Re: [PATCH for-4.16] x86/cpuid: do not shrink number of leaves in max policies

[Ian Jackson <iwj@xxxxxxxxxxxxxx>]

Roger Pau Monne writes ("[PATCH for-4.16] x86/cpuid: do not shrink number of 
leaves in max policies"):
> Shrinking max policies can lead to failures in migration as previous
> versions of Xen didn't shrink the number of leaves in any case, so
> it's possible for a guest created on previous versions of Xen that
> pass CPUID data on the migration stream to contain a max leaf number
> greatest than the one present on the max policies in versions of Xen
> containing 540d911c28.
> 
> Such failure was seen by osstest when doing a migration from Xen
> 4.15 to Xen 4.16-rc on a pair of equal boxes, the noceras.
> 
> Fix this by preventing any shrinking of the max CPUID policies, so
> that previously built guest CPUID policies are compatible.
> 
> Fixes: 540d911c28 ('x86/CPUID: shrink max_{,sub}leaf fields according to 
> actual leaf contents')
> Signed-off-by: Roger Pau Monné <roger.pau@xxxxxxxxxx>
...
> This is a regression from Xen 4.15, so should be considered for Xen
> 4.16. The main risks would be to mess up with the CPUID policy in a
> different way, that would also lead to brokenness. Strictly speaking
> the change here removes the shrinking of max leaves and restores the
> previous behavior, but it's obviously not completely risk free.
> 
> It has proven to fix the regression seen on the noceras.

Ouch.

Questions from my RM hat:

Is there a workaround ?
What proportion of machines do we think this might affect ?

Jan, Andy, do you have an opinion ?

Ian.