Re: [PATCH 4/4] kvm, mem-hotplug: Update apic access page when it is migrated.

[Tang Chen <tangchen@cn.fujitsu.com>]

Hi Gleb,

Thanks for all the advices. Please see below.

On 07/04/2014 06:13 PM, Gleb Natapov wrote:
......
>>>> +static void vmx_set_apic_access_page_addr(struct kvm *kvm, hpa_t hpa)
>>>> +{
>>>> +	if (vm_need_virtualize_apic_accesses(kvm))
>>> This shouldn't even been called if apic access page is not supported. Nor
>>> mmu_notifier path neither tdp_page_fault path should ever see 0xfee00000
>>> address. BUG() is more appropriate here.
>>>
>>
>> I don't quite understand. Why calling this function here will leed to bug ?
>> (Sorry, I'm not quite understand the internal of KVM. Please help.)
> I didn't say that calling this function here will lead to a bug. I am saying that
> if vm_need_virtualize_apic_accesses() is false this function should not be called
> at all, so this check is redundant.
>

Do you mean when vm_need_virtualize_apic_accesses() is false, it should 
not be called ?
It has to be true ?

......
>>>> +	if (kvm->arch.apic_access_page_migrated) {
>>>> +		if (kvm->arch.apic_access_page)
>>>> +			kvm->arch.apic_access_page = pfn_to_page(0);
>>> All vcpus will access apic_access_page without locking here. May be
>>> set kvm->arch.apic_access_page to zero in mmu_notifier and here call
>>>   kvm_x86_ops->set_apic_access_page_addr(kvm, kvm->arch.apic_access_page);
>>>
>>
>> I'm a little confused. apic access page's phys_addr is stored in vmcs, and
>> I think it will be used by vcpu directly to access the physical page.
>> Setting kvm->arch.apic_access_page to zero will not stop it, right ?
>>
> Right, kvm->arch.apic_access_page is just a shadow value for whatever is written
> in vmcs. After setting it all vcpus need to update their vmcs values.
>
>> I'm wondering what happens when apic page is migrated, but the vmcs is still
>> holding its old phys_addr before the vcpu request is handled.
>>
> apic page should not be migrated untill all vpus are forced out of a guest mode and
> instructed to reload new value on a next guest entry. That's what we are trying to
> achieve here.
>

So, setting VMCS APIC_ACCESS_ADDR pointer to zero will not stop vcpu to 
access
apic access page, right ?

If so, all the vcpus have to stop till apic page finishes its migration, 
and new
value is set in each vcpu, which means we should stop guest, right ?

Thanks.