Re: Proposal to further clarify prefix length issues in I-D.ietf-6man-rfc4291bis

[神明達哉 <[e-mail]>]

At Thu, 09 Mar 2017 19:30:11 +0100,
Philip Homburg <[e-mail]> wrote:

> If your use of the term 'subnet prefix' only refers to (stateless) address
> configuration and not to being onlink or not, then I agree about the
> concept (but not the term).
>
> However, where IPv6 clearly separates address configuration from being onlink,
> calling a prefix that is only used for (stateless) address configuration
> 'a subnet prefix' is very confusing.
>
> At the same time, calling some onlink prefixes 'subnet prefixes' and others
> 'other unicast routing prefixes' is also very confusing if they are
> announced using the same L flag in a PIO option in a router advertisement.
> I.e, whether or not the A flag is set in PIO option, should not have an effect
> on what we call an onlink prefix. Or on how we describe onlink prefixes and
> what their properties are.
[...]
> So if we want to say that prefixes used for stateless address configuration
> MUST/SHOULD be 64 bits long, whereas onlink prefixes can be of any length,
> then we should at least use clear terminology that separates prefixes for
> stateless address configuration from other uses.

+1.  I've also been feeling that one source of seeming confusion in
this discussion may be confusion between "subnet prefixes" and
prefixes for on-link determination as described in RFC4861.  In fact,
the latter (I'll use the term "onlink prefix" in this message) has
nothing to do with "subnet prefixes" - it simply specifies a set of
addresses that should be supposed to be on-link for the corresponding
link.  Whether or not we adopt "conservative" or "liberal" view of the
length of IIDs or "subnet prefix" in the context of addressing
architecture, an onlink prefix can be longer or shorter than a
subnet prefix even if these two share some leading bits.

So, for example:
- if we adopt the "conservative" address architecture,
  The subnet prefix for 2001:db8:1:2:3:4:5:6 must be 2001:db8:1:2::/64.
- if we adopt the "liberal" address architecture,
  The subnet prefix for 2001:db8:1:2:3:4:5:6 can be 2001::/16,
  2001:db8::/32, 2001:db8:1::/48, 2001:db8:1:2::/64,
  ...2001:db8:1:2:3:4:5:0/120, etc.
and, in either case,
- onlink prefx in this link advertised via an RA with L=1 can be
  2001::/16, 2001:db8::/32, 2001:db8:1::/48, ...2001:db8:1:2:3:4:5:0/120, etc.

In particular, the fact that onlink prefix can be of any length
doesn't mean it violates the conservative address architecture.

Perhaps one source of confusion is the fact or expectation that RA PIO
often both sets A and L bits on.  So, in the above example, if we have
a PIO for 2001:db8:1::/48 or 2001:db8:1:2:3:4:5:0/120 with both L and
A bits on, these are valid on link prefix.  But they can't be used for
SLAAC (assuming its link-layer spec specifies the consistent IID length
with the "conservative" address architecture, i.e., 64).  That is, for
the same PIO,
- if we consider it as on-link prefix, it can coexist with any of the
  address architecture
- if we consider it for SLAAC, it can only coexist with the
  conservative architecture, or liberal one with an exception of
  requiring 64-bit subnet prefix length specifically for SLAAC

BTW while writing this message I saw a warning from Brian:

>> Isn't it time we all shut up and wait for the document editor to
>> propose text?

I don't want to contribute to making this thread even longer, so I
promise I'll stop here.  But I wanted to make this one remark since
I believe it's important to clarify that ("on-link prefix" is
irrelevant to the address architecture discussion) and I see why this
can be so confusing.  Even if this attempt of mine fails, I'll shut up
now (maybe I should write a separate draft just focusing on this
particular confusion).

--
JINMEI, Tatuya