Hi Bob,
Thanks for the review. We authors will have to go though your comments in details. It is unfortunate that you did not review HTTP Datagram specification (RFC 9297) prior as that is allowing sending of datagrams over HTTP/3. In my initial read through of the comments I think most still applies as they are not directly related to if the tunneled IP are in datagrams (unreliable in delivery order) or in streams (reliable in order delivery). I do know some of the issues you bring up have been discussed but was not documented. We will have to consider if the choice to not document them should be changed or not.
Cheers
Magnus
From: Bob Briscoe via Datatracker <noreply@xxxxxxxx>
Date: Wednesday, 29 March 2023 at 09:04
To: tsv-art@xxxxxxxx <tsv-art@xxxxxxxx>
Cc: draft-ietf-masque-connect-ip.all@xxxxxxxx <draft-ietf-masque-connect-ip.all@xxxxxxxx>, last-call@xxxxxxxx <last-call@xxxxxxxx>, masque@xxxxxxxx <masque@xxxxxxxx>
Subject: Tsvart last call review of draft-ietf-masque-connect-ip-08Reviewer: Bob Briscoe
Review result: Not Ready
This document has been reviewed as part of the transport area review team's
ongoing effort to review key IETF documents. These comments were written
primarily for the transport area directors, but are copied to the document's
authors and WG to allow them to address any issues raised and also to the IETF
discussion list for information.
When done at the time of IETF Last Call, the authors should consider this
review as part of the last-call comments they receive. Please always CC
tsv-art@xxxxxxxx if you reply to or forward this review.
-----------------------------------------------------------------------
==Overall comments==
How can I say this politely? The HTTP aspects of this document are great work
and near-complete. However, the tunnelling aspects, in particular the layering
interactions need a lot of work - possibly even a partial redesign in places.
Also, the security aspects seem to have been viewed through rose-coloured
spectacles; sthg like "Most people who like privacy are nice".
Well-written, but not an easy read in a single pass (not clear what the point
is for quite a few pages and lots of forward x-references). Early overview of
the scheme would help. Need to bear in mind that cross-layer work has to be
addressed at a wider community who don't necessarily all subscribe to the same
mind-share.
-----------------------------------------------------------------------
==Document Structure==
I believe §§3-4 are about the tunnel config, and §§5-7 are the per-packet aspects of the tunnel. Might be worth explaining that in a doc roadmap.
EDITORS: That's a reasonable editorial ask, we'll look into it as editors and see what we can do.
§6 "Payload Format"
This section needs to be split and/or re-titled, 'cos in the middle (from "Client MAY optimistically start sending...") it switches to operation, not format. At "Note that endpoints will decrement..." it switches to TTL handling. Then at "IPv6 requires that.." it switches to MTU discovery, which is another aspect of operation.
Just because TTL and packet size are fields of an IP packet, doesn't mean these aspects fall under the heading of "Payload Format" - these paras are about handling these fields, not defining their format.
EDITORS - We agree, filed an issue to track adding a new Performance Considerations section:
https://github.com/ietf-wg-masque/draft-ietf-masque-connect-ip/issues/144
-----------------------------------------------------------------------
==Gaps==
===A) No discussion of packet ordering ===
Although one of the variants of HTTP (http/3) is over UDP, which is unordered, QUIC still delivers an ordered stream to the app (which, in this case, is the inner of the tunnel). So, presumably as packets arrive at the proxy, if there has been misordering or a loss, QUIC will buffer until the next packet in order can be delivered, i.e. head-of-line blocking. Hence, it's not really advisable to forward connectionless packets through a connection-oriented tunnel. The privacy benefit might make the poor latency performance worth it. But this trade off at least needs to be stated, perhaps in an applicability section.
I don't think an http3 stream is any less granular than an http2 stream or an http1.1 connection in this proxy arrangement. But if it is, each need to be discussed separately.
If there is a vision to use an unordered variant of QUIC, then there could be problems with Context IDs being processed out of order, which could produce all sorts of unexpected side-effects, given Context IDs are potentially stateful (caveat: you will see that I don't really understand what Context IDs might be used for).
EDITORS: As discussed, this section is incorrect.
===B) Conflicting Congestion Controllers (CCs) ===
Needs a section sthg like §12.1 of RFC8229, except:
That text mostly just listed the problems, but the implication of publishing such a section in a spec is essentially, "We've told you the problems, but now go ahead anyway." This is not just a TCP in TCP problem. It's any nesting of two or more CC-controlled protocols (e.g. TCP within QUIC, QUIC within QUIC, etc)
As above, if this is a trade off between privacy and performance, that needs stating.
EDITORS: This makes sense, we'll add some text. Tracking in this issue:
https://github.com/ietf-wg-masque/draft-ietf-masque-connect-ip/issues/145
===C) Overlapping connection life-cycles ===
If a proxy crashes, powers down, etc. can it be brought back up with all the connection state restored that it held before it failed? Otherwise, all the connections within the tunnel inner depend on the survival of their outer connection. There is no reason to believe that all the inner connections are dispensable.
Would a soft-state design be more resilient than the hard-state design described?
EDITORS: This is true of all tunnels, for example IKEv2/IPsec. Luckily here we can use ADDRESS_REQUEST to recreate the same state and avoid connections breaking
[The above points (A-C) are examples of the problems I mentioned when using a connection-oriented protocol to tunnel a connectionless protocol.]
===D) Applicable Next Headers (aka IP Protocols) ===
No discussion of whether some IPv6 'Next Headers' (or IPv4 Protocols) are not applicable / unsafe / poor performing.
The last time that I proposed a generic UDP tunnel to the Intarea as an alternative to GUE, the criticism was that it was "too generic". I was told that the choice of tunnel protocols is about whether they are sufficiently limiting; the aim is not totally generic.
EDITORS: This document is about creating an IP tunnel over HTTP so we have a tight scope based on the underlying protocol instead of on the high layer protocols.
=== E) Applicable Address Ranges ===
Are anycast addresses in scope? IP multicast?
Link-local multicast is mentioned. What about ARP? Service Discovery?
EDITORS: We've documented the address types that require implementation involvement. Anycast and (non-link-local) unicast are equal to unicast as far as a single link is concerned. ARP operates at L2 so it doesn't apply to an L3 tunnel. Similarly, service discovery protocols don't need any special guidance.
===F) Propagation of IP headers and header fields between inner and outer. ===
Some IPv6 extension headers are copied from inner to outer on encapsulation; others aren't [RFC2473; §5.1]. The outer IPv6 flow label is often zeroed. The contents of the DSCP, and ECN fields can be propagated from inner to outer and the ECN outer is propagated back to the forwarded header [RFC2983], [RFC6040].
Because there's a whole stack of headers between the inner and outer IP headers (e.g. IP, UDP, QUIC, HTTP, IP), the draft needs to make it clear that these copying and propagation rules still apply between the two IP headers.
EDITORS: This is a tunnel operating at L7, not an L3/L4 encapsulation format. Copying these fields doesn't make sense because it is possible to have multiple IP packets inside a single IP packet.
===G) No explanation of pros & cons relative to other packet tunnelling ===
Yes, we can tunnel anything over anything, but why this one? AFAIK, it's to hide traffic in the HTTP crowd for privacy. But what are the merits relative to hiding in other crowds? And the drawbacks (e.g. HoL blocking as mentioned above). In particular, the pros & cons vs UDP in HTTP.
EDITORS: There are many motivations for tunneling IP over HTTP, one of them being that it allows traveling through HTTP load balancers. This is discussed in our charter and doesn't need to be added to the document.
===H) Role reversal? ===
It's not made clear whether there's anything intrinsically 'clientish' about a client. Or 'proxyish' about a proxy. Especially in a bump in the wire topology (like the site-site VPN example), does it matter which is which? Is it just determined by the administrator setting a proxy up in listen mode? Is there anything that one or the other cannot do by virtue of its role, other than initiate a connection to the other one? Shouldn't the draft speak about this if it requires correct manual set up? Particularly for tunnels consisting of chains (or trees?) of intermediaries.
EDITORS: The terms "client" and "proxy" are HTTP terminology. The client sends the request, the proxy responds. Once the tunnel is up, the draft talks about endpoints and peers because then it's symmetric.
-----------------------------------------------------------------------
==Section by Section==
1. Introduction (and Abstract)
"...updates RFC9298"
It's not at all clear to me what aspect of RFC9298 this RFC updates. It seems unlikely, given IP Proxying is an alternative to UDP Proxying, not an update. But if it does, the draft needs to say what it updates. And, if RFC9298 were updated by this draft, it would surely be a normative reference, not informative.
EDITORS: Fair enough, we'll add text. Tracked in this issue:
https://github.com/ietf-wg-masque/draft-ietf-masque-connect-ip/issues/146
3. Configuration of Clients
Reasoning for each each URI template rule?
I can't really judge whether this set of rules leaves the URI template flexible enough, without knowing the reason for each constraint. Is this explained in "Proxying UDP in HTTP"? If so, a ref would be useful. "IP proxy deployments SHOULD offer service at this location if they need to interoperate with such clients." Surely "MUST" to interoperate?
EDITORS: That's right, the MUST was to simplify interoperating. We made these decisions for CONNECT-UDP as a WG and then decided to keep the same ones for CONNECT-IP.
4. Tunnelling IP over HTTP
"When sending its IP proxying request, the client SHALL perform URI template expansion..." Surely just "... the client performs..." (just a necessary process step, not an interop requirement)
EDITORS: We disagree, making this normative is clearer. This matches our editorial style in 9298.
"IP proxying requests [responses] do not carry any message content." Is it an error if they do?
EDITORS: The stream is taken over by the capsule protocol. Data on the stream is not message content as the data stream has been taken over. See 9297 for details.
4.1. IP Proxy Handling
"...the IP proxy MUST perform DNS resolution..." Surely "... the IP proxy performs..." (just a necessary process step, not an interop requirement)
EDITORS: We disagree, making this normative is clearer. This matches our editorial style in 9298.
"IP proxies MAY choose to tear down the tunnel due to a period of inactivity" Better? "IP proxies MAY choose to tear down the tunnel, e.g. due to a period of inactivity" Rationale: there are other possible reasons, e.g responding to an attack.
EDITORS: It's already clear that the proxy can tear down the tunnel, this bit is about lifetime
"Any response other than a successful response indicates that the request has
failed; thus, the client MUST abort the request." This is repeated at each instance in the subsequent sections. Either not needed here as well, or not needed in each section as well.
EDITORS: This isn't repeated, they're separate requirements.
4.6. Limiting Request Scope
"When the IP proxy knows that a request is scoped to a target prefix or protocol, it can leverage this information to optimize its resource allocation..." This highlights that the protocol relies on a perverse incentive for the client to scope its request so that the server can optimize its resources. That smells like potentially the opposite of resource optimization, i.e. a vulnerability; clients can scope their requests to force a server to frag its resources. Perhaps consider redesigning the protocol so that the server initiates the scoping (or at least it can somehow negotiate the outcome)?
EDITORS: The server is always able to abort the tunnel if it doesn't want to spend resources. The protocol as currently designed is what we need in practice, your proposed redesign is not.
"target: ...the IP proxy is expected to perform DNS resolution ..." Might this potentially amplify a resource exhaustion vulnerability, esp. if the client continually gives the proxy DNSSEC work?
"IP proxies MAY perform access control using the scoping information provided by the client: if the client is not authorized to access any of the destinations included in the scope, then the IP proxy can immediately fail the request." I think the clause after ':' ought to be tagged as an example. I think it's intended to be a degenerate/extreme example. Whatever, its relation with the earlier words is unclear. It's possible this would be better in §4.1, which is where it would fit into the process steps. However, at that point in the draft, scope limiting hasn't been explained. I'm in two minds.
EDITORS: I agree with you that this change doesn't really improve things, so we'll keep them as-is.
4.7. Capsules
Are all control messages to the tunnel on one stream, or can each capsule arrive in a separate stream, or is it up to the operator? IOW, do capsules have to be acted on in the order they are sent, or does it depend?
EDITORS: Please read RFC 9297 § 3. Capsules are sent on the one (only) stream, and are sent in order. Capsule protocol does not require they be processed in order, but in almost all cases should be.
§4.7.1 The last para starting "Note that the IP forwarding tunnels described in this document..." seems more like an architectural point that ought to be stated earlier. It's related to ADDRESS_ASSIGN, but is it really appropriate to bury it here?
EDITORS: Agreed, tracked in this issue:
https://github.com/ietf-wg-masque/draft-ietf-masque-connect-ip/issues/144
§4.7.2
IP version:
Address assignment syntax presumes that two request prefixes will return two addresses. I think this means that an endpoint cannot request just one address, when it doesn't mind whether it's within a certain IPv4 address range or a certain IPv6 address range. In the happy eyeballs example, this returns two addresses. Similarly, I don't think the syntax can express no preference for a prefix length. Only a specific prefix length is possible. Correct? Should these limitations of the syntax be pointed out?
EDITORS: If a client wants exactly one address, but doesn't care which address family, it should send an ADDRESS_REQUEST for v4 as 0.0.0.0/32, and if it does not get one assigned, then send a request for v6 as ::/128 (or other mask). (This algorithm can be run in the reverse order as well) If it wants _at least_ one address of either address family, it should send a single ADDRESS_REQUEST asking for both, and simply not use the assigned address of the address family it is uninterested in.
§4.7.3
"...the most recently received ROUTE_ADVERTISEMENT capsule..." I assume this means most recently received after reordering has been corrected. See my first point under Capsules about control message ordering.
EDITORS: Please read 9297. These capsules are on the reliable stream and are delivered in order.
§5 Context Identifiers
The point below is in a similar vein to the earlier point about whether generic support of IP Protocols (Next Headers) is a boon or a bane...
This section seems like it's early research. It's vague, with no examples and no clear purpose (to me). So one cannot reason about what security vulnerabilities it might open up (let alone what flexibility it potentially affords). There is no citation of [HTTP-DGRAM] in this section, nor explanation of the purpose of a Context ID.
EDITORS: The first paragraph explains the rationale. This feature was discussed at length in the WG and we decided to add context IDs to both CONNECT-UDP and CONNECT-IP.
"The Context ID value of 0 is reserved for IP payloads, " This clearly means that non-IP payloads cannot use 0, but does it also mean that IP payloads cannot use non-zero values? It's not clear, but none of the examples (that all use IP) show a non-zero Context ID. A later para (2nd para under Fig 13 in §6) seems to more unambiguously define a 1:1 mapping between IP and Context ID=0. If so, I don't understand the point of a Context ID in these tunnelling cases. What am I missing?
EDITORS: Context IDs are registered and can be for any purpose. Context ID 0 is pre-registered for full IP payloads. Future extensions can perform registration to create new context IDs with different payload types
"it is possible for datagrams to be received with Context IDs that have not yet been registered. For instance, this can be due to reordering of the packet containing the datagram and the packet containing the registration message during transmission." So..., what happens then?
EDITORS: Datagrams referring to context ids that are not yet registered can either be buffered or dropped by the receiver, at their discretion. We don't provide specific guidance because this behavior will be specified by extensions that use the context ID and that guidance might vary between extensions.
6. HTTP Datagram Payload Format
"Context ID:
A variable-length integer that contains the value of the Context ID. If an HTTP/3 datagram which carries an unknown Context ID is received, the receiver SHALL either drop that datagram silently or buffer it temporarily (on the order of a round trip) while awaiting the registration of the corresponding Context ID."
A number of problems here:
* undefined what happens if an http/2 or http/1.1 datagram carries an unknown Context ID?
* For TCP, the app layer doesn't know the RTT.
EDITORS: This won't happen in HTTP/2 or HTTP/1.1, because everything is delivered in order by TCP.
* Isn't this a resource exhaustion vulnerability? the sender contrives packets with Context IDs that cause the receiver to buffer everything, exhausting the memory available for other senders.
EDITORS: We're still bound by TCP flow control. Additionally, the receiver is allowed to drop so it'll do that.
"Endpoints MAY implement additional filtering policies on the IP packets they forward." Eh? This is the first mention of filtering policies.
EDITORS: Agreed, this is the first mention. Implementers of IP tunnels are familiar with the concept.
7. Error Signalling
For a bump in the wire topology, can an ICMP error be sent back, from one IP side to the other, through the HTTP tunnel and out the other side? That will probably work. However, presumably an ICMP error raised by the network within the extent of the HTTP tunnel will be directed back to the ingress tunnel endpoint but no further.
EDITORS: CONNECT-IP conveys ICMP inside the tunnel only and is not responsible for any ICMP messages on the outer connection. Existing handling rules apply and are out of scope of this document.
8.4. Proxied Connection Racing
It might be worth pointing out in a note at the end of this example that, as well as the proxy deciding to set up parallel tunnels, the approach would also allow the client to initiate parallel tunnels.
EDITORS: We don't see value in adding this note to this example. The proposed note is a completely different example. The proxy cannot set up parallel tunnels; only the client issuing CONNECT-IP can set up a tunnel. The right-hand side are not tunnels in this example.
10. Security Considerations
I find it quite lame/naïve that this section essentially says, "This Proxy design gives arbitrary clients great power, both to toast proxies and to make them appear to be the senders of arbitrary messages to arbitrary destinations" BUT, "That's not a problem, because a proxy SHOULD restrict its use to authorized users."
This massive "get out of jail free" card begs the question, "What management
processes are necessary for this proxy to monitor users to determine whether to remove their authorization?" It's akin to saying, "We're going to issue automatic assault rifles to everyone over the age of 6, but it's not a problem, because it's illegal to shoot someone without a permit."
EDITORS: Your comment is both ableist and violent. Please rephrase this to be professional and actionable and then we'll discuss more.
-----------------------------------------------------------------------
==Nits==
EDITORS: We've looked through these nits and made minor changes to the document for the ones we agree with.
"HTTP provides the CONNECT method (see Section 9.3.6 of [HTTP]) for creating a
TCP [TCP] tunnel..."
Citation of IETF core protocols like [TCP] is unnecessary. And definitely
not normative - this sentence is informatively comparing this draft with a
TCP-based alternative.
§3
s/an URI/a URI/
(unless we're meant to pronounce URI like Cockney "In an 'urry luv?" :)
§4
Para 1: "To allow negotiation..."
Para 3: "To initiate an IP tunnel..."
I think these two paras repeat each other. But para 3 is better: a) initiate is
a better word than negotiate (AFAICT there is no negotiation), and b) it
explains that it's a /single/ HTTP stream.
"IPv6 scoped addressing zone identifiers..." ref? [RFC6874]?
§4.7.3
Using 'IP Protocol' rather than 'Next Header' as the keyword within an IP
Address Range is rather politically incorrect for IPv6 zealots, isn't it? :)
Is 'equal than' (2 occurrences) correct in US English? It's certainly not in
British. I can't find any reference to it.
§6
s/since receiving addresses and routes is/
/since receipt of addresses and routes is/
"When an endpoint receives an HTTP Datagram containing an IP packet, it..."
a tunnel endpoint?
MTU sometimes ought to be PMTU.
"HTTP Datagrams with payloads of at least 1280 bytes"
Fig 13 potentially labels two things as HTTP Datagram Payload. Which is meant
here? I assume the inner one, but it needs to be clear.
"...the endpoints can pad the QUIC INITIAL packets of the underlying QUIC
connection that IP proxying is running over. (Assuming QUIC version 1 is in
use,..." The parenthesis needs to be tagged as an example.
§8.1
In the format defined in §1.3 of [QUIC] and used in the examples, a value after
'=' is normally numeric. the following format threw me:
Payload = Encapsulated IP Packet
I thought that must imply a field announcing that the type of the payload is
the string "Encapsulated IP Packet". Any chance this could be bracketed off
somehow, e.g.
Payload = <Encapsulated IP Packet>
Is there any precedent in RFC9000 for an unquoted/unbracketed description of a
binary blob on the RHS of '='?
§8.4
"The IP proxy assigns the client both an IPv4 address (192.0.2.3) and an
IPv6 address (2001:db8:1234::a) to the client."
Delete first occurrence of 'the client'.
§10
s/IP proxies SHOULD restrict its use to authenticated users./
/IP proxies SHOULD restrict their use to authorized users./
Reason: as well as the incorrect singular, following the literal rule as
originally written would give access to authentic criminals.
§11.1
There's a few places in the IANA section where editor's marks might be useful
to identify text that will need to be updated once approved.
Cheers
Bob
-- last-call mailing list last-call@xxxxxxxx https://www.ietf.org/mailman/listinfo/last-call
