I'm looking at the audit userspace implications of adding two new kernel audit records. AUDIT_MAC_TASK_CONTEXTS and AUDIT_MAC_OBJ_CONTEXTS are used when there are multiple security modules with a "security context" active on the system. This design has been discussed here at length. The records will look like: AUDIT_MAC_TASK_CONTEXTS subj_<lsmname>=value subj_<lsmname>=value ... Looking at the audit user-space code I see several things that have me concerned. The first is the use of WITH_APPARMOR. Going forward what behavior would we want if subj_apparmor=something shows up on a system that has not got WITH_APPARMOR defined?

The code is inconsistent in that it does not use WITH_SELINUX, but that's hardly a surprise given its origins. There is also no WITH_SMACK, but that's unlikely to be an issue since Smack's use of audit is very much like SELinux's.

The question is what to do about filtering when subj=foo is specified. I suggest that if any of subj_selinux, subj_smack or subj_something is "foo", it is a match.

But the SELinux components of a label (level, user, ...) are also available for filtering. If someone wrote a simple Bell & LaPadula LSM filtering by some of those fields could be useful there, too. I would like guidance on whether I ought to go the route of more extensive use of WITH_APPARMOR (and WITH_SMACK, WITH_MUMBLE) or take the path of greater generalization. Or, whether I should treat each case individually and give it my best whack.

On 6/14/2021 2:13 PM, Steve Grubb wrote: > Hello, > > On Monday, June 14, 2021 3:34:33 PM EDT Casey Schaufler wrote: >> I'm looking at the audit userspace implications of adding two >> new kernel audit records. AUDIT_MAC_TASK_CONTEXTS and >> AUDIT_MAC_OBJ_CONTEXTS are used when there are multiple security >> modules with a "security context" active on the system. This >> design has been discussed here at length. The records will look >> >> like: >> AUDIT_MAC_TASK_CONTEXTS >> subj_<lsmname>=value >> subj_<lsmname>=value >> ... >> >> Looking at the audit user-space code I see several things >> that have me concerned. The first is the use of WITH_APPARMOR. >> Going forward what behavior would we want if subj_apparmor=something >> shows up on a system that has not got WITH_APPARMOR defined? > > I think it should be ignored. > >> The code is inconsistent in that it does not use WITH_SELINUX, >> but that's hardly a surprise given its origins. There is also no >> WITH_SMACK, but that's unlikely to be an issue since Smack's use >> of audit is very much like SELinux's. > > We can add those WITH_* if you like. > >> The question is what to >> do about filtering when subj=foo is specified. I suggest that if >> any of subj_selinux, subj_smack or subj_something is "foo", it is >> a match. > > I think that's how we already treat things. There is a linked list for > AVC's and we match on any of. > >> But the SELinux components of a label (level, user, ...) >> are also available for filtering. If someone wrote a simple Bell & >> LaPadula LSM filtering by some of those fields could be useful >> there, too. >> >> I would like guidance on whether I ought to go the route of >> more extensive use of WITH_APPARMOR (and WITH_SMACK, WITH_MUMBLE) >> or take the path of greater generalization. Or, whether I should >> treat each case individually and give it my best whack. > > To be honest, I have no idea how well the audit system works with any MAC > system except SE Linux. Understood. Part of what I'm looking at is ensuring that as multiple concurrent LSMs come in that the audit user-space isn't mucked up. ausearch has these options: -o,--object <SE Linux Object context> -se,--context <SE Linux context> -su,--subject <SE Linux context> Without multiple LSMs we can easily ignore "SE Linux" in these options and use whatever kind of "context" is available. If I have SELinux and AppArmor, the implication is that you can't search on AppArmor information. Should we be adding -aa,--apparmorcontext <AppArmor context> -as,--apparmorsubject <AppArmor subject context> or should we change -se to look at all "contexts", and change the description to reflect that? Basicaly, I'm asking whether you'd rather add options for other LSMs or remove descriptions that specify SELinux.

> I don't really know if its doing the right thing. > > Ausearch and report share a parser. It is time sensitive. I usually test > it on 4 or 5 Gb of logs. We also have the ausearch-test program which > can be used to test any changes to the parser. > > http://people.redhat.com/sgrubb/audit/ausearch-test-0.6.tar.gz > > Once that is squared away, there is the auparse library. It has a table > that classifies a field name into what it is for interpretation > purposes. You will find a #ifdef WITH_APPARMOR. I don't know if that > table is complete or if it needs to be extended for any other MAC > system. > > That then leads to the auparse normalizer. I don't know if we need to > make > any changes there. You can trigger its code with ausearch --format csv or > -- format text. > > Also, we have some size limits in user space. How big can an event record > be if the file is MAX_PATH name length and it has a space in its name or > directory and each context is it's maximum size? We may need to think > about how this might change the whole userspace ecosystem's size > definition, MAX_AUDIT_MESSAGE_LENGTH, since this is part of the ABI. And > the kernel also has AUDIT_MESSAGE_TEXT_MAX. What would you get with: > > # /usr/sbin/auditctl -m `perl -e 'print "A"x8880'` > > And last...what about auditctl? Is the syscall filter going to allow > filtering on these other subject/object components? > > -Steve