Re: [PATCH 2/4] KVM: Extended shared_msr_global to per CPU

[Avi Kivity <avi@redhat.com>]

On 12/17/2009 05:01 PM, Sheng Yang wrote:
>
>>> -void kvm_define_shared_msr(unsigned slot, u32 msr)
>>> +static void shared_msr_update(unsigned slot, u32 msr)
>>>    {
>>> -	int cpu;
>>> +	struct kvm_shared_msrs *smsr;
>>>    	u64 value;
>>>
>>> +	smsr =&__get_cpu_var(shared_msrs);
>>> +	/* only read, and nobody should modify it at this time,
>>> +	 * so don't need lock */
>>> +	if (slot>= shared_msrs_global.nr) {
>>> +		printk(KERN_ERR "kvm: invalid MSR slot!");
>>> +		return;
>>> +	}
>>> +	if (smsr->values[slot].initialized)
>>> +		return;
>>>        
>> I don't think .initialized is worthwhile.  shared_msr_update is run very
>> rarely.
>>      
> The reason is, after cpu hotplug, the MSR_TSC_AUX would be rewritten by
> vsyscall_init again. But in the hotplug notifier chain, KVM has higher
> priority(20 vs 0 for vsyscall_init), so maybe the rdmsr() here would get a
> bogus value... Then I think prevent it from initializing again should be
> safer.
>
>    

So let's raise vsyscall_init's priority?

> But I just think of another issue: if we hot plug in a cpu(without hot plug
> off), it would have a bogus value as well in the same path? Sound
> troublesome...
>    

Removing .initialized would take care of this issue.

-- 
error compiling committee.c: too many arguments to function