RE: 6MAN WG Last Call:draft-ietf-6man-ipv6-subnet-model-00.txt

Tatuya,

Thanks for all your email. Please see in line between <hs> and </hs>.

-----Original Message-----
From: Jinmei_Tatuya@isc.org [mailto:Jinmei_Tatuya@isc.org] 
Sent: Thursday, July 17, 2008 9:44 PM
To: Wes Beebee (wbeebee)
Cc: Hemant Singh (shemant); Suresh Krishnan; Thomas Narten; Brian
Haberman; Bob Hinden; ipv6@ietf.org
Subject: Re: 6MAN WG Last Call:draft-ietf-6man-ipv6-subnet-model-00.txt

At Thu, 10 Jul 2008 16:21:35 -0400,
"Wes Beebee (wbeebee)" <wbeebee@cisco.com> wrote:

> The problem is that one problem is FAR more likely to happen than the
other.
> 
> I shutdown my machine every night and power it on again in the morning

> when I come to work.  Therefore, every night of every workday I 
> experience the type of outage described in our draft.
> Furthermore, I occasionally go on vacations too - so the outage may
Tatuya
> last more than a day.
> 
> What this means for an administrator is that he has to predict, in 
> advance, how long I may be on vacation so that the RA deprecating the 
> old prefix can last long enough.  That puts an unreasonable 
> expectation on the network administrator.  Furthermore, I don't want 
> to have to get permission from my network administrator in order to go

> on vacation.

You're using inappropriate examples to justify the proposed text:

<hs>
Agreed. However, Wes did make a point that these are problems FAR more
likely to happen and also gave examples of timeframes that are much
longer than the reload time for a host. Like weeks of vacation or 6-8
hours every night. 
</hs>

   Using cached on-link determination information without first
   verifying that the information is still valid after IPv6 interface
   re-initialization may lead to lack of IPv6 network connectivity.  For
   example, a host receives an RA from a router with on-link prefix A.
   The host reboots.  During the reboot, the router sends out prefix A
   with on-link bit set and a zero lifetime to indicate a renumbering.
   The host misses the renumbering.  The host comes online.  Then, the
   router sends an RA with no PIO.  The host uses cached on-link prefix
   A and issues NS's instead of sending traffic to a default router.
   The "Observed Incorrect Implementation Behavior" section below
   describes how this can result in lack of IPv6 connectivity.

This reads to me that the outage is a pretty short time (i.e., while the
host is rebooting), while assuming the administrator stops advertising
the 0-lifetime RAs so quickly.  That's why I said "what's wrong in this
scenario is that the router doesn't keep advertising 0-lifetime-prefixes
sufficiently long".

Even if in the vacation case, the administrator shouldn't stop
advertising 0-lifetime RAs as long as some hosts may keep an old address
or prefix.  Note that they don't have to predict anything about the
users' vacation plan to do so: the necessary period can be calculated
from the previously advertised lifetime and the time when the
renumbering procedure starts.

Having said that, I see your point if we are mainly considering a case
of reconnecting after a long term absence such as a vacation.  Even
though the administrator should keep advertising 0-lifetime RAs to avoid
confusion, it should also be advisable for the host to purge the old
information (or at least confirm whether it's still valid).

<hs>
We just don't want to completely rely on an admin of the router network
to do the right thing and not say, fat-finger any configuration. Soon as
any admin mis-configuration is done, we can get into the problem of
Section 3 and data forwarding stops. The host should be defensive as
well. The goal of our draft it to cover any case that can cause the
problem of Section 3. We see blind caching of on-link determination as a
candidate that may cause the problem.
</hs>

So, for example, if the proposed text were something like this:

   Using cached on-link determination information without first
   verifying that the information is still valid after IPv6 interface
   re-initialization may lead to lack of IPv6 network connectivity.  For
   example, consider the case where a host caches an on-link prefix
   and leaves the subnet for weeks.  If the network renumbers during
   the period but the host continues to keep the cached (already
   invalid) information when it returns, it may lead to a problem
   described in Section 3 below.

that would make sense to me.  I'd then be wondering whether this is
really something to be noted explicitly since it may sound something
pretty obvious.

<hs>
The renumbering may sound obvious to you after we have discussed the
case of overnight shutdown and vacation and gave a specific example of
prior and post RA (with PIO and without PIO). Anyway, this is the
example in the paragraph. The main point is that blind caching of
on-link determination is not recommended host behavior. We would like to
keep this text  - your proposed text looks fine to me. We may modify
your proposed text by giving a specific prior and post RA example.

Further, since RFC4862 and RFC4861 differ in the 2-hour rule for address
autoconfiguration vs. on-link determination, why is it not OK to only
mention on-link determination blind caching in our draft and not mention
any address with blind caching? I am not too strong on this one. If we
are allowed to keep our text and you still insist we add address to
blind caching paragraph, I can do that. But note, even when we add
address to the blind caching paragraph, there is no normative text in
our paragraph that overrules anything related to caching addresses in
RFC4862. So why do any host implementors have to worry about anything?
Wes is on vacation till Wednesday next week. My statements here only
represent me. I need to talk to Erik and Wes too about this one. 
</hs>

Hemant

---
JINMEI, Tatuya
Internet Systems Consortium, Inc.