DHCP support for InfiniBand

Barr Hibbs rbhibbs at pacbell.net
Thu Mar 16 19:01:31 UTC 2006

> -----Original Message-----
> From: David W. Hankins
> Sent: Thursday, March 16, 2006 08:34
> I'm not seeing anything in there that relies on the server
to echo the
> client id back.
...I had totally missed the discussion about InfiniBand on
the DHC Working Group mailing list and offered some comments
to the author of that draft and the IP Over InfiniBand
Working Group.  Unfortunately, the draft had passed last
call and is advancing towards RFC status.

I very carefully read the draft and commented on practical
issues, not the least of which is that some DHCP servers
(and probably relay agents, as well) are known to discard
"ill-formed" packets, including those with 'hlen' zero and
'chaddr' either all zero bits or all one bits.  I also noted
that implementors cannot rely on all DHCP servers to
"support" the 3315-style Client Identifiers called for in
the draft [not that it would matter, really, as servers are
supposed to treat the Client Identifier as an opaque value,
so I remain puzzled why they called for it to be formatted
that way.]

Finally, I agree with both Eli and David that servers are
NOT required to return the Client Identifier and that the
draft doesn't call for it, so other than the use of
BROADCAST for DHCPOFFERs and DHCPACKs, it doesn't look like
any server modifications would be necessary.  [David--I
haven't checked the source code in a long time, but DHCPD
did not validate 'htype' in the past:  if it does today,
that might be the only change necessary, to permit address
type 32.]


More information about the dhcp-users mailing list