X-Git-Url: https://git.sur5r.net/?a=blobdiff_plain;f=servers%2Fslapd%2Fback-bdb%2Fcache.c;h=6ca332131c9850f05e49984cdb3cd940d95717ec;hb=8f02c3f57578f6e35e89fcf73574c34108b41195;hp=f68f89e4f01d4759183ca5646ac47ed7049ccb9a;hpb=f696f7ab55cbd526bc6e008aa3c89f41bf2c8fb9;p=openldap

diff --git a/servers/slapd/back-bdb/cache.c b/servers/slapd/back-bdb/cache.c
index f68f89e4f0..6ca332131c 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 2000-2004 The OpenLDAP Foundation.
+ * Copyright 2000-2011 The OpenLDAP Foundation.
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
@@ -26,9 +26,32 @@
 
 #include "back-bdb.h"
 
-static int	bdb_cache_delete_internal(Cache *cache, EntryInfo *e);
+#include "ldap_rq.h"
+
+#ifdef BDB_HIER
+#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
+#define SLAPD_UNUSED
+#ifdef SLAPD_UNUSED
 static void	bdb_lru_print(Cache *cache);
+static void	bdb_idtree_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 *
@@ -37,35 +60,123 @@ 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;
+			ei->bei_finders = 0;
+			ei->bei_lrunext = NULL;
 		}
-		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;
 }
 
+static void
+bdb_cache_entryinfo_free( Cache *cache, EntryInfo *ei )
+{
+	free( ei->bei_nrdn.bv_val );
+	BER_BVZERO( &ei->bei_nrdn );
+#ifdef BDB_HIER
+	free( ei->bei_rdn.bv_val );
+	BER_BVZERO( &ei->bei_rdn );
+	ei->bei_modrdns = 0;
+	ei->bei_ckids = 0;
+	ei->bei_dkids = 0;
+#endif
+	ei->bei_parent = NULL;
+	ei->bei_kids = NULL;
+	ei->bei_lruprev = NULL;
+
+#if 0
+	ldap_pvt_thread_mutex_lock( &cache->c_eifree_mutex );
+	ei->bei_lrunext = cache->c_eifree;
+	cache->c_eifree = ei;
+	ldap_pvt_thread_mutex_unlock( &cache->c_eifree_mutex );
+#else
+	ldap_pvt_thread_mutex_destroy( &ei->bei_kids_mutex );
+	ch_free( ei );
+#endif
+}
+
+#define LRU_DEL( c, e ) do { \
+	if ( e == e->bei_lruprev ) { \
+		(c)->c_lruhead = (c)->c_lrutail = NULL; \
+	} else { \
+		if ( e == (c)->c_lruhead ) (c)->c_lruhead = e->bei_lruprev; \
+		if ( e == (c)->c_lrutail ) (c)->c_lrutail = e->bei_lruprev; \
+		e->bei_lrunext->bei_lruprev = e->bei_lruprev; \
+		e->bei_lruprev->bei_lrunext = e->bei_lrunext; \
+	} \
+	e->bei_lruprev = NULL; \
+} while ( 0 )
+
+/* 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( struct bdb_info *bdb, EntryInfo *ei )
+{
+
+	/* Already linked, ignore */
+	if ( ei->bei_lruprev )
+		return;
+
+	/* Insert into circular LRU list */
+	ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_lru_mutex );
+
+	ei->bei_lruprev = bdb->bi_cache.c_lrutail;
+	if ( bdb->bi_cache.c_lrutail ) {
+		ei->bei_lrunext = bdb->bi_cache.c_lrutail->bei_lrunext;
+		bdb->bi_cache.c_lrutail->bei_lrunext = ei;
+		if ( ei->bei_lrunext )
+			ei->bei_lrunext->bei_lruprev = ei;
+	} else {
+		ei->bei_lrunext = ei->bei_lruprev = ei;
+		bdb->bi_cache.c_lruhead = ei;
+	}
+	bdb->bi_cache.c_lrutail = ei;
+	ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_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,
-	u_int32_t locker,
+	struct bdb_info *bdb,
+	DB_TXN *txn,
 	EntryInfo *ei,
 	int rw,
 	int tryOnly,
 	DB_LOCK *lock )
 {
-#ifdef NO_THREADS
+#ifdef NO_DB_LOCK
 	return 0;
 #else
 	int	rc;
@@ -74,6 +185,7 @@ bdb_cache_entry_db_relock(
 
 	if ( !lock ) return 0;
 
+	DBTzero( &lockobj );
 	lockobj.data = &ei->bei_id;
 	lockobj.size = sizeof(ei->bei_id) + 1;
 
@@ -83,19 +195,13 @@ 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, TXN_ID(txn), tryOnly ? DB_LOCK_NOWAIT : 0,
 		list, 2, NULL );
 
-	if (rc) {
-#ifdef NEW_LOGGING
-		LDAP_LOG( CACHE, DETAIL1, 
-			"bdb_cache_entry_db_relock: entry %ld, rw %d, rc %d\n",
-			ei->bei_id, rw, rc );
-#else
+	if (rc && !tryOnly) {
 		Debug( LDAP_DEBUG_TRACE,
 			"bdb_cache_entry_db_relock: entry %ld, rw %d, rc %d\n",
 			ei->bei_id, rw, rc );
-#endif
 	} else {
 		*lock = list[1].lock;
 	}
@@ -104,10 +210,10 @@ 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, DB_TXN *txn, EntryInfo *ei,
 	int rw, int tryOnly, DB_LOCK *lock )
 {
-#ifdef NO_THREADS
+#ifdef NO_DB_LOCK
 	return 0;
 #else
 	int       rc;
@@ -121,41 +227,70 @@ bdb_cache_entry_db_lock( DB_ENV *env, u_int32_t locker, EntryInfo *ei,
 	else
 		db_rw = DB_LOCK_READ;
 
+	DBTzero( &lockobj );
 	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, TXN_ID(txn), tryOnly ? DB_LOCK_NOWAIT : 0,
 					&lockobj, db_rw, lock);
-	if (rc) {
-#ifdef NEW_LOGGING
-		LDAP_LOG( CACHE, DETAIL1, 
-			"bdb_cache_entry_db_lock: entry %ld, rw %d, rc %d\n",
-			ei->bei_id, rw, rc );
-#else
+	if (rc && !tryOnly) {
 		Debug( LDAP_DEBUG_TRACE,
 			"bdb_cache_entry_db_lock: entry %ld, rw %d, rc %d\n",
 			ei->bei_id, rw, rc );
-#endif
 	}
 	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
 }
 
+void
+bdb_cache_return_entry_rw( struct bdb_info *bdb, Entry *e,
+	int rw, DB_LOCK *lock )
+{
+	EntryInfo *ei;
+	int free = 0;
+
+	ei = e->e_private;
+	if ( ei && ( ei->bei_state & CACHE_ENTRY_NOT_CACHED )) {
+		bdb_cache_entryinfo_lock( ei );
+		if ( ei->bei_state & CACHE_ENTRY_NOT_CACHED ) {
+			/* Releasing the entry can only be done when
+			 * we know that nobody else is using it, i.e we
+			 * should have an entry_db writelock.  But the
+			 * flag is only set by the thread that loads the
+			 * entry, and only if no other threads has found
+			 * it while it was working.  All other threads
+			 * clear the flag, which mean that we should be
+			 * the only thread using the entry if the flag
+			 * is set here.
+			 */
+			ei->bei_e = NULL;
+			ei->bei_state ^= CACHE_ENTRY_NOT_CACHED;
+			free = 1;
+		}
+		bdb_cache_entryinfo_unlock( ei );
+	}
+	bdb_cache_entry_db_unlock( bdb, lock );
+	if ( free ) {
+		e->e_private = NULL;
+		bdb_entry_return( e );
+	}
+}
+
 static int
 bdb_cache_entryinfo_destroy( EntryInfo *e )
 {
@@ -168,31 +303,6 @@ 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; \
-	} \
-} 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 )
@@ -213,6 +323,14 @@ bdb_id_cmp( const void *v_e1, const void *v_e2 )
 	return e1->bei_id - e2->bei_id;
 }
 
+static int
+bdb_id_dup_err( void *v1, void *v2 )
+{
+	EntryInfo *e2 = v2;
+	e2->bei_lrunext = v1;
+	return -1;
+}
+
 /* Create an entryinfo in the cache. Caller must release the locks later.
  */
 static int
@@ -227,20 +345,23 @@ bdb_entryinfo_add_internal(
 
 	ei2 = bdb_cache_entryinfo_new( &bdb->bi_cache );
 
-	ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
 	bdb_cache_entryinfo_lock( ei->bei_parent );
+	ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
 
 	ei2->bei_id = ei->bei_id;
 	ei2->bei_parent = ei->bei_parent;
 #ifdef BDB_HIER
 	ei2->bei_rdn = ei->bei_rdn;
 #endif
+#ifdef SLAP_ZONE_ALLOC
+	ei2->bei_bdb = bdb;
+#endif
 
 	/* Add to cache ID tree */
-	if (avl_insert( &bdb->bi_cache.c_idtree, ei2, bdb_id_cmp, avl_dup_error )) {
-		EntryInfo *eix;
-		eix = avl_find( bdb->bi_cache.c_idtree, ei2, bdb_id_cmp );
-		bdb_cache_entryinfo_destroy( ei2 );
+	if (avl_insert( &bdb->bi_cache.c_idtree, ei2, bdb_id_cmp,
+		bdb_id_dup_err )) {
+		EntryInfo *eix = ei2->bei_lrunext;
+		bdb_cache_entryinfo_free( &bdb->bi_cache, ei2 );
 		ei2 = eix;
 #ifdef BDB_HIER
 		/* It got freed above because its value was
@@ -249,11 +370,24 @@ bdb_entryinfo_add_internal(
 		ei->bei_rdn.bv_val = NULL;
 #endif
 	} else {
+		int rc;
+
+		bdb->bi_cache.c_eiused++;
 		ber_dupbv( &ei2->bei_nrdn, &ei->bei_nrdn );
-		avl_insert( &ei->bei_parent->bei_kids, ei2, bdb_rdn_cmp,
+
+		/* This is a new leaf node. But if parent had no kids, then it was
+		 * a leaf and we would be decrementing that. So, only increment if
+		 * the parent already has kids.
+		 */
+		if ( ei->bei_parent->bei_kids || !ei->bei_parent->bei_id )
+			bdb->bi_cache.c_leaves++;
+		rc = avl_insert( &ei->bei_parent->bei_kids, ei2, bdb_rdn_cmp,
 			avl_dup_error );
 #ifdef BDB_HIER
-		ei->bei_parent->bei_ckids++;
+		/* it's possible for hdb_cache_find_parent to beat us to it */
+		if ( !rc ) {
+			ei->bei_parent->bei_ckids++;
+		}
 #endif
 	}
 
@@ -304,35 +438,57 @@ 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 ) {
+			DBC *cursor;
 			int len = ei.bei_nrdn.bv_len;
 				
+			if ( BER_BVISEMPTY( ndn )) {
+				*res = eip;
+				return LDAP_SUCCESS;
+			}
+
 			ei.bei_nrdn.bv_len = ndn->bv_len -
 				(ei.bei_nrdn.bv_val - ndn->bv_val);
+			eip->bei_finders++;
 			bdb_cache_entryinfo_unlock( eip );
 
-			rc = bdb_dn2id( op, txn, &ei.bei_nrdn, &ei );
+			BDB_LOG_PRINTF( bdb->bi_dbenv, NULL, "slapd Reading %s",
+				ei.bei_nrdn.bv_val );
+
+			cursor = NULL;
+			rc = bdb_dn2id( op, &ei.bei_nrdn, &ei, txn, &cursor );
 			if (rc) {
 				bdb_cache_entryinfo_lock( eip );
+				eip->bei_finders--;
+				if ( cursor ) cursor->c_close( cursor );
 				*res = eip;
 				return rc;
 			}
 
+			BDB_LOG_PRINTF( bdb->bi_dbenv, NULL, "slapd Read got %s(%d)",
+				ei.bei_nrdn.bv_val, ei.bei_id );
+
 			/* DN exists but needs to be added to cache */
 			ei.bei_nrdn.bv_len = len;
 			rc = bdb_entryinfo_add_internal( bdb, &ei, &ei2 );
 			/* add_internal left eip and c_rwlock locked */
+			eip->bei_finders--;
 			ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
+			if ( cursor ) cursor->c_close( cursor );
 			if ( rc ) {
 				*res = eip;
 				return rc;
 			}
-		} else if ( ei2->bei_state & CACHE_ENTRY_DELETED ) {
+		}
+		bdb_cache_entryinfo_lock( ei2 );
+		if ( ei2->bei_state & CACHE_ENTRY_DELETED ) {
 			/* In the midst of deleting? Give it a chance to
 			 * complete.
 			 */
+			bdb_cache_entryinfo_unlock( ei2 );
 			bdb_cache_entryinfo_unlock( eip );
 			ldap_pvt_thread_yield();
 			bdb_cache_entryinfo_lock( eip );
@@ -340,7 +496,6 @@ bdb_cache_find_ndn(
 			return DB_NOTFOUND;
 		}
 		bdb_cache_entryinfo_unlock( eip );
-		bdb_cache_entryinfo_lock( ei2 );
 
 		eip = ei2;
 
@@ -365,22 +520,20 @@ bdb_cache_find_ndn(
 /* Walk up the tree from a child node, looking for an ID that's already
  * been linked into the cache.
  */
-static int
+int
 hdb_cache_find_parent(
 	Operation *op,
-	DB_TXN *txn,
+	DB_TXN	*txn,
 	ID id,
 	EntryInfo **res )
 {
 	struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
 	EntryInfo ei, eip, *ei2 = NULL, *ein = NULL, *eir = NULL;
-	char ndn[SLAP_LDAPDN_MAXLEN];
-	ID parent;
-	int rc;
-	int addlru = 1;
+	int rc, add;
 
 	ei.bei_id = id;
 	ei.bei_kids = NULL;
+	ei.bei_ckids = 0;
 
 	for (;;) {
 		rc = hdb_dn2id_parent( op, txn, &ei, &eip.bei_id );
@@ -395,69 +548,104 @@ hdb_cache_find_parent(
 		ein->bei_kids = ei.bei_kids;
 		ein->bei_nrdn = ei.bei_nrdn;
 		ein->bei_rdn = ei.bei_rdn;
+		ein->bei_ckids = ei.bei_ckids;
+#ifdef SLAP_ZONE_ALLOC
+		ein->bei_bdb = bdb;
+#endif
+		ei.bei_ckids = 0;
+		add = 1;
 		
 		/* This node is not fully connected yet */
-		ein->bei_state = CACHE_ENTRY_NOT_LINKED;
+		ein->bei_state |= CACHE_ENTRY_NOT_LINKED;
+
+		/* If this is the first time, save this node
+		 * to be returned later.
+		 */
+		if ( eir == NULL ) {
+			eir = ein;
+			ein->bei_finders++;
+		}
 
+again:
 		/* Insert this node into the ID tree */
 		ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
 		if ( avl_insert( &bdb->bi_cache.c_idtree, (caddr_t)ein,
-			bdb_id_cmp, avl_dup_error ) ) {
+			bdb_id_cmp, bdb_id_dup_err ) ) {
+			EntryInfo *eix = ein->bei_lrunext;
+
+			if ( bdb_cache_entryinfo_trylock( eix )) {
+				ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
+				ldap_pvt_thread_yield();
+				goto again;
+			}
+			ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
 
 			/* Someone else created this node just before us.
 			 * Free our new copy and use the existing one.
 			 */
-			bdb_cache_entryinfo_destroy( ein );
-			ein = (EntryInfo *)avl_find( bdb->bi_cache.c_idtree,
-				(caddr_t) &ei, bdb_id_cmp );
-			
-			/* Link in any kids we've already processed */
-			if ( ei2 ) {
-				bdb_cache_entryinfo_lock( ein );
-				avl_insert( &ein->bei_kids, (caddr_t)ei2,
-					bdb_rdn_cmp, avl_dup_error );
-				bdb_cache_entryinfo_unlock( ein );
-			}
+			bdb_cache_entryinfo_free( &bdb->bi_cache, ein );
 
-			if ( !eir ) {
-				addlru = 0;
+			/* if it was the node we were looking for, just return it */
+			if ( eir == ein ) {
+				*res = eix;
+				rc = 0;
+				break;
 			}
-		}
 
-		/* If this is the first time, save this node
-		 * to be returned later.
-		 */
-		if ( eir == NULL ) eir = ein;
+			ein = ei2;
+			ei2 = eix;
+			add = 0;
+
+			/* otherwise, link up what we have and return */
+			goto gotparent;
+		}
 
 		/* If there was a previous node, link it to this one */
 		if ( ei2 ) ei2->bei_parent = ein;
 
 		/* Look for this node's parent */
+par2:
 		if ( eip.bei_id ) {
 			ei2 = (EntryInfo *) avl_find( bdb->bi_cache.c_idtree,
 					(caddr_t) &eip, bdb_id_cmp );
 		} else {
 			ei2 = &bdb->bi_cache.c_dntree;
 		}
+		if ( ei2 && bdb_cache_entryinfo_trylock( ei2 )) {
+			ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
+			ldap_pvt_thread_yield();
+			ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
+			goto par2;
+		}
+		if ( add )
+			bdb->bi_cache.c_eiused++;
+		if ( ei2 && ( ei2->bei_kids || !ei2->bei_id ))
+			bdb->bi_cache.c_leaves++;
 		ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
 
+gotparent:
 		/* Got the parent, link in and we're done. */
 		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);
-			bdb_cache_entryinfo_unlock( ei2 );
 			bdb_cache_entryinfo_lock( eir );
+			ein->bei_parent = ei2;
+
+			if ( avl_insert( &ei2->bei_kids, (caddr_t)ein, bdb_rdn_cmp,
+				avl_dup_error) == 0 )
+				ei2->bei_ckids++;
 
 			/* 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 );
+			eir->bei_finders--;
+
 			*res = eir;
 			break;
 		}
 		ei.bei_kids = NULL;
 		ei.bei_id = eip.bei_id;
+		ei.bei_ckids = 1;
 		avl_insert( &ei.bei_kids, (caddr_t)ein, bdb_rdn_cmp,
 			avl_dup_error );
 	}
@@ -500,82 +688,177 @@ int hdb_cache_load(
 }
 #endif
 
-/* caller must have lru_mutex locked. mutex
- * will be unlocked on return.
+/* This is best-effort only. If all entries in the cache are
+ * busy, they will all be kept. This is unlikely to happen
+ * unless the cache is very much smaller than the working set.
  */
 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 = NULL;
+	int islocked;
+	ID eicount, ecount;
+	ID count, efree, eifree = 0;
+#ifdef LDAP_DEBUG
+	int iter;
+#endif
+
+	/* Wait for the mutex; we're the only one trying to purge. */
+	ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_lru_mutex );
+
+	if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize ) {
+		efree = bdb->bi_cache.c_cursize - bdb->bi_cache.c_maxsize;
+		efree += bdb->bi_cache.c_minfree;
+	} else {
+		efree = 0;
+	}
+
+	/* maximum number of EntryInfo leaves to cache. In slapcat
+	 * we always free all leaf nodes.
+	 */
 
-	if ( locker ) {
+	if ( slapMode & SLAP_TOOL_READONLY ) {
+		eifree = bdb->bi_cache.c_leaves;
+	} else if ( bdb->bi_cache.c_eimax &&
+		bdb->bi_cache.c_leaves > bdb->bi_cache.c_eimax ) {
+		eifree = bdb->bi_cache.c_minfree * 10;
+		if ( eifree >= bdb->bi_cache.c_leaves )
+			eifree /= 2;
+	}
+
+	if ( !efree && !eifree ) {
+		ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
+		bdb->bi_cache.c_purging = 0;
+		return;
+	}
+
+	if ( bdb->bi_cache.c_txn ) {
 		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;
+	eicount = 0;
+	ecount = 0;
+#ifdef LDAP_DEBUG
+	iter = 0;
+#endif
 
-		/* Look for an unused entry to remove */
-		for (elru = bdb->bi_cache.c_lrutail; elru; elru = elprev, i++ ) {
-			elprev = elru->bei_lruprev;
+	/* Look for an unused entry to remove */
+	for ( elru = bdb->bi_cache.c_lruhead; elru; elru = elnext ) {
+		elnext = elru->bei_lrunext;
 
-			/* Too many probes, not enough idle, give up */
-			if (i > 10) break;
+		if ( bdb_cache_entryinfo_trylock( elru ))
+			goto bottom;
 
-			/* 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 there's no entry, or this node is in
-				 * the process of linking into the cache,
-				 * skip it.
+		/* 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 );
+			goto bottom;
+		}
+
+		/* 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 | CACHE_ENTRY_LOADING |
+			CACHE_ENTRY_ONELEVEL )) ||
+			elru->bei_finders > 0 ) {
+			bdb_cache_entryinfo_unlock( elru );
+			goto bottom;
+		}
+
+		if ( bdb_cache_entryinfo_trylock( elru->bei_parent )) {
+			bdb_cache_entryinfo_unlock( elru );
+			goto bottom;
+		}
+
+		/* entryinfo is locked */
+		islocked = 1;
+
+		/* If we can successfully writelock it, then
+		 * the object is idle.
+		 */
+		if ( bdb_cache_entry_db_lock( bdb,
+			bdb->bi_cache.c_txn, elru, 1, 1, lockp ) == 0 ) {
+
+			/* Free entry for this node if it's present */
+			if ( elru->bei_e ) {
+				ecount++;
+
+				/* the cache may have gone over the limit while we
+				 * weren't looking, so double check.
 				 */
-				if ( !elru->bei_e || (elru->bei_state & CACHE_ENTRY_NOT_LINKED) ) {
-					bdb_cache_entry_db_unlock( bdb->bi_dbenv, lockp );
-					continue;
+				if ( !efree && ecount > bdb->bi_cache.c_maxsize )
+					efree = bdb->bi_cache.c_minfree;
+
+				if ( count < efree ) {
+					elru->bei_e->e_private = NULL;
+#ifdef SLAP_ZONE_ALLOC
+					bdb_entry_return( bdb, elru->bei_e, elru->bei_zseq );
+#else
+					bdb_entry_return( elru->bei_e );
+#endif
+					elru->bei_e = NULL;
+					count++;
+				} else {
+					/* Keep this node cached, skip to next */
+					bdb_cache_entry_db_unlock( bdb, lockp );
+					goto next;
 				}
-				LRU_DELETE( &bdb->bi_cache, elru );
-				elru->bei_e->e_private = NULL;
-				bdb_entry_return( elru->bei_e );
-				elru->bei_e = NULL;
-				bdb_cache_entry_db_unlock( bdb->bi_dbenv, lockp );
-				--bdb->bi_cache.c_cursize;
-				if (bdb->bi_cache.c_cursize < bdb->bi_cache.c_maxsize)
-					break;
 			}
+			bdb_cache_entry_db_unlock( bdb, lockp );
+
+			/* 
+			 * If it is a leaf node, and we're over the limit, free it.
+			 */
+			if ( elru->bei_kids ) {
+				/* Drop from list, we ignore it... */
+				LRU_DEL( &bdb->bi_cache, elru );
+			} else if ( eicount < eifree ) {
+				/* Too many leaf nodes, free this one */
+				bdb_cache_delete_internal( &bdb->bi_cache, elru, 0 );
+				bdb_cache_delete_cleanup( &bdb->bi_cache, elru );
+				islocked = 0;
+				eicount++;
+			}	/* Leave on list until we need to free it */
 		}
-	}
-	LRU_ADD( &bdb->bi_cache, ei );
-	ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_mutex );
-}
 
-EntryInfo *
-bdb_cache_find_info(
-	struct bdb_info *bdb,
-	ID id )
-{
-	EntryInfo ei, *ei2;
+next:
+		if ( islocked ) {
+			bdb_cache_entryinfo_unlock( elru );
+			bdb_cache_entryinfo_unlock( elru->bei_parent );
+		}
 
-	ei.bei_id = id;
+		if ( count >= efree && eicount >= eifree )
+			break;
+bottom:
+		if ( elnext == bdb->bi_cache.c_lruhead )
+			break;
+#ifdef LDAP_DEBUG
+		iter++;
+#endif
+	}
 
-	ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
-	ei2 = (EntryInfo *) avl_find( bdb->bi_cache.c_idtree,
-					(caddr_t) &ei, bdb_id_cmp );
-	ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
-	return ei2;
+	if ( count || ecount > bdb->bi_cache.c_cursize ) {
+		ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
+		/* HACK: we seem to be losing track, fix up now */
+		if ( ecount > bdb->bi_cache.c_cursize )
+			bdb->bi_cache.c_cursize = ecount;
+		bdb->bi_cache.c_cursize -= count;
+		ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
+	}
+	bdb->bi_cache.c_lruhead = elnext;
+	ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
+	bdb->bi_cache.c_purging = 0;
 }
 
 /*
  * cache_find_id - find an entry in the cache, given id.
- * The entry is locked for Read upon return. Call with islocked TRUE if
+ * The entry is locked for Read upon return. Call with flag ID_LOCKED if
  * the supplied *eip was already locked.
  */
 
@@ -585,17 +868,19 @@ bdb_cache_find_id(
 	DB_TXN	*tid,
 	ID				id,
 	EntryInfo	**eip,
-	int		islocked,
-	u_int32_t	locker,
+	int		flag,
 	DB_LOCK		*lock )
 {
 	struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
 	Entry	*ep = NULL;
-	int	rc = 0;
-	EntryInfo ei;
+	int	rc = 0, load = 0;
+	EntryInfo ei = { 0 };
 
 	ei.bei_id = id;
 
+#ifdef SLAP_ZONE_ALLOC
+	slap_zh_rlock(bdb->bi_cache.c_zctx);
+#endif
 	/* If we weren't given any info, see if we have it already cached */
 	if ( !*eip ) {
 again:	ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
@@ -603,13 +888,13 @@ again:	ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
 			(caddr_t) &ei, bdb_id_cmp );
 		if ( *eip ) {
 			/* If the lock attempt fails, the info is in use */
-			if ( ldap_pvt_thread_mutex_trylock(
-					&(*eip)->bei_kids_mutex )) {
+			if ( bdb_cache_entryinfo_trylock( *eip )) {
+				int del = (*eip)->bei_state & CACHE_ENTRY_DELETED;
 				ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
 				/* If this node is being deleted, treat
 				 * as if the delete has already finished
 				 */
-				if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
+				if ( del ) {
 					return DB_NOTFOUND;
 				}
 				/* otherwise, wait for the info to free up */
@@ -625,7 +910,7 @@ again:	ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
 				ldap_pvt_thread_yield();
 				goto again;
 			}
-			islocked = 1;
+			flag |= ID_LOCKED;
 		}
 		ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
 	}
@@ -637,93 +922,185 @@ again:	ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
 		if ( rc == 0 ) {
 			rc = bdb_cache_find_ndn( op, tid,
 				&ep->e_nname, eip );
-			if ( *eip ) islocked = 1;
+			if ( *eip ) flag |= ID_LOCKED;
 			if ( rc ) {
+				ep->e_private = NULL;
+#ifdef SLAP_ZONE_ALLOC
+				bdb_entry_return( bdb, ep, (*eip)->bei_zseq );
+#else
 				bdb_entry_return( ep );
+#endif
 				ep = NULL;
 			}
 		}
 #else
 		rc = hdb_cache_find_parent(op, tid, id, eip );
-		if ( rc == 0 && *eip ) islocked = 1;
+		if ( rc == 0 ) flag |= ID_LOCKED;
 #endif
 	}
 
 	/* Ok, we found the info, do we have the entry? */
-	if ( *eip && rc == 0 ) {
+	if ( rc == 0 ) {
+		if ( !( flag & ID_LOCKED )) {
+			bdb_cache_entryinfo_lock( *eip );
+			flag |= ID_LOCKED;
+		}
+
 		if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
 			rc = DB_NOTFOUND;
 		} else {
-			rc = bdb_cache_entry_db_lock( bdb->bi_dbenv, locker, *eip, 0, 0, lock );
-			/* entry is protected now, we don't need to hold the entryinfo */
-			if ( islocked ) {
+			(*eip)->bei_finders++;
+			(*eip)->bei_state |= CACHE_ENTRY_REFERENCED;
+			if ( flag & ID_NOENTRY ) {
 				bdb_cache_entryinfo_unlock( *eip );
-				islocked = 0;
+				return 0;
 			}
-			if ( rc == 0 ) {
-				if ( !(*eip)->bei_e ) {
-					if (!ep) {
+			/* Make sure only one thread tries to load the entry */
+load1:
+#ifdef SLAP_ZONE_ALLOC
+			if ((*eip)->bei_e && !slap_zn_validate(
+					bdb->bi_cache.c_zctx, (*eip)->bei_e, (*eip)->bei_zseq)) {
+				(*eip)->bei_e = NULL;
+				(*eip)->bei_zseq = 0;
+			}
+#endif
+			if ( !(*eip)->bei_e && !((*eip)->bei_state & CACHE_ENTRY_LOADING)) {
+				load = 1;
+				(*eip)->bei_state |= CACHE_ENTRY_LOADING;
+				flag |= ID_CHKPURGE;
+			}
+
+			if ( !load ) {
+				/* Clear the uncached state if we are not
+				 * loading it, i.e it is already cached or
+				 * another thread is currently loading it.
+				 */
+				if ( (*eip)->bei_state & CACHE_ENTRY_NOT_CACHED ) {
+					(*eip)->bei_state ^= CACHE_ENTRY_NOT_CACHED;
+					flag |= ID_CHKPURGE;
+				}
+			}
+
+			if ( flag & ID_LOCKED ) {
+				bdb_cache_entryinfo_unlock( *eip );
+				flag ^= ID_LOCKED;
+			}
+			rc = bdb_cache_entry_db_lock( bdb, tid, *eip, load, 0, lock );
+			if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
+				rc = DB_NOTFOUND;
+				bdb_cache_entry_db_unlock( bdb, lock );
+				bdb_cache_entryinfo_lock( *eip );
+				(*eip)->bei_finders--;
+				bdb_cache_entryinfo_unlock( *eip );
+			} else if ( rc == 0 ) {
+				if ( load ) {
+					if ( !ep) {
 						rc = bdb_id2entry( op->o_bd, tid, id, &ep );
 					}
 					if ( rc == 0 ) {
-						bdb_cache_entry_db_relock( bdb->bi_dbenv, locker,
-							*eip, 1, 0, lock );
-						/* Make sure no other modifier beat us to it */
-						if ( (*eip)->bei_e ) {
-							bdb_entry_return( ep );
-							ep = NULL;
+						ep->e_private = *eip;
 #ifdef BDB_HIER
-							/* Check for subtree renames */
-							rc = bdb_fix_dn( (*eip)->bei_e, 1 );
-							if ( rc ) rc = bdb_fix_dn( (*eip)->bei_e, 2 );
+						while ( (*eip)->bei_state & CACHE_ENTRY_NOT_LINKED )
+							ldap_pvt_thread_yield();
+						bdb_fix_dn( ep, 0 );
 #endif
-						} else {
-							ep->e_private = *eip;
-#ifdef BDB_HIER
-							bdb_fix_dn( ep, 0 );
+						bdb_cache_entryinfo_lock( *eip );
+
+						(*eip)->bei_e = ep;
+#ifdef SLAP_ZONE_ALLOC
+						(*eip)->bei_zseq = *((ber_len_t *)ep - 2);
 #endif
-							(*eip)->bei_e = ep;
+						ep = NULL;
+						if ( flag & ID_NOCACHE ) {
+							/* Set the cached state only if no other thread
+							 * found the info while we were loading the entry.
+							 */
+							if ( (*eip)->bei_finders == 1 ) {
+								(*eip)->bei_state |= CACHE_ENTRY_NOT_CACHED;
+								flag ^= ID_CHKPURGE;
+							}
 						}
-						bdb_cache_entry_db_relock( bdb->bi_dbenv, locker,
+						bdb_cache_entryinfo_unlock( *eip );
+						bdb_cache_lru_link( bdb, *eip );
+					}
+					if ( rc == 0 ) {
+						/* If we succeeded, downgrade back to a readlock. */
+						rc = bdb_cache_entry_db_relock( bdb, tid,
 							*eip, 0, 0, lock );
+					} else {
+						/* Otherwise, release the lock. */
+						bdb_cache_entry_db_unlock( bdb, lock );
 					}
+				} else if ( !(*eip)->bei_e ) {
+					/* Some other thread is trying to load the entry,
+					 * wait for it to finish.
+					 */
+					bdb_cache_entry_db_unlock( bdb, lock );
+					bdb_cache_entryinfo_lock( *eip );
+					flag |= ID_LOCKED;
+					goto load1;
 #ifdef BDB_HIER
 				} else {
 					/* Check for subtree renames
 					 */
 					rc = bdb_fix_dn( (*eip)->bei_e, 1 );
 					if ( rc ) {
-						bdb_cache_entry_db_relock( bdb->bi_dbenv,
-							locker, *eip, 1, 0, lock );
+						bdb_cache_entry_db_relock( bdb,
+							tid, *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,
-							locker, *eip, 0, 0, lock );
+						bdb_cache_entry_db_relock( bdb,
+							tid, *eip, 0, 0, lock );
 					}
 #endif
 				}
-
+				bdb_cache_entryinfo_lock( *eip );
+				(*eip)->bei_finders--;
+				if ( load )
+					(*eip)->bei_state ^= CACHE_ENTRY_LOADING;
+				bdb_cache_entryinfo_unlock( *eip );
 			}
 		}
 	}
+	if ( flag & ID_LOCKED ) {
+		bdb_cache_entryinfo_unlock( *eip );
+	}
+	if ( ep ) {
+		ep->e_private = NULL;
+#ifdef SLAP_ZONE_ALLOC
+		bdb_entry_return( bdb, ep, (*eip)->bei_zseq );
+#else
+		bdb_entry_return( ep );
+#endif
+	}
 	if ( rc == 0 ) {
-		/* set lru mutex */
-		ldap_pvt_thread_mutex_lock( &bdb->bi_cache.lru_mutex );
-		/* if entry is on LRU list, remove from old spot */
-		if ( (*eip)->bei_lrunext || (*eip)->bei_lruprev ) {
-			LRU_DELETE( &bdb->bi_cache, *eip );
-		} else {
-		/* if entry is new, bump cache size */
-			bdb->bi_cache.c_cursize++;
+		int purge = 0;
+
+		if (( flag & ID_CHKPURGE ) || bdb->bi_cache.c_eimax ) {
+			ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
+			if ( flag & ID_CHKPURGE ) {
+				bdb->bi_cache.c_cursize++;
+				if ( !bdb->bi_cache.c_purging && bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize ) {
+					purge = 1;
+					bdb->bi_cache.c_purging = 1;
+				}
+			} else if ( !bdb->bi_cache.c_purging && bdb->bi_cache.c_eimax && bdb->bi_cache.c_leaves > bdb->bi_cache.c_eimax ) {
+				purge = 1;
+				bdb->bi_cache.c_purging = 1;
+			}
+			ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
 		}
-		/* lru_mutex is unlocked for us */
-		bdb_cache_lru_add( bdb, locker, *eip );
+		if ( purge )
+			bdb_cache_lru_purge( bdb );
 	}
 
-	if ( islocked ) {
-		bdb_cache_entryinfo_unlock( *eip );
+#ifdef SLAP_ZONE_ALLOC
+	if (rc == 0 && (*eip)->bei_e) {
+		slap_zn_rlock(bdb->bi_cache.c_zctx, (*eip)->bei_e);
 	}
+	slap_zh_runlock(bdb->bi_cache.c_zctx);
+#endif
 	return rc;
 }
 
@@ -755,73 +1132,103 @@ bdb_cache_add(
 	EntryInfo *eip,
 	Entry *e,
 	struct berval *nrdn,
-	u_int32_t locker )
+	DB_TXN *txn,
+	DB_LOCK *lock )
 {
 	EntryInfo *new, ei;
+	int rc, purge = 0;
+#ifdef BDB_HIER
 	struct berval rdn = e->e_name;
-	DB_LOCK lock;
-	int rc;
+#endif
 
 	ei.bei_id = e->e_id;
 	ei.bei_parent = eip;
 	ei.bei_nrdn = *nrdn;
 	ei.bei_lockpad = 0;
 
+#if 0
 	/* 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, txn, &ei, 0, 0, lock );
 	if ( rc ) {
 		bdb_cache_entryinfo_unlock( eip );
 		return rc;
 	}
+#endif
 
 #ifdef BDB_HIER
 	if ( nrdn->bv_len != e->e_nname.bv_len ) {
-		char *ptr = strchr( rdn.bv_val, ',' );
+		char *ptr = ber_bvchr( &rdn, ',' );
+		assert( ptr != NULL );
 		rdn.bv_len = ptr - rdn.bv_val;
 	}
 	ber_dupbv( &ei.bei_rdn, &rdn );
 	if ( eip->bei_dkids ) eip->bei_dkids++;
 #endif
 
+	if (eip->bei_parent) {
+		bdb_cache_entryinfo_lock( eip->bei_parent );
+		eip->bei_parent->bei_state &= ~CACHE_ENTRY_NO_GRANDKIDS;
+		bdb_cache_entryinfo_unlock( eip->bei_parent );
+	}
+
 	rc = bdb_entryinfo_add_internal( bdb, &ei, &new );
 	/* bdb_csn_commit can cause this when adding the database root entry */
 	if ( new->bei_e ) {
 		new->bei_e->e_private = NULL;
+#ifdef SLAP_ZONE_ALLOC
+		bdb_entry_return( bdb, new->bei_e, new->bei_zseq );
+#else
 		bdb_entry_return( new->bei_e );
+#endif
 	}
 	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 );
 
-	/* set lru mutex */
-	ldap_pvt_thread_mutex_lock( &bdb->bi_cache.lru_mutex );
+	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;
-	/* lru_mutex is unlocked for us */
-	bdb_cache_lru_add( bdb, locker, new );
+	if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize &&
+		!bdb->bi_cache.c_purging ) {
+		purge = 1;
+		bdb->bi_cache.c_purging = 1;
+	}
+	ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
+
+	new->bei_finders = 1;
+	bdb_cache_lru_link( bdb, new );
+
+	if ( purge )
+		bdb_cache_lru_purge( bdb );
 
-	bdb_cache_entryinfo_unlock( eip );
-	ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
 	return rc;
 }
 
+void bdb_cache_deref(
+	EntryInfo *ei
+	)
+{
+	bdb_cache_entryinfo_lock( ei );
+	ei->bei_finders--;
+	bdb_cache_entryinfo_unlock( ei );
+}
+
 int
 bdb_cache_modify(
+	struct bdb_info *bdb,
 	Entry *e,
 	Attribute *newAttrs,
-	DB_ENV *env,
-	u_int32_t locker,
+	DB_TXN *txn,
 	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, txn, 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.
@@ -840,20 +1247,22 @@ bdb_cache_modify(
  */
 int
 bdb_cache_modrdn(
+	struct bdb_info *bdb,
 	Entry *e,
 	struct berval *nrdn,
 	Entry *new,
 	EntryInfo *ein,
-	DB_ENV *env,
-	u_int32_t locker,
+	DB_TXN *txn,
 	DB_LOCK *lock )
 {
 	EntryInfo *ei = BEI(e), *pei;
-	struct berval rdn;
 	int rc;
+#ifdef BDB_HIER
+	struct berval rdn;
+#endif
 
 	/* Get write lock on data */
-	rc =  bdb_cache_entry_db_relock( env, locker, ei, 1, 0, lock );
+	rc =  bdb_cache_entry_db_relock( bdb, txn, ei, 1, 0, lock );
 	if ( rc ) return rc;
 
 	/* If we've done repeated mods on a cached entry, then e_attrs
@@ -878,15 +1287,27 @@ 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 );
+
 #ifdef BDB_HIER
 	free( ei->bei_rdn.bv_val );
 
 	rdn = e->e_name;
 	if ( nrdn->bv_len != e->e_nname.bv_len ) {
-		char *ptr = strchr(rdn.bv_val, ',');
+		char *ptr = ber_bvchr(&rdn, ',');
+		assert( ptr != NULL );
 		rdn.bv_len = ptr - rdn.bv_val;
 	}
 	ber_dupbv( &ei->bei_rdn, &rdn );
+
+	/* If new parent, decrement kid counts */
+	if ( ein ) {
+		pei->bei_ckids--;
+		if ( pei->bei_dkids ) {
+			pei->bei_dkids--;
+			if ( pei->bei_dkids < 2 )
+				pei->bei_state |= CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
+		}
+	}
 #endif
 
 	if (!ein) {
@@ -895,17 +1316,32 @@ bdb_cache_modrdn(
 		ei->bei_parent = ein;
 		bdb_cache_entryinfo_unlock( pei );
 		bdb_cache_entryinfo_lock( ein );
-	}
+
+		/* new parent now has kids */
+		if ( ein->bei_state & CACHE_ENTRY_NO_KIDS )
+			ein->bei_state ^= CACHE_ENTRY_NO_KIDS;
+		/* grandparent has grandkids */
+		if ( ein->bei_parent )
+			ein->bei_parent->bei_state &= ~CACHE_ENTRY_NO_GRANDKIDS;
 #ifdef BDB_HIER
-	{
-		int max = ei->bei_modrdns;
-		/* Record the generation number of this change */
-		for ( pei = ein; pei->bei_parent; pei = pei->bei_parent ) {
-			if ( pei->bei_modrdns > max ) max = pei->bei_modrdns;
-		}
-		ei->bei_modrdns = max + 1;
+		/* 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;
+
+		ein->bei_ckids++;
+		if ( ein->bei_dkids ) ein->bei_dkids++;
+#endif
 	}
+
+#ifdef BDB_HIER
+	/* Record the generation number of this change */
+	ldap_pvt_thread_mutex_lock( &bdb->bi_modrdns_mutex );
+	bdb->bi_modrdns++;
+	ei->bei_modrdns = bdb->bi_modrdns;
+	ldap_pvt_thread_mutex_unlock( &bdb->bi_modrdns_mutex );
 #endif
+
 	avl_insert( &ein->bei_kids, ei, bdb_rdn_cmp, avl_dup_error );
 	bdb_cache_entryinfo_unlock( ein );
 	return rc;
@@ -919,57 +1355,63 @@ bdb_cache_modrdn(
  */
 int
 bdb_cache_delete(
-    Cache	*cache,
+	struct bdb_info *bdb,
     Entry		*e,
-    DB_ENV	*env,
-    u_int32_t	locker,
+    DB_TXN *txn,
     DB_LOCK	*lock )
 {
 	EntryInfo *ei = BEI(e);
-	int	rc;
+	int	rc, busy = 0;
+
+	assert( e->e_private != NULL );
 
-	assert( e->e_private );
+	/* Lock the entry's info */
+	bdb_cache_entryinfo_lock( ei );
 
 	/* Set this early, warn off any queriers */
 	ei->bei_state |= CACHE_ENTRY_DELETED;
 
-	/* Lock the entry's info */
-	bdb_cache_entryinfo_lock( ei );
+	if (( ei->bei_state & ( CACHE_ENTRY_NOT_LINKED |
+		CACHE_ENTRY_LOADING | CACHE_ENTRY_ONELEVEL )) ||
+		ei->bei_finders > 0 )
+		busy = 1;
+
+	bdb_cache_entryinfo_unlock( ei );
+
+	while ( busy ) {
+		ldap_pvt_thread_yield();
+		busy = 0;
+		bdb_cache_entryinfo_lock( ei );
+		if (( ei->bei_state & ( CACHE_ENTRY_NOT_LINKED |
+			CACHE_ENTRY_LOADING | CACHE_ENTRY_ONELEVEL )) ||
+			ei->bei_finders > 0 )
+			busy = 1;
+		bdb_cache_entryinfo_unlock( 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, txn, ei, 1, 0, lock );
 	if ( rc ) {
+		bdb_cache_entryinfo_lock( ei );
 		/* couldn't lock, undo and give up */
 		ei->bei_state ^= CACHE_ENTRY_DELETED;
 		bdb_cache_entryinfo_unlock( ei );
 		return rc;
 	}
 
-	/* set cache write lock */
-	ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
-
-	/* Lock the parent's kids tree */
-	bdb_cache_entryinfo_lock( ei->bei_parent );
-
-#ifdef NEW_LOGGING
-	LDAP_LOG( CACHE, ENTRY, 
-		"bdb_cache_delete: delete %ld.\n", e->e_id, 0, 0 );
-#else
 	Debug( LDAP_DEBUG_TRACE, "====> bdb_cache_delete( %ld )\n",
 		e->e_id, 0, 0 );
-#endif
 
 	/* set lru mutex */
-	ldap_pvt_thread_mutex_lock( &cache->lru_mutex );
-	rc = bdb_cache_delete_internal( cache, e->e_private );
-	/* free lru mutex */
-	ldap_pvt_thread_mutex_unlock( &cache->lru_mutex );
+	ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_lru_mutex );
 
-	/* free cache write lock */
-	ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
-	bdb_cache_entryinfo_unlock( ei->bei_parent );
+	bdb_cache_entryinfo_lock( ei->bei_parent );
+	bdb_cache_entryinfo_lock( ei );
+	rc = bdb_cache_delete_internal( &bdb->bi_cache, e->e_private, 1 );
+	bdb_cache_entryinfo_unlock( ei );
 
-	/* Leave entry info locked */
+	/* free lru mutex */
+	ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
 
 	return( rc );
 }
@@ -977,71 +1419,81 @@ bdb_cache_delete(
 void
 bdb_cache_delete_cleanup(
 	Cache *cache,
-	Entry *e )
+	EntryInfo *ei )
 {
-	EntryInfo *ei = BEI(e);
+	/* Enter with ei locked */
 
-	ei->bei_e = NULL;
-	e->e_private = NULL;
-	bdb_entry_return( e );
+	/* already freed? */
+	if ( !ei->bei_parent ) return;
 
-	free( ei->bei_nrdn.bv_val );
-	ei->bei_nrdn.bv_val = NULL;
-#ifdef BDB_HIER
-	free( ei->bei_rdn.bv_val );
-	ei->bei_rdn.bv_val = NULL;
-	ei->bei_modrdns = 0;
-	ei->bei_ckids = 0;
-	ei->bei_dkids = 0;
+	if ( ei->bei_e ) {
+		ei->bei_e->e_private = NULL;
+#ifdef SLAP_ZONE_ALLOC
+		bdb_entry_return( ei->bei_bdb, ei->bei_e, ei->bei_zseq );
+#else
+		bdb_entry_return( ei->bei_e );
 #endif
-	ei->bei_parent = NULL;
-	ei->bei_kids = NULL;
-	ei->bei_lruprev = NULL;
+		ei->bei_e = NULL;
+	}
 
-	ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
-	ei->bei_lrunext = cache->c_eifree;
-	cache->c_eifree = ei;
-	ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
 	bdb_cache_entryinfo_unlock( ei );
+	bdb_cache_entryinfo_free( cache, ei );
 }
 
 static int
 bdb_cache_delete_internal(
     Cache	*cache,
-    EntryInfo		*e )
+    EntryInfo		*e,
+    int		decr )
 {
 	int rc = 0;	/* return code */
+	int decr_leaf = 0;
+
+	/* already freed? */
+	if ( !e->bei_parent ) {
+		assert(0);
+		return -1;
+	}
 
 #ifdef BDB_HIER
 	e->bei_parent->bei_ckids--;
-	if ( e->bei_parent->bei_dkids ) e->bei_parent->bei_dkids--;
+	if ( decr && e->bei_parent->bei_dkids ) e->bei_parent->bei_dkids--;
 #endif
 	/* dn tree */
 	if ( avl_delete( &e->bei_parent->bei_kids, (caddr_t) e, bdb_rdn_cmp )
 		== NULL )
 	{
 		rc = -1;
+		assert(0);
 	}
+	if ( e->bei_parent->bei_kids )
+		decr_leaf = 1;
 
+	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;
+		assert(0);
 	}
+	ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
+	bdb_cache_entryinfo_unlock( e->bei_parent );
 
-	if (rc != 0) {
-		return rc;
-	}
-
-	/* lru */
-	LRU_DELETE( cache, e );
-	cache->c_cursize--;
+	if ( rc == 0 ){
+		/* lru */
+		LRU_DEL( cache, e );
 
-	/*
-	 * flag entry to be freed later by a call to cache_return_entry()
-	 */
-	e->bei_state |= CACHE_ENTRY_DELETED;
+		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( 0 );
+	return( rc );
 }
 
 static void
@@ -1053,7 +1505,11 @@ bdb_entryinfo_release( void *data )
 	}
 	if ( ei->bei_e ) {
 		ei->bei_e->e_private = NULL;
+#ifdef SLAP_ZONE_ALLOC
+		bdb_entry_return( ei->bei_bdb, ei->bei_e, ei->bei_zseq );
+#else
 		bdb_entry_return( ei->bei_e );
+#endif
 	}
 	bdb_cache_entryinfo_destroy( ei );
 }
@@ -1064,81 +1520,133 @@ 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->c_lru_mutex );
 
-#ifdef NEW_LOGGING
-	LDAP_LOG( CACHE, ENTRY, "bdb_cache_release_all: enter\n", 0, 0, 0 );
-#else
 	Debug( LDAP_DEBUG_TRACE, "====> bdb_cache_release_all\n", 0, 0, 0 );
-#endif
 
 	avl_free( cache->c_dntree.bei_kids, NULL );
 	avl_free( cache->c_idtree, bdb_entryinfo_release );
+	for (;cache->c_eifree;cache->c_eifree = cache->c_lruhead) {
+		cache->c_lruhead = cache->c_eifree->bei_lrunext;
+		bdb_cache_entryinfo_destroy(cache->c_eifree);
+	}
+	cache->c_cursize = 0;
+	cache->c_eiused = 0;
+	cache->c_leaves = 0;
+	cache->c_idtree = NULL;
 	cache->c_lruhead = NULL;
 	cache->c_lrutail = NULL;
+	cache->c_dntree.bei_kids = NULL;
 
 	/* free lru mutex */
-	ldap_pvt_thread_mutex_unlock( &cache->lru_mutex );
+	ldap_pvt_thread_mutex_unlock( &cache->c_lru_mutex );
 	/* free cache write lock */
 	ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
 }
 
 #ifdef LDAP_DEBUG
 static void
+bdb_lru_count( Cache *cache )
+{
+	EntryInfo	*e;
+	int ei = 0, ent = 0, nc = 0;
+
+	for ( e = cache->c_lrutail; ; ) {
+		ei++;
+		if ( e->bei_e ) {
+			ent++;
+			if ( e->bei_state & CACHE_ENTRY_NOT_CACHED )
+				nc++;
+			fprintf( stderr, "ei %d entry %p dn %s\n", ei, (void *) e->bei_e, e->bei_e->e_name.bv_val );
+		}
+		e = e->bei_lrunext;
+		if ( e == cache->c_lrutail )
+			break;
+	}
+	fprintf( stderr, "counted %d entryInfos and %d entries, %d notcached\n",
+		ei, ent, nc );
+	ei = 0;
+	for ( e = cache->c_lrutail; ; ) {
+		ei++;
+		e = e->bei_lruprev;
+		if ( e == cache->c_lrutail )
+			break;
+	}
+	fprintf( stderr, "counted %d entryInfos (on lruprev)\n", ei );
+}
+
+#ifdef SLAPD_UNUSED
+static void
 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", (void *) 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",
+			(void *) e, (void *) 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",
+			(void *) e, (void *) e->bei_e, e->bei_id, e->bei_nrdn.bv_val );
+		e = e->bei_lruprev;
+		if ( e == cache->c_lrutail )
+			break;
 	}
 }
+
+static int
+bdb_entryinfo_print(void *data, void *arg)
+{
+	EntryInfo *e = data;
+	fprintf( stderr, "\t%p, %p id %ld rdn \"%s\"\n",
+		(void *) e, (void *) e->bei_e, e->bei_id, e->bei_nrdn.bv_val );
+	return 0;
+}
+
+static void
+bdb_idtree_print(Cache *cache)
+{
+	avl_apply( cache->c_idtree, bdb_entryinfo_print, NULL, -1, AVL_INORDER );
+}
+#endif
 #endif
 
-#ifdef BDB_REUSE_LOCKERS
 static void
-bdb_locker_id_free( void *key, void *data )
+bdb_reader_free( void *key, void *data )
 {
-	DB_ENV *env = key;
-	int lockid = (int) data;
-	int rc;
+	/* DB_ENV *env = key; */
+	DB_TXN *txn = data;
 
-	rc = XLOCK_ID_FREE( env, lockid );
-	if ( rc == EINVAL ) {
-		DB_LOCKREQ lr;
-#ifdef NEW_LOGGING
-		LDAP_LOG( BACK_BDB, ERR,
-			"bdb_locker_id_free: %d err %s(%d)\n",
-			lockid, db_strerror(rc), rc );
-#else
-		Debug( LDAP_DEBUG_ANY,
-			"bdb_locker_id_free: %d err %s(%d)\n",
-			lockid, db_strerror(rc), rc );
-#endif
-		memset( &lr, 0, sizeof(lr) );
+	if ( txn ) TXN_ABORT( txn );
+}
+
+/* free up any keys used by the main thread */
+void
+bdb_reader_flush( DB_ENV *env )
+{
+	void *data;
+	void *ctx = ldap_pvt_thread_pool_context();
 
-		/* release all locks held by this locker. */
-		lr.op = DB_LOCK_PUT_ALL;
-		env->lock_vec( env, lockid, 0, &lr, 1, NULL );
-		XLOCK_ID_FREE( env, lockid );
+	if ( !ldap_pvt_thread_pool_getkey( ctx, env, &data, NULL ) ) {
+		ldap_pvt_thread_pool_setkey( ctx, env, NULL, 0, NULL, NULL );
+		bdb_reader_free( env, data );
 	}
 }
 
 int
-bdb_locker_id( Operation *op, DB_ENV *env, int *locker )
+bdb_reader_get( Operation *op, DB_ENV *env, DB_TXN **txn )
 {
-	int i, rc, lockid;
+	int i, rc;
 	void *data;
 	void *ctx;
 
-	if ( !env || !locker ) return -1;
+	if ( !env || !txn ) return -1;
 
 	/* If no op was provided, try to find the ctx anyway... */
 	if ( op ) {
@@ -1149,58 +1657,29 @@ bdb_locker_id( Operation *op, DB_ENV *env, int *locker )
 
 	/* Shouldn't happen unless we're single-threaded */
 	if ( !ctx ) {
-		*locker = 0;
+		*txn = NULL;
 		return 0;
 	}
 
 	if ( ldap_pvt_thread_pool_getkey( ctx, env, &data, NULL ) ) {
 		for ( i=0, rc=1; rc != 0 && i<4; i++ ) {
-			rc = XLOCK_ID( env, &lockid );
+			rc = TXN_BEGIN( env, NULL, txn, DB_READ_COMMITTED );
 			if (rc) ldap_pvt_thread_yield();
 		}
 		if ( rc != 0) {
 			return rc;
 		}
-		data = (void *)lockid;
+		data = *txn;
 		if ( ( rc = ldap_pvt_thread_pool_setkey( ctx, env,
-			data, bdb_locker_id_free ) ) ) {
-			XLOCK_ID_FREE( env, lockid );
-#ifdef NEW_LOGGING
-			LDAP_LOG( BACK_BDB, ERR, "bdb_locker_id: err %s(%d)\n",
+			data, bdb_reader_free, NULL, NULL ) ) ) {
+			TXN_ABORT( *txn );
+			Debug( LDAP_DEBUG_ANY, "bdb_reader_get: err %s(%d)\n",
 				db_strerror(rc), rc, 0 );
-#else
-			Debug( LDAP_DEBUG_ANY, "bdb_locker_id: err %s(%d)\n",
-				db_strerror(rc), rc, 0 );
-#endif
 
 			return rc;
 		}
 	} else {
-		lockid = (int)data;
+		*txn = data;
 	}
-	*locker = lockid;
 	return 0;
 }
-#endif
-
-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;
-			bdb_entry_return( ei->bei_e );
-			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 );
-}