Re: [Lse-tech] Re: [PATCH] O(1) scheduler set_cpus_allowed for non-current tasks

[Rusty Russell <rusty@rustcorp.com.au>]

On Mon, 25 Feb 2002 19:59:49 -0800
Paul Jackson <pj@engr.sgi.com> wrote:

> On Mon, 25 Feb 2002, Dipankar Sarma wrote:
> >
> > If these are processes that are bound to the CPU to be shut down,
> > wouldn't it make sense to fail the CPU shut down operation ? If you
> > are giving enough control to the user to make CPU affinity decisions,
> > they better know how to cleanup before shutting down a CPU.
> 
> I can imagine some users (applications) wanting to insist on
> staying on a particular CPU (Pike's Peak or Bust), and some
> content to be migrated automatically, and some wanting to
> receive and act on requests to migrate.
> 
> One of these policies might be default, with others as options.
> 
> Some CPU shut down operations _can't_ fail ... if they are motivated
> say by hardware about to fail.

Exactly.  If I run the RC5 challenge, one per cpu (using a mythical
oncpu(1) program, say), I'd be very upset if my whole machine dies because
it don't take down a faulty CPU!

I think SIGPWR is appropriate here.  If that doesn't work, then SIGKILL.
Of course, ksoftirqd is a special case, etc.

Rusty.
-- 
  Anyone who quotes me in their sig is an idiot. -- Rusty Russell.