Re: [External] Re: [RFC: timer passthrough 1/9] KVM: vmx: hook set_next_event for getting the host tscd

[Zhimin Feng]

Hi tglx

This question is very nice,  we should be considered to judge whether 
the current active device is the tsc deadline timer. I will fix this in V2.

Thanks

Zhimin

在 2021/2/6 上午2:11, Thomas Gleixner 写道:
> On Fri, Feb 05 2021 at 18:03, Zhimin Feng wrote:
> > @@ -520,6 +521,24 @@ struct kvm_vcpu_hv {
> >   	cpumask_t tlb_flush;
> >   };
> >   
> > +enum tick_device_mode {
> > +	TICKDEV_MODE_PERIODIC,
> > +	TICKDEV_MODE_ONESHOT,
> > +};
> > +
> > +struct tick_device {
> > +	struct clock_event_device *evtdev;
> > +	enum tick_device_mode mode;
> > +};
> There is a reason why these things are defined in a header file which is
> not public. Nothing outside of kernel/time/ has to fiddle with
> this. Aside of that how are these things supposed to stay in sync?
>
> > diff --git a/kernel/time/tick-common.c b/kernel/time/tick-common.c
> > index 6c9c342dd0e5..bc50f4a1a7c0 100644
> > --- a/kernel/time/tick-common.c
> > +++ b/kernel/time/tick-common.c
> > @@ -26,6 +26,7 @@
> >    * Tick devices
> >    */
> >   DEFINE_PER_CPU(struct tick_device, tick_cpu_device);
> > +EXPORT_SYMBOL_GPL(tick_cpu_device);
> Not going to happen ever.
>
> > +#define TSC_DIVISOR  8
> > +static DEFINE_PER_CPU(struct timer_passth_info, passth_info);
> > +
> > +static int override_lapic_next_event(unsigned long delta,
> > +		struct clock_event_device *evt)
> > +{
> > +	struct timer_passth_info *local_timer_info;
> > +	u64 tsc;
> > +	u64 tscd;
> > +
> > +	local_timer_info = &per_cpu(passth_info, smp_processor_id());
> > +	tsc = rdtsc();
> > +	tscd = tsc + (((u64) delta) * TSC_DIVISOR);
> > +	local_timer_info->host_tscd = tscd;
> > +	wrmsrl(MSR_IA32_TSCDEADLINE, tscd);
> > +	return 0;
> > +}
> > +
> > +static void vmx_host_timer_passth_init(void *junk)
> > +{
> > +	struct timer_passth_info *local_timer_info;
> > +	int cpu = smp_processor_id();
> > +
> > +	local_timer_info = &per_cpu(passth_info, cpu);
> > +	local_timer_info->curr_dev = per_cpu(tick_cpu_device, cpu).evtdev;
> > +	local_timer_info->orig_set_next_event =
> > +		local_timer_info->curr_dev->set_next_event;
> > +	local_timer_info->curr_dev->set_next_event = override_lapic_next_event;
> So when loading the KVM module you steal the set_next_event pointer of
> the clock event device which is currently active. What guarantees that
>
>      1) The current active device is the tsc deadline timer
>      2) The active device does not change
>
> Nothing.
>
> Thanks,
>
>          tglx