NT 3.51 dhcp client and server ip address

[Tim Rowe <[e-mail]>]

While testing our ISC dhcpd port, I ran into some unfortunate behavior.
It appears that the Windows/NT 3.51 dhcp client uses the Server
Identifier as the IP address of the dhcp server instead of using the
source IP address of the DHCPACK (or any other message) sent from the
server.  This is particularly problematic for multiple interface
support, since the Server Identifier may not be on the same subnet as
the client.  In fact, the client may not have any way to get to a
network where the Server Identifier is valid.  As a result, unicast
DHCPREQUEST messages sent to renew a lease may not reach the server
who holds the lease, ultimately requiring the client to fall back to
another DHCPDISCOVER.

As I read the RFC, there is no requirement that the Server Identifier
be a valid IP address of the server for any subnet, much less the one
the client is on.  Unfortunately, nothing in the standard prohibits
clients from using it that way.  My understanding is that the Server
Identifier is an opaque key that the client shouldn't try to interpret
literally.  Conceivably, it would be sufficient to have it be unique
on a per-subnet basis, though per-server would be ideal.

In summary, though I don't think the Windows/NT client behavior is
currently prohibited, it should be.  Server Identifiers should be
defined in the RFC as opaque, with clients disallowed from using them
as the IP address of the dhcp server.

Any other opinions?
-- 
Tim Rowe                                        Sequent Computer Systems
[phone]                                  15450 SW Koll Parkway
[e-mail]                                Beaverton, OR 97006-6063