Re: [patch] kvmclock fix

[Gerd Hoffmann <kraxel@redhat.com>]

Glauber Costa wrote:
>> +static unsigned int  ref_freq;
>> +static unsigned long loops_per_jiffy_ref;
>> +static unsigned long tsc_khz_ref;
> 
> Doesn't these need to be percpu too ?
> 
> Otherwise we could be calculating the new frequency based on a foreign
> reference.

I doubt you can have a SMP systems with cpus having different freqs.
Also tsc.c does the same.

> I believe a possible approach is to move those that refer to tsc
> calculation to tsc.c, and make all uses
> of tsc_khz comply to a per_cpu use. A global tsc_khz does not make sense anyway.
> 
> Then we'll have a specific kvm notifier that just forces the guest out
> of its execution so we can update
> the clock state.

Then we have an ordering issue with the two notifier calls.  I don't
thing there is a way to force a specific call order.

cheers,
  Gerd