[RFC][PATCH] audit: add feature audit_lost reset

Fri Dec 9 07:00:14 UTC 2016

On 2016-12-08 09:05, Paul Moore wrote:
> On Wed, Dec 7, 2016 at 10:53 PM, Richard Guy Briggs <rgb at redhat.com> wrote:
> > On 2016-12-07 18:45, Paul Moore wrote:
> >> On Wed, Dec 7, 2016 at 6:30 PM, Steve Grubb <sgrubb at redhat.com> wrote:
> >> > On Wednesday, December 7, 2016 6:10:49 PM EST Paul Moore wrote:
> >> >> On Wed, Dec 7, 2016 at 10:58 AM, Richard Guy Briggs <rgb at redhat.com> wrote:
> >> >> > On 2016-12-07 10:53, Steve Grubb wrote:
> >> >> >> On Wednesday, December 7, 2016 10:05:30 AM EST Paul Moore wrote:
> >> >> >> > On Tue, Dec 6, 2016 at 10:32 PM, Richard Guy Briggs <rgb at redhat.com>
> >> > wrote:
> >> >> >> > > On 2016-12-06 19:17, Paul Moore wrote:
> >> >> >> > >> On Tue, Dec 6, 2016 at 12:13 AM, Richard Guy Briggs <rgb at redhat.com>
> >> >> >> > >> Okay, back up ... this whole mess about atomic_xchg() was always
> >> >> >> > >> unrelated to my original suggestion, let's focus on my original
> >> >> >> > >> comment ... don't reset the counter on a AUDIT_GET, reset it on a
> >> >> >> > >> AUDIT_SET with an AUDIT_STATUS_LOST, does that make sense?
> >> >> >> > >
> >> >> >> > > I understood that.  It sounds like a nice simple and straightforward
> >> >> >> > > method to do it but for the question of accuracy.  Please rewind to
> >> >> >> > > my
> >> >> >> > > fundamental point: How do we get an accurate reading of the last
> >> >> >> > > value
> >> >> >> > > of audit_lost before resetting it?
> >> >> >> >
> >> >> >> > Okay, I thought you were worried about a different race, which is why
> >> >> >> > this discussion wasn't making much sense to me.  I understand your
> >> >> >> > point, but I really dislike the API; although that's not your fault,
> >> >> >> > it's really the only way to do it via AUDIT_GET.
> >> >> >> >
> >> >> >> > I'd much prefer we go with the cleaner AUDIT_SET approach and just not
> >> >> >> > worry about the small race window.  It would only be an issue if you
> >> >> >> > reset the count under heavy audit load, and why would you reset the
> >> >> >> > lost value if you were under a heavy audit load?  That just doesn't
> >> >> >> > make sense.
> >> >> >> >
> >> >> >> > I suppose we should hear from Steve on this since he was the one who
> >> >> >> > has been asking for this feature, although I'm pretty sure I know what
> >> >> >> > he is going to say.
> >> >> >>
> >> >> >> To start with, this request comes from users of the audit system. I just
> >> >> >> passed along the request. The issue is that when you do auditctl -s, you
> >> >> >> get the number of records lost. If you do it the next day, you have to
> >> >> >> do math to see what the one day delta is. So, to make reporting easy,
> >> >> >> they want it to be reset whenever they do audictl -s.
> >> >> >>
> >> >> >> You could also make a AUDIT_GET_RESET that gets the status and resets the
> >> >> >> number atomically. Then I can add another commandline option to auditctl
> >> >> >> that allows an admin to say also reset the counters. If that command
> >> >> >> line option is passed, I call AUDIT_GET_RESET otherwise I call
> >> >> >> AUDIT_GET. Thought?>
> >> >> > This would be slightly simpler in kernel implementation than the method
> >> >> > I proposed and would work fine, off the top of my head.
> >> >>
> >> >> I'd prefer not to introduce another command message type for something
> >> >> small like this.
> >> >>
> >> >> Steve, do you have any objection to the AUDIT_SET based approach?
> >> >
> >> > Either way, we'd need a feature flag so that I can tell if the kernel supports
> >> > this or not.
> >>
> >> I think we are okay without a specific feature flag as sending a
> >> AUDIT_SET/reset on an old kernel will be harmless; it won't do
> >> anything, but it shouldn't return an error either.
> >
> > Ok, so userspace is still left wondering if it worked until the next
> > time it reads that value, and even then it can't be certain if that
> > value is the same or higher than it was when userspace thought it reset
> > that value.  At this point, an old kernel will not return an error and
> > simply ignore any new AUDIT_STATUS_* flag since each flag is treated
> > independently and extra flags are ignored and not blocked.  This seems
> > sloppy since we have two ways of fixing this uncertainty pretty easily.
> 
> Comments like the above aren't helpful, they are just annoying.  The
> drawbacks to the AUDIT_SET/reset approach have already been discussed
> on this thread, if you want to do something constructive think about
> how you can resolve these limitations within the context of others
> comments/feedback.

I'm sorry to offend.  That wasn't my intent.  I was trying to bring new
information and observations into the discussion.  Is there anything
factually wrong in my observations other than the opinion of it being
sloppy?

> I've already mentioned that I didn't like the AUDIT_GET/reset approach
> because I thought the interface was bad.  As I'm sure you know, the
> audit kernel/userspace interface is a bit of a hot-button topic with
> me; I think it has a lot of problems and I'm very intent on not making
> it worse (in my opinion, I will admit that API design is not entirely
> objective).  Continuing to argue for a interface design that I've
> already expressed a dislike for is not likely to win me over to your
> side; regardless of the outcome you will end up frustrating both
> yourself and the maintainer, neither are good things.

I've re-read the thread from the beginning.  I guess I must have missed
what was the fundamental problem with the AUDIT_GET/reset method other
than taste.  What don't you like about the API that precludes
using/abusing it this way?  Is it the issue that a GET would
surprisingly change a value rather than just reading it?  I didn't like
it either, but it was the next obvious way to tackle the issue without
losing information.

> We all agree that there is a potential race window with respect to
> reading/resetting the lost counter, where we disagree is the
> likelihood of that happening in practice.

We've had other races that looked pretty unlikely in practice that were
deemed to be unacceptable.  Granted the risks were higher if they were
exploited, but they were mitigated to the best of our ability.

None of this should matter now since there are ideas below that should
work.

> You feel very strongly that
> the window is of grave concern, Steve and myself much less so.  If you
> still feel strongly about this, think about some different ways in
> which you can avoid losing a lost message counter bump.  Off the top
> of my head, there are really only two ways for the kernel's audit
> subsystem to send information back to userspace in this case, via a
> netlink return/error message or an audit record.  We could possibly do
> something with the netlink error message by returning the lost counter
> as a positive integer (negative integer is a failure code, zero is
> success), but that might get tricky in the future, although we could
> mitigate that risk by forcing the AUDIT_SET/reset to happen by itself
> (in other words, don't simply check to see if the bit is set in the
> bitmask, e.g. (s.mask & AUDIT_STATUS_LOST), check to see it is equal,
> e.g. (s.mask == AUDIT_STATUS_LOST)).

This could work.  What risk do you see in doing it with other flags?
That another set failure could usurp the return code?  If so, yes, I
agree with requiring it to be a lone flag.

> We could also mitigate the race
> via an audit record by emitting a record indicating that the lost
> counter was reset and record the lost counter (before the reset) in
> that record; honestly, now that I'm writing this, it seems like
> something we should be doing regardless, as tampering with the lost
> counter seems like a security relevant event.

Agreed.  This should be added regardless of reset method.

> There you go, two possible solutions for eliminating/mitigating the
> potential race while sticking with the simpler AUDIT_SET/reset
> interface.  I suppose you could even implement both of the solutions
> above, they aren't mutually exclusive; that would depend on what
> Steve/userspace would prefer.  Finally, as for the feature bitmap to
> signal to userspace that we support this new feature: if you can't
> live without it, go ahead and add it in.  As I said before, I'm a
> little concerned at the rate we are consuming this bitmap, but I'll
> admit we still have plenty of room before we have to start worrying
> about alternatives.

I would suggest that the return value (presuming it was reset when
non-zero) or the audit record generated reporting the lost value
reset would be sufficient confirmation that the feature exists on the
running kernel and the addition to the feature bitmap is not strictly
necessary, but you only find this out upon attempting that lost reset.

Well, we haven't used much of that bitmap space and if it isn't to be
used when needed, why is it there?  If there is a relatively simple
alternate non-destructive way to discover the presence of a feature use
of the bitmap isn't necessary.

> paul moore

- RGB

--
Richard Guy Briggs <rgb at redhat.com>
Kernel Security Engineering, Base Operating Systems, Red Hat
Remote, Ottawa, Canada
Voice: +1.647.777.2635, Internal: (81) 32635