]> git.sur5r.net Git - openldap/blobdiff - servers/slapd/back-bdb/cache.c
Bypass a few redundant locks
[openldap] / servers / slapd / back-bdb / cache.c
index d1e8fe1648a1fa4f80f0915574ca309390d252e6..5c9dc31490405d4a4e79bcadc1cc97e1e9213e57 100644 (file)
@@ -1,8 +1,17 @@
 /* cache.c - routines to maintain an in-core cache of entries */
 /* $OpenLDAP$ */
-/*
- * Copyright 1998-2003 The OpenLDAP Foundation, All Rights Reserved.
- * COPYING RESTRICTIONS APPLY, see COPYRIGHT file
+/* This work is part of OpenLDAP Software <http://www.openldap.org/>.
+ *
+ * Copyright 2000-2006 The OpenLDAP Foundation.
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted only as authorized by the OpenLDAP
+ * Public License.
+ *
+ * A copy of this license is available in the file LICENSE in the
+ * top-level directory of the distribution or, alternatively, at
+ * <http://www.OpenLDAP.org/license.html>.
  */
 
 #include "portable.h"
 
 #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
+#ifdef SLAPD_UNUSED
 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( )
+bdb_cache_entryinfo_new( Cache *cache )
 {
-       EntryInfo *ei;
+       EntryInfo *ei = NULL;
 
-       ei = ch_calloc(1, sizeof(struct bdb_entry_info));
-       ldap_pvt_thread_mutex_init( &ei->bei_kids_mutex );
+       if ( cache->c_eifree ) {
+               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_mutex_unlock( &cache->c_eifree_mutex );
+       }
+       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 */
-static int
+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;
@@ -52,8 +136,8 @@ bdb_cache_entry_db_relock(
 
        if ( !lock ) return 0;
 
-       lockobj.data = ei;
-       lockobj.size = sizeof(ei->bei_parent) + sizeof(ei->bei_id);
+       lockobj.data = &ei->bei_id;
+       lockobj.size = sizeof(ei->bei_id) + 1;
 
        list[0].op = DB_LOCK_PUT;
        list[0].lock = *lock;
@@ -61,30 +145,25 @@ 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) {
-#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;
        }
        return rc;
 #endif
 }
+
 static int
-bdb_cache_entry_db_lock
-( DB_ENV *env, u_int32_t locker, EntryInfo *ei, int rw, int tryOnly, DB_LOCK *lock )
+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;
@@ -98,36 +177,31 @@ bdb_cache_entry_db_lock
        else
                db_rw = DB_LOCK_READ;
 
-       lockobj.data = ei;
-       lockobj.size = sizeof(ei->bei_parent) + sizeof(ei->bei_id);
+       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) {
-#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;
 
-       rc = LOCK_PUT ( env, lock );
+       if ( !lock || lock->mode == DB_LOCK_NG ) return 0;
+
+       rc = LOCK_PUT ( bdb->bi_dbenv, lock );
        return rc;
 #endif
 }
@@ -144,39 +218,16 @@ 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 )
 {
        const EntryInfo *e1 = v_e1, *e2 = v_e2;
        int rc = e1->bei_nrdn.bv_len - e2->bei_nrdn.bv_len;
-       if (rc == 0) rc = strncmp( e1->bei_nrdn.bv_val, e2->bei_nrdn.bv_val,
-               e1->bei_nrdn.bv_len );
+       if (rc == 0) {
+               rc = strncmp( e1->bei_nrdn.bv_val, e2->bei_nrdn.bv_val,
+                       e1->bei_nrdn.bv_len );
+       }
        return rc;
 }
 
@@ -193,107 +244,32 @@ static int
 bdb_entryinfo_add_internal(
        struct bdb_info *bdb,
        EntryInfo *ei,
-       EntryInfo **res,
-       u_int32_t locker
-)
+       EntryInfo **res )
 {
-       Cache *cache = &bdb->bi_cache;
-       DB_ENV *env = bdb->bi_dbenv;
        EntryInfo *ei2 = NULL;
-       int incr = 1;
-       int addkid = 1;
-       int rc;
-       DB_LOCK lock;
 
        *res = NULL;
 
+       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 );
 
-       /* if parent was previously considered a leaf node,
-        * it was on the LRU list. Now it's going to have
-        * kids, take it off the LRU list.
-        */
-       ldap_pvt_thread_mutex_lock( &cache->lru_mutex );
-       if ( ei->bei_parent->bei_id && !ei->bei_parent->bei_kids ) {
-               LRU_DELETE( cache, ei->bei_parent );
-               incr = 0;
-       }
-
-       cache->c_cursize += incr;
-
-       /* See if we're above the cache size limit */
-       if ( cache->c_cursize > cache->c_maxsize ) {
-               EntryInfo *elru, *elprev;
-               int i = 0;
-
-               /* Look for an unused entry to remove */
-               for (elru = cache->c_lrutail; elru; elru = elprev, i++ ) {
-                       elprev = elru->bei_lruprev;
-
-                       /* Too many probes, not enough idle, give up */
-                       if (i > 10) break;
-
-                       /* If we can successfully writelock it, then
-                        * the object is idle.
-                        */
-                       if ( bdb_cache_entry_db_lock( env, locker, elru, 1, 1,
-                               &lock ) == 0 ) {
-                               if ( !elru->bei_e ) {
-                                       bdb_cache_entry_db_unlock( env, &lock );
-                                       continue;
-                               }
-                               /* Need to lock parent to delete child */
-                               if ( ldap_pvt_thread_mutex_trylock(
-                                       &elru->bei_parent->bei_kids_mutex )) {
-                                       bdb_cache_entry_db_unlock( env, &lock );
-                                       continue;
-                               }
-                               bdb_cache_delete_internal( cache, elru );
-                               bdb_cache_entryinfo_unlock( elru->bei_parent );
-                               elru->bei_e->e_private = NULL;
-                               bdb_entry_return( elru->bei_e );
-                               bdb_cache_entry_db_unlock( env, &lock );
-                               if (ei2) {
-                                       bdb_cache_entryinfo_destroy( elru );
-                               } else {
-                                       /* re-use this one */
-                                       ch_free(elru->bei_nrdn.bv_val);
-                                       elru->bei_nrdn.bv_val = NULL;
-                                       elru->bei_e = NULL;
-                                       elru->bei_kids = NULL;
-                                       elru->bei_lrunext = NULL;
-                                       elru->bei_lruprev = NULL;
-                                       elru->bei_state = 0;
-#ifdef BDB_HIER
-                                       ch_free(elru->bei_rdn.bv_val);
-                                       elru->bei_rdn.bv_val = NULL;
-                                       elru->bei_modrdns = 0;
-#endif
-                                       ei2 = elru;
-                               }
-                               if (cache->c_cursize < cache->c_maxsize)
-                                       break;
-                       }
-               }
-       }
-       if (!ei2) {
-               ei2 = bdb_cache_entryinfo_new();
-       }
        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( &cache->c_idtree, ei2, bdb_id_cmp, avl_dup_error )) {
+       if (avl_insert( &bdb->bi_cache.c_idtree, ei2, bdb_id_cmp, avl_dup_error )) {
                EntryInfo *eix;
-               eix = avl_find( cache->c_idtree, ei2, bdb_id_cmp );
+               eix = avl_find( bdb->bi_cache.c_idtree, ei2, bdb_id_cmp );
                bdb_cache_entryinfo_destroy( ei2 );
                ei2 = eix;
-               addkid = 0;
-               cache->c_cursize -= incr;
 #ifdef BDB_HIER
                /* It got freed above because its value was
                 * assigned to ei2.
@@ -301,16 +277,22 @@ bdb_entryinfo_add_internal(
                ei->bei_rdn.bv_val = NULL;
 #endif
        } else {
-               LRU_ADD( cache, ei2 );
+               bdb->bi_cache.c_eiused++;
                ber_dupbv( &ei2->bei_nrdn, &ei->bei_nrdn );
-       }
 
-       if ( addkid ) {
+               /* 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++;
                avl_insert( &ei->bei_parent->bei_kids, ei2, bdb_rdn_cmp,
                        avl_dup_error );
+#ifdef BDB_HIER
+               ei->bei_parent->bei_ckids++;
+#endif
        }
-
-       ldap_pvt_thread_mutex_unlock( &cache->lru_mutex );
+       bdb_cache_lru_link( &bdb->bi_cache, ei2 );
 
        *res = ei2;
        return 0;
@@ -327,9 +309,7 @@ bdb_cache_find_ndn(
        Operation       *op,
        DB_TXN          *txn,
        struct berval   *ndn,
-       EntryInfo       **res,
-       u_int32_t       locker
-)
+       EntryInfo       **res )
 {
        struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
        EntryInfo       ei, *eip, *ei2;
@@ -347,16 +327,33 @@ bdb_cache_find_ndn(
                ptr = ndn->bv_val + ndn->bv_len - op->o_bd->be_nsuffix[0].bv_len;
                ei.bei_nrdn.bv_val = ptr;
                ei.bei_nrdn.bv_len = op->o_bd->be_nsuffix[0].bv_len;
+               /* Skip to next rdn if suffix is empty */
+               if ( ei.bei_nrdn.bv_len == 0 ) {
+                       for (ptr = ei.bei_nrdn.bv_val - 2; ptr > ndn->bv_val
+                               && !DN_SEPARATOR(*ptr); ptr--) /* empty */;
+                       if ( ptr >= ndn->bv_val ) {
+                               if (DN_SEPARATOR(*ptr)) ptr++;
+                               ei.bei_nrdn.bv_len = ei.bei_nrdn.bv_val - ptr;
+                               ei.bei_nrdn.bv_val = ptr;
+                       }
+               }
                eip = &bdb->bi_cache.c_dntree;
        }
        
        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 ) {
                        int len = ei.bei_nrdn.bv_len;
                                
-                       ei.bei_nrdn.bv_len = ndn->bv_len - (ei.bei_nrdn.bv_val - ndn->bv_val);
+                       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);
                        bdb_cache_entryinfo_unlock( eip );
 
                        rc = bdb_dn2id( op, txn, &ei.bei_nrdn, &ei );
@@ -368,8 +365,7 @@ bdb_cache_find_ndn(
 
                        /* DN exists but needs to be added to cache */
                        ei.bei_nrdn.bv_len = len;
-                       rc = bdb_entryinfo_add_internal( bdb, &ei, &ei2,
-                               locker );
+                       rc = bdb_entryinfo_add_internal( bdb, &ei, &ei2 );
                        /* add_internal left eip and c_rwlock locked */
                        ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
                        if ( rc ) {
@@ -393,7 +389,7 @@ bdb_cache_find_ndn(
 
                /* Advance to next lower RDN */
                for (ptr = ei.bei_nrdn.bv_val - 2; ptr > ndn->bv_val
-                       && !DN_SEPARATOR(*ptr); ptr--);
+                       && !DN_SEPARATOR(*ptr); ptr--) /* empty */;
                if ( ptr >= ndn->bv_val ) {
                        if (DN_SEPARATOR(*ptr)) ptr++;
                        ei.bei_nrdn.bv_len = ei.bei_nrdn.bv_val - ptr - 1;
@@ -412,54 +408,66 @@ 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,
+       u_int32_t       locker,
        ID id,
-       EntryInfo **res
-)
+       EntryInfo **res )
 {
        struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
        EntryInfo ei, eip, *ei2 = NULL, *ein = NULL, *eir = NULL;
-       ID parent;
        int rc;
 
        ei.bei_id = id;
        ei.bei_kids = NULL;
+       ei.bei_ckids = 0;
 
        for (;;) {
-               rc = hdb_dn2id_parent( op, txn, &ei, &eip.bei_id );
+               rc = hdb_dn2id_parent( op, txn, locker, &ei, &eip.bei_id );
                if ( rc ) break;
 
                /* Save the previous node, if any */
                ei2 = ein;
 
                /* Create a new node for the current ID */
-               ein = bdb_cache_entryinfo_new();
+               ein = bdb_cache_entryinfo_new( &bdb->bi_cache );
                ein->bei_id = ei.bei_id;
                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;
                
                /* 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 );
                if ( avl_insert( &bdb->bi_cache.c_idtree, (caddr_t)ein,
                        bdb_id_cmp, avl_dup_error ) ) {
 
-                       /* Hm, can this really happen? */
+                       /* 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 );
+                               ein->bei_ckids++;
                                bdb_cache_entryinfo_unlock( ein );
                        }
+               } else {
+                       bdb_cache_lru_link( &bdb->bi_cache, ein );
                }
 
                /* If this is the first time, save this node
@@ -470,37 +478,202 @@ hdb_cache_find_parent(
                /* If there was a previous node, link it to this one */
                if ( ei2 ) ei2->bei_parent = ein;
 
+               /* Look for this node's parent */
                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;
                }
+               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 );
 
+               /* Got the parent, link in and we're done. */
                if ( ei2 ) {
-                       ein->bei_parent = 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 );
-                       *res = eir;
-                       bdb_cache_entryinfo_lock( eir );
-               }
-               ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
-               if ( ei2 ) {
-                       /* Found a link. Reset all the state info */
+                       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 );
+                       bdb_cache_entryinfo_lock( eir );
+
+                       *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 );
        }
        return rc;
 }
+
+/* Used by hdb_dn2idl when loading the EntryInfo for all the children
+ * of a given node
+ */
+int hdb_cache_load(
+       struct bdb_info *bdb,
+       EntryInfo *ei,
+       EntryInfo **res )
+{
+       EntryInfo *ei2;
+       int rc;
+
+       /* See if we already have this one */
+       bdb_cache_entryinfo_lock( ei->bei_parent );
+       ei2 = (EntryInfo *)avl_find( ei->bei_parent->bei_kids, ei, bdb_rdn_cmp );
+       bdb_cache_entryinfo_unlock( ei->bei_parent );
+
+       if ( !ei2 ) {
+               /* Not found, add it */
+               struct berval bv;
+
+               /* bei_rdn was not malloc'd before, do it now */
+               ber_dupbv( &bv, &ei->bei_rdn );
+               ei->bei_rdn = bv;
+
+               rc = bdb_entryinfo_add_internal( bdb, ei, res );
+               bdb_cache_entryinfo_unlock( ei->bei_parent );
+               ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
+       } else {
+               /* Found, return it */
+               *res = ei2;
+               return 0;
+       }
+       return rc;
+}
+#endif
+
+static void
+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 ( 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;
+       }
+
+       count = 0;
+       /* Look for an unused entry to remove */
+       for (elru = bdb->bi_cache.c_lruhead; elru; elru = elnext ) {
+               elnext = elru->bei_lrunext;
+
+               if ( ldap_pvt_thread_mutex_trylock( &elru->bei_kids_mutex ))
+                       continue;
+
+               /* 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 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;
+               }
+
+               /* 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_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 );
+#else
+                               bdb_entry_return( elru->bei_e );
 #endif
+                               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;
+                               }
+                               /* 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;
+               }
+       }
+
+       bdb->bi_cache.c_lruhead = elnext;
+       ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_head_mutex );
+}
+
+EntryInfo *
+bdb_cache_find_info(
+       struct bdb_info *bdb,
+       ID id )
+{
+       EntryInfo       ei = { 0 },
+                       *ei2;
+
+       ei.bei_id = id;
+
+       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;
+}
 
 /*
  * cache_find_id - find an entry in the cache, given id.
@@ -516,25 +689,44 @@ bdb_cache_find_id(
        EntryInfo       **eip,
        int             islocked,
        u_int32_t       locker,
-       DB_LOCK         *lock
-)
+       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 );
+again: ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
                *eip = (EntryInfo *) avl_find( bdb->bi_cache.c_idtree,
-                                       (caddr_t) &ei, bdb_id_cmp );
+                       (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 )) {
                                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 ) {
+                                       return DB_NOTFOUND;
+                               }
+                               /* otherwise, wait for the info to free up */
+                               ldap_pvt_thread_yield();
+                               goto again;
+                       }
+                       /* If this info isn't hooked up to its parent yet,
+                        * unlock and wait for it to be fully initialized
+                        */
+                       if ( (*eip)->bei_state & CACHE_ENTRY_NOT_LINKED ) {
+                               bdb_cache_entryinfo_unlock( *eip );
+                               ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
                                ldap_pvt_thread_yield();
                                goto again;
                        }
@@ -546,73 +738,139 @@ again:           ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
        /* See if the ID exists in the database; add it to the cache if so */
        if ( !*eip ) {
 #ifndef BDB_HIER
-               rc = bdb_id2entry( op->o_bd, tid, id, &ep );
+               rc = bdb_id2entry( op->o_bd, tid, locker, id, &ep );
                if ( rc == 0 ) {
                        rc = bdb_cache_find_ndn( op, tid,
-                               &ep->e_nname, eip, locker );
-                       if ( *eip )
-                               islocked = 1;
+                               &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
                                bdb_entry_return( ep );
+#endif
                                ep = NULL;
                        }
                }
 #else
-               rc = hdb_cache_find_parent(op, tid, id, eip );
-               if ( rc == 0 && *eip )
-                       islocked = 1;
+               rc = hdb_cache_find_parent(op, tid, locker, id, eip );
+               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 if ( !(*eip)->bei_e ) {
-                       if (!ep) {
-                               rc = bdb_id2entry( op->o_bd, tid, id, &ep );
+               } else {
+                       /* 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;
                        }
-                       if ( rc == 0 ) {
-                               bdb_cache_entry_db_lock( bdb->bi_dbenv, locker,
-                                       *eip, 1, 0, lock );
-                               ep->e_private = *eip;
-#ifdef BDB_HIER
-                               bdb_fix_dn( ep, 0 );
 #endif
-                               (*eip)->bei_e = ep;
-                               bdb_cache_entry_db_relock( bdb->bi_dbenv, locker,
-                                       *eip, 0, 0, lock );
+                       if ( !(*eip)->bei_e && !((*eip)->bei_state & CACHE_ENTRY_LOADING)) {
+                               load = 1;
+                               (*eip)->bei_state |= CACHE_ENTRY_LOADING;
                        }
-               } else {
+                       if ( islocked ) {
+                               bdb_cache_entryinfo_unlock( *eip );
+                               islocked = 0;
+                       }
+                       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, lock );
+                       } else if ( rc == 0 ) {
+                               if ( load ) {
+                                       if ( !ep) {
+                                               rc = bdb_id2entry( op->o_bd, tid, locker, id, &ep );
+                                       }
+                                       if ( rc == 0 ) {
+                                               ep->e_private = *eip;
 #ifdef BDB_HIER
-                       rc = bdb_fix_dn( (*eip)->bei_e, 1 );
-                       if ( rc ) {
-                               bdb_cache_entry_db_lock( bdb->bi_dbenv,
-                                       locker, *eip, 1, 0, lock );
-                               rc = bdb_fix_dn( (*eip)->bei_e, 2 );
-                               bdb_cache_entry_db_relock( bdb->bi_dbenv,
-                                       locker, *eip, 0, 0, lock );
-                       } else {
-                               bdb_cache_entry_db_lock( bdb->bi_dbenv,
-                                       locker, *eip, 0, 0, lock );
+                                               bdb_fix_dn( ep, 0 );
+#endif
+                                               (*eip)->bei_e = ep;
+#ifdef SLAP_ZONE_ALLOC
+                                               (*eip)->bei_zseq = *((ber_len_t *)ep - 2);
+#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, locker,
+                                                       *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 );
+                                       islocked = 1;
+                                       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,
+                                                       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,
+                                                       locker, *eip, 0, 0, lock );
+                                       }
+#endif
+                               }
+
                        }
+               }
+       }
+       if ( islocked ) {
+               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_cache_entry_db_lock( bdb->bi_dbenv, locker,
-                                       *eip, 0, 0, lock );
+               bdb_entry_return( ep );
 #endif
-               }
        }
-       if ( rc == 0 && (*eip)->bei_kids == NULL ) {
-               /* set lru mutex */
-               ldap_pvt_thread_mutex_lock( &bdb->bi_cache.lru_mutex );
-               LRU_DELETE( &bdb->bi_cache, *eip );
-               LRU_ADD( &bdb->bi_cache, *eip );
-               ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_mutex );
+       if ( rc == 0 ) {
+               int purge = 0;
+
+               if ( load ) {
+                       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 );
+               }
+               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;
 }
 
@@ -620,8 +878,7 @@ int
 bdb_cache_children(
        Operation *op,
        DB_TXN *txn,
-       Entry *e
-)
+       Entry *e )
 {
        int rc;
 
@@ -633,7 +890,7 @@ bdb_cache_children(
        }
        rc = bdb_dn2id_children( op, txn, e );
        if ( rc == DB_NOTFOUND ) {
-               BEI(e)->bei_state |= CACHE_ENTRY_NO_KIDS;
+               BEI(e)->bei_state |= CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
        }
        return rc;
 }
@@ -645,56 +902,94 @@ bdb_cache_add(
        EntryInfo *eip,
        Entry *e,
        struct berval *nrdn,
-       u_int32_t locker
-)
+       u_int32_t locker,
+       DB_LOCK *lock )
 {
        EntryInfo *new, ei;
+       int rc, purge = 0;
+#ifdef BDB_HIER
        struct berval rdn = e->e_name;
-       int rc;
+#endif
 
        ei.bei_id = e->e_id;
        ei.bei_parent = eip;
        ei.bei_nrdn = *nrdn;
+       ei.bei_lockpad = 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, locker, &ei, 0, 0, lock );
+       if ( rc ) {
+               bdb_cache_entryinfo_unlock( eip );
+               return rc;
+       }
+
 #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
+
+       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
-       rc = bdb_entryinfo_add_internal( bdb, &ei, &new, locker );
+       }
        new->bei_e = e;
        e->e_private = new;
-       new->bei_state = CACHE_ENTRY_NO_KIDS;
+       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 );
+
        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 );
+
+       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
-)
+       DB_LOCK *lock )
 {
        EntryInfo *ei = BEI(e);
-       
+       int rc;
        /* Get write lock on data */
-       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.
         */
-       if ( (void *)e->e_attrs != (void *)(e+1) ) {
-               attrs_free( e->e_attrs );
+       if ( ! rc ) {
+               if ( (void *)e->e_attrs != (void *)(e+1) ) {
+                       attrs_free( e->e_attrs ); 
+               }
+               e->e_attrs = newAttrs;
        }
-       e->e_attrs = newAttrs;
-
-       return 0;
+       return rc;
 }
 
 /*
@@ -702,21 +997,23 @@ 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_LOCK *lock
-)
+       DB_LOCK *lock )
 {
        EntryInfo *ei = BEI(e), *pei;
+       int rc;
+#ifdef BDB_HIER
        struct berval rdn;
-       int rc = 0;
+#endif
 
        /* Get write lock on data */
-       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 ) return rc;
 
        /* If we've done repeated mods on a cached entry, then e_attrs
         * is no longer contiguous with the entry, and must be freed.
@@ -725,8 +1022,9 @@ bdb_cache_modrdn(
                attrs_free( e->e_attrs );
        }
        e->e_attrs = new->e_attrs;
-       if( e->e_nname.bv_val < e->e_bv.bv_val || e->e_nname.bv_val >
-               e->e_bv.bv_val + e->e_bv.bv_len ) {
+       if( e->e_nname.bv_val < e->e_bv.bv_val ||
+               e->e_nname.bv_val > e->e_bv.bv_val + e->e_bv.bv_len )
+       {
                ch_free(e->e_name.bv_val);
                ch_free(e->e_nname.bv_val);
        }
@@ -739,15 +1037,22 @@ 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 );
 
        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 );
+       pei->bei_ckids--;
+       if ( pei->bei_dkids ) pei->bei_dkids--;
 #endif
 
        if (!ein) {
@@ -757,15 +1062,25 @@ 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
-       { 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;
+
+       {
+               /* 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 );
        }
+       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 );
@@ -780,100 +1095,140 @@ 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
-)
+    DB_LOCK    *lock )
 {
        EntryInfo *ei = BEI(e);
        int     rc;
 
-       assert( e->e_private );
+       assert( e->e_private != NULL );
 
        /* Set this early, warn off any queriers */
        ei->bei_state |= CACHE_ENTRY_DELETED;
 
-       /* Get write lock on the data */
-       bdb_cache_entry_db_relock( env, locker, ei, 1, 0, lock );
+       /* Lock the entry's info */
+       bdb_cache_entryinfo_lock( ei );
 
-       /* 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 );
+       /* Get write lock on the data */
+       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;
+               bdb_cache_entryinfo_unlock( ei );
+               return rc;
+       }
 
-#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 );
+       ldap_pvt_thread_mutex_lock( &bdb->bi_cache.lru_head_mutex );
+
+       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 */
 
-       /* free cache write lock */
-       ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
-       bdb_cache_entryinfo_unlock( ei->bei_parent );
        return( rc );
 }
 
 void
 bdb_cache_delete_cleanup(
-       Entry *e
-)
+       Cache *cache,
+       EntryInfo *ei )
 {
-       bdb_cache_entryinfo_destroy( e->e_private );
-       e->e_private = NULL;
-       bdb_entry_return( e );
+       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_e = NULL;
+       }
+
+       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;
+#endif
+       ei->bei_parent = NULL;
+       ei->bei_kids = NULL;
+       ei->bei_lruprev = NULL;
+
+       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 );
+       bdb_cache_entryinfo_unlock( ei );
 }
-       
+
 static int
 bdb_cache_delete_internal(
     Cache      *cache,
-    EntryInfo          *e
-)
+    EntryInfo          *e,
+    int                decr )
 {
        int rc = 0;     /* return code */
+       int decr_leaf = 0;
 
+       /* Lock the parent's kids tree */
+       bdb_cache_entryinfo_lock( e->bei_parent );
+
+#ifdef BDB_HIER
+       e->bei_parent->bei_ckids--;
+       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 )
+       if ( avl_delete( &e->bei_parent->bei_kids, (caddr_t) e, bdb_rdn_cmp )
+               == NULL )
        {
                rc = -1;
        }
+       if ( e->bei_parent->bei_kids )
+               decr_leaf = 1;
 
-       /* If parent has no more kids, put in on LRU list */
-       if ( e->bei_parent->bei_kids == NULL ) {
-               LRU_ADD( cache, e->bei_parent );
-               cache->c_cursize++;
-       }
+       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) {
-               return rc;
-       }
+       if ( rc == 0 ){
+               /* lru */
+               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 );
+               }
 
-       /* lru */
-       LRU_DELETE( cache, e );
-       cache->c_cursize--;
+               if ( e->bei_lrunext ) e->bei_lrunext->bei_lruprev = e->bei_lruprev;
+               if ( e->bei_lruprev ) e->bei_lruprev->bei_lrunext = e->bei_lrunext;
 
-       /*
-        * 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
@@ -885,7 +1240,11 @@ 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 );
 }
@@ -896,78 +1255,85 @@ 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 );
 
-#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->lru_tail_mutex );
        /* free cache write lock */
        ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
 }
 
 #ifdef LDAP_DEBUG
+#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", 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
+#endif
 
 #ifdef BDB_REUSE_LOCKERS
 static void
 bdb_locker_id_free( void *key, void *data )
 {
        DB_ENV *env = key;
-       int lockid = (int) data;
+       u_int32_t lockid = (long)data;
        int rc;
 
-
        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) );
-
+                       "bdb_locker_id_free: %lu err %s(%d)\n",
+                       (unsigned long) lockid, db_strerror(rc), rc );
                /* release all locks held by this locker. */
                lr.op = DB_LOCK_PUT_ALL;
+               lr.obj = NULL;
                env->lock_vec( env, lockid, 0, &lr, 1, NULL );
                XLOCK_ID_FREE( env, lockid );
        }
 }
 
 int
-bdb_locker_id( Operation *op, DB_ENV *env, int *locker )
+bdb_locker_id( Operation *op, DB_ENV *env, u_int32_t *locker )
 {
-       int i, rc, lockid;
+       int i, rc;
+       u_int32_t lockid;
        void *data;
        void *ctx;
 
@@ -994,24 +1360,19 @@ bdb_locker_id( Operation *op, DB_ENV *env, int *locker )
                if ( rc != 0) {
                        return rc;
                }
-               data = (void *)lockid;
+               data = (void *)((long)lockid);
                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",
-                               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;
+               lockid = (long)data;
        }
        *locker = lockid;
        return 0;
 }
-#endif
+#endif /* BDB_REUSE_LOCKERS */