void
bdb_cache_return_entry_rw( struct bdb_info *bdb, Entry *e,
int rw, DB_LOCK *lock )
{
EntryInfo *ei;
int free = 0;
ei = e->e_private;
if ( ei &&
( ei->bei_state & CACHE_ENTRY_NOT_CACHED ) &&
( bdb_cache_entryinfo_trylock( ei ) == 0 )) {
if ( ei->bei_state & CACHE_ENTRY_NOT_CACHED ) {
/* Releasing the entry can only be done when
* we know that nobody else is using it, i.e we
* should have an entry_db writelock. But the
* flag is only set by the thread that loads the
* entry, and only if no other threads has found
* it while it was working. All other threads
* clear the flag, which mean that we should be
* the only thread using the entry if the flag
* is set here.
*/
ei->bei_e = NULL;
ei->bei_state ^= CACHE_ENTRY_NOT_CACHED;
free = 1;
}
bdb_cache_entryinfo_unlock( ei );
}
bdb_cache_entry_db_unlock( bdb, lock );
if ( free ) {
e->e_private = NULL;
bdb_entry_return( e );
}
}
int
bdb_cache_find_id(
Operation *op,
DB_TXN *tid,
ID id,
EntryInfo **eip,
int flag,
DB_LOCK *lock )
{
struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
Entry *ep = NULL;
int rc = 0, load = 0;
EntryInfo ei = { 0 };
ei.bei_id = id;
#ifdef SLAP_ZONE_ALLOC
slap_zh_rlock(bdb->bi_cache.c_zctx);
#endif
/* If we weren't given any info, see if we have it already cached */
if ( !*eip ) {
again: ldap_pvt_thread_rdwr_rlock( &bdb->bi_cache.c_rwlock );
*eip = (EntryInfo *) avl_find( bdb->bi_cache.c_idtree,
(caddr_t) &ei, bdb_id_cmp );
if ( *eip ) {
/* If the lock attempt fails, the info is in use */
if ( bdb_cache_entryinfo_trylock( *eip )) {
int del = (*eip)->bei_state & CACHE_ENTRY_DELETED;
ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
/* If this node is being deleted, treat
* as if the delete has already finished
*/
if ( del ) {
return DB_NOTFOUND;
}
/* otherwise, wait for the info to free up */
ldap_pvt_thread_yield();
goto again;
}
/* If this info isn't hooked up to its parent yet,
* unlock and wait for it to be fully initialized
*/
if ( (*eip)->bei_state & CACHE_ENTRY_NOT_LINKED ) {
bdb_cache_entryinfo_unlock( *eip );
ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
ldap_pvt_thread_yield();
goto again;
}
flag |= ID_LOCKED;
}
ldap_pvt_thread_rdwr_runlock( &bdb->bi_cache.c_rwlock );
}
/* See if the ID exists in the database; add it to the cache if so */
if ( !*eip ) {
#ifndef BDB_HIER
rc = bdb_id2entry( op->o_bd, tid, id, &ep );
if ( rc == 0 ) {
rc = bdb_cache_find_ndn( op, tid,
&ep->e_nname, eip );
if ( *eip ) flag |= ID_LOCKED;
if ( rc ) {
ep->e_private = NULL;
#ifdef SLAP_ZONE_ALLOC
bdb_entry_return( bdb, ep, (*eip)->bei_zseq );
#else
bdb_entry_return( ep );
#endif
ep = NULL;
}
}
#else
rc = hdb_cache_find_parent(op, tid, id, eip );
if ( rc == 0 ) flag |= ID_LOCKED;
#endif
}
/* Ok, we found the info, do we have the entry? */
if ( rc == 0 ) {
if ( !( flag & ID_LOCKED )) {
bdb_cache_entryinfo_lock( *eip );
flag |= ID_LOCKED;
}
if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
rc = DB_NOTFOUND;
} else {
(*eip)->bei_finders++;
(*eip)->bei_state |= CACHE_ENTRY_REFERENCED;
if ( flag & ID_NOENTRY ) {
bdb_cache_entryinfo_unlock( *eip );
return 0;
}
/* Make sure only one thread tries to load the entry */
load1:
#ifdef SLAP_ZONE_ALLOC
if ((*eip)->bei_e && !slap_zn_validate(
bdb->bi_cache.c_zctx, (*eip)->bei_e, (*eip)->bei_zseq)) {
(*eip)->bei_e = NULL;
(*eip)->bei_zseq = 0;
}
#endif
if ( !(*eip)->bei_e && !((*eip)->bei_state & CACHE_ENTRY_LOADING)) {
load = 1;
(*eip)->bei_state |= CACHE_ENTRY_LOADING;
flag |= ID_CHKPURGE;
}
if ( !load ) {
/* Clear the uncached state if we are not
* loading it, i.e it is already cached or
* another thread is currently loading it.
*/
if ( (*eip)->bei_state & CACHE_ENTRY_NOT_CACHED ) {
(*eip)->bei_state ^= CACHE_ENTRY_NOT_CACHED;
flag |= ID_CHKPURGE;
}
}
if ( flag & ID_LOCKED ) {
bdb_cache_entryinfo_unlock( *eip );
flag ^= ID_LOCKED;
}
rc = bdb_cache_entry_db_lock( bdb, tid, *eip, load, 0, lock );
if ( (*eip)->bei_state & CACHE_ENTRY_DELETED ) {
rc = DB_NOTFOUND;
bdb_cache_entry_db_unlock( bdb, lock );
bdb_cache_entryinfo_lock( *eip );
(*eip)->bei_finders--;
bdb_cache_entryinfo_unlock( *eip );
} else if ( rc == 0 ) {
if ( load ) {
if ( !ep) {
rc = bdb_id2entry( op->o_bd, tid, id, &ep );
}
if ( rc == 0 ) {
ep->e_private = *eip;
#ifdef BDB_HIER
while ( (*eip)->bei_state & CACHE_ENTRY_NOT_LINKED )
ldap_pvt_thread_yield();
bdb_fix_dn( ep, 0 );
#endif
bdb_cache_entryinfo_lock( *eip );
(*eip)->bei_e = ep;
#ifdef SLAP_ZONE_ALLOC
(*eip)->bei_zseq = *((ber_len_t *)ep - 2);
#endif
ep = NULL;
if ( flag & ID_NOCACHE ) {
/* Set the cached state only if no other thread
* found the info while we were loading the entry.
*/
if ( (*eip)->bei_finders == 1 ) {
(*eip)->bei_state |= CACHE_ENTRY_NOT_CACHED;
flag ^= ID_CHKPURGE;
}
}
bdb_cache_entryinfo_unlock( *eip );
bdb_cache_lru_link( bdb, *eip );
}
if ( rc == 0 ) {
/* If we succeeded, downgrade back to a readlock. */
rc = bdb_cache_entry_db_relock( bdb, tid,
*eip, 0, 0, lock );
} else {
/* Otherwise, release the lock. */
bdb_cache_entry_db_unlock( bdb, lock );
}
} else if ( !(*eip)->bei_e ) {
/* Some other thread is trying to load the entry,
* wait for it to finish.
*/
bdb_cache_entry_db_unlock( bdb, lock );
bdb_cache_entryinfo_lock( *eip );
flag |= ID_LOCKED;
goto load1;
#ifdef BDB_HIER
} else {
/* Check for subtree renames
*/
rc = bdb_fix_dn( (*eip)->bei_e, 1 );
if ( rc ) {
bdb_cache_entry_db_relock( bdb,
tid, *eip, 1, 0, lock );
/* check again in case other modifier did it already */
if ( bdb_fix_dn( (*eip)->bei_e, 1 ) )
rc = bdb_fix_dn( (*eip)->bei_e, 2 );
bdb_cache_entry_db_relock( bdb,
tid, *eip, 0, 0, lock );
}
#endif
}
bdb_cache_entryinfo_lock( *eip );
(*eip)->bei_finders--;
if ( load )
(*eip)->bei_state ^= CACHE_ENTRY_LOADING;
bdb_cache_entryinfo_unlock( *eip );
}
}
}
if ( flag & ID_LOCKED ) {
bdb_cache_entryinfo_unlock( *eip );
}
if ( ep ) {
ep->e_private = NULL;
#ifdef SLAP_ZONE_ALLOC
bdb_entry_return( bdb, ep, (*eip)->bei_zseq );
#else
bdb_entry_return( ep );
#endif
}
if ( rc == 0 ) {
int purge = 0;
if (( flag & ID_CHKPURGE ) || bdb->bi_cache.c_eimax ) {
ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
if ( flag & ID_CHKPURGE ) {
bdb->bi_cache.c_cursize++;
if ( !bdb->bi_cache.c_purging && bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize ) {
purge = 1;
bdb->bi_cache.c_purging = 1;
}
} else if ( !bdb->bi_cache.c_purging && bdb->bi_cache.c_eimax && bdb->bi_cache.c_leaves > bdb->bi_cache.c_eimax ) {
purge = 1;
bdb->bi_cache.c_purging = 1;
}
ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
}
if ( purge )
bdb_cache_lru_purge( bdb );
}
#ifdef SLAP_ZONE_ALLOC
if (rc == 0 && (*eip)->bei_e) {
slap_zn_rlock(bdb->bi_cache.c_zctx, (*eip)->bei_e);
}
slap_zh_runlock(bdb->bi_cache.c_zctx);
#endif
return rc;
}
/* This is best-effort only. If all entries in the cache are
* busy, they will all be kept. This is unlikely to happen
* unless the cache is very much smaller than the working set.
*/
static void
bdb_cache_lru_purge( struct bdb_info *bdb )
{
DB_LOCK lock, *lockp;
EntryInfo *elru, *elnext = NULL;
int islocked;
ID eicount, ecount;
ID count, efree, eifree = 0;
#ifdef LDAP_DEBUG
int iter;
#endif
/* Wait for the mutex; we're the only one trying to purge. */
ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_lru_mutex );
if ( bdb->bi_cache.c_cursize > bdb->bi_cache.c_maxsize ) {
efree = bdb->bi_cache.c_cursize - bdb->bi_cache.c_maxsize;
efree += bdb->bi_cache.c_minfree;
} else {
efree = 0;
}
/* maximum number of EntryInfo leaves to cache. In slapcat
* we always free all leaf nodes.
*/
if ( slapMode & SLAP_TOOL_READONLY ) {
eifree = bdb->bi_cache.c_leaves;
} else if ( bdb->bi_cache.c_eimax &&
bdb->bi_cache.c_leaves > bdb->bi_cache.c_eimax ) {
eifree = bdb->bi_cache.c_minfree * 10;
if ( eifree >= bdb->bi_cache.c_leaves )
eifree /= 2;
}
if ( !efree && !eifree ) {
ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
bdb->bi_cache.c_purging = 0;
return;
}
if ( bdb->bi_cache.c_txn ) {
lockp = &lock;
} else {
lockp = NULL;
}
count = 0;
eicount = 0;
ecount = 0;
#ifdef LDAP_DEBUG
iter = 0;
#endif
/* Look for an unused entry to remove */
for ( elru = bdb->bi_cache.c_lruhead; elru; elru = elnext ) {
elnext = elru->bei_lrunext;
if ( bdb_cache_entryinfo_trylock( elru ))
goto bottom;
/* This flag implements the clock replacement behavior */
if ( elru->bei_state & ( CACHE_ENTRY_REFERENCED )) {
elru->bei_state &= ~CACHE_ENTRY_REFERENCED;
bdb_cache_entryinfo_unlock( elru );
goto bottom;
}
/* If this node is in the process of linking into the cache,
* or this node is being deleted, skip it.
*/
if (( elru->bei_state & ( CACHE_ENTRY_NOT_LINKED |
CACHE_ENTRY_DELETED | CACHE_ENTRY_LOADING |
CACHE_ENTRY_ONELEVEL )) ||
elru->bei_finders > 0 ) {
bdb_cache_entryinfo_unlock( elru );
goto bottom;
}
if ( bdb_cache_entryinfo_trylock( elru->bei_parent )) {
bdb_cache_entryinfo_unlock( elru );
goto bottom;
}
/* entryinfo is locked */
islocked = 1;
/* If we can successfully writelock it, then
* the object is idle.
*/
if ( bdb_cache_entry_db_lock( bdb,
bdb->bi_cache.c_txn, elru, 1, 1, lockp ) == 0 ) {
/* Free entry for this node if it's present */
if ( elru->bei_e ) {
ecount++;
/* the cache may have gone over the limit while we
* weren't looking, so double check.
*/
if ( !efree && ecount > bdb->bi_cache.c_maxsize )
efree = bdb->bi_cache.c_minfree;
if ( count < efree ) {
elru->bei_e->e_private = NULL;
#ifdef SLAP_ZONE_ALLOC
bdb_entry_return( bdb, elru->bei_e, elru->bei_zseq );
#else
bdb_entry_return( elru->bei_e );
#endif
elru->bei_e = NULL;
count++;
} else {
/* Keep this node cached, skip to next */
bdb_cache_entry_db_unlock( bdb, lockp );
goto next;
}
}
bdb_cache_entry_db_unlock( bdb, lockp );
/*
* If it is a leaf node, and we're over the limit, free it.
*/
if ( elru->bei_kids ) {
/* Drop from list, we ignore it... */
LRU_DEL( &bdb->bi_cache, elru );
} else if ( eicount < eifree ) {
/* Too many leaf nodes, free this one */
bdb_cache_delete_internal( &bdb->bi_cache, elru, 0 );
bdb_cache_delete_cleanup( &bdb->bi_cache, elru );
islocked = 0;
eicount++;
} /* Leave on list until we need to free it */
}
next:
if ( islocked ) {
bdb_cache_entryinfo_unlock( elru );
bdb_cache_entryinfo_unlock( elru->bei_parent );
}
if ( count >= efree && eicount >= eifree )
break;
bottom:
if ( elnext == bdb->bi_cache.c_lruhead )
break;
#ifdef LDAP_DEBUG
iter++;
#endif
}
if ( count || ecount > bdb->bi_cache.c_cursize ) {
ldap_pvt_thread_mutex_lock( &bdb->bi_cache.c_count_mutex );
/* HACK: we seem to be losing track, fix up now */
if ( ecount > bdb->bi_cache.c_cursize )
bdb->bi_cache.c_cursize = ecount;
bdb->bi_cache.c_cursize -= count;
ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_count_mutex );
}
bdb->bi_cache.c_lruhead = elnext;
ldap_pvt_thread_mutex_unlock( &bdb->bi_cache.c_lru_mutex );
bdb->bi_cache.c_purging = 0;
}
/* Walk up the tree from a child node, looking for an ID that's already
* been linked into the cache.
*/
int
hdb_cache_find_parent(
Operation *op,
DB_TXN *txn,
ID id,
EntryInfo **res )
{
struct bdb_info *bdb = (struct bdb_info *) op->o_bd->be_private;
EntryInfo ei, eip, *ei2 = NULL, *ein = NULL, *eir = NULL;
int rc, add;
ei.bei_id = id;
ei.bei_kids = NULL;
ei.bei_ckids = 0;
for (;;) {
rc = hdb_dn2id_parent( op, txn, &ei, &eip.bei_id );
if ( rc ) break;
/* Save the previous node, if any */
ei2 = ein;
/* Create a new node for the current ID */
ein = bdb_cache_entryinfo_new( &bdb->bi_cache );
ein->bei_id = ei.bei_id;
ein->bei_kids = ei.bei_kids;
ein->bei_nrdn = ei.bei_nrdn;
ein->bei_rdn = ei.bei_rdn;
ein->bei_ckids = ei.bei_ckids;
#ifdef SLAP_ZONE_ALLOC
ein->bei_bdb = bdb;
#endif
ei.bei_ckids = 0;
add = 1;
/* This node is not fully connected yet */
ein->bei_state |= CACHE_ENTRY_NOT_LINKED;
/* If this is the first time, save this node
* to be returned later.
*/
if ( eir == NULL ) {
eir = ein;
ein->bei_finders++;
}
again:
/* Insert this node into the ID tree */
ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
if ( avl_insert( &bdb->bi_cache.c_idtree, (caddr_t)ein,
bdb_id_cmp, bdb_id_dup_err ) ) {
EntryInfo *eix = ein->bei_lrunext;
if ( bdb_cache_entryinfo_trylock( eix )) {
ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
ldap_pvt_thread_yield();
goto again;
}
ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
/* Someone else created this node just before us.
* Free our new copy and use the existing one.
*/
bdb_cache_entryinfo_free( &bdb->bi_cache, ein );
/* if it was the node we were looking for, just return it */
if ( eir == ein ) {
*res = eix;
rc = 0;
break;
}
ein = ei2;
ei2 = eix;
add = 0;
/* otherwise, link up what we have and return */
goto gotparent;
}
/* If there was a previous node, link it to this one */
if ( ei2 ) ei2->bei_parent = ein;
/* Look for this node's parent */
par2:
if ( eip.bei_id ) {
ei2 = (EntryInfo *) avl_find( bdb->bi_cache.c_idtree,
(caddr_t) &eip, bdb_id_cmp );
} else {
ei2 = &bdb->bi_cache.c_dntree;
}
if ( ei2 && bdb_cache_entryinfo_trylock( ei2 )) {
ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
ldap_pvt_thread_yield();
ldap_pvt_thread_rdwr_wlock( &bdb->bi_cache.c_rwlock );
goto par2;
}
if ( add )
bdb->bi_cache.c_eiused++;
if ( ei2 && ( ei2->bei_kids || !ei2->bei_id ))
bdb->bi_cache.c_leaves++;
ldap_pvt_thread_rdwr_wunlock( &bdb->bi_cache.c_rwlock );
gotparent:
/* Got the parent, link in and we're done. */
if ( ei2 ) {
bdb_cache_entryinfo_lock( eir );
ein->bei_parent = ei2;
if ( avl_insert( &ei2->bei_kids, (caddr_t)ein, bdb_rdn_cmp,
avl_dup_error) == 0 )
ei2->bei_ckids++;
/* Reset all the state info */
for (ein = eir; ein != ei2; ein=ein->bei_parent)
ein->bei_state &= ~CACHE_ENTRY_NOT_LINKED;
bdb_cache_entryinfo_unlock( ei2 );
eir->bei_finders--;
*res = eir;
break;
}
ei.bei_kids = NULL;
ei.bei_id = eip.bei_id;
ei.bei_ckids = 1;
avl_insert( &ei.bei_kids, (caddr_t)ein, bdb_rdn_cmp,
avl_dup_error );
}
return rc;
}