Re: [PATCH ghak25 v4 3/3] audit: add subj creds to NETFILTER_CFG record to cover async unregister

[Paul Moore <paul@paul-moore.com>]

On Wed, May 6, 2020 at 6:43 PM Richard Guy Briggs <rgb@redhat.com> wrote:
> On 2020-05-06 17:26, Steve Grubb wrote:
> > On Wednesday, April 29, 2020 5:32:47 PM EDT Richard Guy Briggs wrote:
> > > On 2020-04-29 14:47, Steve Grubb wrote:
> > > > On Wednesday, April 29, 2020 10:31:46 AM EDT Richard Guy Briggs wrote:
> > > > > On 2020-04-28 18:25, Paul Moore wrote:
> > > > > > On Wed, Apr 22, 2020 at 5:40 PM Richard Guy Briggs <rgb@redhat.com>
> > > >
> > > > wrote:
> > > > > > > Some table unregister actions seem to be initiated by the kernel to
> > > > > > > garbage collect unused tables that are not initiated by any
> > > > > > > userspace
> > > > > > > actions.  It was found to be necessary to add the subject
> > > > > > > credentials
> > > > > > > to  cover this case to reveal the source of these actions.  A
> > > > > > > sample
> > > > > > > record:
> > > > > > > type=NETFILTER_CFG msg=audit(2020-03-11 21:25:21.491:269) :
> > > > > > > table=nat
> > > > > > > family=bridge entries=0 op=unregister pid=153 uid=root auid=unset
> > > > > > > tty=(none) ses=unset subj=system_u:system_r:kernel_t:s0
> > > > > > > comm=kworker/u4:2 exe=(null)>
> > > > > >
> > > > > > [I'm going to comment up here instead of in the code because it is a
> > > > > > bit easier for everyone to see what the actual impact might be on the
> > > > > > records.]
> > > > > >
> > > > > > Steve wants subject info in this case, okay, but let's try to trim
> > > > > > out
> > > > > > some of the fields which simply don't make sense in this record; I'm
> > > > > > thinking of fields that are unset/empty in the kernel case and are
> > > > > > duplicates of other records in the userspace/syscall case.  I think
> > > > > > that means we can drop "tty", "ses", "comm", and "exe" ... yes?
> > > > >
> > > > > From the ghak28 discussion, this list and order was selected due to
> > > > > Steve's preference for the "kernel" record convention, so deviating
> > > > > from this will create yet a new field list.  I'll defer to Steve on
> > > > > this. It also has to do with the searchability of fields if they are
> > > > > missing.
> > > > >
> > > > > I do agree that some fields will be superfluous in the kernel case.
> > > > > The most important field would be "subj", but then "pid" and "comm", I
> > > > > would think.  Based on this contents of the "subj" field, I'd think
> > > > > that "uid", "auid", "tty", "ses" and "exe" are not needed.
> > > >
> > > > We can't be adding deleting fields based on how its triggered. If they
> > > > are unset, that is fine. The main issue is they have to behave the same.
> > >
> > > I don't think the intent was to have fields swing in and out depending
> > > on trigger.  The idea is to potentially permanently not include them in
> > > this record type only.  The justification is that where they aren't
> > > needed for the kernel trigger situation it made sense to delete them
> > > because if it is a user context event it will be accompanied by a
> > > syscall record that already has that information and there would be no
> > > sense in duplicating it.
> >
> > We should not be adding syscall records to anything that does not result from
> > a syscall rule triggering the event. Its very wasteful. More wasteful than
> > just adding the necessary fields.
>
> So what you are saying is you want all the fields that are being
> proposed to be added to this record?
>
> If the records are all from one event, they all should all have the same
> timestamp/serial number so that the records are kept together and not
> mistaken for multiple events.  One reason for having information in
> seperate records is to be able to filter them either in kernel or in
> userspace if you don't need certain records.

Yes, I'm opposed to duplicating fields across records in a single
event.  If there are cases where we have a standalone record, such as
with "unregister", then there is an argument to be made about
duplicating some fields that are important in the standalone
unregister case.  However, this is *only* for those fields which make
sense in the standalone kernel unregister event; if the field isn't
useful in this unregister corner case *and* it is duplicated in
another record type which normally accompanies this record in an event
there is no reason it needs to be in this record.

> > I also wished we had a coding specification that put this in writing so that
> > every event is not a committee decision. That anyone can look at the document
> > and Do The Right Thing ™.
> >
> > If I add a section to Writing-Good-Events outlining the expected ordering of
> > fields, would that be enough that we do not have long discussions about event
> > format? I'm thinking this would also help new people that want to contribute.

To be clear, we are not changing any existing record formats; they are
part of the kernel/userspace ABI and changing them would break the
ABI.

In a perfect world both the audit kernel and userspace would have been
designed, implemented, and documented better.  Unfortunately it wasn't
and we have to live with what we have.

-- 
paul moore
www.paul-moore.com