[PATCH V33 24/30] bpf: Restrict bpf when kernel lockdown is in confidentiality mode

[Andy Lutomirski]

On Thu, Jun 27, 2019 at 7:35 AM Stephen Smalley <sds at tycho.nsa.gov> wrote:
>
> On 6/26/19 8:57 PM, Andy Lutomirski wrote:
> >
> >
> >> On Jun 26, 2019, at 1:22 PM, James Morris <jmorris at namei.org> wrote:
> >>
> >> [Adding the LSM mailing list: missed this patchset initially]
> >>
> >>> On Thu, 20 Jun 2019, Andy Lutomirski wrote:
> >>>
> >>> This patch exemplifies why I don't like this approach:
> >>>
> >>>> @@ -97,6 +97,7 @@ enum lockdown_reason {
> >>>>         LOCKDOWN_INTEGRITY_MAX,
> >>>>         LOCKDOWN_KCORE,
> >>>>         LOCKDOWN_KPROBES,
> >>>> +       LOCKDOWN_BPF,
> >>>>         LOCKDOWN_CONFIDENTIALITY_MAX,
> >>>
> >>>> --- a/security/lockdown/lockdown.c
> >>>> +++ b/security/lockdown/lockdown.c
> >>>> @@ -33,6 +33,7 @@ static char *lockdown_reasons[LOCKDOWN_CONFIDENTIALITY_MAX+1] = {
> >>>>         [LOCKDOWN_INTEGRITY_MAX] = "integrity",
> >>>>         [LOCKDOWN_KCORE] = "/proc/kcore access",
> >>>>         [LOCKDOWN_KPROBES] = "use of kprobes",
> >>>> +       [LOCKDOWN_BPF] = "use of bpf",
> >>>>         [LOCKDOWN_CONFIDENTIALITY_MAX] = "confidentiality",
> >>>
> >>> The text here says "use of bpf", but what this patch is *really* doing
> >>> is locking down use of BPF to read kernel memory.  If the details
> >>> change, then every LSM needs to get updated, and we risk breaking user
> >>> policies that are based on LSMs that offer excessively fine
> >>> granularity.
> >>
> >> Can you give an example of how the details might change?
> >>
> >>> I'd be more comfortable if the LSM only got to see "confidentiality"
> >>> or "integrity".
> >>
> >> These are not sufficient for creating a useful policy for the SELinux
> >> case.
> >>
> >>
> >
> > I may have misunderstood, but I thought that SELinux mainly needed to allow certain privileged programs to bypass the policy.  Is there a real example of what SELinux wants to do that can’t be done in the simplified model?
> >
> > The think that specifically makes me uneasy about exposing all of these precise actions to LSMs is that they will get exposed to userspace in a way that forces us to treat them as stable ABIs.
>
> There are two scenarios where finer-grained distinctions make sense:
>
> - Users may need to enable specific functionality that falls under the
> umbrella of "confidentiality" or "integrity" lockdown.  Finer-grained
> lockdown reasons free them from having to make an all-or-nothing choice
> between lost functionality or no lockdown at all. This can be supported
> directly by the lockdown module without any help from SELinux or other
> security modules; we just need the ability to specify these
> finer-grained lockdown levels via the boot parameters and securityfs nodes.
>
> - Different processes/programs may need to use different sets of
> functionality restricted via lockdown confidentiality or integrity
> categories.  If we have to allow all-or-none for the set of
> interfaces/functionality covered by the generic confidentiality or
> integrity categories, then we'll end up having to choose between lost
> functionality or overprivileged processes, neither of which is optimal.
>
> Is it truly the case that everything under the "confidentiality"
> category poses the same level of risk to kernel confidentiality, and
> similarly for everything under the "integrity" category?  If not, then
> being able to distinguish them definitely has benefit.
>

They're really quite similar in my mind.  Certainly some things in the
"integrity" category give absolutely trivial control over the kernel
(e.g. modules) while others make it quite challenging (ioperm), but
the end result is very similar.  And quite a few "confidentiality"
things genuinely do allow all kernel memory to be read.

I agree that finer-grained distinctions could be useful. My concern is
that it's a tradeoff, and the other end of the tradeoff is an ABI
stability issue.  If someone decides down the road that some feature
that is currently "integrity" can be split into a narrow "integrity"
feature and a "confidentiality" feature then, if the user policy knows
about the individual features, there's a risk of breaking people's
systems.  If we keep the fine-grained control, do we have a clear
compatibility story?