On 19/02/2018 14:35, David Woodhouse wrote:

On Mon, 2018-02-19 at 14:10 +0100, Paolo Bonzini wrote:

...

...

Hardware seems like a reasonable place to get the default value (cf. the VMX capability MSRs).

There are some differences:

• a zero value for ARCH_CAPABILITIES should be safe, while a zero value

for VMX capabilities doesn't really make sense.  On the contrary, a nonzero value for ARCH_CAPABILITIES is not safe across live migration.

Any VMM which is going to support live migration surely needs to pay at least a small amount of attention to the features it exposes? Exposing the ARCH_CAPABILITIES CPUID bit without actually looking at the contents of the associated MSR which that bit advertises would be... a little strange, would it not?

I think what we should do is simply backport Tom Lendacky's series to 4.14 and 4.9 ASAP, and add ARCH_CAPABILITIES support there. Then the question of the default becomes moot, more or less.

Paolo

I don't see why we care so much about the *default* value, in that context.