Re: [PATCHSET] workqueue: reimplement CPU hotplug to keep idle workers

[Peter De Schrijver <pdeschrijver@nvidia.com>]

On Fri, Jul 20, 2012 at 09:44:04PM +0200, Rafael J. Wysocki wrote:
> On Friday, July 20, 2012, Tejun Heo wrote:
> > Hello,
> > 
> > On Fri, Jul 20, 2012 at 08:22:30PM +0200, Peter Zijlstra wrote:
> > > I really think people who use hotplug at high frequencies are on drugs
> > > and doing it wrong.
> > 
> > I don't know.  It does make some sense.  It's not like we have any
> > other mechanism to keep some processors completely quiesient, which
> > could make a noticeable difference from powersaving POV compared to
> > mostly idle.  Rafael, can you please chime in and explain how / where
> > / how freqeuntly / etc CPU hotplug is used for powersaving?
> 
> Well, there are use cases I'm not really familiar with.
> 
> Pretty much the only use case I'm sufficiently familiar with is
> suspend/hibernate where we unplug all of the nonboot CPUs at one point.
> 
> The other use cases, which I don't really think are entirely valid,
> are on some ARM platforms where CPUs are unplugged instead of being put into
> C-states or equivalent (because we don't have a good mechanism for handling
> multiprocessor C-states; there's a set of patches for that waiting for
> the merge window in the Len's tree).  I'm hoping to get rid of those
> use cases in future entirely.

I don't see how we can get rid of hotplug (or some other state where
some processors are completely quiesient) in systems which have CPU clusters
with a different number of CPUs in each cluster. (eg 4 and 1 in case of
Tegra30).

Cheers,

Peter.