[v6ops] Re: Ketan Talaulikar's Discuss on draft-ietf-v6ops-framework-md-ipv6only-underlay-27: (with DISCUSS and COMMENT)
Chongfeng Xie <chongfeng.xie@foxmail.com> Fri, 09 October 2026 08:16 UTC
Received: from out162-62-58-216.mail.qq.com (out162-62-58-216.mail.qq.com [162.62.58.216]) (using TLSv1.2 with cipher ECDHE-ECDSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by mx.ietf.org (Postfix) with ESMTPS id 6CEAD46; Fri, 09 Oct 2026 08:16:51 +0000 (UTC)
Authentication-Results: mx.ietf.org; dkim=pass header.d=foxmail.com header.s=s201512 header.b="Rq6X4/m1"; dmarc=pass (policy=none) header.from=foxmail.com; spf=pass (mx.ietf.org: domain of chongfeng.xie@foxmail.com designates 162.62.58.216 as permitted sender) smtp.mailfrom=chongfeng.xie@foxmail.com
DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=foxmail.com; s=s201512; t=1791533801; bh=6YmDXSmDj0KRs6PMZHEZ4ZieplRR+aI+2miqjlmjnBY=; h=Date:From:To:Cc:Subject:References; b=Rq6X4/m1sr8uox1m9CFiX6mWZTzxvFM6fOddMNNW6HMJ6VH83OoP6Wi09QuTpGV55 L95JRSQHsiT10CIbcEt9nabr7xZ1tfDFmLFgo2lIdD+n8/RC1IjIClr8zDKn5o0WRo tDVN3UCwVHxjKElXgiRpr7d22nPXYSF2BXdm6YK8=
Received: from LAPTOP-BOBOCIFS ([1.203.112.206]) by newxmesmtplogicsvrszc50-0.qq.com (NewEsmtp) with SMTP id 42687E40; Fri, 09 Oct 2026 16:16:38 +0800
X-QQ-mid: xmsmtpt1791533798t7924vkak
Message-ID: <tencent_70B06B3240268E38B3FB58834AE5428F7C07@qq.com>
X-QQ-XMAILINFO: Nte9/BsRzcszEuzoqDQkLnzuMOfd18L0F5DhT0OAsqoYes+CaUD1ethSVbbdih pRRQbnYIr9sRFQ+bplT2ChUIO4YngmgcVzL0v39PJWwqfRJOcXkXw0Yuea+/qNp9pXchtvy8jPfO O4LMKxryKw9JAVNVWL0A7xY16rFBgJNLCmxwxBhn2IDgP3khuXE/56eOuiLt8JancYIDpb7R1v4N vyyLFPWRTCuFAVo6t3zarXaybDgqEgKWgPnSuUuk5511l7Jhdpx3AvX296dY7eF8cJwGiSr+4RVF udxVsxhj1UT3/eSYJg62Wcw+makZVdWQ3GlrJKBNZlEBkQFF9hSMVfgbDmgpzfeqs/m8kUbJRFAr LOgARRuFNHIoMDl0JkAwWX6pN86qcuyByytaFSuiuChwnH6061Q+WYwHNqyJDDU4jAEcn6Hl4/U1 0EwK5GCeaIrX1j+EDuaJm4QUoNvC26NrM0U6k/SBavliMUproKuNr4AvPGLnSTUqNAi1aPBlgXfu zGD+GbXgxOLs52eSwjthuTCdXLQwBeBVOgc+UYHqAP4J2nus48z67ReZQcIbyJu5VuqRoCNto/q9 j0mgn8gdAq+SJahrkCACsxIo9rNxdOzhWeJuRB5nXb+OMcEGlL04zT5vH7xghDiERfTYz9Psw95W DwaCRlh9drwUCWfJWg6DE9KYAFov3Q1INolpAF6n972JxkIfwRAOC2gq5AbtOGM1xSWOc6qsaUWu QtwLP8LWkTNyPrf+bjJxCmKyGykc3yhjc8vgircddGzbTTTRHa93GczfAcXIXbpsuBySQKIiu3Uu HbxIm9ez9I4SOkyjFEhcwafdqBDprsYjd8P3hPv/Odvg0YcQmWu6uoHZXcqR96tzJcK+h2vRQ7MQ 5NEc5eYcWLVJ/gSDxbzMRUsBvDye589+Tmqk4qkUCdbOaH0i1oZWN98VMJkfVPszEdlfbY92oLmw 9dpbfRfkbXhzHvmifUEQlOVsRJRm2jfiI7m5cHvpxL3xHSJyURebjLkSXgzKVynK4Au6L/zcLuvG Zjhi0vJ8B3I/PCCSjfhyj9J7cVjl4fXzxCWIEMby9iJY1f6pBXkKTGqFO1xeW3wNiaPs5LWuKVbk sCJf2KQgBRDir1/oiqDKlH2w6HEeT/SnoTQtXpelgrjBBV2Kfxdz/yF1ValEWUiNT9bMHg
X-QQ-XMRINFO: Nq+8W0+stu50tPAe92KXseR0ZZmBTk3gLg==
Date: Fri, 09 Oct 2026 16:16:38 +0800
From: Chongfeng Xie <chongfeng.xie@foxmail.com>
To: "ketant.ietf" <ketant.ietf@gmail.com>
References: <178714897973.773353.8850189455856515301@dt-datatracker-7c6ddbc678-86d5j>, <tencent_A738D6CCE062D52A430B11F2E9AEAF01A309@qq.com>, <CAH6gdPyUS5kR4UKAgN4x1fGY6L4wf=5Er1vGKvrQjTXy+Ry2Lg@mail.gmail.com>
X-Priority: 3
X-GUID: 76263260-54C7-4BEF-8A77-DE615A5ABF26
X-Has-Attach: no
X-Mailer: Foxmail 7.2.25.542[cn]
Mime-Version: 1.0
X-OQ-MSGID: <202610091616376671892@foxmail.com>
Content-Type: multipart/alternative; boundary="----=_001_NextPart413673571142_=----"
X-Spamd-Bar: --------------------
Message-ID-Hash: 3J4BCOYW7GU63WK45BRJSS3FOEJF3OWL
X-Message-ID-Hash: 3J4BCOYW7GU63WK45BRJSS3FOEJF3OWL
X-MailFrom: forwardingalgorithm@dmarc.ietf.org
X-Mailman-Rule-Misses: dmarc-mitigation; no-senders; approved; loop; banned-address; header-match-v6ops.ietf.org-0; emergency; member-moderation; nonmember-moderation; administrivia; implicit-dest; max-recipients; max-size; news-moderation; no-subject; digests; suspicious-header
CC: IESG <iesg@ietf.org>, draft-ietf-v6ops-framework-md-ipv6only-underlay <draft-ietf-v6ops-framework-md-ipv6only-underlay@ietf.org>, ron <ron@bonica.org>, v6ops-chairs <v6ops-chairs@ietf.org>, list <v6ops@ietf.org>
X-Mailman-Version: 3.3.10
Precedence: list
Subject: [v6ops] Re: Ketan Talaulikar's Discuss on draft-ietf-v6ops-framework-md-ipv6only-underlay-27: (with DISCUSS and COMMENT)
List-Id: v6ops discussion list <v6ops.ietf.org>
Archived-At: <https://mailarchive.ietf.org/arch/msg/v6ops/gJ2-GHsTbbaLntEFCAzSTDRNlbw>
List-Archive: <https://mailarchive.ietf.org/arch/browse/v6ops>
List-Help: <mailto:v6ops-request@ietf.org?subject=help>
List-Owner: <mailto:v6ops-owner@ietf.org>
List-Post: <mailto:v6ops@ietf.org>
List-Subscribe: <mailto:v6ops-join@ietf.org>
List-Unsubscribe: <mailto:v6ops-leave@ietf.org>
Hi Ketan,
We have just submitted -30 (dated 9 October 2026). With this new version, the following changes have been made,
-As promised in our previous reply, -30 adds RFC 4364 and RFC 4925 (DISCUSS 3)as informative references and rewrites the Section 1 paragraph to acknowledge the encapsulation-based prior art (RFC 4925, RFC 5565, RFC 4364) informationally .
-Editorial refinements, a few minor wording and formatting clean-ups.
Your other DISCUSS items from the review of -27 were addressed in -28 and -29:
We would be glad to make any further adjustments you may suggest, and we look forward to your review of -30.
Best regards,
Chongfeng
On behalf of all the co-authors
From: Ketan Talaulikar
Date: 2026-09-16 16:44
To: Chongfeng Xie
CC: IESG; draft-ietf-v6ops-framework-md-ipv6only-underlay; ron; v6ops-chairs; list
Subject: Re: [v6ops] Ketan Talaulikar's Discuss on draft-ietf-v6ops-framework-md-ipv6only-underlay-27: (with DISCUSS and COMMENT)
Hi Chongfeng,
Thanks for your response. Please see my follow-up comments and clarifications inline below. For the rest, we largely agree, and I will await the document update to perform a full review pass and update my ballot.
On Sat, Sep 12, 2026 at 1:46 PM Chongfeng Xie <chongfeng.xie@foxmail.com> wrote:
Hi Ketan,
Thank you for your thorough review and raising so many comments and suggestions, which
are very constructive. Before I give my reply to your comments one-by-one, I would like to
provide several major points about the considerations and future revisions, if there are any
inaccuracies, feel free to correct me.
• This document illustrates operators' considerations and requirements for deploying
IPv6-only networks, addressing the challenges faced in the transition to IPv6-only,
in particular, it raises the considerations about how to handle the relationship between
the IPv4 and IPv6 address spaces when evolving toward an IPv6-only from the perspective
of several operators. Its primary purpose is to provide guidance for IPv6-only deployment
in backbone networks. As a "Framework" document—it describes, at a high level, "what
needs to be done," rather than attempting to prescribe" which specific protocols should be used."
KT> Thanks for that clarification, and that makes sense to me. Please update the document accordingly.
• The scope is "networks formed by interconnecting multiple IPv6-only ASes," rather
than the entire Internet. We are considering adding a separate "Scope of Applicability"
subsection in the new version to explicitly clarify that this framework applies to a "
Limited and network environment," and to articulate the trust relationships among
participating parties.
KT> "networks formed by interconnecting multiple IPv6-only ASes" is ambiguous. The questions that come to mind are: (a) How are these networks connected? If connected over the Internet, this requires advertising the mapped IPv6 prefixes and their associated mapping rules over the Internet. (b) If connected via a private peering relationship between trusted providers, this requires advertising similar information only on such peering sessions. This makes it operationally closer to a VPN service between cooperating providers (somewhat like a CSC model). I guess the objective is (b), and if so, this needs to be clarified. If not, then it opens up many more questions about the resultant advertisements over the Internet, and you may wish to discuss this with the GROW community.
• Based the suggestion of you and Gunter, BGP-related content will be moved out of
this document, primarily from Sections 5.2 and 8.3. We will keep the v6ops framework
document focused on operational guidelines, while ensuring that routing complexities
are properly specified and reviewed by the routing area experts.
KT> Thank you. I would suggest moving the routing parts into the IDR document, provided the IDR WG agrees. It will fill the solution gap and get the necessary reviews from IDR.
• RFC 5565 addresses IPv4 over IPv6 by encapsulation, however it focuses on the scenarios
within a single AS. In contrast, the new draft focuses on a larger-scale internetworking
environment consisting of multiple ASes networks, with the goal of achieving "usability"
and "large-scale deployment" (cross-domain, efficient, and controllable IPv4-as-a-Service).
KT> That is a very narrow view of the RFC 5565 solution. I would draw your attention to https://www.rfc-editor.org/info/rfc5565/#section-12. I consider it a lego block that must be acknowledged in this work. The components built around this lego-block for the overall solution deployment can be covered by this document. My point was that the document is missing references to deployed prior standards.
• Technical clarification: RFC6052-based IP address translation is not limited to supporting
packet translation only, it can also support tunnel-based encapsulation for packets.
Nevertheless, following your suggestion, this document will demote encapsulation approach
and primarily illustrates the framework based on the packet translation mechanism.
In situations where encapsulation is necessary, this document only mentions encapsulation mode.
KT> Thanks for agreeing to remove encapsulation and focus on translation based solution. I disagree with your view that RFC6052 is not limited to translation - that view conflicts with the very title of the document and you may also refer to https://www.rfc-editor.org/rfc/rfc6052.html#section-1.1 about relationship with encapsulation.
Then please see my feedback inline [Chongfeng],
From: Ketan Talaulikar via Datatracker
Date: 2026-08-19 22:16
To: The IESG
CC: draft-ietf-v6ops-framework-md-ipv6only-underlay; ron; v6ops-chairs; v6ops
Subject: [v6ops] Ketan Talaulikar's Discuss on draft-ietf-v6ops-framework-md-ipv6only-underlay-27: (with DISCUSS and COMMENT)
Ketan Talaulikar has entered the following ballot position for
draft-ietf-v6ops-framework-md-ipv6only-underlay-27: Discuss
When responding, please keep the subject line intact and reply to all
email addresses included in the To and CC lines. (Feel free to cut this
introductory paragraph, however.)
Please refer to https://www.ietf.org/about/groups/iesg/statements/handling-ballot-positions/
for more information about how to handle DISCUSS and COMMENT positions.
The document, along with other ballot positions, can be found here:
https://datatracker.ietf.org/doc/draft-ietf-v6ops-framework-md-ipv6only-underlay/
----------------------------------------------------------------------
DISCUSS:
----------------------------------------------------------------------
Thanks to the authors and the WG for their work on this document.
As someone that supports IPv6 and especially IPv6-only underlay deployments, I
find myself conflicted with this document's form and proposals.
This is a routing-heavy document and it's unfortunate that this never
got an RTGDIR (early or any) review. I don't know if it was socialized with
GROW WG given its potential intersection with Internet routing (unless I've
got the document's applicability wrong).
This framework document coming from V6OPS depends on specifications that have
not yet matured or published. This gives the impression that operational
experience on it is pending and that strikes me as highly unusual (I echo
some of the sentiments of GENART reviewer) - is there any precedence for
V6OPS asking for publication of such kind of document? This is what gets
me to the first discussion point.
Please find below some discussion points and comments on this document, inline
in the idnits output of v27. Lookout for the <EoRv27> tag at the end to ensure
you are seeing the full review.
<discuss-1> Nature of this document and V6OPS charter fit
I would like to discuss what kind of document this is. It is not a
deployment description, since (I believe) nothing described has been
deployed and the (BGP or DNS-based) rule distribution mechanisms (I believe)
are neither mature nor implemented. It is not a requirements document, since
it states no requirements on the mechanism it depends on - which is what the
INTDIR early review asked for. It is not a problem statement of deployment
or operational issues. Neither is it something that is driving protocol
development work in V6OPS. It is a framework that leaves important things
unspecified and normatively requires work that is not yet mature or even
adopted. How does this fit into V6OPS charter? I don't consider milestone
as a charter fitment.
I looked for a precedent among V6OPS prior work and did not find one.
Please see if the document could be rescoped to the problem statement
(leaning upon deployment or operational challenges), the deployment
requirements, and the required properties of any conforming
rule-distribution mechanism. This fits within V6OPS charter and can help
organize work in other WGs - notably related BGP and routing work in
other WGs.
[Chongfeng]: This document presents a framework for building and operating
a large-scale IPv6-only network from the perspective of network operators.
In particular it proposes stateless address mapping as the basis for enabling
IPv4 service data transmission in a multi-domain IPv6-only environment
(i.e., IPv4-as-a-Service). This document serves a distinct and critical purpose
for the v6ops WG: it defines the operational requirements and architectural
constraints that evolution of operator networks towards IPv6-only. It identifies
the gaps in existing solutions and provides a high-level, technology-agnostic
blueprint. This is precisely the role of a 'Framework' document in V6OPS: to
guide the development of operational solutions and transition strategies rather
than to report on already-deployed ones. I think it is within the scope of v6ops.
KT> Yes, it is definitely within the V6OPS charter once its scope is clarified and content updated. Thanks.
<discuss-2> The applicability of the framework is never stated - whether it
is for multiple ASes of a single provider, a closed set of cooperating
providers, or an open capability reaching the global IPv4 Internet.
603 Although this document illustrates the framework for a multi-domain
604 IPv6-only network operated by multiple NPs, this framework is also
605 applicable to an IPv6-only network operated by a single NP.
This is the document's only applicability statement, and it widens the scope
downward without ever bounding it upward. What is the maximum extent of a
multi-domain IPv6-only underlay network? The Internet?
[Chongfeng]: Regarding the applicability, a multi-domain IPv6-only underlay network in
this document refers to a network composed of multiple interconnected IPv6-only ASes,
where the devices in the core run only the IPv6 protocol. If these ASes belong to a
single operator, the network is operated by that one operator; if they belong to
multiple operators, it is operated by multiple operators with mutual agreement.
It has limited scope and is not the Internet, and it is required not to influence
the running of the IPv4 Internet and IPv6 Internet, based on the
"IPv6-as-a-Service" requirements proposed in this document, A multi-domain
IPv6-only underlay network can meet its users' needs for accessing IPv4 Internet services.
I plan to add the text above into the new version.
The evolution of this form of network towards IPv6-only is currently the focus of IPv6
development. This draft mainly presents considerations and requirements, rather than
advancing specific technical specifications.
KT> Agree, and please see the detailed aspect of the deployment model and applicability - (a) or (b) in my comment above. I believe the intention is the model (b) based on your explanation, but please correct me if I am wrong.
The text supports opposite answers. Section 1 motivates the work by network
providers requiring access to the global IPv4 Internet, Section 8.1 has the
egress PE forwarding converted packets into the IPv4 Internet, and Section
5.2 requires the distribution procedure to scale across multiple
independently administered NPs. Against that, Section 6 requires mapping
prefix uniqueness across the whole domain, Section 7 requires the
participating carriers to negotiate management-level SLAs for ICMPv6
traversal, and Section 8.3 makes the default rule conditional on explicit
bilateral policy. Every one of the second set is a cooperating-parties
assumption; none of the first set is bounded.
The scale, routing and security analyses all differ between "several ASes
of one provider" and "an open capability reaching the global IPv4 Internet",
and several of my comments below cannot be weighed without the answer. It is
also the point on which this framework differs from 464XLAT, MAP-E, MAP-T and
DS-Lite, which Section 1 places it alongside: those are single-administration,
their mapping rules are provisioned rather than exchanged, and each has one
well-defined edge to the rest of the Internet. This framework has none of those
properties.
[Chongfeng]: As mentioned earlier, this framework targets backbone networks, covering both
a single operator's internal backbone and a large backbone for multi-operator
collaboration. We will clarify this scope explicitly in the document. The two are
complementary rather than substitutes of the same kind. We will also revise
Section 1 to distinguish our work from 464XLAT, MAP-E, MAP-T, and DS-Lite:
those are single-administration access technologies, while our framework
addresses multi-domain backbones where mapping rules are exchanged across
AS boundaries.
KT> The framing of solutions is not always consistent across technologies at the IETF. Many solutions do not detail routing aspects, such as within a single AS, multiple ASes within an administrative context, cooperating ASes with peering/trust relationships, or routing over the Internet. This is often because they don't affect routing. That is not the case here; therefore, a clear description of the applicability from a routing context is very important for this document. When comparing or distinguishing from those other techniques, please cover this aspect as well.
Please add an applicability section stating the intended deployment scope,
the cooperation and trust assumptions that scope implies, and the boundary
of the domain - including what a border router does with IPv4-embedded IPv6
packets arriving from outside it. If the answer is a limited domain, RFC
8799 gives the shape of what needs saying. Was this ever discussed in the WG?
[Chongfeng]:Yes, an applicability section about the intended deployment scope will be added.
This has been discussed in the WG, please see below,
https://mailarchive.ietf.org/arch/msg/v6ops/WlNIy_H_EidjvY9fYoOhdfWY1NE/
KT> I don't seem to understand the relation of that specific thread in this context. However, I think we have made good progress in understanding the document's context.
<discuss-3> The document does not situate itself correctly against the
existing IETF work for the same problem. Instead it has misleading, incorrect
and unhelpful comparisons.
160 solutions. When transmitting IPv4 service data, this framework
161 enables end-to-end tunneling or translation across multiple network
162 providers, and therefore is different from existing SRv6/MPLS VPN
163 solutions which are for a single NP. Unlike IP-in-IP tunnel
164 mechanism, where tunnels often operate alongside routing protocols
165 that constitute a control plane (e.g., routing adjacencies over the
166 tunnel), this framework includes a specific mechanism for
167 distributing IPv4-to-IPv6 mapping rules across domains.
First, "existing SRv6/MPLS VPN solutions which are for a single NP" is not
correct. RFC 4364 Section 10, "Multi-AS Backbones", exists precisely for VPN
sites connected to different Autonomous Systems and different SPs, and gives
the three procedures universally known as Inter-AS Options A, B and C. RFC
4364 Section 9, "Carriers' Carriers", covers an SP obtaining backbone
service from another SP. Both have been in the base L3VPN specification
since decades and both are widely deployed between distinct providers.
Second, the contrast with IP-in-IP tunnelling omits the closest prior art.
RFC 5565, the Softwire Mesh Framework, frames exactly this problem in
Section 3.2 - IPv4 islands across an IPv6-only transit core - covers the
multi-AS case in Section 12, and in Section 5 has the AFBRs use BGP to
distribute the other family's reachability with a next hop in the transit
family. That is the requirement RFC 5549 and then RFC 8950 were written to
satisfy; RFC 8950 is cited here while RFC 5565 and RFC 4925 are not. RFC
5565 Section 5 is also explicit that the BGP sessions are not tunnelled, so
softwire mesh does not have the property this paragraph contrasts against.
I don't think this document needs to make these false assertions. What it
does need to acknowledge is the existing prior art in routing technology
deployments when it comes IPv4 over IPv6 and then identify the deployment
or operational issues or improvements that motivated the definition of the
new framework in this document.
Please rewrite the paragraph accordingly and add RFC 4364 and RFC 5565 as
informative references. I would like to check with these prior art were
discussed in the WG?
[Chongfeng]: Below we provide our responses and proposed actions.
Issue 1: Comparison with SRv6/MPLS VPN solutions
We acknowledge that this statement is technically inaccurate. RFC 4364 does clearly
specify Inter-AS VPN options for connecting sites across different Autonomous Systems
and service providers. We agree with the AD that this claim should not stand in the draft.
We will delete the incorrect qualifier "which are for a single NP" from the text in the next
revision. The new framework operates on a different technical basis than MPLS-based
VPNs. The revised paragraph (around lines 160-167) will read as follows:
"...therefore is different from existing SRv6/MPLS VPN solutions. Unlike IP-in-IP tunnel
mechanism, where tunnels often operate alongside routing protocols that constitute a
control plane, this framework includes a specific mechanism for distributing IPv4-to-IPv6
mapping rules across domains."
Issue 2: Reference to RFC 5565
We agree that RFC 5565 is the closest prior art and should have been cited.
We will address this as follows:
1. Add citation to RFC 5565 (and RFC 4925 as the foundational document) in the
Introduction and/or the comparison section.
2. Revise the comparison paragraph to accurately position our work relative to RFC 5565.
We propose the following revised text:
"Softwire Mesh Framework (RFC 5565) aims to addresses the fundamental problem of
connecting IPv4 islands across an IPv6-only transit core. However, RFC 5565 primarily
focuses on the single-domain case, with Section 12 clearly states that,
“Since the procedures described above require the address of a remote
softwire endpoint to be the same as the address of the BGP next hop,
those procedures do not work as specified when the transit core
consists of multiple ASes.”
The new framework is a stateless, address-mapping-based mechanism that performs translation
at the network edge (PE), while the internal network focuses solely on IPv6 forwarding,
it is designed to operate consistently across single or multiple administrative domains.
This design achieves higher forwarding efficiency and aligns with the current trend toward
lightweight, stateless transition technologies.
KT> Thanks for those detail responses. If the document focuses mainly on the translation approach, then it should be sufficient to cite all those prior specifications and their technologies as informational and term them out of scope since all of them are encapsulation based and do not use translation. If this document only focusses on translation, then that should be sufficient. I will leave it to the authors/WG if they wish to get into a detailed comparison between translation and encapsulation, but my recommendation would be to simply acknowledge the alternative (encapsulation) approach, focus on translation and indicate that comparison is out of scope for this document.
<discuss-4> Vulnerability without source validation and BCP 38 filtering
676 In this framework, as the receiver of IPv4-embedded IPv6 packets,
677 each egress PE assumes that all ingress PEs are legal and authorized
678 to send IPv4-embedded IPv6 packets to it. After the egress PE
679 receives IPv4-embedded IPv6 packets, it will convert them into IPv4
680 packets and forward them into the IPv4 Internet. If IPv6 packets
681 cannot guarantee their authenticity or integrity, then there may be a
682 spoofing attack. A malicious ingress PE could send IPv6 packets
683 converted from IPv4 packets to attack an egress PE. Since the PEs in
684 this framework are stateless, even when receiving large-volume
685 traffic flows, they will not increase mapping session counts within
686 the device like a stateful NAT device would, thus avoiding
687 significant consequences. Even if no per-flow state exists, it
The egress PE's behaviour is: match my mapping prefix in the destination,
extract the embedded IPv4 source and destination, emit an IPv4 packet into
the IPv4 Internet. So anywhere the mapping prefix is IPv6-reachable, an
attacker can choose both the IPv4 source and the IPv4 destination of a
packet the egress PE will originate on its behalf, statelessly and at line
rate. That is a source-address-spoofing service rather than a resource-
exhaustion problem, and statelessness makes it cheaper rather than safer.
RFC 6052 Section 5.1 already identifies this case and names the mitigation -
reverse path checks, and verifying that packets come from an authorized
location. The only pointer here is the generic reference to RFC 6052 in
Section 8.2, which does not put an operator anywhere near the control they
need. Note too that the RFC 7915 Section 8 statement adopted in Section 8.2
carries the caveat "and in the routing protocols that are used to make the
packets reach the translator" - and in this framework those are inter-
domain.
Suggest to add: that egress PEs MUST apply source address validation, accepting
only IPv4-embedded IPv6 packets whose IPv6 source is formed from an
authorized PE's mapping prefix and whose embedded IPv4 source is consistent
with that PE's advertised mapping rule, with the behaviour on failure
stated; whether mapping prefixes are expected to be reachable from outside
the participating NPs, and the border filtering that follows; and direct
references to RFC 6052 Sections 5.1 and 5.3 and to BCP 38.
[Chongfeng]: OK, thank you for your suggestion. The following statement will be added in the
security considerations section,
“As stated in section 5.1 of RFC6052, an attacker could use an IPv4-embedded IPv6 address
as the source address of malicious packets. After translation, the packets will appear as IPv4
packets from the specified source, and the attacker may be hard to track. If left without
mitigation, the attack would allow malicious IPv6 nodes to spoof arbitrary IPv4 addresses.
To guarantee the authenticity of the source address of IPv4-embedded IPv6 packets, an
egress PEs must apply source address validation, accepting only IPv4-embedded IPv6
packets whose IPv6 source is formed from an authorized PE's mapping prefix and whose
embedded IPv4 source is consistent with that PE's advertised mapping rule, packets
failing this validation should be discarded and counted; whether mapping prefixes are
expected to be reachable from outside the participating operators, and the border
filtering that follows; The analysis of these security risks can be found in RFC 6052 Sections
5.1 and 5.3 and to BCP 38(RFC2827).“
I will add BCP 38(RFC2827) as informative references.”
KT> Thank you.
<discuss-5> Normative dependency on protocol specifications that don't seem
to have matured, implemented and deployed
432 [I-D.ietf-idr-mpbgp-extension-4map6] can be implemented in PE1 and
433 PE3 to support the MR-DB operations in the framework. In addition,
434 for the mapping rules to propagate from PE3 to PE1 across the
435 network, the intermediate BGP speakers (e.g., route reflectors /
436 ASBRs) that propagate the relevant NLRI will support [RFC8950].
This places a deployment requirement on route reflectors and ASBRs, which
are not part of the framework's own device model - Section 5.1 lists edge PE
devices, core P devices and customer-side IPv4 routers. It states that
requirement as "will support", which carries no normative force, and it does
so for an NLRI whose encoding this document does not define. The same "will"
pattern appears at line 453 in Section 5.1 and at line 585 in Section 5.3.
The underlying claim is right - RFC 8950 Section 5 anticipates the route
reflector case and requires that a next-hop encoding passed along unchanged
is not altered. It is the framing that needs work. Please add the control-
plane devices to the framework's component list in Section 5.1 and replace
"will support" with a determinate formulation. See also my comments below on
Section 5.2, which is the same defect on a different sentence, and on the
classification of RFC 8950.
The other and larger concern is the normative dependency on the BGP that is
placed as informative when much of this framework is heavily dependent
on routing and the mechanics in the IDR document. There is a similar situation
with DNS-based solution, but that seems like an individual draft that I
have not even looked at.
Please add a short subsection to Section 5.2 stating the properties that any
conforming rule-distribution mechanism must have: which fields are carried,
that the key fields including the Forwarding Type are not modified in
transit, the scoping and filtering expectations at administrative
boundaries, and what origin authentication is assumed. Then remove all of
BGP and routing heavy/specifics from this document to the IDR WG document.
Stick to the high-level functionality and the forwarding aspects in this
document. This will eliminate the reference entirely. It also helps perhaps
with the DNS one - though I don't quite follow that one.
[Chongfeng]: OK, I will add a short subsection to Section 5.2 stating the properties
that any conforming rule-distribution mechanism must have: which fields are carried,
that the key fields including the Forwarding Type are not modified in
transit, the scoping and filtering expectations at administrative
boundaries, and what origin authentication is assumed. Then remove all of
BGP and routing heavy/specifics from this document to the IDR WG document.
Enable this document stick to the high-level functionality and the forwarding aspects in this
document.
KT> Thank you.
<discuss-6> Potential for routing loops
496 With the availability of a default egress PE, the ingress PE can
497 deliver the IPv4 packets to the default egress PE when it does not
498 obtain the address mapping rule for that IPv4 address block. From
After the default egress PE translates or decapsulates, the recovered IPv4
packet carries no indication that it has already traversed the framework
once, and it is forwarded on that PE's IPv4 routing information. If that
PE's best IPv4 path for the destination points back toward another
participating PE - entirely possible where the default egress PE's IPv4 view
is partial, or where two NPs each treat the other as their default - the
packet is mapped again and the cycle repeats. Section 8.3 covers traffic
attraction and hijack, but not this.
Please state that a default egress PE is expected to have a complete IPv4
forwarding view for the traffic it attracts, and should not resolve traffic
received via the framework back onto a path that re-enters the framework. It
is worth noting explicitly that TTL and hop limit handling bounds the loop
but does not prevent it.
[Chongfeng]: We agree with your analysis that a recovered IPv4 packet
at the default egress PE could, in certain scenarios, be re-forwarded
back into the framework, thereby creating a forwarding loop. To address
this, we will add the following statement to section 8.3,
“After the default egress PE translates or decapsulates, the recovered IPv4
packet carries no indication that it has already traversed the framework
once, and it is forwarded on that PE's IPv4 routing information. If that
PE's best IPv4 path for the destination points back toward another
participating PE - entirely possible where the default egress PE's IPv4 view
is partial, or where two operators each treat the other as their default - the
packet is mapped again and the cycle repeats. To prevent these
potential routing loops, a default egress PE is expected to have a complete IPv4
forwarding view for the traffic it attracts, and should not resolve traffic
received via the framework back onto a path that re-enters the framework.
It is worth noting explicitly that TTL and hop limit handling bounds the loop
but does not prevent it.”
KT> This is good. However, please consider making this a stricter and normative (MUST?) operational requirement. Without that, the issue of routing loops will not get addressed adequately.
<discuss-7> Another issue with applicability - does it work on PEs, CEs, or UEs?
393 To enable IPv4 service data forwarding in a multi-domain IPv6-only
394 network, IPv4 packets need to be converted to IPv6 packets - either
395 at the UEs/CEs or at the PEs located at the edge of the network.
"either at the UEs/CEs or at the PEs" describes two architectures, and
everything after this sentence describes only the second. Section 5.1's
component list has the customer-side routers as IPv4 speakers, not
converters.
This matters more than a loose sentence, because a UE/CE-based model puts
the mapping prefix and the MR-DB at the customer edge, which changes the
applicability answer, the scale, the security model and the addressing plan
together.
Please either delete "at the UEs/CEs or", or state explicitly that UE/CE-
based conversion is covered by the existing access-side technologies listed
in Section 1 and is out of scope for this framework, which is about the
underlay. I would prefer the former, but I would like the document to
answers the question of how this relates to what already exists - which
Section 1 raises and does not close.
[Chongfeng]: For the applicability of this framework, this sentence is not accurate.
I will delete "at the UEs/CEs or". UE/CE-based conversion is covered by the
existing access-side technologies listed in Section 1.
KT> Thank you.
<discuss-8> The Operational Considerations are written for the translation
path only, and there is no manageability content.
345 * The IPv6 source address is derived by appending the IPv4 source
346 address to its local IPv6 mapping prefix, i.e., Pref6(ingress PE).
Section 6 says that one or more distinct IPv6 mapping prefixes are assigned
to each PE device, so "its local IPv6 mapping prefix" is ambiguous where a
PE has several. Which one is used for source derivation, and must it be one
the PE has advertised in a mapping rule?
More broadly, the document has no manageability content at all: nothing on
detecting MR-DB inconsistency between PEs, nothing on what an operator
observes when a mapping rule is withdrawn, no counters or diagnostics, and
nothing on OAM across the framework. The last of those is not merely an
omission - see my comment on Section 7 below, where ICMPv6 errors generated
by core P routers are addressed to IPv4-embedded IPv6 addresses and the
document never says who translates them back. For a framework whose stated
motivation in Section 1 is that dual-stack doubles the troubleshooting
effort, this is a conspicuous gap.
Please state which mapping prefix is used and whether it must have been
advertised, and add a manageability subsection covering MR-DB monitoring,
the behaviour on mapping rule withdrawal, and OAM across the framework.
[Chongfeng]:We agree that when multiple distinct IPv6 mapping prefixes are assigned to a PE
(as described in Section 6), the statement "its local IPv6 mapping prefix" is unclear.
This is indeed a policy decision that should be left to the network operator. However,
we will revise the document to provide a precise operational framework for this
selection, while explicitly preserving local policy flexibility.
Specifically, we propose the following changes:
1. Revise Section 4.1 (IPv4/IPv6 Address Mapping):
OLD:
"The IPv6 source address is derived by appending the IPv4 source address to its local
IPv6 mapping prefix, i.e., Pref6(ingress PE)."
NEW:
"The IPv6 source address is derived by appending the IPv4 source address to an
IPv6 mapping prefix that is assigned to the ingress PE and has been advertised
via the mapping rule distribution mechanism. The specific mapping prefix to use
for a given packet is a matter of local policy. For example, an ingress PE may use
different source mapping prefixes depending on the egress PE's domain (e.g.,
intra-provider vs. inter-provider)."
2. Add a paragraph in Section 6 (IPv6 Mapping Prefix Allocation):
After describing that one or more prefixes may be assigned, add:
"When a PE is assigned multiple IPv6 mapping prefixes, operators should define a clear
policy for selecting which prefix to use as the source when constructing IPv4-embedded
IPv6 packets. The policy for selecting among multiple IPv6 mapping prefixes on an
ingress PE is deployment-specific. The selection may be based on, for example, the
destination domain, the type of service, or other administrative boundaries. However,
this document does not mandate a single selection approach, but expects operators to
develop and refine their strategies based on operational needs."
3. Add a new Section 8 (Manageability Considerations) to address the
absence of manageability content:
8. Manageability Considerations
This section outlines manageability requirements for the framework, covering MR-DB
monitoring, PE behavior upon rule withdrawal, and OAM considerations.
8.1. MR-DB Consistency Monitoring
Each PE maintains a local MR-DB containing address mapping rules. The following
capabilities are REQUIRED:
a) Each mapping rule SHOULD be identifiable by a version or timestamp
to allow comparison of rule freshness across PEs.
b) PEs SHOULD be able to exchange MR-DB summary information to
verify rule set consistency.
c) PEs SHOULD generate notifications upon significant MR-DB changes
(addition, modification, or withdrawal of a mapping rule).
d) PEs SHOULD maintain counters for packets forwarded using each
mapping rule to enable detection of anomalies (e.g., sustained high
volume through a default rule may indicate missing explicit rules).
e) The MR-DB SHOULD be accessible via management interfaces
(e.g., NETCONF/YANG) for query and audit purposes.
8.2. Behavior on Mapping Rule Withdrawal
When a mapping rule is withdrawn, the following behaviors are REQUIRED:
f) The withdrawn rule MUST be removed immediately and not used for
new packet conversions.
g) A configurable graceful period (e.g., operator-configurable, on the order
of seconds) MAY be supported for packets already in the forwarding pipeline to
avoid disrupting established sessions.
h) If no explicit rule exists for a destination, the packet SHOULD be
forwarded via the default mapping rule if configured; otherwise, it SHOULD
be dropped and an ICMP/ICMPv6 error MUST be generated.
i) A notification MUST be generated upon rule withdrawal, including the
affected IPv4 prefix and the reason.
8.3. OAM Requirements
8.3.1. ICMP/ICMPv6 Error Handling
When a core P router generates an ICMPv6 error message destined to an
IPv4-embedded IPv6 address, the message MUST be forwarded to the egress
PE identified by the destination mapping prefix. Upon receiving such a
message, the egress PE MUST translate it to the corresponding ICMP message
and forward it toward the original IPv4 source, in compliance with
[RFC7915]. Inter-provider SLAs SHOULD permit such ICMPv6
messages to traverse administrative boundaries to avoid PMTUD black holes.
8.3.2. Connectivity Verification
Operators MUST be able to initiate diagnostic probes (e.g., ping, traceroute) from
an ingress PE to an IPv4 destination behind a remote egress PE. These tools
SHOULD support tracing the full path across the IPv6 underlay. Each PE
SHOULD support a diagnostic mode to verify local translation/encapsulation
functions using test packets.
8.3.3. Performance Monitoring
The following metrics MUST be monitored at PE devices:
j) Conversion latency
k) Packet drop rate due to MR-DB lookup failures
l) Throughput per egress PE prefix
Operators SHOULD establish baselines and alerts for significant deviations.
8.3.4. Cross-Domain Coordination
For multi-domain deployments, the following coordination is REQUIRED:
m)Inter-provider SLAs SHOULD include provisions for ICMPv6
traversal across administrative boundaries and cooperative troubleshooting.
n) Out-of-band communication channels SHOULD be established for coordinating
on inter-domain issues.
o) PEs SHOULD report the reachability and health of their mapping rules to a
central management system for unified monitoring.
Also, do you think this would be too detailed?
KT> Thanks. This is in the right direction. However, the use of BCP14 keywords seems excessive even if all the points are pertinent to the framework. Perhaps just rephrase them as considerations rather than mandates? That said, this is up to the authors and the WG. The content addresses my concern either way.
<discuss-9> What's the framework about? Translation or Encapsulation or both?
563 2. Encapsulation
565 The address mapping process for encapsulation follows the same
566 procedure as translation: When the ingress PE receives an IPv4
Translation and encapsulation are presented as two realisations of
one framework, and lines 598 to 601 assert they are equivalent from the
control plane's and the core's point of view. This is not correct and in
fact the rest of the document does not treat them as equals. RFC 6052 and
RFC 7915 are normative; RFC 2473, the only thing this mode rests on, is
informative - which is wrong if encapsulation is a supported mode. The
Abstract does not mention encapsulation. Section 4.1 states the proposal
as translation. Section 8.2 says flatly that the Stateless IP/ICMP
Translation Algorithm is used, and there is no security analysis of the
tunnel mode anywhere. This subsection is defined by reference to the
translation one and never describes outer header construction, which the
INTDIR early review raised and I do not see addressed. Section 7's
operational guidance is entirely RFC 7915-based.
Then there is the routing part (lines 432 to 443) that do not do justice
to the encapsulation mode which uses RFC 5565 and RFC 8950. It misleads
by referring to both the new BGP mapping prefix extensions and the existing
work without getting into which of the two modes they work on and how.
There is also the sense that I get which is that mapping rules are
required when doing encapsulation, and I fail to understand that. Why is
translation required when two IPv4 hosts are talking over an IPv6-only
underlay using an encapsulation mode? I get it that mapping prefixes are
needed for translation. This seems like another area where the two modes
blur into each other and create ambiguity (and for me, confusion).
[Chongfeng]: Yes, the stateless method of address mapping mentioned in
the abstract supports both translation and encapsulation. For translation,
IPv6 addresses generated by stateless address mapping can be used
as the outer source and destination addresses.
I can see two options.
Promote encapsulation: make RFC 2473 normative, specify outer header
construction including Traffic Class, Flow Label and Hop Limit, add ECN
handling per RFC 6040, add tunnel MTU and ICMP relay per RFC 2473 Sections
7.1 and 8, extend Sections 7 and 8 to cover the mode, resolve the mode
selection rule at lines 584 to 590, add the routing part, and position the
result against RFC 5565 rather than re-deriving it.
OR demote encapsulation: state that this framework specifies stateless
translation, and that encapsulating IPv4 in IPv6 across a multi-AS core
is an alternative realisation of the same address-mapping control plane,
already framed by RFC 5565 and out of scope here.
My sense from the document is that the second is preferred, but I could be
wrong. The second option resolves the mode selection question outright
and reduces most of the ECN question and parts of the MTU and ICMP ones to
scoping statements, and it makes the document's actual contribution -
stateless per-prefix mapping with no per-destination tunnel state - stand
out instead of being one of two half-described options.
[Chongfeng]: OK, we prefer the second option, that is, demote encapsulation.
However, as mentioned before, RFC5565 does not focus on implementing end-to-end
encapsulating IPv4 in IPv6 across a multi-AS core. The following statement will be
added in section 4.2,
“this framework specifies stateless translation IPv4/IPv6 address translation, which can be
used both for packet translation or encapsulation at the data layer. However, this
document mainly illustrates the framework using the packet translation mechanism,
encapsulating IPv4 in IPv6 across a multi-AS core using the same address-mapping
control plane is out of scope here. ”
KT> This is ok, but I still don't follow why translation is at all required when doing encapsulation. Can you please clarify? When I suggested "demote encapsulation", my point was to indicate that encapsulation approach also exists (with reference to all the relevant existing specs and deployments), but that this document focuses on the translation based approach.
The subsequent text in this document will mainly elaborate on the packet translation mode.
If the authors/WG chose two modes deliberately then so be it, but then the
first option is the obligation, and the reference classification needs
fixing either way. The document needs to clearly disambiguate both modes,
their details, and then do a comparison analysis between them.
Was carrying both an explicit WG choice, and if so what deployment need does
one mode serve that the other does not? It goes both ways.
Note: The comments also indicate some serious defects that could be considered
to be worthy of DISCUSS, but I didn't want to load up everything in this block.
Would appreciate if the comments were also given a thorough treatment. Many of
them are related to the discussion points.
----------------------------------------------------------------------
COMMENT:
----------------------------------------------------------------------
A few general comments first:
1) <major> Zero-checksum UDP is not addressed. RFC 7915 Section 4.5 requires
the translator to update the transport checksum for UDP packets that carry
one, and for those that do not it says the translator SHOULD provide a
configuration function whose options are to drop the packet and raise a
management event, or to compute the checksum. IPv4 hosts do legitimately
send UDP with a zero checksum. So in translation mode this traffic is either
dropped or costs a full-payload checksum computation at the PE, while in
encapsulation mode it passes untouched - one more way in which the two modes
are not equivalent. This was pointed out in the GENART Last Call review.
[Chongfeng]: This is a framework draft, zero-checksum UDP is specific to the mechanisms of
translation or encapsulation themselves, this issue that can be addressed by
RFC7915-bis or other documents, and will not be covered by this draft. In addition,
the framework in this document mainly be illustrated based on stateless packet
translation mode.
KT> If the document is only focussing on translation, then I agree that it does not need to get into this topic and we can leave it to RFC7915 (or its bis).
Thanks,
Ketan
2) <editorial> In Section 5.3, lines 554 to 561 and 576 to 582 are the same
paragraph with translation and decapsulation swapped, and lines 545 to 552
and 565 to 574 are similarly parallel. Since the egress check needs
rewriting in both places anyway (see my comment at lines 554 to 560), this
is the moment to factor the common procedure out once and state only the
delta for each mode.
[Chongfeng]: OK, this will be addressed by describing the procedure based on
the packet translation mode, there will be no overlap.
More detailed inline comments follow:
94 with IPv6 traffic growing at a faster rate than IPv4. As of 2022,
95 most IPv6 deployments rely on dual-stack [RFC4213]. However, dual-
<minor> A 2022 datum in a document that will publish in 2026 or later. Either
refresh it with a current reference or drop the sentence or drop the year and
make the statement timeless.
[Chongfeng]: “As of 2022,” will be changed to “Currently”.
152 the definition of "IPv6-only". Unless otherwise stated, the term
153 “IPv6-only network” in this document refers specifically to
154 “IPv6-only underlay network”. This document presents a framework for
<minor> The definition of "IPv6-only network" sits mid-paragraph in Section
1 rather than in the terminology of Section 2. Separately, "IPv6 mapping
prefix" and Pref6(PE) are used throughout the document and are defined
nowhere. Please move this definition into Section 2 and add entries for the
other two.
[Chongfeng]:
Based on your suggestion, we will make the following changes in the next revision:
Relocate the definition of "IPv6-only network": The definition currently in Section 1
will be moved to Section 2 (Terminology). In its original location, we will add a brief c
ross-reference, for example: "This document uses the term 'IPv6-only network' as
defined in Section 2."
Add definitions for missing terms: We will add the following entries to the Terminology
section (Section 2):
IPv6-only network: Unless otherwise stated, the term “IPv6-only network” in this
document refers specifically to an “IPv6-only underlay network” (i.e., a network where
the forwarding plane and transport infrastructure support only IPv6, with no native
IPv4 capability).
IPv6 mapping prefix: A specific IPv6 prefix allocated to a PE device. It is used to construct
an IPv4-embedded IPv6 address by combining with an IPv4 address, as described in
Section 4.1 of this document.
Pref6(PE): The IPv6 mapping prefix assigned to a specific Provider Edge (PE) device. This
prefix uniquely identifies the PE within the IPv6-only underlay network and is used by other
PEs to determine the correct egress PE for an IPv4-over-IPv6 flow.
365 "Network-Specific Prefix". Note that [RFC7915] also allows the use
366 of unicast addresses without u-bit (as long as they are not derived
367 from an IEEE MAC-layer address).
<major> RFC 7915 does not allow this; RFC 7915 Section 6 notes that RFC 7136
updates RFC 4291 to allow it, and neither RFC 7136 nor RFC 4291 is referenced
here. RFC 6052 Section 2.2 says that bits 64 to 71 MUST be set to zero, and
RFC 7136's relaxation of RFC 4291 does not lift that. Placed immediately
after Table 1, in the section that defines how PEs construct addresses,
this note tells an implementer the u-bit constraint is optional. If that is
the intent it needs to be stated as a deliberate deviation and justified.
The reason it matters: Section 6 never constrains mapping prefixes at all.
It does not say that a mapping prefix must be one of the six lengths RFC
6052 Section 2.2 permits, that bits 64 to 71 are zero - which for a /96
mapping prefix is a constraint on the prefix itself, and therefore on NSP
assignment - or that the suffix is zero. The summary at lines 361 to 362 and
the terminology entry at lines 191 to 192 both omit the u octet as well, so
the document's three descriptions of the address format are mutually
inconsistent. Please align them with RFC 6052 Section 2.2, and point at RFC
6052 Section 3.3 for length selection - this is important as those prefixes
get into routing.
[Chongfeng]: We acknowledge the issues you have raised regarding the address
format definition and the incorrect citation of RFC 7915. We will address them as follows:
1) We will remove the incorrect reference to RFC 7915 in line 367. The description of
the IPv4-embedded IPv6 address format will be aligned strictly with RFC 6052 Section 2.2,
making it clear that the u-octet MUST be zero.
2) Regarding that three descriptions of the address format (lines 191–192, 361–362,
and 365–367) that are mutually inconsistent, as they all omit the u-octet constraint, we
will revise all three locations to ensure they consistently and accurately reflect the address
format defined in RFC 6052 Section 2.2, including the zero u-octet requirement.
In addition, this section mainly talks about address format, instead of IPv4/IPv6 packet
translation, so I think RFC7915 can be discussed less here.
OLD:
"IPv4-embedded IPv6 address: An IPv6 address embedded with a 32-bit
IPv4 address used to represent an IPv4 node in an IPv6 network
([RFC6052]). In this document, the IPv4-embedded IPv6 address is
composed of an IPv6 mapping prefix and an IPv4 address."
NEW:
"IPv4-embedded IPv6 address: An IPv6 address formed by a variable-length
prefix, a zero-u-octet (bits 64-71), an embedded 32-bit IPv4 address, and a
variable-length zero suffix, as defined in Section 2.2 of [RFC6052]."
At line 361–362, the following change will be made,
OLD:
“It is the same as Table 1 of [RFC6052], except that the first column of the latter is "Network-Specific Prefix".”
NEW:
“Table 1 shows examples of IPv4-embedded IPv6 addresses constructed
according to Section 2.2 of [RFC6052]. It is the same as Table 1 of
[RFC6052], except that the first column is referred to as the "IPv6
mapping prefix" rather than the "Network-Specific Prefix".”
In addition, the following sentence will be removed((lines 365–367)),
“Note that [RFC7915] also allows the use of unicast addresses without u-bit
(as long as they are not derived from an IEEE MAC-layer address).”
3)Regarding the constraints on the mapping prefix itself,we will update Section 6 to explicitly state that:
-The IPv6 mapping prefix MUST be one of the lengths permitted by RFC 6052 Section 2.2.
-For a /96 mapping prefix, bits 64–71 MUST be zero.
-The suffix portion MUST be zero.
-We will also reference RFC 6052 Section 3.3 for guidance on prefix length selection.
398 e.g., NP-3) from a client-facing interface, it queries its mapping
399 rule database (i.e., MR-DB) to find the rule that best matches the
400 packet's destination IPv4 address. The IPv6 mapping prefix in the
<question> "Best matches" gives impression of longest-prefix match behaviour
(am I correct?) but not the tie-break when two PEs advertise mapping rules
for the same IPv4 block - multihomed customers, anycast IPv4 services, or
simply two PEs with a route to the same aggregate. Is rule selection delegated
entirely to the distribution protocol's best path process? And is multipath
across egress PEs intended to be supported? If either answer is yes, a sentence
saying so belongs here or in Section 5.2. Please clarify in any case.
[Chongfeng]: Yes, "best matches" here means the longest-prefix match. The rule
selection is delegated to the underlying routing protocol's best-path selection
procedure (e.g., BGP's decision process). The details of that procedure are outside
the scope of this document. Multipath forwarding (i.e., load balancing across
multiple egress PEs) is not supported by this framework in its current version.
Based on your comments, we plan to add the following text in Section 5.2 (Address
Mapping Rule Processing):
"When multiple mapping rules match the same IPv4 destination address block
(e.g., due to multihoming or anycast services), the ingress PE selects the single
best rule according to the underlying routing protocol's best-path selection procedure
(e.g., BGP's decision process). This ensures deterministic egress PE selection.
The detailed selection procedure is outside the scope of this document. Multipath
forwarding (i.e., load balancing across multiple egress PEs) is not considered in this framework."
476 For IPv4 service delivery, IPv4/IPv6 address mapping rules need
477 to be generated. In the network shown in Figure 1, when PE3
478 receives an IPv4 BGP route advertisement from an IPv4 border
479 router, e.g., BR1, it extracts IPv4 address blocks and generates
480 address mapping rules by combining them with its own IPv6 mapping
481 prefix. All the address mapping rules, whether locally generated
482 or received from other PEs, are stored in its local MR-DB. PE
483 devices also support rule management operations, such as
484 insertion, modification, and deletion of address mapping rules.
<major> Neither this nor the statement in Section 4.1 that a PE's IPv4
blocks are "extracted from local IPv4 routing table or address pool" bounds
how many mapping rules exist. The document never says which IPv4 blocks a PE
should originate rules for, whether rules track received BGP routes one for
one or are aggregated or configured, or what an MR-DB is expected to hold.
So a reader cannot tell whether the intended operating point is tens of
rules per PE or several orders of magnitude more, and an operator cannot
size for either. Note that the default egress PE is introduced here as a
fallback for rules that have not arrived, not as a policy for limiting which
rules are generated.
Please state the rule-origination policy. My suggestion: a PE originates
mapping rules only for IPv4 address blocks for which it is the authoritative
or aggregating egress - customer blocks, address pools, site aggregates -
and not for transit routes learned from external peers, with the default
rule covering the remainder; then give the expected order of magnitude of
an MR-DB under that policy. This all has significant implications on routing.
Was any of this discussed within the WG? The answer also depends on the
applicability question in my DISCUSS.
Please give a proper thought on how much of such and other routing aspects
this framework wants to get into and if all of this should be moved into
a routing document (e.g., the BGP spec).
[Chongfeng]:
We agree that the document should clearly state a rule-origination policy to
bound the size of the MR-DB. Based on your suggestion, we propose to amend
the following text in section 5.2:
“A PE originates mapping rules only for IPv4 address blocks for which it serves
as the authoritative or aggregating egress point. These blocks include, but are
not limited to, its directly connected customer prefixes, locally provisioned address
pools, and site aggregates. A PE must not originate per-prefix mapping rules for
transit routes learned from external BGP peers. Traffic destined for all other IPv4
addresses (i.e., the remainder of the global IPv4 space) is covered by a default
mapping rule pointing to a configured default egress PE.”
We will also add the following text at the end of Section 5.2 to provide an
order-of-magnitude estimate for the MR-DB size,
“Under the rule-origination policy above, a PE's MR-DB contains mapping entries
only for locally significant IPv4 prefixes—i.e., its own customer blocks, address
pools, and site aggregates. In typical deployments, this yields an MR-DB size on
the order of tens to a few hundred entries, substantially smaller than the global
routing table. The default egress PE mechanism covers all non-local traffic, preventing
unbounded MR-DB growth. This ensures that MR-DB lookups, synchronization, and
rule management impose minimal overhead on PE control and data planes, making the
design operationally manageable even in networks with complex external peering.”
489 correct egress PE. To mitigate this issue, the framework
490 introduces a default egress PE, which advertises a default
491 address mapping rule to all other PEs. The format of default
<major-editorial> The default rule is defined here, and the only constraint
on its use - the document's only MUST, at lines 731 to 733 - is in Section
8.3. A reader implementing or deploying from this section gets the mechanism
and a forward pointer to a security discussion, but not the scoping rule.
Please move the normative scoping statement here, immediately after the rule
format at line 494, and leave Section 8.3 with the threat analysis and the
monitoring and rate-limiting advice, cross-referencing back. The substance
does not change; only the location.
[Chongfeng]:
The normative "MUST" statement at Section 8.3 has been removed from the document.
In light of this, we believe the most effective way to address the underlying concern—that
readers encountering the rule format in Section 5.2 lack immediate context—is to
augment the mechanism description with a clear statement of the rule's purpose and
intended use. Specifically, we intend to make the following replacement immediately
after the rule format (0.0.0.0/0: Pref6(PE)) at line 494:
OLD:
“With the availability of a default egress PE, the ingress PE can
deliver the IPv4 packets to the default egress PE when it does not
obtain the address mapping rule for that IPv4 address block.”
NEW:
“The default rule associated with the designated default egress PE serves as a fallback
mechanism to ensure connectivity when a more specific address mapping rule for
a given IPv4 destination is not available in the ingress PE's MR-DB. When such a
rule is present, the ingress PE forwards the corresponding IPv4-embedded IPv6
packets to the default egress PE. This ensures that IPv4 traffic is not dropped due to
missing mapping information, though it may lead to suboptimal forwarding paths.
Operators should carefully consider the placement and capacity of the default
egress PE to avoid congestion or single points of failure.”
The subsequent sentences in Section 5.2 that introduce the security implications and
point to Section 8.3 remain unchanged, ensuring continuity with the threat analysis.
506 Address mapping rules generated on a PE device need to be
507 distributed to PE devices in other domains. During the
508 transmission of these rules, the key information
509 elements,including the Forwarding Type field defined in
510 Section 4.1, must not be modified by any intermediate node.
511 Furthermore, the transmission procedure must support scalability
512 across multiple independently administered network providers
513 (NPs).
<major> Three things in one sentence, in a document that has a BCP 14
section.
"must not be modified by any intermediate node" is a requirement on third-
party BGP speakers, and whether they honour it determines whether the
framework works. Two implementations reading this differently - one treating
the Forwarding Type as a modifiable transitive attribute - would produce a
working-looking control plane that selects the wrong transformation mode.
That is an interoperability consequence, so the lowercase here is a problem.
Next, the cross-reference is wrong: Section 4.1 does not define the
Forwarding Type field, it names it and delegates the structure to draft-
ietf-idr-mpbgp-extension-4map6 - yet another normative dependency. Third,
"must support scalability" at lines 511 to 513 is a design goal in
requirement clothing, and nothing can be tested against it.
Perhaps: "The key information elements of an address mapping rule, including
the Forwarding Type field, MUST NOT be modified by any intermediate node."
Then correct or remove the Section 4.1 pointer, and reword lines 511 to 513
as a design goal. Better still, fold this into the properties subsection
suggested in my first general comment above, where it is a statement about
what a conforming distribution mechanism must guarantee.
[Chongfeng]:
We agree with all three points raised and have revised the document accordingly.
Details are below.
(1) Corrected the normative language to use BCP 14 keywords;
The following changes will be made in section 5.2,
OLD:
“During the transmission of these rules, the key information
elements, including the Forwarding Type field defined in
Section 4.1, must not be modified by any intermediate node”
NEW:
“During the transmission of these rules, the key information
elements, including the Forwarding Type field mentioned in
Section 4.1, MUST NOT be modified by any intermediate node”
In addition, a more detailed illustration of “Forwarding
Type” will be added in section 4.1, please see below,
“The Forwarding Type field indicates the IPv4/IPv6 forwarding capability
of the egress PE. Specifically, it informs the ingress PE whether the egress PE
supports translation-only, encapsulation-only, or both. Based on the value of
this field, the ingress PE can select an IPv4/IPv6 transformation mechanism
that is supported by both itself and the egress PE, ensuring interoperation
when multiple transformation modes are available in the network.”
(2) Regarding the incorrect cross-reference;
Regarding the cross-reference error, we have removed the pointer to
draft-ietf-idr-mpbgp-extension-4map6 in Section 4.1 to clarify that
this framework document does not specify BGP routing or the detailed
protocol structures. We have replaced it with a sentence stating that the
detailed structure is out of scope for this document, please see below.
OLD:
“The detailed structure of the address mapping rule is defined
in [I-D.ietf-idr-mpbgp-extension-4map6].”
NEW:
“The detailed structure of the address mapping rule, including the
definition of the Forwarding Type field, is out of the scope of this
document and is expected to be defined by the specific control-plane
protocol(s) used for rule distribution.”
3) For the issue of "must support scalability". We acknowledge that this phrase
is a design goal rather than a testable requirement. We will reword lines
511-513, as you suggested, we will revise it to a clear design goal, as below.
"The distribution mechanism should be designed with the goal of supporting
scalability across multiple independently administered network operators."
We use lowercase ‘should’ deliberately, as this is a design goal and not a protocol
requirement, per BCP 14.
554 Upon receiving the IPv6 packet, the egress PE checks whether the
555 destination IPv6 prefix matches its own IPv6 mapping prefix. If
556 not, it discards it or forwards it as a regular IPv6 packet.
557 Otherwise, the egress PE extracts the original IPv4 source and
558 destination addresses from the IPv4-embedded IPv6 addresses and
559 reconstructs the original IPv4 packet, and this process complies
560 with [RFC7915]. The IPv4 packet is then forwarded based on the
<major> Two error-handling gaps, both of which appear twice - the same
paragraph recurs at lines 576 to 578.
The only test is on the destination. Line 557 then assumes the source is an
IPv4-embedded IPv6 address, but nothing has checked that. A native IPv6 host
- or an attacker, per my DISCUSS on Section 8.1 - can send to Pref6(PE) with
an arbitrary IPv6 source, and a stateless translator cannot derive an IPv4
source from it. The document needs to say the packet is discarded.
And "it discards it or forwards it as a regular IPv6 packet" is not a
specification. Two implementations can legitimately choose differently, with
different observable behaviour: a silent drop, or a packet re-entering the
IPv6 forwarding path possibly toward the same PE.
Perhaps, for both occurrences: if the destination address does not match one
of the PE's own IPv6 mapping prefixes, the packet is not subject to this
framework and is forwarded according to the normal IPv6 forwarding rules; if
the destination matches but the source is not an IPv4-embedded IPv6 address
formed from an authorized PE's mapping prefix, the packet MUST be discarded
and SHOULD be counted.
[Chongfeng]: Firstly, this document will mainly center around translation
mode, so the encapsulation-related content will be demoted, the second
occurrence of the above text will be removed as well.
Secondly, based on your suggestion, the following change will be made,
OLD:
“Upon receiving the IPv6 packet, the egress PE checks whether the
destination IPv6 prefix matches its own IPv6 mapping prefix. If
not, it discards it or forwards it as a regular IPv6 packet.
Otherwise, the egress PE extracts the original IPv4 source and
destination addresses from the IPv4-embedded IPv6 addresses and
reconstructs the original IPv4 packet, and this process complies
with [RFC7915]. The IPv4 packet is then forwarded based on the
IPv4 routing information maintained at the egress PE.”
NEW:
“Upon receiving the IPv6 packet, the egress PE processes the packet
according to the following cases:
-If the destination address does NOT match one of the PE's own IPv6
mapping prefixes:
The packet is not subject to this framework and must be forwarded
according to the normal IPv6 forwarding rules.
-If the destination address DOES match, but the source address is
NOT a valid IPv4-embedded IPv6 address(i.e., whose embedded
IPv4 prefix corresponds to the PE's own mapping prefix):
The packet MUST be discarded, and implementations SHOULD maintain
a counter for such drops (for operational visibility and security monitoring).
-If BOTH the destination matches AND the source is a valid
IPv4-embedded IPv6 address:
The existing packet translation logic (extracting IPv4 addresses and reconstructing
the packet per RFC7915) applies.”
572 procedure described in Section 4.1. The framework utilizes the
573 IPv4-in-IPv6 tunnel technique defined in [RFC2473], which enables
574 IPv4 packets to traverse through native IPv6 networks.
<major> RFC 6040 specifies ECN handling for IP-in-IP tunnels, and its scope
covers any IP-in-IP tunnelling irrespective of which version is used for the
inner or outer header, so this mode is covered. RFC 2473 predates it and is
not conformant. RFC 2473 Section 5 lets the tunnel entry point set the outer
Traffic Class to a pre-configured value, overwriting the ECN codepoint that
RFC 6040 Section 4.1 requires be copied. And RFC 2473 has no rule for
propagating congestion experienced inside the tunnel, so a CE mark applied
by a P router in the IPv6-only core is discarded at lines 579 to 580 when
the outer header is removed, where RFC 6040 Section 4.2 requires the more
severe of the inner and outer markings to be propagated.
This is not a corner case: the whole point of the framework is that traffic
crosses several IPv6-only ASes, which is where the congestion happens.
Section 7 already worries about ECN for the translation path; the
encapsulation path has the same problem by a different route and is not
mentioned.
Please add a pointer that PEs implementing encapsulation follow RFC 6040 for
ECN field construction and propagation, notwithstanding RFC 2473's traffic
class text; add RFC 6040 to the references; and extend the ECN paragraph in
Section 7 to cover both modes.
Separately, on the translation side: the "administrative bleaching" concern
in Section 7 is a property of RFC 7915 Sections 4.1 and 5.1 rather than of
this document, and draft-ietf-v6ops-rfc7915-bis is a V6OPS WG document. The
durable fix is to have the bis require that the ECN bits be preserved where
the DSCP is administratively rewritten, and that is worth raising there.
[Chongfeng]:
We fully agree with your technical analysis. Specifically:
• We acknowledge that RFC 2473 predates RFC 6040 and that its Traffic Class
handling is not conformant with the current ECN requirements for IP-in-IP tunnels.
• We also agree that this is not a corner case, as the framework's encapsulation
mode is indeed intended to carry traffic across multiple IPv6-only ASes where congestion
is likely to occur.
• We concur that Section 7 currently only addresses the translation path and omits
the encapsulation path, which is a gap that needs to be fixed.
To address these issues, and considering that, we will make the following changes
in the next revision:
1. In Section 4.1 (or wherever RFC 2473 encapsulation is described), we will add a
clear statement that tunnel entry and exit points MUST follow RFC 6040 for ECN
field construction and propagation, notwithstanding the older Traffic Class text in
RFC 2473 Section 5.
2. In the References section, we will add RFC 6040 as a normative reference.
3. In Section 7, we will extend the existing ECN paragraph to explicitly cover both
the translation mode and the encapsulation mode, making it clear that the RFC 6040
requirements apply to the latter as well.
Question: Do we still need to add these when encapsulation mode will be demoted?
Additionally, we note your comment regarding the 'administrative bleaching' issue on
the translation side. We agree that the correct venue for a durable fix is the ongoing
work on draft-ietf-v6ops-rfc7915-bis, and we will ensure that the discussion in our
Section 7 accurately reflects this provenance without attempting to prescribe a solution
for that document.
584 For IPv4 packet delivery across one IPv6-only network, the ingress PE
585 and the egress PE will use the same IPv4/IPv6 transformation
586 mechanism (translation or encapsulation). When the address mapping
587 rule corresponding to the destination address of a given IPv4 packet
588 is available, based on the value of Forwarding Type of the egress PE,
589 the ingress PE can select the IPv4/IPv6 transformation mechanism
590 which are supported by itself and the egress PE.
<major> Where both PEs advertise support for both mechanisms, there is no
tie-break. Two ingress PEs sending to the same egress PE may choose
differently, and the forward and reverse directions of a single flow may use
different mechanisms.
Lines 598 to 601 are true of the control plane and of the core, and
misleading about the data plane. The modes differ in header overhead - 40
bytes against 20 or 28 - in whether IPv4 options survive, in the treatment
of zero-checksum UDP, in ECN behaviour, and in fragmentation and ICMP relay.
This was raised in the INTDIR early review I do not see it addressed.
Please either state a deterministic selection rule - for example,
encapsulation preferred where both PEs support it, and translation used only
where an egress PE is translation-only - and state that the choice is per
egress PE and applies in both directions of a flow; or, if the choice is
intended to be a local policy matter, say so explicitly and replace lines
598 to 601 with an honest list of the ways in which the two modes differ, so
that an operator can make the choice on informed grounds.
[Chongfeng]:We acknowledge that the current text in lines 598-601 is indeed
ambiguous and fails to provide operational clarity. After careful consideration,
we have chosen to adopt implementing a deterministic selection rule, and
the following subsection will be added,
5.4. Mode Selection Between Translation and Encapsulation
When both the ingress PE and egress PE support both translation and encapsulation
mechanisms (as indicated by the Forwarding Type field in the address mapping rule),
the following deterministic selection rule MUST be applied:
Encapsulation SHALL be preferred over translation when both PEs support encapsulation.
Translation SHOULD be used only in either of the following cases:
-The egress PE does not support encapsulation (i.e., it is translation-only)
-The ingress PE does not support encapsulation (i.e., it is translation-only)
-Encapsulation is administratively disabled by local policy
This selection MUST be made on a per-egress PE basis, and for a given flow, both forward and
reverse directions MUST use the same mechanism to ensure symmetry and avoid issues with
stateful middleboxes.
592 For IPv4-embedded IPv6 packets, regardless of whether translation or
593 encapsulation is used, the Pref6 part of the IPv6 destination address
594 identifies the egress point. Therefore, packet forwarding can be
595 performed by P devices solely based on the Pref6 part of the
596 destination address.
<major> True of the forwarding decision, and it omits the load balancing
decision, which is where multi-domain backbones actually live. In
encapsulation mode the outer header has Next Header 4, and RFC 2473 Section
5 says the Flow Label is typically set to zero, so a P router or a LAG
hashing on the outer 5-tuple sees a zero flow label and, for an unparsed
inner IPv4 header, no ports. The only entropy available is the embedded IPv4
source and destination pair, which gives per-address-pair granularity -
adequate for aggregate traffic, poor for a small number of large flows
between the same pair, which is the data centre to data centre case Section
3 calls out. Translation mode is better off, since the outer header carries
the real transport ports at their normal offsets - which is one more
asymmetry between the two modes.
Please add a sentence that ingress PEs performing encapsulation should set
the IPv6 Flow Label per RFC 6437 from the inner packet's flow
identification, so that ECMP and LAG hashing in the IPv6-only core has
usable entropy, and note the difference between the modes.
And then, related to my DISCUSS point, why is the Pref6 coming into play
when IPv4 traffic is being encapsulated across and IPv6-only core?
[Chongfeng]: Based on your suggestion that this document should mainly center
around translation mode, we have decided to remove the encapsulation-related part.
This eliminates the entropy concern you raised. In addition, the following change will be made:
OLD:
"For IPv4-embedded IPv6 packets, regardless of whether translation or
encapsulation is used, the Pref6 part of the IPv6 destination address
identifies the egress point. Therefore, packet forwarding can be
performed by P devices solely based on the Pref6 part of the
destination address."
NEW:
"For IPv4-embedded IPv6 packets, the Pref6 part of the IPv6 destination address
identifies the egress point. Therefore, packet forwarding can be
performed by P devices solely based on the Pref6 part of the
destination address."
616 PE's IPv6 mapping prefix should be allocated as a sub-prefix of a
617 larger IPv6 prefix that is already advertised in the network and
618 whose next hop reaches that PE (or its site). This allows
619 IPv4-embedded IPv6 packets to be forwarded using existing aggregate
620 routes, avoiding the need for additional FIB entries in the IPv6
621 core.
<major> This is the only place where the relationship between a mapping
prefix and IPv6 reachability is addressed, and it is a lowercase "should"
with no failure analysis. Three consequences the document does not draw.
Nothing requires the ingress PE to verify that Pref6(egress PE) resolves to
a usable IPv6 route before installing and using a mapping rule. In BGP terms
this is a next hop with no recursive resolution requirement. A rule whose
Pref6 is unreachable - the covering aggregate withdrawn, filtered at a
boundary, or never advertised toward this NP - silently drops traffic for the
entire IPv4 block it covers, and the ingress PE has no reason to fall back
to the default egress PE, because it has a rule.
If the covering aggregate is advertised from more than one site, which is
normal practice for a multihomed NP, the packet can be delivered to a PE
that did not advertise the rule. That PE hits the branch at lines 555 to 556
and either discards it or forwards it as a regular IPv6 packet - silently
either way.
And per-PE mapping prefixes are longer than anything most operators will
accept inter-domain, so the framework depends entirely on the covering
aggregate being advertised and accepted across the boundary. That is a real
addressing and routing policy prerequisite and the document does not name
it.
Please state that a mapping rule is usable only while Pref6(PE) resolves to
a reachable IPv6 route, and what a PE does when it does not; and add a
sentence that mapping prefixes are not expected to be advertised inter-
domain, so the framework depends on the covering aggregate propagating to
all participating NPs.
Again, related to my DISCUSS point, there is a lot of routing specifics being
tackled in the document that need a thorough review; seriously consider
the scope of this document and whether these topics are best covered in the
BGP spec.
[Chongfeng]: Agree with your assessment, routing-related content will be moved to BGP spec.
631 network. When IPv4 packets are encapsulated in IPv6, the resulting
632 packet is 40 bytes larger than the original; when translated, the
633 IPv6 header is 20 bytes larger than the IPv4 header. In a multi-
<major> "20 bytes larger" is the best case rather than the general one. RFC
7915 Section 4.1 adds a Fragment Header when the incoming IPv4 packet is
already a fragment, or when the DF bit is clear and the resulting IPv6
packet would exceed the configured lowest-ipv6-mtu. A Fragment Header is 8
octets, so the translation overhead is 28 bytes precisely in the large-
packet, non-DF case that drives the rest of this paragraph. RFC 7915 Section
1.4 itself says "20+ octets" for the IPv4 header, since options are possible
(I doubt they are used though) and are not translated. Since this paragraph
tells operators how much headroom to provision, the number needs to be right.
Separately, all the guidance at lines 638 to 642 is translation-only - every
cross-reference is to RFC 7915. The encapsulation mode's MTU and ICMP
behaviour is in RFC 2473 Section 7.1 and Section 8, neither of which is
cited, so an operator following this section for an encapsulation deployment
gets no guidance at all. RFC 7915 Section 7, on ICMPv6 Packet Too Big, is
also uncited and is directly relevant to the paragraph that follows.
Consider giving a concrete number as well - for example an IPv6-only core
MTU of at least 1500 plus 48 bytes if 1500-byte IPv4 payloads are to survive
without fragmentation in either mode - since "ensuring consistent MTU across
domains" is not actionable without one.
[Chongfeng]: I agree with your analysis that the statements of the increased number
of bytes are inaccurate, so the following change will be made,
OLD:
"When IPv4 packets are encapsulated in IPv6, the resulting
packet is 40 bytes larger than the original; when translated, the
IPv6 header is 20 bytes larger than the IPv4 header."
NEW:
"When IPv4 packets are translated (as specified in RFC 7915), the IPv6 header
is 20 octets larger than the IPv4 header in the best case; however, the overhead
is not fixed. As per RFC 7915 Section 4.1, a Fragment Header (8 octets) is additionally
inserted when the incoming IPv4 packet is already a fragment, or when the DF bit
is clear and the resulting IPv6 packet would exceed the configured lowest-ipv6-mtu.
Moreover, IPv4 options, if present, contribute further overhead (hence RFC 7915
Section 1.4 uses "20+ octets"). Operators provisioning headroom for MTU
should therefore account for up to 28 octets of translation overhead, plus
any IPv4 options that may be present."
In addition, I will give a concrete number as below,
“A concrete recommendation is to provision an IPv6-only core MTU of at
least 1548 bytes (1500-byte IPv4 payload + 40-byte IPv6 header +
8-byte Fragment Header) to allow a 1500-byte IPv4 packet to traverse without
fragmentation in the common translation case.”
642 translation handling. It should be noted that the IPv4 data delivery
643 across IPv6-only network specified in [RFC7915] has already been
644 tested on some large-scale networks, such as CERNET, and there are no
645 major issues observed. In addition, section 4.1 and section 5.1 of
646 [RFC7915] give translators a SHOULD-level option to zero the TOS/
<minor> Three problems in one sentence. RFC 7915 specifies a translation
algorithm, not IPv4 data delivery across an IPv6-only network - what was
deployed on CERNET is a deployment, not a specification. "No major issues
observed" carries no reference and no scope, and it directly undercuts the
three sentences before it, which describe MTU differences between domains as
a fundamental operational concern for any IPv4-over-IPv6 framework. And
CERNET is a single-operator network (but I could be wrong?), so it is not
evidence about the multi-domain case that is this document's whole subject.
Either scope the claim honestly - for example, that stateless translation
per RFC 7915 has been deployed at scale within single operators, while the
multi-domain case in this document has not been operationally validated -
or delete the sentence. The paragraph is stronger without it. Does any
implementation of the framework exist, and has any inter-provider deployment
been attempted or planned? I realize one of the authors is from CERNET
and they will correct me; if I were to name deployments, I would do it
as a separate section in the appendix and give far more details than just
one such blanket statement.
[Chongfeng]: RFC7915 is a translation mechanism, but it does support IPv4
service delivery over IPv6-only. Precisely speaking, RFC7915 based IPv4-as-a-Service
is a double-translation mechanism, one RFC7915-based PE is responsible for
translation of IPv4 to IPv6 for deliver over IPv6 network, the other RFC7915-based PE
is responsible for translation of IPv4 to IPv6 to resort IPv4 packets.
Yes, CERNET is a single-operator network. To clarify the confusion caused by this, we will remove the
following sentence,
“It should be noted that the IPv4 data delivery
across IPv6-only network specified in [RFC7915] has already been
tested on some large-scale networks, such as CERNET, no major issues observed”
<nit> "SHOULD-level" is an all-capitals keyword occurrence in a document
whose Section 1.1 says such occurrences are to be interpreted per BCP 14.
Perhaps "a normatively recommended configuration option". The citation is
also a little loose: RFC 7915 Section 4.1 says a translator SHOULD support a
configurable option to ignore the IPv4 TOS and set the IPv6 traffic class to
zero, while Section 5.1 says the ability to set the IPv4 TOS octet to a
specified value, not necessarily zero.
[Chongfeng]:
The term of “SHOULD” appears in the following text of section 7,
“In addition, section 4.1 and section 5.1 of [RFC7915] give translators a
SHOULD-level option to zero the TOS/Traffic Class octet entirely on
translation ("administrative bleaching"), which silently kills ECN if an operator
enables it for DSCP hygiene.
Based on the previous discussion (https://mailarchive.ietf.org/arch/msg/v6ops/pwd9BTDc5Efp3nQnhKo3NsKnIlk/)
since these issues can be addressed by RFC7915-bis or other documents, it will
not be covered by this draft, so the text will be removed in the next version.
651 One potential risk specific to multi-domain is Path MTU Discovery
652 reliability, an ICMPv6 Type 2 (Packet Too Big) message from a P
653 router in one IPv6 carrier's AS has to reach the originating IPv4
654 host across administrative boundaries with independent ICMP-filtering
655 policies. In this case, IPv4 hosts may not receive "Packet Too Big"
656 notifications, and will keep trying to send large packets. These
<major> This treats cross-domain ICMP filtering as the risk and assumes the
message otherwise reaches the originating IPv4 host. It cannot, as written,
because nothing says who converts it.
In translation mode, a P router in the core drops an oversized IPv4-embedded
IPv6 packet and emits ICMPv6 Packet Too Big. Its destination is the IPv6
source of the offending packet, so it routes back to the ingress PE - but
its source is the P router's own IPv6 address, which in an IPv6-only core
has no IPv4 representation at all. To hand an ICMPv4 message to the IPv4
host, the ingress PE must translate ICMPv6 to ICMPv4 per RFC 7915 Sections
5.2 and 5.3 and synthesise an IPv4 source address for it - the case RFC 7915
Section 6 points to RFC 6791 for. The document never says ingress PEs do
this, never mentions RFC 6791 or any pool for it, and Section 5.3 describes
the ingress PE only as a producer of IPv4-embedded IPv6 packets. In
encapsulation mode the problem appears differently: RFC 2473 Section 8 puts
the relay obligation on the tunnel entry-point node, which must recover the
inner IPv4 header from the ICMPv6 payload. Also uncited.
The consequences are the PMTUD drops this paragraph itself describes,
arriving by a route it does not consider, and that traceroute and ICMP-based
troubleshooting across the framework do not work - which is worth weighing
against the motivation in Section 1.
Please state that PEs implement the full RFC 7915 behaviour including ICMPv6
to ICMPv4 translation for errors generated inside the core, that an RFC 6791
pool or equivalent is required at each PE for the ICMPv4 source address, and
the RFC 2473 Section 8 equivalent for encapsulation. Then rewrite the
paragraph so the filtering concern sits on a described mechanism rather than
an assumed one.
This all will add to the normative references.
[Chongfeng]: Based on the comments above, the following sentences will be added in section 7,
“This issue has been mentioned in RFC7915 section 7. In this framework,
ingress PEs must implement the complete behavior specified in RFC 7915.
To hand an ICMPv4 message to the IPv4 host, the ingress PE must
translate ICMPv6 to ICMPv4 per RFC 7915 Sections 5.2 and 5.3 and
synthesise an IPv4 source address for it — a case for which RFC 7915
Section 6 points to RFC 6791.”
Also, the following sentences will be removed,
“To address this issue, it is recommended to coordinate cross-domain
ICMP policies. The two operators need to negotiate and establish
management-level Service Level Agreements (SLAs) or security
policy exceptions to ensure that critical ICMPv6 Type 2 messages
are allowed to traverse the administrative boundary in both directions.
For example, ICMP traffic originating from the trusted AS of the peer
can be rate-limited and deeply inspected before being permitted, rather than
being completely blocked.”
In addition, RFC6791 will be added as a normative reference.
Encapsulation related efforts will not be discussed here.
713 When the capability to advertise address mapping rules via BGP is
714 introduced, attackers may alter the IPv6 mapping prefix within these
715 rules, leading to improper delivery of IPv4 service traffic over an
716 IPv6-only network. Such an attack differs from pre-existing
717 vulnerabilities in that traffic could be forwarded to a remote target
718 across an intervening network infrastructure (e.g., an IPv6 core),
719 allowing an attack to potentially succeed more easily since less
720 infrastructure needs to be compromised. To mitigate this risk,
721 [I-D.ietf-sidrops-moa-profile] proposes an approach by leveraging
722 RPKI [RFC6480] architecture to verify the authenticity of the address
723 mapping rule associated with an IPv4 address block.
<major> This identifies a security risk that the framework creates, states
that it is easier to exploit than the pre-existing case, and then offers
exactly one mitigation for it: a draft in SIDROPS (which also brings the
question whether the scope is the Internet or some limited domain(s)).
There is no second mitigation for this risk here - the paragraph that
follows is about the default rule, a different exposure.
So the document is on a fork. Either the mitigation is load-bearing, in
which case the reference is normative; or it is not, in which case the
document has identified a novel security risk and left it unmitigated.
The wording "proposes an approach" is what keeps the reference off the
normative list, and it is not a resolution.
There is a third path and I think it is the right one. State in Sections 5.2
and 8.3 that any conforming mapping-rule distribution mechanism must provide
origin authentication for mapping rules, and that a PE must not accept a
mapping rule binding an IPv4 address block to a mapping prefix belonging to
a PE not authorized for that block. That is a property statement - the same
thing my first general comment asks Section 5.2 to supply - and it is
implementable by bilateral filtering or provisioning as well as by RPKI.
The there is no need to reference draft-ietf-sidrops-moa-profile (same as
no need to reference the BGP spec). This connects with my DISCUSS on
Section 8.1: source validation at the egress PE and origin validation of the
mapping rule are two halves of one control, and the document requires
neither.
Separately, RFC 8950 should be a normative reference. Section 4.2 states a
deployment dependency on it - mapping rules cannot propagate unless the
intermediate BGP speakers support it - which is a condition for the
framework working rather than background. But again, it is possible to keep
BGP and routing out of this document and avoid this.
[Chongfeng]: I adopt the third path, BGP and routing will be moved out of
this document, so the text you pointed out will be removed, draft-ietf-sidrops-moa-profile
will not be referenced, either. However, in this document, state it directly as property
requirements in section 5.2, as follows,
1) The distribution mechanism MUST support origin authentication. Specifically, any
distribution mechanism that conforms to this framework and is used to distribute
mapping rules should be capable of verifying the authentic origin of the mapping
rules, thereby ensuring the legitimacy and trustworthiness of their source.
2)PE devices should reject unauthorized mapping rules. Specifically, when a mapping
rule attempts to bind an IPv4 address block to a mapping prefix that is not within the
authorized scope of that address block, the PE device must detect such a violation and
explicitly reject the mapping rule.
728 obvious implications, such as traffic attraction (posing a DoS
729 concentration risk) and becoming a "catch-all" hijack target if rule
730 distribution is compromised. To mitigate these issues, the default
731 rule MUST only be used within a single administrative domain unless
732 explicit bilateral policy exists. Even in those cases, its use is
733 OPTIONAL. Network providers can consider applying monitoring and
734 rate-limiting/ACLs on the default egress PE, and avoid exporting the
735 default rule across inter-domain boundaries.
<major> Setting aside the placement, which I raise on Section 5.2 above,
these three sentences do not cohere.
The placement of "only" leaves it open whether the restriction is on where
the default rule may be used or on what it may be used for. "Even in those
cases, its use is OPTIONAL" has two available antecedents, and OPTIONAL adds
nothing, since nothing preceding it made the default rule mandatory - so it
reads as retracting the MUST rather than qualifying it. The INTDIR early
review flagged this ambiguity. And "avoid exporting the default rule across
inter-domain boundaries" restates, at advisory strength, the constraint that
the MUST has just imposed.
Perhaps:
The default address mapping rule is optional. Where it is used, it
SHOULD NOT be advertised across an administrative boundary unless the
participating network providers have an explicit bilateral agreement
covering its use. Network providers using a default egress PE should
apply monitoring and rate-limiting or ACLs on that PE.
It is also worth asking whether the document wants a MUST here at all.
[Chongfeng]: Yes, the current phrasing is a bit twisted and inconsistent. Based
on your analysis and suggestion, the following change will be made,
OLD:
“To mitigate these issues, the default rule MUST only be used within
a single administrative domain unless explicit bilateral policy exists.
Even in those cases, its use is OPTIONAL. Network providers
can consider applying monitoring and rate-limiting/ACLs on the default
egress PE, and avoid exporting the default rule across inter-domain boundaries.”
NEW:
“Where default address mapping rule is used,
it should not be advertised across an administrative boundary unless the
participating network providers have an explicit bilateral agreement covering
its use. Network operators using a default egress PE should apply monitoring
and rate-limiting or ACLs on that PE.”
No, “MUST” is not wanted by this informational document, “SHOULD” (upper case) isn’t,
either. Nevertheless, I’d like to hear your opinion on this.
I also understand that a one-time response is unlikely to address all your concerns, so I look
forward to continuing our discussion. Thank you.
Best regards
Chongfeng
On behalf of all the co-authors
<EoRv27>
_______________________________________________
v6ops mailing list -- v6ops@ietf.org
To unsubscribe send an email to v6ops-leave@ietf.org
- [v6ops] Ketan Talaulikar's Discuss on draft-ietf-… Ketan Talaulikar via Datatracker
- [v6ops] Re: Ketan Talaulikar's Discuss on draft-i… Chongfeng Xie
- [v6ops] Re: Ketan Talaulikar's Discuss on draft-i… Ketan Talaulikar
- [v6ops] Re: Ketan Talaulikar's Discuss on draft-i… Ketan Talaulikar
- [v6ops] Re: Ketan Talaulikar's Discuss on draft-i… Chongfeng Xie
- [v6ops] Re: Ketan Talaulikar's Discuss on draft-i… Chongfeng Xie
- [v6ops] Re: Ketan Talaulikar's Discuss on draft-i… Chongfeng Xie
- [v6ops] Re: Ketan Talaulikar's Discuss on draft-i… Ketan Talaulikar