On 2019-07-15 16:38, Paul Moore wrote:

On Mon, Jul 8, 2019 at 1:51 PM Richard Guy Briggs [off-list ref] wrote:

quoted

On 2019-05-29 11:29, Paul Moore wrote:

...

quoted

quoted

The idea is that only container orchestrators should be able to
set/modify the audit container ID, and since setting the audit
container ID can have a significant effect on the records captured
(and their routing to multiple daemons when we get there) modifying
the audit container ID is akin to modifying the audit configuration
which is why it is gated by CAP_AUDIT_CONTROL.  The current thinking
is that you would only change the audit container ID from one
set/inherited value to another if you were nesting containers, in
which case the nested container orchestrator would need to be granted
CAP_AUDIT_CONTROL (which everyone to date seems to agree is a workable
compromise).  We did consider allowing for a chain of nested audit
container IDs, but the implications of doing so are significant
(implementation mess, runtime cost, etc.) so we are leaving that out
of this effort.

We had previously discussed the idea of restricting
orchestrators/engines from only being able to set the audit container
identifier on their own descendants, but it was discarded.  I've added a
check to ensure this is now enforced.

When we weren't allowing nested orchestrators it wasn't necessary, but
with the move to support nesting I believe this will be a requirement.
We might also need/want to restrict audit container ID changes if a
descendant is acting as a container orchestrator and managing one or
more audit container IDs; although I'm less certain of the need for
this.

I was of the opinion it was necessary before with single-layer parallel
orchestrators/engines.

quoted

I've also added a check to ensure that a process can't set its own audit
container identifier ...

What does this protect against, or what problem does this solve?
Considering how easy it is to fork/exec, it seems like this could be
trivially bypassed.

Well, for starters, it would remove one layer of nesting.  It would
separate the functional layers of processes.  Other than that, it seems
like a gut feeling that it is just wrong to allow it.  It seems like a
layer violation that one container orchestrator/engine could set its own
audit container identifier and then set its children as well.  It would
be its own parent.  It would make it harder to verify adherance to
descendancy and inheritance rules.

quoted

... and that if the identifier is already set, then the
orchestrator/engine must be in a descendant user namespace from the
orchestrator that set the previously inherited audit container
identifier.

You lost me here ... although I don't like the idea of relying on X
namespace inheritance for a hard coded policy on setting the audit
container ID; we've worked hard to keep this independent of any
definition of a "container" and it would sadden me greatly if we had
to go back on that.

This would seem to be the one concession I'm reluctantly making to try
to solve this nested container orchestrator/engine challenge.

Would backing off on that descendant user namespace requirement and only
require that a nested audit container identifier only be permitted on a
descendant task be sufficient?  It may for this use case, but I suspect
not for additional audit daemons (we're not there yet) and message
routing to those daemons.

The one difference here is that it does not depend on this if the audit
container identifier has not already been set.

paul moore

- RGB

--
Richard Guy Briggs [off-list ref]
Sr. S/W Engineer, Kernel Security, Base Operating Systems
Remote, Ottawa, Red Hat Canada
IRC: rgb, SunRaycer
Voice: +1.647.777.2635, Internal: (81) 32635