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

[Peter Zijlstra <peterz@infradead.org>]

On Fri, 2009-11-27 at 13:21 +0100, Peter Zijlstra wrote:
> +static struct rq *
> +balance_task(struct task_struct *p, int sd_flags, int wake_flags)
> +{
> +       struct rq *rq, *old_rq;
> +       u64 vdelta;
> +
> +               rq = old_rq = task_rq(p);
> +
> +       if (p->sched_class == &fair_sched_class)
> +               vdelta = task_cfs_rq(p)->min_vruntime;
> +
> +       __task_rq_unlock(old_rq);
> +
> +       cpu = select_task_rq(p, sd_flags, wake_flags);
> +
> +       rq = cpu_rq(cpu);
> +       spin_lock(&rq->lock);
> +       if (rq == old_rq)
> +               return rq;
> +
> +       update_rq_clock(rq);
> +
> +       set_task_cpu_all(p, task_cpu(p), cpu);
> +
> +       if (p->sched_class == &fair_sched_class) {
> +               vdelta -= task_cfs_rq(p)->min_vruntime;
> +               p->se.vruntime -= vdelta;
>         }
>  
> +       return rq;
> +} 

Feh, there's a much easier way to deal with that min_vruntime crap.

Do se->vruntime -= cfs_rq->min_vruntime, on dequeue, and
   se->vruntime += cfs_rq->min_vruntime, on enqueue

That leaves the whole thing invariant to the cfs_rq when its not
enqueued, and we don't have to fix it up when moving it around.

Also, note that I ripped out the clock_offset thingy, because with the
current sched_clock.c stuff clocks should get synchronized when we do a
remote clock update (or at least appear monotonic).

Getting these two things sorted returns set_task_cpu() to sanity.

/me tosses patch and starts over.