Re: [PATCH 0/3] Hook up powerclamp with PM QOS and cpuidle

["Rafael J. Wysocki" <rjw@rjwysocki.net>]

On Wednesday, November 27, 2013 08:47:32 AM jacob pan wrote:
> On Wed, 27 Nov 2013 12:56:34 +0100
> Peter Zijlstra <peterz@infradead.org> wrote:
> 
> > On Tue, Nov 26, 2013 at 03:20:08PM -0800, Jacob Pan wrote:
> > > This patchset is intended to address the behavior change and
> > > efficiency loss introduced by using consolidated idle routine in
> > > powerclamp driver.
> > > 
> > > Specifically,
> > > [PATCH 3/8] idle, thermal, acpi: Remove home grown idle
> > > implementations
> > > 
> > > The motivation is that after using common idle routine, powerclamp
> > > driver can no longer pick the deepest idle state needed to conserve
> > > power. Idle state is selected by governors which can be influenced
> > > by PM QOS and other factors. This patchset hooks up powerclamp idle
> > > injection with PM QOS and eventually influce idle governors to pick
> > > the power saving target states.
> > > 
> > > There are some downside of this approach. Due to overhead,
> > > communication with PM QOS is at enable/disable idle injection time
> > > instead of each injection period. The implication is that if the
> > > system natual idle is more than target injected idle, powerclamp
> > > will skip some injection period. During this period however,
> > > deepest idle state may still be chosen necessarily regardless the
> > > latency constraint.
> > 
> > Does the QoS stuff have a means of notifying its users of constraints
> > violation? I suspect some applications might light to be told if their
> > requests aren't honoured.
> > 
> Each class has a notifier. This patchset is calling the notifier
> when the qos class is disable/enable. the receiver of these
> notifications are in the kernel.
> 
> I don't see the qos core code has a way to signal userspace about
> target change.

But user space can add constraints and expect them to be actually satisfied.
If those constraints cannot be satisfied, there should be a mechanism to
notify the processes who set them about that.  That mechanism is not present
currently.

Thanks,
Rafael