Re: [PATCH RFC] reduce runqueue lock contention

[Chris Mason <chris.mason@oracle.com>]

On Thu, May 20, 2010 at 11:23:12PM +0200, Peter Zijlstra wrote:
> On Thu, 2010-05-20 at 23:09 +0200, Peter Zijlstra wrote:
> 
> > int try_to_wake_up(struct task *p, unsigned int mask, wake_flags)
> > {
> >   int state = atomic_read(&p->state);
> > 
> >   do {
> >     if (!(state & mask))
> >       return 0;
> > 
> >     state = atomic_cmpxchg(&p->state, state, TASK_WAKING);
> >   } while (state != TASK_WAKING);
> 
>     cpu = select_task_rq()
> 
> and then somehow see we get set_task_cpu() done without races :-)
> 
> >   /* do this pending queue + ipi thing */
> > 
> >   return 1;
> > }

I tried not to set the task waking, since I was worried about races with
us getting queued somewhere else.  But, I don't have a good handle on
all of that so I kind of chickened out.  That's why my code falls back
to the full ttwu in a few cases.

Do you think the above could be an addition to my patch or that it's
required for my patch to work well?

-chris