Re: [Qemu-devel] Re: [PATCH 0/8] kvm: Fixes, cleanups and live migration

[Gleb Natapov]

On Sun, May 03, 2009 at 11:01:34AM +0300, Avi Kivity wrote:
> Gleb Natapov wrote:
>>> I don't want the migration protocol to encode vendor specific   
>>> information. The architectural state is complicated enough, we don't  
>>> want microarchitectural state as well.
>>>     
>> Then I don't see how migration can work correctly. How do you expect
>> migration to work if you don't migrate part of a processor state? Why
>> not drop non migratable state immediately after exit then? (that is
>> essentially what happens if we don't migrate it).
>>   
>
> If we can roll back the state to before the software interrupt executed,  
> we are never in the situation where the instruction length is needed.
>
Not according to Intel :)

> The whole mess is needed because vmx allows exiting after a software  
> interrupt instruction has been executed, but before the software  
> interrupt was processed by the cpu. If we unexecute the instruction and  
> forget the software interrupt, everything will continue to work.
>
VMX exits with RIP pointing to software interrupt instruction (i.e before
instruction execution), so no need to "unexecute" it. Intel advice to
inject software interrupt as opposite to reexecute instruction. If we
will not migrate information needed to inject soft interrupt we will
have to reexecute it after migration. May be this is not a big deal.

--
                        Gleb.