Re: [certid] Review of draft-saintandre-tls-server-id-check-12

[Peter Saint-Andre <[e-mail]>]

I just realized that we never replied publicly. Jeff and I had a phone
chat with Cullen (and Alexey) about this before the holidays, and we
plan to submit a revised I-D this week. Cullen raised some very good
points, which we've attempted to address in the forthcoming version.

On 12/16/10 8:22 AM, Peter Saint-Andre wrote:
> Thanks for your comments. My co-author and I will need to confer before
> replying, and that might take a few days given the length of your review.
> 
> Peter
> 
> On 12/16/10 12:17 AM, Cullen Jennings wrote:
>>
>> So let me start with I think there is great information in here and I
>> think it should be published as a standards track RFC however I do
>> think there are some issues with the relation with this draft and the
>> realities of what would help improve security in deployment of SIP,
>> HTTP, IMAP, XMPP etc.
>>
>> There are many places where this draft makes choices to improve the
>> security from many current practices. At face value this seems like a
>> good thing but it's not always. The thing reducing the overall
>> security available to users of TLS is not if certs use CN-ID or
>> DNS-ID, it is that it's such a PITA to deploy a TLS  server that
>> people choose to not use TLS at all. Everywhere there is a trade off
>> between making things marginally more secure, or making things
>> cheaper and easer to deploy, I think we need to seriously consider
>> the cheaper and easier approach. Yes, some things are just broken
>> even if they are easy and obviously we can't do those. Let me give an
>> example of this. I looked at the cert for the domains for the authors
>> of this draft. www.cisco.com has 3 DNS names even though as fas as I
>> can tell one of these are for something that would typically be used
>> in a ftp URI and the other HTTP URI. This is because it makes it
>> easier and cheaper for them to use TLS yet seems to go against the
>> recommendation of this "BCP". Then I went over the www.paypal.com
>> domain which uses, gasp, a CN-ID. Do we really believe that paypal is
>> seriously compromising their security by using a CN instead of
>> URI-ID? If so, how? I'm pretty sure the paypal guys know how to run a
>> secure web server. With the exception of Microsoft small business
>> server certificates (which are outrageously expensive by the way) it
>> pretty hard to get SRV name certs. In making these recommendations,
>> did the TLS WG consider the relative prices of various types of
>> certificates? Lets say I had a certificate for the domain example.org
>> because I was using it for email and it has a CN because I got it
>> years ago. Now lets say I am going to go deploy a SIP service on
>> example.org. My position is that best way to encourage the use of
>> security on the internet is to just reuse the certificate I have. It
>> cheap, it's easy, it secure enough even if it does make you feel a
>> bit dirty. I think Jeff disagrees w ith me, we argued for years about
>> this topic and my understanding is his position is that it would be
>> better to say that all new deployments MUST not use a CN name because
>> it's less secure. Give the prevalence of CN on the internet today, I
>> think it is fine to tell people how DNS-ID is better but I don't
>> think it's OK to tell them they should not use CN-ID and I definitely
>> don't think it is OK to tell implementors they don't need to
>> implement CN-ID.
>>
>> I encouraged Chris to write this draft long ago and what we had
>> discussed at the time was forming a RFC with one or more boiler plate
>> pieces of text that could be used in creation of the name matching
>> section of new protocols that got developed. I was thinking of
>> something similar to the way we use rfc 5226 for writing IANA
>> consideration section. Instead we have a document that is creating a
>> very complex situations about whats normative. This draft is a BCP
>> level, and it says you have to do everything in PKIX and PKIX takes
>> precedence. That is basically elevating PKIX to a BCP without
>> appropriate process review. Next this draft contradicts the
>> procedures in existing protocols and says that it does not apply to
>> the existing protocols but that it would take precedence over any
>> future updates of existing protocols that use TLS within the scope
>> specified here. I do not believe you have the consensus of the people
>> woking on SIP that the next time some specification is marked as an 
>> UPDATE to 3261, that implementations need to implement the procedures
>> in this draft. Furthermore, I think that would be counter productive.
>> I think you should make it clear this guidelines for designers of new
>> protocol and people updating existing protocol and that these
>> protocols could make their own choices but would want to take into
>> account the information in this draft. When I read the sentence, 
>> "However, the best current practices described here can be referenced
>> by future specifications, including updates to specifications for
>> existing application protocols if the relevant technology communities
>> agree to do so." I think that is exactly the right solution to the
>> problem. However, that not a BCP, thats a standards track spec.
>> Furthermore, I think this draft is going to have all the normal bugs
>> etc of any other spec that defines algorithms and such it should
>> proceed through the standards track process. If this draft is going
>> to go as a BCP, that text contradicts what a BCP is and needs to come
>> out and the rest of the draft be adjusted appropriately. My
>> preference would be that this draft be standards track. Its writing
>> exactly the same sort of normative algorithm text that we put in all
>> kinds of other thing like SIP, HTTP, and even TLS. They are all
>> standards track. This should be the same.
>>
>> Most RFCs today that use TLS have a page plus or minus that tells an
>> implementer what they need to know about matching names in certs.
>> This draft move that to 30 to 50 pages depending on how you count.
>> Most implementers are just going to ignore this thus worsening the
>> security situation. Think about why is the part implementers need to
>> read 10x longer than existing deceptions - this just seems wrong. Now
>> it's easy to blow off this type concern and say get over it, it's the
>> same number of lines of code they need to write. But the problem is
>> when an implementors goes to start doing this and encounters
>> something that is 50 pages long, they instantly decide this is a big
>> task and down it goes on the priority list of actually happening. The
>> other problem is that even thous it is long, it is still very
>> confusing on how to do things (such at URI). I'll provide more
>> detailed examples of this later in this email. If the document was
>> restructured to have all the normative text in one s hort simple
>> description and the rest moved to an appendix, it would be much
>> easier to get people to take this seriously and easier to review that
>> it was correct.
>>
>> My final big issue is the use of normative language. Lets say there
>> are two procedures A and B and we 100% consensus that B is better
>> than A but we still have to support A for existing deployment
>> reasons. To describe this, the text this draft would use is is MUST
>> do A and SHOULD NOT do B. Now reading 2119 it is pretty clear that
>> SHOULD NOT means you don't implement it unless there are real good
>> reasons to implement it. So on the things were we agree A is
>> preferred to B but you need both for backwards compatibility, this
>> draft needs to say MUST implement A and MUST implement B but
>> deployments are encouraged to use B as we are trying to move away
>> from A. I think the whole document needs a careful read checking for
>> this issue. You can insert the usual rant here about why SHOULD is a
>> awful word in specs 90% of the time it used because implementer
>> thinks it means "ignore rest of sentence". For example,  section 5.4
>> discusses they this spec continues to mandate protocols MUST suppo rt
>> CN yet this draft continually use "SHOULD NOT" when what it really
>> means is MUST implement. This is going to confuse implementors of
>> IETF specification and be ignored by operators. Given the goals of
>> this spec it would be much better if it was clearly defining what
>> IETF required implementers of protocols to do instead of confusing
>> that with how we wish security was deployed.
>>
>>
>> On to the nits.
>>
>> Take an applications like a web server. Is the preferred thing in a
>> cert a DNS-ID or a URI-ID. My reading of the 3.3 is that URI-ID is
>> preferred over DNS-ID yet the examples don't match that. I think
>> point 3 in section 3.1 tries to explain this away but I don't
>> understand that - clearly web browsers use a URI.
>>
>> The rules in section 3.1 don't make sense for a CA. How will the CA
>> know if the cert I want is going to be used for a protocol that uses
>> SRV?
>>
>> In section 3.2, in the imap example, you are saying that if I
>> configure my imap server to mail.example.com and it presents a
>> certificate with a DNS-ID of example.com that this is OK. That does
>> not sound OK to me but I don't know how IMAP works. In the SIP
>> example, the cert should have a SRV and DNS name too. As well as a CN
>> if you actually want it to work in the real world.
>>
>> In section 4.2.1 you have a long discussion on how the information
>> used must come from a way that can't be tampered with over the
>> internet. I generally agree with this but would like to point out
>> that protocols like LOST (see section 18 of rfc5222) specially do the
>> opposite of this and actually match the cert agains the output of
>> NAPTR process not the input to the NAPTR process.
>>
>> The example just seem plain wrong in some cases. Take for example
>> section 4.2.2 where the SIP example has only a URI reference
>> identifiers and no DNS yet the section right before this has said the
>> list MUST include a DNS-ID. This text has been through how many
>> reviews and Last Calls? The problem here is that this draft is too
>> long to review for stuff like this. Even after the IESG is done
>> reviewing it, statistics suggest it will still be littered with bugs
>> like this and implementors will use the examples to guide them. If
>> someone implements what is in the example, it will break in lots of
>> sip deployments.
>>
>> There is a whole algorithm about matching various ID types, but the
>> note about you ignore CN if you have other things is off in "Security
>> Warning" very much out of any flow of the algorithm description then
>> pointed out again in some other section. It's not wrong, but it's a
>> bit weird from an implementer point of view.
>>
>> Many applications do need to deal with IP matching as well as domain
>> names. The way this text is written here makes it harder to figure
>> out how and where to mix that in. I'd rather see it just dealt with
>> than instead of making it out of scope. Obviously it's not common on
>> internet but it is common on private networks and walled gardens
>> where many of the protocols were are talking about are deployed. Many
>> of the "internet of things" people I work with have no intention of
>> using DNS at all. I scoffed at multiple large service providers 10
>> years ago when they said they were not using DNS with SIP but many
>> still use IPs. This sounds less insane when you consider the major OS
>> don't ship with an asynchronous DNS library.
>>
>> I'm baffled on why checking the service name in a SRV record is a
>> SHOULD not a MUST. Could you add text explain why and when one would
>> not check it. If you were in a really good mood you could do that for
>> all the SHOULDs. Actually, when I read section 4.5 carefully, I think
>> it literally says that when using a URI, checking the domain name is
>> a SHOULD not a MUST. Surely check the domain name matches is not a
>> SHOULD level sort of thing.
>>
>> Section 5.4. I have no idea why it matters that some major OS does
>> not support SNI. Even if that OS did support SNI, many many
>> applications running on that OS and the others would not support SNI.
>> It seems like it is the applications acting as TLS servers and
>> clients that are the important thing, not the OS.
>>
>> How you process URI-ID needs work. I could not figure out how to
>> implement given the text in the draft as is. Even ignoring the
>> special tar pit the SIP guys dug for themselves with tel URL
>> processing, just the normal sip, sips issues seems unclear.
>>
>> This seems like a long list of complaints delivered fairly late but,
>> once again, I really do like much of the information in here and
>> think it should be published as PS - it just would be significantly
>> improved with a bit of a re-factored and clean up. If this had been
>> run through the TLS working group, I would have caught all of this in
>> the WG LC. This is a draft that, as a BCP,  profoundly effects many
>> of the protocols I work on including SIP but as far as I can tell has
>> not done much to gather the consensus of the people working on
>> protocol that this draft changes - I don't recall hearing about it
>> until after it went to the IESG so I'm pretty un apologetic about
>> providing these comments during IETF LC.
>>
>> In summary, I like the information in this but I think it still has
>> many small things that need fixing and needs to be changed to get
>> crisp about what implementors need to do and drop the confusing stuff
>> about how we wish operators and CA might use certificates. I also
>> feel strongly that the right way to look at this draft is, as that
>> draft says "practices described here can be referenced by future
>> specifications, including updates to specifications for existing
>> application protocols if the relevant technology communities agree to
>> do so" and that for that reason it has to be standards track not BCP.
>> If it was not being written and pushed by two IESG members, I don't
>> think we would even be discussing if it should be a BCP.
>>