/* $OpenLDAP$ */
/* This work is part of OpenLDAP Software <http://www.openldap.org/>.
*
- * Copyright 2000-2004 The OpenLDAP Foundation.
+ * Copyright 2000-2006 The OpenLDAP Foundation.
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
#include "back-bdb.h"
-static int bdb_cache_delete_internal(Cache *cache, EntryInfo *e);
+#include "ldap_rq.h"
+
+#ifdef BDB_HIER
+#define bdb_cache_lru_purge hdb_cache_lru_purge
+#endif
+static void bdb_cache_lru_purge( struct bdb_info *bdb );
+
+static int bdb_cache_delete_internal(Cache *cache, EntryInfo *e, int decr);
#ifdef LDAP_DEBUG
+#ifdef SLAPD_UNUSED
static void bdb_lru_print(Cache *cache);
#endif
+#endif
+
+/* For concurrency experiments only! */
+#if 0
+#define ldap_pvt_thread_rdwr_wlock(a) 0
+#define ldap_pvt_thread_rdwr_wunlock(a) 0
+#define ldap_pvt_thread_rdwr_rlock(a) 0
+#define ldap_pvt_thread_rdwr_runlock(a) 0
+#endif
-static int bdb_txn_get( Operation *op, DB_ENV *env, DB_TXN **txn );
+#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(
- DB_ENV *env,
+ struct bdb_info *bdb,
u_int32_t locker,
EntryInfo *ei,
int rw,
int tryOnly,
DB_LOCK *lock )
{
-#ifdef NO_THREADS
+#ifdef NO_DB_LOCK
return 0;
#else
int rc;
list[1].lock = *lock;
list[1].mode = rw ? DB_LOCK_WRITE : DB_LOCK_READ;
list[1].obj = &lockobj;
- rc = env->lock_vec(env, locker, tryOnly ? DB_LOCK_NOWAIT : 0,
+ rc = bdb->bi_dbenv->lock_vec(bdb->bi_dbenv, locker, tryOnly ? DB_LOCK_NOWAIT : 0,
list, 2, NULL );
if (rc && !tryOnly) {
}
static int
-bdb_cache_entry_db_lock( DB_ENV *env, u_int32_t locker, EntryInfo *ei,
+bdb_cache_entry_db_lock( struct bdb_info *bdb, 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;
lockobj.data = &ei->bei_id;
lockobj.size = sizeof(ei->bei_id) + 1;
- rc = LOCK_GET(env, locker, tryOnly ? DB_LOCK_NOWAIT : 0,
+ rc = LOCK_GET(bdb->bi_dbenv, locker, tryOnly ? DB_LOCK_NOWAIT : 0,
&lockobj, db_rw, lock);
if (rc && !tryOnly) {
Debug( LDAP_DEBUG_TRACE,
ei->bei_id, rw, rc );
}
return rc;
-#endif /* NO_THREADS */
+#endif /* NO_DB_LOCK */
}
int
-bdb_cache_entry_db_unlock ( DB_ENV *env, DB_LOCK *lock )
+bdb_cache_entry_db_unlock ( struct bdb_info *bdb, DB_LOCK *lock )
{
-#ifdef NO_THREADS
+#ifdef NO_DB_LOCK
return 0;
#else
int rc;
- if ( !lock ) return 0;
+ if ( !lock || lock->mode == DB_LOCK_NG ) return 0;
- rc = LOCK_PUT ( env, lock );
+ rc = LOCK_PUT ( bdb->bi_dbenv, lock );
return rc;
#endif
}
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 )
#ifdef BDB_HIER
ei2->bei_rdn = ei->bei_rdn;
#endif
+#ifdef SLAP_ZONE_ALLOC
+ ei2->bei_bdb = bdb;
+#endif
/* Add to cache ID tree */
if (avl_insert( &bdb->bi_cache.c_idtree, ei2, bdb_id_cmp, avl_dup_error )) {
ei->bei_rdn.bv_val = NULL;
#endif
} else {
+ bdb->bi_cache.c_eiused++;
ber_dupbv( &ei2->bei_nrdn, &ei->bei_nrdn );
+
+ /* This is a new leaf node. But if parent had no kids, then it was
+ * a leaf and we would be decrementing that. So, only increment if
+ * the parent already has kids.
+ */
+ if ( ei->bei_parent->bei_kids || !ei->bei_parent->bei_id )
+ bdb->bi_cache.c_leaves++;
avl_insert( &ei->bei_parent->bei_kids, ei2, bdb_rdn_cmp,
avl_dup_error );
#ifdef BDB_HIER
ei->bei_parent->bei_ckids++;
#endif
}
+ bdb_cache_lru_link( &bdb->bi_cache, ei2 );
*res = ei2;
return 0;
}
for ( bdb_cache_entryinfo_lock( eip ); eip; ) {
+ eip->bei_state |= CACHE_ENTRY_REFERENCED;
ei.bei_parent = eip;
ei2 = (EntryInfo *)avl_find( eip->bei_kids, &ei, bdb_rdn_cmp );
if ( !ei2 ) {
int len = ei.bei_nrdn.bv_len;
+ if ( BER_BVISEMPTY( ndn )) {
+ *res = eip;
+ return LDAP_SUCCESS;
+ }
+
ei.bei_nrdn.bv_len = ndn->bv_len -
(ei.bei_nrdn.bv_val - ndn->bv_val);
bdb_cache_entryinfo_unlock( eip );
/* Walk up the tree from a child node, looking for an ID that's already
* been linked into the cache.
*/
-static int
+int
hdb_cache_find_parent(
Operation *op,
DB_TXN *txn,
+ u_int32_t locker,
ID id,
EntryInfo **res )
{
struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
EntryInfo ei, eip, *ei2 = NULL, *ein = NULL, *eir = NULL;
- char ndn[SLAP_LDAPDN_MAXLEN];
- ID parent;
int rc;
- int addlru = 1;
ei.bei_id = id;
ei.bei_kids = NULL;
+ ei.bei_ckids = 0;
for (;;) {
- rc = hdb_dn2id_parent( op, txn, &ei, &eip.bei_id );
+ rc = hdb_dn2id_parent( op, txn, locker, &ei, &eip.bei_id );
if ( rc ) break;
/* Save the previous node, if any */
ein->bei_kids = ei.bei_kids;
ein->bei_nrdn = ei.bei_nrdn;
ein->bei_rdn = ei.bei_rdn;
+ ein->bei_ckids = ei.bei_ckids;
+#ifdef SLAP_ZONE_ALLOC
+ ein->bei_bdb = bdb;
+#endif
+ ei.bei_ckids = 0;
/* This node is not fully connected yet */
- ein->bei_state = CACHE_ENTRY_NOT_LINKED;
+ ein->bei_state |= CACHE_ENTRY_NOT_LINKED;
/* Insert this node into the ID tree */
ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
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 ( !eir ) {
- addlru = 0;
- }
+ } else {
+ bdb_cache_lru_link( &bdb->bi_cache, ein );
}
/* If this is the first time, save this node
} else {
ei2 = &bdb->bi_cache.c_dntree;
}
+ bdb->bi_cache.c_eiused++;
+ if ( ei2 && ( ei2->bei_kids || !ei2->bei_id ))
+ bdb->bi_cache.c_leaves++;
ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
/* Got the parent, link in and we're done. */
if ( ei2 ) {
bdb_cache_entryinfo_lock( ei2 );
ein->bei_parent = ei2;
+
avl_insert( &ei2->bei_kids, (caddr_t)ein, bdb_rdn_cmp,
avl_dup_error);
- bdb_cache_entryinfo_unlock( ei2 );
- bdb_cache_entryinfo_lock( eir );
+ ei2->bei_ckids++;
/* Reset all the state info */
for (ein = eir; ein != ei2; ein=ein->bei_parent)
ein->bei_state &= ~CACHE_ENTRY_NOT_LINKED;
+
+ bdb_cache_entryinfo_unlock( ei2 );
+ bdb_cache_entryinfo_lock( eir );
+
*res = eir;
break;
}
ei.bei_kids = NULL;
ei.bei_id = eip.bei_id;
+ ei.bei_ckids = 1;
avl_insert( &ei.bei_kids, (caddr_t)ein, bdb_rdn_cmp,
avl_dup_error );
}
}
#endif
-/* caller must have lru_mutex locked. mutex
- * will be unlocked on return.
- */
static void
-bdb_cache_lru_add(
- struct bdb_info *bdb,
- u_int32_t locker,
- EntryInfo *ei )
+bdb_cache_lru_purge( struct bdb_info *bdb )
{
DB_LOCK lock, *lockp;
+ EntryInfo *elru, *elnext;
+ int count, islocked;
+
+ /* Don't bother if we can't get the lock */
+ if ( ldap_pvt_thread_mutex_trylock( &bdb->bi_cache.lru_head_mutex ) )
+ return;
- if ( locker ) {
+ if ( bdb->bi_cache.c_cursize <= bdb->bi_cache.c_maxsize ) {
+ ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_head_mutex );
+ return;
+ }
+
+ if ( bdb->bi_cache.c_locker ) {
lockp = &lock;
} else {
lockp = NULL;
}
- /* See if we're above the cache size limit */
- if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize ) {
- EntryInfo *elru, *elprev;
- int i = 0;
+ count = 0;
+ /* Look for an unused entry to remove */
+ for (elru = bdb->bi_cache.c_lruhead; elru; elru = elnext ) {
+ elnext = elru->bei_lrunext;
- /* Look for an unused entry to remove */
- for (elru = bdb->bi_cache.c_lrutail; elru; elru = elprev, i++ ) {
- elprev = elru->bei_lruprev;
+ if ( ldap_pvt_thread_mutex_trylock( &elru->bei_kids_mutex ))
+ continue;
- /* Too many probes, not enough idle, give up */
- if (i > 10) break;
+ /* 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 we can successfully writelock it, then
- * the object is idle.
- */
- if ( bdb_cache_entry_db_lock( bdb->bi_dbenv, bdb->bi_cache.c_locker, elru, 1, 1,
- lockp ) == 0 ) {
- /* If there's no entry, or this node is in
- * the process of linking into the cache,
- * skip it.
- */
- if ( !elru->bei_e || (elru->bei_state & CACHE_ENTRY_NOT_LINKED) ) {
- bdb_cache_entry_db_unlock( bdb->bi_dbenv, lockp );
- continue;
- }
- LRU_DELETE( &bdb->bi_cache, elru );
+ /* If this node is in the process of linking into the cache,
+ * or this node is being deleted, skip it.
+ *
+ * Also, if this node has no entry attached, skip it, there's
+ * nothing to purge anyway.
+ */
+ if (( elru->bei_state & ( CACHE_ENTRY_NOT_LINKED |
+ CACHE_ENTRY_DELETED | CACHE_ENTRY_LOADING )) ||
+ !elru->bei_e ) {
+ bdb_cache_entryinfo_unlock( elru );
+ continue;
+ }
+
+ /* entryinfo is locked */
+ islocked = 1;
+
+ /* If we can successfully writelock it, then
+ * the object is idle.
+ */
+ if ( bdb_cache_entry_db_lock( bdb,
+ bdb->bi_cache.c_locker, elru, 1, 1, lockp ) == 0 ) {
+
+ /* Free entry for this node if it's present */
+ if ( elru->bei_e ) {
elru->bei_e->e_private = NULL;
+#ifdef SLAP_ZONE_ALLOC
+ bdb_entry_return( bdb, elru->bei_e, elru->bei_zseq );
+#else
bdb_entry_return( elru->bei_e );
+#endif
elru->bei_e = NULL;
- bdb_cache_entry_db_unlock( bdb->bi_dbenv, lockp );
- --bdb->bi_cache.c_cursize;
- if (bdb->bi_cache.c_cursize < bdb->bi_cache.c_maxsize)
- break;
+ count++;
}
+ bdb_cache_entry_db_unlock( bdb, lockp );
+
+ /* ITS#4010 if we're in slapcat, and this node is a leaf
+ * node, free it.
+ *
+ * FIXME: we need to do this for slapd as well, (which is
+ * why we compute bi_cache.c_leaves now) but at the moment
+ * we can't because it causes unresolvable deadlocks.
+ */
+ if ( slapMode & SLAP_TOOL_READONLY ) {
+ if ( !elru->bei_kids ) {
+ bdb_cache_delete_internal( &bdb->bi_cache, elru, 0 );
+ bdb_cache_delete_cleanup( &bdb->bi_cache, elru );
+ islocked = 0;
+ }
+ /* Leave node on LRU list for a future pass */
+ }
+ }
+
+ if ( islocked )
+ bdb_cache_entryinfo_unlock( elru );
+
+ if ( count >= bdb->bi_cache.c_minfree ) {
+ ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
+ bdb->bi_cache.c_cursize -= count;
+ ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
+ break;
}
}
- LRU_ADD( &bdb->bi_cache, ei );
- ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_mutex );
+
+ bdb->bi_cache.c_lruhead = elnext;
+ ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.lru_head_mutex );
}
EntryInfo *
{
struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
Entry *ep = NULL;
- int rc = 0;
+ int rc = 0, load = 0;
EntryInfo ei = { 0 };
ei.bei_id = id;
+#ifdef SLAP_ZONE_ALLOC
+ slap_zh_rlock(bdb->bi_cache.c_zctx);
+#endif
/* If we weren't given any info, see if we have it already cached */
if ( !*eip ) {
again: ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
/* See if the ID exists in the database; add it to the cache if so */
if ( !*eip ) {
#ifndef BDB_HIER
- rc = bdb_id2entry( op->o_bd, tid, id, &ep );
+ rc = bdb_id2entry( op->o_bd, tid, locker, id, &ep );
if ( rc == 0 ) {
rc = bdb_cache_find_ndn( op, tid,
&ep->e_nname, eip );
if ( *eip ) islocked = 1;
if ( rc ) {
+ ep->e_private = NULL;
+#ifdef SLAP_ZONE_ALLOC
+ bdb_entry_return( bdb, ep, (*eip)->bei_zseq );
+#else
bdb_entry_return( ep );
+#endif
ep = NULL;
}
}
#else
- rc = hdb_cache_find_parent(op, tid, id, eip );
- if ( rc == 0 && *eip ) islocked = 1;
+ rc = hdb_cache_find_parent(op, tid, locker, id, eip );
+ if ( rc == 0 ) islocked = 1;
#endif
}
/* Ok, we found the info, do we have the entry? */
- if ( *eip && rc == 0 ) {
+ if ( rc == 0 ) {
if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
rc = DB_NOTFOUND;
} else {
- int load = 0;
/* Make sure only one thread tries to load the entry */
-load1: if ( !(*eip)->bei_e && !((*eip)->bei_state & CACHE_ENTRY_LOADING)) {
+load1:
+#ifdef SLAP_ZONE_ALLOC
+ if ((*eip)->bei_e && !slap_zn_validate(
+ bdb->bi_cache.c_zctx, (*eip)->bei_e, (*eip)->bei_zseq)) {
+ (*eip)->bei_e = NULL;
+ (*eip)->bei_zseq = 0;
+ }
+#endif
+ if ( !(*eip)->bei_e && !((*eip)->bei_state & CACHE_ENTRY_LOADING)) {
load = 1;
(*eip)->bei_state |= CACHE_ENTRY_LOADING;
}
bdb_cache_entryinfo_unlock( *eip );
islocked = 0;
}
- rc = bdb_cache_entry_db_lock( bdb->bi_dbenv, locker, *eip, 0, 0, lock );
+ rc = bdb_cache_entry_db_lock( bdb, locker, *eip, load, 0, lock );
if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
rc = DB_NOTFOUND;
- bdb_cache_entry_db_unlock( bdb->bi_dbenv, lock );
+ bdb_cache_entry_db_unlock( bdb, lock );
} else if ( rc == 0 ) {
if ( load ) {
- DB_TXN *ltid;
- u_int32_t locker2 = locker;
-
- /* We don't wrap entire read operations in txn's, but
- * we need our cache entry lock and any DB page locks
- * to be associated, in order for deadlock detection
- * to work properly. So if we need to read from the DB,
- * we use a long-lived per-thread txn for this step.
- */
- if ( !ep && !tid ) {
- rc = bdb_txn_get( op, bdb->bi_dbenv, <id );
- if ( ltid )
- locker2 = TXN_ID( ltid );
- } else {
- ltid = tid;
- }
- /* Give up original read lock, obtain write lock with
- * (possibly) new locker ID.
- */
- if ( rc == 0 ) {
- rc = bdb_cache_entry_db_relock( bdb->bi_dbenv, locker2,
- *eip, 1, 0, lock );
- }
- if ( rc == 0 && !ep) {
- rc = bdb_id2entry( op->o_bd, ltid, id, &ep );
+ if ( !ep) {
+ rc = bdb_id2entry( op->o_bd, tid, locker, id, &ep );
}
if ( rc == 0 ) {
ep->e_private = *eip;
bdb_fix_dn( ep, 0 );
#endif
(*eip)->bei_e = ep;
+#ifdef SLAP_ZONE_ALLOC
+ (*eip)->bei_zseq = *((ber_len_t *)ep - 2);
+#endif
ep = NULL;
}
+ bdb_cache_entryinfo_lock( *eip );
(*eip)->bei_state ^= CACHE_ENTRY_LOADING;
+ bdb_cache_entryinfo_unlock( *eip );
if ( rc == 0 ) {
/* If we succeeded, downgrade back to a readlock. */
- rc = bdb_cache_entry_db_relock( bdb->bi_dbenv, locker,
+ rc = bdb_cache_entry_db_relock( bdb, locker,
*eip, 0, 0, lock );
} else {
/* Otherwise, release the lock. */
- bdb_cache_entry_db_unlock( bdb->bi_dbenv, lock );
- }
- if ( locker2 != locker ) {
- /* If we're using the per-thread txn, release all
- * of its page locks now.
- */
- DB_LOCKREQ list;
- list.op = DB_LOCK_PUT_ALL;
- list.obj = NULL;
- bdb->bi_dbenv->lock_vec( bdb->bi_dbenv, locker2,
- 0, &list, 1, NULL );
+ bdb_cache_entry_db_unlock( bdb, lock );
}
} else if ( !(*eip)->bei_e ) {
/* Some other thread is trying to load the entry,
- * give it a chance to finish.
+ * wait for it to finish.
*/
- bdb_cache_entry_db_unlock( bdb->bi_dbenv, lock );
- ldap_pvt_thread_yield();
+ bdb_cache_entry_db_unlock( bdb, lock );
bdb_cache_entryinfo_lock( *eip );
islocked = 1;
goto load1;
*/
rc = bdb_fix_dn( (*eip)->bei_e, 1 );
if ( rc ) {
- bdb_cache_entry_db_relock( bdb->bi_dbenv,
+ bdb_cache_entry_db_relock( bdb,
locker, *eip, 1, 0, lock );
/* check again in case other modifier did it already */
if ( bdb_fix_dn( (*eip)->bei_e, 1 ) )
rc = bdb_fix_dn( (*eip)->bei_e, 2 );
- bdb_cache_entry_db_relock( bdb->bi_dbenv,
+ bdb_cache_entry_db_relock( bdb,
locker, *eip, 0, 0, lock );
}
#endif
bdb_cache_entryinfo_unlock( *eip );
}
if ( ep ) {
+ ep->e_private = NULL;
+#ifdef SLAP_ZONE_ALLOC
+ bdb_entry_return( bdb, ep, (*eip)->bei_zseq );
+#else
bdb_entry_return( ep );
+#endif
}
if ( rc == 0 ) {
- /* set lru mutex */
- ldap_pvt_thread_mutex_lock( &bdb->bi_cache.lru_mutex );
- /* if entry is on LRU list, remove from old spot */
- if ( (*eip)->bei_lrunext || (*eip)->bei_lruprev ) {
- LRU_DELETE( &bdb->bi_cache, *eip );
- } else {
- /* if entry is new, bump cache size */
+ 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 );
}
- /* lru_mutex is unlocked for us */
- bdb_cache_lru_add( bdb, locker, *eip );
+ if ( purge )
+ bdb_cache_lru_purge( bdb );
}
+#ifdef SLAP_ZONE_ALLOC
+ if (rc == 0 && (*eip)->bei_e) {
+ slap_zn_rlock(bdb->bi_cache.c_zctx, (*eip)->bei_e);
+ }
+ slap_zh_runlock(bdb->bi_cache.c_zctx);
+#endif
return rc;
}
EntryInfo *eip,
Entry *e,
struct berval *nrdn,
- u_int32_t locker )
+ u_int32_t locker,
+ DB_LOCK *lock )
{
EntryInfo *new, ei;
+ int rc, purge = 0;
+#ifdef BDB_HIER
struct berval rdn = e->e_name;
- DB_LOCK lock;
- int rc;
+#endif
ei.bei_id = e->e_id;
ei.bei_parent = eip;
/* 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, locker, &ei, 0, 0, lock );
if ( rc ) {
bdb_cache_entryinfo_unlock( eip );
return rc;
#ifdef BDB_HIER
if ( nrdn->bv_len != e->e_nname.bv_len ) {
- char *ptr = strchr( rdn.bv_val, ',' );
+ char *ptr = ber_bvchr( &rdn, ',' );
+ assert( ptr != NULL );
rdn.bv_len = ptr - rdn.bv_val;
}
ber_dupbv( &ei.bei_rdn, &rdn );
/* bdb_csn_commit can cause this when adding the database root entry */
if ( new->bei_e ) {
new->bei_e->e_private = NULL;
+#ifdef SLAP_ZONE_ALLOC
+ bdb_entry_return( bdb, new->bei_e, new->bei_zseq );
+#else
bdb_entry_return( new->bei_e );
+#endif
}
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 );
- /* set lru mutex */
- ldap_pvt_thread_mutex_lock( &bdb->bi_cache.lru_mutex );
+ 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;
- /* lru_mutex is unlocked for us */
- bdb_cache_lru_add( bdb, locker, new );
+ 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 );
- bdb_cache_entryinfo_unlock( eip );
- ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
return rc;
}
int
bdb_cache_modify(
+ struct bdb_info *bdb,
Entry *e,
Attribute *newAttrs,
- DB_ENV *env,
u_int32_t locker,
DB_LOCK *lock )
{
EntryInfo *ei = BEI(e);
int rc;
/* Get write lock on data */
- rc = bdb_cache_entry_db_relock( env, locker, ei, 1, 0, lock );
+ rc = bdb_cache_entry_db_relock( bdb, locker, ei, 1, 0, lock );
/* If we've done repeated mods on a cached entry, then e_attrs
* is no longer contiguous with the entry, and must be freed.
*/
int
bdb_cache_modrdn(
+ struct bdb_info *bdb,
Entry *e,
struct berval *nrdn,
Entry *new,
EntryInfo *ein,
- DB_ENV *env,
u_int32_t locker,
DB_LOCK *lock )
{
EntryInfo *ei = BEI(e), *pei;
- struct berval rdn;
int rc;
+#ifdef BDB_HIER
+ struct berval rdn;
+#endif
/* Get write lock on data */
- rc = bdb_cache_entry_db_relock( env, locker, ei, 1, 0, lock );
+ rc = bdb_cache_entry_db_relock( bdb, locker, ei, 1, 0, lock );
if ( rc ) return rc;
/* If we've done repeated mods on a cached entry, then e_attrs
avl_delete( &pei->bei_kids, (caddr_t) ei, bdb_rdn_cmp );
free( ei->bei_nrdn.bv_val );
ber_dupbv( &ei->bei_nrdn, nrdn );
+
+ if ( !pei->bei_kids )
+ pei->bei_state |= CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
+
#ifdef BDB_HIER
free( ei->bei_rdn.bv_val );
rdn = e->e_name;
if ( nrdn->bv_len != e->e_nname.bv_len ) {
- char *ptr = strchr(rdn.bv_val, ',');
+ char *ptr = ber_bvchr(&rdn, ',');
+ assert( ptr != NULL );
rdn.bv_len = ptr - rdn.bv_val;
}
ber_dupbv( &ei->bei_rdn, &rdn );
+ pei->bei_ckids--;
+ if ( pei->bei_dkids ) pei->bei_dkids--;
#endif
if (!ein) {
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;
+
{
- int max = ei->bei_modrdns;
/* Record the generation number of this change */
- for ( pei = ein; pei->bei_parent; pei = pei->bei_parent ) {
- if ( pei->bei_modrdns > max ) max = pei->bei_modrdns;
- }
- ei->bei_modrdns = max + 1;
+ ldap_pvt_thread_mutex_lock( &bdb->bi_modrdns_mutex );
+ bdb->bi_modrdns++;
+ ei->bei_modrdns = bdb->bi_modrdns;
+ ldap_pvt_thread_mutex_unlock( &bdb->bi_modrdns_mutex );
}
+ ein->bei_ckids++;
+ if ( ein->bei_dkids ) ein->bei_dkids++;
#endif
avl_insert( &ein->bei_kids, ei, bdb_rdn_cmp, avl_dup_error );
bdb_cache_entryinfo_unlock( ein );
*/
int
bdb_cache_delete(
- Cache *cache,
+ struct bdb_info *bdb,
Entry *e,
- DB_ENV *env,
u_int32_t locker,
DB_LOCK *lock )
{
EntryInfo *ei = BEI(e);
int rc;
- assert( e->e_private );
+ assert( e->e_private != NULL );
/* Set this early, warn off any queriers */
ei->bei_state |= CACHE_ENTRY_DELETED;
bdb_cache_entryinfo_lock( ei );
/* Get write lock on the data */
- rc = bdb_cache_entry_db_relock( env, locker, ei, 1, 0, lock );
+ rc = bdb_cache_entry_db_relock( bdb, locker, ei, 1, 0, lock );
if ( rc ) {
/* couldn't lock, undo and give up */
ei->bei_state ^= CACHE_ENTRY_DELETED;
return rc;
}
- /* set cache write lock */
- ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
-
- /* Lock the parent's kids tree */
- bdb_cache_entryinfo_lock( ei->bei_parent );
-
Debug( LDAP_DEBUG_TRACE, "====> bdb_cache_delete( %ld )\n",
e->e_id, 0, 0 );
/* set lru mutex */
- ldap_pvt_thread_mutex_lock( &cache->lru_mutex );
- rc = bdb_cache_delete_internal( cache, e->e_private );
- /* free lru mutex */
- ldap_pvt_thread_mutex_unlock( &cache->lru_mutex );
+ ldap_pvt_thread_mutex_lock( &bdb->bi_cache.lru_head_mutex );
- /* free cache write lock */
- ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
- bdb_cache_entryinfo_unlock( ei->bei_parent );
+ 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 );
/* Leave entry info locked */
void
bdb_cache_delete_cleanup(
Cache *cache,
- Entry *e )
+ EntryInfo *ei )
{
- EntryInfo *ei = BEI(e);
-
- ei->bei_e = NULL;
- e->e_private = NULL;
- bdb_entry_return( e );
+ if ( ei->bei_e ) {
+ ei->bei_e->e_private = NULL;
+#ifdef SLAP_ZONE_ALLOC
+ bdb_entry_return( ei->bei_bdb, ei->bei_e, ei->bei_zseq );
+#else
+ bdb_entry_return( ei->bei_e );
+#endif
+ ei->bei_e = NULL;
+ }
free( ei->bei_nrdn.bv_val );
ei->bei_nrdn.bv_val = NULL;
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 );
}
static int
bdb_cache_delete_internal(
Cache *cache,
- EntryInfo *e )
+ EntryInfo *e,
+ int decr )
{
int rc = 0; /* return code */
+ int decr_leaf = 0;
+
+ /* Lock the parent's kids tree */
+ bdb_cache_entryinfo_lock( e->bei_parent );
#ifdef BDB_HIER
e->bei_parent->bei_ckids--;
- if ( e->bei_parent->bei_dkids ) e->bei_parent->bei_dkids--;
+ if ( decr && e->bei_parent->bei_dkids ) e->bei_parent->bei_dkids--;
#endif
/* dn tree */
if ( avl_delete( &e->bei_parent->bei_kids, (caddr_t) e, bdb_rdn_cmp )
{
rc = -1;
}
+ 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 ) == NULL ) {
+ if ( avl_delete( &cache->c_idtree, (caddr_t) e, bdb_id_cmp )) {
+ cache->c_eiused--;
+ if ( decr_leaf )
+ cache->c_leaves--;
+ } else {
rc = -1;
}
+ ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
- if (rc != 0) {
- return rc;
- }
+ if ( rc == 0 ){
+ /* lru */
+ if ( e == cache->c_lruhead ) cache->c_lruhead = e->bei_lrunext;
+ if ( e == cache->c_lrutail ) {
+ ldap_pvt_thread_mutex_lock( &cache->lru_tail_mutex );
+ if ( e == cache->c_lrutail ) cache->c_lrutail = e->bei_lruprev;
+ ldap_pvt_thread_mutex_unlock( &cache->lru_tail_mutex );
+ }
- /* lru */
- LRU_DELETE( cache, e );
- cache->c_cursize--;
+ if ( e->bei_lrunext ) e->bei_lrunext->bei_lruprev = e->bei_lruprev;
+ if ( e->bei_lruprev ) e->bei_lruprev->bei_lrunext = e->bei_lrunext;
- /*
- * flag entry to be freed later by a call to cache_return_entry()
- */
- e->bei_state |= CACHE_ENTRY_DELETED;
+ if ( e->bei_e ) {
+ ldap_pvt_thread_mutex_lock( &cache->c_count_mutex );
+ cache->c_cursize--;
+ ldap_pvt_thread_mutex_unlock( &cache->c_count_mutex );
+ }
+ }
- return( 0 );
+ return( rc );
}
static void
}
if ( ei->bei_e ) {
ei->bei_e->e_private = NULL;
+#ifdef SLAP_ZONE_ALLOC
+ bdb_entry_return( ei->bei_bdb, ei->bei_e, ei->bei_zseq );
+#else
bdb_entry_return( ei->bei_e );
+#endif
}
bdb_cache_entryinfo_destroy( ei );
}
/* set cache write lock */
ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
/* set lru mutex */
- ldap_pvt_thread_mutex_lock( &cache->lru_mutex );
+ ldap_pvt_thread_mutex_lock( &cache->lru_tail_mutex );
Debug( LDAP_DEBUG_TRACE, "====> bdb_cache_release_all\n", 0, 0, 0 );
cache->c_lruhead = cache->c_eifree->bei_lrunext;
bdb_cache_entryinfo_destroy(cache->c_eifree);
}
+ cache->c_cursize = 0;
+ cache->c_eiused = 0;
+ cache->c_leaves = 0;
+ cache->c_idtree = NULL;
cache->c_lruhead = NULL;
cache->c_lrutail = NULL;
+ cache->c_dntree.bei_kids = NULL;
/* free lru mutex */
- ldap_pvt_thread_mutex_unlock( &cache->lru_mutex );
+ ldap_pvt_thread_mutex_unlock( &cache->lru_tail_mutex );
/* free cache write lock */
ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
}
#ifdef LDAP_DEBUG
+#ifdef SLAPD_UNUSED
static void
bdb_lru_print( Cache *cache )
{
EntryInfo *e;
- fprintf( stderr, "LRU queue (head to tail):\n" );
- for ( e = cache->c_lruhead; e != NULL; e = e->bei_lrunext ) {
- fprintf( stderr, "\trdn \"%20s\" id %ld\n",
- e->bei_nrdn.bv_val, e->bei_id );
+ fprintf( stderr, "LRU circle head: %p\n", cache->c_lruhead );
+ fprintf( stderr, "LRU circle (tail forward):\n" );
+ for ( e = cache->c_lrutail; ; ) {
+ fprintf( stderr, "\t%p, %p id %ld rdn \"%s\"\n",
+ e, e->bei_e, e->bei_id, e->bei_nrdn.bv_val );
+ e = e->bei_lrunext;
+ if ( e == cache->c_lrutail )
+ break;
}
- fprintf( stderr, "LRU queue (tail to head):\n" );
- for ( e = cache->c_lrutail; e != NULL; e = e->bei_lruprev ) {
- fprintf( stderr, "\trdn \"%20s\" id %ld\n",
- e->bei_nrdn.bv_val, e->bei_id );
+ fprintf( stderr, "LRU circle (tail backward):\n" );
+ for ( e = cache->c_lrutail; ; ) {
+ fprintf( stderr, "\t%p, %p id %ld rdn \"%s\"\n",
+ e, e->bei_e, e->bei_id, e->bei_nrdn.bv_val );
+ e = e->bei_lruprev;
+ if ( e == cache->c_lrutail )
+ break;
}
}
#endif
-
-static void
-bdb_txn_free( void *key, void *data )
-{
- DB_TXN *txn = data;
- TXN_ABORT( txn );
-}
-
-/* Obtain a long-lived transaction for the current thread */
-static int
-bdb_txn_get( Operation *op, DB_ENV *env, DB_TXN **txn )
-{
- int i, rc, lockid;
- void *ctx, *data;
-
- /* If no op was provided, try to find the ctx anyway... */
- if ( op ) {
- ctx = op->o_threadctx;
- } else {
- ctx = ldap_pvt_thread_pool_context();
- }
-
- /* Shouldn't happen unless we're single-threaded */
- if ( !ctx ) {
- *txn = NULL;
- return 0;
- }
-
- if ( ldap_pvt_thread_pool_getkey( ctx, ((char *)env)+1, &data, NULL ) ) {
- for ( i=0, rc=1; rc != 0 && i<4; i++ ) {
- rc = TXN_BEGIN( env, NULL, txn, 0 );
- if (rc) ldap_pvt_thread_yield();
- }
- if ( rc != 0) {
- return rc;
- }
- if ( ( rc = ldap_pvt_thread_pool_setkey( ctx, ((char *)env)+1,
- *txn, bdb_txn_free ) ) ) {
- TXN_ABORT( *txn );
- Debug( LDAP_DEBUG_ANY, "bdb_txn_get: err %s(%d)\n",
- db_strerror(rc), rc, 0 );
-
- return rc;
- }
- } else {
- *txn = data;
- }
- return 0;
-}
+#endif
#ifdef BDB_REUSE_LOCKERS
static void
bdb_locker_id_free( void *key, void *data )
{
DB_ENV *env = key;
- int lockid = (int) data;
+ u_int32_t lockid = (long)data;
int rc;
rc = XLOCK_ID_FREE( env, lockid );
if ( rc == EINVAL ) {
DB_LOCKREQ lr;
Debug( LDAP_DEBUG_ANY,
- "bdb_locker_id_free: %d err %s(%d)\n",
- lockid, db_strerror(rc), rc );
+ "bdb_locker_id_free: %lu err %s(%d)\n",
+ (unsigned long) lockid, db_strerror(rc), rc );
/* release all locks held by this locker. */
lr.op = DB_LOCK_PUT_ALL;
lr.obj = NULL;
}
int
-bdb_locker_id( Operation *op, DB_ENV *env, int *locker )
+bdb_locker_id( Operation *op, DB_ENV *env, u_int32_t *locker )
{
- int i, rc, lockid;
+ int i, rc;
+ u_int32_t lockid;
void *data;
void *ctx;
if ( rc != 0) {
return rc;
}
- data = (void *)lockid;
+ data = (void *)((long)lockid);
if ( ( rc = ldap_pvt_thread_pool_setkey( ctx, env,
data, bdb_locker_id_free ) ) ) {
XLOCK_ID_FREE( env, lockid );
return rc;
}
} else {
- lockid = (int)data;
+ lockid = (long)data;
}
*locker = lockid;
return 0;
}
-#endif
-
-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;
- bdb_entry_return( ei->bei_e );
- 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 );
-}
+#endif /* BDB_REUSE_LOCKERS */
projects / openldap / blobdiff