RE: Index corruption and crash in back-ldbm (ITS#2348)

[vek@pharmapartners.nl]

On Wed, 5 Mar 2003 hyc@highlandsun.com wrote:

> Date: Wed, 5 Mar 2003 23:08:18 GMT
> From: hyc@highlandsun.com
> To: openldap-its@OpenLDAP.org
> Subject: RE: Index corruption and crash in back-ldbm (ITS#2348)
>
> Ignore that previous message. Your bug analysis and patch are exactly right.
> I am committing it to both 2.1 and 2.0.
>

It wasn't complete, though.  The events shown below can be reproduced
by running slapindex without erasing the indexes first, thus when
all the id entries already exist.  If you lower db->dbc_maxids while testing
it would happen much sooner.


Must select the block where id is greater OR EQUEAL to the lower
bound for this block.  If id is equal to the lower bound and
the previous block is selected for insert then the following message
may be triggered.

   "idl_insert_key: id %ld is already in next block\n"

The condition for this message is that the previous block is full
and the id already exist in the following block.  If the previous
block is not full then the id is inserted there even when it may
already exist in the following block, thus an inconsistency.  When
this happens no message is generated.

The patch is for CVS id 1.80.


--- idl.c-1.80	Mon Mar 10 10:43:31 2003
+++ idl.c	Mon Mar 10 10:48:07 2003
@@ -553,11 +553,11 @@

 #ifndef USE_INDIRECT_NIDS
 	/* select the block to try inserting into *//* XXX linear search XXX */
-	for ( i = 0; !ID_BLOCK_NOID(idl, i) && id > ID_BLOCK_ID(idl, i); i++ )
+	for ( i = 0; !ID_BLOCK_NOID(idl, i) && id >= ID_BLOCK_ID(idl, i); i++ )
 		;	/* NULL */
 #else
 	i = idl_find(idl, id);
-	if (ID_BLOCK_ID(idl, i) < id)
+	if (ID_BLOCK_ID(idl, i) <= id)
 		i++;
 #endif

@@ -567,6 +567,11 @@
 	} else {
 		first = 1;
 	}
+
+	/* At this point  the following condition must be true:
+	 * ID_BLOCK_ID(idl, i) <= id && id < ID_BLOCK_ID(idl, i+1)
+	 * except when i is the first or the last block.
+	 */

 	/* get the block */
 	cont_alloc( &k2, &key );





--

Villy