Re: [PATCH v2 05/13] sched: Enable SD_BALANCE_WAKE for asymmetric capacity systems
From: Morten Rasmussen
Date: Tue Jul 12 2016 - 10:24:49 EST
On Mon, Jul 11, 2016 at 01:24:04PM +0200, Peter Zijlstra wrote:
> On Mon, Jul 11, 2016 at 12:04:58PM +0100, Morten Rasmussen wrote:
>
> > One alternative to setting ASYM_CAP bottom up would be to set it only
> > where the asymmetry can be observed, and instead come up with a more
> > complicated way of setting BALANCE_WAKE bottom up until and including
> > the first level having the ASYM_CAP.
>
> Right, that is what I was thinking.
>
> > I looked at it briefly an realized that I couldn't find a clean way of
> > implementing it as I don't think we have visibility of which flags that
> > will be set at higher levels in the sched_domain hierarchy when the
> > lower levels are initialized. IOW, we have behavioural flags settings
> > depend on topology flags settings at a different level.
>
> Looks doable if we pass @child into sd_init() in build_sched_domain().
> Then we could simply do:
>
> *sd = (struct sched_domain){
> /* ... */
> .child = child,
> };
>
> if (sd->flags & ASYM_CAP) {
> struct sched_domain *t = sd;
> while (t) {
> t->sd_flags |= BALANCE_WAKE;
> t = t->child;
> }
> }
>
> Or something like that.
It appears to be working fine. I will roll it into v3 along with the
simpler and more sane ASYM_CAP semantics :)