audit: fix a race condition with the auditd tracking code

Originally reported by Adam and Dusty, it appears we have a small
race window in kauditd_thread(), as documented in the Fedora BZ:

 * https://bugzilla.redhat.com/show_bug.cgi?id=1459326#c35

 "This issue is partly due to the read-copy nature of RCU, and
  partly due to how we sync the auditd_connection state across
  kauditd_thread and the audit control channel.  The kauditd_thread
  thread is always running so it can service the record queues and
  emit the multicast messages, if it happens to be just past the
  "main_queue" label, but before the "if (sk == NULL || ...)"
  if-statement which calls auditd_reset() when the new auditd
  connection is registered it could end up resetting the auditd
  connection, regardless of if it is valid or not.  This is a rather
  small window and the variable nature of multi-core scheduling
  explains why this is proving rather difficult to reproduce."

The fix is to have functions only call auditd_reset() when they
believe that the kernel/auditd connection is still valid, e.g.
non-NULL, and to have these callers pass their local copy of the
auditd_connection pointer to auditd_reset() where it can be compared
with the current connection state before resetting.  If the caller
has a stale state tracking pointer then the reset is ignored.

We also make a small change to kauditd_thread() so that if the
kernel/auditd connection is dead we skip the retry queue and send the
records straight to the hold queue.  This is necessary as we used to
rely on auditd_reset() to occasionally purge the retry queue but we
are going to be calling the reset function much less now and we want
to make sure the retry queue doesn't grow unbounded.

Reported-by: Adam Williamson <awilliam@redhat.com>
Reported-by: Dusty Mabe <dustymabe@redhat.com>
Reviewed-by: Richard Guy Briggs <rgb@redhat.com>
Signed-off-by: Paul Moore <paul@paul-moore.com>