Re: [PATCH 1/2] audit: stop an old auditd being starved out by a new auditd

[Richard Guy Briggs <rgb@redhat.com>]

On 15/09/28, Paul Moore wrote:
> On Monday, September 28, 2015 07:17:31 AM Richard Guy Briggs wrote:
> > On 15/09/25, Paul Moore wrote:
> > > The audit_make_reply() function is the wrong thing to be using here, we
> > > should create our own buffer from scratch like most other records.  Also,
> > > yes, we want to include the new pid, but I really don't think there is
> > > any value in including the seqno of the AUDIT_SET/AUDIT_STATUS_PID
> > > message.
> > 
> > Most other records use audit_log_start(), which isn't what we want here,
> > since we want to bypass the queue to test if it is still alive.  We
> > don't care if it is delivered.  We just care if the socket is still
> > alive.  We don't want a context either.
> 
> Yes, that is why I mentioned creating the buffer from scratch.
> 
> > So, I believ audit_make_reply() can be used just fine, setting portid,
> > seq, done and multi to zero.
> 
> The 'multi' flag should definitely be set to zero, 'seq' is fine at zero, but 
> I think we can do better with 'portid'; we know the 'portid' value so just use 
> it in the call to audit_make_reply().

Most other audit_log_start() created messages set portid to zero except
user messages, and those are set using the initiating process' portid
and not the destination id.  So here I think portid should be zero.  The
target task should know its own portid and the netlink field for portid
isn't used for routing to that destination that I can discern from the
netlink code.

> I don't like that we are reusing audit_make_reply() for non-reply netlink 
> messages, but I'll get over that.  This will likely get a revamp when we get 
> around to a proper fix of the queuing system.

This could even be renamed audit_make_message() and possibly be
generalized to be useful to audit_log_start(), or rather
audit_buffer_alloc().  Later...

> > > > > Also, this is more of a attempted hijack message and not a
> > > > > simple ping, right?
> > > > 
> > > > Ok, so maybe AUDIT_PING is not the appropriate name for it.  I don't
> > > > have a problem changing it, but I think the pid of the hijacker would be
> > > > useful information to the ping-ee unless the ping message was only ever
> > > > issues in a contextless kernel-initiated message.
> > > 
> > > Let's change the message name, this isn't a ping message and we may want
> > > to have a ping message at some point in the future.
> > 
> > Ok, how about AUDIT_HIJACK_TEST, with a payload of the u32
> > representation of the PID of the task attempting to replace it.
> 
> Why add the TEST?  It is a hijack attempt, or at least it is if the record is 
> emitted successfully :)  I would go simply with AUDIT_HIJACK or maybe 
> AUDIT_REPLACE (or similar) if "hijack" is a bit too inflammatory (it probably 
> is ...).

I had actually named it AUDIT_REPLACE_TEST, but your repeated use of the
term "hijack" swayed me...  I'd still lean towards *_TEST since it is
testing to replace a stale socket and not a live one.

> paul moore

- RGB

--
Richard Guy Briggs <rbriggs@redhat.com>
Senior Software Engineer, Kernel Security, AMER ENG Base Operating Systems, Red Hat
Remote, Ottawa, Canada
Voice: +1.647.777.2635, Internal: (81) 32635, Alt: +1.613.693.0684x3545