Re: [Anima] WGLC on draft-ietf-anima-stable-connectivity-03 - Respond by July 28, 2017

Toerless Eckert <> Thu, 03 August 2017 01:08 UTC

Return-Path: <>
Received: from localhost (localhost []) by (Postfix) with ESMTP id 96A0B128BC8 for <>; Wed, 2 Aug 2017 18:08:18 -0700 (PDT)
X-Virus-Scanned: amavisd-new at
X-Spam-Flag: NO
X-Spam-Score: -4.2
X-Spam-Status: No, score=-4.2 tagged_above=-999 required=5 tests=[BAYES_00=-1.9, HEADER_FROM_DIFFERENT_DOMAINS=0.001, RCVD_IN_DNSWL_MED=-2.3, RP_MATCHES_RCVD=-0.001] autolearn=ham autolearn_force=no
Received: from ([]) by localhost ( []) (amavisd-new, port 10024) with ESMTP id dqkbNWZj2JED for <>; Wed, 2 Aug 2017 18:08:15 -0700 (PDT)
Received: from ( []) (using TLSv1.2 with cipher AECDH-AES256-SHA (256/256 bits)) (No client certificate requested) by (Postfix) with ESMTPS id CE3CD131B79 for <>; Wed, 2 Aug 2017 18:08:14 -0700 (PDT)
Received: from ( []) by (Postfix) with ESMTP id 1B7BF58C4BC; Thu, 3 Aug 2017 03:08:10 +0200 (CEST)
Received: by (Postfix, from userid 10463) id EBEABB0C792; Thu, 3 Aug 2017 03:08:09 +0200 (CEST)
Date: Thu, 3 Aug 2017 03:08:09 +0200
From: Toerless Eckert <>
To: Brian E Carpenter <>
Cc: Anima WG <>
Message-ID: <>
References: <> <>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <>
User-Agent: Mutt/1.5.21 (2010-09-15)
Archived-At: <>
Subject: Re: [Anima] WGLC on draft-ietf-anima-stable-connectivity-03 - Respond by July 28, 2017
X-Mailman-Version: 2.1.22
Precedence: list
List-Id: Autonomic Networking Integrated Model and Approach <>
List-Unsubscribe: <>, <>
List-Archive: <>
List-Post: <>
List-Help: <>
List-Subscribe: <>, <>
X-List-Received-Date: Thu, 03 Aug 2017 01:08:19 -0000

Thanks a lot for the thorogh review Brian. I hope your was the last 
mayor content change rev i have to do for the stable connectivity.

Turned out that when i went through your comments, the mayority of changes
i had to do was in the ACP draft because that was underspecified (ACP connect)
and in result there was a bunch of loose suggestive stuff in stable-connectivity.
In the end i could delete hpefully all underspecified stuff from stable-connectivity
and this is now in ACP draft -09 well specified.

Mohamed changes where into -04, but there was no conflict with what you
wrote, so yours is now -05, changes:

Everything else inline below (responding to your points).

Btw: If there are any additional points, i highly recommend to open issues in github:

That github is for both the ACP doc and stable connectivity. Maybe not the best choice to combine them, but it is what it is.


> Hi,
> Here are my comments on this draft. In general it seems to be ready,
> but there are some issues that IMHO need fixing. Here they are,
> followed by a few nits that I noticed.
> Technical issues:
> -----------------
> > 2.1.3.  Simultaneous ACP and data plane connectivity
> ...>    If the data-plane of the network is also supporting IPv6, then the
> >    NOC devices that need access to the ACP should have a dual-homing
> >    IPv6 setup.  One option is to make the NOC devices multi-homed with
> >    one logical or physical IPv6 interface connecting to the data-plane,
> >    and another into the ACP.
> I don't understand the need to call this "dual-homing". That generally
> implies a physical topology with multiple links and/or routers. I think
> all you mean is that the nodes happen to have at least two addresses
> in different IPv6 prefixes (one for the data plane and one for the
> ACP). Having multiple addresses is a standard feature of IPv6. It might
> be done with multiple (virtual or real) interfaces, but it doesn't need
> to be. So I suggest:
>   If the data-plane of the network also supports IPv6, then the
>   NOC devices that need access to the ACP should have both data-plane
>   and ACP IPv6 addresses. One option is to set up the NOC devices with
>   one logical or physical IPv6 interface connecting to the data-plane,
>   and another into the ACP.

The ACP is meant to be an isolated network, so the fundamental idea is that
whenever possible, all adddresses routed across the ACP are non-overlapping
with Data Plane addresses so that its clear what traffic is for ACP and what

The first thing which for security reasons should be out of scope is a
non-ACP router that connects Data-Plane and ACP  and could therefore
arbitrarily leak ACP traffic:

                     ACP Connect
            ACP      ---------  Wrong
            Edge                Router ---- NMS host
            Router   ---------
                     Data plane

There is of course only limited abilities to prohibit this, but it is explicitly
not desired.

When i started to answer to your questions below, there was really no way to
fix up the stable connectivity draft because it could only hint or suggest things
that where not written correctly in the ACP draft. And thats wrong. So i went back
and effectively had to fix up a lot of things in the ACP draft. So see ACP-09
changelong in ACP-09 first, then go back to the following inline answers:

> >    The LAN that provides access to the ACP
> >    should then be given an IPv6 prefix that shares a common prefix with
> >    the IPv6 ULA...
> 1) It is surely a virtual interface, not a LAN.

Well, it can be physical or virtual. Replaced LAN with subnet. Thats the correct term.
The ACP-09 text is now also more specific to physical vs. virtual so those
are two clear separate options re. their securitty impact.

> 2) I think it is better to say
>  "should then be given an IPv6 prefix that lies within the ULA prefix..."

Done. I now also use the therm "covered by" in the acp doc. I hope
that term is also correct.

> (In fact, it doesn't matter that it's a ULA - what matters is that
> it's the prefix that covers the whole ACP. That really goes for the whole
> document; a ULA prefix is just another prefix, after all. It might be
> clearer to just say "ACP prefix" everywhere.)

Right. In the main text. the whole security section is specific to ULA address benefits (isolation),
so i didn't change it on behalf of this point.

> >    ... so that the standard IPv6
> >    interface selection rules on the NOC host
> Cite [RFC6724] for complete clarity.


> >    If this can not be achieved
> >    automatically, then it needs to be done via simple IPv6 static routes
> >    in the NOC host.
> But it can. That's why RFC6724 exists. Do we really need to say this?

As outlined above, the explanation for this all is now in the ACP connect section
of ACP-09. Primarily because it didn't make sense to describe a lot of wonderful
functional aspects of the ACP edge device in an informational document when they
are missing in the normative standards track document.

But there is a single sentence summarizing the issue in the stable connectivity doc now
as well:

| This setup may not work
| with autoconfiguration and all NOC host network stacks due to limitations in
| those network stacks. They need to be able to perform RFC6724 source address selection
| rule 5.5 including caching of next-hop information. See the ACP document text for more details.

> >    Providing two virtual (eg: dot1q subnet) connections into NOC hosts
> >    may be seen as undesired complexity. 
> Either you have to explain 'dot1q' and give a reference, or delete it.
> I'd delete it.

I always annoyed by unspecific terms like virtual, but it clicks for
me when i read practical descriptions like 802.1q. So there is a reference to
it now.

> >    In that case the routing policy
> >    to provide access to both ACP and data-plane via IPv6 needs to happen
> >    in the NOC network itself: The NMS host gets a single attachment
> >    interface but still with the same two IPv6 addresses as in before -
> >    one for use towards the ACP, one towards the data-plane.  The first-
> >    hop router connecting to the NMS host would then have separate
> >    interfaces: one towards the data-plane, one towards the ACP.  Routing
> >    of traffic from NMS hosts would then have to be based on the source
> >    IPv6 address of the host: Traffic from the address designated for ACP
> >    use would get routed towards the ACP, traffic from the designated
> >    data-plane address towards the data-plane.
> That seems like an explanation of very basic routing - does it need to
> be said?

This is all gone from stable connectivity now and hopefully better structured text
in the ACP connect section in acp-09.

> > In the most simple case, we get the following topology:...
> This part isn't very clear to follow. The ASCII art of Fig. 1
> and Fig. 2 is a bit scrappy, and the explanation is a bit short.
> It would be better to start with an explanation of the terms
> like 'NOClan' and 'Rtr1'. And you suddenly start discssing VRFs
> without any definition.

Likewise gone, replaced by better picture in acp-09.

> > 2.1.4.  IPv4 only NMS hosts
> ...
> >    The downside of this architectural decision is the potential need for
> >    short-term workarounds when the operational practices in a network
> >    that can not meet these target expectations.  This section motivates
> >    when and why these workarounds may be necessary and describes them.
> >    All the workarounds described in this section are HIGHLY UNDESIRABLE.
> >    The only long term solution is to enable IPv6 on NMS hosts.
> I full agree with the message but I think the wording has the wrong tone.
> The goal should be to welcome NOCs to the new world of IPv6, not to
> tell them they are dinosaurs. Something like:
>    The implication of this architectural decision is the potential need for
>    short-term workarounds when the operational practices in a network
>    do not yet meet these target expectations.  This section explains
>    when and why these workarounds may be operationally necessary and
>    describes them. However, the long term goal is to upgrade all
>    NMS hosts to native IPv6, so the workarounds described in this
>    section should not be considered permanent.

Love text suggestions. Copied.

> >    To bridge an IPv4 only management plane with the ACP, IPv4 to IPv6
> >    NAT can be used.  This NAT setup could for example be done in Rt1r1
> >    in above picture to also support IPv4 only NMS hots connected to
> >    NOClan.
> I think this (and the following paragraph) is underspecified. It isn't
> clear to me whether this would be described as a NAT64 or a NAT46 scenario
> (which side is the client and which side is the server, in other words).
> There's a lot more to specify to make this work - maybe the details are
> out of scope, which should be stated if so. Also, it would be better
> to say 'IP/ICMP translation' not 'NAT' and cite RFC7915. Make that
> clearly separate from the issue of how addresses are mapped.
> (Incidentally, recall that (a) IPv4 addresses are only 32 bits and
> (b) we own the ACP address plan. So algorithmic mapping of IPv4
> addresses into a special /96 in the ACP address plan is possible.
> Of course that would be an insecure zone in ACP terms.)

*sigh*. Yes. As i tried to explain to Mohamed, i was specifically not trying to
get sucked into individual NAT method description. But i gave up now with your
feedback as well and improved the NAT text to refer to SIIT and explain
the requirements better, eg: v6 initiator -> v4 responder requirement and
v4 initiator -> v4 responder requirement. And then suggesting to use EAM as
the most flexible mapping option.

Hope this improves the IPv4 section a lot.

> > 2.1.5.  Path selection policies
> A lot of this section comes across almost as hand-waving. I did
> wonder whether it would have been enough just to state the
> problem.

Well, the primary problem is that first round implementations of ACP may not
be high performance, so the use of ACP should be prioritized for critical actions
and otherwise Data Plane be used. I though the text says that.

Then the text goes into mechanisms how NOC equipment would select how to decide
whether to use ACP and/or data plane for some action and it suggests that without
changning NOC equipment software, it might be possible to achieve this by using
different names for the devices, eg: names that map to ACP or to data plane or first
to data-plane then to ACP addresses. And then use the right name in a particular
piece of NOC equipment or action on a NOC equipment to get the desired use
of ACP and/or data plane.

If the above explanation makes sense to you but you do not see this best reflected,
maybe you can suggest better text to explain this.

> ...
> >    MP-TCP (Multipath TCP -see [RFC6824]) is a very attractive candidate
> >    to automate the use of both data-plane and ACP...
> I'm not sure... you seem to be asking for new intelligence in
> MPTCP's choice of candidate addresses, not just a policy but
> a whole mechanism in support of that policy. And will you really
> benefit from MPTCP's main point, which is automatic load management
> between alternative paths. Wouldn't SCTP be a better match?

This was just one example for an alternative mechanism to the use of different names,
eg: embody the path selection policy into the transport protocol. Of course
the way how exactly to do this is hand waiving, and MP-TCP was just the one most
well known protocol that also most likely can be stuffed under existing
TCP applications without them knowing.

I would of course like that this is not seen as handwaiving but as
interesting areas of further work/research.

> > 2.1.8.  Long term direction of the solution
> ...>    1.  NMS hosts should at least support IPv6.  IPv4/IPv6 NAT in the
> >        network to enable use of ACP is long term undesirable.  Having
> >        IPv4 only applications automatically leverage IPv6 connectivity
> >        via host-stack options is likely non-feasible (NOTE: this has
> >        still to be vetted more).
> That NOTE needs to be cleared up. Something like 464XLAT (RFC6877)
> might be a good compromise.

See the rewritten SIIT section. IMHO, there can be no simpler "network" based
address translation. Where network based means that the translation happens
in some device he network operator needs to provision. Like the ACP edge device.
Or even an additional address translation device.

So, the only IMHO easier option is when the OS of the NMS host would internally
have IPv4/IPv6 translation so the device/VM looks to the outside like full IPv6.
Alas, i didn't have the time to investigate these options. And most likely if at
all you could only make those work for linux.

So, for now i just remove the note and clarified the last sentence a bit.

If there is anything specific to be said bout why 464XLAT might be better
longer term, let me know and i can add it. For now it looks like yet another
network device configured option to me, but i have not tried to understand it
all the way.

> > 3.  Security Considerations
> ...
> >    ULA addressing as proposed in this document is preferred over...
> Was this pasted from the ACP draft? Surely that is where ULA is proposed.

Probably long time ago. This paragraph was already improved in -04 by Mohameds

> >    Randomn ULA addressing provides more than sufficient protection
> >    against address collision even though there is no central assignment
> >    authority.
> I don't like that phrase: it isn't the *address* that's random,
> it's the /48 prefix. Also, somebody is always ready to complain
> about the collision risk. Suggest:
>    The random nature of a ULA prefix provides strong protection
>    against address collision even though there is no central assignment
>    authority.

Thanks, fixed.

> >    If packets with unexpected ULA addresses are seen and one expects
> >    them to be from another networks ACP from which they leaked, then
> >    some form of ULA prefix registrastion (not allocation) can be
> >    beneficial.  Some voluntary registries exist, for example
> >, although none of them is
> >    preferrable because of being operated by some recognized authority.
> >    If an operator would want to make its ULA prefix known, it might need
> >    to register it with multiple existing registries.
> > 
> >    ULA Centrally assigned ULA addresses (ULA-C) was an attempt to
> >    introduce centralized registration of randomly assigned addresses and
> >    potentially even carve out a different ULA prefix for such addresses.
> >    This proposal is currently not proceeding, and it is questionable
> >    whether the stable connectivity use case provides sufficient
> >    motivation to revive this effort.
> I think all that is a red herring. I suggest replacing it by a simple
> statement:
>    If packets with unexpected ULA addresses are seen they should
>    be discarded.
> (Even that is not really needed, since all border routers should
> discard such packets anyway.)

I remember we had long discussions about this so i kinda want to refrain from
removing the text. Instead:

The filtering you mentioned should vbe in ACP spec but where not. So i added
into acp-09:

  - mandate that ACP edge devices must filter based on source IPv6 address.
  - Mandating/explain that RPL roots can diagnose unknown destination IPv6
    addresses in the ACP.
    (Aka: if your filtering did mean the destinationa address, then this is NOT
     possible on the border router due to RPL. Can't have minimum routing table
     and full route prefix visibility at the same time ;-(

So, with that text there, i changed the stable connectivity paragraphs above to

<t>The ACP specification demands that only packets from configured ACP prefixes
are permitted from ACP connect interfaces. It also requires that RPL root
ACP devices need to be able to diagnose unknown ACP destination addresses.<t>

<t>To help diagnose packets that unexpectedly leaked for example from another
ACP (that was meant to be deployed separately), it can be useful to voluntarily
list your own the ULA ACP prefixes in one of the sites on the Internet,
for example Note that this does not
constitute registration and if you want to ensure that your leaked ACP packets
can be recognized to come from you, you may need to list your prefixes in
multiple of those sites.</t>

<t>Note that there is a provision in <xref target="RFC4193"/> for non-locally 
assigned address space (L bit = 0), but there is no existing
standardization for this, so these ULA prefixes must not be used.</t>

> >    Using current registration options implies that there will not be
> >    reverse DNS mapping for ACP addresses.
> Really? I assume we're talking about two-faced DNS, and afaik nothing
> stops an operator providing reverse mapping in the private DNS.
> That seems to be implied by the following paragraphs, so the text
> seems inconsistent anyway.

I know it under the name "split-horizon DNS". Is there any reference ?

I changed the two paragraphs to read as follows, i hope this eliminates any
real or perceived conflicts:

<t>According to RFC4193 section 4.4, PTR records for ULA addresses should not
be installed into the global DNS (no guaranteed ownership). Hence also the 
need to rely on voluntary lists (and in prior paragraph) to make the use of
an ULA prefix globally known.</t>

<t>Nevertheless, some legacy OAM applications running across the ACP may rely
on reverse DNS lookup for authentication of requests (eg: TFTP for download
of network firmware/config/software). Operators may therefore use split horizon
DNS to provide global PTR records for their own ULA prefixes only into their
own domain to continue relying on this method. Given the security of the ACP,
this may even increase the security of such legacy methods.</t>

> > 4.  No IPv4 for ACP
> This section seems out of place stuck between Security Considerations
> and IANA Considerations. Maybe it should be an Appendix, referenced
> in section 2.1.4. "IPv4 only NMS hosts".

Michael Richardson made the argument that nobody reads beyond the list of
authors, and i kinda agree, so i am trying to avoid appendices. Like in the
ACP document i have now moved this section before the standard Security/IANA
considerations and put it under the label "Architectural considerations". Which
kinda should show its closeness to all the other following "Considerations".

Hope this is fine. If not, yell.

> Nits/English:
> -------------
> (There are probably other nits as well as these, which the RFC Editor
> will fix...)

Right. Lets agree on "code complete", i will do more spell checker later as well.

> > Abstract
> ...
> >    ...  Provisioning during device/network bring
> >    up tends to be far less easy to automate than service provisioning
> >    later on, changes in core network functions impacting reachability
> >    can not be automated either because of ongoing connectivity
> >    requirements for the OAM equipment itself, and widely used OAM
> >    protocols are not secure enough to be carried across the network
> >    without security concerns.
> Suggested:
>   Provisioning while bringing up devices and networks
>   tends to be more difficult to automate than service provisioning
>   later on, changes in core network functions impacting reachability
>   cannot be automated because of ongoing connectivity
>   requirements for the OAM equipment itself, and widely used OAM
>   protocols are not secure enough to be carried across the network
>   without security concerns.

Thanks. Copied.

> >    This document describes how to integrate OAM processes with the
> >    autonomic control plane (ACP) in Autonomic Networks (AN). to provide
> >    stable and secure connectivity for those OAM processes.
> Suggested:
>   This document describes how to integrate OAM processes with the
>   autonomic control plane (ACP) in Autonomic Networks (AN) in order to provide
>   stable and secure connectivity for those OAM processes.

Thanks. Copied.

> > 1.2.  Data Communication Networks (DCNs)
> > 
> >    In the late 1990'th and early 2000, IP networks became the method of
> >    choice to build separate OAM networks for the communications
> >    infrastructure in service providers.  This concept was standardized
> >    in G.7712/Y.1703 
> This needs a complete informational reference. Also, according to
> it is a superseded
> reference, so maybe there is a better one?

Reference now points to the oldest version of the series because that is the
purpose of the text (since when is this done). Annotation in reference points to
the home page of the series -

> > 2.1.1.  Simple connectivity for non-autonomic NMS hosts
> ...
> >  For example, if DNS in the network was set up with
> >    names for network devices as, then the ACP
> >    address of that device could be mapped to devicename-
> >
> ...


> > 2.1.2.  Challenges and limitation of simple connectivity
> ...>    Note that these challenges and limitations exist because the ACP is
> >    primarily designed to support distributed ASA ...
> Define "ASA" when first used.


> Regards,
>     Brian