Re: [bess] Benjamin Kaduk's Discuss on draft-ietf-bess-nsh-bgp-control-plane-13: (with DISCUSS and COMMENT)

Benjamin Kaduk <> Sun, 21 June 2020 21:08 UTC

Return-Path: <>
Received: from localhost (localhost []) by (Postfix) with ESMTP id 409BD3A0886; Sun, 21 Jun 2020 14:08:16 -0700 (PDT)
X-Virus-Scanned: amavisd-new at
X-Spam-Flag: NO
X-Spam-Score: -1.897
X-Spam-Status: No, score=-1.897 tagged_above=-999 required=5 tests=[BAYES_00=-1.9, RCVD_IN_MSPIKE_H4=0.001, RCVD_IN_MSPIKE_WL=0.001, SPF_HELO_NONE=0.001, SPF_PASS=-0.001, URIBL_BLOCKED=0.001] autolearn=ham autolearn_force=no
Received: from ([]) by localhost ( []) (amavisd-new, port 10024) with ESMTP id t6B9wJW4cQd4; Sun, 21 Jun 2020 14:08:13 -0700 (PDT)
Received: from ( []) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by (Postfix) with ESMTPS id 3CB563A0885; Sun, 21 Jun 2020 14:08:12 -0700 (PDT)
Received: from ([]) (authenticated bits=56) (User authenticated as kaduk@ATHENA.MIT.EDU) by (8.14.7/8.12.4) with ESMTP id 05LL87Ul016839 (version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-GCM-SHA384 bits=256 verify=NOT); Sun, 21 Jun 2020 17:08:10 -0400
Date: Sun, 21 Jun 2020 14:08:07 -0700
From: Benjamin Kaduk <>
To: Adrian Farrel <>
Cc:,,, "'The IESG'" <>
Message-ID: <>
References: <> <0a1e01d635ab$8def5380$a9cdfa80$>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <0a1e01d635ab$8def5380$a9cdfa80$>
User-Agent: Mutt/1.12.1 (2019-06-15)
Archived-At: <>
Subject: Re: [bess] Benjamin Kaduk's Discuss on draft-ietf-bess-nsh-bgp-control-plane-13: (with DISCUSS and COMMENT)
X-Mailman-Version: 2.1.29
Precedence: list
List-Id: BGP-Enabled ServiceS working group discussion list <>
List-Unsubscribe: <>, <>
List-Archive: <>
List-Post: <>
List-Help: <>
List-Subscribe: <>, <>
X-List-Received-Date: Sun, 21 Jun 2020 21:08:16 -0000

Hi Adrian,

My apologies for the tardiness of the response.

Also my apologies for the silly questions contained in the following; I
seem to have lost a lot of the understanding of this technology that I
gained while performing the initial review.

On Fri, May 29, 2020 at 12:23:17PM +0100, Adrian Farrel wrote:
> Hi Ben,
> Many thanks for the detailed review.
> Herewith responses to your Discuss and Comments.
> > ------------------------------------------------------------------
> >
> > Section seems to talk about intermingling SFIR-RDs and
> > SFIR Pool Identifiers in a common list, but I do not see how it's
> > defined to intermingle the (six-octet) Pool Identifier Value with
> > the (eight-octet) RDs.
> That's because we include the entire extended community not just the value field.  The first two octets indicate whether the extended community is an RD or SFIR Pool Identifier. Each entry has to be read and processed so there is no question of needing to know the length to step over it.

Thank you for reassuring me that this is not an issue.  However, I'm still
not sure which words I should be looking at to know that this ("we include
the entire extended community") is the case.  I don't see much discussion
of what the contents of the SFIR-RD List is at all, just a note that each
entry in the list is eight octets long, and that we can use "SFIR-RDs from
the advertisements of the SFIs in the SFIRs", and that the list can
"contai[n] one or more SFIR Pool Identifiers", and that an SFIR-RD "of
value zero" (64 bits set to zero?) has special semantics.

Even if I do somehow make the leap that I need the full 8-octet SFIR Pool
Identifier Extended community for use in this list, I'm still confused
about whether there is potential ambiguity.  This extended community starts
with type/subtype octets of TBD6 and TBD7, allocated from the BGP
Transitive Extended Community Type and (new) SFC Extended Community
Sub-Types Registry.  These pool identifiers are comingled in the SFIR-RD
List with the 8-octet SFIR-RD values, which are defined as "the RD used to
advertise an SFI".  If these are to be RDs in the sense of RFC 4364, then
they start with a two-byte type field, allocated from the "Route
Distinguisher Type Field" registry created by RFC 4364.  (Currently only
types 0, 1, 2, and 65535 have assigned semantics.)  Do we need to then
allocate the Route Distinguisher Type Field value that corresponds to the
two octet sequence TBD6 TBD7 in order to guarantee that the SFIR-RDs in
this list can be distinguished from SFIR Pool Identifiers?  (Is our usage
of "all bits zero" in conflict with some other legitimate usage of RDs?)

Furthermore (and quite removed from the original Discuss point), the -14
added examples using IPv6, and seems to be saying that the RD value is used
to indicate (e.g.) 2001:db8::192:0:2:1/1.  But the RD is only 8 octets --
how can the full 127 bits of IPv6 prefix be represented in 8 octets?

> > Section 3.1 seems to allow multiple SFIRs associated with a given
> > RD, but the rest of the document seems to assume that any RD
> > has at most one associated SFIR (as is stated explicitly for SFPRs).
> > (A few specific mentions in the COMMENT.)
> Right. There is NOT a 1:1 mapping between an SFIR and an RD. But we have some legacy text that is not clear on that point and we are scrubbing it.

Okay, happy to see that happen.

> Note that Section 3.1 describes two cases, advertising a single SFIR to represent more than one SFI of a given SFT and advertising an SFIR for each SFI of a given SFT.
> > Within my own limited understanding, it seems like this document is
> > expanding the boundaries of the SFC Architecture in ways not envisioned
> > by RFC 7665 or 8300, and the shepherd writeup is pretty quiet on to what
> > extent this architectural change is accepted by the WG (as opposed to
> > being contained to just the BGP control plane mechanism).  I'd like to
> > get a positive affirmation from someone more familiar with the
> > discussions that this is moving the architecture in the right direction
> > with respect to things like:
> > - the introduction of the Service Function Type intermediate
> >   classification
> > - the more prominent treatment of looping, jumping, and branching as
> >  operations within a single SFP without reclassification by using the
> >  "Change Sequence" SFT entries to indicate these behaviors within the
> >  SFP definition itself (I note that RFC 7665 does not mention "jump" at all)
> > - the introduction of "gaps" in the SI sequence of a given SFP
> Note that "classification" and "reclassification" are slightly fudged (my opinion) in the 8300 architecture. They may be performed by a separate component, but that is a functional component that can be built into any other component of the system. Thus, an SF might reasonably have a reclassifier that makes decisions to continue forwarding a packet or to send it to a penalty box. All of the functions of looping, jumping, and branching (the penalty box example is a branch) are done through reclassification.

Okay.  (To be honest, the first and third points seemed more significant to
me, and I wasn't as sure on the reclassification stuff.)

> We took particular care to socialise this work with the SFC working group presenting it to the WG at IETF meetings. The BESS working group last call was also copied to the SFC list. We did receive some comments that helped improve the work, but no complaints about infringing the architecture.
> I guess the SFC chairs can confirm this if you're still worried.

It would be good to see this, but I won't hold the discuss point for it.

> > With respect to SFT in particular, it sounds like this is intended to
> > help with scalability, which makes the genart reviewer's comment about
> > lack of implementation experience particularly poingant.  It seems like
> > SFIR pools perform a similar role, though of course not identical; I
> > didn't get a clear sense of why pools without SFTs are insufficient.
> SFT allows for an SFF to not need to advertise more than one instance of an SF of that type, which is a scalability thing. But, more importantly, it allows a controller to not need to specify a specific SFI at an SFF just saying "any SFI of this type" making the work of the controller a bit easier, and leaving local load balancing choices to the SFF.
> The pool is a more diffused tool. All SIs still need to be advertised so that they can be included in the pool. The pool tells two SFFs to make a choice:
> - an SFF may have to choose between multiple local SFIs in the same pool
> - an SFF may have to choose between multiple next hop SFFs to satisfy the next hop pool
> Pools also reduce the re-advertisement of SFPRs which would otherwise occur when the SFIR-RD list changes (so that is a scalability thing, I suppose).
> Note that it is not a requirement that all SFIs in a pool are of the same type. For example, there might be FooCorp's firewall and BarCorp's firewall with different SFT values but in the same pool.

Thanks for pointing that last bit out (including in the examples in the
document!); I don't think it was fully clear to me originally.

> See sections 1.2, 2.2, and 3.1 for additional info.
> > -----------------------------------------------------------------
> >
> > I support Roman's Discuss.
> I copied you on the response to that with proposed new text for section 9.
> > Section 2.1
> >
> >   In fact, each SI is mapped to one or more SFs that are implemented by
> >   one or more Service Function Instances (SFIs) that support those
> >   specified SFs.  Thus an SI may represent a choice of SFIs of one or
> >   more Service Function Types.  [...]
> >
> > In my head a SI that maps to more than one SF implies a requirement to
> > do all of the listed SFs and in a specific order; "a choice of SFIs of
> > one or more types" seems to relax that a bit so that the order is
> > relaxed and perhaps the requirement to do all is also relaxed.  I think
> > "choice of SFIs for one or more" would alleviate that concern (if it is
> > even a real concern).
> Thanks for the glimpse into your head 😊
> An SI is an instruction to execute *a* service function. There may be a choice of SF, but only one is executed per SI. Several SIs in a row may cause execution of SFs at the same SFF.I think the current wording holds.
> This is all RFC 8300 stuff.
> There is more detail in section 5 of this document.

Sounds like my head was wrong, then.  Thanks for the explanation.

> >   in the underlay network.  How this indicator is generated and
> >   supplied, and how an SFF generates a new entropy indicator when it
> >   forwards a packet to the next SFF are out of scope of this document.
> >
> > nit: comma after "next SFF".
> OK
> > Section 2.2
> >
> >   This approach assumes that there is an underlay network that provides
> >   connectivity between SFFs and Controllers, and that the SFFs are
> >   grouped to form one or more service function overlay networks through
> >   which SFPs are built.  We assume BGP connectivity between the
> >   Controllers and all SFFs within each service function overlay
> >   network.
> >
> > Are we thus implicitly not assuming BGP connectivity between arbitrary
> > pairs of SFFs?
> Depends on what you mean by "connectivity". I normally expect that word to mean a BGP peering: that is not needed. But there does need to be a chain of connectivity so BGP information can be passed along. The use of a Route Reflector is a key tool in this.

I believe I was also thinking "direct peering", so we seem to be in
agreement.  It's not clear whether there's much value in explicitly saying
anything more about this in the text.

> > Section 3
> >
> >   The nexthop field of the MP_REACH_NLRI attribute of the SFC NLRI MUST
> >   be set to loopback address of the advertising SFF.
> >
> > nit: "a loopback address" (?)
> ack
> > Section 3.1
> >
> >   Per [RFC4364] the RD field comprises a two byte Type field and a six
> >   byte Value field.  If two SFIRs are originated from different
> >   administrative domains, they MUST have different RDs.  In particular,
> >   SFIRs from different VPNs (for different service function overlay
> >   networks) MUST have different RDs, and those RDs MUST be different
> >   from any non-VPN SFIRs.
> >
> > I think we should probably be a little more explicit that we reuse the
> > RD element (and the RD type/value semantics!) wholesale from RFC 4364
> > and
> > That is, we are not just reusing the layout, but the logical type as
> > well.
> It's not reuse, it is use. This is what an RD is. We are using it as designed.

Ah, I think I see now.  (FWIW, this helps me be more concrete about how I
am confused, regarding the Discuss section...)

> > Also, are all of these "MUST"s indicating requirements that are new with
> > this specification, or are some of them inherited from RFC 4364?
> They are new with this specification. Hence the mention of "SFIR" in both cases.

Okay, thanks.

> > The discussion later on in Section 4.2 suggests that perhaps these RDs
> > need to be unique per SFI, akin to how we say they are unique per SFP in
> > Section 3.2?
> Per your Discuss, 4.2 contains some legacy stuff wrt the 1:1 mapping. 3.1 was accurate.
> The RD indicates the application (e.g., the overlay).
> >   The Service Function Type identifies the functions/features of
> >   service function can offer, e.g., classifier, firewall, load
> >   balancer, etc.  There may be several SFIs that can perform a given
> >   Service Function.  Each node hosting an SFI MUST originate an SFIR
> >
> > This seems a bit of a non-sequitur: we go straight from SF*T*s to
> > discussion of the SFI/SF relationship, and we only mention "type" (but
> > not "SFT") in the rest of the paragraph, so the connection is harder to
> > return to than it need be.
> I think we can clean up type/SFT.


> >   Note that a node may advertise all SFIs of one SFT in one shot using
> >   normal BGP Update packing.  That is, all of the SFIRs in an Update
> >   share a common Tunnel Encapsulation and RT attribute.  See also
> >
> > I think we need to expand "route target".
> > Also, I suggest s/may advertise all/may advertise all its/
> Ack
> > Section 3.1.2
> >
> > I suggest explicitly stating that these MPLS labels are used to direct
> > incoming traffic to the SFI in question (fulfilling the promise made by
> > the abstract).
> >
> >   Note that it is assumed that each SFF has one or more globally unique
> >   SFC Context Labels and that the context label space and the SPI
> >   address space are disjoint.
> >
> > What I'm taking away from the last clause is roughly "we do not assume
> > that the values of SFC context labels have any structured relationship
> > to the corresponding SPI values", even though a strict reading of
> > "disjoint" would seem to be that "if a value is used in one setting it
> > is guaranteed to not be used in the other setting".  I think some
> > rewording may be in order.
> The "strict" meaning was the intended meaning. We have added "(i.e., a label value cannot be used both to indicate an SFC context and an SPI, and it can be determined from knowledge of the label spaces whether a label indicates an SFC context or an SPI)"

Ah, thank you; that's quite clear.

> > Section 3.2
> >
> > [our treatment of RD should presumably be consistent with Section 3.1,
> > if we make any changes there]
> >
> >   byte Value field.  All SFPs MUST be associated with different RDs.
> >   The association of an SFP with an RD is determined by provisioning.
> >
> > So this means that any given RD value can have at most one SFP
> > associated with it?  I guess that's needed in order to provide the level
> > of separation between SFPs that we need...
> The RD provides uniqueness between SFPRs
> > Section 3.2.1
> >
> > Is there intended to be a distinction between the error-handling
> > behavior for error cases 1/2/6/7 and error case 4?  (If there is, I'm
> > not seeing it.)
> Yes. Can't recall the history of how this ended up like this, but you're right and we have merged the list entries.
> >      5., 8.:  The absence of an RD with which to correlate is nothing
> >         more than a soft error.  The receiver SHOULD store the
> >         information from the SFP attribute until a corresponding
> >         advertisement is received.
> >
> > Should any sort of timeout also be applied?
> This was discussed and I was over-ruled 😊
> The theory is that the configuration of an SFC system might be taken over a long period, and nodes might be down for a while. A timer, it was felt, was too rigid.
> The resultant "SHOULD" allows (of course) an implementation to not store information until the corresponding advertisement is received if (for example) it finds itself memory-constrained.

Ah, thanks for the explanation (and sorry that you got over-ruled :)

> > Section
> >
> >      SFP.  An SFP attribute SHOULD NOT contain more than one
> >      Association TLV with Association Type 1: if more than one is
> >      present, the first one MUST be processed and subsequent instances
> >      MUST be ignored.  Note that documents that define new Association
> >
> > Do we want to say anything about checking consistency that (per these
> > rules) the forward SFP is associated with the backward, and the backward
> > SFP is associated with the forward one?  [I guess Section 7.1 is
> > probably enough.]
> We think 7.1 is enough. There is a forward pointer at the end of
> > Section
> >
> > [It's not clear to me that grouping by type gives any efficiency savings
> > here, as grouping by pool is available, and the semantics seem to be
> > equivalent to if the list of allowed SFI (pool)s was explicitly given
> > without the container.  A single pool containing all the SFIs in the
> > type would replace the SFIR-RD-value-zero semantics.]
> Grouping by type gives scalability wrt the number of SFIRs - it reduces their number.
> Pool ID gives scalability by reducing the frequency of SFPR - i.e., a new instance of an SFPR doesn't need to be advertised in response to the change in the SFIR population.  It comes w/ a configuration cost which RD-value-zero does not have. 
> But while we're here. should describe sub-TLVs not TLVs.
> > Section
> >
> >   The SFP Traversal With MPLS Label Stack TLV (Type value 5) is a zero
> >   length sub-TLV that can be carried in the SFP Attribute and indicates
> >
> > If it's carried directly in the SFP Attribute, doesn't that make it a
> > non-sub TLV?
> Yes.
> Rats' nest of when is a TLV not a TLV?
> > Section 3.2.2
> >
> >   In all cases, there is no way for the receiving SFF to know which
> >   SFPR to process, and the SFPRs could be received in any order.  At
> >   any point in time, when multiple SFPRs have the same SPI but
> >   different SFPR-RDs, the SFF MUST use the SFPR with the numerically
> >   lowest SFPR-RD.  The SFF SHOULD log this occurrence to assist with
> >
> > (The "interpreted as an 8-octet integer in big-endian form" is
> > implicit?)
> I thought we'd all moved on from this and had a general understanding of how to eat our eggs.
> But it never hurts to remind, so added (although "network byte ordering" is the term du jour)
> > Section 4.1
> >
> >   The main feature introduced by this document is the ability to create
> >   multiple service function overlay networks through the use of Route
> >   Targets (RTs) [RFC4364].
> >
> > As mentioned above, this seems like the second (not first) usage of
> > "RT".
> Fixed as above, but left the expansion/abbreviation here because of the title of the section.
> > Section 4.5
> >
> >   When an SFF gets an SFPR advertisement it will first determine
> >   whether to import the route by examining the RT.  If the SFPR is
> >   imported the SFF then determines whether it is on the SFP by looking
> >   for its own SFIR-RDs in the SFPR.  For each occurrence in the SFP,
> >
> > Also for non-specific SFT mentions that would apply, right?
> Yes, should add "and RD zero"
> > Section 5
> >
> >   o  There may be an obvious branch needed in an SFP such as the
> >      processing after a firewall where admitted packets continue along
> >      the SFP, but suspect packets are diverted to a "penalty box".  In
> >      this case, the next hop in the SFP will be indicated with two
> >      different SFT values.
> >
> > My (admittedly poor) understanding of the SFC Architecture was that this
> > was best done as a reclassification, and that the various SFI options
> > for a given SI were intended to be roughly equivalent.  Am I just
> > imagining that?
> Yes, reclassification is the word. Now the choice here is:
> - program the reclassifier
> - include the information in the SFP
> The former essentially means that you have three SFP fragments. We are describing the second of these to make an SFP that branches.
> >   2.  From the SFP attribute of that SFPR, it finds the Hop TLV with SI
> >       value set to SI-y.  If there is no such Hop TLV, then such
> >       packets cannot be forwarded.
> >
> > This seems to exclude the "gaps" in SI space that this document is
> > trying to allow.  Or is the gap processing to have already happened by
> > the time the SFF is at the "needs to forward" stage?
> The sequence of SI values may have gaps, but the SFP contains those explicit values and must be matched.
> The next value is found in the Hop TLV.


> >       A.  An SFIR is relevant if it carries RT-z, the SFT in its NLRI
> >           matches the SFT value in one of the SFT TLVs, and the RD
> >           value in its NLRI matches an entry in the list of SFIR-RDs in
> >           that SFT TLV.
> >
> >       B.  If an entry in the SFIR-RD list of an SFT TLV contains the
> >           value zero, then an SFIR is relevant if it carries RT-z and
> >           the SFT in its NLRI matches the SFT value in that SFT TLV.
> >           I.e., any SFIR in the service function overlay network
> >           defined by RT-z and with the correct SFT is relevant.
> >
> > Do we need to talk about expansion of pools as well?
> Thanks. Good catch.
> >   Thus, at any point in time when an SFF selects its next hop, it
> >   chooses from the intersection of the set of next hop RDs contained in
> >   the SFPR and the RDs contained in its local set of SFIRs.  If the
> >   intersection is null, the SFPR is unusable.  Similarly, when this
> >   condition obtains the originator of the SFPR SHOULD either withdraw
> >   the SFPR or re-advertise it with a new set of RDs for the affected
> >   hop.
> >
> > (This seems to require a definition of "contained in" that includes
> > various types of expansion having been performed.)
> OK
> > nit: s/obtains/is detected/ (??)
> Damn our fancy use of language ;o)
> > Section 7.4
> >
> >   o  If an SFC Classifiers Extended Community is received with SFT = 0
> >      then there are two sub-cases:
> >
> >      *  If there is a choice of SFT in the hop indicated by the value
> >         of the SI (including SI = 0) then SFT = 0 means there is a free
> >         choice according to local policy of which SFT to use).
> >
> >      *  If there is no choice of SFT in the hop indicated by the value
> >         of SI, then SFT = 0 means that the value of the SFT at that hop
> >         as indicated in the SFPR for the indicated SPI MUST be used.
> >
> > side note(?): I think it would be possible to include the second case as
> > a special case of the first, as a "free choice" among all one allowed
> > values by the SFP for that SI.
> True, but clearer to spell it out.
> >   which the traffic belongs.  Additionally, note that to put the
> >   indicated SPI into context when multiple SFC overlays are present in
> >   one network, each FlowSpec update MUST be tagged with the route
> >   target of the overlay or VPN network for which it is intended.
> >
> > I'd suggest making this requirement more prominent by putting it earlier
> > in the section.
> Yes

(It looks like now we duplicate the normative MUST, but I'm not going to
make a big deal about it.)

> > Section 7.5
> >
> > The description of these flags as describing "how the originating SFF
> > expects to see the SPI/SI represented" has some connotation of
> > exclusivity ("I expect to see an NSH" implying "I don't expect to see
> > anything else, including MPLS"), which does not seem like a good fit for
> > a bitmask where multiple bits can be set (as opposed to just a codepoint
> > for the encoding to use).
> >
> >   The following bits are defined by this document:
> >
> >   Bit 0:  If this bit is set the NSH is to be used to carry the SPI/SI
> >      in the data plane.
> >
> > How are we numbering the bits?
> Thanks. Yes, from most significant as zero.
> > Section 7.6
> >
> >   When a classifier constructs an MPLS label stack for an SFP it starts
> >   with that SFP' last hop.  If the last hop requires an {SPI, SI} label
> >
> > nit: "SFP's"
> Ack
> > Section 8.1
> >
> >   SFF1 followed by an SF of type 43 located at SFF2.  This path is
> >   fully explicit and each SFF is offered no choice in forwarding packet
> >   along the path.
> >
> > nit: "forwarding packets" or "forwarding a packet"
> Ack
> > Section 8.4
> >
> >   This example provides a choice of SF type in the second hop in the
> >   path.  The SI of 250 indicates a choice between SF type 43 located
> >   through SF2 and SF type 44 located at SF3.
> >
> > nit: s/located through/located at/?
> Ack
> >   When the packets are returned to SFF1 by the SFI the SI will be
> >   decreased to 250 for the next hop.  SFF1 now has a free choice of
> >   next hop SFF to execute the next hop in the path selecting between
> >   all SFF2 that support an SF of type 43 and SFF3 that supports an SF
> >   of type 44.  These may be completely different functions that are to
> >
> > nit: I think s/all SFF2 that support/SFF2 that supports/ ?
> Actually, s/SFF2/SFFs/
> > Section 8.8
> >
> > Some pedantic part of me wants to complain that this is not a "branch",
> > since switching to SPF10 is the only option allowed at SI=250 on SPF11.
> We welcome your inner pedant. Please name it and give it an email address.
> It's an optional branch.
> > Section 8.9.1
> >
> > It might make more sense to mvoe SPF13's definition before the paragraph
> > that talks about "sufficiently precisely specified SFPs", since SPF13
> > does not actually "specify [...] exact SFIs to use"
> Done
> > Section 9
> >
> > I really like the way this section is written; thank you!
> Although not well enough to prevent you supporting Roman's Discuss 😉

Liking the layout and style doesn't mean that there's not more content to add :)

> > I think the core SFC documents cover enough of the considerations
> > inherent to looping, jumping, and branching that there's probably not
> > more to say here.
> >
> > I'll list a handful of potential security topics I can think of after
> > reading the document, but that's not meant to imply that they all need
> > to be covered individually in the document: they are *potential* topics,
> > but may not end up being worth talking about.
> [snip]
> We factored a lot of these good suggestions into the reworking of Section 9 sent out to Roman and cc'ed to you.
> > Section 10.4
> >
> > [When I first saw "association TLV" I was expecting it to share a
> > registry with some other association grouping usages.  But I don't have
> > a particular complaint about doing it this way.]
> So no action on that one.
> Thanks once again,

And thanks to you as well for the updates.