Re: [Qemu-devel] [PATCH v2 2/5] i386: Add CPUID bit and feature words for IA32_ARCH_CAPABILITIES MSR

[Eduardo Habkost]

On Tue, Jul 03, 2018 at 03:35:13PM +0800, Robert Hoo wrote:
> On Thu, 2018-06-28 at 15:28 -0300, Eduardo Habkost wrote:
> > On Wed, Jun 27, 2018 at 07:27:21PM +0800, Robert Hoo wrote:
> > > Support of IA32_PRED_CMD MSR already be enumerated by same CPUID bit as
> > > SPEC_CTRL.
> > > 
> > > Signed-off-by: Robert Hoo <address@hidden>
> > 
> > Based on kernel commit 1eaafe91, it looks like we must always set
> > IA32_ARCH_CAPABILITIES.RSBA[bit 2] unless we're really sure the
> > VM will not be migrated to a vulnerable processor.
> > 
> > Considering this, I'd like to make "+arch-capabilities" set
> > IA32_ARCH_CAPABILITIES.RSBA by default, unless RSBA is explicitly
> > disabled by management software.
> > 
> Agree. But this seems beyond Icelake CPU model scope. How about I think
> about this carefully and compose another patch (set) for this?

This plan makes sense to me, as I don't want to make this
decision block IceLake from being in QEMU 3.0.

However, enabling CPUID_7_0_EDX_ARCH_CAPABILITIES in IceLake but
setting the MSR to 0 seems pointless.

I think we should add IceLake without
CPUID_7_0_EDX_ARCH_CAPABILITIES first, and later (after deciding
on a reasonable default value for MSR_IA32_ARCH_CAPABILITIES),
enable the CPUID bit on IceLake (hopefully in time for QEMU 3.0).


> And you'd like to set  IA32_ARCH_CAPABILITIES.RSBA by default in qemu or
> kvm layer?

Probably we need to make this decision in QEMU.  If KVM set RSBA
automatically on .get_msr_feature(), QEMU won't be able to
differentiate a host with RSBA set from a host with RSBA unset.

-- 
Eduardo