Re: [Qemu-devel] Re: [PATCH 3/4] mask out forbidden cpuid features with kvm ioctl

[Anthony Liguori <anthony@codemonkey.ws>]

Avi Kivity wrote:
> Anthony Liguori wrote:
>> Glauber Costa wrote:
>>> KVM has a (so far unused) ioctl to inform userspace
>>> about supported cpuid bits in the current host.
>>>
>>> The lack of this kind of checking can lead to bugs in which
>>> a cpuid bit is exposed but not supported by the host kernel
>>> (an example is fedora bug 
>>> https://bugzilla.redhat.com/show_bug.cgi?id=481274)
>>>
>>> The current host_cpuid code is replaced by this general check.
>>>
>>> Signed-off-by: Glauber Costa <glommer@redhat.com>
>>>   
>>
>> So what do we do about live migration?  
>
> We need a way to completely specify cpuid from the command line.  The 
> management utility would figure out what is supported (driven by both 
> cpu capabilities and installed kvm capabilities) and fgure out a 
> greatest common denominator, or something.

What is the default case?  Pass everything through or pick a reasonable 
small, constant subset?  That is, is -cpuid a union or an intersection.

Regards,

Anthony Liguori

>> Does KVM mask a set of bits unconditionally?  
>
> The kernel doesn't touch cpuid (except nx due to an historical mistake).
>
>> Before, we were masking unconditionally which should insulate 
>> migration reasonable well.
>
> That suits migration, but not other uses of qemu.
>