Date:   Tue, 27 Oct 2020 21:35:17 -0400
From:   Paul Moore <paul@...l-moore.com>
To:     Richard Guy Briggs <rgb@...hat.com>
Cc:     nhorman@...driver.com, linux-api@...r.kernel.org,
        containers@...ts.linux-foundation.org,
        LKML <linux-kernel@...r.kernel.org>, dhowells@...hat.com,
        Linux-Audit Mailing List <linux-audit@...hat.com>,
        netfilter-devel@...r.kernel.org, ebiederm@...ssion.com,
        simo@...hat.com, netdev@...r.kernel.org,
        linux-fsdevel@...r.kernel.org, Eric Paris <eparis@...isplace.org>,
        mpatel@...hat.com, Serge Hallyn <serge@...lyn.com>
Subject: Re: [PATCH ghak90 V9 05/13] audit: log container info of syscalls

On Fri, Oct 23, 2020 at 4:40 PM Richard Guy Briggs <rgb@...hat.com> wrote:
> On 2020-10-22 21:21, Paul Moore wrote:
> > On Wed, Oct 21, 2020 at 12:39 PM Richard Guy Briggs <rgb@...hat.com> wrote:
> > > Here is an exmple I was able to generate after updating the testsuite
> > > script to include a signalling example of a nested audit container
> > > identifier:
> > >
> > > ----
> > > type=PROCTITLE msg=audit(2020-10-21 10:31:16.655:6731) : proctitle=/usr/bin/perl -w containerid/test
> > > type=CONTAINER_ID msg=audit(2020-10-21 10:31:16.655:6731) : contid=7129731255799087104^3333941723245477888
> > > type=OBJ_PID msg=audit(2020-10-21 10:31:16.655:6731) : opid=115583 oauid=root ouid=root oses=1 obj=unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023 ocomm=perl
> > > type=CONTAINER_ID msg=audit(2020-10-21 10:31:16.655:6731) : contid=3333941723245477888
> > > type=OBJ_PID msg=audit(2020-10-21 10:31:16.655:6731) : opid=115580 oauid=root ouid=root oses=1 obj=unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023 ocomm=perl
> > > type=CONTAINER_ID msg=audit(2020-10-21 10:31:16.655:6731) : contid=8098399240850112512^3333941723245477888
> > > type=OBJ_PID msg=audit(2020-10-21 10:31:16.655:6731) : opid=115582 oauid=root ouid=root oses=1 obj=unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023 ocomm=perl
> > > type=SYSCALL msg=audit(2020-10-21 10:31:16.655:6731) : arch=x86_64 syscall=kill success=yes exit=0 a0=0xfffe3c84 a1=SIGTERM a2=0x4d524554 a3=0x0 items=0 ppid=115564 pid=115567 auid=root uid=root gid=root euid=root suid=root fsuid=root egid=root sgid=root fsgid=root tty=ttyS0 ses=1 comm=perl exe=/usr/bin/perl subj=unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023 key=testsuite-1603290671-AcLtUulY
> > > ----
> > >
> > > There are three CONTAINER_ID records which need some way of associating with OBJ_PID records.  An additional CONTAINER_ID record would be present if the killing process itself had an audit container identifier.  I think the most obvious way to connect them is with a pid= field in the CONTAINER_ID record.
> >
> > Using a "pid=" field as a way to link CONTAINER_ID records to other
> > records raises a few questions.  What happens if/when we need to
> > represent those PIDs in the context of a namespace?  Are we ever going
> > to need to link to records which don't have a "pid=" field?  I haven't
> > done the homework to know if either of these are a concern right now,
> > but I worry that this might become a problem in the future.
>
> Good point about PID namespaces in the future but those accompanying
> records will already have to be conditioned for the PID namespace
> context that is requesting it, so I don't see this as a showstopper.

Possibly, it just gets very messy.  Doubly so when you start looking
at potentially adjusting for multiple audit daemons.  Thankfully it
doesn't look like using the PID is a good idea for other reasons.

> I've forgotten about an important one we already hit, which is a network
> event that only has a NETFILTER_PKT record, but in that case, there is
> no ambiguity since there are no other records associated with that
> event.  So the second is already an issue now.  Using
> task_tgid_nr(current), in the contid testsuite script network event it
> attributed it to ping which caused the event, but we cannot use this
> since it wasn't triggered by a syscall and doesn't accurately reflect
> the kernel thread that received it.  It could just be set to zero for
> network events.

Possibly.  It just seems like too much hackery to start; that's the
stuff you do once it has been in a kernel release for years and need
to find a workaround that doesn't break everything.  I think we should
aim a bit higher right now.

> > The idea of using something like "item=" is interesting.  As you
> > mention, the "item=" field does present some overlap problems with the
> > PATH record, but perhaps we can do something similar.  What if we
> > added a "record=" (or similar, I'm not worried about names at this
> > point) to each record, reset to 0/1 at the start of each event, and
> > when we needed to link records somehow we could add a "related=1,..,N"
> > field.  This would potentially be useful beyond just the audit
> > container ID work.
>
> Does it make any sense to use the same keyword in each type of record
> such as record/records as in PATH/SYSCALL: item/items ?

That was mentioned above, if you can assure yourself and the rest of
us that it can be safely reused I think that might be okay, but I'm
not convinced that is safe at the moment.  Although I will admit those
are fears are not based on an exhaustive search through the code (or a
determined "think").

> (I prefer 0-indexed like item=...)

I have no preference on where we start the index, but it makes sense
to keep the same index starting point as the PATH records.

-- 
paul moore
www.paul-moore.com

Powered by blists - more mailing lists