[Freeipa-devel] Certificate Identity Mapping

Mon Jan 2 07:06:04 UTC 2017

On 19.12.2016 12:13, Sumit Bose wrote:
> On Mon, Dec 19, 2016 at 10:02:58AM +0100, Jan Cholasta wrote:
>> I agree with *almost* everything Sumit said. See my inline comments below.
>>
>> On 16.12.2016 11:53, Sumit Bose wrote:
>>> On Tue, Dec 06, 2016 at 04:39:10PM +0100, Florence Blanc-Renaud wrote:
>>>> Hi,
>>>>
>>>> I have started a feature description for the Certificate Identity Mapping at
>>>> the following location:
>>>> http://www.freeipa.org/page/V4/Certificate_Identity_Mapping
>>>>
>>>> This is a first step, focusing on the interface we would like to provide. It
>>>> still contains open questions, some of which are linked to the corresponding
>>>> design on SSSD side:
>>>> https://fedorahosted.org/sssd/wiki/DesignDocs/MatchingAndMappingCertificates
>>>> https://fedorahosted.org/sssd/wiki/DesignDocs/SmartcardsAndMultipleIdentities
>>>>
>>>> Comments, concerns and suggestions are welcome. Thanks!
>>>
>>> Hi Flo,
>>>
>>> thank you very much for setting up the page.
>>>
>>> My comments are mostly about the commands.
>>>
>>> certmappingconfig-mod:
>>>
>>> * --enable=Boolean: if this option is 'False' SSSD will basically show
>>>   the current behavior and just look up the certificates directly. But I
>>>   wonder if the option is needed at all because not adding any mapping
>>>   rules would have the same effect.
>>>
>>>   What is the scope here, only the IPA domain, or all trusted domains as
>>>   well? If it is for trusted domains as well will the certmappingrule-*
>>>   commands and user-{add/remove}-certmapping return an error?
>>>
>>>   So, in general I see an overlap with the mapping rules and I think it
>>>   would be clearer to drop this option and do the lookups according to
>>>   the mapping rules.
>>>
>>> * --prompt-username=Boolean: the description implies that this option is
>>>   synonymous to 1:1 mapping, but it is not. On Linux authentication in
>>>   most cases use a user name either by directly asking (e.g. /bin/login)
>>>   or using the current user name (e.g. sudo). So, according to its name
>>>   it would only control if gdm is allowed to ask for an (optional) user
>>>   name.
>>>
>>>   If the option is renamed to e.g. --force-1-to-1-mapping to really
>>>   enforce a 1:1 mapping then it would make sense to derived to gdm
>>>   behavior. I.e. if 1:1 mapping is enforce it makes no sense for gdm to
>>>   ask for a user name and if it is not enforced then it makes sense to
>>>   offer and optional user name input field.
>>>
>>> * --enable-username-mismatch=Boolean: I think this option can be
>>>   dropped. My test so far show that if a non-matching hint is given on a
>>>   Windows client authentication fails.
>>>
>>> * --alternate-attribute=STRING: I think this option isn't needed as
>>>   well. For IPA server-side we should decide on an attribute name and
>>>   add it to the schema for user objects. On the client side the
>>>   attribute name can be taken from the mapping rule.A
>>>
>>>
>>> certmappingrule.*:
>>>
>>> * ISSUERDN: it looks like you want to use issuerName here. In
>>>   certificateRecord it it used with LDAP ordering and I would prefer
>>>   LDAP ordering at all points where we have a choice. Unfortunately in the
>>>   issuer-subject mapping AD dictates X.500 ordering.
>>
>> LDAP ordering should indeed be preferred, as it is used everywhere else in
>> IPA. We can convert to/from X.500 ordering where necessary, when possible.
>>
>>>
>>> * DOMAINDN: does this refer to the nsslapd-certmap-basedn attribute in
>>>   the example? My intention in the SSSD design-page was to specify the
>>>   domain (as in DNS domain/IPA domain/trusted domain) where the matching
>>>   user should be searched. Different domains might certificates from
>>>   different issuers and some domains might not even use certificates.
>>>   With this information SSSD does not have to search any domain trusted
>>>   by IPA from a given certificate, but look only at domains listed here
>>>   (the attribute should be a multi-value one).
>>>
>>>   There are objects in the LDAP tree for each trusted domain which are
>>>   used by SSSD so using a DN syntax would be valid here.
>>
>> We use domain names rather than DNs to refer to domains everywhere else in
>> the framework. I don't think this place should be an exception.
>
> I'm fine with domain names as well. In fact I didn't thought of using
> DNs for this before I read DOMAINDN on the design page.
>
>>
>>>
>>> * LDAPSEARCHFILTER: I think a separate option is not need. LDAP search
>>>   filters should just be a special kind of mapping rules. I can image in
>>>   syntax like: <LDAPFILTER:(&(cn=%A)(email=%B)(authType=pkinit))>. I
>>>   think the difficult part with the LDAP filters will to define sensible
>>>   templates.
>>
>> I'm not sure I understand. Could you please elaborate a little bit?
>
> A LDAP search filter which would cover the AD behavior would look like:
>
> (|(altSecurityIdentities=<I>%A<S>%B)(userPrincipalName=%C)(samAccountName=%D))
>
> where
>
> %A: must be replaced with the issuer of the certificate in X.500 order
> %B: must be replaced with the subject of the certificate in X.500 order
>
> it would be possible of course to use a specific template here which
> would generate the complete search attribute value.
>
> %C: must be replaced by the principal from AD's SAN
>     szOID_NT_PRINCIPAL_NAME
> %D: must be replaced with only then name component (the part before the
>     realm) of the principal from szOID_NT_PRINCIPAL_NAME
>
> As %C and %D imply this filter will only work for certificates which
> have szOID_NT_PRINCIPAL_NAME but for those it must be used to be
> compatible with AD. For certificates without
>
> (altSecurityIdentities=<I>%A<S>%B)
>
> is sufficient. It is possible to select the right filter with matching
> rules.

Right.

>
> So we have to find suitable names for the %A, %B, %C and %D templates
> and also allow different representations (e.g. LDAP or X.500 order for
> DNs).

I would personally prefer if we used Python-style formatting strings for 
the templates, I find them much more pleasant to work with than C-style 
formatting strings. The filter which covers AD behavior could be written as:

(|(altSecurityIdentities=<I>{issuer_dn!x500}<S>{subject_dn!x500})(userPrincipalName={subject_nt_principal})(samAccountName={subject_nt_principal.name}))

>
>>
>>>   But as long as we keep the general mapping rule syntax
>>>   flexible the LDAP filter rules can be added in a later version.
>>
>> IMHO it should be the other way round and LDAP filters should be implemented
>> first, as they offer all the flexibility we need (all of the other fields
>> can be easily implemented on top of LDAP filters) and are by default
>> extensible without having to update servers and clients.
>
> In general I agree, as long we can find a suitable scheme to handle the
> templates to add content from the certificate in a specific format to
> the search filters.
>
> But from the user/admin perspective there should be special rules for
> common use-cases which do not require to know too much details about
> certificates and LDAP trees. E.g. for AD (either via direct or indirect
> integration) there should be a <AD-LIKE> rule which just does all which
> AD would do depending on the certificate type. For IPA something like
> <ALT-SEC-ID-I-S> might be a good start for handling external
> certificates which do not contain user specific data which can be mapped
> to user object because the syntax is already known from AD.

This could be handled in the IPA plugin by converting from the 
user-friendly representation to LDAP filter template internally when a 
mapping rule is added or modified.

>
>>
>>>
>>> * enable/disable: I think this is a good idea and would be consistent
>>>   with other rules like HBAC and sudo
>>>
>>> * user-{add/mod} LOGIN --certmappingdata DATA: I think it might be
>>>   better to not add this option and only implement the
>>>   'user-{add/remove}-certmapping' commands
>>>
>>> * user-{add/remove}-certmapping: you say '... almost any type of mapping,
>>>   or a more user-friendly API ...'. I would not say 'or' but 'and' and
>>>   implement both
>>>
>>> * ipaCertMappingEnableMismatch and ipaCertMappingAlternateIdAttribute; I
>>>   think both are note needed, see above
>>>
>>> * altSecurityIdentities: I would prefer to use a different name and OID.
>>>   Using the same definition as AD would imo imply that it can be used in
>>>   the same way as in AD. But e.g. AD also supports other content like
>>>   KERBEROS:alternative_user_principal at AD.DOMAIN which we will not
>>>   support.
>>>
>>> * issuerName vs ipaCAIssuerDN: I would prefer issuerName because it is
>>>   general UTF-8 and not DN syntax (1.3.6.1.4.1.1466.115.121.1.12). Since
>>>   the issuer DN in general will not be a DN from the local LDAP tree I
>>>   think the UTF-8 version fits better.
>>
>> I think it's worth mentioning that if the attribute used DN syntax and
>> matching, we wouldn't have to worry about normalizing the issuer name before
>> searching for it, as DS would do that for us.
>
> Good point, but I think the main use case for this attribute is on the
> client side to determine if a rule should be applied to a certificate or
> not. So I guess LDAP searches with this attribute would be rare because
> the client will load all rules in one run.
>
>>
>>>
>>> * nsslapd-certmap-basedn, see DOMAINDN above
>>>
>>> * altSecurityIdentities example: X.500 ordering is used by AD here and
>>>   unfortunately I think we have to adopt it at least for this specific
>>>   usage, here is an ldapsearch output from AD:
>>>
>>> altSecurityIdentities:
>>> X509:<I>DC=devel,DC=ad,CN=ad-AD-SERVER-CA<S>DC=devel,DC
>>>  =ad,CN=Users,CN=t u,E=test.user at email.domain
>>> altSecurityIdentities: X509:<I>O=Red Hat,OU=prod,CN=Certificate
>>> Authority<S>DC
>>>  =com,DC=redhat,OU=users,OID.0.9.2342.19200300.100.1.1=sbose,E=sbose at redhat.co
>>>  m,CN=Sumit Bose Sumit Bose
>>>
>>> * Certificate Mapping Administrators or re-use Certificate
>>>   Administrators: I would prefer a new 'Certificate Mapping
>>>   Administrators'
>>>
>>> * Users can manage their own X.509 certificate mappings? I'm not sure
>>>   here, at the first glance I would say no. How are OTP tokens handled?
>>>   Maybe this would be a candidate for certmappingconfig-* option?
>>
>> I think a better question is "How is userCertificate handled?"
>>
>> Anyway, self-service permissions can be enabled/disabled, so there is really
>> no need for a new certmappingconfig option.
>
> Yes, this makes sense.
>
> bye,
> Sumit
>>
>>>
>>> That's all :-)
>>>
>>> bye,
>>> Sumit
>>>
>>
>>
>> --
>> Jan Cholasta

-- 
Jan Cholasta