Re: [RFC 08/20] ima: Move measurement list related variables into ima_namespace

On 12/2/21 12:44, James Bottomley wrote:

On Thu, 2021-12-02 at 11:45 -0500, Stefan Berger wrote:

quoted

On 12/2/21 11:29, James Bottomley wrote:

quoted

On Thu, 2021-12-02 at 08:41 -0500, Stefan Berger wrote:

quoted

On 12/2/21 07:46, James Bottomley wrote:

quoted

On Tue, 2021-11-30 at 11:06 -0500, Stefan Berger wrote:

quoted

Move measurement list related variables into the
ima_namespace.
This
way a
front-end like SecurityFS can show the measurement list
inside an
IMA
namespace.

Implement ima_free_measurements() to free a list of
measurements
and call it when an IMA namespace is deleted.

This one worries me quite a lot.  What seems to be happening in
this
code:

quoted

@@ -107,7 +100,7 @@ static int ima_add_digest_entry(struct

ima_namespace *ns,
           qe->entry = entry;
    
           INIT_LIST_HEAD(&qe->later);
-       list_add_tail_rcu(&qe->later, &ima_measurements);
+       list_add_tail_rcu(&qe->later, &ns->ima_measurements);
    
           atomic_long_inc(&ns->ima_htable.len);
           if (update_htable) {

is that we now only add the measurements to the namespace list,
but
that list is freed when the namespace dies.  However, the
measurement
is still extended through the PCRs meaning we have incomplete
information for a replay after the namespace dies?

*Not at all.* The measurement list of the namespace is
independent
of
the host.

The cover letter states:

I get that the host can set up a policy to log everything in the
namespace, but that wasn't my question.  My question is can the
guest
set up a policy to log something that doesn't go into the host log
(because the host hasn't asked for it to be logged) but extends a
PCR
anyway, thus destroying the ability of the host to do log replay.

host log goes with host TPM and vice versa

guest log goes with (optional) vTPM and vice version

But that's what doesn't seem to happen ... ima_pcr_extend isn't
virtualized and it's always called from ima_add_template_entry()
meaning the physical TPM is always extended even for a namespace only
entry.

You cannot set a measurement rule in the namespace. That is prevented 
per 9/20: ima: Only accept AUDIT rules for IMA non-init_ima_ns 
namespaces for now.

Also, with the tests that I have done with IMA namespaces I have not 
seen any 'evmctl ima_measurement ...' failures.

Have you been able to cause the IMA namespace to do measurements? It 
would be an easy thing to move the tpm_chip into the ima_namespace as 
well, but per 9/20 this shouldn't be necessary at this point.

quoted

Extending the PCR of the host's TPM would require the data to be
logged in the host log as well. So, no, it's not possible.

Well, exactly: if you don't have or want a vTPM per container the only
way to attest is via the physical TPM which means all entries in the
namespace must be in the host log, so the host owner can quote and
reply and they can split the attested log and give assurance to the
namespaces that their entries are correct.

Yes, this series allows you to log into the system log and along with 
this extend the TPM PCR.

James