Re: [PATCH 1/4] KVM MMU: fix race in invlpg code

Re: [PATCH 1/4] KVM MMU: fix race in invlpg code

[Avi Kivity]

On 04/30/2010 12:00 PM, Xiao Guangrong wrote:
> It has race in invlpg code, like below sequences:
>
> A: hold mmu_lock and get 'sp'
> B: release mmu_lock and do other things
> C: hold mmu_lock and continue use 'sp'
>
> if other path freed 'sp' in stage B, then kernel will crash
>
> This patch checks 'sp' whether lived before use 'sp' in stage C
>    

> Signed-off-by: Xiao Guangrong<xiaoguangrong@cn.fujitsu.com>
> ---
>   arch/x86/kvm/paging_tmpl.h |   18 +++++++++++++++++-
>   1 files changed, 17 insertions(+), 1 deletions(-)
>
> diff --git a/arch/x86/kvm/paging_tmpl.h b/arch/x86/kvm/paging_tmpl.h
> index 624b38f..641d844 100644
> --- a/arch/x86/kvm/paging_tmpl.h
> +++ b/arch/x86/kvm/paging_tmpl.h
> @@ -462,11 +462,15 @@ out_unlock:
>
>   static void FNAME(invlpg)(struct kvm_vcpu *vcpu, gva_t gva)
>   {
> -	struct kvm_mmu_page *sp = NULL;
> +	struct kvm_mmu_page *sp = NULL, *s;
>   	struct kvm_shadow_walk_iterator iterator;
> +	struct hlist_head *bucket;
> +	struct hlist_node *node, *tmp;
>   	gfn_t gfn = -1;
>   	u64 *sptep = NULL, gentry;
>   	int invlpg_counter, level, offset = 0, need_flush = 0;
> +	unsigned index;
> +	bool live = false;
>
>   	spin_lock(&vcpu->kvm->mmu_lock);
>
> @@ -519,10 +523,22 @@ static void FNAME(invlpg)(struct kvm_vcpu *vcpu, gva_t gva)
>
>   	mmu_guess_page_from_pte_write(vcpu, gfn_to_gpa(gfn) + offset, gentry);
>   	spin_lock(&vcpu->kvm->mmu_lock);
> +	index = kvm_page_table_hashfn(gfn);
> +	bucket =&vcpu->kvm->arch.mmu_page_hash[index];
> +	hlist_for_each_entry_safe(s, node, tmp, bucket, hash_link)
> +		if (s == sp) {
>    

At this point, sp might have been freed and re-allocated, now pointing 
at something completely different.  So need to check role etc.

Alternatively, increase root_count.  Then sp is guaranteed to be live 
(though it may have role.invalid set).

> +			live = true;
> +			break;
> +		}
> +
> +	if (!live)
> +		goto unlock_exit;
> +
>   	if (atomic_read(&vcpu->kvm->arch.invlpg_counter) == invlpg_counter) {
>   		++vcpu->kvm->stat.mmu_pte_updated;
>   		FNAME(update_pte)(vcpu, sp, sptep,&gentry);
>   	}
> +unlock_exit:
>   	spin_unlock(&vcpu->kvm->mmu_lock);
>   	mmu_release_page_from_pte_write(vcpu);
>   }
>    


-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.

Re: [PATCH 3/4] KVM MMU: allow shadow page become unsync at creating time

[Avi Kivity]

On 04/30/2010 12:03 PM, Xiao Guangrong wrote:
> Allow new shadow page become unsync when is created, then we no need
> write-protect the 'sp->gfn', this idea is from Avi:
>
> |Another interesting case is to create new shadow pages in the unsync
> |state. That can help when the guest starts a short lived process: we
> |can avoid write protecting its pagetables completely
>
>    

Any idea how this improves performance?  A kernel build has a lot of 
short lived processes, might see an improvement there.

-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.

Re: [PATCH 4/4] KVM MMU: do not intercept invlpg if 'oos_shadow' is disabled

[Avi Kivity]

On 04/30/2010 12:05 PM, Xiao Guangrong wrote:
> If 'oos_shadow' == 0, intercepting invlpg command is really
> unnecessary.
>
> And it's good for us to compare the performance between enable 'oos_shadow'
> and disable 'oos_shadow'
>
> @@ -74,8 +74,9 @@ static int dbg = 0;
>   module_param(dbg, bool, 0644);
>   #endif
>
> -static int oos_shadow = 1;
> +int __read_mostly oos_shadow = 1;
>   module_param(oos_shadow, bool, 0644);
> +EXPORT_SYMBOL_GPL(oos_shadow);
>    

Please rename to kvm_oos_shadow to reduce potential for conflict with 
other global names.

But really, this is a debug option, I don't expect people to run with 
oos_shadow=0, so there's not much motivation to optimize it.

-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.

Re: [PATCH 4/4] KVM MMU: do not intercept invlpg if 'oos_shadow' is disabled

[Xiao Guangrong]

Avi Kivity wrote:
> On 04/30/2010 12:05 PM, Xiao Guangrong wrote:
>> If 'oos_shadow' == 0, intercepting invlpg command is really
>> unnecessary.
>>
>> And it's good for us to compare the performance between enable
>> 'oos_shadow'
>> and disable 'oos_shadow'
>>
>> @@ -74,8 +74,9 @@ static int dbg = 0;
>>   module_param(dbg, bool, 0644);
>>   #endif
>>
>> -static int oos_shadow = 1;
>> +int __read_mostly oos_shadow = 1;
>>   module_param(oos_shadow, bool, 0644);
>> +EXPORT_SYMBOL_GPL(oos_shadow);
>>    
> 
> Please rename to kvm_oos_shadow to reduce potential for conflict with
> other global names.
> 
> But really, this is a debug option, I don't expect people to run with
> oos_shadow=0, so there's not much motivation to optimize it.

Agreed, but, 'oos_shadow' option is document in Documentation/kernel-parameters.txt,
if it's just a debug option, i think we do better not document it.

Thanks,
Xiao

Re: [PATCH 4/4] KVM MMU: do not intercept invlpg if 'oos_shadow' is disabled

[Avi Kivity]

On 05/05/2010 03:54 PM, Xiao Guangrong wrote:
>
> Avi Kivity wrote:
>    
>> On 04/30/2010 12:05 PM, Xiao Guangrong wrote:
>>      
>>> If 'oos_shadow' == 0, intercepting invlpg command is really
>>> unnecessary.
>>>
>>> And it's good for us to compare the performance between enable
>>> 'oos_shadow'
>>> and disable 'oos_shadow'
>>>
>>> @@ -74,8 +74,9 @@ static int dbg = 0;
>>>    module_param(dbg, bool, 0644);
>>>    #endif
>>>
>>> -static int oos_shadow = 1;
>>> +int __read_mostly oos_shadow = 1;
>>>    module_param(oos_shadow, bool, 0644);
>>> +EXPORT_SYMBOL_GPL(oos_shadow);
>>>
>>>        
>> Please rename to kvm_oos_shadow to reduce potential for conflict with
>> other global names.
>>
>> But really, this is a debug option, I don't expect people to run with
>> oos_shadow=0, so there's not much motivation to optimize it.
>>      
> Agreed, but, 'oos_shadow' option is document in Documentation/kernel-parameters.txt,
> if it's just a debug option, i think we do better not document it.
>    

It has to be documented, otherwise people complain :)

Anyway the variable name and the option name don't have to be the same 
(I think).

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