1 /* cache.c - routines to maintain an in-core cache of entries */
3 /* This work is part of OpenLDAP Software <http://www.openldap.org/>.
5 * Copyright 2000-2007 The OpenLDAP Foundation.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted only as authorized by the OpenLDAP
12 * A copy of this license is available in the file LICENSE in the
13 * top-level directory of the distribution or, alternatively, at
14 * <http://www.OpenLDAP.org/license.html>.
22 #include <ac/string.h>
23 #include <ac/socket.h>
32 #define bdb_cache_lru_purge hdb_cache_lru_purge
34 static void bdb_cache_lru_purge( struct bdb_info *bdb );
36 static int bdb_cache_delete_internal(Cache *cache, EntryInfo *e, int decr);
40 static void bdb_lru_print(Cache *cache);
44 /* For concurrency experiments only! */
46 #define ldap_pvt_thread_rdwr_wlock(a) 0
47 #define ldap_pvt_thread_rdwr_wunlock(a) 0
48 #define ldap_pvt_thread_rdwr_rlock(a) 0
49 #define ldap_pvt_thread_rdwr_runlock(a) 0
53 #define ldap_pvt_thread_mutex_trylock(a) 0
57 bdb_cache_entryinfo_new( Cache *cache )
61 if ( cache->c_eifree ) {
62 ldap_pvt_thread_mutex_lock( &cache->c_eifree_mutex );
63 if ( cache->c_eifree ) {
65 cache->c_eifree = ei->bei_lrunext;
68 ldap_pvt_thread_mutex_unlock( &cache->c_eifree_mutex );
71 ei = ch_calloc(1, sizeof(EntryInfo));
72 ldap_pvt_thread_mutex_init( &ei->bei_kids_mutex );
75 ei->bei_state = CACHE_ENTRY_REFERENCED;
81 bdb_cache_entryinfo_free( Cache *cache, EntryInfo *ei )
83 free( ei->bei_nrdn.bv_val );
84 ei->bei_nrdn.bv_val = NULL;
86 free( ei->bei_rdn.bv_val );
87 ei->bei_rdn.bv_val = NULL;
92 ei->bei_parent = NULL;
94 ei->bei_lruprev = NULL;
96 ldap_pvt_thread_mutex_lock( &cache->c_eifree_mutex );
97 ei->bei_lrunext = cache->c_eifree;
99 ldap_pvt_thread_mutex_unlock( &cache->c_eifree_mutex );
102 #define LRU_DEL( c, e ) do { \
103 if ( e == (c)->c_lruhead ) (c)->c_lruhead = e->bei_lruprev; \
104 if ( e == (c)->c_lrutail ) (c)->c_lrutail = e->bei_lruprev; \
105 e->bei_lrunext->bei_lruprev = e->bei_lruprev; \
106 e->bei_lruprev->bei_lrunext = e->bei_lrunext; \
107 e->bei_lruprev = NULL; \
110 /* Note - we now use a Second-Chance / Clock algorithm instead of
111 * Least-Recently-Used. This tremendously improves concurrency
112 * because we no longer need to manipulate the lists every time an
113 * entry is touched. We only need to lock the lists when adding
114 * or deleting an entry. It's now a circular doubly-linked list.
115 * We always append to the tail, but the head traverses the circle
116 * during a purge operation.
119 bdb_cache_lru_link( struct bdb_info *bdb, EntryInfo *ei )
122 /* Insert into circular LRU list */
123 ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_lru_mutex );
125 /* Still linked, remove */
126 if ( ei->bei_lruprev ) {
127 LRU_DEL( &bdb->bi_cache, ei );
129 ei->bei_lruprev = bdb->bi_cache.c_lrutail;
130 if ( bdb->bi_cache.c_lrutail ) {
131 ei->bei_lrunext = bdb->bi_cache.c_lrutail->bei_lrunext;
132 bdb->bi_cache.c_lrutail->bei_lrunext = ei;
133 if ( ei->bei_lrunext )
134 ei->bei_lrunext->bei_lruprev = ei;
136 ei->bei_lrunext = ei->bei_lruprev = ei;
137 bdb->bi_cache.c_lruhead = ei;
139 bdb->bi_cache.c_lrutail = ei;
140 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
147 /* #define NO_DB_LOCK 1 */
148 /* Note: The BerkeleyDB locks are much slower than regular
149 * mutexes or rdwr locks. But the BDB implementation has the
150 * advantage of using a fixed size lock table, instead of
151 * allocating a lock object per entry in the DB. That's a
152 * key benefit for scaling. It also frees us from worrying
153 * about undetectable deadlocks between BDB activity and our
154 * own cache activity. It's still worth exploring faster
155 * alternatives though.
158 #if DB_VERSION_FULL >= 0x04060012
159 #define BDB_LOCKID(locker) locker->id
161 #define BDB_LOCKID(locker) locker
164 /* Atomically release and reacquire a lock */
166 bdb_cache_entry_db_relock(
167 struct bdb_info *bdb,
181 if ( !lock ) return 0;
183 lockobj.data = &ei->bei_id;
184 lockobj.size = sizeof(ei->bei_id) + 1;
186 list[0].op = DB_LOCK_PUT;
187 list[0].lock = *lock;
188 list[1].op = DB_LOCK_GET;
189 list[1].lock = *lock;
190 list[1].mode = rw ? DB_LOCK_WRITE : DB_LOCK_READ;
191 list[1].obj = &lockobj;
192 rc = bdb->bi_dbenv->lock_vec(bdb->bi_dbenv, BDB_LOCKID(locker), tryOnly ? DB_LOCK_NOWAIT : 0,
195 if (rc && !tryOnly) {
196 Debug( LDAP_DEBUG_TRACE,
197 "bdb_cache_entry_db_relock: entry %ld, rw %d, rc %d\n",
198 ei->bei_id, rw, rc );
200 *lock = list[1].lock;
207 bdb_cache_entry_db_lock( struct bdb_info *bdb, BDB_LOCKER locker, EntryInfo *ei,
208 int rw, int tryOnly, DB_LOCK *lock )
217 if ( !lock ) return 0;
220 db_rw = DB_LOCK_WRITE;
222 db_rw = DB_LOCK_READ;
224 lockobj.data = &ei->bei_id;
225 lockobj.size = sizeof(ei->bei_id) + 1;
227 rc = LOCK_GET(bdb->bi_dbenv, BDB_LOCKID(locker), tryOnly ? DB_LOCK_NOWAIT : 0,
228 &lockobj, db_rw, lock);
229 if (rc && !tryOnly) {
230 Debug( LDAP_DEBUG_TRACE,
231 "bdb_cache_entry_db_lock: entry %ld, rw %d, rc %d\n",
232 ei->bei_id, rw, rc );
235 #endif /* NO_DB_LOCK */
239 bdb_cache_entry_db_unlock ( struct bdb_info *bdb, DB_LOCK *lock )
246 if ( !lock || lock->mode == DB_LOCK_NG ) return 0;
248 rc = LOCK_PUT ( bdb->bi_dbenv, lock );
254 bdb_cache_entryinfo_destroy( EntryInfo *e )
256 ldap_pvt_thread_mutex_destroy( &e->bei_kids_mutex );
257 free( e->bei_nrdn.bv_val );
259 free( e->bei_rdn.bv_val );
265 /* Do a length-ordered sort on normalized RDNs */
267 bdb_rdn_cmp( const void *v_e1, const void *v_e2 )
269 const EntryInfo *e1 = v_e1, *e2 = v_e2;
270 int rc = e1->bei_nrdn.bv_len - e2->bei_nrdn.bv_len;
272 rc = strncmp( e1->bei_nrdn.bv_val, e2->bei_nrdn.bv_val,
273 e1->bei_nrdn.bv_len );
279 bdb_id_cmp( const void *v_e1, const void *v_e2 )
281 const EntryInfo *e1 = v_e1, *e2 = v_e2;
282 return e1->bei_id - e2->bei_id;
286 bdb_id_dup_err( void *v1, void *v2 )
289 e2->bei_lrunext = v1;
293 /* Create an entryinfo in the cache. Caller must release the locks later.
296 bdb_entryinfo_add_internal(
297 struct bdb_info *bdb,
301 EntryInfo *ei2 = NULL;
305 ei2 = bdb_cache_entryinfo_new( &bdb->bi_cache );
307 ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
308 bdb_cache_entryinfo_lock( ei->bei_parent );
310 ei2->bei_id = ei->bei_id;
311 ei2->bei_parent = ei->bei_parent;
313 ei2->bei_rdn = ei->bei_rdn;
315 #ifdef SLAP_ZONE_ALLOC
319 /* Add to cache ID tree */
320 if (avl_insert( &bdb->bi_cache.c_idtree, ei2, bdb_id_cmp,
322 EntryInfo *eix = ei2->bei_lrunext;
323 bdb_cache_entryinfo_free( &bdb->bi_cache, ei2 );
326 /* It got freed above because its value was
329 ei->bei_rdn.bv_val = NULL;
332 bdb->bi_cache.c_eiused++;
333 ber_dupbv( &ei2->bei_nrdn, &ei->bei_nrdn );
335 /* This is a new leaf node. But if parent had no kids, then it was
336 * a leaf and we would be decrementing that. So, only increment if
337 * the parent already has kids.
339 if ( ei->bei_parent->bei_kids || !ei->bei_parent->bei_id )
340 bdb->bi_cache.c_leaves++;
341 avl_insert( &ei->bei_parent->bei_kids, ei2, bdb_rdn_cmp,
344 ei->bei_parent->bei_ckids++;
352 /* Find the EntryInfo for the requested DN. If the DN cannot be found, return
353 * the info for its closest ancestor. *res should be NULL to process a
354 * complete DN starting from the tree root. Otherwise *res must be the
355 * immediate parent of the requested DN, and only the RDN will be searched.
356 * The EntryInfo is locked upon return and must be unlocked by the caller.
365 struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
366 EntryInfo ei, *eip, *ei2;
370 /* this function is always called with normalized DN */
372 /* we're doing a onelevel search for an RDN */
373 ei.bei_nrdn.bv_val = ndn->bv_val;
374 ei.bei_nrdn.bv_len = dn_rdnlen( op->o_bd, ndn );
377 /* we're searching a full DN from the root */
378 ptr = ndn->bv_val + ndn->bv_len - op->o_bd->be_nsuffix[0].bv_len;
379 ei.bei_nrdn.bv_val = ptr;
380 ei.bei_nrdn.bv_len = op->o_bd->be_nsuffix[0].bv_len;
381 /* Skip to next rdn if suffix is empty */
382 if ( ei.bei_nrdn.bv_len == 0 ) {
383 for (ptr = ei.bei_nrdn.bv_val - 2; ptr > ndn->bv_val
384 && !DN_SEPARATOR(*ptr); ptr--) /* empty */;
385 if ( ptr >= ndn->bv_val ) {
386 if (DN_SEPARATOR(*ptr)) ptr++;
387 ei.bei_nrdn.bv_len = ei.bei_nrdn.bv_val - ptr;
388 ei.bei_nrdn.bv_val = ptr;
391 eip = &bdb->bi_cache.c_dntree;
394 for ( bdb_cache_entryinfo_lock( eip ); eip; ) {
395 eip->bei_state |= CACHE_ENTRY_REFERENCED;
397 ei2 = (EntryInfo *)avl_find( eip->bei_kids, &ei, bdb_rdn_cmp );
399 int len = ei.bei_nrdn.bv_len;
401 if ( BER_BVISEMPTY( ndn )) {
406 ei.bei_nrdn.bv_len = ndn->bv_len -
407 (ei.bei_nrdn.bv_val - ndn->bv_val);
408 bdb_cache_entryinfo_unlock( eip );
410 rc = bdb_dn2id( op, txn, &ei.bei_nrdn, &ei );
412 bdb_cache_entryinfo_lock( eip );
417 /* DN exists but needs to be added to cache */
418 ei.bei_nrdn.bv_len = len;
419 rc = bdb_entryinfo_add_internal( bdb, &ei, &ei2 );
420 /* add_internal left eip and c_rwlock locked */
421 ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
426 } else if ( ei2->bei_state & CACHE_ENTRY_DELETED ) {
427 /* In the midst of deleting? Give it a chance to
430 bdb_cache_entryinfo_unlock( eip );
431 ldap_pvt_thread_yield();
432 bdb_cache_entryinfo_lock( eip );
436 bdb_cache_entryinfo_unlock( eip );
437 bdb_cache_entryinfo_lock( ei2 );
441 /* Advance to next lower RDN */
442 for (ptr = ei.bei_nrdn.bv_val - 2; ptr > ndn->bv_val
443 && !DN_SEPARATOR(*ptr); ptr--) /* empty */;
444 if ( ptr >= ndn->bv_val ) {
445 if (DN_SEPARATOR(*ptr)) ptr++;
446 ei.bei_nrdn.bv_len = ei.bei_nrdn.bv_val - ptr - 1;
447 ei.bei_nrdn.bv_val = ptr;
449 if ( ptr < ndn->bv_val ) {
459 /* Walk up the tree from a child node, looking for an ID that's already
460 * been linked into the cache.
463 hdb_cache_find_parent(
470 struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
471 EntryInfo ei, eip, *ei2 = NULL, *ein = NULL, *eir = NULL;
479 rc = hdb_dn2id_parent( op, txn, locker, &ei, &eip.bei_id );
482 /* Save the previous node, if any */
485 /* Create a new node for the current ID */
486 ein = bdb_cache_entryinfo_new( &bdb->bi_cache );
487 ein->bei_id = ei.bei_id;
488 ein->bei_kids = ei.bei_kids;
489 ein->bei_nrdn = ei.bei_nrdn;
490 ein->bei_rdn = ei.bei_rdn;
491 ein->bei_ckids = ei.bei_ckids;
492 #ifdef SLAP_ZONE_ALLOC
497 /* This node is not fully connected yet */
498 ein->bei_state |= CACHE_ENTRY_NOT_LINKED;
500 /* Insert this node into the ID tree */
501 ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
502 if ( avl_insert( &bdb->bi_cache.c_idtree, (caddr_t)ein,
503 bdb_id_cmp, bdb_id_dup_err ) ) {
504 EntryInfo *eix = ein->bei_lrunext;
506 /* Someone else created this node just before us.
507 * Free our new copy and use the existing one.
509 bdb_cache_entryinfo_free( &bdb->bi_cache, ein );
512 /* Link in any kids we've already processed */
514 bdb_cache_entryinfo_lock( ein );
515 avl_insert( &ein->bei_kids, (caddr_t)ei2,
516 bdb_rdn_cmp, avl_dup_error );
518 bdb_cache_entryinfo_unlock( ein );
522 /* If this is the first time, save this node
523 * to be returned later.
525 if ( eir == NULL ) eir = ein;
527 /* If there was a previous node, link it to this one */
528 if ( ei2 ) ei2->bei_parent = ein;
530 /* Look for this node's parent */
532 ei2 = (EntryInfo *) avl_find( bdb->bi_cache.c_idtree,
533 (caddr_t) &eip, bdb_id_cmp );
535 ei2 = &bdb->bi_cache.c_dntree;
537 bdb->bi_cache.c_eiused++;
538 if ( ei2 && ( ei2->bei_kids || !ei2->bei_id ))
539 bdb->bi_cache.c_leaves++;
540 ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
542 /* Got the parent, link in and we're done. */
544 bdb_cache_entryinfo_lock( ei2 );
545 ein->bei_parent = ei2;
547 avl_insert( &ei2->bei_kids, (caddr_t)ein, bdb_rdn_cmp,
551 /* Reset all the state info */
552 for (ein = eir; ein != ei2; ein=ein->bei_parent)
553 ein->bei_state &= ~CACHE_ENTRY_NOT_LINKED;
555 bdb_cache_entryinfo_unlock( ei2 );
556 bdb_cache_entryinfo_lock( eir );
562 ei.bei_id = eip.bei_id;
564 avl_insert( &ei.bei_kids, (caddr_t)ein, bdb_rdn_cmp,
570 /* Used by hdb_dn2idl when loading the EntryInfo for all the children
574 struct bdb_info *bdb,
581 /* See if we already have this one */
582 bdb_cache_entryinfo_lock( ei->bei_parent );
583 ei2 = (EntryInfo *)avl_find( ei->bei_parent->bei_kids, ei, bdb_rdn_cmp );
584 bdb_cache_entryinfo_unlock( ei->bei_parent );
587 /* Not found, add it */
590 /* bei_rdn was not malloc'd before, do it now */
591 ber_dupbv( &bv, &ei->bei_rdn );
594 rc = bdb_entryinfo_add_internal( bdb, ei, res );
595 bdb_cache_entryinfo_unlock( ei->bei_parent );
596 ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
598 /* Found, return it */
606 /* This is best-effort only. If all entries in the cache are
607 * busy, they will all be kept. This is unlikely to happen
608 * unless the cache is very much smaller than the working set.
611 bdb_cache_lru_purge( struct bdb_info *bdb )
613 DB_LOCK lock, *lockp;
614 EntryInfo *elru, *elnext = NULL;
615 int count, islocked, eimax;
617 /* Wait for the mutex; we're the only one trying to purge. */
618 ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_lru_mutex );
620 if ( bdb->bi_cache.c_cursize <= bdb->bi_cache.c_maxsize ) {
621 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
622 bdb->bi_cache.c_purging = 0;
626 if ( bdb->bi_cache.c_locker ) {
634 /* maximum number of EntryInfo leaves to cache. In slapcat
635 * we always free all leaf nodes.
637 if ( slapMode & SLAP_TOOL_READONLY )
640 eimax = bdb->bi_cache.c_eimax;
642 /* Look for an unused entry to remove */
643 for ( elru = bdb->bi_cache.c_lruhead; elru; elru = elnext ) {
644 elnext = elru->bei_lrunext;
646 if ( bdb_cache_entryinfo_trylock( elru ))
649 /* This flag implements the clock replacement behavior */
650 if ( elru->bei_state & ( CACHE_ENTRY_REFERENCED )) {
651 elru->bei_state &= ~CACHE_ENTRY_REFERENCED;
652 bdb_cache_entryinfo_unlock( elru );
656 /* If this node is in the process of linking into the cache,
657 * or this node is being deleted, skip it.
659 if (( elru->bei_state & ( CACHE_ENTRY_NOT_LINKED |
660 CACHE_ENTRY_DELETED | CACHE_ENTRY_LOADING )) ||
661 elru->bei_finders > 0 ) {
662 bdb_cache_entryinfo_unlock( elru );
666 /* entryinfo is locked */
669 /* If we can successfully writelock it, then
670 * the object is idle.
672 if ( bdb_cache_entry_db_lock( bdb,
673 bdb->bi_cache.c_locker, elru, 1, 1, lockp ) == 0 ) {
675 /* Free entry for this node if it's present */
677 elru->bei_e->e_private = NULL;
678 #ifdef SLAP_ZONE_ALLOC
679 bdb_entry_return( bdb, elru->bei_e, elru->bei_zseq );
681 bdb_entry_return( elru->bei_e );
686 bdb_cache_entry_db_unlock( bdb, lockp );
689 * If it is a leaf node, and we're over the limit, free it.
691 if ( elru->bei_kids ) {
692 /* Drop from list, we ignore it... */
693 LRU_DEL( &bdb->bi_cache, elru );
694 } else if ( bdb->bi_cache.c_leaves > eimax ) {
695 /* Too many leaf nodes, free this one */
696 bdb_cache_delete_internal( &bdb->bi_cache, elru, 0 );
697 bdb_cache_delete_cleanup( &bdb->bi_cache, elru );
699 } /* Leave on list until we need to free it */
703 bdb_cache_entryinfo_unlock( elru );
705 if ( count >= bdb->bi_cache.c_minfree ) {
706 ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
707 bdb->bi_cache.c_cursize -= count;
708 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
712 if ( elnext == bdb->bi_cache.c_lruhead )
716 bdb->bi_cache.c_lruhead = elnext;
717 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
718 bdb->bi_cache.c_purging = 0;
723 struct bdb_info *bdb,
726 EntryInfo ei = { 0 },
731 ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
732 ei2 = (EntryInfo *) avl_find( bdb->bi_cache.c_idtree,
733 (caddr_t) &ei, bdb_id_cmp );
734 ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
739 * cache_find_id - find an entry in the cache, given id.
740 * The entry is locked for Read upon return. Call with islocked TRUE if
741 * the supplied *eip was already locked.
754 struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
756 int rc = 0, load = 0;
757 EntryInfo ei = { 0 };
761 #ifdef SLAP_ZONE_ALLOC
762 slap_zh_rlock(bdb->bi_cache.c_zctx);
764 /* If we weren't given any info, see if we have it already cached */
766 again: ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
767 *eip = (EntryInfo *) avl_find( bdb->bi_cache.c_idtree,
768 (caddr_t) &ei, bdb_id_cmp );
770 /* If the lock attempt fails, the info is in use */
771 if ( bdb_cache_entryinfo_trylock( *eip )) {
772 ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
773 /* If this node is being deleted, treat
774 * as if the delete has already finished
776 if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
779 /* otherwise, wait for the info to free up */
780 ldap_pvt_thread_yield();
783 /* If this info isn't hooked up to its parent yet,
784 * unlock and wait for it to be fully initialized
786 if ( (*eip)->bei_state & CACHE_ENTRY_NOT_LINKED ) {
787 bdb_cache_entryinfo_unlock( *eip );
788 ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
789 ldap_pvt_thread_yield();
794 ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
797 /* See if the ID exists in the database; add it to the cache if so */
800 rc = bdb_id2entry( op->o_bd, tid, locker, id, &ep );
802 rc = bdb_cache_find_ndn( op, tid,
804 if ( *eip ) islocked = 1;
806 ep->e_private = NULL;
807 #ifdef SLAP_ZONE_ALLOC
808 bdb_entry_return( bdb, ep, (*eip)->bei_zseq );
810 bdb_entry_return( ep );
816 rc = hdb_cache_find_parent(op, tid, locker, id, eip );
817 if ( rc == 0 ) islocked = 1;
821 /* Ok, we found the info, do we have the entry? */
823 if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
826 (*eip)->bei_finders++;
827 (*eip)->bei_state |= CACHE_ENTRY_REFERENCED;
828 /* Make sure only one thread tries to load the entry */
830 #ifdef SLAP_ZONE_ALLOC
831 if ((*eip)->bei_e && !slap_zn_validate(
832 bdb->bi_cache.c_zctx, (*eip)->bei_e, (*eip)->bei_zseq)) {
833 (*eip)->bei_e = NULL;
834 (*eip)->bei_zseq = 0;
837 if ( !(*eip)->bei_e && !((*eip)->bei_state & CACHE_ENTRY_LOADING)) {
839 (*eip)->bei_state |= CACHE_ENTRY_LOADING;
843 bdb_cache_entryinfo_unlock( *eip );
846 rc = bdb_cache_entry_db_lock( bdb, locker, *eip, load, 0, lock );
847 if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
849 bdb_cache_entry_db_unlock( bdb, lock );
850 } else if ( rc == 0 ) {
853 rc = bdb_id2entry( op->o_bd, tid, locker, id, &ep );
856 ep->e_private = *eip;
861 #ifdef SLAP_ZONE_ALLOC
862 (*eip)->bei_zseq = *((ber_len_t *)ep - 2);
865 bdb_cache_lru_link( bdb, *eip );
868 /* If we succeeded, downgrade back to a readlock. */
869 rc = bdb_cache_entry_db_relock( bdb, locker,
872 /* Otherwise, release the lock. */
873 bdb_cache_entry_db_unlock( bdb, lock );
875 } else if ( !(*eip)->bei_e ) {
876 /* Some other thread is trying to load the entry,
877 * wait for it to finish.
879 bdb_cache_entry_db_unlock( bdb, lock );
880 bdb_cache_entryinfo_lock( *eip );
885 /* Check for subtree renames
887 rc = bdb_fix_dn( (*eip)->bei_e, 1 );
889 bdb_cache_entry_db_relock( bdb,
890 locker, *eip, 1, 0, lock );
891 /* check again in case other modifier did it already */
892 if ( bdb_fix_dn( (*eip)->bei_e, 1 ) )
893 rc = bdb_fix_dn( (*eip)->bei_e, 2 );
894 bdb_cache_entry_db_relock( bdb,
895 locker, *eip, 0, 0, lock );
899 bdb_cache_entryinfo_lock( *eip );
900 (*eip)->bei_finders--;
902 (*eip)->bei_state ^= CACHE_ENTRY_LOADING;
903 bdb_cache_entryinfo_unlock( *eip );
908 bdb_cache_entryinfo_unlock( *eip );
911 ep->e_private = NULL;
912 #ifdef SLAP_ZONE_ALLOC
913 bdb_entry_return( bdb, ep, (*eip)->bei_zseq );
915 bdb_entry_return( ep );
922 ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
923 bdb->bi_cache.c_cursize++;
924 if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize &&
925 !bdb->bi_cache.c_purging ) {
927 bdb->bi_cache.c_purging = 1;
929 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
932 bdb_cache_lru_purge( bdb );
935 #ifdef SLAP_ZONE_ALLOC
936 if (rc == 0 && (*eip)->bei_e) {
937 slap_zn_rlock(bdb->bi_cache.c_zctx, (*eip)->bei_e);
939 slap_zh_runlock(bdb->bi_cache.c_zctx);
952 if ( BEI(e)->bei_kids ) {
955 if ( BEI(e)->bei_state & CACHE_ENTRY_NO_KIDS ) {
958 rc = bdb_dn2id_children( op, txn, e );
959 if ( rc == DB_NOTFOUND ) {
960 BEI(e)->bei_state |= CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
965 /* Update the cache after a successful database Add. */
968 struct bdb_info *bdb,
978 struct berval rdn = e->e_name;
986 /* Lock this entry so that bdb_add can run to completion.
987 * It can only fail if BDB has run out of lock resources.
989 rc = bdb_cache_entry_db_lock( bdb, locker, &ei, 0, 0, lock );
991 bdb_cache_entryinfo_unlock( eip );
996 if ( nrdn->bv_len != e->e_nname.bv_len ) {
997 char *ptr = ber_bvchr( &rdn, ',' );
998 assert( ptr != NULL );
999 rdn.bv_len = ptr - rdn.bv_val;
1001 ber_dupbv( &ei.bei_rdn, &rdn );
1002 if ( eip->bei_dkids ) eip->bei_dkids++;
1005 rc = bdb_entryinfo_add_internal( bdb, &ei, &new );
1006 /* bdb_csn_commit can cause this when adding the database root entry */
1008 new->bei_e->e_private = NULL;
1009 #ifdef SLAP_ZONE_ALLOC
1010 bdb_entry_return( bdb, new->bei_e, new->bei_zseq );
1012 bdb_entry_return( new->bei_e );
1017 new->bei_state |= CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
1018 eip->bei_state &= ~CACHE_ENTRY_NO_KIDS;
1019 if (eip->bei_parent) {
1020 eip->bei_parent->bei_state &= ~CACHE_ENTRY_NO_GRANDKIDS;
1022 bdb_cache_entryinfo_unlock( eip );
1024 ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
1025 ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
1026 ++bdb->bi_cache.c_cursize;
1027 if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize &&
1028 !bdb->bi_cache.c_purging ) {
1030 bdb->bi_cache.c_purging = 1;
1032 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
1034 bdb_cache_lru_link( bdb, new );
1037 bdb_cache_lru_purge( bdb );
1044 struct bdb_info *bdb,
1046 Attribute *newAttrs,
1050 EntryInfo *ei = BEI(e);
1052 /* Get write lock on data */
1053 rc = bdb_cache_entry_db_relock( bdb, locker, ei, 1, 0, lock );
1055 /* If we've done repeated mods on a cached entry, then e_attrs
1056 * is no longer contiguous with the entry, and must be freed.
1059 if ( (void *)e->e_attrs != (void *)(e+1) ) {
1060 attrs_free( e->e_attrs );
1062 e->e_attrs = newAttrs;
1068 * Change the rdn in the entryinfo. Also move to a new parent if needed.
1072 struct bdb_info *bdb,
1074 struct berval *nrdn,
1080 EntryInfo *ei = BEI(e), *pei;
1086 /* Get write lock on data */
1087 rc = bdb_cache_entry_db_relock( bdb, locker, ei, 1, 0, lock );
1088 if ( rc ) return rc;
1090 /* If we've done repeated mods on a cached entry, then e_attrs
1091 * is no longer contiguous with the entry, and must be freed.
1093 if ( (void *)e->e_attrs != (void *)(e+1) ) {
1094 attrs_free( e->e_attrs );
1096 e->e_attrs = new->e_attrs;
1097 if( e->e_nname.bv_val < e->e_bv.bv_val ||
1098 e->e_nname.bv_val > e->e_bv.bv_val + e->e_bv.bv_len )
1100 ch_free(e->e_name.bv_val);
1101 ch_free(e->e_nname.bv_val);
1103 e->e_name = new->e_name;
1104 e->e_nname = new->e_nname;
1106 /* Lock the parent's kids AVL tree */
1107 pei = ei->bei_parent;
1108 bdb_cache_entryinfo_lock( pei );
1109 avl_delete( &pei->bei_kids, (caddr_t) ei, bdb_rdn_cmp );
1110 free( ei->bei_nrdn.bv_val );
1111 ber_dupbv( &ei->bei_nrdn, nrdn );
1113 if ( !pei->bei_kids )
1114 pei->bei_state |= CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
1117 free( ei->bei_rdn.bv_val );
1120 if ( nrdn->bv_len != e->e_nname.bv_len ) {
1121 char *ptr = ber_bvchr(&rdn, ',');
1122 assert( ptr != NULL );
1123 rdn.bv_len = ptr - rdn.bv_val;
1125 ber_dupbv( &ei->bei_rdn, &rdn );
1127 if ( pei->bei_dkids ) pei->bei_dkids--;
1131 ein = ei->bei_parent;
1133 ei->bei_parent = ein;
1134 bdb_cache_entryinfo_unlock( pei );
1135 bdb_cache_entryinfo_lock( ein );
1137 /* parent now has kids */
1138 if ( ein->bei_state & CACHE_ENTRY_NO_KIDS )
1139 ein->bei_state ^= CACHE_ENTRY_NO_KIDS;
1142 /* parent might now have grandkids */
1143 if ( ein->bei_state & CACHE_ENTRY_NO_GRANDKIDS &&
1144 !(ei->bei_state & (CACHE_ENTRY_NO_KIDS)))
1145 ein->bei_state ^= CACHE_ENTRY_NO_GRANDKIDS;
1148 /* Record the generation number of this change */
1149 ldap_pvt_thread_mutex_lock( &bdb->bi_modrdns_mutex );
1151 ei->bei_modrdns = bdb->bi_modrdns;
1152 ldap_pvt_thread_mutex_unlock( &bdb->bi_modrdns_mutex );
1155 if ( ein->bei_dkids ) ein->bei_dkids++;
1157 avl_insert( &ein->bei_kids, ei, bdb_rdn_cmp, avl_dup_error );
1158 bdb_cache_entryinfo_unlock( ein );
1162 * cache_delete - delete the entry e from the cache.
1164 * returns: 0 e was deleted ok
1165 * 1 e was not in the cache
1166 * -1 something bad happened
1170 struct bdb_info *bdb,
1175 EntryInfo *ei = BEI(e);
1178 assert( e->e_private != NULL );
1180 /* Set this early, warn off any queriers */
1181 ei->bei_state |= CACHE_ENTRY_DELETED;
1183 /* Lock the entry's info */
1184 bdb_cache_entryinfo_lock( ei );
1186 /* Get write lock on the data */
1187 rc = bdb_cache_entry_db_relock( bdb, locker, ei, 1, 0, lock );
1189 /* couldn't lock, undo and give up */
1190 ei->bei_state ^= CACHE_ENTRY_DELETED;
1191 bdb_cache_entryinfo_unlock( ei );
1195 Debug( LDAP_DEBUG_TRACE, "====> bdb_cache_delete( %ld )\n",
1199 ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_lru_mutex );
1201 rc = bdb_cache_delete_internal( &bdb->bi_cache, e->e_private, 1 );
1203 /* free lru mutex */
1204 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
1206 /* Leave entry info locked */
1212 bdb_cache_delete_cleanup(
1217 ei->bei_e->e_private = NULL;
1218 #ifdef SLAP_ZONE_ALLOC
1219 bdb_entry_return( ei->bei_bdb, ei->bei_e, ei->bei_zseq );
1221 bdb_entry_return( ei->bei_e );
1226 bdb_cache_entryinfo_free( cache, ei );
1227 bdb_cache_entryinfo_unlock( ei );
1231 bdb_cache_delete_internal(
1236 int rc = 0; /* return code */
1239 /* Lock the parent's kids tree */
1240 bdb_cache_entryinfo_lock( e->bei_parent );
1243 e->bei_parent->bei_ckids--;
1244 if ( decr && e->bei_parent->bei_dkids ) e->bei_parent->bei_dkids--;
1247 if ( avl_delete( &e->bei_parent->bei_kids, (caddr_t) e, bdb_rdn_cmp )
1252 if ( e->bei_parent->bei_kids )
1255 bdb_cache_entryinfo_unlock( e->bei_parent );
1257 ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
1259 if ( avl_delete( &cache->c_idtree, (caddr_t) e, bdb_id_cmp )) {
1266 ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
1270 LRU_DEL( cache, e );
1273 ldap_pvt_thread_mutex_lock( &cache->c_count_mutex );
1275 ldap_pvt_thread_mutex_unlock( &cache->c_count_mutex );
1283 bdb_entryinfo_release( void *data )
1285 EntryInfo *ei = (EntryInfo *)data;
1286 if ( ei->bei_kids ) {
1287 avl_free( ei->bei_kids, NULL );
1290 ei->bei_e->e_private = NULL;
1291 #ifdef SLAP_ZONE_ALLOC
1292 bdb_entry_return( ei->bei_bdb, ei->bei_e, ei->bei_zseq );
1294 bdb_entry_return( ei->bei_e );
1297 bdb_cache_entryinfo_destroy( ei );
1301 bdb_cache_release_all( Cache *cache )
1303 /* set cache write lock */
1304 ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
1306 ldap_pvt_thread_mutex_lock( &cache->c_lru_mutex );
1308 Debug( LDAP_DEBUG_TRACE, "====> bdb_cache_release_all\n", 0, 0, 0 );
1310 avl_free( cache->c_dntree.bei_kids, NULL );
1311 avl_free( cache->c_idtree, bdb_entryinfo_release );
1312 for (;cache->c_eifree;cache->c_eifree = cache->c_lruhead) {
1313 cache->c_lruhead = cache->c_eifree->bei_lrunext;
1314 bdb_cache_entryinfo_destroy(cache->c_eifree);
1316 cache->c_cursize = 0;
1317 cache->c_eiused = 0;
1318 cache->c_leaves = 0;
1319 cache->c_idtree = NULL;
1320 cache->c_lruhead = NULL;
1321 cache->c_lrutail = NULL;
1322 cache->c_dntree.bei_kids = NULL;
1324 /* free lru mutex */
1325 ldap_pvt_thread_mutex_unlock( &cache->c_lru_mutex );
1326 /* free cache write lock */
1327 ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
1333 bdb_lru_print( Cache *cache )
1337 fprintf( stderr, "LRU circle head: %p\n", (void *) cache->c_lruhead );
1338 fprintf( stderr, "LRU circle (tail forward):\n" );
1339 for ( e = cache->c_lrutail; ; ) {
1340 fprintf( stderr, "\t%p, %p id %ld rdn \"%s\"\n",
1341 (void *) e, (void *) e->bei_e, e->bei_id, e->bei_nrdn.bv_val );
1343 if ( e == cache->c_lrutail )
1346 fprintf( stderr, "LRU circle (tail backward):\n" );
1347 for ( e = cache->c_lrutail; ; ) {
1348 fprintf( stderr, "\t%p, %p id %ld rdn \"%s\"\n",
1349 (void *) e, (void *) e->bei_e, e->bei_id, e->bei_nrdn.bv_val );
1351 if ( e == cache->c_lrutail )
1358 #ifdef BDB_REUSE_LOCKERS
1360 bdb_locker_id_free( void *key, void *data )
1366 #if DB_VERSION_FULL >= 0x04060012
1367 BDB_LOCKER lptr = data;
1370 lockid = (long)data;
1372 rc = XLOCK_ID_FREE( env, lockid );
1373 if ( rc == EINVAL ) {
1375 Debug( LDAP_DEBUG_ANY,
1376 "bdb_locker_id_free: %lu err %s(%d)\n",
1377 (unsigned long) lockid, db_strerror(rc), rc );
1378 /* release all locks held by this locker. */
1379 lr.op = DB_LOCK_PUT_ALL;
1381 env->lock_vec( env, lockid, 0, &lr, 1, NULL );
1382 XLOCK_ID_FREE( env, lockid );
1386 /* free up any keys used by the main thread */
1388 bdb_locker_flush( DB_ENV *env )
1391 void *ctx = ldap_pvt_thread_pool_context();
1393 if ( !ldap_pvt_thread_pool_getkey( ctx, env, &data, NULL ) ) {
1394 ldap_pvt_thread_pool_setkey( ctx, env, NULL, NULL );
1395 bdb_locker_id_free( env, data );
1400 bdb_locker_id( Operation *op, DB_ENV *env, BDB_LOCKER *locker )
1407 if ( !env || !locker ) return -1;
1409 /* If no op was provided, try to find the ctx anyway... */
1411 ctx = op->o_threadctx;
1413 ctx = ldap_pvt_thread_pool_context();
1416 /* Shouldn't happen unless we're single-threaded */
1422 if ( ldap_pvt_thread_pool_getkey( ctx, env, &data, NULL ) ) {
1423 for ( i=0, rc=1; rc != 0 && i<4; i++ ) {
1424 rc = XLOCK_ID( env, &lockid );
1425 if (rc) ldap_pvt_thread_yield();
1430 #if DB_VERSION_FULL >= 0x04060012
1432 __lock_getlocker( env->lk_handle, lockid, 0, &lptr );
1436 data = (void *)((long)lockid);
1438 if ( ( rc = ldap_pvt_thread_pool_setkey( ctx, env,
1439 data, bdb_locker_id_free ) ) ) {
1440 XLOCK_ID_FREE( env, lockid );
1441 Debug( LDAP_DEBUG_ANY, "bdb_locker_id: err %s(%d)\n",
1442 db_strerror(rc), rc, 0 );
1447 lockid = (long)data;
1449 #if DB_VERSION_FULL >= 0x04060012
1456 #endif /* BDB_REUSE_LOCKERS */