Re: [PATCH v2 1/1] target-i386: Fix default Hypervisor level for hypervisor-vendor=kvm.

[Eduardo Habkost <ehabkost@redhat.com>]

On Fri, Sep 21, 2012 at 05:28:27PM -0400, Don Slutz wrote:
> 
> On 09/21/12 16:49, Eduardo Habkost wrote:
> >On Fri, Sep 21, 2012 at 04:26:58PM -0400, Don Slutz wrote:
> >>On 09/21/12 10:18, Eduardo Habkost wrote:
> >>>On Thu, Sep 20, 2012 at 04:06:27PM -0400, Don Slutz wrote:
> >>>> From http://lkml.indiana.edu/hypermail/linux/kernel/1205.0/00100.html
> >>>>EAX should be KVM_CPUID_FEATURES (0x40000001) not 0.
> >>>>
> >>>>Added hypervisor-vendor=kvm0 to get the older CPUID result. kvm1 selects the newer one.
> >>>Why not just make "hypervisor-vendor=kvm" control only the hypervisor
> >>>vendor string, and support something like "kvm-hypervisor-level=0" to
> >>>restore the old cpuid_hv_level=0 behavior?
> >>   -cpu host,hypervisor-vendor=kvm,hypervisor-level=0
> >>
> >>  Does this.
> >Good.  :-)
> >
> >>>This is similar to the kvmclock case: it would allow us to make
> >>>"hypervisor-vendor=kvm" use saner values as default, but letting old
> >>>machine-types to override it for compatibility if required.
> >>Right now since I am using env->cpuid_hv_level == 0 as a flag. This
> >>means that:
> >>
> >>   -cpu host,hypervisor-level=0,hypervisor-vendor=kvm
> >>
> >>   -cpu host,hypervisor-vendor=kvm,hypervisor-level=0
> >>
> >>end up with different CPUID data (Which I do not like). I will fix this in the next round.
> >Right. This has to be fixed.
> >
> >>Did you want me to drop kvm0 and kvm1?
> >Yes, if level is already configurable using the hypervisor-level
> >property, I don't see the need for kvm0 and kvm1.
> >
> >If you make kvm_arch_init_vcpu() actually use those fields, you will end
> >up implementing what's required to allow migration compatibility to be
> >kept (the only thing missing is to make the CPU class a child of
> >DeviceState, and add hypervisor-level=0 to the existing machine-types).
> >:-)
> >
> You mean like
>   http://lists.gnu.org/archive/html/qemu-devel/2012-09/msg03402.html
> and
>   http://lists.gnu.org/archive/html/qemu-devel/2012-09/msg03405.html
> already change kvm_arch_init_vcpu().

Yes! Sorry, I hadn't reviewed all of your previous series yet.  :-)

> I did not know that I need to
> make the CPU class a child of DeviceState.

You don't. But we plan to do that, eventually, so we can put
CPU compatibility properties into machine-types.

> Nor that I needed to add hypervisor-vendor=kvm,hypervisor-level=0 to
> the existing machine-types.  Since without specifying
> hypervisor-level=0 it defaults to 0 and kvm_arch_init_vcpu() will
> default to setting hypervisor-vendor=kvm.

What I would like to do is: to change the default to 0x40000001 (that's
the correct value), but make the pc-1.1 and older machine-types keep the
hypervisor-level=0 default for compatibility.

(But to be able to do that, we need to first make the CPU class a child
of DeviceState, so that's something to be done later)

-- 
Eduardo