Re: [patch] sched: fix b5d9d734 blunder in task_new_fair()

[Mike Galbraith <efault@gmx.de>]

On Tue, 2009-11-24 at 19:27 +0100, Peter Zijlstra wrote:
> On Tue, 2009-11-24 at 19:21 +0100, Mike Galbraith wrote:
> > On Tue, 2009-11-24 at 18:54 +0100, Peter Zijlstra wrote:
> > > On Tue, 2009-11-24 at 18:35 +0100, Peter Zijlstra wrote:
> > > > On Tue, 2009-11-24 at 18:07 +0100, Mike Galbraith wrote:
> > > > > >       if (p->sched_class->task_new) {
> > > > > >               /* can detect migration through: task_cpu(p) != smp_processor_id() */
> > > > > 
> > > > > What if the parent was migrated before we got here?
> > > > 
> > > > Well, the only case it really matters for is the child_runs_first crap,
> > > > which is basically broken on SMP anyway, so I don't think we care too
> > > > much if we race here.
> > > > 
> > > > Unless I missed some detail that is ;-)
> > > 
> > > 
> > > Also, we're running all this from the parent context, and we have
> > > preemption disabled, we're not going anywhere.
> > 
> > In sched_fork() and wake_up_new_process(), but in between?
> 
> Hmm, right, back to the previous argument then ;-)

Yeah.

We can be preempted between original task struct copy and getting to
sched_fork(), and after leaving copy_process(), so I don't see any way
around lock parent, update and copy vruntime.  Whether we race in
placing the child wrt parent isn't a big deal, but the child's vruntime
is, as is fiddling with the parent's task struct and runqueue.

	-Mike