Re: [PATCH v2 02/12] KVM: x86: tag the instructions which are used to write page table

[Xiao Guangrong <xiaoguangrong@cn.fujitsu.com>]

Hi Avi,

On 08/03/2011 04:09 PM, Avi Kivity wrote:
> On 08/03/2011 09:02 AM, Xiao Guangrong wrote:
>> >  Note we don't need to actually emulate, just decode, since page_fault can tell us whether a write failed due to page tables or mmio.
>> >
>>
>> This is a interesting feature. If it happens, i will just drop the shadow pages
>> and retry these instructions directly.
> 
> Note it's a little dangerous.  If the guest uses a non-page-table modifying instruction on the PDE that points to the instruction, then we will unmap the instruction and go to an infinite loop.
> 

Yes, it is.

> Maybe it's better to emulate if we can't find a fix for that.
> 
> One way would be to emulate every 20 instructions; this breaks us out of the loop but reduces costly emulations to 5%.
> 

After much thought about this, may be this optimization is not good since:
- it is little complex
- this optimization is only applied to the instruction emulation caused by #PF
- it does not improve too much:
  if we emulate the instruction, we need to do:
  - decode instruction
  - emulate it
  - zap shadow pages
  And do this, it can return to the guest, the guest can run the next instruction

  if we retry the instruction, we need to do:
  - decode instruction
  - zap shadow pages
  then return to the guest and retry the instruction, however, we will get page fault
  again(since the mapping is still read-only), so we will get another VM-exit and need
  to do:
  # trigger page fault
  - handle the page fault and change the mapping to writable
  - retry the instruction
  until now, the guest can run the next instruction

So, i do not think the new way is better, your opinion？