]> git.sur5r.net Git - openldap/blobdiff - servers/slapd/back-bdb/cache.c
Ditch LRU cache replacement in favor of 2nd-chance/clock.
[openldap] / servers / slapd / back-bdb / cache.c
index db71ca70cea7e19f6ae097993412cb76ecc99242..86acd1316fb121c3786611f89add02aab3d16b1d 100644 (file)
@@ -29,9 +29,9 @@
 #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
@@ -40,30 +40,83 @@ 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;
                }
-               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;
 }
 
+/* Note - we now use a Second-Chance / Clock algorithm instead of
+ * Least-Recently-Used. This tremendously improves concurrency
+ * because we no longer need to manipulate the lists every time an
+ * entry is touched. We only need to lock the lists when adding
+ * or deleting an entry. It's now a circular doubly-linked list.
+ * We always append to the tail, but the head traverses the circle
+ * during a purge operation.
+ */
+static void
+bdb_cache_lru_link( Cache *cache, EntryInfo *ei )
+{
+       /* Insert into circular LRU list */
+       ldap_pvt_thread_mutex_lock( &cache->lru_tail_mutex );
+       ei->bei_lruprev = cache->c_lrutail;
+       if ( cache->c_lrutail ) {
+               ei->bei_lrunext = cache->c_lrutail->bei_lrunext;
+               cache->c_lrutail->bei_lrunext = ei;
+               if ( ei->bei_lrunext )
+                       ei->bei_lrunext->bei_lruprev = ei;
+       } else {
+               ei->bei_lrunext = ei->bei_lruprev = ei;
+               cache->c_lruhead = ei;
+       }
+       cache->c_lrutail = ei;
+       ldap_pvt_thread_mutex_unlock( &cache->lru_tail_mutex );
+}
+
+#ifdef NO_THREADS
+#define NO_DB_LOCK
+#endif
+
+/* #define NO_DB_LOCK 1 */
+/* Note: The BerkeleyDB locks are much slower than regular
+ * mutexes or rdwr locks. But the BDB implementation has the
+ * advantage of using a fixed size lock table, instead of
+ * allocating a lock object per entry in the DB. That's a
+ * key benefit for scaling. It also frees us from worrying
+ * about undetectable deadlocks between BDB activity and our
+ * own cache activity. It's still worth exploring faster
+ * alternatives though.
+ */
+
 /* Atomically release and reacquire a lock */
 int
 bdb_cache_entry_db_relock(
@@ -74,7 +127,7 @@ bdb_cache_entry_db_relock(
        int tryOnly,
        DB_LOCK *lock )
 {
-#ifdef NO_THREADS
+#ifdef NO_DB_LOCK
        return 0;
 #else
        int     rc;
@@ -110,7 +163,7 @@ static int
 bdb_cache_entry_db_lock( DB_ENV *env, u_int32_t locker, EntryInfo *ei,
        int rw, int tryOnly, DB_LOCK *lock )
 {
-#ifdef NO_THREADS
+#ifdef NO_DB_LOCK
        return 0;
 #else
        int       rc;
@@ -135,18 +188,18 @@ 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 )
 {
-#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 );
        return rc;
@@ -165,32 +218,6 @@ 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 )
@@ -265,6 +292,7 @@ bdb_entryinfo_add_internal(
                ei->bei_parent->bei_ckids++;
 #endif
        }
+       bdb_cache_lru_link( &bdb->bi_cache, ei2 );
 
        *res = ei2;
        return 0;
@@ -313,6 +341,7 @@ 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 ) {
@@ -390,7 +419,6 @@ 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 addlru = 0;
 
        ei.bei_id = id;
        ei.bei_kids = NULL;
@@ -416,7 +444,7 @@ 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 );
@@ -438,8 +466,8 @@ hdb_cache_find_parent(
                                ein->bei_ckids++;
                                bdb_cache_entryinfo_unlock( ein );
                        }
-                       addlru = 0;
-
+               } else {
+                       bdb_cache_lru_link( &bdb->bi_cache, ein );
                }
 
                /* If this is the first time, save this node
@@ -462,25 +490,21 @@ 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( ei2 );
                        ein->bei_parent = ei2;
+
+                       /* Reset all the state info */
+                       for (ein = eir; ein != ei2; ein=ein->bei_parent)
+                               ein->bei_state &= ~CACHE_ENTRY_NOT_LINKED;
+
                        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;
                        *res = eir;
                        break;
                }
@@ -529,144 +553,87 @@ int hdb_cache_load(
 }
 #endif
 
-static void *
-bdb_cache_lru_purge(void *ctx, void *arg)
+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;
-
-       if ( bdb->bi_cache.c_locker ) {
-               lockp = &lock;
-       } else {
-               lockp = NULL;
-       }
+       EntryInfo *elru, *elnext;
+       int count, islocked;
 
-       ldap_pvt_thread_mutex_lock( &bdb->bi_cache.lru_tail_mutex );
+       /* Don't bother if we can't get the lock */
+       if ( ldap_pvt_thread_mutex_trylock( &bdb->bi_cache.lru_head_mutex ) )
+               return;
+
+       if ( bdb->bi_cache.c_cursize <= bdb->bi_cache.c_maxsize ) {
+               ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_head_mutex );
+               return;
+       }
 
+       count = 0;
        /* 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 we can successfully writelock it, then
-                * the object is idle.
+               if ( ldap_pvt_thread_mutex_trylock( &elru->bei_kids_mutex ))
+                       continue;
+
+               /* This flag implements the clock replacement behavior */
+               if ( elru->bei_state & ( CACHE_ENTRY_REFERENCED )) {
+                       elru->bei_state &= ~CACHE_ENTRY_REFERENCED;
+                       bdb_cache_entryinfo_unlock( elru );
+                       continue;
+               }
+
+               /* If this node is in the process of linking into the cache,
+                * or this node is being deleted, skip it.
                 */
-               if ( bdb_cache_entry_db_lock( bdb->bi_dbenv,
-                               bdb->bi_cache.c_locker, elru, 1, 1, lockp ) == 0 ) {
+               if ( elru->bei_state & ( CACHE_ENTRY_NOT_LINKED |
+                       CACHE_ENTRY_DELETED | CACHE_ENTRY_LOADING )) {
+                       bdb_cache_entryinfo_unlock( elru );
+                       continue;
+               }
 
-                       int stop = 0;
+               islocked = 1;
 
-                       /* 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;
+               /* Free entry for this node if it's present */
+               if ( elru->bei_e ) {
+                       elru->bei_e->e_private = NULL;
 #ifdef SLAP_ZONE_ALLOC
-                               bdb_entry_return( bdb, elru->bei_e, elru->bei_zseq );
+                       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 );
-                               }
-                               /* 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 );
+                       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 ) {
+                               bdb_cache_delete_internal( &bdb->bi_cache, elru, 0 );
+                               bdb_cache_delete_cleanup( &bdb->bi_cache, elru );
+                               islocked = 0;
                        }
-                       if (stop) break;
+                       /* Leave node on LRU list for a future pass */
                }
-       }
 
-       ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_tail_mutex );
+               if ( islocked )
+                       bdb_cache_entryinfo_unlock( elru );
 
-       /* 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 );
+               if ( count >= bdb->bi_cache.c_minfree ) {
+                       ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
+                       bdb->bi_cache.c_cursize -= count;
+                       ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
+                       break;
+               }
        }
 
-       return NULL;
-}
-
-/* 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 );
+       bdb->bi_cache.c_lruhead = elru;
        ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_head_mutex );
-
-       /* 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;
-
-                       rtask.arg = bdb;
-                       bdb_cache_lru_purge( NULL, &rtask );
-               } else {
-                       int wake = 0;
-                       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 )) {
-                                       /* We want it to start right now */
-                                       bdb->bi_cache_task->interval.tv_sec = 0;
-                                       ldap_pvt_runqueue_resched( &slapd_rq, bdb->bi_cache_task,
-                                               0 );
-                                       /* But don't try to reschedule it while it's running */
-                                       bdb->bi_cache_task->interval.tv_sec = 3600;
-                                       wake = 1;
-                               }
-                       } else {
-                               bdb->bi_cache_task = ldap_pvt_runqueue_insert( &slapd_rq, 3600,
-                                       bdb_cache_lru_purge, bdb, "bdb_cache_lru_purge",
-                                       bdb->bi_dbenv_home );
-                               wake = 1;
-                       }
-                       ldap_pvt_thread_mutex_unlock( &slapd_rq.rq_mutex );
-                       /* Don't bother waking if the purge task is already running */
-                       if ( wake )
-                               slap_wake_listener();
-               }
-       }
 }
 
 EntryInfo *
@@ -755,6 +722,7 @@ again:      ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
                                &ep->e_nname, eip );
                        if ( *eip ) islocked = 1;
                        if ( rc ) {
+                               ep->e_private = NULL;
 #ifdef SLAP_ZONE_ALLOC
                                bdb_entry_return( bdb, ep, (*eip)->bei_zseq );
 #else
@@ -817,7 +785,9 @@ load1:
 #endif
                                                ep = NULL;
                                        }
+                                       bdb_cache_entryinfo_lock( *eip );
                                        (*eip)->bei_state ^= CACHE_ENTRY_LOADING;
+                                       bdb_cache_entryinfo_unlock( *eip );
                                        if ( rc == 0 ) {
                                                /* If we succeeded, downgrade back to a readlock. */
                                                rc = bdb_cache_entry_db_relock( bdb->bi_dbenv, locker,
@@ -859,6 +829,7 @@ load1:
                bdb_cache_entryinfo_unlock( *eip );
        }
        if ( ep ) {
+               ep->e_private = NULL;
 #ifdef SLAP_ZONE_ALLOC
                bdb_entry_return( bdb, ep, (*eip)->bei_zseq );
 #else
@@ -866,31 +837,17 @@ load1:
 #endif
        }
        if ( rc == 0 ) {
+               int purge = 0;
 
                if ( load ) {
-                       ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
+                       ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
                        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 );
-                       }
-                       /* lru_head_mutex is unlocked for us */
-                       bdb_cache_lru_add( bdb, *eip );
+                       if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize )
+                               purge = 1;
+                       ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
                }
+               if ( purge )
+                       bdb_cache_lru_purge( bdb );
        }
 
 #ifdef SLAP_ZONE_ALLOC
@@ -930,11 +887,11 @@ bdb_cache_add(
        EntryInfo *eip,
        Entry *e,
        struct berval *nrdn,
-       u_int32_t locker )
+       u_int32_t locker,
+       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
@@ -947,7 +904,7 @@ 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->bi_dbenv, locker, &ei, 0, 0, lock );
        if ( rc ) {
                bdb_cache_entryinfo_unlock( eip );
                return rc;
@@ -975,21 +932,22 @@ 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 )
+               purge = 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 );
-
-       /* lru_head_mutex is unlocked for us */
-       bdb_cache_lru_add( bdb, new );
+       if ( purge )
+               bdb_cache_lru_purge( bdb );
 
        return rc;
 }
@@ -1064,6 +1022,10 @@ bdb_cache_modrdn(
        avl_delete( &pei->bei_kids, (caddr_t) ei, bdb_rdn_cmp );
        free( ei->bei_nrdn.bv_val );
        ber_dupbv( &ei->bei_nrdn, nrdn );
+
+       if ( !pei->bei_kids )
+               pei->bei_state |= CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
+
 #ifdef BDB_HIER
        free( ei->bei_rdn.bv_val );
 
@@ -1074,6 +1036,8 @@ 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--;
 #endif
 
        if (!ein) {
@@ -1083,7 +1047,16 @@ bdb_cache_modrdn(
                bdb_cache_entryinfo_unlock( pei );
                bdb_cache_entryinfo_lock( ein );
        }
+       /* parent now has kids */
+       if ( ein->bei_state & CACHE_ENTRY_NO_KIDS )
+               ein->bei_state ^= CACHE_ENTRY_NO_KIDS;
+
 #ifdef BDB_HIER
+       /* 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 );
@@ -1091,6 +1064,8 @@ bdb_cache_modrdn(
                ei->bei_modrdns = bdb->bi_modrdns;
                ldap_pvt_thread_mutex_unlock( &bdb->bi_modrdns_mutex );
        }
+       ein->bei_ckids++;
+       if ( ein->bei_dkids ) ein->bei_dkids++;
 #endif
        avl_insert( &ein->bei_kids, ei, bdb_rdn_cmp, avl_dup_error );
        bdb_cache_entryinfo_unlock( ein );
@@ -1135,18 +1110,12 @@ bdb_cache_delete(
                e->e_id, 0, 0 );
 
        /* set lru mutex */
-       ldap_pvt_thread_mutex_lock( &cache->lru_tail_mutex );
-
-       /* set cache write lock */
-       ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
+       ldap_pvt_thread_mutex_lock( &cache->lru_head_mutex );
 
        rc = bdb_cache_delete_internal( cache, e->e_private, 1 );
 
-       /* free cache write lock */
-       ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
-
        /* free lru mutex */
-       ldap_pvt_thread_mutex_unlock( &cache->lru_tail_mutex );
+       ldap_pvt_thread_mutex_unlock( &cache->lru_head_mutex );
 
        /* Leave entry info locked */
 
@@ -1181,10 +1150,10 @@ bdb_cache_delete_cleanup(
        ei->bei_kids = NULL;
        ei->bei_lruprev = NULL;
 
-       ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
+       ldap_pvt_thread_mutex_lock( &cache->c_eifree_mutex );
        ei->bei_lrunext = cache->c_eifree;
        cache->c_eifree = ei;
-       ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
+       ldap_pvt_thread_mutex_unlock( &cache->c_eifree_mutex );
        bdb_cache_entryinfo_unlock( ei );
 }
 
@@ -1195,6 +1164,7 @@ 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 );
@@ -1210,22 +1180,39 @@ 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--;
-       }
+               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 );
+               }
 
-       bdb_cache_entryinfo_unlock( e->bei_parent );
+               if ( e->bei_lrunext ) e->bei_lrunext->bei_lruprev = e->bei_lruprev;
+               if ( e->bei_lruprev ) e->bei_lruprev->bei_lrunext = e->bei_lrunext;
+
+               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 );
 }
@@ -1368,29 +1355,3 @@ bdb_locker_id( Operation *op, DB_ENV *env, u_int32_t *locker )
        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 );
-}