[RFC] KVM MMU: thinking of shadow page cache

[RFC] KVM MMU: thinking of shadow page cache

[Xiao Guangrong]

When we cached shadow page tables, one guest page table may have
many shadow pages, take below case for example:

 (RO+U) ---> |------| __     |------|
 (W+U ) ---> |  GP1 |   |    |  GP2 |
 (W+P ) ---> |------|   |--> |------|

There have 3 kinds of permission mapping to GP1, so we should
allocate 3 shadow pages for GP1 and 3 shadow pages for GP2.
And it has 3 class permissions(R/W, U/S, X/NX) in x86's architecture,
for the worst case, we should allocate 2^3 pages for every paging
mapping level.

This waste is caused by that we only set the permission bits in PTE,
not in the middle mapping level.

So, i think we can mapping guest page table's permission into cache
shadow page table, then it can be shared between many shadow page tables
if their map to the same gust physics address. For above case, we only need 2
pages.

Any comments?

Re: [RFC] KVM MMU: thinking of shadow page cache

[Avi Kivity]

On 03/30/2010 04:59 AM, Xiao Guangrong wrote:
> When we cached shadow page tables, one guest page table may have
> many shadow pages, take below case for example:
>
>   (RO+U) --->  |------| __     |------|
>   (W+U ) --->  |  GP1 |   |    |  GP2 |
>   (W+P ) --->  |------|   |-->  |------|
>
> There have 3 kinds of permission mapping to GP1, so we should
> allocate 3 shadow pages for GP1 and 3 shadow pages for GP2.
> And it has 3 class permissions(R/W, U/S, X/NX) in x86's architecture,
> for the worst case, we should allocate 2^3 pages for every paging
> mapping level.
>
> This waste is caused by that we only set the permission bits in PTE,
> not in the middle mapping level.
>
> So, i think we can mapping guest page table's permission into cache
> shadow page table, then it can be shared between many shadow page tables
> if their map to the same gust physics address. For above case, we only need 2
> pages.
>
> Any comments?
>    

We've considered this in the past, it makes sense.  The big question is 
whether any guests actually map the same page table through PDEs with 
different permissions (mapping the same page table through multiple PDEs 
is very common, but always with the same permissions).  Do you know of 
any such guest?

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

Re: [RFC] KVM MMU: thinking of shadow page cache

[Xiao Guangrong]

Avi Kivity wrote:

>>    
> 
> We've considered this in the past, it makes sense.  The big question is
> whether any guests actually map the same page table through PDEs with
> different permissions (mapping the same page table through multiple PDEs
> is very common, but always with the same permissions).  Do you know of
> any such guest?

I also don't know whether have such guest.
Maybe my idea is no good for current OS, thanks for your comments.

Xiao

Re: [RFC] KVM MMU: thinking of shadow page cache

[Avi Kivity]

On 04/01/2010 12:05 PM, Xiao Guangrong wrote:
>
>> We've considered this in the past, it makes sense.  The big question is
>> whether any guests actually map the same page table through PDEs with
>> different permissions (mapping the same page table through multiple PDEs
>> is very common, but always with the same permissions).  Do you know of
>> any such guest?
>>      
> I also don't know whether have such guest.
> Maybe my idea is no good for current OS, thanks for your comments.
>    

In fact there are plans to make kvm such a guest (when running in nested 
mode) - when we perform live migration we write-protect all guest pages, 
and it's reasonable to use the top-level shadow page to write protect 
all memory at once instead of iterating over all mmu pages.  When that 
goes in, we should also implement your idea.

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