if ( ep ) {
+		ep->e_private = NULL;
 #ifdef SLAP_ZONE_ALLOC
 		bdb_entry_return( bdb, ep, (*eip)->bei_zseq );
 #else
@@ -813,30 +852,17 @@ load1:
 #endif
 	}
 	if ( rc == 0 ) {
+		int purge = 0;
 
 		if ( load ) {
-			ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
+			ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
 			bdb->bi_cache.c_cursize++;
-			ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
-		}
-
-		ldap_pvt_thread_mutex_lock( &bdb->bi_cache.lru_mutex );
-
-		/* If the LRU list has only one entry and this is it, it
-		 * doesn't need to be added again.
-		 */
-		if ( bdb->bi_cache.c_lruhead == bdb->bi_cache.c_lrutail &&
-			bdb->bi_cache.c_lruhead == *eip ) {
-			ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_mutex );
-		} else {
-
-			/* if entry is on LRU list, remove from old spot */
-			if ( (*eip)->bei_lrunext || (*eip)->bei_lruprev ) {
-				LRU_DELETE( &bdb->bi_cache, *eip );
-			}
-			/* lru_mutex is unlocked for us */
-			bdb_cache_lru_add( bdb, locker, *eip );
+			if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize )
+				purge = 1;
+			ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
 		}
+		if ( purge )
+			bdb_cache_lru_purge( bdb );
 	}
 
 #ifdef SLAP_ZONE_ALLOC
@@ -876,11 +902,11 @@ bdb_cache_add(
 	EntryInfo *eip,
 	Entry *e,
 	struct berval *nrdn,
-	u_int32_t locker )
+	u_int32_t locker,
+	DB_LOCK *lock )
 {
 	EntryInfo *new, ei;
-	DB_LOCK lock;
-	int rc;
+	int rc, purge = 0;
 #ifdef BDB_HIER
 	struct berval rdn = e->e_name;
 #endif
@@ -893,7 +919,7 @@ bdb_cache_add(
 	/* Lock this entry so that bdb_add can run to completion.
 	 * It can only fail if BDB has run out of lock resources.
 	 */
-	rc = bdb_cache_entry_db_lock( bdb->bi_dbenv, locker, &ei, 1, 0, &lock );
+	rc = bdb_cache_entry_db_lock( bdb, locker, &ei, 0, 0, lock );
 	if ( rc ) {
 		bdb_cache_entryinfo_unlock( eip );
 		return rc;
@@ -921,37 +947,38 @@ bdb_cache_add(
 	}
 	new->bei_e = e;
 	e->e_private = new;
-	new->bei_state = CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
+	new->bei_state |= CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
 	eip->bei_state &= ~CACHE_ENTRY_NO_KIDS;
 	if (eip->bei_parent) {
 		eip->bei_parent->bei_state &= ~CACHE_ENTRY_NO_GRANDKIDS;
 	}
 	bdb_cache_entryinfo_unlock( eip );
 
-	++bdb->bi_cache.c_cursize;
 	ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
+	ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
+	++bdb->bi_cache.c_cursize;
+	if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize )
+		purge = 1;
+	ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
 
-	/* set lru mutex */
-	ldap_pvt_thread_mutex_lock( &bdb->bi_cache.lru_mutex );
-
-	/* lru_mutex is unlocked for us */
-	bdb_cache_lru_add( bdb, locker, new );
+	if ( purge )
+		bdb_cache_lru_purge( bdb );
 
 	return rc;
 }
 
 int
 bdb_cache_modify(
+	struct bdb_info *bdb,
 	Entry *e,
 	Attribute *newAttrs,
-	DB_ENV *env,
 	u_int32_t locker,
 	DB_LOCK *lock )
 {
 	EntryInfo *ei = BEI(e);
 	int rc;
 	/* Get write lock on data */
-	rc = bdb_cache_entry_db_relock( env, locker, ei, 1, 0, lock );
+	rc = bdb_cache_entry_db_relock( bdb, locker, ei, 1, 0, lock );
 
 	/* If we've done repeated mods on a cached entry, then e_attrs
 	 * is no longer contiguous with the entry, and must be freed.
@@ -985,7 +1012,7 @@ bdb_cache_modrdn(
 #endif
 
 	/* Get write lock on data */
-	rc =  bdb_cache_entry_db_relock( bdb->bi_dbenv, locker, ei, 1, 0, lock );
+	rc =  bdb_cache_entry_db_relock( bdb, locker, ei, 1, 0, lock );
 	if ( rc ) return rc;
 
 	/* If we've done repeated mods on a cached entry, then e_attrs
@@ -1010,6 +1037,10 @@ bdb_cache_modrdn(
 	avl_delete( &pei->bei_kids, (caddr_t) ei, bdb_rdn_cmp );
 	free( ei->bei_nrdn.bv_val );
 	ber_dupbv( &ei->bei_nrdn, nrdn );
+
+	if ( !pei->bei_kids )
+		pei->bei_state |= CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
+
 #ifdef BDB_HIER
 	free( ei->bei_rdn.bv_val );
 
@@ -1020,6 +1051,8 @@ bdb_cache_modrdn(
 		rdn.bv_len = ptr - rdn.bv_val;
 	}
 	ber_dupbv( &ei->bei_rdn, &rdn );
+	pei->bei_ckids--;
+	if ( pei->bei_dkids ) pei->bei_dkids--;
 #endif
 
 	if (!ein) {
@@ -1029,7 +1062,16 @@ bdb_cache_modrdn(
 		bdb_cache_entryinfo_unlock( pei );
 		bdb_cache_entryinfo_lock( ein );
 	}
+	/* parent now has kids */
+	if ( ein->bei_state & CACHE_ENTRY_NO_KIDS )
+		ein->bei_state ^= CACHE_ENTRY_NO_KIDS;
+
 #ifdef BDB_HIER
+	/* parent might now have grandkids */
+	if ( ein->bei_state & CACHE_ENTRY_NO_GRANDKIDS &&
+		!(ei->bei_state & (CACHE_ENTRY_NO_KIDS)))
+		ein->bei_state ^= CACHE_ENTRY_NO_GRANDKIDS;
+
 	{
 		/* Record the generation number of this change */
 		ldap_pvt_thread_mutex_lock( &bdb->bi_modrdns_mutex );
@@ -1037,6 +1079,8 @@ bdb_cache_modrdn(
 		ei->bei_modrdns = bdb->bi_modrdns;
 		ldap_pvt_thread_mutex_unlock( &bdb->bi_modrdns_mutex );
 	}
+	ein->bei_ckids++;
+	if ( ein->bei_dkids ) ein->bei_dkids++;
 #endif
 	avl_insert( &ein->bei_kids, ei, bdb_rdn_cmp, avl_dup_error );
 	bdb_cache_entryinfo_unlock( ein );
@@ -1051,9 +1095,8 @@ bdb_cache_modrdn(
  */
 int
 bdb_cache_delete(
-    Cache	*cache,
+	struct bdb_info *bdb,
     Entry		*e,
-    DB_ENV	*env,
     u_int32_t	locker,
     DB_LOCK	*lock )
 {
@@ -1069,7 +1112,7 @@ bdb_cache_delete(
 	bdb_cache_entryinfo_lock( ei );
 
 	/* Get write lock on the data */
-	rc = bdb_cache_entry_db_relock( env, locker, ei, 1, 0, lock );
+	rc = bdb_cache_entry_db_relock( bdb, locker, ei, 1, 0, lock );
 	if ( rc ) {
 		/* couldn't lock, undo and give up */
 		ei->bei_state ^= CACHE_ENTRY_DELETED;
@@ -1081,18 +1124,12 @@ bdb_cache_delete(
 		e->e_id, 0, 0 );
 
 	/* set lru mutex */
-	ldap_pvt_thread_mutex_lock( &cache->lru_mutex );
-
-	/* set cache write lock */
-	ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
+	ldap_pvt_thread_mutex_lock( &bdb->bi_cache.lru_head_mutex );
 
-	rc = bdb_cache_delete_internal( cache, e->e_private, 1 );
-
-	/* free cache write lock */
-	ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
+	rc = bdb_cache_delete_internal( &bdb->bi_cache, e->e_private, 1 );
 
 	/* free lru mutex */
-	ldap_pvt_thread_mutex_unlock( &cache->lru_mutex );
+	ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_head_mutex );
 
 	/* Leave entry info locked */
 
@@ -1127,10 +1164,10 @@ bdb_cache_delete_cleanup(
 	ei->bei_kids = NULL;
 	ei->bei_lruprev = NULL;
 
-	ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
+	ldap_pvt_thread_mutex_lock( &cache->c_eifree_mutex );
 	ei->bei_lrunext = cache->c_eifree;
 	cache->c_eifree = ei;
-	ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
+	ldap_pvt_thread_mutex_unlock( &cache->c_eifree_mutex );
 	bdb_cache_entryinfo_unlock( ei );
 }
 
@@ -1141,6 +1178,7 @@ bdb_cache_delete_internal(
     int		decr )
 {
 	int rc = 0;	/* return code */
+	int decr_leaf = 0;
 
 	/* Lock the parent's kids tree */
 	bdb_cache_entryinfo_lock( e->bei_parent );
@@ -1156,22 +1194,39 @@ bdb_cache_delete_internal(
 		rc = -1;
 	}
 	if ( e->bei_parent->bei_kids )
-		cache->c_leaves--;
+		decr_leaf = 1;
 
+	bdb_cache_entryinfo_unlock( e->bei_parent );
+
+	ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
 	/* id tree */
-	if ( avl_delete( &cache->c_idtree, (caddr_t) e, bdb_id_cmp ) == NULL ) {
+	if ( avl_delete( &cache->c_idtree, (caddr_t) e, bdb_id_cmp )) {
+		cache->c_eiused--;
+		if ( decr_leaf )
+			cache->c_leaves--;
+	} else {
 		rc = -1;
 	}
+	ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
 
 	if ( rc == 0 ){
-		cache->c_eiused--;
-
 		/* lru */
-		LRU_DELETE( cache, e );
-		if ( e->bei_e ) cache->c_cursize--;
-	}
+		if ( e == cache->c_lruhead ) cache->c_lruhead = e->bei_lrunext;
+		if ( e == cache->c_lrutail ) {
+			ldap_pvt_thread_mutex_lock( &cache->lru_tail_mutex );
+			if ( e == cache->c_lrutail ) cache->c_lrutail = e->bei_lruprev;
+			ldap_pvt_thread_mutex_unlock( &cache->lru_tail_mutex );
+		}
 
-	bdb_cache_entryinfo_unlock( e->bei_parent );
+		if ( e->bei_lrunext ) e->bei_lrunext->bei_lruprev = e->bei_lruprev;
+		if ( e->bei_lruprev ) e->bei_lruprev->bei_lrunext = e->bei_lrunext;
+
+		if ( e->bei_e ) {
+			ldap_pvt_thread_mutex_lock( &cache->c_count_mutex );
+			cache->c_cursize--;
+			ldap_pvt_thread_mutex_unlock( &cache->c_count_mutex );
+		}
+	}
 
 	return( rc );
 }
@@ -1200,7 +1255,7 @@ bdb_cache_release_all( Cache *cache )
 	/* set cache write lock */
 	ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
 	/* set lru mutex */
-	ldap_pvt_thread_mutex_lock( &cache->lru_mutex );
+	ldap_pvt_thread_mutex_lock( &cache->lru_tail_mutex );
 
 	Debug( LDAP_DEBUG_TRACE, "====> bdb_cache_release_all\n", 0, 0, 0 );
 
@@ -1219,7 +1274,7 @@ bdb_cache_release_all( Cache *cache )
 	cache->c_dntree.bei_kids = NULL;
 
 	/* free lru mutex */
-	ldap_pvt_thread_mutex_unlock( &cache->lru_mutex );
+	ldap_pvt_thread_mutex_unlock( &cache->lru_tail_mutex );
 	/* free cache write lock */
 	ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
 }
@@ -1231,15 +1286,22 @@ bdb_lru_print( Cache *cache )
 {
 	EntryInfo	*e;
 
-	fprintf( stderr, "LRU queue (head to tail):\n" );
-	for ( e = cache->c_lruhead; e != NULL; e = e->bei_lrunext ) {
-		fprintf( stderr, "\trdn \"%20s\" id %ld\n",
-			e->bei_nrdn.bv_val, e->bei_id );
+	fprintf( stderr, "LRU circle head: %p\n", cache->c_lruhead );
+	fprintf( stderr, "LRU circle (tail forward):\n" );
+	for ( e = cache->c_lrutail; ; ) {
+		fprintf( stderr, "\t%p, %p id %ld rdn \"%s\"\n",
+			e, e->bei_e, e->bei_id, e->bei_nrdn.bv_val );
+		e = e->bei_lrunext;
+		if ( e == cache->c_lrutail )
+			break;
 	}
-	fprintf( stderr, "LRU queue (tail to head):\n" );
-	for ( e = cache->c_lrutail; e != NULL; e = e->bei_lruprev ) {
-		fprintf( stderr, "\trdn \"%20s\" id %ld\n",
-			e->bei_nrdn.bv_val, e->bei_id );
+	fprintf( stderr, "LRU circle (tail backward):\n" );
+	for ( e = cache->c_lrutail; ; ) {
+		fprintf( stderr, "\t%p, %p id %ld rdn \"%s\"\n",
+			e, e->bei_e, e->bei_id, e->bei_nrdn.bv_val );
+		e = e->bei_lruprev;
+		if ( e == cache->c_lrutail )
+			break;
 	}
 }
 #endif
@@ -1314,29 +1376,3 @@ bdb_locker_id( Operation *op, DB_ENV *env, u_int32_t *locker )
 	return 0;
 }
 #endif /* BDB_REUSE_LOCKERS */
-
-void
-bdb_cache_delete_entry(
-	struct bdb_info *bdb,
-	EntryInfo *ei,
-	u_int32_t locker,
-	DB_LOCK *lock )
-{
-	ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
-	if ( bdb_cache_entry_db_lock( bdb->bi_dbenv, bdb->bi_cache.c_locker, ei, 1, 1, lock ) == 0 )
-	{
-		if ( ei->bei_e && !(ei->bei_state & CACHE_ENTRY_NOT_LINKED )) {
-			LRU_DELETE( &bdb->bi_cache, ei );
-			ei->bei_e->e_private = NULL;
-#ifdef SLAP_ZONE_ALLOC
-			bdb_entry_return( bdb, ei->bei_e, ei->bei_zseq );
-#else
-			bdb_entry_return( ei->bei_e );
-#endif
-			ei->bei_e = NULL;
-			--bdb->bi_cache.c_cursize;
-		}
-		bdb_cache_entry_db_unlock( bdb->bi_dbenv, lock );
-	}
-	ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
-}