[Qemu-devel] Re: POLL: Why do you use kqemu?

[Jan Kiszka <jan.kiszka@siemens.com>]

Avi Kivity wrote:
> Jan Kiszka wrote:
>> Avi Kivity wrote:
>>  
>>> Jan Kiszka wrote:
>>>    
>>>> And the fact that kqemu has to use tcg in order to achieve a reasonable
>>>> performance is rather a disadvantage. The complexity and overhead for
>>>> synchronizing tcg with the in-kernel accelerator is enormous. If there
>>>> were a feasible way to overcome this with kqemu, it would benefit a
>>>> lot.
>>>> But unfortunately there is none (given you don't want to invest
>>>> reasonable efforts).
>>>>         
>>> Note that kvm suffers from something similar (to a smaller magnitude) as
>>> well: if a guest pages in its page tables, kvm knows nothing about it
>>> and will thus have outdated shadows.  To date we haven't encountered a
>>> problem with it, but it's conceivable.  I think Windows can page its
>>> page tables, but maybe it's disabled by default, or maybe it doesn't dma
>>> directly into the page tables.
>>>     
>>
>> Can't follow, always thought that kernel space gets informed when some
>> I/O operation handled by user space modified an "interesting" page.
>>   
> 
> It doesn't.  Host userspace has unrestricted access to guest memory.
> 
>>> Not sure how to fix.  Maybe write protect the host page tables when we
>>>     
>>
>> You mean guest page table?
>>   
> 
> Both :)
> 
> When kvm write protects a guest page table in the shadow page table
> entries pointing to that guest page, it should also write protect the
> guest page table in the host page table entries to the same guest page.

Ah, now I got it. What do other hypervisors do?

> 
>>> shadow a page table, and get an mmu notifier to tell us when its made
>>> writable?  Seems expensive.  Burying head in sand is much easier.
>>>
>>>     
>>
>> Does this still apply to nested paging? I guess (hope) not...
>>   
> 
> No, nested paging brings cancer and cures world peace.  Or something.
> 

Well, then it's probably not worth bothering, at least until a real
guest problem is explainable with this limitation. Are there any
suspicious reports floating around (maybe not only about Windows)?

Jan

-- 
Siemens AG, Corporate Technology, CT SE 2
Corporate Competence Center Embedded Linux