Re: [PATCH] cpuidle: Allow configuration of the polling interval before cpuidle enters a c-state

[Mel Gorman <mgorman@techsingularity.net>]

On Fri, Nov 27, 2020 at 11:08:11AM -0300, Marcelo Tosatti wrote:
> > Well, IMV this is not straightforward at all.
> > 
> > It requires the admin to know how cpuidle works and why this
> > particular polling limit is likely to be suitable for the given
> > workload.  And whether or not the default polling limit should be
> > changed at all.
> 
> KVM polling (virt/kvm/kvm_main.c grow_halt_poll_ns/shrink_halt_poll_ns)
> tries to adjust the polling window based on poll success/failure. 
> 
> The cpuidle haltpoll governor (for KVM guests) uses the same adjustment
> logic.
> 
> Perhaps a similar (or improved) scheme can be adapted to baremetal.
> 
> https://www.kernel.org/doc/Documentation/virtual/kvm/halt-polling.txt

I'm aware of the haltpoll driver and why it's needed there. Given that
it adds cost to the refect callback and increases exit latency that it
would be unpopular on bare metal.

I'm prototyping a v2 that simply picks different balance point to see if
that gets a better reception.

-- 
Mel Gorman
SUSE Labs