On Tue, 2020-12-29 at 10:46 -0800, Casey Schaufler wrote:
> >>>>>> -int security_audit_rule_match(u32 secid, u32 field, u32 op, void *lsmrule)
> >>>>>> +int security_audit_rule_match(u32 secid, u32 field, u32 op, void **lsmrule)
> >>>>>> {
> >>>>>> - return call_int_hook(audit_rule_match, 0, secid, field, op, lsmrule);
> >>>>>> + struct security_hook_list *hp;
> >>>>>> + int rc;
> >>>>>> +
> >>>>>> + hlist_for_each_entry(hp, &security_hook_heads.audit_rule_match, list) {
> >>>>>> + if (WARN_ON(hp->lsmid->slot < 0 || hp->lsmid->slot >= lsm_slot))
> >>>>>> + continue;
> >>>>>> + rc = hp->hook.audit_rule_match(secid, field, op,
> >>>>>> + &lsmrule[hp->lsmid->slot]);
> >>>>>> + if (rc)
> >>>>>> + return rc;
> >>>>> Suppose that there is an IMA dont_measure or dont_appraise rule, if one
> >>>>> LSM matches, then this returns true, causing any measurement or
> >>>>> integrity verification to be skipped.
> >>>> Yes, that is correct. Like the audit system, you're doing a string based
> >>>> lookup, which pretty well has to work this way. I have proposed compound
> >>>> label specifications in the past, but even if we accepted something like
> >>>> "apparmor=dates,selinux=figs" we'd still have to be compatible with the
> >>>> old style inputs.
> >>>>
> >>>>> Sample policy rules:
> >>>>> dont_measure obj_type=foo_log
> >>>>> dont_appraise obj_type=foo_log
> >>> IMA could extend the existing policy rules like "lsm=[selinux] |
> >>> [smack] | [apparmor]", but that assumes that the underlying
> >>> infrastructure supports it.
> >> Yes, but you would still need rational behavior in the
> >> case where someone has old IMA policy rules.
> > From an IMA perspective, allowing multiple LSMs to define the same
> > policy label is worse than requiring the label be constrained to a
> > particular LSM.
>
> Just to be sure we're talking about the same thing,
> the case I'm referring to is something like a file with
> two extended attributes:
>
> security.apparmor MacAndCheese
> security.SMACK64 MacAndCheese
>
> and an IMA rule that says
>
> dont_measure obj_type=MacAndCheese
>
> In this case the dont_measure will be applied to both.
> On the other hand,
>
> security.apparmor MacAndCheese
> security.SMACK64 FranksAndBeans
>
> would also apply the rule to both, which is not
> what you want. Unfortunately, there is no way to
> differentiate which LSM hit the rule.
>
> So now I'm a little confused. The case where both LSMs
> use the same label looks like it works right, where the
> case where they're different doesn't.
I'm more concerned about multiple LSMs using the same label. The
label's meaning is LSM specific.
>
> I'm beginning to think that identifying which LSMs matched
> a rule (it may be none, either or both) is the right solution.
> I don't think that audit is as sensitive to this.
If the label's meaning is LSM specific, then the rule needs to be LSM
specific.
