Open Stack

Hi Marco

1. I agree that a discovery service is needed somewhere in the
federation architecture. But the discovery service should be independent
of the endpoint URL that is used to access OpenStack services via
Keystone. It is not a good design to mix up these two aspects, which
appears to have been done in the current design. So Keystone could still
offer a discovery service to end users, by returning details of the IDPs
that it trusts, and information about how to contact those IdPs. But all
the federated users should still be redirected back to the same endpoint
URL of Keystone. Then if discovery is offered by something other than
Keystone, Keystone can still validate federated user authentication.

2. Discovery is different from trust management. The discovery service
could be separate from Keystone (and Apache), but Keystone (or Apache)
would still need a way of specifying which IdPs are trusted, so as to
reject users authenticated by untrusted IdPs.

3. In ABFAB, discovery is part of the ABFAB protocol suite, since
usernames contain the domain name of the user (e.g. d.chad at kent.ac.uk)
which allows Radius to route the user to the correct IDP. Hence ABFAB
does not require Keystone to be the discovery service. But Keystone (or
Apache) still needs to perform trust management to reject users from
untrusted IdPs.

More comments below about the current design, which is clearly sub-optimal.


On 05/01/2015 10:33, Marco Fargetta wrote:
> Hi David,
> 
> in principle I agree with your comments. The current design mixes
> different aspect up and it is not manageable when the number of IdPs
> get bigger, like in the case you should allow access from users in a
> country federation, especially compared to other tools supporting
> identity federation.
> 
> Nevertheless, I think you have to consider the current implementation
> like a fusion between the discovery protocol and the
> authentication.

this is still mixing up aspects!

 Users, instead of being re-directed to a Discovery
> Service providing the list of IdPs, receive the list from keystone
> itself. 

The list of IdPs and the endpoint URL are two different concepts and
should not be mixed up together (which they currently are). You should
be able to tell the user the list of IdPs to choose from, and still have
a single endpoint URL. You could in our original federation
implementation :-)


> Each endpoint URL is an IdP the user can use to authenticate

Sorry it is not quite that. The endpoint is not an IDP, but rather it is
the endpoint that the user has to return to after being authenticated by
the IDP.

> and when selected the user should go directly to the IdP

the discovery service should tell the user how to get to this IDP, but
not where in Keystone to return to, since in general, the Disco service
should be usable by multiple service providers. Therefore the
information it provides should be independent of the service providers
(which in the current implementation it is not).

 and not into
> the DS. Of course I am not saying this is good but it is acceptable
> from the user point of view. There is not the problem to map IdPs in
> the DS with endpoint URL because it is made in advance.

This mapping is not needed in my opinion.

> 
> By the way, if you change the approach and create a single URL for the
> authentication then I cannot see the use of a list of trusted
> IdPs.

as stated above trust management is separate from and different to
discovery. Dont confuse the two.

 You should disable the not accepted IdPs at higher level so to
> avoid the situation where a user authenticate into the IdP but cannot
> access the service.

In general you can never prevent this, since the endpoint is public
information (which is usually published).

 You may work at apache and DS level to enable only
> trusted IdPs. Then you need a better mapping in order to put your
> logic there.
> 
> I think this is a significant change and if there is agreement I think
> it is possible to end with a more flexible design.

this is needed in my opinion.

> 
> Do you plan to propose a new spec?

Not at the current time. There is little point in producing designs that
no-one is willing to implement :-) Once the core implementers accept
that the current design is poor and needs re-engineering, then I will be
happy to propose a new design

regards

David

> 
> Marco
> 
> 
> 
> 
> 
> 
> 
> On Fri, Jan 02, 2015 at 09:51:55PM +0000, David Chadwick wrote:
>> Hi Marco
>>
>> I think the current design is wrong because it is mixing up access
>> control with service endpoint location. The endpoint of a service should
>> be independent of the access control rules determining who can contact
>> the service. Any entity should be able to contact a service endpoint
>> (subject to firewall rules of course, but this is out of scope of
>> Keystone), and once connected, access control should then be enforced.
>> Unfortunately the current design directly ties access control (which
>> IdP) to the service endpoint by building the IDP name into the URL. This
>> is fundamentally a bad design. Not only is it too limiting, but also it
>> is mixing up different concerns, rather than separating them out, which
>> is a good computer science principle.
>>
>> So, applying the separation of concerns principle to Keystone, the
>> federated login endpoint should not be tied to any specific IdP. There
>> are many practical reasons for this, such as:
>>
>> a) in the general case the users of an openstack service could be from
>> multiple different organisations, and hence multiple different IdPs, but
>> they may all need to access the same service and hence same endpoint,
>> b) users who are authorised to access an openstack service might be
>> authorised based on their identity attributes that are not IdP specific
>> (e.g. email address), so they might have a choice of IDP to use
>> c) federations are getting larger and larger, and interfederations are
>> exploding the number of IdPs that users can use. The GEANT eduGAIN
>> interfederation for example now has IdPs from about 20 countries, and
>> each country can have over a 100 IdPs. So we are talking about thousands
>> of IdPs in a federation. It is conceivable that users from all of these
>> might wish to access a given cloud service.
>>
>> Here is my proposal for how federation should be re-engineered
>>
>> 1. The federation endpoint URL for Keystone can be anything intuitive
>> and in keeping with existing guidelines, and should be IDP independent
>>
>> 2. Apache will protect this endpoint with whatever federation
>> protocol(s) it is able to. The Keystone administrator and Apache
>> administrator will liaise out of band to determine the name of the
>> endpoint and the federation protocol and IDPs that will be able to
>> access it.
>>
>> 3. Keystone will have its list of trusted IdPs as now.
>>
>> 4. Keystone will have its mapping rules as now (although I still believe
>> it would be better for mapping rules to be IDP independent, and to have
>> lists of trusted attributes from trusted IDPs instead)
>>
>> 5. Apache will return to Keystone two new parameters indicating the IdP
>> and protocol that were used by the user in connecting to the endpoint.
>> Apache knows what these are.
>>
>> 6. Keystone will use these new parameters for access control and mapping
>> rules. i.e. it will reject any users who are from untrusted IdPs, and it
>> will determine the right mapping rule to use based on the values of the
>> two new parameters. A simple table in Keystone will map the IdPs and
>> protocols into the correct mapping rule to use.
>>
>> This is not a huge change to make, in fact it should be a rather simple
>> re-engineering task.
>>
>> regards
>>
>> David
>>
>>
>> On 24/12/2014 17:50, Marco Fargetta wrote:
>>>
>>>> On 24 Dec 2014, at 17:34, David Chadwick <d.w.chadwick at kent.ac.uk> wrote:
>>>>
>>>> If I understand the bug fix correctly, it is firmly tying the URL to the
>>>> IDP to the mapping rule. But I think this is going in the wrong
>>>> direction for several reasons:
>>>>
>>>> 1. With Shibboleth, if you use a WAYF service, then anyone from hundreds
>>>> of different federated IDPs may end up being used to authenticate the
>>>> user who is accessing OpenStack/Keystone. We dont want to have hundreds
>>>> of URLs. One is sufficient. Plus we dont know which IDP the user will
>>>> eventually choose, as this is decided by the WAYF service. So the
>>>> "correct" URL cannot be pre-chosen by the user.
>>>>
>>>
>>>
>>> With the proposed configuration of shibboleth when you access the URL then you are
>>> redirect only to the IdP configured for the URL. Since a URL is tied to only an IDP there
>>> is not need of a WAYF.
>>>
>>> Anyway, this is a change only in the documentation and it was the first fix because there was
>>> an agreement to provide a solution also for Juno with the minimal change in the code.
>>>
>>> The other fix I proposed, which is under review, requires an additional parameter when you
>>> configure the IdP in OS-Federation. This accepts one or more EntityIDs so you can map the entities
>>> with the URL. This also requires to specify the http variable where you can get the entityID (this
>>> is a parameter so it can be compatible with different SAML plug-ins).
>>> If you do not specify these values the behaviour is like the current implementation otherwise
>>> providing the list of entities and the parameter the access to the URL is allowed only to the
>>> IDP included in the list and the other are rejected.
>>>
>>> I tried to be more compatible with the current implementation as possible.
>>>
>>> Is this in the right direction? Could you comment on the review page? It will be better to understand
>>> it the patch need extra work. The link is: https://review.openstack.org/#/c/142743/
>>>
>>>> 2. With ABFAB, the IDP to be used is not known by the SP (Keystone)
>>>> until after authentication. This is because the realm is incorporated in
>>>> the user's ID (user at real.com) and this is not visible to Keystone. So it
>>>> is not possible to have different URLs for different IDPs. They all have
>>>> to use the same URL.
>>>>
>>>> So there should be one URL protecting Keystone, and when the response
>>>> comes from Apache, Keystone needs to be able to reliably determine
>>>>
>>>> a) which IDP was used by the user
>>>> b) which protocol was used
>>>>
>>>> and from this, choose which mapping rule to use
>>>>
>>>
>>>
>>> This would require a new design of the OS-Federation and you have proposed several specs
>>> I was agreeing with. Nevertheless, it seems there was not consensus in the community so
>>> I think you have to find a way to integrate ABFAB with the current model.
>>>
>>> Is it possible to have a single mapping with many rules and keystone chose according to the
>>> information coming after the authentication? Maybe this require to work on the mapping but it does not
>>> require changes in the overall architecture, just an idea.
>>>
>>>
>>>
>>>
>>>> regards
>>>>
>>>> david
>>>>
>>>>
>>>> On 24/12/2014 10:19, Marco Fargetta wrote:
>>>>> Hi All,
>>>>>
>>>>> this bug was already reported and fixed in two steps:
>>>>>
>>>>> https://bugs.launchpad.net/ossn/+bug/1390124
>>>>>
>>>>>
>>>>> The first step is in the documentation. There should be also an OSS advice for previous
>>>>> version of OpenStack. The solution consist in configuring shibboleth to use different IdPs for 
>>>>> different URLs.
>>>>>
>>>>> The second step, still in progress, is to include an ID in the IdP configuration. My patch is under review here:
>>>>>
>>>>> https://review.openstack.org/#/c/142743/
>>>>>
>>>>> Let me know if it is enough to solve the issue in your case.
>>>>>
>>>>> Marco
>>>>>
>>>>>> On 24 Dec 2014, at 10:08, David Chadwick <d.w.chadwick at kent.ac.uk> wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>> On 23/12/2014 21:56, Morgan Fainberg wrote:
>>>>>>>
>>>>>>>> On Dec 23, 2014, at 1:08 PM, Dolph Mathews <dolph.mathews at gmail.com
>>>>>>>> <mailto:dolph.mathews at gmail.com>> wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>> On Tue, Dec 23, 2014 at 1:33 PM, David
>>>>>>>> Chadwick <d.w.chadwick at kent.ac.uk <mailto:d.w.chadwick at kent.ac.uk>> wrote:
>>>>>>>>
>>>>>>>>   Hi Adam
>>>>>>>>
>>>>>>>>   On 23/12/2014 17:34, Adam Young wrote:
>>>>>>>>> On 12/23/2014 11:34 AM, David Chadwick wrote:
>>>>>>>>>> Hi guys
>>>>>>>>>>
>>>>>>>>>> we now have the ABFAB federation protocol working with Keystone, using a
>>>>>>>>>> modified mod_auth_kerb plugin for Apache (available from the project
>>>>>>>>>> Moonshot web site). However, we did not change Keystone configuration
>>>>>>>>>> from its original SAML federation configuration, when it was talking to
>>>>>>>>>> SAML IDPs, using mod_shibboleth. Neither did we modify the Keystone code
>>>>>>>>>> (which I believe had to be done for OpenID connect.) We simply replaced
>>>>>>>>>> mod_shibboleth with mod_auth_kerb and talked to a completely different
>>>>>>>>>> IDP with a different protocol. And everything worked just fine.
>>>>>>>>>>
>>>>>>>>>> Consequently Keystone is broken, since you can configure it to trust a
>>>>>>>>>> particular IDP, talking a particular protocol, but Apache will happily
>>>>>>>>>> talk to another IDP, using a different protocol, and Keystone cannot
>>>>>>>>>> tell the difference and will happily accept the authenticated user.
>>>>>>>>>> Keystone should reject any authenticated user who does not come from the
>>>>>>>>>> trusted IDP talking the correct protocol. Otherwise there is no point in
>>>>>>>>>> configuring Keystone with this information, if it is ignored by Keystone.
>>>>>>>>> The IDP and the Protocol should be passed from HTTPD in env vars. Can
>>>>>>>>> you confirm/deny that this is the case now?
>>>>>>>>
>>>>>>>>   What is passed from Apache is the 'PATH_INFO' variable, and it is
>>>>>>>>   set to
>>>>>>>>   the URL of Keystone that is being protected, which in our case is
>>>>>>>>   /OS-FEDERATION/identity_providers/KentProxy/protocols/saml2/auth
>>>>>>>>
>>>>>>>>   There are also the following arguments passed to Keystone
>>>>>>>>   'wsgiorg.routing_args': (<routes.util.URLGenerator object at
>>>>>>>>   0x7ffaba339190>, {'identity_provider': u'KentProxy', 'protocol':
>>>>>>>>   u'saml2'})
>>>>>>>>
>>>>>>>>   and
>>>>>>>>
>>>>>>>>   'PATH_TRANSLATED':
>>>>>>>>   '/var/www/keystone/main/v3/OS-FEDERATION/identity_providers/KentProxy/protocols/saml2/auth'
>>>>>>>>
>>>>>>>>   So Apache is telling Keystone that it has protected the URL that
>>>>>>>>   Keystone has configured to be protected.
>>>>>>>>
>>>>>>>>   However, Apache has been configured to protect this URL with the ABFAB
>>>>>>>>   protocol and the local Radius server, rather than the KentProxy
>>>>>>>>   IdP and
>>>>>>>>   the SAML2 protocol. So we could say that Apache is lying to Keystone,
>>>>>>>>   and because Keystone trusts Apache, then Keystone trusts Apache's lies
>>>>>>>>   and wrongly thinks that the correct IDP and protocol were used.
>>>>>>>>
>>>>>>>>   The only sure way to protect Keystone from a wrongly or mal-configured
>>>>>>>>   Apache is to have end to end security, where Keystone gets a token
>>>>>>>>   from
>>>>>>>>   the IDP that it can validate, to prove that it is the trusted IDP that
>>>>>>>>   it is talking to. In other words, if Keystone is given the original
>>>>>>>>   signed SAML assertion from the IDP, it will know for definite that the
>>>>>>>>   user was authenticated by the trusted IDP using the trusted protocol
>>>>>>>>
>>>>>>>>
>>>>>>>> So the "bug" is a misconfiguration, not an actual bug. The goal was to
>>>>>>>> trust and leverage httpd, not reimplement it and all it's extensions.
>>>>>>>
>>>>>>> Fixing this “bug” would be moving towards Keystone needing to implement
>>>>>>> all of the various protocols to avoid “misconfigurations”. There are
>>>>>>> probably some more values that can be passed down from the Apache layer
>>>>>>> to help provide more confidence in the IDP that is being used. I don’t
>>>>>>> see a real tangible benefit to moving away from leveraging HTTPD for
>>>>>>> handling the heavy lifting when handling federated Identity. 
>>>>>>
>>>>>> Its not as heavy as you suggest. Apache would still do all the protocol
>>>>>> negotiation and validation. Keystone would only need to verify the
>>>>>> signature of the incoming SAML assertion in order to validate who the
>>>>>> IDP was, and that it was SAML. (Remember that Keystone already
>>>>>> implements SAML for sending out SAML assertions, which is much more
>>>>>> heavyweight.) ABFAB sends an unsigned SAML assertion embedded in a
>>>>>> Radius attribute, so obtaining this and doing a minimum of field
>>>>>> checking would be sufficient. There will be something similar that can
>>>>>> be done for OpenID Connect.
>>>>>>
>>>>>> So we are not talking about redoing all the protocol handling, simply
>>>>>> checking that the trust rules that have already been configured into
>>>>>> Keystone, are actually being followed by Apache. "Trust but verify" in
>>>>>> the words of Ronald Regan.
>>>>>>
>>>>>> regards
>>>>>>
>>>>>> David
>>>>>>
>>>>>>>
>>>>>>> —Morgan
>>>>>>>
>>>>>>>>
>>>>>>>>   regards
>>>>>>>>
>>>>>>>>   David
>>>>>>>>
>>>>>>>>>
>>>>>>>>> On the Apache side we are looking to expand the set of variables set.
>>>>>>>>> http://www.freeipa.org/page/Environment_Variables#Proposed_Additional_Variables
>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>>   The original SAML assertion
>>>>>>>>>
>>>>>>>>> mod_shib does support Shib-Identity-Provider :
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> https://wiki.shibboleth.net/confluence/display/SHIB2/NativeSPAttributeAccess#NativeSPAttributeAccess-CustomSPVariables
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Which should be sufficient: if the user is coming in via
>>>>>>>>   mod_shib, they
>>>>>>>>> are using SAML.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> BTW, we are using the Juno release. We should fix this bug in Kilo.
>>>>>>>>>>
>>>>>>>>>> As I have been saying for many months, Keystone does not know
>>>>>>>>   anything
>>>>>>>>>> about SAML or ABFAB or OpenID Connect protocols, so there is
>>>>>>>>   currently
>>>>>>>>>> no point in configuring this information into Keystone.
>>>>>>>>   Keystone is only
>>>>>>>>>> aware of environmental parameters coming from Apache. So this
>>>>>>>>   is the
>>>>>>>>>> protocol that Keystone recognises. If you want Keystone to try to
>>>>>>>>>> control the federation protocol and IDPs used by Apache, then
>>>>>>>>   you will
>>>>>>>>>> need the Apache plugins to pass the name of the IDP and the
>>>>>>>>   protocol
>>>>>>>>>> being used as environmental parameters to Keystone, and then
>>>>>>>>   Keystone
>>>>>>>>>> can check that the ones that it has been configured to trust, are
>>>>>>>>>> actually being used by Apache.
>>>>>>>>>>
>>>>>>>>>> regards
>>>>>>>>>>
>>>>>>>>>> David
>>>>>>>>>>
>>>>>>>>>> _______________________________________________
>>>>>>>>>> OpenStack-dev mailing list
>>>>>>>>>> OpenStack-dev at lists.openstack.org
>>>>>>>>   <mailto:OpenStack-dev at lists.openstack.org>
>>>>>>>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> _______________________________________________
>>>>>>>>> OpenStack-dev mailing list
>>>>>>>>> OpenStack-dev at lists.openstack.org
>>>>>>>>   <mailto:OpenStack-dev at lists.openstack.org>
>>>>>>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>>>>>>>
>>>>>>>>
>>>>>>>>   _______________________________________________
>>>>>>>>   OpenStack-dev mailing list
>>>>>>>>   OpenStack-dev at lists.openstack.org
>>>>>>>>   <mailto:OpenStack-dev at lists.openstack.org>
>>>>>>>>   http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>>>>>>
>>>>>>>>
>>>>>>>> _______________________________________________
>>>>>>>> OpenStack-dev mailing list
>>>>>>>> OpenStack-dev at lists.openstack.org
>>>>>>>> <mailto:OpenStack-dev at lists.openstack.org>
>>>>>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> OpenStack-dev mailing list
>>>>>>> OpenStack-dev at lists.openstack.org
>>>>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> OpenStack-dev mailing list
>>>>>> OpenStack-dev at lists.openstack.org
>>>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>>>
>>>>> ====================================================
>>>>> Eng. Marco Fargetta, PhD
>>>>>
>>>>> Istituto Nazionale di Fisica Nucleare (INFN)
>>>>> Catania, Italy
>>>>>
>>>>> EMail: Marco.Fargetta at ct.infn.it
>>>>> ====================================================
>>>>>
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> OpenStack-dev mailing list
>>>>> OpenStack-dev at lists.openstack.org
>>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>>>
>>>>
>>>> _______________________________________________
>>>> OpenStack-dev mailing list
>>>> OpenStack-dev at lists.openstack.org
>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>
>>> ====================================================
>>> Eng. Marco Fargetta, PhD
>>>
>>> Istituto Nazionale di Fisica Nucleare (INFN)
>>> Catania, Italy
>>>
>>> EMail: Marco.Fargetta at ct.infn.it
>>> ====================================================
>>>
>>>
>>>
>>> _______________________________________________
>>> OpenStack-dev mailing list
>>> OpenStack-dev at lists.openstack.org
>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>
>>
>> _______________________________________________
>> OpenStack-dev mailing list
>> OpenStack-dev at lists.openstack.org
>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
> 
> 
> 
> _______________________________________________
> OpenStack-dev mailing list
> OpenStack-dev at lists.openstack.org
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>