> It is a serious concern.  Inventing new, subtle behavior differences 
> between user and kernel threads is, in a word, gross.  It's certain
> to bite people in the future.
So you are suggesting that it's better to slow down *all* threads so
that it's possible to have kernel threads with automatically shared
credentials?
BTW, signals and rescheduling (unless PREEMPT=y) don't work
automatically for the same reason.