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 2000-2009 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 /* Atomically release and reacquire a lock */
160 bdb_cache_entry_db_relock(
161 struct bdb_info *bdb,
175 if ( !lock ) return 0;
177 lockobj.data = &ei->bei_id;
178 lockobj.size = sizeof(ei->bei_id) + 1;
180 list[0].op = DB_LOCK_PUT;
181 list[0].lock = *lock;
182 list[1].op = DB_LOCK_GET;
183 list[1].lock = *lock;
184 list[1].mode = rw ? DB_LOCK_WRITE : DB_LOCK_READ;
185 list[1].obj = &lockobj;
186 rc = bdb->bi_dbenv->lock_vec(bdb->bi_dbenv, TXN_ID(txn), tryOnly ? DB_LOCK_NOWAIT : 0,
189 if (rc && !tryOnly) {
190 Debug( LDAP_DEBUG_TRACE,
191 "bdb_cache_entry_db_relock: entry %ld, rw %d, rc %d\n",
192 ei->bei_id, rw, rc );
194 *lock = list[1].lock;
201 bdb_cache_entry_db_lock( struct bdb_info *bdb, DB_TXN *txn, EntryInfo *ei,
202 int rw, int tryOnly, DB_LOCK *lock )
211 if ( !lock ) return 0;
214 db_rw = DB_LOCK_WRITE;
216 db_rw = DB_LOCK_READ;
218 lockobj.data = &ei->bei_id;
219 lockobj.size = sizeof(ei->bei_id) + 1;
221 rc = LOCK_GET(bdb->bi_dbenv, TXN_ID(txn), tryOnly ? DB_LOCK_NOWAIT : 0,
222 &lockobj, db_rw, lock);
223 if (rc && !tryOnly) {
224 Debug( LDAP_DEBUG_TRACE,
225 "bdb_cache_entry_db_lock: entry %ld, rw %d, rc %d\n",
226 ei->bei_id, rw, rc );
229 #endif /* NO_DB_LOCK */
233 bdb_cache_entry_db_unlock ( struct bdb_info *bdb, DB_LOCK *lock )
240 if ( !lock || lock->mode == DB_LOCK_NG ) return 0;
242 rc = LOCK_PUT ( bdb->bi_dbenv, lock );
248 bdb_cache_return_entry_rw( struct bdb_info *bdb, Entry *e,
249 int rw, DB_LOCK *lock )
256 ( ei->bei_state & CACHE_ENTRY_NOT_CACHED ) &&
257 ( bdb_cache_entryinfo_trylock( ei ) == 0 )) {
258 if ( ei->bei_state & CACHE_ENTRY_NOT_CACHED ) {
259 /* Releasing the entry can only be done when
260 * we know that nobody else is using it, i.e we
261 * should have an entry_db writelock. But the
262 * flag is only set by the thread that loads the
263 * entry, and only if no other threads has found
264 * it while it was working. All other threads
265 * clear the flag, which mean that we should be
266 * the only thread using the entry if the flag
270 ei->bei_state ^= CACHE_ENTRY_NOT_CACHED;
273 bdb_cache_entryinfo_unlock( ei );
275 bdb_cache_entry_db_unlock( bdb, lock );
278 bdb_entry_return( e );
283 bdb_cache_entryinfo_destroy( EntryInfo *e )
285 ldap_pvt_thread_mutex_destroy( &e->bei_kids_mutex );
286 free( e->bei_nrdn.bv_val );
288 free( e->bei_rdn.bv_val );
294 /* Do a length-ordered sort on normalized RDNs */
296 bdb_rdn_cmp( const void *v_e1, const void *v_e2 )
298 const EntryInfo *e1 = v_e1, *e2 = v_e2;
299 int rc = e1->bei_nrdn.bv_len - e2->bei_nrdn.bv_len;
301 rc = strncmp( e1->bei_nrdn.bv_val, e2->bei_nrdn.bv_val,
302 e1->bei_nrdn.bv_len );
308 bdb_id_cmp( const void *v_e1, const void *v_e2 )
310 const EntryInfo *e1 = v_e1, *e2 = v_e2;
311 return e1->bei_id - e2->bei_id;
315 bdb_id_dup_err( void *v1, void *v2 )
318 e2->bei_lrunext = v1;
322 /* Create an entryinfo in the cache. Caller must release the locks later.
325 bdb_entryinfo_add_internal(
326 struct bdb_info *bdb,
330 EntryInfo *ei2 = NULL;
334 ei2 = bdb_cache_entryinfo_new( &bdb->bi_cache );
336 ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
337 bdb_cache_entryinfo_lock( ei->bei_parent );
339 ei2->bei_id = ei->bei_id;
340 ei2->bei_parent = ei->bei_parent;
342 ei2->bei_rdn = ei->bei_rdn;
344 #ifdef SLAP_ZONE_ALLOC
348 /* Add to cache ID tree */
349 if (avl_insert( &bdb->bi_cache.c_idtree, ei2, bdb_id_cmp,
351 EntryInfo *eix = ei2->bei_lrunext;
352 bdb_cache_entryinfo_free( &bdb->bi_cache, ei2 );
355 /* It got freed above because its value was
358 ei->bei_rdn.bv_val = NULL;
363 bdb->bi_cache.c_eiused++;
364 ber_dupbv( &ei2->bei_nrdn, &ei->bei_nrdn );
366 /* This is a new leaf node. But if parent had no kids, then it was
367 * a leaf and we would be decrementing that. So, only increment if
368 * the parent already has kids.
370 if ( ei->bei_parent->bei_kids || !ei->bei_parent->bei_id )
371 bdb->bi_cache.c_leaves++;
372 rc = avl_insert( &ei->bei_parent->bei_kids, ei2, bdb_rdn_cmp,
375 /* This should never happen; entry cache is corrupt */
376 bdb->bi_dbenv->log_flush( bdb->bi_dbenv, NULL );
380 ei->bei_parent->bei_ckids++;
388 /* Find the EntryInfo for the requested DN. If the DN cannot be found, return
389 * the info for its closest ancestor. *res should be NULL to process a
390 * complete DN starting from the tree root. Otherwise *res must be the
391 * immediate parent of the requested DN, and only the RDN will be searched.
392 * The EntryInfo is locked upon return and must be unlocked by the caller.
401 struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
402 EntryInfo ei, *eip, *ei2;
406 /* this function is always called with normalized DN */
408 /* we're doing a onelevel search for an RDN */
409 ei.bei_nrdn.bv_val = ndn->bv_val;
410 ei.bei_nrdn.bv_len = dn_rdnlen( op->o_bd, ndn );
413 /* we're searching a full DN from the root */
414 ptr = ndn->bv_val + ndn->bv_len - op->o_bd->be_nsuffix[0].bv_len;
415 ei.bei_nrdn.bv_val = ptr;
416 ei.bei_nrdn.bv_len = op->o_bd->be_nsuffix[0].bv_len;
417 /* Skip to next rdn if suffix is empty */
418 if ( ei.bei_nrdn.bv_len == 0 ) {
419 for (ptr = ei.bei_nrdn.bv_val - 2; ptr > ndn->bv_val
420 && !DN_SEPARATOR(*ptr); ptr--) /* empty */;
421 if ( ptr >= ndn->bv_val ) {
422 if (DN_SEPARATOR(*ptr)) ptr++;
423 ei.bei_nrdn.bv_len = ei.bei_nrdn.bv_val - ptr;
424 ei.bei_nrdn.bv_val = ptr;
427 eip = &bdb->bi_cache.c_dntree;
430 for ( bdb_cache_entryinfo_lock( eip ); eip; ) {
431 eip->bei_state |= CACHE_ENTRY_REFERENCED;
433 ei2 = (EntryInfo *)avl_find( eip->bei_kids, &ei, bdb_rdn_cmp );
436 int len = ei.bei_nrdn.bv_len;
438 if ( BER_BVISEMPTY( ndn )) {
443 ei.bei_nrdn.bv_len = ndn->bv_len -
444 (ei.bei_nrdn.bv_val - ndn->bv_val);
445 bdb_cache_entryinfo_unlock( eip );
447 BDB_LOG_PRINTF( bdb->bi_dbenv, NULL, "slapd Reading %s",
448 ei.bei_nrdn.bv_val );
450 lock.mode = DB_LOCK_NG;
451 rc = bdb_dn2id( op, &ei.bei_nrdn, &ei, txn, &lock );
453 bdb_cache_entryinfo_lock( eip );
454 bdb_cache_entry_db_unlock( bdb, &lock );
459 BDB_LOG_PRINTF( bdb->bi_dbenv, NULL, "slapd Read got %s(%d)",
460 ei.bei_nrdn.bv_val, ei.bei_id );
462 /* DN exists but needs to be added to cache */
463 ei.bei_nrdn.bv_len = len;
464 rc = bdb_entryinfo_add_internal( bdb, &ei, &ei2 );
465 /* add_internal left eip and c_rwlock locked */
466 ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
467 bdb_cache_entry_db_unlock( bdb, &lock );
472 } else if ( ei2->bei_state & CACHE_ENTRY_DELETED ) {
473 /* In the midst of deleting? Give it a chance to
476 bdb_cache_entryinfo_unlock( eip );
477 ldap_pvt_thread_yield();
478 bdb_cache_entryinfo_lock( eip );
482 bdb_cache_entryinfo_unlock( eip );
483 bdb_cache_entryinfo_lock( ei2 );
487 /* Advance to next lower RDN */
488 for (ptr = ei.bei_nrdn.bv_val - 2; ptr > ndn->bv_val
489 && !DN_SEPARATOR(*ptr); ptr--) /* empty */;
490 if ( ptr >= ndn->bv_val ) {
491 if (DN_SEPARATOR(*ptr)) ptr++;
492 ei.bei_nrdn.bv_len = ei.bei_nrdn.bv_val - ptr - 1;
493 ei.bei_nrdn.bv_val = ptr;
495 if ( ptr < ndn->bv_val ) {
505 /* Walk up the tree from a child node, looking for an ID that's already
506 * been linked into the cache.
509 hdb_cache_find_parent(
515 struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
516 EntryInfo ei, eip, *ei2 = NULL, *ein = NULL, *eir = NULL;
524 rc = hdb_dn2id_parent( op, txn, &ei, &eip.bei_id );
527 /* Save the previous node, if any */
530 /* Create a new node for the current ID */
531 ein = bdb_cache_entryinfo_new( &bdb->bi_cache );
532 ein->bei_id = ei.bei_id;
533 ein->bei_kids = ei.bei_kids;
534 ein->bei_nrdn = ei.bei_nrdn;
535 ein->bei_rdn = ei.bei_rdn;
536 ein->bei_ckids = ei.bei_ckids;
537 #ifdef SLAP_ZONE_ALLOC
542 /* This node is not fully connected yet */
543 ein->bei_state |= CACHE_ENTRY_NOT_LINKED;
545 /* Insert this node into the ID tree */
546 ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
547 if ( avl_insert( &bdb->bi_cache.c_idtree, (caddr_t)ein,
548 bdb_id_cmp, bdb_id_dup_err ) ) {
549 EntryInfo *eix = ein->bei_lrunext;
551 /* Someone else created this node just before us.
552 * Free our new copy and use the existing one.
554 bdb_cache_entryinfo_free( &bdb->bi_cache, ein );
557 /* Link in any kids we've already processed */
559 bdb_cache_entryinfo_lock( ein );
560 avl_insert( &ein->bei_kids, (caddr_t)ei2,
561 bdb_rdn_cmp, avl_dup_error );
563 bdb_cache_entryinfo_unlock( ein );
567 /* If this is the first time, save this node
568 * to be returned later.
570 if ( eir == NULL ) eir = ein;
572 /* If there was a previous node, link it to this one */
573 if ( ei2 ) ei2->bei_parent = ein;
575 /* Look for this node's parent */
577 ei2 = (EntryInfo *) avl_find( bdb->bi_cache.c_idtree,
578 (caddr_t) &eip, bdb_id_cmp );
580 ei2 = &bdb->bi_cache.c_dntree;
582 bdb->bi_cache.c_eiused++;
583 if ( ei2 && ( ei2->bei_kids || !ei2->bei_id ))
584 bdb->bi_cache.c_leaves++;
585 ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
587 /* Got the parent, link in and we're done. */
589 bdb_cache_entryinfo_lock( eir );
590 bdb_cache_entryinfo_lock( ei2 );
591 ein->bei_parent = ei2;
593 avl_insert( &ei2->bei_kids, (caddr_t)ein, bdb_rdn_cmp,
597 /* Reset all the state info */
598 for (ein = eir; ein != ei2; ein=ein->bei_parent)
599 ein->bei_state &= ~CACHE_ENTRY_NOT_LINKED;
601 bdb_cache_entryinfo_unlock( ei2 );
607 ei.bei_id = eip.bei_id;
609 avl_insert( &ei.bei_kids, (caddr_t)ein, bdb_rdn_cmp,
615 /* Used by hdb_dn2idl when loading the EntryInfo for all the children
619 struct bdb_info *bdb,
626 /* See if we already have this one */
627 bdb_cache_entryinfo_lock( ei->bei_parent );
628 ei2 = (EntryInfo *)avl_find( ei->bei_parent->bei_kids, ei, bdb_rdn_cmp );
629 bdb_cache_entryinfo_unlock( ei->bei_parent );
632 /* Not found, add it */
635 /* bei_rdn was not malloc'd before, do it now */
636 ber_dupbv( &bv, &ei->bei_rdn );
639 rc = bdb_entryinfo_add_internal( bdb, ei, res );
640 bdb_cache_entryinfo_unlock( ei->bei_parent );
641 ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
643 /* Found, return it */
651 /* This is best-effort only. If all entries in the cache are
652 * busy, they will all be kept. This is unlikely to happen
653 * unless the cache is very much smaller than the working set.
656 bdb_cache_lru_purge( struct bdb_info *bdb )
658 DB_LOCK lock, *lockp;
659 EntryInfo *elru, *elnext = NULL;
660 int count, islocked, eimax;
661 int efree = 0, eifree = 0, eicount, ecount;
666 /* Wait for the mutex; we're the only one trying to purge. */
667 ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_lru_mutex );
669 /* maximum number of EntryInfo leaves to cache. In slapcat
670 * we always free all leaf nodes.
672 if ( slapMode & SLAP_TOOL_READONLY )
675 eimax = bdb->bi_cache.c_eimax;
677 if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize )
678 efree = bdb->bi_cache.c_minfree;
679 if ( bdb->bi_cache.c_leaves > eimax ) {
680 eifree = bdb->bi_cache.c_minfree * 10;
681 if ( eifree >= eimax )
685 if ( !efree && !eifree ) {
686 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
687 bdb->bi_cache.c_purging = 0;
691 if ( bdb->bi_cache.c_txn ) {
704 /* Look for an unused entry to remove */
705 for ( elru = bdb->bi_cache.c_lruhead; elru; elru = elnext ) {
706 elnext = elru->bei_lrunext;
708 if ( bdb_cache_entryinfo_trylock( elru ))
711 /* This flag implements the clock replacement behavior */
712 if ( elru->bei_state & ( CACHE_ENTRY_REFERENCED )) {
713 elru->bei_state &= ~CACHE_ENTRY_REFERENCED;
714 bdb_cache_entryinfo_unlock( elru );
718 /* If this node is in the process of linking into the cache,
719 * or this node is being deleted, skip it.
721 if (( elru->bei_state & ( CACHE_ENTRY_NOT_LINKED |
722 CACHE_ENTRY_DELETED | CACHE_ENTRY_LOADING )) ||
723 elru->bei_finders > 0 ) {
724 bdb_cache_entryinfo_unlock( elru );
728 /* entryinfo is locked */
731 /* If we can successfully writelock it, then
732 * the object is idle.
734 if ( bdb_cache_entry_db_lock( bdb,
735 bdb->bi_cache.c_txn, elru, 1, 1, lockp ) == 0 ) {
737 /* Free entry for this node if it's present */
740 if ( count < efree ) {
741 elru->bei_e->e_private = NULL;
742 #ifdef SLAP_ZONE_ALLOC
743 bdb_entry_return( bdb, elru->bei_e, elru->bei_zseq );
745 bdb_entry_return( elru->bei_e );
750 /* Keep this node cached, skip to next */
751 bdb_cache_entry_db_unlock( bdb, lockp );
755 bdb_cache_entry_db_unlock( bdb, lockp );
758 * If it is a leaf node, and we're over the limit, free it.
760 if ( elru->bei_kids ) {
761 /* Drop from list, we ignore it... */
762 LRU_DEL( &bdb->bi_cache, elru );
763 } else if ( eicount < eifree ) {
764 /* Too many leaf nodes, free this one */
765 bdb_cache_delete_internal( &bdb->bi_cache, elru, 0 );
766 bdb_cache_delete_cleanup( &bdb->bi_cache, elru );
769 } /* Leave on list until we need to free it */
774 bdb_cache_entryinfo_unlock( elru );
776 if ( count >= efree && eicount >= eifree ) {
777 if ( count || ecount > bdb->bi_cache.c_cursize ) {
778 ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
779 /* HACK: we seem to be losing track, fix up now */
780 if ( ecount > bdb->bi_cache.c_cursize )
781 bdb->bi_cache.c_cursize = ecount;
782 bdb->bi_cache.c_cursize -= count;
783 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
788 if ( elnext == bdb->bi_cache.c_lruhead )
795 bdb->bi_cache.c_lruhead = elnext;
796 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
797 bdb->bi_cache.c_purging = 0;
802 struct bdb_info *bdb,
805 EntryInfo ei = { 0 },
810 ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
811 ei2 = (EntryInfo *) avl_find( bdb->bi_cache.c_idtree,
812 (caddr_t) &ei, bdb_id_cmp );
813 ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
818 * cache_find_id - find an entry in the cache, given id.
819 * The entry is locked for Read upon return. Call with flag ID_LOCKED if
820 * the supplied *eip was already locked.
832 struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
834 int rc = 0, load = 0, added = 0;
835 EntryInfo ei = { 0 };
839 #ifdef SLAP_ZONE_ALLOC
840 slap_zh_rlock(bdb->bi_cache.c_zctx);
842 /* If we weren't given any info, see if we have it already cached */
844 again: ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
845 *eip = (EntryInfo *) avl_find( bdb->bi_cache.c_idtree,
846 (caddr_t) &ei, bdb_id_cmp );
848 /* If the lock attempt fails, the info is in use */
849 if ( bdb_cache_entryinfo_trylock( *eip )) {
850 ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
851 /* If this node is being deleted, treat
852 * as if the delete has already finished
854 if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
857 /* otherwise, wait for the info to free up */
858 ldap_pvt_thread_yield();
861 /* If this info isn't hooked up to its parent yet,
862 * unlock and wait for it to be fully initialized
864 if ( (*eip)->bei_state & CACHE_ENTRY_NOT_LINKED ) {
865 bdb_cache_entryinfo_unlock( *eip );
866 ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
867 ldap_pvt_thread_yield();
872 ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
875 /* See if the ID exists in the database; add it to the cache if so */
878 rc = bdb_id2entry( op->o_bd, tid, id, &ep );
880 rc = bdb_cache_find_ndn( op, tid,
882 if ( *eip ) flag |= ID_LOCKED;
884 ep->e_private = NULL;
885 #ifdef SLAP_ZONE_ALLOC
886 bdb_entry_return( bdb, ep, (*eip)->bei_zseq );
888 bdb_entry_return( ep );
894 rc = hdb_cache_find_parent(op, tid, id, eip );
895 if ( rc == 0 ) flag |= ID_LOCKED;
899 /* Ok, we found the info, do we have the entry? */
901 if ( !( flag & ID_LOCKED )) {
902 bdb_cache_entryinfo_lock( *eip );
906 if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
909 (*eip)->bei_finders++;
910 (*eip)->bei_state |= CACHE_ENTRY_REFERENCED;
911 /* Make sure only one thread tries to load the entry */
913 #ifdef SLAP_ZONE_ALLOC
914 if ((*eip)->bei_e && !slap_zn_validate(
915 bdb->bi_cache.c_zctx, (*eip)->bei_e, (*eip)->bei_zseq)) {
916 (*eip)->bei_e = NULL;
917 (*eip)->bei_zseq = 0;
920 if ( !(*eip)->bei_e && !((*eip)->bei_state & CACHE_ENTRY_LOADING)) {
922 (*eip)->bei_state |= CACHE_ENTRY_LOADING;
926 /* Clear the uncached state if we are not
927 * loading it, i.e it is already cached or
928 * another thread is currently loading it.
930 if ( (*eip)->bei_state & CACHE_ENTRY_NOT_CACHED ) {
931 (*eip)->bei_state &= ~CACHE_ENTRY_NOT_CACHED;
937 if ( flag & ID_LOCKED ) {
938 bdb_cache_entryinfo_unlock( *eip );
941 rc = bdb_cache_entry_db_lock( bdb, tid, *eip, load, 0, lock );
942 if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
944 bdb_cache_entry_db_unlock( bdb, lock );
945 } else if ( rc == 0 ) {
948 rc = bdb_id2entry( op->o_bd, tid, id, &ep );
951 ep->e_private = *eip;
956 #ifdef SLAP_ZONE_ALLOC
957 (*eip)->bei_zseq = *((ber_len_t *)ep - 2);
960 bdb_cache_lru_link( bdb, *eip );
962 if (( flag & ID_NOCACHE ) &&
963 ( bdb_cache_entryinfo_trylock( *eip ) == 0 )) {
964 /* Set the cached state only if no other thread
965 * found the info while we were loading the entry.
967 if ( (*eip)->bei_finders == 1 )
968 (*eip)->bei_state |= CACHE_ENTRY_NOT_CACHED;
969 bdb_cache_entryinfo_unlock( *eip );
973 /* If we succeeded, downgrade back to a readlock. */
974 rc = bdb_cache_entry_db_relock( bdb, tid,
977 /* Otherwise, release the lock. */
978 bdb_cache_entry_db_unlock( bdb, lock );
980 } else if ( !(*eip)->bei_e ) {
981 /* Some other thread is trying to load the entry,
982 * wait for it to finish.
984 bdb_cache_entry_db_unlock( bdb, lock );
985 bdb_cache_entryinfo_lock( *eip );
990 /* Check for subtree renames
992 rc = bdb_fix_dn( (*eip)->bei_e, 1 );
994 bdb_cache_entry_db_relock( bdb,
995 tid, *eip, 1, 0, lock );
996 /* check again in case other modifier did it already */
997 if ( bdb_fix_dn( (*eip)->bei_e, 1 ) )
998 rc = bdb_fix_dn( (*eip)->bei_e, 2 );
999 bdb_cache_entry_db_relock( bdb,
1000 tid, *eip, 0, 0, lock );
1005 bdb_cache_entryinfo_lock( *eip );
1006 (*eip)->bei_finders--;
1008 (*eip)->bei_state ^= CACHE_ENTRY_LOADING;
1009 bdb_cache_entryinfo_unlock( *eip );
1012 if ( flag & ID_LOCKED ) {
1013 bdb_cache_entryinfo_unlock( *eip );
1016 ep->e_private = NULL;
1017 #ifdef SLAP_ZONE_ALLOC
1018 bdb_entry_return( bdb, ep, (*eip)->bei_zseq );
1020 bdb_entry_return( ep );
1027 ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
1028 if ( !( flag & ID_NOCACHE )) {
1029 bdb->bi_cache.c_cursize++;
1030 if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize &&
1031 !bdb->bi_cache.c_purging ) {
1033 bdb->bi_cache.c_purging = 1;
1035 } else if ( bdb->bi_cache.c_leaves > bdb->bi_cache.c_eimax && !bdb->bi_cache.c_purging ) {
1037 bdb->bi_cache.c_purging = 1;
1039 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
1042 bdb_cache_lru_purge( bdb );
1045 #ifdef SLAP_ZONE_ALLOC
1046 if (rc == 0 && (*eip)->bei_e) {
1047 slap_zn_rlock(bdb->bi_cache.c_zctx, (*eip)->bei_e);
1049 slap_zh_runlock(bdb->bi_cache.c_zctx);
1062 if ( BEI(e)->bei_kids ) {
1065 if ( BEI(e)->bei_state & CACHE_ENTRY_NO_KIDS ) {
1068 rc = bdb_dn2id_children( op, txn, e );
1069 if ( rc == DB_NOTFOUND ) {
1070 BEI(e)->bei_state |= CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
1075 /* Update the cache after a successful database Add. */
1078 struct bdb_info *bdb,
1081 struct berval *nrdn,
1088 struct berval rdn = e->e_name;
1091 ei.bei_id = e->e_id;
1092 ei.bei_parent = eip;
1093 ei.bei_nrdn = *nrdn;
1096 /* Lock this entry so that bdb_add can run to completion.
1097 * It can only fail if BDB has run out of lock resources.
1099 rc = bdb_cache_entry_db_lock( bdb, txn, &ei, 0, 0, lock );
1101 bdb_cache_entryinfo_unlock( eip );
1106 if ( nrdn->bv_len != e->e_nname.bv_len ) {
1107 char *ptr = ber_bvchr( &rdn, ',' );
1108 assert( ptr != NULL );
1109 rdn.bv_len = ptr - rdn.bv_val;
1111 ber_dupbv( &ei.bei_rdn, &rdn );
1112 if ( eip->bei_dkids ) eip->bei_dkids++;
1115 rc = bdb_entryinfo_add_internal( bdb, &ei, &new );
1116 /* bdb_csn_commit can cause this when adding the database root entry */
1118 new->bei_e->e_private = NULL;
1119 #ifdef SLAP_ZONE_ALLOC
1120 bdb_entry_return( bdb, new->bei_e, new->bei_zseq );
1122 bdb_entry_return( new->bei_e );
1127 new->bei_state |= CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
1128 eip->bei_state &= ~CACHE_ENTRY_NO_KIDS;
1129 if (eip->bei_parent) {
1130 eip->bei_parent->bei_state &= ~CACHE_ENTRY_NO_GRANDKIDS;
1132 bdb_cache_entryinfo_unlock( eip );
1134 ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
1135 ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
1136 ++bdb->bi_cache.c_cursize;
1137 if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize &&
1138 !bdb->bi_cache.c_purging ) {
1140 bdb->bi_cache.c_purging = 1;
1142 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
1144 bdb_cache_lru_link( bdb, new );
1147 bdb_cache_lru_purge( bdb );
1154 struct bdb_info *bdb,
1156 Attribute *newAttrs,
1160 EntryInfo *ei = BEI(e);
1162 /* Get write lock on data */
1163 rc = bdb_cache_entry_db_relock( bdb, txn, ei, 1, 0, lock );
1165 /* If we've done repeated mods on a cached entry, then e_attrs
1166 * is no longer contiguous with the entry, and must be freed.
1169 if ( (void *)e->e_attrs != (void *)(e+1) ) {
1170 attrs_free( e->e_attrs );
1172 e->e_attrs = newAttrs;
1178 * Change the rdn in the entryinfo. Also move to a new parent if needed.
1182 struct bdb_info *bdb,
1184 struct berval *nrdn,
1190 EntryInfo *ei = BEI(e), *pei;
1196 /* Get write lock on data */
1197 rc = bdb_cache_entry_db_relock( bdb, txn, ei, 1, 0, lock );
1198 if ( rc ) return rc;
1200 /* If we've done repeated mods on a cached entry, then e_attrs
1201 * is no longer contiguous with the entry, and must be freed.
1203 if ( (void *)e->e_attrs != (void *)(e+1) ) {
1204 attrs_free( e->e_attrs );
1206 e->e_attrs = new->e_attrs;
1207 if( e->e_nname.bv_val < e->e_bv.bv_val ||
1208 e->e_nname.bv_val > e->e_bv.bv_val + e->e_bv.bv_len )
1210 ch_free(e->e_name.bv_val);
1211 ch_free(e->e_nname.bv_val);
1213 e->e_name = new->e_name;
1214 e->e_nname = new->e_nname;
1216 /* Lock the parent's kids AVL tree */
1217 pei = ei->bei_parent;
1218 bdb_cache_entryinfo_lock( pei );
1219 avl_delete( &pei->bei_kids, (caddr_t) ei, bdb_rdn_cmp );
1220 free( ei->bei_nrdn.bv_val );
1221 ber_dupbv( &ei->bei_nrdn, nrdn );
1224 free( ei->bei_rdn.bv_val );
1227 if ( nrdn->bv_len != e->e_nname.bv_len ) {
1228 char *ptr = ber_bvchr(&rdn, ',');
1229 assert( ptr != NULL );
1230 rdn.bv_len = ptr - rdn.bv_val;
1232 ber_dupbv( &ei->bei_rdn, &rdn );
1234 /* If new parent, decrement kid counts */
1237 if ( pei->bei_dkids ) {
1239 if ( pei->bei_dkids < 2 )
1240 pei->bei_state |= CACHE_ENTRY_NO_KIDS | CACHE_ENTRY_NO_GRANDKIDS;
1246 ein = ei->bei_parent;
1248 ei->bei_parent = ein;
1249 bdb_cache_entryinfo_unlock( pei );
1250 bdb_cache_entryinfo_lock( ein );
1252 /* new parent now has kids */
1253 if ( ein->bei_state & CACHE_ENTRY_NO_KIDS )
1254 ein->bei_state ^= CACHE_ENTRY_NO_KIDS;
1255 /* grandparent has grandkids */
1256 if ( ein->bei_parent )
1257 ein->bei_parent->bei_state &= ~CACHE_ENTRY_NO_GRANDKIDS;
1259 /* parent might now have grandkids */
1260 if ( ein->bei_state & CACHE_ENTRY_NO_GRANDKIDS &&
1261 !(ei->bei_state & CACHE_ENTRY_NO_KIDS))
1262 ein->bei_state ^= CACHE_ENTRY_NO_GRANDKIDS;
1265 if ( ein->bei_dkids ) ein->bei_dkids++;
1270 /* Record the generation number of this change */
1271 ldap_pvt_thread_mutex_lock( &bdb->bi_modrdns_mutex );
1273 ei->bei_modrdns = bdb->bi_modrdns;
1274 ldap_pvt_thread_mutex_unlock( &bdb->bi_modrdns_mutex );
1277 avl_insert( &ein->bei_kids, ei, bdb_rdn_cmp, avl_dup_error );
1278 bdb_cache_entryinfo_unlock( ein );
1282 * cache_delete - delete the entry e from the cache.
1284 * returns: 0 e was deleted ok
1285 * 1 e was not in the cache
1286 * -1 something bad happened
1290 struct bdb_info *bdb,
1295 EntryInfo *ei = BEI(e);
1298 assert( e->e_private != NULL );
1300 /* Lock the entry's info */
1301 bdb_cache_entryinfo_lock( ei );
1303 /* Set this early, warn off any queriers */
1304 ei->bei_state |= CACHE_ENTRY_DELETED;
1306 bdb_cache_entryinfo_unlock( ei );
1308 /* Get write lock on the data */
1309 rc = bdb_cache_entry_db_relock( bdb, txn, ei, 1, 0, lock );
1311 /* couldn't lock, undo and give up */
1312 ei->bei_state ^= CACHE_ENTRY_DELETED;
1316 Debug( LDAP_DEBUG_TRACE, "====> bdb_cache_delete( %ld )\n",
1320 ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_lru_mutex );
1322 rc = bdb_cache_delete_internal( &bdb->bi_cache, e->e_private, 1 );
1324 /* free lru mutex */
1325 ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
1331 bdb_cache_delete_cleanup(
1335 /* Enter with ei locked */
1338 ei->bei_e->e_private = NULL;
1339 #ifdef SLAP_ZONE_ALLOC
1340 bdb_entry_return( ei->bei_bdb, ei->bei_e, ei->bei_zseq );
1342 bdb_entry_return( ei->bei_e );
1347 bdb_cache_entryinfo_free( cache, ei );
1348 bdb_cache_entryinfo_unlock( ei );
1352 bdb_cache_delete_internal(
1357 int rc = 0; /* return code */
1360 /* Lock the parent's kids tree */
1361 bdb_cache_entryinfo_lock( e->bei_parent );
1364 e->bei_parent->bei_ckids--;
1365 if ( decr && e->bei_parent->bei_dkids ) e->bei_parent->bei_dkids--;
1368 if ( avl_delete( &e->bei_parent->bei_kids, (caddr_t) e, bdb_rdn_cmp )
1373 if ( e->bei_parent->bei_kids )
1376 bdb_cache_entryinfo_unlock( e->bei_parent );
1378 ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
1380 if ( avl_delete( &cache->c_idtree, (caddr_t) e, bdb_id_cmp )) {
1387 ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
1391 LRU_DEL( cache, e );
1394 ldap_pvt_thread_mutex_lock( &cache->c_count_mutex );
1396 ldap_pvt_thread_mutex_unlock( &cache->c_count_mutex );
1404 bdb_entryinfo_release( void *data )
1406 EntryInfo *ei = (EntryInfo *)data;
1407 if ( ei->bei_kids ) {
1408 avl_free( ei->bei_kids, NULL );
1411 ei->bei_e->e_private = NULL;
1412 #ifdef SLAP_ZONE_ALLOC
1413 bdb_entry_return( ei->bei_bdb, ei->bei_e, ei->bei_zseq );
1415 bdb_entry_return( ei->bei_e );
1418 bdb_cache_entryinfo_destroy( ei );
1422 bdb_cache_release_all( Cache *cache )
1424 /* set cache write lock */
1425 ldap_pvt_thread_rdwr_wlock( &cache->c_rwlock );
1427 ldap_pvt_thread_mutex_lock( &cache->c_lru_mutex );
1429 Debug( LDAP_DEBUG_TRACE, "====> bdb_cache_release_all\n", 0, 0, 0 );
1431 avl_free( cache->c_dntree.bei_kids, NULL );
1432 avl_free( cache->c_idtree, bdb_entryinfo_release );
1433 for (;cache->c_eifree;cache->c_eifree = cache->c_lruhead) {
1434 cache->c_lruhead = cache->c_eifree->bei_lrunext;
1435 bdb_cache_entryinfo_destroy(cache->c_eifree);
1437 cache->c_cursize = 0;
1438 cache->c_eiused = 0;
1439 cache->c_leaves = 0;
1440 cache->c_idtree = NULL;
1441 cache->c_lruhead = NULL;
1442 cache->c_lrutail = NULL;
1443 cache->c_dntree.bei_kids = NULL;
1445 /* free lru mutex */
1446 ldap_pvt_thread_mutex_unlock( &cache->c_lru_mutex );
1447 /* free cache write lock */
1448 ldap_pvt_thread_rdwr_wunlock( &cache->c_rwlock );
1454 bdb_lru_print( Cache *cache )
1458 fprintf( stderr, "LRU circle head: %p\n", (void *) cache->c_lruhead );
1459 fprintf( stderr, "LRU circle (tail forward):\n" );
1460 for ( e = cache->c_lrutail; ; ) {
1461 fprintf( stderr, "\t%p, %p id %ld rdn \"%s\"\n",
1462 (void *) e, (void *) e->bei_e, e->bei_id, e->bei_nrdn.bv_val );
1464 if ( e == cache->c_lrutail )
1467 fprintf( stderr, "LRU circle (tail backward):\n" );
1468 for ( e = cache->c_lrutail; ; ) {
1469 fprintf( stderr, "\t%p, %p id %ld rdn \"%s\"\n",
1470 (void *) e, (void *) e->bei_e, e->bei_id, e->bei_nrdn.bv_val );
1472 if ( e == cache->c_lrutail )
1480 bdb_reader_free( void *key, void *data )
1482 /* DB_ENV *env = key; */
1488 /* free up any keys used by the main thread */
1490 bdb_reader_flush( DB_ENV *env )
1493 void *ctx = ldap_pvt_thread_pool_context();
1495 if ( !ldap_pvt_thread_pool_getkey( ctx, env, &data, NULL ) ) {
1496 ldap_pvt_thread_pool_setkey( ctx, env, NULL, 0, NULL, NULL );
1497 bdb_reader_free( env, data );
1502 bdb_reader_get( Operation *op, DB_ENV *env, DB_TXN **txn )
1508 if ( !env || !txn ) return -1;
1510 /* If no op was provided, try to find the ctx anyway... */
1512 ctx = op->o_threadctx;
1514 ctx = ldap_pvt_thread_pool_context();
1517 /* Shouldn't happen unless we're single-threaded */
1523 if ( ldap_pvt_thread_pool_getkey( ctx, env, &data, NULL ) ) {
1524 for ( i=0, rc=1; rc != 0 && i<4; i++ ) {
1525 rc = TXN_BEGIN( env, NULL, txn, DB_READ_COMMITTED );
1526 if (rc) ldap_pvt_thread_yield();
1532 if ( ( rc = ldap_pvt_thread_pool_setkey( ctx, env,
1533 data, bdb_reader_free, NULL, NULL ) ) ) {
1535 Debug( LDAP_DEBUG_ANY, "bdb_reader_get: err %s(%d)\n",
1536 db_strerror(rc), rc, 0 );