Re: [PATCH 2/3] KVM: x86: let reexecute_instruction work for tdp

[Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>]

On 11/28/2012 10:01 PM, Gleb Natapov wrote:
> On Wed, Nov 28, 2012 at 11:15:13AM +0800, Xiao Guangrong wrote:
>> On 11/28/2012 07:32 AM, Marcelo Tosatti wrote:
>>> On Tue, Nov 27, 2012 at 11:13:11AM +0800, Xiao Guangrong wrote:
>>>>>> +static bool reexecute_instruction(struct kvm_vcpu *vcpu, unsigned long cr2)
>>>>>>  {
>>>>>> -	gpa_t gpa;
>>>>>> +	gpa_t gpa = cr2;
>>>>>>  	pfn_t pfn;
>>>>>>
>>>>>> -	if (tdp_enabled)
>>>>>> +	if (!ACCESS_ONCE(vcpu->kvm->arch.indirect_shadow_pages))
>>>>>>  		return false;
>>>>>
>>>>> How is indirect_shadow_pages protected? Why is ACCESS_ONCE() being used
>>>>> to read it?
>>>>
>>>> Hi Marcelo,
>>>>
>>>> It is protected by mmu-lock for it only be changed when mmu-lock is hold. And
>>>> ACCESS_ONCE is used on read path avoiding magic optimization from compiler.
>>>
>>> Please switch to mmu_lock protection, there is no reason to have access
>>> to this variable locklessly - not performance critical.
>>>
>>> For example, there is no use of barriers when modifying the variable.
>>
>> This is not bad, the worst case is, the direct mmu failed to unprotect the shadow
>> pages, (meet indirect_shadow_pages = 0, but there has shadow pages being shadowed.),
>> after enter to guest, we will go into reexecute_instruction again, then it will
>> remove shadow pages.
>>
> Isn't the same scenario can happen even with mmu lock around
> indirect_shadow_pages access?

Hmm..., i also think it is no different. Even using mmu-lock, we can not
prevent the target pfn can not be write-protected later. Marcelo?