[Freeipa-devel] [freeipa] #3668: CA-less install fails when intermediate CA is used

[Petr Viktorin]

On 06/10/2013 05:34 PM, Dmitri Pal wrote:
> On 06/10/2013 09:56 AM, Petr Viktorin wrote:
>> On 06/10/2013 03:35 PM, Rob Crittenden wrote:
>>> Dmitri Pal wrote:
>>>> On 06/10/2013 05:54 AM, Jan Cholasta wrote:
>>>>> On 7.6.2013 15:36, John Dennis wrote:
>>>>>> On 06/07/2013 09:26 AM, Jan Cholasta wrote:
>>>>>>> On 7.6.2013 15:17, John Dennis wrote:
>>>>>>>> On 06/07/2013 08:57 AM, Jan Cholasta wrote:
>>>>>>>>> Yes, this is correct. The DS certificate must be directly signed
>>>>>>>>> by the
>>>>>>>>> CA trusted by IPA (specified by --root-ca-cert in
>>>>>>>>> ipa-server-install),
>>>>>>>>> there may be no intermediate CAs, because ldapsearch and
>>>>>>>>> friends and
>>>>>>>>> python-ldap don't like them.
>>>>>>>>
>>>>>>>> That doesn't sound right. Do we understand why a chain length >
>>>>>>>> 1 is
>>>>>>>> failing?
>>>>>>>>
>>>>>>>
>>>>>>> LDAP utilities and python-ldap only trust certificates directly
>>>>>>> issued
>>>>>>> by CAs you point them to (at least on Fedora 18).
>>>>>>
>>>>>> This sounds like a bug in MozLDAP (i.e. the NSS LDAP crypto
>>>>>> provider).
>>>>>> Have we filed a bug? Let's file the bug here in the Red Hat bugzilla,
>>>>>> not upstream, we're the maintainers of MozLDAP and upstream is
>>>>>> already
>>>>>> frustrated with it.
>>>>>>
>>>>>
>>>>> I have just tested with libldap compiled with OpenSSL on my Arch Linux
>>>>> box, it does not work as well:
>>>>>
>>>>> $ LDAPTLS_CACERT=$PWD/ca.crt ldapsearch -H
>>>>> ldap://vm-131.idm.lab.bos.redhat.com -ZZ -D 'cn=Directory Manager' -W
>>>>> -b '' -s base
>>>>> ldap_start_tls: Connect error (-11)
>>>>>       additional info: error:14090086:SSL
>>>>> routines:SSL3_GET_SERVER_CERTIFICATE:certificate verify failed
>>>>> (invalid CA certificate)
>>>>>
>>>>> For the record, this is what happens with NSS on Fedora:
>>>>>
>>>>> $ LDAPTLS_CACERT=$PWD/ca.crt ldapsearch -H
>>>>> ldap://vm-131.idm.lab.bos.redhat.com -ZZ -D 'cn=Directory Manager' -W
>>>>> -b '' -s base
>>>>> ldap_start_tls: Connect error (-11)
>>>>>       additional info: TLS error -8179:Peer's Certificate issuer is not
>>>>> recognized.
>>>>>
>>>>> Honza
>>>>>
>>>> If the options does not work we should hide it for now and clearly
>>>> state
>>>> in the docs and man pages that the case when certs come from different
>>>> CA is not supported for the time being.
>>>>
>>>
>>> IIRC it was added for two reasons:
>>>
>>> 1. Because the CA is not required to be included in the PKCS#12 file.
>>
>> Well, that's not a necessary requirement. A tester did have a bit of
>> trouble creating a PKCS#12 file with both the server cert and the CA
>> cert, but the trouble is comparable to having to create a PEM file.
>>
>>> 2. Because previous attempts to figure out the nickname of the signing
>>> CA was problematic.
>>
>> This is the main reason. It could be done, but I believe we should
>> avoid any guessing to determine the root of trust.
>>
>
> I am not sure I get it.
> So we inspect the certificate PKCS#12 file and there can be ambiguity
> about the root of trust. How? This part is not clear to me.
> Sorry for being slow here. Can someone explain it in more details?

A PKCS#12 file can contain any number of arbitrary certificates (or 
other objects).
There is already some ambiguity in determining which cert to use for the 
server. Here we error out if we don't find exactly one cert with 
appropriate usage flags.
The root CA cert can be the one that signed the server cert (the only 
scenario that works now), or it can be any one along the chain.
If you have an intermediate CA signed by a well-known authority, you 
would probably not want to trust *everything* signed by that authority; 
you'd want to trust "your" root CA which can be anywhere along the chain.

If we exclude scenarios with Intermediate CAs (which don't currently 
work), the situation is much simpler, but we don't want to limit the CLI 
to that.

Note that users can't usually carefully control what's inside the 
PKCS#12 file: the tools will typically allow generating a file with 
either only one cert, or with the entire chain.

> Regardless...
> If something can't be reliably determined then we should probably do
> following:
> If we have any issues with determining the root of trust we should fail
> and request the caller to run the command again with additional command
> line option that would explicitly provide the the name of the root CA.
>
> Will that help?

That would be any time there's more than one CA in the trust chain.
Is it worth special-casing this? Are there real-world deployments that 
only use one CA? Note that in this case you can just use a single-cert 
PKCS#12 plus a PEM file for the CA, which is simple enough IMO.

-- 
Petr³