[RFC PATCH ghak10 2/3] audit: Add the audit_adjtime() function

[Ondrej Mosnacek]

ut 19. 6. 2018 o 22:52 Steve Grubb <sgrubb at redhat.com> napísal(a):
> On Tuesday, June 19, 2018 4:57:37 AM EDT Ondrej Mosnacek wrote:
> > 2018-06-18 20:39 GMT+02:00 Steve Grubb <sgrubb at redhat.com>:
> > > On Friday, June 15, 2018 8:45:22 AM EDT Ondrej Mosnacek wrote:
> > >> This patch adds a new function that shall be used to log any
> > >> modification of the system's clock (via the adjtimex() syscall).
> > >>
> > >> The function logs an audit record of type AUDIT_TIME_ADJUSTED with the
> > >> following fields:
> > >> * txmodes  (the 'modes' field of struct timex)
> > >> * txoffset (the 'offset' field of struct timex)
> > >> * txfreq   (the 'freq' field of struct timex)
> > >> * txmaxerr (the 'maxerror' field of struct timex)
> > >> * txesterr (the 'esterror' field of struct timex)
> > >> * txstatus (the 'status' field of struct timex)
> > >> * txconst  (the 'constant' field of struct timex)
> > >> * txsec, txusec (the 'tv_sec' and 'tv_usec' fields of the 'time' field
> > >> of struct timex (respectively))
> > >> * txtick   (the 'tick' field of struct timex)
> > >
> > > Are all of these fields security relevant? Primarily what we need to know
> > > is if time is being adjusted. This is relevant because a bad guy may
> > > adjust system time to make something appear to happen earlier or later
> > > than it really did which make correlation hard or impossible.
> >
> > This is still an open question for me... On the one hand, we might
> > want to know exactly what the bad guy was trying to do ("He changed
> > the offset to 500 ms." vs. just "He adjusted the clock."), on the
> > other hand, we may not really care and consider it yet another junk
> > data in the logs... A possible compromise could be to log only
> > relevant fields (see 'Possible improvements' in the commit message).
> > Assuming ntpd or other authorized applications would only modify
> > one/few variables at a time, this would add only a few fields to the
> > record each time.
>
> I think we want the modes field so that we know what was changed. But do we
> really need to know maxerror or status? I think we should limit this to the
> modes and any value of a time adjustment.

Well, the status field can be at the very least used to set the
STA_INS/STA_DEL status flags [1], which seem to coordinate the
insertion of leap second. Now this may be a bit far-fetched, but I can
imagine how bogus insertions of leap seconds could be used to
gradually shift the clock to a bad value. I think we can safely drop
maxerror/esterror (I believe these are just informational of how much
the clock is 'out of sync'), but the rest seems like it could at least
indirectly influence the time (IOW, I wasn't able to find a proof that
it can't...).

[1] https://elixir.bootlin.com/linux/v4.17/source/include/uapi/linux/timex.h#L128

>
> > Note that this new auxiliary record gets only logged on *modifying*
> > operations, which should not be that frequent, and thus it shouldn't
> > be a problem to output a bit of potentially useful information.
>
> We're after just security information.
>
> > That said, I don't mind logging just an empty record if that is
> > preferred.
>
> An empty buffer doesn't tell us what was adjusted.

Yes, sorry, that should have been "a record with only modes field".

>
> Thanks,
> -Steve

>
>--
Ondrej Mosnacek <omosnace at redhat dot com>
Associate Software Engineer, Security Technologies
Red Hat, Inc.