]> git.sur5r.net Git - openldap/blobdiff - servers/slapd/back-bdb/cache.c
ITS#5860
[openldap] / servers / slapd / back-bdb / cache.c
index cb1ff8e70286545908a5dcaf4132e75a6e47f2d0..6eddd00856cfc1cd8a68d02de4789aa411803229 100644 (file)
@@ -2,7 +2,7 @@
 /* $OpenLDAP$ */
 /* This work is part of OpenLDAP Software <http://www.openldap.org/>.
  *
- * Copyright 2000-2005 The OpenLDAP Foundation.
+ * Copyright 2000-2009 The OpenLDAP Foundation.
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
 #include "ldap_rq.h"
 
 #ifdef BDB_HIER
-#define bdb_cache_lru_add      hdb_cache_lru_add
+#define bdb_cache_lru_purge    hdb_cache_lru_purge
 #endif
-static void bdb_cache_lru_add( struct bdb_info *bdb, EntryInfo *ei );
+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);
 #endif
 #endif
 
+/* For concurrency experiments only! */
+#if 0
+#define        ldap_pvt_thread_rdwr_wlock(a)   0
+#define        ldap_pvt_thread_rdwr_wunlock(a) 0
+#define        ldap_pvt_thread_rdwr_rlock(a)   0
+#define        ldap_pvt_thread_rdwr_runlock(a) 0
+#endif
+
+#if 0
+#define ldap_pvt_thread_mutex_trylock(a) 0
+#endif
+
 static EntryInfo *
 bdb_cache_entryinfo_new( Cache *cache )
 {
        EntryInfo *ei = NULL;
 
        if ( cache->c_eifree ) {
-               ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
+               ldap_pvt_thread_mutex_lock( &cache->c_eifree_mutex );
                if ( cache->c_eifree ) {
                        ei = cache->c_eifree;
                        cache->c_eifree = ei->bei_lrunext;
+                       ei->bei_finders = 0;
                }
-               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 );
+       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 );
+}
+
+#define LRU_DEL( c, e ) do { \
+       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;
@@ -92,7 +183,7 @@ bdb_cache_entry_db_relock(
        list[1].lock = *lock;
        list[1].mode = rw ? DB_LOCK_WRITE : DB_LOCK_READ;
        list[1].obj = &lockobj;
-       rc = env->lock_vec(env, locker, tryOnly ? DB_LOCK_NOWAIT : 0,
+       rc = bdb->bi_dbenv->lock_vec(bdb->bi_dbenv, TXN_ID(txn), tryOnly ? DB_LOCK_NOWAIT : 0,
                list, 2, NULL );
 
        if (rc && !tryOnly) {
@@ -107,10 +198,10 @@ 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;
@@ -127,7 +218,7 @@ bdb_cache_entry_db_lock( DB_ENV *env, u_int32_t locker, EntryInfo *ei,
        lockobj.data = &ei->bei_id;
        lockobj.size = sizeof(ei->bei_id) + 1;
 
-       rc = LOCK_GET(env, locker, tryOnly ? DB_LOCK_NOWAIT : 0,
+       rc = LOCK_GET(bdb->bi_dbenv, TXN_ID(txn), tryOnly ? DB_LOCK_NOWAIT : 0,
                                        &lockobj, db_rw, lock);
        if (rc && !tryOnly) {
                Debug( LDAP_DEBUG_TRACE,
@@ -135,24 +226,59 @@ bdb_cache_entry_db_lock( DB_ENV *env, u_int32_t locker, EntryInfo *ei,
                        ei->bei_id, rw, rc );
        }
        return rc;
-#endif /* NO_THREADS */
+#endif /* NO_DB_LOCK */
 }
 
 int
-bdb_cache_entry_db_unlock ( DB_ENV *env, DB_LOCK *lock )
+bdb_cache_entry_db_unlock ( struct bdb_info *bdb, DB_LOCK *lock )
 {
-#ifdef NO_THREADS
+#ifdef NO_DB_LOCK
        return 0;
 #else
        int rc;
 
-       if ( !lock ) return 0;
+       if ( !lock || lock->mode == DB_LOCK_NG ) return 0;
 
-       rc = LOCK_PUT ( env, lock );
+       rc = LOCK_PUT ( bdb->bi_dbenv, lock );
        return rc;
 #endif
 }
 
+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_trylock( ei ) == 0 )) {
+               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 )
 {
@@ -165,32 +291,6 @@ bdb_cache_entryinfo_destroy( EntryInfo *e )
        return 0;
 }
 
-#define LRU_DELETE( cache, ei ) do { \
-       if ( (ei)->bei_lruprev != NULL ) { \
-               (ei)->bei_lruprev->bei_lrunext = (ei)->bei_lrunext; \
-       } else { \
-               (cache)->c_lruhead = (ei)->bei_lrunext; \
-       } \
-       if ( (ei)->bei_lrunext != NULL ) { \
-               (ei)->bei_lrunext->bei_lruprev = (ei)->bei_lruprev; \
-       } else { \
-               (cache)->c_lrutail = (ei)->bei_lruprev; \
-       } \
-       (ei)->bei_lrunext = (ei)->bei_lruprev = NULL; \
-} while(0)
-
-#define LRU_ADD( cache, ei ) do { \
-       (ei)->bei_lrunext = (cache)->c_lruhead; \
-       if ( (ei)->bei_lrunext != NULL ) { \
-               (ei)->bei_lrunext->bei_lruprev = (ei); \
-       } \
-       (cache)->c_lruhead = (ei); \
-       (ei)->bei_lruprev = NULL; \
-       if ( (cache)->c_lrutail == NULL ) { \
-               (cache)->c_lrutail = (ei); \
-       } \
-} while(0)
-
 /* Do a length-ordered sort on normalized RDNs */
 static int
 bdb_rdn_cmp( const void *v_e1, const void *v_e2 )
@@ -211,6 +311,14 @@ 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
@@ -238,10 +346,10 @@ bdb_entryinfo_add_internal(
 #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
@@ -250,6 +358,8 @@ 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 );
 
@@ -259,8 +369,13 @@ bdb_entryinfo_add_internal(
                 */
                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,
+               rc = avl_insert( &ei->bei_parent->bei_kids, ei2, bdb_rdn_cmp,
                        avl_dup_error );
+               if ( rc ) {
+                       /* This should never happen; entry cache is corrupt */
+                       bdb->bi_dbenv->log_flush( bdb->bi_dbenv, NULL );
+                       assert( !rc );
+               }
 #ifdef BDB_HIER
                ei->bei_parent->bei_ckids++;
 #endif
@@ -313,9 +428,11 @@ 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 ) {
+                       DB_LOCK lock;
                        int len = ei.bei_nrdn.bv_len;
                                
                        if ( BER_BVISEMPTY( ndn )) {
@@ -327,18 +444,27 @@ bdb_cache_find_ndn(
                                (ei.bei_nrdn.bv_val - ndn->bv_val);
                        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 );
+
+                       lock.mode = DB_LOCK_NG;
+                       rc = bdb_dn2id( op, &ei.bei_nrdn, &ei, txn, &lock );
                        if (rc) {
                                bdb_cache_entryinfo_lock( eip );
+                               bdb_cache_entry_db_unlock( bdb, &lock );
                                *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 */
                        ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
+                       bdb_cache_entry_db_unlock( bdb, &lock );
                        if ( rc ) {
                                *res = eip;
                                return rc;
@@ -382,22 +508,20 @@ bdb_cache_find_ndn(
 int
 hdb_cache_find_parent(
        Operation *op,
-       DB_TXN *txn,
-       u_int32_t       locker,
+       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;
        int rc;
-       int addlru = 0;
 
        ei.bei_id = id;
        ei.bei_kids = NULL;
        ei.bei_ckids = 0;
 
        for (;;) {
-               rc = hdb_dn2id_parent( op, txn, locker, &ei, &eip.bei_id );
+               rc = hdb_dn2id_parent( op, txn, &ei, &eip.bei_id );
                if ( rc ) break;
 
                /* Save the previous node, if any */
@@ -416,19 +540,19 @@ hdb_cache_find_parent(
                ei.bei_ckids = 0;
                
                /* This node is not fully connected yet */
-               ein->bei_state = CACHE_ENTRY_NOT_LINKED;
+               ein->bei_state |= CACHE_ENTRY_NOT_LINKED;
 
                /* Insert this node into the ID tree */
                ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
                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;
 
                        /* 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 );
+                       bdb_cache_entryinfo_free( &bdb->bi_cache, ein );
+                       ein = eix;
                        
                        /* Link in any kids we've already processed */
                        if ( ei2 ) {
@@ -438,8 +562,6 @@ hdb_cache_find_parent(
                                ein->bei_ckids++;
                                bdb_cache_entryinfo_unlock( ein );
                        }
-                       addlru = 0;
-
                }
 
                /* If this is the first time, save this node
@@ -462,25 +584,22 @@ hdb_cache_find_parent(
                                bdb->bi_cache.c_leaves++;
                ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
 
-               if ( addlru ) {
-                       ldap_pvt_thread_mutex_lock( &bdb->bi_cache.lru_head_mutex );
-                       bdb_cache_lru_add( bdb, ein );
-               }
-               addlru = 1;
-
                /* Got the parent, link in and we're done. */
                if ( ei2 ) {
+                       bdb_cache_entryinfo_lock( eir );
                        bdb_cache_entryinfo_lock( ei2 );
                        ein->bei_parent = ei2;
+
                        avl_insert( &ei2->bei_kids, (caddr_t)ein, bdb_rdn_cmp,
                                avl_dup_error);
                        ei2->bei_ckids++;
-                       bdb_cache_entryinfo_unlock( ei2 );
-                       bdb_cache_entryinfo_lock( eir );
 
                        /* Reset all the state info */
                        for (ein = eir; ein != ei2; ein=ein->bei_parent)
                                ein->bei_state &= ~CACHE_ENTRY_NOT_LINKED;
+
+                       bdb_cache_entryinfo_unlock( ei2 );
+
                        *res = eir;
                        break;
                }
@@ -529,131 +648,153 @@ int hdb_cache_load(
 }
 #endif
 
-static void *
-bdb_cache_lru_purge(void *ctx, void *arg)
+/* 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_purge( struct bdb_info *bdb )
 {
-       struct re_s *rtask = arg;
-       struct bdb_info *bdb = rtask->arg;
        DB_LOCK         lock, *lockp;
-       EntryInfo *elru, *elprev;
-       int count = 0;
+       EntryInfo *elru, *elnext = NULL;
+       int count, islocked, eimax;
+       int efree = 0, eifree = 0, eicount, ecount;
+#ifdef LDAP_DEBUG
+       int iter;
+#endif
 
-       if ( bdb->bi_cache.c_locker ) {
+       /* Wait for the mutex; we're the only one trying to purge. */
+       ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_lru_mutex );
+
+       /* maximum number of EntryInfo leaves to cache. In slapcat
+        * we always free all leaf nodes.
+        */
+       if ( slapMode & SLAP_TOOL_READONLY )
+               eimax = 0;
+       else
+               eimax = bdb->bi_cache.c_eimax;
+
+       if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize )
+               efree = bdb->bi_cache.c_minfree;
+       if ( bdb->bi_cache.c_leaves > eimax ) {
+               eifree = bdb->bi_cache.c_minfree * 10;
+               if ( eifree >= eimax )
+                       eifree = eimax / 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;
        }
 
-       ldap_pvt_thread_mutex_lock( &bdb->bi_cache.lru_tail_mutex );
+       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 ) {
-               elprev = elru->bei_lruprev;
+       for ( elru = bdb->bi_cache.c_lruhead; elru; elru = elnext ) {
+               elnext = elru->bei_lrunext;
+
+               if ( bdb_cache_entryinfo_trylock( elru ))
+                       goto bottom;
+
+               /* 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 )) ||
+                       elru->bei_finders > 0 ) {
+                       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->bi_dbenv,
-                               bdb->bi_cache.c_locker, elru, 1, 1, lockp ) == 0 ) {
-
-                       int stop = 0;
+               if ( bdb_cache_entry_db_lock( bdb,
+                       bdb->bi_cache.c_txn, elru, 1, 1, lockp ) == 0 ) {
 
-                       /* If this node is in the process of linking into the cache,
-                        * or this node is being deleted, skip it.
-                        */
-                       if ( elru->bei_state &
-                               ( CACHE_ENTRY_NOT_LINKED | CACHE_ENTRY_DELETED )) {
-                               bdb_cache_entry_db_unlock( bdb->bi_dbenv, lockp );
-                               continue;
-                       }
                        /* Free entry for this node if it's present */
                        if ( elru->bei_e ) {
-                               elru->bei_e->e_private = NULL;
+                               ecount++;
+                               if ( count < efree ) {
+                                       elru->bei_e->e_private = NULL;
 #ifdef SLAP_ZONE_ALLOC
-                               bdb_entry_return( bdb, elru->bei_e, elru->bei_zseq );
+                                       bdb_entry_return( bdb, elru->bei_e, elru->bei_zseq );
 #else
-                               bdb_entry_return( elru->bei_e );
+                                       bdb_entry_return( elru->bei_e );
 #endif
-                               elru->bei_e = NULL;
-                               count++;
-                       }
-                       /* ITS#4010 if we're in slapcat, and this node is a leaf
-                        * node, free it.
-                        *
-                        * FIXME: we need to do this for slapd as well, (which is
-                        * why we compute bi_cache.c_leaves now) but at the moment
-                        * we can't because it causes unresolvable deadlocks. 
-                        */
-                       if ( slapMode & SLAP_TOOL_READONLY ) {
-                               if ( !elru->bei_kids ) {
-                                       /* This does LRU_DELETE for us */
-                                       bdb_cache_delete_internal( &bdb->bi_cache, elru, 0 );
-                                       bdb_cache_delete_cleanup( &bdb->bi_cache, elru );
+                                       elru->bei_e = NULL;
+                                       count++;
+                               } else {
+                                       /* Keep this node cached, skip to next */
+                                       bdb_cache_entry_db_unlock( bdb, lockp );
+                                       goto next;
                                }
-                               /* Leave node on LRU list for a future pass */
-                       } else {
-                               LRU_DELETE( &bdb->bi_cache, elru );
-                       }
-                       bdb_cache_entry_db_unlock( bdb->bi_dbenv, lockp );
-
-                       if ( count == bdb->bi_cache.c_minfree ) {
-                               ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
-                               bdb->bi_cache.c_cursize -= bdb->bi_cache.c_minfree;
-                               if ( bdb->bi_cache.c_maxsize - bdb->bi_cache.c_cursize >=
-                                       bdb->bi_cache.c_minfree )
-                                       stop = 1;
-                               count = 0;
-                               ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
                        }
-                       if (stop) break;
-               }
-       }
-
-       ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_tail_mutex );
-
-       /* If we're running as a task, drop the task */
-       if ( ctx ) {
-               ldap_pvt_thread_mutex_lock( &slapd_rq.rq_mutex );
-               ldap_pvt_runqueue_stoptask( &slapd_rq, rtask );
-               /* Defer processing till we're needed again */
-               ldap_pvt_runqueue_resched( &slapd_rq, rtask, 1 );
-               ldap_pvt_thread_mutex_unlock( &slapd_rq.rq_mutex );
-       }
-}
-
-/* caller must have lru_head_mutex locked. mutex
- * will be unlocked on return.
- */
-static void
-bdb_cache_lru_add(
-       struct bdb_info *bdb,
-       EntryInfo *ei )
-{
-       LRU_ADD( &bdb->bi_cache, ei );
-       ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_head_mutex );
+                       bdb_cache_entry_db_unlock( bdb, lockp );
 
-       /* See if we're above the cache size limit */
-       if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize ) {
-               if ( slapMode & SLAP_TOOL_MODE ) {
-                       struct re_s rtask;
+                       /* 
+                        * 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 */
+               }
 
-                       rtask.arg = bdb;
-                       bdb_cache_lru_purge( NULL, &rtask );
-               } else {
-                       ldap_pvt_thread_mutex_lock( &slapd_rq.rq_mutex );
-                       if ( bdb->bi_cache_task ) {
-                               if ( !ldap_pvt_runqueue_isrunning( &slapd_rq,
-                                       bdb->bi_cache_task ))
-                                       ldap_pvt_runqueue_resched( &slapd_rq, bdb->bi_cache_task,
-                                               0 );
-                       } else {
-                               bdb->bi_cache_task = ldap_pvt_runqueue_insert( &slapd_rq, 0,
-                                       bdb_cache_lru_purge, bdb, "bdb_cache_lru_purge",
-                                       bdb->bi_dbenv_home );
+next:
+               if ( islocked )
+                       bdb_cache_entryinfo_unlock( elru );
+
+               if ( count >= efree && eicount >= eifree ) {
+                       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 );
                        }
-                       ldap_pvt_thread_mutex_unlock( &slapd_rq.rq_mutex );
+                       break;
                }
+bottom:
+               if ( elnext == bdb->bi_cache.c_lruhead )
+                       break;
+#ifdef LDAP_DEBUG
+               iter++;
+#endif
        }
+
+       bdb->bi_cache.c_lruhead = elnext;
+       ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
+       bdb->bi_cache.c_purging = 0;
 }
 
 EntryInfo *
@@ -675,7 +816,7 @@ bdb_cache_find_info(
 
 /*
  * 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.
  */
 
@@ -685,8 +826,7 @@ 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;
@@ -706,8 +846,7 @@ 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 )) {
                                ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
                                /* If this node is being deleted, treat
                                 * as if the delete has already finished
@@ -728,7 +867,7 @@ 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 );
        }
@@ -736,12 +875,13 @@ 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, locker, id, &ep );
+               rc = bdb_id2entry( op->o_bd, tid, id, &ep );
                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
@@ -751,16 +891,23 @@ again:    ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
                        }
                }
 #else
-               rc = hdb_cache_find_parent(op, tid, locker, id, eip );
-               if ( rc == 0 && *eip ) islocked = 1;
+               rc = hdb_cache_find_parent(op, tid, id, eip );
+               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 {
+                       (*eip)->bei_finders++;
+                       (*eip)->bei_state |= CACHE_ENTRY_REFERENCED;
                        /* Make sure only one thread tries to load the entry */
 load1:
 #ifdef SLAP_ZONE_ALLOC
@@ -774,24 +921,28 @@ load1:
                                load = 1;
                                (*eip)->bei_state |= CACHE_ENTRY_LOADING;
                        }
-                       if ( islocked ) {
+
+                       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.
+                                */
+                               (*eip)->bei_state &= ~CACHE_ENTRY_NOT_CACHED;
+                               flag &= ~ID_NOCACHE;
+                       }
+
+                       if ( flag & ID_LOCKED ) {
                                bdb_cache_entryinfo_unlock( *eip );
-                               islocked = 0;
+                               flag ^= ID_LOCKED;
                        }
-                       rc = bdb_cache_entry_db_lock( bdb->bi_dbenv, locker, *eip, 0, 0, lock );
+                       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->bi_dbenv, lock );
+                               bdb_cache_entry_db_unlock( bdb, lock );
                        } else if ( rc == 0 ) {
                                if ( load ) {
-                                       /* Give up original read lock, obtain write lock
-                                        */
-                                   if ( rc == 0 ) {
-                                               rc = bdb_cache_entry_db_relock( bdb->bi_dbenv, locker,
-                                                       *eip, 1, 0, lock );
-                                       }
-                                       if ( rc == 0 && !ep) {
-                                               rc = bdb_id2entry( op->o_bd, tid, locker, id, &ep );
+                                       if ( !ep) {
+                                               rc = bdb_id2entry( op->o_bd, tid, id, &ep );
                                        }
                                        if ( rc == 0 ) {
                                                ep->e_private = *eip;
@@ -803,24 +954,32 @@ load1:
                                                (*eip)->bei_zseq = *((ber_len_t *)ep - 2);
 #endif
                                                ep = NULL;
+                                               bdb_cache_lru_link( bdb, *eip );
+                                               if (( flag & ID_NOCACHE ) &&
+                                                       ( bdb_cache_entryinfo_trylock( *eip ) == 0 )) {
+                                                       /* 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;
+                                                       bdb_cache_entryinfo_unlock( *eip );
+                                               }
                                        }
-                                       (*eip)->bei_state ^= CACHE_ENTRY_LOADING;
                                        if ( rc == 0 ) {
                                                /* If we succeeded, downgrade back to a readlock. */
-                                               rc = bdb_cache_entry_db_relock( bdb->bi_dbenv, locker,
+                                               rc = bdb_cache_entry_db_relock( bdb, tid,
                                                        *eip, 0, 0, lock );
                                        } else {
                                                /* Otherwise, release the lock. */
-                                               bdb_cache_entry_db_unlock( bdb->bi_dbenv, lock );
+                                               bdb_cache_entry_db_unlock( bdb, lock );
                                        }
                                } else if ( !(*eip)->bei_e ) {
                                        /* Some other thread is trying to load the entry,
-                                        * give it a chance to finish.
+                                        * wait for it to finish.
                                         */
-                                       bdb_cache_entry_db_unlock( bdb->bi_dbenv, lock );
-                                       ldap_pvt_thread_yield();
+                                       bdb_cache_entry_db_unlock( bdb, lock );
                                        bdb_cache_entryinfo_lock( *eip );
-                                       islocked = 1;
+                                       flag |= ID_LOCKED;
                                        goto load1;
 #ifdef BDB_HIER
                                } else {
@@ -828,24 +987,29 @@ load1:
                                         */
                                        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 ( islocked ) {
+       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
@@ -853,31 +1017,25 @@ load1:
 #endif
        }
        if ( rc == 0 ) {
+               int purge = 0;
 
                if ( load ) {
-                       ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
-                       bdb->bi_cache.c_cursize++;
-                       ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
-               }
-
-               ldap_pvt_thread_mutex_lock( &bdb->bi_cache.lru_head_mutex );
-
-               /* If the LRU list has only one entry and this is it, it
-                * doesn't need to be added again.
-                */
-               if ( bdb->bi_cache.c_lruhead == bdb->bi_cache.c_lrutail &&
-                       bdb->bi_cache.c_lruhead == *eip ) {
-                       ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_head_mutex );
-               } else {
-                       /* if entry is on LRU list, remove from old spot */
-                       if ( (*eip)->bei_lrunext || (*eip)->bei_lruprev ) {
-                               ldap_pvt_thread_mutex_lock( &bdb->bi_cache.lru_tail_mutex );
-                               LRU_DELETE( &bdb->bi_cache, *eip );
-                               ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_tail_mutex );
+                       ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
+                       if ( !( flag & ID_NOCACHE )) {
+                               bdb->bi_cache.c_cursize++;
+                               if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize &&
+                                       !bdb->bi_cache.c_purging ) {
+                                       purge = 1;
+                                       bdb->bi_cache.c_purging = 1;
+                               }
+                       } else if ( bdb->bi_cache.c_leaves > bdb->bi_cache.c_eimax && !bdb->bi_cache.c_purging ) {
+                               purge = 1;
+                               bdb->bi_cache.c_purging = 1;
                        }
-                       /* lru_head_mutex is unlocked for us */
-                       bdb_cache_lru_add( bdb, *eip );
+                       ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
                }
+               if ( purge )
+                       bdb_cache_lru_purge( bdb );
        }
 
 #ifdef SLAP_ZONE_ALLOC
@@ -917,11 +1075,11 @@ bdb_cache_add(
        EntryInfo *eip,
        Entry *e,
        struct berval *nrdn,
-       u_int32_t locker )
+       DB_TXN *txn,
+       DB_LOCK *lock )
 {
        EntryInfo *new, ei;
-       DB_LOCK lock;
-       int rc;
+       int rc, purge = 0;
 #ifdef BDB_HIER
        struct berval rdn = e->e_name;
 #endif
@@ -934,7 +1092,7 @@ bdb_cache_add(
        /* Lock this entry so that bdb_add can run to completion.
         * It can only fail if BDB has run out of lock resources.
         */
-       rc = bdb_cache_entry_db_lock( bdb->bi_dbenv, locker, &ei, 1, 0, &lock );
+       rc = bdb_cache_entry_db_lock( bdb, txn, &ei, 0, 0, lock );
        if ( rc ) {
                bdb_cache_entryinfo_unlock( eip );
                return rc;
@@ -962,37 +1120,43 @@ bdb_cache_add(
        }
        new->bei_e = e;
        e->e_private = new;
-       new->bei_state = CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
+       new->bei_state |= CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
        eip->bei_state &= ~CACHE_ENTRY_NO_KIDS;
        if (eip->bei_parent) {
                eip->bei_parent->bei_state &= ~CACHE_ENTRY_NO_GRANDKIDS;
        }
        bdb_cache_entryinfo_unlock( eip );
 
-       ++bdb->bi_cache.c_cursize;
        ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
+       ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
+       ++bdb->bi_cache.c_cursize;
+       if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize &&
+               !bdb->bi_cache.c_purging ) {
+               purge = 1;
+               bdb->bi_cache.c_purging = 1;
+       }
+       ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
 
-       /* set lru mutex */
-       ldap_pvt_thread_mutex_lock( &bdb->bi_cache.lru_head_mutex );
+       bdb_cache_lru_link( bdb, new );
 
-       /* lru_head_mutex is unlocked for us */
-       bdb_cache_lru_add( bdb, new );
+       if ( purge )
+               bdb_cache_lru_purge( bdb );
 
        return rc;
 }
 
 int
 bdb_cache_modify(
+       struct bdb_info *bdb,
        Entry *e,
        Attribute *newAttrs,
-       DB_ENV *env,
-       u_int32_t locker,
+       DB_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.
@@ -1016,7 +1180,7 @@ bdb_cache_modrdn(
        struct berval *nrdn,
        Entry *new,
        EntryInfo *ein,
-       u_int32_t locker,
+       DB_TXN *txn,
        DB_LOCK *lock )
 {
        EntryInfo *ei = BEI(e), *pei;
@@ -1026,7 +1190,7 @@ bdb_cache_modrdn(
 #endif
 
        /* Get write lock on data */
-       rc =  bdb_cache_entry_db_relock( bdb->bi_dbenv, locker, ei, 1, 0, lock );
+       rc =  bdb_cache_entry_db_relock( bdb, txn, ei, 1, 0, lock );
        if ( rc ) return rc;
 
        /* If we've done repeated mods on a cached entry, then e_attrs
@@ -1051,6 +1215,7 @@ 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 );
 
@@ -1061,6 +1226,16 @@ bdb_cache_modrdn(
                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) {
@@ -1069,16 +1244,32 @@ 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
-       {
-               /* 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 );
+               /* 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;
@@ -1092,10 +1283,9 @@ 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);
@@ -1103,18 +1293,19 @@ bdb_cache_delete(
 
        assert( e->e_private != NULL );
 
+       /* 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 );
+       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 ) {
                /* couldn't lock, undo and give up */
                ei->bei_state ^= CACHE_ENTRY_DELETED;
-               bdb_cache_entryinfo_unlock( ei );
                return rc;
        }
 
@@ -1122,20 +1313,12 @@ bdb_cache_delete(
                e->e_id, 0, 0 );
 
        /* set lru mutex */
-       ldap_pvt_thread_mutex_lock( &cache->lru_tail_mutex );
+       ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_lru_mutex );
 
-       /* set cache write lock */
-       ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
-
-       rc = bdb_cache_delete_internal( cache, e->e_private, 1 );
-
-       /* free cache write lock */
-       ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
+       rc = bdb_cache_delete_internal( &bdb->bi_cache, e->e_private, 1 );
 
        /* free lru mutex */
-       ldap_pvt_thread_mutex_unlock( &cache->lru_tail_mutex );
-
-       /* Leave entry info locked */
+       ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
 
        return( rc );
 }
@@ -1145,6 +1328,8 @@ bdb_cache_delete_cleanup(
        Cache *cache,
        EntryInfo *ei )
 {
+       /* Enter with ei locked */
+
        if ( ei->bei_e ) {
                ei->bei_e->e_private = NULL;
 #ifdef SLAP_ZONE_ALLOC
@@ -1155,23 +1340,7 @@ bdb_cache_delete_cleanup(
                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_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_free( cache, ei );
        bdb_cache_entryinfo_unlock( ei );
 }
 
@@ -1182,6 +1351,7 @@ bdb_cache_delete_internal(
     int                decr )
 {
        int rc = 0;     /* return code */
+       int decr_leaf = 0;
 
        /* Lock the parent's kids tree */
        bdb_cache_entryinfo_lock( e->bei_parent );
@@ -1197,22 +1367,31 @@ bdb_cache_delete_internal(
                rc = -1;
        }
        if ( e->bei_parent->bei_kids )
-               cache->c_leaves--;
+               decr_leaf = 1;
+
+       bdb_cache_entryinfo_unlock( e->bei_parent );
 
+       ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
        /* id tree */
-       if ( avl_delete( &cache->c_idtree, (caddr_t) e, bdb_id_cmp ) == NULL ) {
+       if ( avl_delete( &cache->c_idtree, (caddr_t) e, bdb_id_cmp )) {
+               cache->c_eiused--;
+               if ( decr_leaf )
+                       cache->c_leaves--;
+       } else {
                rc = -1;
        }
+       ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
 
        if ( rc == 0 ){
-               cache->c_eiused--;
-
                /* lru */
-               LRU_DELETE( cache, e );
-               if ( e->bei_e ) cache->c_cursize--;
-       }
+               LRU_DEL( cache, e );
 
-       bdb_cache_entryinfo_unlock( e->bei_parent );
+               if ( e->bei_e ) {
+                       ldap_pvt_thread_mutex_lock( &cache->c_count_mutex );
+                       cache->c_cursize--;
+                       ldap_pvt_thread_mutex_unlock( &cache->c_count_mutex );
+               }
+       }
 
        return( rc );
 }
@@ -1241,7 +1420,7 @@ bdb_cache_release_all( Cache *cache )
        /* set cache write lock */
        ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
        /* set lru mutex */
-       ldap_pvt_thread_mutex_lock( &cache->lru_tail_mutex );
+       ldap_pvt_thread_mutex_lock( &cache->c_lru_mutex );
 
        Debug( LDAP_DEBUG_TRACE, "====> bdb_cache_release_all\n", 0, 0, 0 );
 
@@ -1260,7 +1439,7 @@ bdb_cache_release_all( Cache *cache )
        cache->c_dntree.bei_kids = NULL;
 
        /* free lru mutex */
-       ldap_pvt_thread_mutex_unlock( &cache->lru_tail_mutex );
+       ldap_pvt_thread_mutex_unlock( &cache->c_lru_mutex );
        /* free cache write lock */
        ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
 }
@@ -1272,51 +1451,57 @@ 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;
        }
 }
 #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;
-       u_int32_t lockid = (long)data;
-       int rc;
+       /* DB_ENV *env = key; */
+       DB_TXN *txn = data;
 
-       rc = XLOCK_ID_FREE( env, lockid );
-       if ( rc == EINVAL ) {
-               DB_LOCKREQ lr;
-               Debug( LDAP_DEBUG_ANY,
-                       "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 );
+       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();
+
+       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, u_int32_t *locker )
+bdb_reader_get( Operation *op, DB_ENV *env, DB_TXN **txn )
 {
        int i, rc;
-       u_int32_t lockid;
        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 ) {
@@ -1327,57 +1512,29 @@ bdb_locker_id( Operation *op, DB_ENV *env, u_int32_t *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 *)((long)lockid);
+               data = *txn;
                if ( ( rc = ldap_pvt_thread_pool_setkey( ctx, env,
-                       data, bdb_locker_id_free ) ) ) {
-                       XLOCK_ID_FREE( env, lockid );
-                       Debug( LDAP_DEBUG_ANY, "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 );
 
                        return rc;
                }
        } else {
-               lockid = (long)data;
+               *txn = data;
        }
-       *locker = lockid;
        return 0;
 }
-#endif /* BDB_REUSE_LOCKERS */
-
-void
-bdb_cache_delete_entry(
-       struct bdb_info *bdb,
-       EntryInfo *ei,
-       u_int32_t locker,
-       DB_LOCK *lock )
-{
-       ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
-       if ( bdb_cache_entry_db_lock( bdb->bi_dbenv, bdb->bi_cache.c_locker, ei, 1, 1, lock ) == 0 )
-       {
-               if ( ei->bei_e && !(ei->bei_state & CACHE_ENTRY_NOT_LINKED )) {
-                       LRU_DELETE( &bdb->bi_cache, ei );
-                       ei->bei_e->e_private = NULL;
-#ifdef SLAP_ZONE_ALLOC
-                       bdb_entry_return( bdb, ei->bei_e, ei->bei_zseq );
-#else
-                       bdb_entry_return( ei->bei_e );
-#endif
-                       ei->bei_e = NULL;
-                       --bdb->bi_cache.c_cursize;
-               }
-               bdb_cache_entry_db_unlock( bdb->bi_dbenv, lock );
-       }
-       ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
-}