[Freeipa-devel] [DESIGN][UPDATE] Time-Based HBAC Policies

Thu Sep 1 11:26:07 UTC 2016

On 08/31/2016 12:57 PM, Petr Spacek wrote:
> On 31.8.2016 12:42, Standa Laznicka wrote:
>> On 08/30/2016 03:34 PM, Simo Sorce wrote:
>>> On Tue, 2016-08-30 at 08:47 +0200, Standa Laznicka wrote:
>>>> On 08/26/2016 05:37 PM, Simo Sorce wrote:
>>>>> On Fri, 2016-08-26 at 11:26 -0400, Simo Sorce wrote:
>>>>>> On Fri, 2016-08-26 at 18:09 +0300, Alexander Bokovoy wrote:
>>>>>>> On Fri, 26 Aug 2016, Simo Sorce wrote:
>>>>>>>> On Fri, 2016-08-26 at 12:39 +0200, Martin Basti wrote:
>>>>>>>>>> I miss "why" part of "To be able to handle backward compatibility
>>>>>>>>> with
>>>>>>>>>> ease, a new object called ipaHBACRulev2 is introduced. " in the
>>>>>>>>> design
>>>>>>>>>> page. If the reason is the above - old client's should ignore time
>>>>>>>>> rules
>>>>>>>>>> then it has to be mentioned there. Otherwise I don't see a reason to
>>>>>>>>>> introduce a new object type instead of extending the current.
>>>>>>>>> How do you want to enforce HBAC rule that have set time from 10 to 14
>>>>>>>>> everyday? With the same objectclass old clients will allow this HBAC
>>>>>>>>> for
>>>>>>>>> all day. Isn't this CVE?
>>>>>>>> This is a discussion worth having.
>>>>>>>>
>>>>>>>> In general it is a CVE only if an authorization mechanism fails to work
>>>>>>>> as advertised.
>>>>>>>>
>>>>>>>> If you make it clear that old clients *DO NOT* respect time rules then
>>>>>>>> there is no CVE material, it is working as "described".
>>>>>>>>
>>>>>>>> The admins already have a way to not set those rules for older clients
>>>>>>>> by simply grouping newer clients in a different host group and applying
>>>>>>>> time rules only there.
>>>>>>>>
>>>>>>>> So the question really is: should we allow admins to apply an HBAC Rule
>>>>>>>> potentially to older clients that do not understand it and will
>>>>>>>> therefore allow access at any time of the day, or should we prevent it ?
>>>>>>>>
>>>>>>>> This is a hard question to answer and can go both ways.
>>>>>>>>
>>>>>>>> A time rule may be something that admins want to enforce at all cost or
>>>>>>>> deny access. In this case a client that fails to handle it would be a
>>>>>>>> problem.
>>>>>>>>
>>>>>>>> But it may be something that is just used for defense in depth and not a
>>>>>>>> strictly hard requirement. In this case allowing older clients would
>>>>>>>> make it an easy transition as you just set up the rule and the client
>>>>>>>> will start enforcing the time when it is upgraded but work otherwise
>>>>>>>> with the same rules.
>>>>>>>>
>>>>>>>> I am a bit conflicted on trying to decide what scenario we should
>>>>>>>> target, but the second one appeals to me because host groups do already
>>>>>>>> give admins a good way to apply rules to a specific set of hosts and
>>>>>>>> exclude old clients w/o us making it a hard rule.
>>>>>>>> OTOH if an admin does not understand this difference, they may be
>>>>>>>> surprised to find out there are clients that do not honor it.
>>>>>>>>
>>>>>>>> Perhaps we could find a way to set a flag on the rule such that when set
>>>>>>>> (and only when set) older clients get excluded by way of changing the
>>>>>>>> objectlass or something else to similar effect.
>>>>>>>>
>>>>>>>> Open to discussion.
>>>>>>> At this point using new object class becomes an attractive approach. We
>>>>>>> don't have means to exclude HBAC rules other than applying them
>>>>>>> per-host/hostgroup. We also have no deny rules.
>>>>>>>
>>>>>>> I have another idea: what about enforcing time rules always to apply
>>>>>>> per-host or per-hostgroup by default? Add --force option to override the
>>>>>>> behavior but default to not allow --hostcat=all. This would raise
>>>>>>> awareness and make sure admins are actually applying these rules with
>>>>>>> intention.
>>>>>> This sounds like a good idea, but it is not a silver bullet I am afraid.
>>>>>>
>>>>>> Simo.
>>>>> I was thinking that for future proofing we could add a version field,
>>>>> then reasoned more and realized that changing the object class is
>>>>> basically the same thing.
>>>>>
>>>>> There is only one big problem, ipaHBACRule is a STRUCTURAL objectclass.
>>>>> (I know 389ds allows us to do an LDAPv3 illegal operation and change it,
>>>>> but I do not like to depend on that behavoir).
>>>>>
>>>>> Now looking into this I had an idea to solve the problem of legacy
>>>>> clients without having to swap classes.
>>>>> We can redefine the accessRuleType attribute to be a "capability" type.
>>>>>
>>>>> Ie rules that have a timeAccess component will be of type
>>>>> "allow_with_time" instead of just "allow".
>>>>> Old clients are supposed to search with accessRuleType=allow (and I can
>>>>> see that SSSD does that), so an older client will fail to get those
>>>>> rules as they won't match.
>>>>>
>>>>> New clients instead can recognize both types.
>>>>>
>>>>> Also if we need a future extension we will simpy add a new access rule
>>>>> type and we can have the same effect.
>>>>> The nice thing is that accessRyleType is defined as multivalue (no
>>>>> SINGLE in schema) so we may actually create compatible rules if we want
>>>>> to.
>>>>> Ie we could set both "allow" and "allow_with_time" on an object for
>>>>> cases where the admin wants to enforce the time part only o newer client
>>>>> but otherwise apply the rule to any client.
>>>>>
>>>>> This should give us the best of all options at once.
>>>>>
>>>>> Thoughts ?
>>>>>
>>>>> Simo.
>>>>>
>>>> Sorry to join the discussion so late, I was away yesterday.
>>>>
>>>> I have to say I too like this idea much better than fiddling with the
>>>> objectClasses. Also, I believe that accessRuleType was originally
>>>> actually used to distinguish newer version of HBAC rules from the older
>>>> so we may just do this again and profit from its original purpose. To
>>>> top it off, this change should be really easy to implement to what I
>>>> currently have on SSSD side.
>>>>
>>>> I was just wondering - would you propose for every newly created rule to
>>>> have the new accessRuleType set to "allow_with_time" or should the type
>>>> change with addition of time rules to the HBAC rule as it does
>>>> currently? Also, should the user be able to modify the type so that a
>>>> rule with the new type is also visible for older clients (=> he could
>>>> add "allow" to type anytime)?
>>> Rules of type allow_with_time will not work on older clients, so we
>>> should probably default to just the old "allow" schema.
>>>
>>> I think in the first implementation the framework/cli/ui should not
>>> emphasize this attribute but simply replace allow -> allow_with_time if
>>> a time attribute is added.
>>>
>>> In future we may give control of it and allow even to set multiple
>>> values, after we discuss better if that should be done, and with ample
>>> warnings to admins.
>>>
>>> Also setting a time rule makes a rule incompatible with older clients so
>>> we should spell it clearly in the CLI/UI with a warning message that
>>> this rule will not apply at all to older clients.
>>>
>>>> Thanks for your ideas, I am very happy with what you suggested here :)
>>> Thank you.
>>>
>>> Simo.
>>>
>> So - can we all agree on a solution?
>>
>> I took an extra half an hour and created the accessRuleType solution on top of
>> what I currently have, see patches attached to get the picture what the change
>> would mean for what I currently have in
>> https://github.com/stlaz/freeipa/tree/timerules_2 and
>> https://github.com/stlaz/sssd/tree/freeipa-trac-547_2. Note that the sssd
>> patch is really just to get a picture, it currently causes sssd_be to core
>> dump, not sure why and don't want to waste time debugging it right now.
>>
>> I myself would in the end rather go for objectClasses implementation as new
>> rules are not shown to old clients which seems correct as there's no confusion
>> for admins who might scratch their heads at old clients with no idea why their
>> HBAC rules don't apply otherwise.
> +1, I agree with Standa and Martin Basti. Let me repeat myself:
>
> I like the idea of "capabilities" in general but it needs proper design and
> detailed specification first.
>
> Given that we have to modify SSSD anyway, I would go for ipaHBACRulev2 object
> class with clear definition of "capabilities" (without any obsolete cruft).
>
> That should be future proof and without any negative/unforeseen impact to
> existing clients + it matches what Jan Pazdziora plans to do for HBAC+URI.
>
As there were no further objections, the latest changes with the 
objectclass implementation that were made according to Honza's 
suggestions were pushed to appear in the pull request 
https://github.com/freeipa/freeipa/pull/23.