Re: [PATCH v2] KVM: x86: Do not return host topology information from KVM_GET_SUPPORTED_CPUID

[Paolo Bonzini]

On 1/26/23 01:58, Jim Mattson wrote:
> > You wrote it yourself: any VMM must either populate the topology on its
> > own, or possibly fill it with zeros.  Returning a value that is
> > extremely unlikely to be used is worse in pretty much every way (apart
> > from not breaking your VMM, of course).
>
> I've complained about this particular ioctl more than I can remember.
> This is just one of its many problems.

I agree.  At the very least it should have been a VM ioctl.

> > With a total of six known users (QEMU, crosvm, kvmtool, firecracker,
> > rust-vmm, and the Google VMM), KVM is damned if it reverts the patch and
> > damned if it doesn't.  There is a tension between fixing the one VMM
> > that was using KVM_GET_SUPPORTED_CPUID correctly and now breaks loudly,
> > and fixing 3-4 that were silently broken and are now fixed.  I will
> > probably send a patch to crosvm, though.
> >
> > The VMM being _proprietary_ doesn't really matter, however it does
> > matter to me that it is not _public_: it is only used within Google, and
> > the breakage is neither hard to fix in the VMM nor hard to temporarily
> > avoid by reverting the patch in the Google kernel.
>
> Sadly, there isn't a single kernel involved. People running our VMM on
> their desktops are going to be impacted as soon as this patch hits
> that distro. (I don't know if I can say which distro that is.) So, now
> we have to get the VMM folks to urgently accommodate this change and
> get a new distribution out.

Ok, this is what is needed to make a more informed choice.  To be clear, 
this is _still_ not public (for example it's not ChromeOS), so there is 
at least some control on what version of the VMM they use?  Would it 
make sense to buy you a few months by deferring this patch to Linux 6.3-6.5?

Thanks,

Paolo