[Freeipa-devel] Generic support for unknown DNS RR types (RFC 3597)

[Petr Spacek]

On 10.3.2015 20:04, Simo Sorce wrote:
> On Tue, 2015-03-10 at 19:24 +0100, Petr Spacek wrote:
>> On 10.3.2015 18:36, Simo Sorce wrote:
>>> On Tue, 2015-03-10 at 18:26 +0100, Petr Spacek wrote:
>>>> On 10.3.2015 17:35, Simo Sorce wrote:
>>>>> On Tue, 2015-03-10 at 16:19 +0100, Petr Spacek wrote:
>>>>>> On 10.3.2015 15:53, Simo Sorce wrote:
>>>>>>> On Tue, 2015-03-10 at 15:32 +0100, Petr Spacek wrote:
>>>>>>>> Hello,
>>>>>>>>
>>>>>>>> I would like to discuss Generic support for unknown DNS RR types (RFC 3597
>>>>>>>> [0]). Here is the proposal:
>>>>>>>>
>>>>>>>> LDAP schema
>>>>>>>> ===========
>>>>>>>> - 1 new attribute:
>>>>>>>> ( <OID> NAME 'GenericRecord' DESC 'unknown DNS record, RFC 3597' EQUALITY
>>>>>>>> caseIgnoreIA5Match SYNTAX 1.3.6.1.4.1.1466.115.121.1.26 )
>>>>>>>>
>>>>>>>> The attribute should be added to existing idnsRecord object class as MAY.
>>>>>>>>
>>>>>>>> This new attribute should contain data encoded according to ​RFC 3597 section
>>>>>>>> 5 [5]:
>>>>>>>>
>>>>>>>> The RDATA section of an RR of unknown type is represented as a
>>>>>>>>    sequence of white space separated words as follows:
>>>>>>>>
>>>>>>>>       The special token \# (a backslash immediately followed by a hash
>>>>>>>>       sign), which identifies the RDATA as having the generic encoding
>>>>>>>>       defined herein rather than a traditional type-specific encoding.
>>>>>>>>
>>>>>>>>       An unsigned decimal integer specifying the RDATA length in octets.
>>>>>>>>
>>>>>>>>       Zero or more words of hexadecimal data encoding the actual RDATA
>>>>>>>>       field, each containing an even number of hexadecimal digits.
>>>>>>>>
>>>>>>>>    If the RDATA is of zero length, the text representation contains only
>>>>>>>>    the \# token and the single zero representing the length.
>>>>>>>>
>>>>>>>> Examples from RFC:
>>>>>>>>       a.example.   CLASS32     TYPE731         \# 6 abcd (
>>>>>>>>                                                ef 01 23 45 )
>>>>>>>>       b.example.   HS          TYPE62347       \# 0
>>>>>>>>       e.example.   IN          A               \# 4 0A000001
>>>>>>>>       e.example.   CLASS1      TYPE1           10.0.0.2
>>>>>>>>
>>>>>>>>
>>>>>>>> Open questions about LDAP format
>>>>>>>> ================================
>>>>>>>> Should we include "\#" constant? We know that the attribute contains record in
>>>>>>>> RFC 3597 syntax so it is not strictly necessary.
>>>>>>>>
>>>>>>>> I think it would be better to follow RFC 3597 format. It allows blind
>>>>>>>> copy&pasting from other tools, including direct calls to python-dns.
>>>>>>>>
>>>>>>>> It also eases writing conversion tools between DNS and LDAP format because
>>>>>>>> they do not need to change record values.
>>>>>>>>
>>>>>>>>
>>>>>>>> Another question is if we should explicitly include length of data represented
>>>>>>>> in hexadecimal notation as a decimal number. I'm very strongly inclined to let
>>>>>>>> it there because it is very good sanity check and again, it allows us to
>>>>>>>> re-use existing tools including parsers.
>>>>>>>>
>>>>>>>> I will ask Uninett.no for standardization after we sort this out (they own the
>>>>>>>> OID arc we use for DNS records).
>>>>>>>>
>>>>>>>>
>>>>>>>> Attribute usage
>>>>>>>> ===============
>>>>>>>> Every DNS RR type has assigned a number [1] which is used on wire. RR types
>>>>>>>> which are unknown to the server cannot be named by their mnemonic/type name
>>>>>>>> because server would not be able to do name->number conversion and to generate
>>>>>>>> DNS wire format.
>>>>>>>>
>>>>>>>> As a result, we have to encode the RR type number somehow. Let's use attribute
>>>>>>>> sub-types.
>>>>>>>>
>>>>>>>> E.g. a record with type 65280 and hex value 0A000001 will be represented as:
>>>>>>>> GenericRecord;TYPE65280: \# 4 0A000001
>>>>>>>>
>>>>>>>>
>>>>>>>> CLI
>>>>>>>> ===
>>>>>>>> $ ipa dnsrecord-add zone.example owner \
>>>>>>>>   --generic-type=65280 --generic-data='\# 4 0A000001'
>>>>>>>>
>>>>>>>> $ ipa dnsrecord-show zone.example owner
>>>>>>>> Record name: owner
>>>>>>>> TYPE65280 Record: \# 4 0A000001
>>>>>>>>
>>>>>>>>
>>>>>>>> ACK? :-)
>>>>>>>
>>>>>>> Almost.
>>>>>>> We should refrain from using subtypes when not necessary, and in this
>>>>>>> case it is not necessary.
>>>>>>>
>>>>>>> Use:
>>>>>>> GenericRecord: 65280 \# 4 0A000001
>>>>>>
>>>>>> I was considering that too but I can see two main drawbacks:
>>>>>>
>>>>>> 1) It does not work very well with DS ACI (targetattrfilter, anyone?). Adding
>>>>>> generic write access to GenericRecord == ability to add TLSA records too,
>>>>>> which you may not want. IMHO it is perfectly reasonable to limit write access
>>>>>> to certain types (e.g. to one from private range).
>>>>>>
>>>>>> 2) We would need a separate substring index for emulating filters like
>>>>>> (type==65280). AFAIK GenericRecord;TYPE65280 should work with presence index
>>>>>> which will be handy one day when we decide to handle upgrades like
>>>>>> GenericRecord;TYPE256->UriRecord.
>>>>>>
>>>>>> Another (less important) annoyance is that conversion tools would have to
>>>>>> mangle record data instead of just converting attribute name->record type.
>>>>>>
>>>>>>
>>>>>> I can be convinced that subtypes are not necessary but I do not see clear
>>>>>> advantage of avoiding them. What is the problem with subtypes?
>>>>>
>>>>> Poor support by most clients, so it is generally discouraged.
>>>> Hmm, it does not sound like a thing we should care in this case. DNS tree is
>>>> not meant for direct consumption by LDAP clients (compare with cn=compat).
>>>>
>>>> IMHO the only two clients we should care are FreeIPA framework and
>>>> bind-dyndb-ldap so I do not see this as a problem, really. If someone wants to
>>>> access DNS tree by hand - sure, use a standard compliant client!
>>>>
>>>> Working ACI and LDAP filters sounds like good price for supporting only
>>>> standards compliant clients.
>>>>
>>>> AFAIK OpenLDAP works well and I suspect that ApacheDS will work too because
>>>> Eclipse has nice support for sub-types built-in. If I can draw some
>>>> conclusions from that, sub-types are not a thing aliens forgot here when
>>>> leaving Earth one million years ago :-)
>>>>
>>>>> The problem with subtypes and ACIs though is that I think ACIs do not
>>>>> care about the subtype unless you explicit mention them.
>>>> IMHO that is exactly what I would like to see for GenericRecord. It allows us
>>>> to write ACI which allows admins to add any GenericRecord and at the same time
>>>> allows us to craft ACI which allows access only to GenericRecord;TYPE65280 for
>>>> specific group/user.
>>>>
>>>>> So perhaps bind_dyndb_ldap should refuse to use a generic type that
>>>>> shadows DNSSEC relevant records ?
>>>> Sorry, this cannot possibly work because it depends on up-to-date blacklist.
>>>>
>>>> How would the plugin released in 2015 know that highly sensitive OPENPGPKEY
>>>> type will be standardized in 2016 and assigned number XYZ?
>>>
>>> Ok, show me an example ACI that works and you get my ack :)
>>
>> Am I being punished for something? :-)
>>
>> Anyway, this monstrosity:
>>
>> (targetattr = "objectclass || txtRecord;test")(target =
>> "ldap:///idnsname=*,cn=dns,dc=ipa,dc=example")(version 3.0;acl
>> "permission:luser: Read DNS Entries";allow (compare,read,search) userdn =
>> "ldap:///uid=luser,cn=users,cn=accounts,dc=ipa,dc=example";)
>>
>> Gives 'luser' read access only to txtRecord;test and *not* to the whole
>> txtRecord in general.
>>
>> $ kinit luser
>> $ ldapsearch -Y GSSAPI -s base -b
>> 'idnsname=txt,idnsname=ipa.example.,cn=dns,dc=ipa,dc=example'
>> SASL username: luser at IPA.EXAMPLE
>>
>> # txt, ipa.example., dns, ipa.example
>> dn: idnsname=txt,idnsname=ipa.example.,cn=dns,dc=ipa,dc=example
>> objectClass: top
>> objectClass: idnsrecord
>> tXTRecord;test: Guess what is new here!
>>
>> Filter '(tXTRecord;test=*)' works as expected and returns only objects with
>> subtype ;test.
>>
>> The only weird thing I noticed is that search filter '(tXTRecord=*)' does not
>> return the object if you have access only to an subtype with existing value
>> but not to the 'vanilla' attribute.
>>
>> Maybe it is a bug? I will think about it for a while and possibly open a
>> ticket. Anyway, this is not something we need for implementation.
>>
>>
>> For completeness:
>>
>> $ kinit admin
>> $ ldapsearch -Y GSSAPI -s base -b
>> 'idnsname=txt,idnsname=ipa.example.,cn=dns,dc=ipa,dc=example'
>> SASL username: admin at IPA.EXAMPLE
>>
>> # txt, ipa.example., dns, ipa.example
>> dn: idnsname=txt,idnsname=ipa.example.,cn=dns,dc=ipa,dc=example
>> objectClass: top
>> objectClass: idnsrecord
>> tXTRecord: nothing
>> tXTRecord: something
>> idnsName: txt
>> tXTRecord;test: Guess what is new here!
>>
>>
>> And yes, you assume correctly that (targetattr = "txtRecord") gives access to
>> whole txtRecord including all its subtypes.
>>
>> ACK? :-)
>>
> 
> ACK.

Thank you. Now to the most important and difficult question:
Should the attribute name be "GenericRecord" or "UnknownRecord"?

I like "GenericRecord" but Honza prefers "UnknownRecord" so we need a third
opinion :-)

> Make sure it is abundantly clear in the docs what is the implication of
> giving access to the generic attribute w/o qualifications.

Sure.

-- 
Petr^2 Spacek