[PATCH 1/2] tick: broadcast: Deny per-cpu clockevents from being broadcast sources

[Sören Brinkmann]

Hi Preeti,

On Fri, Sep 13, 2013 at 04:09:46PM +0530, Preeti U Murthy wrote:
> Hi Soren,
> 
> On 09/13/2013 03:50 PM, Preeti Murthy wrote:
> > Hi,
> > 
> > So the patch that Daniel points out http://lwn.net/Articles/566270/ ,
> > enables broadcast functionality
> > without using an external global clock device. It uses one of the per cpu
> > clock devices to enable the broadcast functionality.
> > 
> > The way it achieves this is by creating a pseudo clock device and
> > associating it with one of the cpus clock device and
> > by having a hrtimer queued on the same cpu. This pseudo clock device acts
> > as the broadcast device, and the
> >  per cpu clock device that it is associated with acts as the broadcast
> > source.
> > 
> > The disadvantages that Soren mentions in having a per cpu clock device as
> > the broadcast source can be overcome
> > by following the approach proposed in this patch n the way described below:
> > 
> > 1. What if the cpu, whose clock device is the broadcast source goes offline?
> > 
> > The solution that the above patch proposes is associate the pseudo clock
> > device with another cpu and move the hrtimer
> > whose function is explained in the next point to another cpu. The broadcast
> > functionality continues to remain active transparently.
> > 
> > 2. The cpu that requires broadcast functionality is different from the cpu
> > whose clock device is the broadcast source.
> > So how will the former cpu program/control the clock device of the latter
> > cpu?
> > 
> > The above patch queues a hrtimer on the cpu whose clock device is the
> > broadcast source, which expires at
> > max(tick_broadcast_period,  dev->next_event), where tick_broadcast_period
> > is what we define and dev is the
> > pseudo device whose next event is set by the broadcast framework.
> > 
> > On expiry of this hrtimer, do broadcast handling and reprogram the hrtimer
> > with same as above,
> > max(tick_broadcast_period,  dev->next_event).
> > 
> > This ensures that a cpu that requires broadcast function to be activated
> > need not program the broadcast source,
> > which also happens to be a per cpu clock device. The hrtimer queued on the
> > cpu whose clock device is the
> > broadcast source takes care of when to do broadcast handling.
> > tick_broadcast_period ensures that we do
> > not miss wakeups. This is introduced to overcome the constraint of a cpu
> > not being able to program the clock
> > device of another cpu.
> > 
> > Soren, do let me know if the above approach described in the patch has not
> > addressed any of the challenges
> > that you see with having a  per cpu clock device as the broadcast source.
> > 
> > Regards
> > Preeti U Murthy
> > 
> > 
> > On Fri, Sep 13, 2013 at 1:55 PM, Daniel Lezcano
> > <daniel.lezcano at linaro.org>wrote:
> > 
> >> On 09/12/2013 10:30 PM, Thomas Gleixner wrote:
> >>> On Thu, 12 Sep 2013, Soren Brinkmann wrote:
> >>>> From: Stephen Boyd <sboyd at codeaurora.org>
> >>>>
> >>>> On most ARM systems the per-cpu clockevents are truly per-cpu in
> >>>> the sense that they can't be controlled on any other CPU besides
> >>>> the CPU that they interrupt. If one of these clockevents were to
> >>>> become a broadcast source we will run into a lot of trouble
> >>>> because the broadcast source is enabled on the first CPU to go
> >>>> into deep idle (if that CPU suffers from FEAT_C3_STOP) and that
> >>>> could be a different CPU than what the clockevent is interrupting
> >>>> (or even worse the CPU that the clockevent interrupts could be
> >>>> offline).
> >>>>
> >>>> Theoretically it's possible to support per-cpu clockevents as the
> >>>> broadcast source but so far we haven't needed this and supporting
> >>>> it is rather complicated. Let's just deny the possibility for now
> >>>> until this becomes a reality (let's hope it never does!).
> >>>
> >>> Well, we can't do it this way. There are globally accessible clock
> >>> event devices which deliver only to cpu0. So the mask check might be
> >>> causing failure here.
> >>>
> >>> Just add a feature flag CLOCK_EVT_FEAT_PERCPU to the clock event
> >>> device and check for it.
> >>
> >> It sounds probably more understandable than dealing with the cpumasks.
> >>
> >> I am wondering if this is semantically opposed to
> >> http://lwn.net/Articles/566270/ ?
> >>
> >> [PATCH V3 0/6] cpuidle/ppc: Enable broadcast support for deep idle states
> >>
> >>   -- Daniel
> 
> So the point I am trying to make is that the fix that you have proposed
> on this thread is valid. It is difficult to ensure that a per cpu clock
> device doubles up as the broadcast source without significant code
> changes to the current broadcast code and the timer code.
> 
> But the patch [PATCH V3 0/6] cpuidle/ppc: Enable broadcast support for
> deep idle states, attempts to overcome the disadvantage on certain
> architectures of not having an external clock device to perform
> broadcast *without* significant code changes in broadcast or timer.
> 
> This patch does not conflict with what you are proposing in this thread
> of having a feature flag CLOCK_EVT_FEAT_PERCPU, since the pseudo clock
> device that the patch introduces will not have this flag set anyway.
> 
> So ideally architectures, without having a planned infrastructure in
> them cannot nominate their per cpu clock device as the broadcast source.
> And if they do have some infrastructure to support a per cpu clock
> device as broadcast source, they should ensure that the device passes
> your test as is proposed in this patch. So your fix is valid IMHO. That
> said it is still possible to manage without an external clock device for
> performing broadcast.

Thanks for the explanation but now I'm a little confused. That's a lot of
details and I'm lacking the in depth knowledge to fully understand
everything.

Is it correct to say, that your patch series enables per cpu devices to
be the broadcast device - for PPC?
And that would mean, that even though you have a per cpu device, you'd
deliberately not set the FEAT_PERCPU flag, because on PPC a per cpu
timer is a valid broadcast device?

Assuming that is not going into an utterly wrong direction: How would we
close on this one? AFAIK, ARM does not have this capability and I guess
it won't be added. So, should I go forward with the fix proposed by
Thomas? Should we rename the FEAT_PERCPU flag to something else, given
that PPC may use per cpu devices for broadcasting and the sole usage of
that flag is to prevent such a device from becoming the broadcast device?

	Thanks,
	Sören