]> git.sur5r.net Git - openldap/blobdiff - servers/slapd/back-bdb/cache.c
ITS#5439
[openldap] / servers / slapd / back-bdb / cache.c
index 7064c576b393afafb991b123a34165a237fd9f34..ff512977bc40d83fce6371cc3a793c3a6c854164 100644 (file)
@@ -2,7 +2,7 @@
 /* $OpenLDAP$ */
 /* This work is part of OpenLDAP Software <http://www.openldap.org/>.
  *
- * Copyright 2000-2007 The OpenLDAP Foundation.
+ * Copyright 2000-2008 The OpenLDAP Foundation.
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
 #ifdef BDB_HIER
 #define bdb_cache_lru_purge    hdb_cache_lru_purge
 #endif
-static void bdb_cache_lru_purge( struct bdb_info *bdb, uint32_t locker );
+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
@@ -62,6 +63,7 @@ bdb_cache_entryinfo_new( Cache *cache )
                if ( cache->c_eifree ) {
                        ei = cache->c_eifree;
                        cache->c_eifree = ei->bei_lrunext;
+                       ei->bei_finders = 0;
                }
                ldap_pvt_thread_mutex_unlock( &cache->c_eifree_mutex );
        }
@@ -75,6 +77,36 @@ bdb_cache_entryinfo_new( Cache *cache )
        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
@@ -84,22 +116,28 @@ bdb_cache_entryinfo_new( Cache *cache )
  * during a purge operation.
  */
 static void
-bdb_cache_lru_link( Cache *cache, EntryInfo *ei )
+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( &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;
+       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;
-               cache->c_lruhead = ei;
+               bdb->bi_cache.c_lruhead = ei;
        }
-       cache->c_lrutail = ei;
-       ldap_pvt_thread_mutex_unlock( &cache->lru_tail_mutex );
+       bdb->bi_cache.c_lrutail = ei;
+       ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
 }
 
 #ifdef NO_THREADS
@@ -121,7 +159,7 @@ bdb_cache_lru_link( Cache *cache, EntryInfo *ei )
 int
 bdb_cache_entry_db_relock(
        struct bdb_info *bdb,
-       u_int32_t locker,
+       BDB_LOCKER locker,
        EntryInfo *ei,
        int rw,
        int tryOnly,
@@ -145,7 +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 = bdb->bi_dbenv->lock_vec(bdb->bi_dbenv, locker, tryOnly ? DB_LOCK_NOWAIT : 0,
+       rc = bdb->bi_dbenv->lock_vec(bdb->bi_dbenv, BDB_LOCKID(locker), tryOnly ? DB_LOCK_NOWAIT : 0,
                list, 2, NULL );
 
        if (rc && !tryOnly) {
@@ -160,7 +198,7 @@ bdb_cache_entry_db_relock(
 }
 
 static int
-bdb_cache_entry_db_lock( struct bdb_info *bdb, u_int32_t locker, EntryInfo *ei,
+bdb_cache_entry_db_lock( struct bdb_info *bdb, BDB_LOCKER locker, EntryInfo *ei,
        int rw, int tryOnly, DB_LOCK *lock )
 {
 #ifdef NO_DB_LOCK
@@ -180,7 +218,7 @@ bdb_cache_entry_db_lock( struct bdb_info *bdb, u_int32_t locker, EntryInfo *ei,
        lockobj.data = &ei->bei_id;
        lockobj.size = sizeof(ei->bei_id) + 1;
 
-       rc = LOCK_GET(bdb->bi_dbenv, locker, tryOnly ? DB_LOCK_NOWAIT : 0,
+       rc = LOCK_GET(bdb->bi_dbenv, BDB_LOCKID(locker), tryOnly ? DB_LOCK_NOWAIT : 0,
                                        &lockobj, db_rw, lock);
        if (rc && !tryOnly) {
                Debug( LDAP_DEBUG_TRACE,
@@ -206,6 +244,41 @@ bdb_cache_entry_db_unlock ( struct bdb_info *bdb, DB_LOCK *lock )
 #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 )
 {
@@ -238,6 +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
@@ -265,10 +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
@@ -277,6 +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 );
 
@@ -286,13 +369,17 @@ 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
        }
-       bdb_cache_lru_link( &bdb->bi_cache, ei2 );
 
        *res = ei2;
        return 0;
@@ -307,7 +394,7 @@ bdb_entryinfo_add_internal(
 int
 bdb_cache_find_ndn(
        Operation       *op,
-       DB_TXN          *txn,
+       BDB_LOCKER              locker,
        struct berval   *ndn,
        EntryInfo       **res )
 {
@@ -345,6 +432,7 @@ bdb_cache_find_ndn(
                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 )) {
@@ -356,18 +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, locker, &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;
@@ -411,8 +508,7 @@ bdb_cache_find_ndn(
 int
 hdb_cache_find_parent(
        Operation *op,
-       DB_TXN *txn,
-       u_int32_t       locker,
+       BDB_LOCKER      locker,
        ID id,
        EntryInfo **res )
 {
@@ -425,7 +521,7 @@ hdb_cache_find_parent(
        ei.bei_ckids = 0;
 
        for (;;) {
-               rc = hdb_dn2id_parent( op, txn, locker, &ei, &eip.bei_id );
+               rc = hdb_dn2id_parent( op, locker, &ei, &eip.bei_id );
                if ( rc ) break;
 
                /* Save the previous node, if any */
@@ -449,14 +545,14 @@ hdb_cache_find_parent(
                /* 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 ) {
@@ -466,8 +562,6 @@ hdb_cache_find_parent(
                                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
@@ -492,6 +586,7 @@ hdb_cache_find_parent(
 
                /* 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;
 
@@ -504,7 +599,6 @@ hdb_cache_find_parent(
                                ein->bei_state &= ~CACHE_ENTRY_NOT_LINKED;
 
                        bdb_cache_entryinfo_unlock( ei2 );
-                       bdb_cache_entryinfo_lock( eir );
 
                        *res = eir;
                        break;
@@ -554,48 +648,64 @@ int hdb_cache_load(
 }
 #endif
 
+/* 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, uint32_t locker )
+bdb_cache_lru_purge( struct bdb_info *bdb )
 {
-       DB_LOCK         lock;
-       EntryInfo *elru, *elnext;
-       int count, islocked;
+       DB_LOCK         lock, *lockp;
+       EntryInfo *elru, *elnext = NULL;
+       int count, islocked, eimax;
 
-       /* Don't bother if we can't get the lock */
-       if ( ldap_pvt_thread_mutex_trylock( &bdb->bi_cache.lru_head_mutex ) )
-               return;
+       /* Wait for the mutex; we're the only one trying to purge. */
+       ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_lru_mutex );
 
        if ( bdb->bi_cache.c_cursize <= bdb->bi_cache.c_maxsize ) {
-               ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_head_mutex );
+               ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
+               bdb->bi_cache.c_purging = 0;
                return;
        }
 
+       if ( bdb->bi_cache.c_locker ) {
+               lockp = &lock;
+       } else {
+               lockp = NULL;
+       }
+
        count = 0;
+
+       /* 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;
+
        /* Look for an unused entry to remove */
-       for (elru = bdb->bi_cache.c_lruhead; elru; elru = elnext ) {
+       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;
+               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 );
-                       continue;
+                       goto bottom;
                }
 
                /* 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 ) {
+                       elru->bei_finders > 0 ) {
                        bdb_cache_entryinfo_unlock( elru );
-                       continue;
+                       goto bottom;
                }
 
                /* entryinfo is locked */
@@ -604,7 +714,8 @@ bdb_cache_lru_purge( struct bdb_info *bdb, uint32_t locker )
                /* If we can successfully writelock it, then
                 * the object is idle.
                 */
-               if ( bdb_cache_entry_db_lock( bdb, locker, elru, 1, 1, &lock ) == 0 ) {
+               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 ) {
@@ -617,23 +728,20 @@ bdb_cache_lru_purge( struct bdb_info *bdb, uint32_t locker )
                                elru->bei_e = NULL;
                                count++;
                        }
-                       bdb_cache_entry_db_unlock( bdb, &lock );
+                       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 it is a leaf node, and we're over the limit, free it.
                         */
-                       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 ( elru->bei_kids ) {
+                               /* Drop from list, we ignore it... */
+                               LRU_DEL( &bdb->bi_cache, elru );
+                       } else if ( bdb->bi_cache.c_leaves > eimax ) {
+                               /* 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;
+                       }       /* Leave on list until we need to free it */
                }
 
                if ( islocked )
@@ -645,10 +753,14 @@ bdb_cache_lru_purge( struct bdb_info *bdb, uint32_t locker )
                        ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
                        break;
                }
+bottom:
+               if ( elnext == bdb->bi_cache.c_lruhead )
+                       break;
        }
 
        bdb->bi_cache.c_lruhead = elnext;
-       ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_head_mutex );
+       ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
+       bdb->bi_cache.c_purging = 0;
 }
 
 EntryInfo *
@@ -670,7 +782,7 @@ 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.
  */
 
@@ -680,8 +792,8 @@ bdb_cache_find_id(
        DB_TXN  *tid,
        ID                              id,
        EntryInfo       **eip,
-       int             islocked,
-       u_int32_t       locker,
+       int             flag,
+       BDB_LOCKER      locker,
        DB_LOCK         *lock )
 {
        struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
@@ -701,8 +813,7 @@ 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
@@ -723,7 +834,7 @@ 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 );
        }
@@ -733,9 +844,9 @@ again:      ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
 #ifndef BDB_HIER
                rc = bdb_id2entry( op->o_bd, tid, locker, id, &ep );
                if ( rc == 0 ) {
-                       rc = bdb_cache_find_ndn( op, tid,
+                       rc = bdb_cache_find_ndn( op, locker,
                                &ep->e_nname, eip );
-                       if ( *eip ) islocked = 1;
+                       if ( *eip ) flag |= ID_LOCKED;
                        if ( rc ) {
                                ep->e_private = NULL;
 #ifdef SLAP_ZONE_ALLOC
@@ -747,16 +858,23 @@ 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 ) islocked = 1;
+               rc = hdb_cache_find_parent(op, locker, id, eip );
+               if ( rc == 0 ) flag |= ID_LOCKED;
 #endif
        }
 
        /* Ok, we found the info, do we have the entry? */
        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
@@ -770,9 +888,19 @@ 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, locker, *eip, load, 0, lock );
                        if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
@@ -793,6 +921,16 @@ 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 was loading the entry.
+                                                        */
+                                                       if ( (*eip)->bei_finders == 1 )
+                                                               (*eip)->bei_state |= CACHE_ENTRY_NOT_CACHED;
+                                                       bdb_cache_entryinfo_unlock( *eip );
+                                               }
                                        }
                                        if ( rc == 0 ) {
                                                /* If we succeeded, downgrade back to a readlock. */
@@ -802,16 +940,13 @@ load1:
                                                /* Otherwise, release the lock. */
                                                bdb_cache_entry_db_unlock( bdb, lock );
                                        }
-                                       bdb_cache_entryinfo_lock( *eip );
-                                       (*eip)->bei_state ^= CACHE_ENTRY_LOADING;
-                                       bdb_cache_entryinfo_unlock( *eip );
                                } 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;
+                                       flag |= ID_LOCKED;
                                        goto load1;
 #ifdef BDB_HIER
                                } else {
@@ -829,11 +964,15 @@ load1:
                                        }
 #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 ) {
@@ -848,14 +987,19 @@ load1:
                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 ( !( flag & ID_NOCACHE )) {
+                               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 );
+                       }
                }
                if ( purge )
-                       bdb_cache_lru_purge( bdb, locker );
+                       bdb_cache_lru_purge( bdb );
        }
 
 #ifdef SLAP_ZONE_ALLOC
@@ -895,7 +1039,7 @@ bdb_cache_add(
        EntryInfo *eip,
        Entry *e,
        struct berval *nrdn,
-       u_int32_t locker,
+       BDB_LOCKER locker,
        DB_LOCK *lock )
 {
        EntryInfo *new, ei;
@@ -950,12 +1094,17 @@ bdb_cache_add(
        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 )
+       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 );
 
+       bdb_cache_lru_link( bdb, new );
+
        if ( purge )
-               bdb_cache_lru_purge( bdb, locker );
+               bdb_cache_lru_purge( bdb );
 
        return rc;
 }
@@ -965,7 +1114,7 @@ bdb_cache_modify(
        struct bdb_info *bdb,
        Entry *e,
        Attribute *newAttrs,
-       u_int32_t locker,
+       BDB_LOCKER locker,
        DB_LOCK *lock )
 {
        EntryInfo *ei = BEI(e);
@@ -995,7 +1144,7 @@ bdb_cache_modrdn(
        struct berval *nrdn,
        Entry *new,
        EntryInfo *ein,
-       u_int32_t locker,
+       BDB_LOCKER locker,
        DB_LOCK *lock )
 {
        EntryInfo *ei = BEI(e), *pei;
@@ -1031,9 +1180,6 @@ bdb_cache_modrdn(
        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 );
 
@@ -1044,8 +1190,16 @@ bdb_cache_modrdn(
                rdn.bv_len = ptr - rdn.bv_val;
        }
        ber_dupbv( &ei->bei_rdn, &rdn );
-       pei->bei_ckids--;
-       if ( pei->bei_dkids ) pei->bei_dkids--;
+
+       /* 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) {
@@ -1054,27 +1208,32 @@ bdb_cache_modrdn(
                ei->bei_parent = ein;
                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;
 
+               /* 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
-       /* 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;
+               /* 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
        }
-       ein->bei_ckids++;
-       if ( ein->bei_dkids ) ein->bei_dkids++;
+
+#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;
@@ -1090,7 +1249,7 @@ int
 bdb_cache_delete(
        struct bdb_info *bdb,
     Entry              *e,
-    u_int32_t  locker,
+    BDB_LOCKER locker,
     DB_LOCK    *lock )
 {
        EntryInfo *ei = BEI(e);
@@ -1117,12 +1276,12 @@ bdb_cache_delete(
                e->e_id, 0, 0 );
 
        /* set lru mutex */
-       ldap_pvt_thread_mutex_lock( &bdb->bi_cache.lru_head_mutex );
+       ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_lru_mutex );
 
        rc = bdb_cache_delete_internal( &bdb->bi_cache, e->e_private, 1 );
 
        /* free lru mutex */
-       ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_head_mutex );
+       ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
 
        /* Leave entry info locked */
 
@@ -1144,23 +1303,7 @@ 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_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_free( cache, ei );
        bdb_cache_entryinfo_unlock( ei );
 }
 
@@ -1204,15 +1347,7 @@ bdb_cache_delete_internal(
 
        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 );
-               }
-
-               if ( e->bei_lrunext ) e->bei_lrunext->bei_lruprev = e->bei_lruprev;
-               if ( e->bei_lruprev ) e->bei_lruprev->bei_lrunext = e->bei_lrunext;
+               LRU_DEL( cache, e );
 
                if ( e->bei_e ) {
                        ldap_pvt_thread_mutex_lock( &cache->c_count_mutex );
@@ -1248,7 +1383,7 @@ 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 );
 
@@ -1267,7 +1402,7 @@ 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 );
 }
@@ -1279,11 +1414,11 @@ bdb_lru_print( Cache *cache )
 {
        EntryInfo       *e;
 
-       fprintf( stderr, "LRU circle head: %p\n", cache->c_lruhead );
+       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",
-                       e, e->bei_e, e->bei_id, e->bei_nrdn.bv_val );
+                       (void *) e, (void *) e->bei_e, e->bei_id, e->bei_nrdn.bv_val );
                e = e->bei_lrunext;
                if ( e == cache->c_lrutail )
                        break;
@@ -1291,7 +1426,7 @@ bdb_lru_print( Cache *cache )
        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 );
+                       (void *) e, (void *) e->bei_e, e->bei_id, e->bei_nrdn.bv_val );
                e = e->bei_lruprev;
                if ( e == cache->c_lrutail )
                        break;
@@ -1305,9 +1440,15 @@ static void
 bdb_locker_id_free( void *key, void *data )
 {
        DB_ENV *env = key;
-       u_int32_t lockid = (long)data;
+       u_int32_t lockid;
        int rc;
 
+#if DB_VERSION_FULL >= 0x04060012
+       BDB_LOCKER lptr = data;
+       lockid = lptr->id;
+#else
+       lockid = (long)data;
+#endif
        rc = XLOCK_ID_FREE( env, lockid );
        if ( rc == EINVAL ) {
                DB_LOCKREQ lr;
@@ -1322,8 +1463,21 @@ bdb_locker_id_free( void *key, void *data )
        }
 }
 
+/* free up any keys used by the main thread */
+void
+bdb_locker_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_locker_id_free( env, data );
+       }
+}
+
 int
-bdb_locker_id( Operation *op, DB_ENV *env, u_int32_t *locker )
+bdb_locker_id( Operation *op, DB_ENV *env, BDB_LOCKER *locker )
 {
        int i, rc;
        u_int32_t lockid;
@@ -1353,9 +1507,16 @@ bdb_locker_id( Operation *op, DB_ENV *env, u_int32_t *locker )
                if ( rc != 0) {
                        return rc;
                }
+#if DB_VERSION_FULL >= 0x04060012
+               { BDB_LOCKER lptr;
+               __lock_getlocker( env->lk_handle, lockid, 0, &lptr );
+               data = lptr;
+               }
+#else
                data = (void *)((long)lockid);
+#endif
                if ( ( rc = ldap_pvt_thread_pool_setkey( ctx, env,
-                       data, bdb_locker_id_free ) ) ) {
+                       data, bdb_locker_id_free, NULL, NULL ) ) ) {
                        XLOCK_ID_FREE( env, lockid );
                        Debug( LDAP_DEBUG_ANY, "bdb_locker_id: err %s(%d)\n",
                                db_strerror(rc), rc, 0 );
@@ -1365,7 +1526,11 @@ bdb_locker_id( Operation *op, DB_ENV *env, u_int32_t *locker )
        } else {
                lockid = (long)data;
        }
+#if DB_VERSION_FULL >= 0x04060012
+       *locker = data;
+#else
        *locker = lockid;
+#endif
        return 0;
 }
 #endif /* BDB_REUSE_LOCKERS */