On Fri, 2019-07-19 at 11:00 -0500, Eric W. Biederman wrote: > Paul Moore <paul@xxxxxxxxxxxxxx> writes: > > > On Wed, Jul 17, 2019 at 8:52 PM Richard Guy Briggs <rgb@xxxxxxxxxx> > > wrote: > > > On 2019-07-16 19:30, Paul Moore wrote: > > > > ... > > > > > > We can trust capable(CAP_AUDIT_CONTROL) for enforcing audit > > > > container ID policy, we can not trust > > > > ns_capable(CAP_AUDIT_CONTROL). > > > > > > Ok. So does a process in a non-init user namespace have two (or > > > more) sets of capabilities stored in creds, one in the > > > init_user_ns, and one in current_user_ns? Or does it get > > > stripped of all its capabilities in init_user_ns once it has its > > > own set in current_user_ns? If the former, then we can use > > > capable(). If the latter, we need another mechanism, as > > > you have suggested might be needed. > > > > Unfortunately I think the problem is that ultimately we need to > > allow any container orchestrator that has been given privileges to > > manage the audit container ID to also grant that privilege to any > > of the child process/containers it manages. I don't believe we can > > do that with capabilities based on the code I've looked at, and the > > discussions I've had, but if you find a way I would leave to hear > > it. > > > If some random unprivileged user wants to fire up a container > > > orchestrator/engine in his own user namespace, then audit needs > > > to be namespaced. Can we safely discard this scenario for now? > > > > I think the only time we want to allow a container orchestrator to > > manage the audit container ID is if it has been granted that > > privilege by someone who has that privilege already. In the zero- > > container, or single-level of containers, case this is relatively > > easy, and we can accomplish it using CAP_AUDIT_CONTROL as the > > privilege. If we start nesting container orchestrators it becomes > > more complicated as we need to be able to support granting and > > inheriting this privilege in a manner; this is why I suggested a > > new mechanism *may* be necessary. > > > Let me segway a bit and see if I can get this conversation out of the > rut it seems to have drifted into. > > Unprivileged containers and nested containers exist today and are > going to become increasingly common. Let that be a given. Agree fully. > As I recall the interesting thing for audit to log is actions by > privileged processes. Audit can log more but generally configuring > logging by of the actions of unprivileged users is effectively a self > DOS. > > So I think the initial implementation can safely ignore actions of > nested containers and unprivileged containers because you don't care > about their actions. I don't entirely agree here: remember there might be two consumers for the audit data: the physical system owner (checking up on the tenants) and the tenant themselves who might be watching either their sub tenants or their users (and who, obviously, won't get the full audit stream). In either case, the tenant may or may not be privileged, and if they're privileged, it might be through the user_ns in which case the physical system owner and the kernel would see them as "not privileged". So I think we are ultimately going to need the ability to audit unprivileged containers. I also think audit has a role to play in intrusion detection and forensic analysis for fully unprivileged containers running external services, but I don't think we have to solve that case immediately. > If we start allow running audit in a container then we need to deal > with all of the nesting issues but until then I don't think you folks > care. > > Or am I wrong. Do the requirements for securely auditing things from > the kernel care about the actions of unprivileged users? I think ultimately we have to care, but it could be three phases: first would be genuinely privileged containers (i.e. with real root inside, being our most dangerous problem) the second would be user_ns privileged containers (i.e. with both user_ns and an interior root mapping) and the third would be unprivileged containers (with or without user_ns but no interior root). James
