Re: [PATCH 0/4] kvm, mem-hotplug: Do not pin ept identity pagetable and apic access page.

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

Hi Gleb,

On 07/03/2014 02:04 PM, Gleb Natapov wrote:
> On Thu, Jul 03, 2014 at 09:17:59AM +0800, Tang Chen wrote:
>> Hi Gleb,
>>
>> On 07/02/2014 05:00 PM, Tang Chen wrote:
>>> Hi Gleb, Marcelo,
>>>
>>> Please help to review this patch-set.
>>>
>>> NOTE: This patch-set doesn't work properly.
>>>
>>>
>>> ept identity pagetable and apic access page in kvm are pinned in memory.
>>> As a result, they cannot be migrated/hot-removed.
>>>
>>> But actually they don't need to be pinned in memory.
>>>
>>> This patch-set introduces two new vcpu requests: KVM_REQ_MIGRATE_EPT and KVM_REQ_MIGRATE_APIC.
>>> These two requests are made when the two pages are migrated by the mmu_notifier
>>> to reset the related variable to unusable value. And will also be made when
>>> ept violation happens to reset new pages.
>>>
>>>
>>> [Known problem]
>>> After this patch-set applied, the two pages can be migrated/hot-removed.
>>> But after migrating apic access page, the guest died.
>>>
>>> The host physical address of apic access page is stored in VMCS. I reset
>>> it to 0 to stop guest from accessing it when it is unmapped by
>>> kvm_mmu_notifier_invalidate_page(). And reset it to new page's host physical
>>> address in tdp_page_fault(). But it seems that guest will access apic page
>>> directly by the host physical address.
>>
>> Would you please to give some advice about this problem ?
>>
> I haven't reviewed third patch yet, will do ASAP.
>

I printed some info in the kernel, and I found that mmu_notifier 
unmapped the
apic page and set VMCS APIC_ACCESS_ADDR to 0. But apic page ept 
violation didn't
happen. And the guest stopped running.

I think when guest tried to access apic page, there was no ept violation 
happened.
And as a result, VMCS APIC_ACCESS_ADDR was not correctly set.

Referring to Intel Software Developer's Manuel Vol 3B, when accessing 
apic page
using translation with a large page (2M, 4M, 1G), APIC VM_exit will not 
happen.

How do you think about this ?

Thanks. :)