Re: [Hipsec] rfc5201-bis-04 review

[Ari Keranen <[e-mail]>]

Hi Tobi,

On Mar 13, 2011, at 9:19 PM, Tobias Heer wrote:
> Hello Ari,
> 
> ok, this took a while. Finally I am through. Most of your comments did not require second thoughts. On some I made more detailed comments. Please check if I always got your point or if I misunderstood something.

OK, see inline (clipping away parts that seem to be clear by now).

> Also. A BIG thank you for providing alternative text when appropriate. That did speed up addressing the comments quite much. Excellent work!

You're welcome!

> Am 03.02.2011 um 18:47 schrieb Ari Keranen:
>>    This control is set in packets R1 and/or
>>    I2.  The peer receiving an anonymous HI may choose to refuse it.
>> 
>> How about using anonymous HIs with other packets? Say, NOTIFYs or UPDATEs?
> 
> It doesn't make much sense because the anonymous HI is only important for end-hosts when they bootstrap a new HIP association.

So, if one sends HOST_ID in NOTIFY, even if it's the same anonymous HOST_ID one used in the BeX (with the A-bit), the host would not set the A-bit here? This sounds just a bit inconsistent. Would it be easier to just say that the A-bit is (always) set in packets where the HI is anonymous? That is, not say anything about R1/I2.

> For middleboxes, things may be different because after an UPDATE a middlebox may suddenly be on the primary communication path between two HIP hosts that previously established a connection.
> 
> However, middleboxes are only addressed to a minimal degree in BEX. The semantics and function of the HIT for middleboxes is undefined so far. If we start adding things for middleboxes here we might have to delve much deeper.
> 
> Any opinion?

I'd leave middlebox considerations, as long as we're not breaking anything, for another draft.

>> 5.1.3.  HIP Fragmentation Support
>> 
[...]
>> HIP-aware NAT
>> devices MUST perform any IPv4 reassembly/fragmentation.
>> 
>> Despite the warning in the next paragraph, this would potentially make all HIP-aware IPv4 NAT devices vulnerable for fragment-attacks. Could change the MUST into SHOULD? Also, could clarify that reassembly/fragmentation is a MUST/SHOULD for HIP packets, not for any random packet.
> 
> It should definitely be a SHOULD. The box may or may not be useful without fragmentation support. BEX is not the document to decide that.

Agree.

> I also agree that only HIP packets are to be considered. However, the distinction may be difficult if the fragments arrive out of order. How will the box figure out if the fragment is a HIP control packet or not if the HIP header is in another fragment?

True, but I'd leave that up to the box. That is, if/when it can figure it out, it can stop to attempt reassembly.

>> 
>> it is strongly recommended that the separate document
>> specifying the certificate usage in the HIP Base Exchange defines the
>> usage of "Hash and URL" formats rather than including certificates in
>> exchanges.
>> 
>> This already exists in the CERT draft, so this text could be removed/re-worded. The reference to fragmentation DoS attack is still useful though.
>> 
> Alternative text:
> 
>   Certificate chains can cause the packet to be fragmented and
>   fragmentation can open implementations to denial-of-service attacks
>   [KAU03].  "Hash and URL" schemes as defined in [I-D.ietf-hip-cert]
>   for HIP version 1 may be used to avoid fragmentation and mitigate
>   resulting DoS attacks.
> 
> The implications and security considerations for certficate chains and hash and url are discussed there. Hence, I see no reason to put a MUST or SHOULD here.

Sounds good to me.

>> 
>> 5.2.  HIP Parameters
>> 
>> The HIP Parameters are used to carry the public key associated with
>> the sender's HIT, together with related security and other
>> information.
>> 
>> Considering the amount of different things carried in HIP Parameters today, this section could start with a more generic description of the use of the parameters ("and other information" sounds a bit understatement). Can't come up with the right words for now, but would be good to have something that would explain that Parameters are used to carry the security stuff, but are also one of the main ways for extending HIP.
> 
> Let me try ;-)
> 
>   The HIP parameters carry information that is necessary for
>   establishing and maintaining a HIP association.  For example, the
>   peer's public keys as well as the signaling for negotiating ciphers
>   and payload handling are encapsulated in HIP parameters.  Additional
>   information, meaningful for end-hosts or middleboxes, may also be
>   included in HIP parameters.  The specification of the HIP parameters
>   and their mapping to HIP packets and packet types is flexible to
>   allow HIP extensions to define new parameters and new protocol
>   behavior.
> 
>   In HIP packets, HIP parameters are ordered according to their numeric
>   type number and encoded in TLV format.
> 

Much better.

>> 
>> | ECHO_RESPONSE_SIGNED   | 961   | variable | Opaque data echoed    |
>> |                        |       |          | back; under signature |
>> 
>> s/echoed back/echoed back by request/
>> 
> Do you think this is necessary? To me it is obvious and pronounced in the rest of the text.

Not necessary, but IMHO more clear that way. BTW, -05 revision has now different text (regarding "by request") for SIGNED and UNSIGNED.


>> 
>> |  2048 -  4095 | Parameters related to HIP transport types         |
>> 
>> s/transport types/transport formats/
>> (to be consistent with the later text)

Missed this one?

>> 5.2.1.  TLV Format
>> 
>> The type-field value also describes the order of these
>> fields in the packet, except for type values from 2048 to 4095 which
>> are reserved for new transport forms.
>> 
>> This is a bit strange exception. I wonder if it would make sense to make this consistent with the other ordered lists, i.e., have one parameter that lists the supported transport formats and then, depending on which format was chosen, the corresponding parameter(s) from this range are used (and others ignored).
> 
> I recently sent this issue to the list and unless there are any objections, I'll add the following.
> 
> I won't be able to get this done before the next version but it will be the first thing for the next revision.
> 
> Option: Give all transforms different type numbers, keep the ordering and express preference in a list.
> 
> Example HIP packet excerpt:
> +------+ +------+---------+ +------+-----+ +------+-----+
> |Header| | 2050 | XYZ, ESP| | 2095 | ESP | | 2096 | XYZ | ...
> +------+ +----------------+ +------+-----+ +------+-----+
>              ^
>              |
> List----------+
> 

Makes sense.

>> 5.2.2.  Defining New Parameters
>> 
>>     Implementations operating in a mode
>>     adhering to this specification MUST disable the sending of new
>>     critical parameters.
>> 
>> This sounds a bit strange. Should that be "MUST allow disabling"?
>> 
> I think the original meaning was that you have to enable it by hand and it is disabled by default. This makes sense because otherwise standard compliant implementations will be incompatible by default. To make it clearer I changed the sentence to:
> 
>               Implementations operating in a mode adhering to this
>               specification MUST disable the sending of new critical
>               parameters by default.
> 

That's better.   

>> 
>> 5.2.6.  DIFFIE_HELLMAN
>> 
>> The 160-bit ECP
>> group can be used when lower security is enough (e.g., web surfing)
>> and when the equipment is not powerful enough (e.g., some PDAs).
>> 
>> Should mention which Group ID the 160-bit ECP is (I'd assume 10, but it's not explicit).
>> 
> I rephrased the sentence to mention the group name explicitly. This should clarify it.
> 
>           A HIP implementation MUST implement Group ID 3. The 160-bit
>           SECP160R1 group can be used when lower security is enough ...
> 

That works too.

>> Also, some general recommendation on the order of proposed group IDs would be helpful so that implementations don't end up (unnecessarily) proposing the weakest groups by default.
>> 
> Hmmm.. I think the words of caution above should avoid that. Otherwise, I could write something like:
> 
> In general stronger groups SHOULD be preferred.
> 
> Should we explicit state the order here? Stronger groups are of course more costly and need more space in the packets. So I wouldn't want to propose group 9 as default.

I think the important thing is to tell what is the order of strength between the groups. Intuitively one could assume that the higher the value the stronger the group, but clearly that's not the case with SECP160R1.

>> 
>> 5.2.7.  HIP_CIPHER
>> 
>>   Cipher ID      defines the cipher algorithm to be used for
>>                  encrypting parts of the HIP packet
>> 
>> Is HIP_CIPHER used anywhere else except in the ENCRYPTED parameter? If not, the explanation above could say that explicitly. That is, for example,
>> s/parts of the HIP packet/the contents of the ENCRYPTED parameter/
>> 
> Nope it is nowhere else (at least in this document). However other documents might use it for something else.
> But well, these documents will define it more explicitly anyway.
> 
> Agreed, changed.

Makes sense.

>> 
>> NULL-ENCRYPTION is included
>> for testing purposes.  NULL-ENCRYPTION SHOULD NOT be configurable via
>> the UI.
>> 
>> There may be no UI for the implementation, so could say "by the user" instead of "via the UI".
>> 
> There is (almost) always a UI, right? Maybe its no GUI, maybe its not even a command line tool... but any interface that allows the user to interact is a UI, right. Changing it to "by the user" implies the existence of some UI as well.

I would not consider a config file as "user interface". Or is it OK to config NULL encryption via config file (but not via "UI")? 

>> 
>> 5.2.8.  HOST_ID
>> 
>>  DI-type           type of the following Domain Identifier field
>>  Domain Identifier the identifier of the sender
>> 
>> The actual content of these fields is not clear from the text. Could at least move the DI-types table closer to the figure. Also, "DI Length" could simply say "length (in octets) of the Domain Identifier field"; the FQDN and NAI part just confuses here.
>> 
> I moved it closer and adjusted the order of the field explanations to match the order of the fields in the parameter.
> I hope things are clearer now.

Yes (although the "DI Length" and "Algorithm" field/explanation order doesn't still match). 

>> 
>> If there is no Domain Identifier, i.e., the DI-type field is zero,
>> the DI Length field is set to zero as well.
>> 
>> Should there be some normative text on when to have the domain identifier?
> It is zero if the HIT does not belong to any domain, I guess. I am not fully aware of the reasons that led to the inclusion of the NAI and FQDN here. From my perspective it is some nice-to-have information that adds some weak hierarchy to the HOST_ID (weak because it is not a input for the actual HI or HIT.
> 
> Maybe someone can shed some light? (I will post this as separate mail on the list so that it is not lost in the bulk of the comments and replies here).

Good idea.

>> 5.2.13.  HIP_SIGNATURE
>> 
>> The signature algorithms are defined in Section 5.2.8.
>> 
>> The algorithm identifier field side for HOST_ID (Section 5.2.8) is 16 bits whereas here it is 8 bits. Something wrong here?
> 
> Woops. That's still a remnant of 5201 (no bis). It should be changed, of course. It won't bloat the signature to use a 16 bit field. On the other hand, we're limited in the numbers of algorithms anyway so 256 seems kind of okay. Well, I'll go for the 16 bit field in HIP_SIGNATURE if no one produces valid reasons not to do this.
> 
> This has to be carefully documented in the changelog - otherwise implementers will struggle with this hidden change quite a lot.

Sounds good to me.

>> 
>> Could also reference to exact table with the algorithms since section 5.2.8 has 4 different tables.
> I added a introductory sentence to the table in 5.2.8. I hope this does the trick.

Yes, that helped.

>> Notification information can be error messages specifying why an SA
>> 
>> First occurrence of SA. Expand, and maybe add a reference too.
>> 
> I think this refers to the HIP security association, not IPsec. In that regard, the text is blurry (I'll change that) but does not require a reference.
> 
> The new text is:
> 
>           Notification information can be error messages specifying
>           why an HIP Security Association could not be established.
>           It can also be status data that a HIP implementation
>           wishes to communicate with a peer process.  The table
>           below lists the Notification messages and their
>           corresponding values.

That's better.

> 
>> 
>> Overall, this section is fairly inconsistent with the use of different ways to refer to "Notify Message Type", some suggestions for change:
>> 
>> The table below lists the Notification messages and their
>> corresponding values.
>> 
>> s/Notification messages/Notify Message Types/
>> 
> suggestion: The table
>           below lists the notification messages and their
>           Notification Message Types.
> 

Sounds good.

>> An
>> implementation that receives a NOTIFY packet with a NOTIFICATION
>> error parameter in response to a request packet
>> 
>> s/NOTIFICATION error parameter/(something)/
>> 
> 
> An implementation that receives a
>           NOTIFY packet with a Notify Message Type that indicates an
>           error in response to a request packet ...
> 
> is this appropriate for (something)?

Yes!

>> 
>>   UNSUPPORTED_CRITICAL_PARAMETER_TYPE        1
>> 
>>     Sent if the parameter type has the "critical" bit set and the
>>     parameter type is not recognized.  Notification Data contains the
>>     two-octet parameter type.
>> 
>> What if there are multiple unsupported critical parameter types? Should send multiple NOTIFICATIONs of this Message Type or would it be better to concatenate the types?
> 
> There may be several notify parameters in one HIP packet in that case. I spelled it out explicitly in the introduction:
> 
>           HIP packets MAY contain
>           multiple notify parameters if several problems exist or
>           several independent pieces of information must be transmitted.
> 

OK, that helps. Just s/notify/NOTIFICATION/

>> 
>> The sender has to create the Opaque field
>> so that it can later identify and remove the corresponding
>> ECHO_RESPONSE_UNSIGNED parameter.
>> 
>> Does this apply also to middleboxes, i.e., is "sender" the sender of the packet or sender of the parameter?
>> 
> The creator of the parameter. I'll state this more clearly.
> 
>           The creator of the
>           ECHO_REQUEST_UNSIGNED (end-host or middlebox) has to
>           create the Opaque field so that it can later identify and
>           remove the corresponding ECHO_RESPONSE_UNSIGNED parameter.
> 

That's better.
 
>> 
>> 5.3.1.  I1 - the HIP Initiator Packet
>> 
>>   IP ( HIP ( DH_GROUP_LIST ) )
>> 
>> This is not strictly consistent with the notation definition "X(y)   indicates that y is a parameter of X" (HIP stuff is not really a parameter of IP packet). I'm not entirely sure how to fix this though (or if it even needs to be fixed).
>> 
>> 
> 
> I think it ti rather clear here. if it were IP(HIP) I could imagine that someone might be confused. All other forms of braces are used up anyway ([{}])

Agreed; no point changing this.


>> 
>> 5.3.2.  R1 - the HIP Responder Packet
>> 
>> The Initiator's HIT MUST match the one received in the I1 packet.
>> 
>> ...if the R1 is sent due to I1.
>> 
> The Initiator's HIT MUST match the one received in the I1
>           packet if the R1 is a response to an I1.
> 

That's good.

>> 
>> If a future version of this protocol is considered, we strongly
>> recommend that these issues be studied again.
>> 
>> Has anyone looked into this?
>> 
> I thought about it for a while but did not come up with any good conclusion.

Perhaps this also deserves a separate thread and/or discussion at the WG meeting.

>> 
>> The signature is calculated over the whole HIP envelope
>> 
>> "HIP envelope" is not defined anywhere (but used many times after this occurrence too). Perhaps add it to the definitions part.
>> 
>> 
> We can substitute it with HIP packet. That is more technical and more precise.

OK, that helps.

> 
>> 5.3.3.  I2 - the Second HIP Initiator Packet
>> 
>> The HITs used MUST match the ones used previously.
>> 
>> This is ambiguous. Should that be "used in R1" or are there also other possibilities?
>> 
>> 
> Nope. Changed.
> 
> (Damn... Man, this review is sooo much work. I am not even close to the end. Thanks for doing this in such great detail!)

You're welcome :)

>> 5.3.5.  UPDATE - the HIP Update Packet
>> 
>> An UPDATE that does not contain a SEQ parameter is
>> simply an acknowledgment of a previous UPDATE and itself MUST NOT be
>> acknowledged by a separate ACK.
> 
> The text above seems shaky in itself. The update should nonetheless contain an ACK because otherwise it is not clear which SEQ it acknowledges. I'll change that.
> 
>> 
>> What about UPDATE without ACK nor SEQ? Unreliable UPDATE or a protocol error?
>> 
> I changed the text to:
> 
>           The UPDATE packet contains zero or one SEQ parameter.  The
>           presence of a SEQ parameter indicates that the receiver
>           MUST acknowledge the the UPDATE.  An UPDATE that does not
>           contain a SEQ but only an ACK parameter is simply an acknowledgment of a
>           previous UPDATE and itself MUST NOT be acknowledged by a
>           separate ACK. UPDATE packets without SEQ parameters do not
>           require an acknowledgment and do not require processing in
>           relative order to other UPDATE packets.
> 
> Satisfied?

That's more clear. Although, I'm not sure what would you use "unreliable UPDATEs" for. If you don't care about reliability, that sounds more like a notification, and hence using NOTIFY packet(s) would make more sense, right? I could be missing a use case though.

[...]
>> 6.4.1.  HMAC Calculation
>> 
>> 6.  Set Checksum and Header Length field in the HIP header to
>>     original values.
>> 
>> This could be just an implementation issue, but unless the signatures and MACs are also restored, the length and checksum are invalid after this step. Is this step even needed?
>> 
> Well.. this is a hint for the implementors so that they don't deal with a corrupted packet. I guess its fine.

OK. Could make sense to explicitly make it "an implementor hint" then. Something like (to the end) "Note that the checksum and length field contain incorrect values after these steps". Not sure if that's more clear; I'll leave it up to you. 

>> 
>> 6.5.  HIP KEYMAT Generation
>> 
>> 
>>    HIP-gl encryption key for HOST_g's outgoing HIP packets
>>    HIP-gl integrity (HMAC) key for HOST_g's outgoing HIP packets
>> 
>> It's a bit confusing to use the same "key name" twice (same for -lg). I assume "outgoing HIP packets" means the ENCRYPTED parameter? If so, could say it explicitly.
> 
> Changed it to make ENCRYPTED more obvious but I am not sure what you mean by "same key name twice"... can you explain?

I assume that you draw here two (well, 4 if you see the whole text) different keys -- which both have the name "HIP-gl". Or is that just a single key?

[...]
>> 
>> 6.7.  Processing Incoming I1 Packets
>> 
>> 2.  If the Responder is in ESTABLISHED state, the Responder MAY
>>     respond to this with an R1 packet, prepare to drop existing SAs,
>> 
>> Should prepare to drop only the SAs one has with the I1's sender?
>> 
> New text:
>             If the Responder is in ESTABLISHED state, the Responder
>             MAY respond to this with an R1 packet, prepare to drop
>             an existing HIP security association with the peer, and
>             stay at ESTABLISHED state.

That's better.

>> 
>>     Other than that, selecting the HIT is a
>>     local policy matter.
>> 
>> Shouldn't the Responder use the same as HIT the Initiator used for it? This would otherwise break section 6.8 step 3.
> Actually that is stated two sentences before the clause you cited:
> 
>             If the received Responder's HIT in the I1 is NULL, the
>             Responder selects a HIT with a the same HIT Suite as the
>             Initiator's HIT.  If this HIT Suite is not supported by
>             the Responder, it SHOULD select a REQUIRED HIT Suite
>             from <xref target="hit_suite_list"/>, which is currently
>             RSA/DSA/SHA-1. Other than that, selecting the HIT is a
>             local policy matter.
> 
> The "Other than that, selecting the HIT is a local policy matter" is for the case if several HITs (same suite) are possible.

OK, I guess I just missed that one.

[...]
>> 
>> 6.15.  Processing CLOSE_ACK Packets
>> 
>> A host can map CLOSE messages to CLOSE_ACK messages by using
>> the included ECHO_RESPONSE_SIGNED in response to the sent
>> ECHO_REQUEST_SIGNED in the CLOSE_ACK.
>> 
>> s/using/matching/
>> And the end of the sentence seems to be incorrect anyway (REQ is not in ACK). Could rephrase the whole sentence.
>> 
> 
> New text:
> 
>         A host can map CLOSE messages to
>         CLOSE_ACK messages by using the included
>         ECHO_RESPONSE_SIGNED that was sent in the CLOSE_ACK packet
>         in response to the ECHO_REQUEST_SIGNED in the CLOSE packet.

That's already better, but still a bit confusing. How about something like:

A host can map CLOSE_ACK messages to CLOSE messages by comparing the value of ECHO_REQUEST_SIGNED (in the CLOSE packet) to the value of ECHO_RESPONSE_SIGNED (in the CLOSE_ACK packet).

[...]
>> 7.  HIP Policies
>> 
>> Initiators may use a different HI for different Responders to provide
>> basic privacy.  Such implementations SHOULD keep state for mapping
>> Initiator's HIT to Responder's HIT.  This access control list (ACL)
>> SHOULD also include preferred transform and local lifetimes.
>> 
>> This sounds a bit ambiguous. Why is the mapping needed? What are the lifetimes? Why store the preferred transform?
>> 
> Okay... The text was blurry and weird. I have some equally blurry but clearer alternative here. Blurry because this is really local policy, not HIP.
> 
>       Initiators may use a different HI for different Responders to
>       provide basic privacy. Whether such private HITs are used
>       repeatedly with the same Responder and how long these HITs are
>       used is decided by local policy and depends on the privacy
>       requirements of the Initiator.

Much better. Although, now you have both "HI" and "HITs". I'd use only one consistently in the text (even though they have 1:1 mapping).

>> 
>> The value of K used in the HIP R1 packet can also vary by policy.  K
>> should never be greater than 20, but for trusted partners it could be
>> as low as 0.
>> 
>> s/should never/SHOULD NOT/
>> Also, why 20? Some rough number on "how much is 20" could be helpful. How about, say, 10 years from now, is 20 that big still then or should we comment something on adjusting the maximum value in the future? Or will it be HIP v3 by then? :)
> 
> I would remove the text on 20 altogether. The 20 makes assumptions on the capabilities of the clients and the attacker, which we should not do. The 20 is just an unjustified magic number. The new text says:
> 
>       The value of K used in the HIP R1 must be chosen with care.
>       Too high numbers of K will exclude clients with weak CPUs
>       beause these devices cannot solve the puzzle within reasonable
>       time.  K should only be raised if a Responder is under high
>       load attack, i.e., it cannot process all incoming HIP
>       handshakes any more. If a responder is not under high load, K
>       SHOULD be 0.

That sounds reasonable.

>> 10.  IANA Considerations
>> 
>> This document also creates a set of new namespaces.  These are
>> described below.
>> 
>> Do we create new namespaces for v2 or should we re-use the existing ones?
>> 
> We create new ones because the parameter ranges and the parameter definitions change. So I guess no textual change is needed here.

OK. 

By the way, we seem to have default policy of "IETF Review" for the namespaces. I'd recommend changing that into "IETF Review or IESG Approval", at least for the namespaces that have plenty of space in them (8 bits or more). That way also non-IETF (e.g., IRTF) RFCs or docs from other SDOs can be given values if it makes sense.


Cheers,
Ari