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 /* Already linked, ignore */
123 if ( ei->bei_lruprev )
126 /* Insert into circular LRU list */
127 ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_lru_mutex );
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_return_entry_rw( struct bdb_info *bdb, Entry *e,
255 int rw, DB_LOCK *lock )
260 bdb_cache_entry_db_unlock( bdb, lock );
262 bdb_cache_entryinfo_lock( ei );
263 if ( ei->bei_state & CACHE_ENTRY_NOT_CACHED ) {
265 ei->bei_state ^= CACHE_ENTRY_NOT_CACHED;
268 bdb_cache_entryinfo_unlock( ei );
271 bdb_entry_return( e );
276 bdb_cache_entryinfo_destroy( EntryInfo *e )
278 ldap_pvt_thread_mutex_destroy( &e->bei_kids_mutex );
279 free( e->bei_nrdn.bv_val );
281 free( e->bei_rdn.bv_val );
287 /* Do a length-ordered sort on normalized RDNs */
289 bdb_rdn_cmp( const void *v_e1, const void *v_e2 )
291 const EntryInfo *e1 = v_e1, *e2 = v_e2;
292 int rc = e1->bei_nrdn.bv_len - e2->bei_nrdn.bv_len;
294 rc = strncmp( e1->bei_nrdn.bv_val, e2->bei_nrdn.bv_val,
295 e1->bei_nrdn.bv_len );
301 bdb_id_cmp( const void *v_e1, const void *v_e2 )
303 const EntryInfo *e1 = v_e1, *e2 = v_e2;
304 return e1->bei_id - e2->bei_id;
308 bdb_id_dup_err( void *v1, void *v2 )
311 e2->bei_lrunext = v1;
315 /* Create an entryinfo in the cache. Caller must release the locks later.
318 bdb_entryinfo_add_internal(
319 struct bdb_info *bdb,
323 EntryInfo *ei2 = NULL;
327 ei2 = bdb_cache_entryinfo_new( &bdb->bi_cache );
329 ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
330 bdb_cache_entryinfo_lock( ei->bei_parent );
332 ei2->bei_id = ei->bei_id;
333 ei2->bei_parent = ei->bei_parent;
335 ei2->bei_rdn = ei->bei_rdn;
337 #ifdef SLAP_ZONE_ALLOC
341 /* Add to cache ID tree */
342 if (avl_insert( &bdb->bi_cache.c_idtree, ei2, bdb_id_cmp,
344 EntryInfo *eix = ei2->bei_lrunext;
345 bdb_cache_entryinfo_free( &bdb->bi_cache, ei2 );
348 /* It got freed above because its value was
351 ei->bei_rdn.bv_val = NULL;
354 bdb->bi_cache.c_eiused++;
355 ber_dupbv( &ei2->bei_nrdn, &ei->bei_nrdn );
357 /* This is a new leaf node. But if parent had no kids, then it was
358 * a leaf and we would be decrementing that. So, only increment if
359 * the parent already has kids.
361 if ( ei->bei_parent->bei_kids || !ei->bei_parent->bei_id )
362 bdb->bi_cache.c_leaves++;
363 avl_insert( &ei->bei_parent->bei_kids, ei2, bdb_rdn_cmp,
366 ei->bei_parent->bei_ckids++;
374 /* Find the EntryInfo for the requested DN. If the DN cannot be found, return
375 * the info for its closest ancestor. *res should be NULL to process a
376 * complete DN starting from the tree root. Otherwise *res must be the
377 * immediate parent of the requested DN, and only the RDN will be searched.
378 * The EntryInfo is locked upon return and must be unlocked by the caller.
387 struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
388 EntryInfo ei, *eip, *ei2;
392 /* this function is always called with normalized DN */
394 /* we're doing a onelevel search for an RDN */
395 ei.bei_nrdn.bv_val = ndn->bv_val;
396 ei.bei_nrdn.bv_len = dn_rdnlen( op->o_bd, ndn );
399 /* we're searching a full DN from the root */
400 ptr = ndn->bv_val + ndn->bv_len - op->o_bd->be_nsuffix[0].bv_len;
401 ei.bei_nrdn.bv_val = ptr;
402 ei.bei_nrdn.bv_len = op->o_bd->be_nsuffix[0].bv_len;
403 /* Skip to next rdn if suffix is empty */
404 if ( ei.bei_nrdn.bv_len == 0 ) {
405 for (ptr = ei.bei_nrdn.bv_val - 2; ptr > ndn->bv_val
406 && !DN_SEPARATOR(*ptr); ptr--) /* empty */;
407 if ( ptr >= ndn->bv_val ) {
408 if (DN_SEPARATOR(*ptr)) ptr++;
409 ei.bei_nrdn.bv_len = ei.bei_nrdn.bv_val - ptr;
410 ei.bei_nrdn.bv_val = ptr;
413 eip = &bdb->bi_cache.c_dntree;
416 for ( bdb_cache_entryinfo_lock( eip ); eip; ) {
417 eip->bei_state |= CACHE_ENTRY_REFERENCED;
419 ei2 = (EntryInfo *)avl_find( eip->bei_kids, &ei, bdb_rdn_cmp );
421 int len = ei.bei_nrdn.bv_len;
423 if ( BER_BVISEMPTY( ndn )) {
428 ei.bei_nrdn.bv_len = ndn->bv_len -
429 (ei.bei_nrdn.bv_val - ndn->bv_val);
430 bdb_cache_entryinfo_unlock( eip );
432 rc = bdb_dn2id( op, txn, &ei.bei_nrdn, &ei );
434 bdb_cache_entryinfo_lock( eip );
439 /* DN exists but needs to be added to cache */
440 ei.bei_nrdn.bv_len = len;
441 rc = bdb_entryinfo_add_internal( bdb, &ei, &ei2 );
442 /* add_internal left eip and c_rwlock locked */
443 ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
448 } else if ( ei2->bei_state & CACHE_ENTRY_DELETED ) {
449 /* In the midst of deleting? Give it a chance to
452 bdb_cache_entryinfo_unlock( eip );
453 ldap_pvt_thread_yield();
454 bdb_cache_entryinfo_lock( eip );
458 bdb_cache_entryinfo_unlock( eip );
459 bdb_cache_entryinfo_lock( ei2 );
463 /* Advance to next lower RDN */
464 for (ptr = ei.bei_nrdn.bv_val - 2; ptr > ndn->bv_val
465 && !DN_SEPARATOR(*ptr); ptr--) /* empty */;
466 if ( ptr >= ndn->bv_val ) {
467 if (DN_SEPARATOR(*ptr)) ptr++;
468 ei.bei_nrdn.bv_len = ei.bei_nrdn.bv_val - ptr - 1;
469 ei.bei_nrdn.bv_val = ptr;
471 if ( ptr < ndn->bv_val ) {
481 /* Walk up the tree from a child node, looking for an ID that's already
482 * been linked into the cache.
485 hdb_cache_find_parent(
492 struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
493 EntryInfo ei, eip, *ei2 = NULL, *ein = NULL, *eir = NULL;
501 rc = hdb_dn2id_parent( op, txn, locker, &ei, &eip.bei_id );
504 /* Save the previous node, if any */
507 /* Create a new node for the current ID */
508 ein = bdb_cache_entryinfo_new( &bdb->bi_cache );
509 ein->bei_id = ei.bei_id;
510 ein->bei_kids = ei.bei_kids;
511 ein->bei_nrdn = ei.bei_nrdn;
512 ein->bei_rdn = ei.bei_rdn;
513 ein->bei_ckids = ei.bei_ckids;
514 #ifdef SLAP_ZONE_ALLOC
519 /* This node is not fully connected yet */
520 ein->bei_state |= CACHE_ENTRY_NOT_LINKED;
522 /* Insert this node into the ID tree */
523 ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
524 if ( avl_insert( &bdb->bi_cache.c_idtree, (caddr_t)ein,
525 bdb_id_cmp, bdb_id_dup_err ) ) {
526 EntryInfo *eix = ein->bei_lrunext;
528 /* Someone else created this node just before us.
529 * Free our new copy and use the existing one.
531 bdb_cache_entryinfo_free( &bdb->bi_cache, ein );
534 /* Link in any kids we've already processed */
536 bdb_cache_entryinfo_lock( ein );
537 avl_insert( &ein->bei_kids, (caddr_t)ei2,
538 bdb_rdn_cmp, avl_dup_error );
540 bdb_cache_entryinfo_unlock( ein );
544 /* If this is the first time, save this node
545 * to be returned later.
547 if ( eir == NULL ) eir = ein;
549 /* If there was a previous node, link it to this one */
550 if ( ei2 ) ei2->bei_parent = ein;
552 /* Look for this node's parent */
554 ei2 = (EntryInfo *) avl_find( bdb->bi_cache.c_idtree,
555 (caddr_t) &eip, bdb_id_cmp );
557 ei2 = &bdb->bi_cache.c_dntree;
559 bdb->bi_cache.c_eiused++;
560 if ( ei2 && ( ei2->bei_kids || !ei2->bei_id ))
561 bdb->bi_cache.c_leaves++;
562 ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
564 /* Got the parent, link in and we're done. */
566 bdb_cache_entryinfo_lock( eir );
567 bdb_cache_entryinfo_lock( ei2 );
568 ein->bei_parent = ei2;
570 avl_insert( &ei2->bei_kids, (caddr_t)ein, bdb_rdn_cmp,
574 /* Reset all the state info */
575 for (ein = eir; ein != ei2; ein=ein->bei_parent)
576 ein->bei_state &= ~CACHE_ENTRY_NOT_LINKED;
578 bdb_cache_entryinfo_unlock( ei2 );
584 ei.bei_id = eip.bei_id;
586 avl_insert( &ei.bei_kids, (caddr_t)ein, bdb_rdn_cmp,
592 /* Used by hdb_dn2idl when loading the EntryInfo for all the children
596 struct bdb_info *bdb,
603 /* See if we already have this one */
604 bdb_cache_entryinfo_lock( ei->bei_parent );
605 ei2 = (EntryInfo *)avl_find( ei->bei_parent->bei_kids, ei, bdb_rdn_cmp );
606 bdb_cache_entryinfo_unlock( ei->bei_parent );
609 /* Not found, add it */
612 /* bei_rdn was not malloc'd before, do it now */
613 ber_dupbv( &bv, &ei->bei_rdn );
616 rc = bdb_entryinfo_add_internal( bdb, ei, res );
617 bdb_cache_entryinfo_unlock( ei->bei_parent );
618 ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
620 /* Found, return it */
628 /* This is best-effort only. If all entries in the cache are
629 * busy, they will all be kept. This is unlikely to happen
630 * unless the cache is very much smaller than the working set.
633 bdb_cache_lru_purge( struct bdb_info *bdb )
635 DB_LOCK lock, *lockp;
636 EntryInfo *elru, *elnext = NULL;
637 int count, islocked, eimax;
639 /* Wait for the mutex; we're the only one trying to purge. */
640 ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_lru_mutex );
642 if ( bdb->bi_cache.c_cursize <= bdb->bi_cache.c_maxsize ) {
643 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
644 bdb->bi_cache.c_purging = 0;
648 if ( bdb->bi_cache.c_locker ) {
656 /* maximum number of EntryInfo leaves to cache. In slapcat
657 * we always free all leaf nodes.
659 if ( slapMode & SLAP_TOOL_READONLY )
662 eimax = bdb->bi_cache.c_eimax;
664 /* Look for an unused entry to remove */
665 for ( elru = bdb->bi_cache.c_lruhead; elru; elru = elnext ) {
666 elnext = elru->bei_lrunext;
668 if ( bdb_cache_entryinfo_trylock( elru ))
671 /* This flag implements the clock replacement behavior */
672 if ( elru->bei_state & ( CACHE_ENTRY_REFERENCED )) {
673 elru->bei_state &= ~CACHE_ENTRY_REFERENCED;
674 bdb_cache_entryinfo_unlock( elru );
678 /* If this node is in the process of linking into the cache,
679 * or this node is being deleted, skip it.
681 if (( elru->bei_state & ( CACHE_ENTRY_NOT_LINKED |
682 CACHE_ENTRY_DELETED | CACHE_ENTRY_LOADING )) ||
683 elru->bei_finders > 0 ) {
684 bdb_cache_entryinfo_unlock( elru );
688 /* entryinfo is locked */
691 /* If we can successfully writelock it, then
692 * the object is idle.
694 if ( bdb_cache_entry_db_lock( bdb,
695 bdb->bi_cache.c_locker, elru, 1, 1, lockp ) == 0 ) {
697 /* Free entry for this node if it's present */
699 elru->bei_e->e_private = NULL;
700 #ifdef SLAP_ZONE_ALLOC
701 bdb_entry_return( bdb, elru->bei_e, elru->bei_zseq );
703 bdb_entry_return( elru->bei_e );
708 bdb_cache_entry_db_unlock( bdb, lockp );
711 * If it is a leaf node, and we're over the limit, free it.
713 if ( elru->bei_kids ) {
714 /* Drop from list, we ignore it... */
715 LRU_DEL( &bdb->bi_cache, elru );
716 } else if ( bdb->bi_cache.c_leaves > eimax ) {
717 /* Too many leaf nodes, free this one */
718 bdb_cache_delete_internal( &bdb->bi_cache, elru, 0 );
719 bdb_cache_delete_cleanup( &bdb->bi_cache, elru );
721 } /* Leave on list until we need to free it */
725 bdb_cache_entryinfo_unlock( elru );
727 if ( count >= bdb->bi_cache.c_minfree ) {
728 ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
729 bdb->bi_cache.c_cursize -= count;
730 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
734 if ( elnext == bdb->bi_cache.c_lruhead )
738 bdb->bi_cache.c_lruhead = elnext;
739 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
740 bdb->bi_cache.c_purging = 0;
745 struct bdb_info *bdb,
748 EntryInfo ei = { 0 },
753 ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
754 ei2 = (EntryInfo *) avl_find( bdb->bi_cache.c_idtree,
755 (caddr_t) &ei, bdb_id_cmp );
756 ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
761 * cache_find_id - find an entry in the cache, given id.
762 * The entry is locked for Read upon return. Call with flag ID_LOCKED if
763 * the supplied *eip was already locked.
776 struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
778 int rc = 0, load = 0;
779 EntryInfo ei = { 0 };
783 #ifdef SLAP_ZONE_ALLOC
784 slap_zh_rlock(bdb->bi_cache.c_zctx);
786 /* If we weren't given any info, see if we have it already cached */
788 again: ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
789 *eip = (EntryInfo *) avl_find( bdb->bi_cache.c_idtree,
790 (caddr_t) &ei, bdb_id_cmp );
792 /* If the lock attempt fails, the info is in use */
793 if ( bdb_cache_entryinfo_trylock( *eip )) {
794 ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
795 /* If this node is being deleted, treat
796 * as if the delete has already finished
798 if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
801 /* otherwise, wait for the info to free up */
802 ldap_pvt_thread_yield();
805 /* If this info isn't hooked up to its parent yet,
806 * unlock and wait for it to be fully initialized
808 if ( (*eip)->bei_state & CACHE_ENTRY_NOT_LINKED ) {
809 bdb_cache_entryinfo_unlock( *eip );
810 ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
811 ldap_pvt_thread_yield();
816 ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
819 /* See if the ID exists in the database; add it to the cache if so */
822 rc = bdb_id2entry( op->o_bd, tid, locker, id, &ep );
824 rc = bdb_cache_find_ndn( op, tid,
826 if ( *eip ) flag |= ID_LOCKED;
828 ep->e_private = NULL;
829 #ifdef SLAP_ZONE_ALLOC
830 bdb_entry_return( bdb, ep, (*eip)->bei_zseq );
832 bdb_entry_return( ep );
838 rc = hdb_cache_find_parent(op, tid, locker, id, eip );
839 if ( rc == 0 ) flag |= ID_LOCKED;
843 /* Ok, we found the info, do we have the entry? */
845 if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
848 (*eip)->bei_finders++;
849 (*eip)->bei_state |= CACHE_ENTRY_REFERENCED;
850 /* Make sure only one thread tries to load the entry */
852 #ifdef SLAP_ZONE_ALLOC
853 if ((*eip)->bei_e && !slap_zn_validate(
854 bdb->bi_cache.c_zctx, (*eip)->bei_e, (*eip)->bei_zseq)) {
855 (*eip)->bei_e = NULL;
856 (*eip)->bei_zseq = 0;
859 if ( !(*eip)->bei_e && !((*eip)->bei_state & CACHE_ENTRY_LOADING)) {
861 (*eip)->bei_state |= CACHE_ENTRY_LOADING;
864 /* If the entry was loaded before but uncached, and we need
865 * it again, clear the uncached state
867 if ( (*eip)->bei_state & CACHE_ENTRY_NOT_CACHED ) {
868 (*eip)->bei_state ^= CACHE_ENTRY_NOT_CACHED;
869 if ( flag & ID_NOCACHE )
873 if ( flag & ID_LOCKED ) {
874 bdb_cache_entryinfo_unlock( *eip );
877 rc = bdb_cache_entry_db_lock( bdb, locker, *eip, load, 0, lock );
878 if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
880 bdb_cache_entry_db_unlock( bdb, lock );
881 } else if ( rc == 0 ) {
884 rc = bdb_id2entry( op->o_bd, tid, locker, id, &ep );
887 ep->e_private = *eip;
892 #ifdef SLAP_ZONE_ALLOC
893 (*eip)->bei_zseq = *((ber_len_t *)ep - 2);
896 bdb_cache_lru_link( bdb, *eip );
897 if ( flag & ID_NOCACHE ) {
898 bdb_cache_entryinfo_lock( *eip );
899 (*eip)->bei_state |= CACHE_ENTRY_NOT_CACHED;
900 bdb_cache_entryinfo_unlock( *eip );
904 /* If we succeeded, downgrade back to a readlock. */
905 rc = bdb_cache_entry_db_relock( bdb, locker,
908 /* Otherwise, release the lock. */
909 bdb_cache_entry_db_unlock( bdb, lock );
911 } else if ( !(*eip)->bei_e ) {
912 /* Some other thread is trying to load the entry,
913 * wait for it to finish.
915 bdb_cache_entry_db_unlock( bdb, lock );
916 bdb_cache_entryinfo_lock( *eip );
921 /* Check for subtree renames
923 rc = bdb_fix_dn( (*eip)->bei_e, 1 );
925 bdb_cache_entry_db_relock( bdb,
926 locker, *eip, 1, 0, lock );
927 /* check again in case other modifier did it already */
928 if ( bdb_fix_dn( (*eip)->bei_e, 1 ) )
929 rc = bdb_fix_dn( (*eip)->bei_e, 2 );
930 bdb_cache_entry_db_relock( bdb,
931 locker, *eip, 0, 0, lock );
935 bdb_cache_entryinfo_lock( *eip );
936 (*eip)->bei_finders--;
938 (*eip)->bei_state ^= CACHE_ENTRY_LOADING;
939 bdb_cache_entryinfo_unlock( *eip );
943 if ( flag & ID_LOCKED ) {
944 bdb_cache_entryinfo_unlock( *eip );
947 ep->e_private = NULL;
948 #ifdef SLAP_ZONE_ALLOC
949 bdb_entry_return( bdb, ep, (*eip)->bei_zseq );
951 bdb_entry_return( ep );
958 if ( !( flag & ID_NOCACHE )) {
959 ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
960 bdb->bi_cache.c_cursize++;
961 if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize &&
962 !bdb->bi_cache.c_purging ) {
964 bdb->bi_cache.c_purging = 1;
966 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
970 bdb_cache_lru_purge( bdb );
973 #ifdef SLAP_ZONE_ALLOC
974 if (rc == 0 && (*eip)->bei_e) {
975 slap_zn_rlock(bdb->bi_cache.c_zctx, (*eip)->bei_e);
977 slap_zh_runlock(bdb->bi_cache.c_zctx);
990 if ( BEI(e)->bei_kids ) {
993 if ( BEI(e)->bei_state & CACHE_ENTRY_NO_KIDS ) {
996 rc = bdb_dn2id_children( op, txn, e );
997 if ( rc == DB_NOTFOUND ) {
998 BEI(e)->bei_state |= CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
1003 /* Update the cache after a successful database Add. */
1006 struct bdb_info *bdb,
1009 struct berval *nrdn,
1016 struct berval rdn = e->e_name;
1019 ei.bei_id = e->e_id;
1020 ei.bei_parent = eip;
1021 ei.bei_nrdn = *nrdn;
1024 /* Lock this entry so that bdb_add can run to completion.
1025 * It can only fail if BDB has run out of lock resources.
1027 rc = bdb_cache_entry_db_lock( bdb, locker, &ei, 0, 0, lock );
1029 bdb_cache_entryinfo_unlock( eip );
1034 if ( nrdn->bv_len != e->e_nname.bv_len ) {
1035 char *ptr = ber_bvchr( &rdn, ',' );
1036 assert( ptr != NULL );
1037 rdn.bv_len = ptr - rdn.bv_val;
1039 ber_dupbv( &ei.bei_rdn, &rdn );
1040 if ( eip->bei_dkids ) eip->bei_dkids++;
1043 rc = bdb_entryinfo_add_internal( bdb, &ei, &new );
1044 /* bdb_csn_commit can cause this when adding the database root entry */
1046 new->bei_e->e_private = NULL;
1047 #ifdef SLAP_ZONE_ALLOC
1048 bdb_entry_return( bdb, new->bei_e, new->bei_zseq );
1050 bdb_entry_return( new->bei_e );
1055 new->bei_state |= CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
1056 eip->bei_state &= ~CACHE_ENTRY_NO_KIDS;
1057 if (eip->bei_parent) {
1058 eip->bei_parent->bei_state &= ~CACHE_ENTRY_NO_GRANDKIDS;
1060 bdb_cache_entryinfo_unlock( eip );
1062 ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
1063 ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
1064 ++bdb->bi_cache.c_cursize;
1065 if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize &&
1066 !bdb->bi_cache.c_purging ) {
1068 bdb->bi_cache.c_purging = 1;
1070 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
1072 bdb_cache_lru_link( bdb, new );
1075 bdb_cache_lru_purge( bdb );
1082 struct bdb_info *bdb,
1084 Attribute *newAttrs,
1088 EntryInfo *ei = BEI(e);
1090 /* Get write lock on data */
1091 rc = bdb_cache_entry_db_relock( bdb, locker, ei, 1, 0, lock );
1093 /* If we've done repeated mods on a cached entry, then e_attrs
1094 * is no longer contiguous with the entry, and must be freed.
1097 if ( (void *)e->e_attrs != (void *)(e+1) ) {
1098 attrs_free( e->e_attrs );
1100 e->e_attrs = newAttrs;
1106 * Change the rdn in the entryinfo. Also move to a new parent if needed.
1110 struct bdb_info *bdb,
1112 struct berval *nrdn,
1118 EntryInfo *ei = BEI(e), *pei;
1124 /* Get write lock on data */
1125 rc = bdb_cache_entry_db_relock( bdb, locker, ei, 1, 0, lock );
1126 if ( rc ) return rc;
1128 /* If we've done repeated mods on a cached entry, then e_attrs
1129 * is no longer contiguous with the entry, and must be freed.
1131 if ( (void *)e->e_attrs != (void *)(e+1) ) {
1132 attrs_free( e->e_attrs );
1134 e->e_attrs = new->e_attrs;
1135 if( e->e_nname.bv_val < e->e_bv.bv_val ||
1136 e->e_nname.bv_val > e->e_bv.bv_val + e->e_bv.bv_len )
1138 ch_free(e->e_name.bv_val);
1139 ch_free(e->e_nname.bv_val);
1141 e->e_name = new->e_name;
1142 e->e_nname = new->e_nname;
1144 /* Lock the parent's kids AVL tree */
1145 pei = ei->bei_parent;
1146 bdb_cache_entryinfo_lock( pei );
1147 avl_delete( &pei->bei_kids, (caddr_t) ei, bdb_rdn_cmp );
1148 free( ei->bei_nrdn.bv_val );
1149 ber_dupbv( &ei->bei_nrdn, nrdn );
1151 if ( !pei->bei_kids )
1152 pei->bei_state |= CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
1155 free( ei->bei_rdn.bv_val );
1158 if ( nrdn->bv_len != e->e_nname.bv_len ) {
1159 char *ptr = ber_bvchr(&rdn, ',');
1160 assert( ptr != NULL );
1161 rdn.bv_len = ptr - rdn.bv_val;
1163 ber_dupbv( &ei->bei_rdn, &rdn );
1165 if ( pei->bei_dkids ) pei->bei_dkids--;
1169 ein = ei->bei_parent;
1171 ei->bei_parent = ein;
1172 bdb_cache_entryinfo_unlock( pei );
1173 bdb_cache_entryinfo_lock( ein );
1175 /* parent now has kids */
1176 if ( ein->bei_state & CACHE_ENTRY_NO_KIDS )
1177 ein->bei_state ^= CACHE_ENTRY_NO_KIDS;
1180 /* parent might now have grandkids */
1181 if ( ein->bei_state & CACHE_ENTRY_NO_GRANDKIDS &&
1182 !(ei->bei_state & (CACHE_ENTRY_NO_KIDS)))
1183 ein->bei_state ^= CACHE_ENTRY_NO_GRANDKIDS;
1186 /* Record the generation number of this change */
1187 ldap_pvt_thread_mutex_lock( &bdb->bi_modrdns_mutex );
1189 ei->bei_modrdns = bdb->bi_modrdns;
1190 ldap_pvt_thread_mutex_unlock( &bdb->bi_modrdns_mutex );
1193 if ( ein->bei_dkids ) ein->bei_dkids++;
1195 avl_insert( &ein->bei_kids, ei, bdb_rdn_cmp, avl_dup_error );
1196 bdb_cache_entryinfo_unlock( ein );
1200 * cache_delete - delete the entry e from the cache.
1202 * returns: 0 e was deleted ok
1203 * 1 e was not in the cache
1204 * -1 something bad happened
1208 struct bdb_info *bdb,
1213 EntryInfo *ei = BEI(e);
1216 assert( e->e_private != NULL );
1218 /* Set this early, warn off any queriers */
1219 ei->bei_state |= CACHE_ENTRY_DELETED;
1221 /* Lock the entry's info */
1222 bdb_cache_entryinfo_lock( ei );
1224 /* Get write lock on the data */
1225 rc = bdb_cache_entry_db_relock( bdb, locker, ei, 1, 0, lock );
1227 /* couldn't lock, undo and give up */
1228 ei->bei_state ^= CACHE_ENTRY_DELETED;
1229 bdb_cache_entryinfo_unlock( ei );
1233 Debug( LDAP_DEBUG_TRACE, "====> bdb_cache_delete( %ld )\n",
1237 ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_lru_mutex );
1239 rc = bdb_cache_delete_internal( &bdb->bi_cache, e->e_private, 1 );
1241 /* free lru mutex */
1242 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
1244 /* Leave entry info locked */
1250 bdb_cache_delete_cleanup(
1255 ei->bei_e->e_private = NULL;
1256 #ifdef SLAP_ZONE_ALLOC
1257 bdb_entry_return( ei->bei_bdb, ei->bei_e, ei->bei_zseq );
1259 bdb_entry_return( ei->bei_e );
1264 bdb_cache_entryinfo_free( cache, ei );
1265 bdb_cache_entryinfo_unlock( ei );
1269 bdb_cache_delete_internal(
1274 int rc = 0; /* return code */
1277 /* Lock the parent's kids tree */
1278 bdb_cache_entryinfo_lock( e->bei_parent );
1281 e->bei_parent->bei_ckids--;
1282 if ( decr && e->bei_parent->bei_dkids ) e->bei_parent->bei_dkids--;
1285 if ( avl_delete( &e->bei_parent->bei_kids, (caddr_t) e, bdb_rdn_cmp )
1290 if ( e->bei_parent->bei_kids )
1293 bdb_cache_entryinfo_unlock( e->bei_parent );
1295 ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
1297 if ( avl_delete( &cache->c_idtree, (caddr_t) e, bdb_id_cmp )) {
1304 ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
1308 LRU_DEL( cache, e );
1311 ldap_pvt_thread_mutex_lock( &cache->c_count_mutex );
1313 ldap_pvt_thread_mutex_unlock( &cache->c_count_mutex );
1321 bdb_entryinfo_release( void *data )
1323 EntryInfo *ei = (EntryInfo *)data;
1324 if ( ei->bei_kids ) {
1325 avl_free( ei->bei_kids, NULL );
1328 ei->bei_e->e_private = NULL;
1329 #ifdef SLAP_ZONE_ALLOC
1330 bdb_entry_return( ei->bei_bdb, ei->bei_e, ei->bei_zseq );
1332 bdb_entry_return( ei->bei_e );
1335 bdb_cache_entryinfo_destroy( ei );
1339 bdb_cache_release_all( Cache *cache )
1341 /* set cache write lock */
1342 ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
1344 ldap_pvt_thread_mutex_lock( &cache->c_lru_mutex );
1346 Debug( LDAP_DEBUG_TRACE, "====> bdb_cache_release_all\n", 0, 0, 0 );
1348 avl_free( cache->c_dntree.bei_kids, NULL );
1349 avl_free( cache->c_idtree, bdb_entryinfo_release );
1350 for (;cache->c_eifree;cache->c_eifree = cache->c_lruhead) {
1351 cache->c_lruhead = cache->c_eifree->bei_lrunext;
1352 bdb_cache_entryinfo_destroy(cache->c_eifree);
1354 cache->c_cursize = 0;
1355 cache->c_eiused = 0;
1356 cache->c_leaves = 0;
1357 cache->c_idtree = NULL;
1358 cache->c_lruhead = NULL;
1359 cache->c_lrutail = NULL;
1360 cache->c_dntree.bei_kids = NULL;
1362 /* free lru mutex */
1363 ldap_pvt_thread_mutex_unlock( &cache->c_lru_mutex );
1364 /* free cache write lock */
1365 ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
1371 bdb_lru_print( Cache *cache )
1375 fprintf( stderr, "LRU circle head: %p\n", (void *) cache->c_lruhead );
1376 fprintf( stderr, "LRU circle (tail forward):\n" );
1377 for ( e = cache->c_lrutail; ; ) {
1378 fprintf( stderr, "\t%p, %p id %ld rdn \"%s\"\n",
1379 (void *) e, (void *) e->bei_e, e->bei_id, e->bei_nrdn.bv_val );
1381 if ( e == cache->c_lrutail )
1384 fprintf( stderr, "LRU circle (tail backward):\n" );
1385 for ( e = cache->c_lrutail; ; ) {
1386 fprintf( stderr, "\t%p, %p id %ld rdn \"%s\"\n",
1387 (void *) e, (void *) e->bei_e, e->bei_id, e->bei_nrdn.bv_val );
1389 if ( e == cache->c_lrutail )
1396 #ifdef BDB_REUSE_LOCKERS
1398 bdb_locker_id_free( void *key, void *data )
1404 #if DB_VERSION_FULL >= 0x04060012
1405 BDB_LOCKER lptr = data;
1408 lockid = (long)data;
1410 rc = XLOCK_ID_FREE( env, lockid );
1411 if ( rc == EINVAL ) {
1413 Debug( LDAP_DEBUG_ANY,
1414 "bdb_locker_id_free: %lu err %s(%d)\n",
1415 (unsigned long) lockid, db_strerror(rc), rc );
1416 /* release all locks held by this locker. */
1417 lr.op = DB_LOCK_PUT_ALL;
1419 env->lock_vec( env, lockid, 0, &lr, 1, NULL );
1420 XLOCK_ID_FREE( env, lockid );
1424 /* free up any keys used by the main thread */
1426 bdb_locker_flush( DB_ENV *env )
1429 void *ctx = ldap_pvt_thread_pool_context();
1431 if ( !ldap_pvt_thread_pool_getkey( ctx, env, &data, NULL ) ) {
1432 ldap_pvt_thread_pool_setkey( ctx, env, NULL, NULL );
1433 bdb_locker_id_free( env, data );
1438 bdb_locker_id( Operation *op, DB_ENV *env, BDB_LOCKER *locker )
1445 if ( !env || !locker ) return -1;
1447 /* If no op was provided, try to find the ctx anyway... */
1449 ctx = op->o_threadctx;
1451 ctx = ldap_pvt_thread_pool_context();
1454 /* Shouldn't happen unless we're single-threaded */
1460 if ( ldap_pvt_thread_pool_getkey( ctx, env, &data, NULL ) ) {
1461 for ( i=0, rc=1; rc != 0 && i<4; i++ ) {
1462 rc = XLOCK_ID( env, &lockid );
1463 if (rc) ldap_pvt_thread_yield();
1468 #if DB_VERSION_FULL >= 0x04060012
1470 __lock_getlocker( env->lk_handle, lockid, 0, &lptr );
1474 data = (void *)((long)lockid);
1476 if ( ( rc = ldap_pvt_thread_pool_setkey( ctx, env,
1477 data, bdb_locker_id_free ) ) ) {
1478 XLOCK_ID_FREE( env, lockid );
1479 Debug( LDAP_DEBUG_ANY, "bdb_locker_id: err %s(%d)\n",
1480 db_strerror(rc), rc, 0 );
1485 lockid = (long)data;
1487 #if DB_VERSION_FULL >= 0x04060012
1494 #endif /* BDB_REUSE_LOCKERS */