kacarstensen@csupomona.edu wrote:

I've investigated this issue a little bit more since my initial bug report.

I'm not sure if connection_write is supposed to validate that a stream is active before or after calling slapd_clr_write, but it seems like the assertion wouldn't be an issue if that validation were performed before calling slapd_clr_write. To test this thought, I rebuilt openldap 2.4.28 with the following patch:

--- openldap-2.4.28/servers/slapd/connection.c 2011-11-25 10:52:29.000000000 -0800 +++ openldap-2.4.28-new/servers/slapd/connection.c 2012-01-12 13:35:45.000000000 -0800 @@ -1893,8 +1893,6 @@

 assert( connections != NULL );

• slapd_clr_write( s, 0 );
• c = connection_get( s ); if( c == NULL ) { Debug( LDAP_DEBUG_ANY,

@@ -1903,6 +1901,8 @@ return -1; }

• slapd_clr_write( s, 0 );

#ifdef HAVE_TLS
 if ( c->c_is_tls&&  c->c_needs_tls_accept ) {
     connection_return( c );

and tried to reproduce the problem under the same circumstances as reported in my initial bug report. The master slapd tolerated the misconfigured replicas for 5 days without crashing; before, it would crash reliably within a half hour or so. I didn't notice any regressions due to the patch, though the master slapd wasn't exposed to a typical workload during the experiment.

Any thoughts on this patch?

Sounds OK to me, committed to git master. Thanks.