]> git.sur5r.net Git - openldap/blobdiff - servers/slapd/back-bdb/cache.c
Use refcnt instead of keeping entry locked in Add response
[openldap] / servers / slapd / back-bdb / cache.c
index a20d772026b088dab805783916958dab54c6d497..6ca332131c9850f05e49984cdb3cd940d95717ec 100644 (file)
@@ -2,7 +2,7 @@
 /* $OpenLDAP$ */
 /* This work is part of OpenLDAP Software <http://www.openldap.org/>.
  *
- * Copyright 2000-2009 The OpenLDAP Foundation.
+ * Copyright 2000-2011 The OpenLDAP Foundation.
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
@@ -38,6 +38,7 @@ static int    bdb_cache_delete_internal(Cache *cache, EntryInfo *e, int decr);
 #define SLAPD_UNUSED
 #ifdef SLAPD_UNUSED
 static void    bdb_lru_print(Cache *cache);
+static void    bdb_idtree_print(Cache *cache);
 #endif
 #endif
 
@@ -94,17 +95,26 @@ bdb_cache_entryinfo_free( Cache *cache, EntryInfo *ei )
        ei->bei_kids = NULL;
        ei->bei_lruprev = NULL;
 
+#if 0
        ldap_pvt_thread_mutex_lock( &cache->c_eifree_mutex );
        ei->bei_lrunext = cache->c_eifree;
        cache->c_eifree = ei;
        ldap_pvt_thread_mutex_unlock( &cache->c_eifree_mutex );
+#else
+       ldap_pvt_thread_mutex_destroy( &ei->bei_kids_mutex );
+       ch_free( ei );
+#endif
 }
 
 #define LRU_DEL( c, e ) do { \
-       if ( e == (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; \
+       if ( e == e->bei_lruprev ) { \
+               (c)->c_lruhead = (c)->c_lrutail = NULL; \
+       } else { \
+               if ( e == (c)->c_lruhead ) (c)->c_lruhead = e->bei_lruprev; \
+               if ( e == (c)->c_lrutail ) (c)->c_lrutail = e->bei_lruprev; \
+               e->bei_lrunext->bei_lruprev = e->bei_lruprev; \
+               e->bei_lruprev->bei_lrunext = e->bei_lrunext; \
+       } \
        e->bei_lruprev = NULL; \
 } while ( 0 )
 
@@ -175,6 +185,7 @@ bdb_cache_entry_db_relock(
 
        if ( !lock ) return 0;
 
+       DBTzero( &lockobj );
        lockobj.data = &ei->bei_id;
        lockobj.size = sizeof(ei->bei_id) + 1;
 
@@ -216,6 +227,7 @@ bdb_cache_entry_db_lock( struct bdb_info *bdb, DB_TXN *txn, EntryInfo *ei,
        else
                db_rw = DB_LOCK_READ;
 
+       DBTzero( &lockobj );
        lockobj.data = &ei->bei_id;
        lockobj.size = sizeof(ei->bei_id) + 1;
 
@@ -253,9 +265,8 @@ bdb_cache_return_entry_rw( struct bdb_info *bdb, Entry *e,
        int free = 0;
 
        ei = e->e_private;
-       if ( ei &&
-               ( ei->bei_state & CACHE_ENTRY_NOT_CACHED ) &&
-               ( bdb_cache_entryinfo_trylock( ei ) == 0 )) {
+       if ( ei && ( ei->bei_state & CACHE_ENTRY_NOT_CACHED )) {
+               bdb_cache_entryinfo_lock( ei );
                if ( ei->bei_state & CACHE_ENTRY_NOT_CACHED ) {
                        /* Releasing the entry can only be done when
                         * we know that nobody else is using it, i.e we
@@ -334,8 +345,8 @@ bdb_entryinfo_add_internal(
 
        ei2 = bdb_cache_entryinfo_new( &bdb->bi_cache );
 
-       ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
        bdb_cache_entryinfo_lock( ei->bei_parent );
+       ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
 
        ei2->bei_id = ei->bei_id;
        ei2->bei_parent = ei->bei_parent;
@@ -372,13 +383,11 @@ bdb_entryinfo_add_internal(
                        bdb->bi_cache.c_leaves++;
                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++;
+               /* it's possible for hdb_cache_find_parent to beat us to it */
+               if ( !rc ) {
+                       ei->bei_parent->bei_ckids++;
+               }
 #endif
        }
 
@@ -433,7 +442,7 @@ bdb_cache_find_ndn(
                ei.bei_parent = eip;
                ei2 = (EntryInfo *)avl_find( eip->bei_kids, &ei, bdb_rdn_cmp );
                if ( !ei2 ) {
-                       DB_LOCK lock;
+                       DBC *cursor;
                        int len = ei.bei_nrdn.bv_len;
                                
                        if ( BER_BVISEMPTY( ndn )) {
@@ -443,16 +452,18 @@ bdb_cache_find_ndn(
 
                        ei.bei_nrdn.bv_len = ndn->bv_len -
                                (ei.bei_nrdn.bv_val - ndn->bv_val);
+                       eip->bei_finders++;
                        bdb_cache_entryinfo_unlock( eip );
 
                        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 );
+                       cursor = NULL;
+                       rc = bdb_dn2id( op, &ei.bei_nrdn, &ei, txn, &cursor );
                        if (rc) {
                                bdb_cache_entryinfo_lock( eip );
-                               bdb_cache_entry_db_unlock( bdb, &lock );
+                               eip->bei_finders--;
+                               if ( cursor ) cursor->c_close( cursor );
                                *res = eip;
                                return rc;
                        }
@@ -464,16 +475,20 @@ bdb_cache_find_ndn(
                        ei.bei_nrdn.bv_len = len;
                        rc = bdb_entryinfo_add_internal( bdb, &ei, &ei2 );
                        /* add_internal left eip and c_rwlock locked */
+                       eip->bei_finders--;
                        ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
-                       bdb_cache_entry_db_unlock( bdb, &lock );
+                       if ( cursor ) cursor->c_close( cursor );
                        if ( rc ) {
                                *res = eip;
                                return rc;
                        }
-               } else if ( ei2->bei_state & CACHE_ENTRY_DELETED ) {
+               }
+               bdb_cache_entryinfo_lock( ei2 );
+               if ( ei2->bei_state & CACHE_ENTRY_DELETED ) {
                        /* In the midst of deleting? Give it a chance to
                         * complete.
                         */
+                       bdb_cache_entryinfo_unlock( ei2 );
                        bdb_cache_entryinfo_unlock( eip );
                        ldap_pvt_thread_yield();
                        bdb_cache_entryinfo_lock( eip );
@@ -481,7 +496,6 @@ bdb_cache_find_ndn(
                        return DB_NOTFOUND;
                }
                bdb_cache_entryinfo_unlock( eip );
-               bdb_cache_entryinfo_lock( ei2 );
 
                eip = ei2;
 
@@ -515,7 +529,7 @@ hdb_cache_find_parent(
 {
        struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
        EntryInfo ei, eip, *ei2 = NULL, *ein = NULL, *eir = NULL;
-       int rc;
+       int rc, add;
 
        ei.bei_id = id;
        ei.bei_kids = NULL;
@@ -539,67 +553,92 @@ hdb_cache_find_parent(
                ein->bei_bdb = bdb;
 #endif
                ei.bei_ckids = 0;
+               add = 1;
                
                /* This node is not fully connected yet */
                ein->bei_state |= CACHE_ENTRY_NOT_LINKED;
 
+               /* If this is the first time, save this node
+                * to be returned later.
+                */
+               if ( eir == NULL ) {
+                       eir = ein;
+                       ein->bei_finders++;
+               }
+
+again:
                /* Insert this node into the ID tree */
                ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
                if ( avl_insert( &bdb->bi_cache.c_idtree, (caddr_t)ein,
                        bdb_id_cmp, bdb_id_dup_err ) ) {
                        EntryInfo *eix = ein->bei_lrunext;
 
+                       if ( bdb_cache_entryinfo_trylock( eix )) {
+                               ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
+                               ldap_pvt_thread_yield();
+                               goto again;
+                       }
+                       ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
+
                        /* Someone else created this node just before us.
                         * Free our new copy and use the existing one.
                         */
                        bdb_cache_entryinfo_free( &bdb->bi_cache, ein );
-                       ein = eix;
-                       
-                       /* Link in any kids we've already processed */
-                       if ( ei2 ) {
-                               bdb_cache_entryinfo_lock( ein );
-                               avl_insert( &ein->bei_kids, (caddr_t)ei2,
-                                       bdb_rdn_cmp, avl_dup_error );
-                               ein->bei_ckids++;
-                               bdb_cache_entryinfo_unlock( ein );
+
+                       /* if it was the node we were looking for, just return it */
+                       if ( eir == ein ) {
+                               *res = eix;
+                               rc = 0;
+                               break;
                        }
-               }
 
-               /* If this is the first time, save this node
-                * to be returned later.
-                */
-               if ( eir == NULL ) eir = ein;
+                       ein = ei2;
+                       ei2 = eix;
+                       add = 0;
+
+                       /* otherwise, link up what we have and return */
+                       goto gotparent;
+               }
 
                /* If there was a previous node, link it to this one */
                if ( ei2 ) ei2->bei_parent = ein;
 
                /* Look for this node's parent */
+par2:
                if ( eip.bei_id ) {
                        ei2 = (EntryInfo *) avl_find( bdb->bi_cache.c_idtree,
                                        (caddr_t) &eip, bdb_id_cmp );
                } else {
                        ei2 = &bdb->bi_cache.c_dntree;
                }
-               bdb->bi_cache.c_eiused++;
+               if ( ei2 && bdb_cache_entryinfo_trylock( ei2 )) {
+                       ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
+                       ldap_pvt_thread_yield();
+                       ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
+                       goto par2;
+               }
+               if ( add )
+                       bdb->bi_cache.c_eiused++;
                if ( ei2 && ( ei2->bei_kids || !ei2->bei_id ))
-                               bdb->bi_cache.c_leaves++;
+                       bdb->bi_cache.c_leaves++;
                ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
 
+gotparent:
                /* Got the parent, link in and we're done. */
                if ( ei2 ) {
                        bdb_cache_entryinfo_lock( 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++;
+                       if ( avl_insert( &ei2->bei_kids, (caddr_t)ein, bdb_rdn_cmp,
+                               avl_dup_error) == 0 )
+                               ei2->bei_ckids++;
 
                        /* Reset all the state info */
                        for (ein = eir; ein != ei2; ein=ein->bei_parent)
                                ein->bei_state &= ~CACHE_ENTRY_NOT_LINKED;
 
                        bdb_cache_entryinfo_unlock( ei2 );
+                       eir->bei_finders--;
 
                        *res = eir;
                        break;
@@ -658,8 +697,9 @@ bdb_cache_lru_purge( struct bdb_info *bdb )
 {
        DB_LOCK         lock, *lockp;
        EntryInfo *elru, *elnext = NULL;
-       int count, islocked, eimax;
-       int efree = 0, eifree = 0, eicount, ecount;
+       int islocked;
+       ID eicount, ecount;
+       ID count, efree, eifree = 0;
 #ifdef LDAP_DEBUG
        int iter;
 #endif
@@ -667,24 +707,24 @@ bdb_cache_lru_purge( struct bdb_info *bdb )
        /* Wait for the mutex; we're the only one trying to purge. */
        ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_lru_mutex );
 
+       if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize ) {
+               efree = bdb->bi_cache.c_cursize - bdb->bi_cache.c_maxsize;
+               efree += bdb->bi_cache.c_minfree;
+       } else {
+               efree = 0;
+       }
+
        /* maximum number of EntryInfo leaves to cache. In slapcat
         * we always free all leaf nodes.
         */
-       if ( slapMode & SLAP_TOOL_READONLY )
-               eimax = 0;
-       else
-               eimax = bdb->bi_cache.c_eimax;
-
-       efree = bdb->bi_cache.c_cursize - bdb->bi_cache.c_maxsize;
-       if ( efree < 1 )
-               efree = 0;
-       else 
-               efree += bdb->bi_cache.c_minfree;
 
-       if ( bdb->bi_cache.c_leaves > eimax ) {
+       if ( slapMode & SLAP_TOOL_READONLY ) {
+               eifree = bdb->bi_cache.c_leaves;
+       } else if ( bdb->bi_cache.c_eimax &&
+               bdb->bi_cache.c_leaves > bdb->bi_cache.c_eimax ) {
                eifree = bdb->bi_cache.c_minfree * 10;
-               if ( eifree >= eimax )
-                       eifree = eimax / 2;
+               if ( eifree >= bdb->bi_cache.c_leaves )
+                       eifree /= 2;
        }
 
        if ( !efree && !eifree ) {
@@ -724,12 +764,18 @@ bdb_cache_lru_purge( struct bdb_info *bdb )
                 * or this node is being deleted, skip it.
                 */
                if (( elru->bei_state & ( CACHE_ENTRY_NOT_LINKED |
-                       CACHE_ENTRY_DELETED | CACHE_ENTRY_LOADING )) ||
+                       CACHE_ENTRY_DELETED | CACHE_ENTRY_LOADING |
+                       CACHE_ENTRY_ONELEVEL )) ||
                        elru->bei_finders > 0 ) {
                        bdb_cache_entryinfo_unlock( elru );
                        goto bottom;
                }
 
+               if ( bdb_cache_entryinfo_trylock( elru->bei_parent )) {
+                       bdb_cache_entryinfo_unlock( elru );
+                       goto bottom;
+               }
+
                /* entryinfo is locked */
                islocked = 1;
 
@@ -782,20 +828,13 @@ bdb_cache_lru_purge( struct bdb_info *bdb )
                }
 
 next:
-               if ( islocked )
+               if ( islocked ) {
                        bdb_cache_entryinfo_unlock( elru );
+                       bdb_cache_entryinfo_unlock( elru->bei_parent );
+               }
 
-               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 );
-                       }
+               if ( count >= efree && eicount >= eifree )
                        break;
-               }
 bottom:
                if ( elnext == bdb->bi_cache.c_lruhead )
                        break;
@@ -804,28 +843,19 @@ bottom:
 #endif
        }
 
+       if ( count || ecount > bdb->bi_cache.c_cursize ) {
+               ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
+               /* HACK: we seem to be losing track, fix up now */
+               if ( ecount > bdb->bi_cache.c_cursize )
+                       bdb->bi_cache.c_cursize = ecount;
+               bdb->bi_cache.c_cursize -= count;
+               ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
+       }
        bdb->bi_cache.c_lruhead = elnext;
        ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
        bdb->bi_cache.c_purging = 0;
 }
 
-EntryInfo *
-bdb_cache_find_info(
-       struct bdb_info *bdb,
-       ID id )
-{
-       EntryInfo       ei = { 0 },
-                       *ei2;
-
-       ei.bei_id = id;
-
-       ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
-       ei2 = (EntryInfo *) avl_find( bdb->bi_cache.c_idtree,
-                                       (caddr_t) &ei, bdb_id_cmp );
-       ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
-       return ei2;
-}
-
 /*
  * cache_find_id - find an entry in the cache, given id.
  * The entry is locked for Read upon return. Call with flag ID_LOCKED if
@@ -859,11 +889,12 @@ again:    ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
                if ( *eip ) {
                        /* If the lock attempt fails, the info is in use */
                        if ( bdb_cache_entryinfo_trylock( *eip )) {
+                               int del = (*eip)->bei_state & CACHE_ENTRY_DELETED;
                                ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
                                /* If this node is being deleted, treat
                                 * as if the delete has already finished
                                 */
-                               if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
+                               if ( del ) {
                                        return DB_NOTFOUND;
                                }
                                /* otherwise, wait for the info to free up */
@@ -920,6 +951,10 @@ again:     ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
                } else {
                        (*eip)->bei_finders++;
                        (*eip)->bei_state |= CACHE_ENTRY_REFERENCED;
+                       if ( flag & ID_NOENTRY ) {
+                               bdb_cache_entryinfo_unlock( *eip );
+                               return 0;
+                       }
                        /* Make sure only one thread tries to load the entry */
 load1:
 #ifdef SLAP_ZONE_ALLOC
@@ -932,6 +967,7 @@ load1:
                        if ( !(*eip)->bei_e && !((*eip)->bei_state & CACHE_ENTRY_LOADING)) {
                                load = 1;
                                (*eip)->bei_state |= CACHE_ENTRY_LOADING;
+                               flag |= ID_CHKPURGE;
                        }
 
                        if ( !load ) {
@@ -940,9 +976,9 @@ load1:
                                 * another thread is currently loading it.
                                 */
                                if ( (*eip)->bei_state & CACHE_ENTRY_NOT_CACHED ) {
-                                       (*eip)->bei_state &= ~CACHE_ENTRY_NOT_CACHED;
+                                       (*eip)->bei_state ^= CACHE_ENTRY_NOT_CACHED;
+                                       flag |= ID_CHKPURGE;
                                }
-                               flag &= ~ID_NOCACHE;
                        }
 
                        if ( flag & ID_LOCKED ) {
@@ -953,6 +989,9 @@ load1:
                        if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
                                rc = DB_NOTFOUND;
                                bdb_cache_entry_db_unlock( bdb, lock );
+                               bdb_cache_entryinfo_lock( *eip );
+                               (*eip)->bei_finders--;
+                               bdb_cache_entryinfo_unlock( *eip );
                        } else if ( rc == 0 ) {
                                if ( load ) {
                                        if ( !ep) {
@@ -961,23 +1000,28 @@ load1:
                                        if ( rc == 0 ) {
                                                ep->e_private = *eip;
 #ifdef BDB_HIER
+                                               while ( (*eip)->bei_state & CACHE_ENTRY_NOT_LINKED )
+                                                       ldap_pvt_thread_yield();
                                                bdb_fix_dn( ep, 0 );
 #endif
+                                               bdb_cache_entryinfo_lock( *eip );
+
                                                (*eip)->bei_e = ep;
 #ifdef SLAP_ZONE_ALLOC
                                                (*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 )) {
+                                               if ( flag & ID_NOCACHE ) {
                                                        /* Set the cached state only if no other thread
                                                         * found the info while we were loading the entry.
                                                         */
-                                                       if ( (*eip)->bei_finders == 1 )
+                                                       if ( (*eip)->bei_finders == 1 ) {
                                                                (*eip)->bei_state |= CACHE_ENTRY_NOT_CACHED;
-                                                       bdb_cache_entryinfo_unlock( *eip );
+                                                               flag ^= ID_CHKPURGE;
+                                                       }
                                                }
+                                               bdb_cache_entryinfo_unlock( *eip );
+                                               bdb_cache_lru_link( bdb, *eip );
                                        }
                                        if ( rc == 0 ) {
                                                /* If we succeeded, downgrade back to a readlock. */
@@ -1011,12 +1055,12 @@ load1:
                                        }
 #endif
                                }
+                               bdb_cache_entryinfo_lock( *eip );
+                               (*eip)->bei_finders--;
+                               if ( load )
+                                       (*eip)->bei_state ^= CACHE_ENTRY_LOADING;
+                               bdb_cache_entryinfo_unlock( *eip );
                        }
-                       bdb_cache_entryinfo_lock( *eip );
-                       (*eip)->bei_finders--;
-                       if ( load )
-                               (*eip)->bei_state ^= CACHE_ENTRY_LOADING;
-                       bdb_cache_entryinfo_unlock( *eip );
                }
        }
        if ( flag & ID_LOCKED ) {
@@ -1033,20 +1077,17 @@ load1:
        if ( rc == 0 ) {
                int purge = 0;
 
-               if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize ||
-                       bdb->bi_cache.c_leaves > bdb->bi_cache.c_eimax ) {
+               if (( flag & ID_CHKPURGE ) || bdb->bi_cache.c_eimax ) {
                        ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
-                       if ( !bdb->bi_cache.c_purging ) {
-                               if ( !( flag & ID_NOCACHE )) {
-                                       bdb->bi_cache.c_cursize++;
-                                       if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize ) {
-                                               purge = 1;
-                                               bdb->bi_cache.c_purging = 1;
-                                       }
-                               } else if ( bdb->bi_cache.c_leaves > bdb->bi_cache.c_eimax ) {
+                       if ( flag & ID_CHKPURGE ) {
+                               bdb->bi_cache.c_cursize++;
+                               if ( !bdb->bi_cache.c_purging && bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize ) {
                                        purge = 1;
                                        bdb->bi_cache.c_purging = 1;
                                }
+                       } else if ( !bdb->bi_cache.c_purging && bdb->bi_cache.c_eimax && bdb->bi_cache.c_leaves > bdb->bi_cache.c_eimax ) {
+                               purge = 1;
+                               bdb->bi_cache.c_purging = 1;
                        }
                        ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
                }
@@ -1105,6 +1146,7 @@ bdb_cache_add(
        ei.bei_nrdn = *nrdn;
        ei.bei_lockpad = 0;
 
+#if 0
        /* Lock this entry so that bdb_add can run to completion.
         * It can only fail if BDB has run out of lock resources.
         */
@@ -1113,6 +1155,7 @@ bdb_cache_add(
                bdb_cache_entryinfo_unlock( eip );
                return rc;
        }
+#endif
 
 #ifdef BDB_HIER
        if ( nrdn->bv_len != e->e_nname.bv_len ) {
@@ -1124,6 +1167,12 @@ bdb_cache_add(
        if ( eip->bei_dkids ) eip->bei_dkids++;
 #endif
 
+       if (eip->bei_parent) {
+               bdb_cache_entryinfo_lock( eip->bei_parent );
+               eip->bei_parent->bei_state &= ~CACHE_ENTRY_NO_GRANDKIDS;
+               bdb_cache_entryinfo_unlock( eip->bei_parent );
+       }
+
        rc = bdb_entryinfo_add_internal( bdb, &ei, &new );
        /* bdb_csn_commit can cause this when adding the database root entry */
        if ( new->bei_e ) {
@@ -1138,9 +1187,6 @@ bdb_cache_add(
        e->e_private = new;
        new->bei_state |= CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
        eip->bei_state &= ~CACHE_ENTRY_NO_KIDS;
-       if (eip->bei_parent) {
-               eip->bei_parent->bei_state &= ~CACHE_ENTRY_NO_GRANDKIDS;
-       }
        bdb_cache_entryinfo_unlock( eip );
 
        ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
@@ -1153,6 +1199,7 @@ bdb_cache_add(
        }
        ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
 
+       new->bei_finders = 1;
        bdb_cache_lru_link( bdb, new );
 
        if ( purge )
@@ -1161,6 +1208,15 @@ bdb_cache_add(
        return rc;
 }
 
+void bdb_cache_deref(
+       EntryInfo *ei
+       )
+{
+       bdb_cache_entryinfo_lock( ei );
+       ei->bei_finders--;
+       bdb_cache_entryinfo_unlock( ei );
+}
+
 int
 bdb_cache_modify(
        struct bdb_info *bdb,
@@ -1305,7 +1361,7 @@ bdb_cache_delete(
     DB_LOCK    *lock )
 {
        EntryInfo *ei = BEI(e);
-       int     rc;
+       int     rc, busy = 0;
 
        assert( e->e_private != NULL );
 
@@ -1315,13 +1371,31 @@ bdb_cache_delete(
        /* Set this early, warn off any queriers */
        ei->bei_state |= CACHE_ENTRY_DELETED;
 
+       if (( ei->bei_state & ( CACHE_ENTRY_NOT_LINKED |
+               CACHE_ENTRY_LOADING | CACHE_ENTRY_ONELEVEL )) ||
+               ei->bei_finders > 0 )
+               busy = 1;
+
        bdb_cache_entryinfo_unlock( ei );
 
+       while ( busy ) {
+               ldap_pvt_thread_yield();
+               busy = 0;
+               bdb_cache_entryinfo_lock( ei );
+               if (( ei->bei_state & ( CACHE_ENTRY_NOT_LINKED |
+                       CACHE_ENTRY_LOADING | CACHE_ENTRY_ONELEVEL )) ||
+                       ei->bei_finders > 0 )
+                       busy = 1;
+               bdb_cache_entryinfo_unlock( ei );
+       }
+
        /* Get write lock on the data */
        rc = bdb_cache_entry_db_relock( bdb, txn, ei, 1, 0, lock );
        if ( rc ) {
+               bdb_cache_entryinfo_lock( ei );
                /* couldn't lock, undo and give up */
                ei->bei_state ^= CACHE_ENTRY_DELETED;
+               bdb_cache_entryinfo_unlock( ei );
                return rc;
        }
 
@@ -1331,7 +1405,10 @@ bdb_cache_delete(
        /* set lru mutex */
        ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_lru_mutex );
 
+       bdb_cache_entryinfo_lock( ei->bei_parent );
+       bdb_cache_entryinfo_lock( ei );
        rc = bdb_cache_delete_internal( &bdb->bi_cache, e->e_private, 1 );
+       bdb_cache_entryinfo_unlock( ei );
 
        /* free lru mutex */
        ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
@@ -1373,11 +1450,10 @@ bdb_cache_delete_internal(
        int decr_leaf = 0;
 
        /* already freed? */
-       if ( !e->bei_parent )
+       if ( !e->bei_parent ) {
+               assert(0);
                return -1;
-
-       /* Lock the parent's kids tree */
-       bdb_cache_entryinfo_lock( e->bei_parent );
+       }
 
 #ifdef BDB_HIER
        e->bei_parent->bei_ckids--;
@@ -1388,12 +1464,11 @@ bdb_cache_delete_internal(
                == NULL )
        {
                rc = -1;
+               assert(0);
        }
        if ( e->bei_parent->bei_kids )
                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 )) {
@@ -1402,8 +1477,10 @@ bdb_cache_delete_internal(
                        cache->c_leaves--;
        } else {
                rc = -1;
+               assert(0);
        }
        ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
+       bdb_cache_entryinfo_unlock( e->bei_parent );
 
        if ( rc == 0 ){
                /* lru */
@@ -1468,6 +1545,36 @@ bdb_cache_release_all( Cache *cache )
 }
 
 #ifdef LDAP_DEBUG
+static void
+bdb_lru_count( Cache *cache )
+{
+       EntryInfo       *e;
+       int ei = 0, ent = 0, nc = 0;
+
+       for ( e = cache->c_lrutail; ; ) {
+               ei++;
+               if ( e->bei_e ) {
+                       ent++;
+                       if ( e->bei_state & CACHE_ENTRY_NOT_CACHED )
+                               nc++;
+                       fprintf( stderr, "ei %d entry %p dn %s\n", ei, (void *) e->bei_e, e->bei_e->e_name.bv_val );
+               }
+               e = e->bei_lrunext;
+               if ( e == cache->c_lrutail )
+                       break;
+       }
+       fprintf( stderr, "counted %d entryInfos and %d entries, %d notcached\n",
+               ei, ent, nc );
+       ei = 0;
+       for ( e = cache->c_lrutail; ; ) {
+               ei++;
+               e = e->bei_lruprev;
+               if ( e == cache->c_lrutail )
+                       break;
+       }
+       fprintf( stderr, "counted %d entryInfos (on lruprev)\n", ei );
+}
+
 #ifdef SLAPD_UNUSED
 static void
 bdb_lru_print( Cache *cache )
@@ -1492,6 +1599,21 @@ bdb_lru_print( Cache *cache )
                        break;
        }
 }
+
+static int
+bdb_entryinfo_print(void *data, void *arg)
+{
+       EntryInfo *e = data;
+       fprintf( stderr, "\t%p, %p id %ld rdn \"%s\"\n",
+               (void *) e, (void *) e->bei_e, e->bei_id, e->bei_nrdn.bv_val );
+       return 0;
+}
+
+static void
+bdb_idtree_print(Cache *cache)
+{
+       avl_apply( cache->c_idtree, bdb_entryinfo_print, NULL, -1, AVL_INORDER );
+}
 #endif
 #endif