[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: IPv6 WG Call for Adoption:draft-daniel-ipv6-ra-mo-flags-01.txt



>>>>> On Wed, 27 Oct 2004 20:23:52 +0300 (EEST), 
>>>>> Pekka Savola <pekkas@xxxxxxxxxx> said:

>> Title		: Consideration M and O Flags of IPv6 Router 
>> Advertisement
>> Author(s)	: S. Park, et al.
>> Filename	: draft-daniel-ipv6-ra-mo-flags-01.txt
>> Pages		: 12
>> Date		: 2004-10-25
>> 
>> as a WG document.  Comments and opinions can be sent to the
>> mailing list and/or the chairs.  We will accept comments on this
>> call until 11/05/2004.

> I think this is valuable clarifying work, and a good starting point 
> for the work.  I think it's worth adopting.

I'm glad to hear your support.  I'd also appreciate comments from
others, either positive or negative, since the adoption or
non-adoption will also affect the next step of rfc2462bis regarding AD
comments.

> A couple of comments on the draft for improvement:

> - just a question on how full DHCPv6 client works with stateless 
> DHCPv6 server:

>    If M-Policy is 1, the host SHOULD invoke Host Configuration Behaviour
>     for address and other configuration information, regardless of the
>     change of the state variables.  The host SHOULD NOT invoke
>     Information Configuration Behaviour regardless of O-Policy.  Note,
>     however, that if the available DHCPv6 servers only provide the
>     service for the Information Configuration Behaviour, the host will
>     even not be able to configure other configuration parameters than
>     addresses.  Thus, it is generally inadvisable to set M-Policy to 1,
>     unless there is a particular reason to do so.

> ==> is it really true that if a full DHCPv6 client is initiated, and 
> the server is 'stateless', the DHCPv6 client doesn't get back the 
> stateless information?

> That would seem to be a big problem for DHCPv6 deployment if servers 
> are stateless, and I think someone has claimed that this shouldn't be 
> a problem.

> If the text you wrote is correct, that would mean that to get robust 
> behaviour, the host would have to either rely that the capabilities of 
> the DHCPv6 servers match those that are advertised in the route 
> advertisements, or to always just "prepare for the worst" by firing 
> off a parallel information-request as well.

> And then it would seem to make sense for prudent DHCPv6 implementers 
> to do everything they can to ensure they'll receive information-reply 
> from stateful servers as well.. because that's better than nothing, 
> and better than relying on how the bits in RAs have been configured!

Good points.  In fact, I thought this was highly controversial and
wanted to ask opinions from others.

But, I'd first like to clarify some things.  We intentionally avoided
using the term "stateful DHCPv6" and "stateless DHCPv6" because these
terms are very confusing as we discussed in this list before.  I'm
going to use the following new terms defined in Section 3 of the draft:

  Host Configuration Behaviour:
     configuring address and other config information through
     Solicit/Advertise/Request/Reply or Solicit/Reply (if rapid commit
     is enabled) exchanges. [Note: in this message, I'll forget the
     short-cut exchange with rapid commit for simplicity.]

  Information Configuration Behaviour:
     configuring other information excluding addresses through
     Information-request and Reply exchanges

Secondly, the part you cited is not related to the M/O flags.  Policy
1 means the host SHOULD perform the Host/Information Configuration
Behaviour *regardless of* the M/O flags.

But your main point still stands, and I agree this is a problematic
case.  I'd first note I don't think it's a true-or-false thing; we're
just going to define the most reasonable behavior, and this is its
first step.  The background of this statement is the dependency
described in Section 6, which was originally stated in RFC2462, that
if we invoke Host Configuration Behaviour we basically should not
invoke Information Configuration Behaviour.

As you pointed out, I don't think this restriction is always
reasonable.  But when we wrote revision 01, we basically tried to
follow the original sense of RFC2462, and we could not find a better
behavior.  Hence the current statement.

But if we (the wg) can agree on being more flexible, I can think of a
better approach.  For example, if a host which sets M-Policy to 1
cannot get any Advertisement to Solicit messages for some period, it
may make sense to fall back to Information Configuration Behaviour as
you indicated.

One possible problem of this approach is that RFC3315 requires the
client to send Solicit forever, until it gets an Advertise or Reply,
by specifying MRC and MRD being 0 (Section 17.1.2 of RFC3315).  So, at
least protocol-wise, the fall-back behavior does not (or may not)
conform to the RFC.

But, if we can also agree on modifying RFC3315 and RFC3736 a bit, I
can even think of a bit smarter approach:

  - DHCPv6 servers that only support Information Configuration
    Behaviour should also support the Advertise message and the IA_xxx
    options, but it always return an empty Advertise message to
    Solicit with a status code being NoAddrsAvail.  (This is a
    modification to RFC3736)
  - If a DHCPv6 client that supports both Host/Information
    Configuration Behaviours only receives such an Advertise message,
    it MAY fall-back to Information Configuration Behaviour.  (This is
    a modification to Section 17.2.1 of RFC3315, which requires the
    client ignore such Advertise).

Although the first modification requires additional implementation
complexity to DHCPv6 servers that only conform to RFC3736, the
resulting behavior is still "stateless", and in that sense, should be
lightweight.  So I think it's acceptable.

Anyway, there is no doubt that we need further discussions on this
point.  I'm open to other ideas.

> - minor issue wrt inconsistent M/O information:

>     If the host frequently receives inconsistent M and O flags of Router
>     Advertisement (e.g., in a mobile environment for supporting fast
>     movement detection), it may need complex consideration on an
>     erroneous case.  However, this case is not closely related to this
>     document; rather, it is a general issue on the inconsistent Router
>     Advertisement parameters from multiple routers.  In fact, other
>     configuration parameters such as the MTU size and the hop limit are
>     also possible to be inconsistent in different Router Advertisements.

> ==> actually M/O is IMHO a bit worse than these two, because these are 
> just internal variables, typically implemented in the kernel, and 
> toggling back and forth between them is not a problem.

> Presumably, initiating DHCPv6 requires running a non-trivial user-land 
> process.  Hence, I'd dare to say that the transitions are a bigger 
> problem with M/O flags, as has already been discussed a bit in 
> security considerations.

> Thus I'd like to see some further text, e.g. 1-2 sentences, describing 
> why M/O transitions are a bigger problem than other inconsistent 
> information.

I think I said this before, but anyway, personally I'm not really sure
if it's a good idea to mention such an implementation-specific thing
in a document describing protocol behaviors.  However, I don't oppose
to that idea now that this is in an Appendix.

					JINMEI, Tatuya
					Communication Platform Lab.
					Corporate R&D Center, Toshiba Corp.
					jinmei@xxxxxxxxxxxxxxxxxxxxx

--------------------------------------------------------------------
IETF IPv6 working group mailing list
ipv6@xxxxxxxx
Administrative Requests: https://www1.ietf.org/mailman/listinfo/ipv6
--------------------------------------------------------------------