Re: [PATCH v2] NMI: fix NMI period is not correct when cpu frequency changes issue.

[Don Zickus <dzickus@redhat.com>]

On Thu, Apr 18, 2013 at 02:04:00PM +0200, Stephane Eranian wrote:
> On Thu, Apr 18, 2013 at 1:42 PM, Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:
> > On Tue, 2013-04-16 at 06:57 +0000, Pan, Zhenjie wrote:
> >> Watchdog use performance monitor of cpu clock cycle to generate NMI to detect hard lockup.
> >> But when cpu's frequency changes, the event period will also change.
> >> It's not as expected as the configration.
> >> For example, set the NMI event handler period is 10 seconds when the cpu is 2.0GHz.
> >> If the cpu changes to 800MHz, the period will be 10*(2000/800)=25 seconds.
> >> So it may make hard lockup detect not work if the watchdog timeout is not long enough.
> >> Now, set a notifier to listen to the cpu frequency change.
> >> And dynamic re-config the NMI event to make the event period correct.
> >>
> >
> >
> > Urgh,. does this really matter.. all we really want is for that NMI to
> > hit eventually in the not too distant future. Does the frequency really
> > matter _that_ much?
> >
> I agree, it does not really matter. Set the watchdog to a couple of minutes
> and it should be fine, shouldn't it?

I believe it mattered to the Chrome folks. They want the watchdog to be as
tight as possible so the user experience isn't a hang but a quick reboot
instead.  They like setting the watchdog to something like 2 seconds.

There was a patch a few months ago that tried to hack around this issue
and I suggested this approach as a better solution.  I forgot what the
original problem was.  Perhaps someone can jump in and explain the problem
being solved (other than the watchdog isn't always 10 seconds)?

Cheers,
Don

> 
> > Also, can't we simply pick an event that's invariant to the cpufreq
> > nonsense? Something like CPU_CLK_UNHALTED.REF -- or better the
> > fixed_ctr2 which nobody ever uses anyway.
> >
> You don't want to use fixed counter 2 for NMI watchdog because it's pinned.
> No other counter can count this event. And it is very useful. I use it often.