Re: [PATCH v2 0/2] Introduce the pkill_on_warn parameter

[Casey Schaufler]

On 11/18/2021 9:32 AM, Kees Cook wrote:
> On Tue, Nov 16, 2021 at 11:00:23AM -0800, Casey Schaufler wrote:
> > On 11/16/2021 10:41 AM, Kees Cook wrote:
> > > On Tue, Nov 16, 2021 at 12:12:16PM +0300, Alexander Popov wrote:
> > > > What if the Linux kernel had a LSM module responsible for error handling policy?
> > > > That would require adding LSM hooks to BUG*(), WARN*(), KERN_EMERG, etc.
> > > > In such LSM policy we can decide immediately how to react on the kernel error.
> > > > We can even decide depending on the subsystem and things like that.
> > > That would solve the "atomicity" issue the WARN tracepoint solution has,
> > > and it would allow for very flexible userspace policy.
> > >
> > > I actually wonder if the existing panic_on_* sites should serve as a
> > > guide for where to put the hooks. The current sysctls could be replaced
> > > by the hooks and a simple LSM.
> > Do you really want to make error handling a "security" issue?
> > If you add security_bug(), security_warn_on() and the like
> > you're begging that they be included in SELinux (AppArmor) policy.
> > BPF, too, come to think of it. Is that what you want?
> Yeah, that is what I was thinking. This would give the LSM a view into
> kernel state, which seems a reasonable thing to do. If system integrity
> is compromised, an LSM may want to stop trusting things.

How are you planning to communicate the security relevance of the
warning to the LSM? I don't think that __FILE__, __LINE__ or __func__
is great information to base security policy on. Nor is a backtrace.

> A dedicated error-handling LSM could be added for those hooks that
> implemented the existing default panic_on_* sysctls, and could expand on
> that logic for other actions.

I can see having an interface like LSM for choosing a bug/warn policy.
I worry about expanding the LSM hook list for a case where I would
hope no existing LSM would use them, and the new LSM doesn't use any
of the existing hooks.