Re: [PATCH v9 5/6] IMA: Add support to limit measuring keys

[Mat Martineau <mathew.j.martineau@linux.intel.com>]

[-- Attachment #1: Type: text/plain, Size: 2605 bytes --]


On Wed, 4 Dec 2019, Mimi Zohar wrote:

> [Cc'ing Mat Martineau]
>
> On Tue, 2019-12-03 at 15:37 -0800, Lakshmi Ramasubramanian wrote:
>> On 12/3/2019 12:06 PM, Mimi Zohar wrote:
>>
>>> Suppose both root and uid 1000 define a keyring named "foo".  The
>>> current "keyrings=foo" will measure all keys added to either keyring
>>> named "foo".  There needs to be a way to limit measuring keys to a
>>> particular keyring named "foo".
>>>
>>> Mimi
>>
>> Thanks for clarifying.
>>
>> Suppose two different non-root users create keyring with the same name
>> "foo" and, say, both are measured, how would we know which keyring
>> measurement belongs to which user?
>>
>> Wouldn't it be sufficient to include only keyrings created by "root"
>> (UID value 0) in the key measurement? This will include all the builtin
>> trusted keyrings (such as .builtin_trusted_keys,
>> .secondary_trusted_keys, .ima, .evm, etc.).
>>
>> What would be the use case for including keyrings created by non-root
>> users in key measurement?
>>
>> Also, since the UID for non-root users can be any integer value (greater
>> than 0), can an an administrator craft a generic IMA policy that would
>> be applicable to all clients in an enterprise?
>
> The integrity subsystem, and other concepts upstreamed to support it,
> are being used by different people/companies in different ways.  I
> know some of the ways, but not all, as how it is being used.  For
> example, Mat Martineau gave an LSS2019-NA talk titled "Using and
> Implementing Keyring Restrictions for Userspace".  I don't know if he
> would be interested in measuring keys on these restricted userspace
> keyrings, but before we limit how a new feature works, we should at
> least look to see if that limitation is really necessary.

The use cases I'm most familiar with could have a use for key measurement 
for something like enterprise Wi-Fi root certificates. I'm not sure of the 
best way to uniquely identify a key to measure in that scenario, it could 
be anchored in various ways (process, session, thread, or user keyrings, 
for example) and may be owned by a non-root user. As Lakshmi noted above, 
key names are not unique, and I'll add that namespace considerations may 
come in to play too.

Keys (including keyrings like .builtin_trusted_keys, .ima, etc) can be 
linked to multiple keyrings, maybe you could create a system-level 
.ima_measured keyring. You could measure keys that are accessible from 
that keyring, and opt in more keys for measurement by linking them to 
.ima_measured or a keyring nested within .ima_measured.

--
Mat Martineau
Intel