In src/kdc/network.c, in the function:

process_packet():

We find:

    response = NULL;
    saddr_len = sizeof(saddr);
    cc = recvfrom(port_fd, pktbuf, sizeof(pktbuf), 0,
                  (struct sockaddr *)&saddr, &saddr_len);
    if (cc == -1) {
        if (errno != EINTR
            /* This is how Linux indicates that a previous
               transmission was refused, e.g., if the client timed out
               before getting the response packet.  */
            && errno != ECONNREFUSED
            )
            com_err(prog, errno, "while receiving from network");
        return;
    }
    if (!cc)
        return;         /* zero-length packet? */

Unfortunately, if you receive a datagram of over sizeof(pktbuf)
you will succeed with cc == sizeof(pktbuf) not detecting the fact
that there was additional data.  This results in an ASN.1 parse
error.  What should happen is that the KDC should return an
appropriate error to the client indicating that TCP should be used.
Or maybe the buffer size should be increased to the maximum allowable
for UDP.  I prefer the second option as there is nothing inherently
wrong with 64K UDP packets.

I noticed this while debugging a JGSS problem.  Apparently, the
Java Kerberos libraries do not fail over from UDP to TCP unless
the KDC specifically tells them to.  And they have no default
setting for udp_preference_limit.  And so, if you are constructing
tickets of over 4K because, let's say, a user is in a lot of groups
in Windows, JGSS will just fail against an MIT KDC.

Fix:

change MAX_DGRAM_SIZE in /include/krb5/stock/osconf.h to be the
actual maximum datagram size, 65536.

--
    Roland Dowdeswell                      http://Imrryr.ORG/~elric/