Appeal against IESG decision

[Robert Elz <[e-mail]>]

This is an appeal to the IAB against the IESG decision to reject
my appeal against their earlier decision to approve the publication
of draft-ietf-ipngwg-addr-arch-v3-11.txt as a Draft Standard.

The issues here are very simple, and no lengthy examination of mailing
list archives, taking of evidence, hearing opinions, ... should be
necessary in this case.   I believe that none of the facts are in any
kind of dispute.

Those facts are

1) RFC2026 says, in section 4.1.2 ...

   A specification from which at least two independent and interoperable
   implementations from different code bases have been developed, and
   for which sufficient successful operational experience has been
   obtained, may be elevated to the "Draft Standard" level. [...]

   The requirement for at least two independent and interoperable
   implementations applies to all of the options and features of the
   specification.  In cases in which one or more options or features
   have not been demonstrated in at least two interoperable
   implementations, the specification may advance to the Draft Standard
   level only if those options or features are removed.

2) draft-ietf-ipngwg-addr-arch-v3-11.txt contains at least one (and perhaps
two) features for which there are not two interoperable implementations.

The one is:

	For all unicast addresses, except those that start with binary 
	value 000, Interface IDs are required to be 64 bits long and 
	to be constructed in Modified EUI-64 format.

There's no dispute that there are no interoperable implementations of
this - there are no implementations of it at all (or no documented ones
anyway).

Note that the spec actually gives no option here, other than the exceptions
(the 000 addresses, and multicast), interface IDs are required to be 64
bits long.    While all implementations I'm aware of allow 64 bit IDs,
none have been presented that require it.  The draft *requires* it.

Any reasonable reading of 2026 would require that that feature of the
specification be removed from the draft before the draft is permitted to
be published as a draft standard.   Of course, as an alternative, the WG
or IESG could have the draft, as it is, published as a Proposed Standard,
and await the necessary two implementations of the feature before requesting
advancement.

The IESG's opinion of this seems to be that the "two implementations of
every feature" applies only where they consider it important enough to
bother checking.   I have no problems with drafts advancing when no-one
brings to the attention of the IESG that there is a problem in this area.
But when a problem is pointed out, the clear words of 2026 really must be
enforced.

The rationale for this requirement in 2026 is simple (as the IESG should
know, as the author of 2026 is a member of the IESG).   First, it ensures
that the text in the document is clear enough that it can be implemented
in an interoperable way.   And second, it helps make sure that the
document doesn't get cluttered with requirements in practice no-one
bothers to implement - that is, that the document is a proper specification,
and anyone reading the document can implement from it, with the
expectation that their implementation will interoperate with others.

The quoted text from the draft fails both of those tests.   We have no
implementations so we don't know that the text is clear enough to be
implemented correctly.   It may seem obvious that the text is clear to
any reader - but the IETF has always ignored "seem obvious" and required
actual implementation experience as a demonstration.

Second, an implementation which did faithfully follow the words of the
draft would fail to interoperate correctly with every other known
implementation of it.   It may be claimed that it is the other
implementations, or the way they are configured, that is at fault here,
but that's not relevant - the aim is to get interoperability, and if
we have operators configuring /112, /226, /227 and similar prefix
lengths (that is, interface ID's that are 16, 2, or 1, and other,
numbers of bits long) - and we do - then an implementation that enforced
the 64 bit IID requirement (allowed only /64 prefix on an interface)
would fail to interoperate with other implementations (with all other
existing implementations).

This seems to be a "placeholder" fluff feature, being maintained to
perhaps allow some future design to allow applications to simply "know"
what is the prefix, and what is the interface-ID.   The requirement for
existing implementations in 2026 is a specific requirement that such
fluff be removed from docs before they're allowed to advance to DS status.

The extra requirement should be removed from the document, and then, if
the WG so desires, published as a PS (or Experimental) RFC of its own.
If it then becomes accepted and implemented, it could be merged back with
the main document in a later revision.


The second issue in the appeal to the IESG concerned the 'u' bit, which
is one of the bits of the IID as defined.

The IESG referred to this as ...

 B/ Robert says "The requirement that where the 'u' bit (the inverted L 
 bit from the MAC address) is set, the IID is globally unique."

and eventually concluded ...

 The IESG notes that there is no wording in
 draft-ietf-ipngwg-addr-arch-v3-11.txt requiring that IIDs be globally
 unique.

and then quoted two passages from the draft, only the last part of
which is relevant.

   In
   the resulting Modified EUI-64 format the "u" bit is set to one (1) to
   indicate global scope, and it is set to zero (0) to indicate local
   scope.

It is true that the doc does not expressly say "globally unique", what
it says is "indicate global scope".

The draft also says ...

   Modified EUI-64 format based Interface identifiers may have global
   scope when derived from a global token (e.g., IEEE 802 48-bit MAC or
   IEEE EUI-64 identifiers [EUI64]) or may have local scope [...]

And ...

   The use of the universal/local bit in the Modified EUI-64 format
   identifier is to allow development of future technology that can take
   advantage of interface identifiers with global scope.

I doubt I'm the only reader to come to the conclusion that "have global
scope" actually means "be globally unique".   What's more, a review of
various IPv6 related mailing lists will show this opinion expressed over
and over again.   Clearly there are many readers who have leapt to this
wrong conclusion, if it is in fact wrong, and as the IESG have concluded,
there is no actual expectation that the IID will be any more than probably
globally unique, then the draft should probably be more explicit in saying
that, rather than allowing readers to leap to the incorrect interpretation.

But there is more to this than the IESG apparently understood.   The question
isn't just whether the 'u' bit being set implies that the IID is globally
unique (or has "global scope") but whether there are any implementations at
all that actually enforce the setting of the 'u' bit only when the IID
has been formed from a (probably) globally unique token (a MAC address
or similar).   Here there has been one implementation reported on the
mailing list (but not in the interoperability report) - but for DS status,
one implementation isn't enough.

Other implementations allow the user to configure the 'u' bit set without
having any knowledge or expectation, or reason to assume, that the address
is derived from any kind of globally unique token (or token with global scope).
What's more, they do this for good reason, as without that ability, users have
no way to remove a NIC, and replace it with another, and retain the original
(auto-configured) IPv6 address (which being based upon the MAC address that
the old NIC provided, would have had the 'u' bit set).   In this case the
address is in fact based upon a globally unique token, but the implementation
has no way to know that, and so must also allow the 'u' bit to be set when
the rest of the IID is 0, or 1, or 2, which are most certainly not tokens
with any kind of global scope.    What's more, as the old NIC may be now in
use in some other host, there's no reason in the example cited to assume that
there's any uniqueness at all - for correct IPv6 operation, the NIC can't
be connected to the same subnet as where it used to be, but beyond that
IPv6 works just fine with the same IID on different nets).

Once again, we have a feature of the specification which is either not
implemented, or at best, is not clear.

The IESG's response to all of this is ...

 When considering this appeal, it is clear from the interoperability
 reports that there are implementations that generate the interface ID
 from the EUI-64 identifier, which makes it be 64 bits long. It is
 also clear that the uniqueness properties of EUI-64 based identifiers
 will be the same as the EUI-64 identifiers from which they are derived
 (which is slightly weaker than a requirement for global uniqueness).

Yes - though in practice implementations (bar one) allow addresses to be
generated from EUI-64's, but they also allow indistinguishable addresses
to be manually configured - so in practice (which is what the interoperable
implementation requirement is attempting to ensure that the specification
conforms to) extracting an IID from an IPv6 address, and expecting it to
have any kind of similar properties of global uniqueness as an EUI-64
would be a false expectation, and the draft should not lead people into
expecting otherwise.   Only if implementations actually enforced this
requirement (which they can easily do, as the one which has done it shows,
though of course, this loses functionality) would this expectation be
justified.

 So for at least some implementations, they are capable of acting as
 specified in the document being challenged.

No.   The requirement challenged is a "must only be" - or if you like,
a MUST NOT.   The fact that it is possible to conform with the spec
using existing implementations has nothing to do with the issue at all.

The IESG's response here would be the equivalent of responding to a
requirement that "All cars must be red" by pointing to a few red cars
and saying "see, it can be done".   That it can be done isn't the issue,
the issue is that the specification says it MUST be done.   To be advanced
to DS, all that is required is that there be 2 conforming implementations,
what the rest do is irrelevant (until the doc is ready to be advanced to full
standard).   For some specifications, 2 implementations itself is a large
hurdle, for IPv6, it isn't, there are many implementations.   That none
of them have implemented one of the requirements of the doc, and only one
another, should be a pretty obvious red flag that these requirements should
not remain in the document.

The IESG also says ...

 We traditionally require that things interoperate when configured 
 correctly, not that they interoperate when configured incorrectly, or 
 that it be impossible to configure them incorrectly.

Of course, that's as it should be.   That is, except where the specification
explicitly says that something must not be possible.   There's no point
keeping a prohibition in the specification if no-one takes any notice of
it.   There's no difference here to keeping some other feature that no-one
bothers to implement.

And again from the IESG ...

 Implementation reports are used to verify that independent 
 implementations can succesfully interoperate. This is a quality check 
 on the clarity of the documents.

Yes.   But the IESG have managed to conveniently forget the other purpose
for the requirement - that is, as a check that the features are actually
being implemented, and that the document isn't describing things which in
practice everyone ignores.

But even without that, here we have no quality check on the clarity of
the relevant statements in the documents - no-one has implemented them.
(No-one has implemented one, there's only one implementation of the other).
We're only guessing if we assume that the statements are clear enough.

IESG again:

 Requiring explicit verification on all statements would be a change to 
 existing practice and one that would likely increase the difficulty in 
 advancing documents on the standards track.

That's what was intended.   If existing practice has been to ignore
the two implementation requirement, when it is known not to be met,
then I submit that the IESG has been operating contrary to the clear
instructions of 2026.

It need not be onerous to enforce this however - it is entirely reasonable
to expect the community to point out any flaws in the implementation
reports as published.   If there are no reported problems, the IESG,
and the community, are justified in assuming that the reports fully
document the required interoperability of every feature.   Where there
are reports that interoperability of some feature has not been properly
documented, then it should be easy for the implementation report to be
corrected, if the feature has in fact been implemented and tested.  If it
has been implemented, but not tested, then the report should be seen as
being of benefit, in showing a potential trouble area, not as a burden.
If testing shows that all works, then there's real harm done, and the
implementation report can be corrected.   If testing shows 
non-interoperability, then clearly there's something that needs fixing
(in some cases, just implementation bugs, after which further testing
will show the specification is fine).   On the other hand if testing is
not possible, because the feature has not been implemented, then it really
should be removed from the document before it is advanced.   That's what
the requirement in 2026 is there for.

The IESG again ...

 There are many places in IETF standards where a field is stated to be
 a specific length or a value to be within a range. Requiring that the
 limits be enforced in software for all of these cases would put a
 significant extra burden on the implementers and the documenters of
 the implementations for questionable benefit.

This is once again based upon a misunderstanding of what is required.
No-one is requiring that the limits be enforced in general.   What is
being asked is that someone (sometwo really) has done it.   What's the
point of a requirement that is universally ignored?   There is none,
it is misleading, and should not be permitted in a Draft Standard.

I would ask that the IAB instruct the IESG to overturn their decision
to publish the draft (draft-ietf-ipngwg-addr-arch-v3-11.txt) as a
Draft Standard, and at their choice, either publish it as a Proposed
Standard, or return it to the working group for amendments that will
allow it to be published as a Draft Standard.

kre

ps: I would also request that the RFC editor continue to defer
publication of this draft until the IAB has dealt with this appeal.