Richard Guy Briggs <rgb@xxxxxxxxxx> wrote:
> Well, we are only logging "some change", so is it necessary to log the
> generation count to show that? Is the generation count of specific
> interest?
No, its of no specific interest. I just worded this poorly.
If the generation id increments, then something has been changed by the
batch, thats all.
> > (After that, kernel can't back down anymore, i.e. all errors are
> > caught/handled beforehand).
>
> I did think of recording all failed attempts too, but coding that would
> be more effort. It is worth doing if it is deemed important,
> particularly for permission issues (as opposed to resource limits or
> packet format errors. This would be more of interest to a security
> officer rather than a network technician, but the latter may find it
> useful for debugging.
The permission check is done early, in nfnetlink_rcv() (search for
EPERM), you would need to add an audit call there if thats relevant
for audit purposes.
> > If its 'any config change', then you also need to handle adds
> > or delete from sets/maps, since that may allow something that wasn't
> > allowed before, e.g. consider
> >
> > ip saddr @trused accept
> >
> > and then, later on,
> > nft add element ip filter @trusted { 10.0.0.0/8, 192.168.0.1 }
> >
> > This would not add a table, or chain, or set, but it does implicitly
> > alter the ruleset.
>
> Ah, ok, so yes, we would need that too. I see family and table in
> there, op is evident. Is there a useful value we can use in the
> "entries" field?
Maybe the handle of the set that the element was added to.
Each set, rule, chain, ... has a kernel-assigned number that
serves as a unique identifier.
> > Is that record format expected to emit the current number of chains?
>
> I was aiming for a relevant value such as perhaps the new rule number or
> the rule number being deleted.
In that case, use the handle, which is a u64 with a unique value (for a
given table).
> > Since table names can be anything in nf_tables (they have no special
> > properties anymore), the table name is interesting from a informational
> > pov, but not super interesting.
>
> I don't think we need to be able to completely reconstruct the
> tables/chains/rules from the information in the audit log, but be aware
> of who is changing what when.
Ok. Have a look at nf_tables_fill_gen_info() in that case, you probably
want to emit at least the pid and task info, unless audit doesn't add
that already anyway.
> > Consider a batch update that commits 100 new rules in chain x,
> > this would result in 100 audit_log_nfcfg() calls, each with the
> > same information.
>
> So rule number would be a useful differentiator here.
Ok. Yes, that is available (rule->handle).
