Re: [Qemu-devel] [PATCH 5/5] target-i386: fix kvm_default_features overwriting kvm_arch_get_supported_cpuid()

[Igor Mammedov]

On Tue, 5 Feb 2013 19:20:04 -0200
Eduardo Habkost <address@hidden> wrote:

> On Tue, Feb 05, 2013 at 08:47:29PM +0100, Igor Mammedov wrote:
> > On Tue, 5 Feb 2013 17:04:26 -0200
> > Eduardo Habkost <address@hidden> wrote:
> > 
> > > On Tue, Feb 05, 2013 at 07:46:09PM +0100, Igor Mammedov wrote:
> > > > On Tue, 5 Feb 2013 15:53:04 -0200
> > > > Eduardo Habkost <address@hidden> wrote:
> > > > 
> > > > > On Tue, Feb 05, 2013 at 05:39:24PM +0100, Igor Mammedov wrote:
> > > > > > ORing kvm_default_features to def->kvm_features might set
> > > > > > unsupported bits if kvm_arch_get_supported_cpuid() returns less
> > > > > > bits set than in kvm_default_features. Fix it by removing
> > > > > > kvm_default_features and using only what
> > > > > > kvm_arch_get_supported_cpuid() returns.
> > > > > 
> > > > > This is exactly what we _must not_ do! We can't change CPUID bits
> > > > > in a machine-type or we will change the guest ABI. We must keep
> > > > > them stable for a machine-type and:
> > > > > 
> > > > > * Not change if QEMU is upgraded;
> > > > > * Not change if the host kernel is upgraded;
> > > > > * Not change if running on different host hardware.
> > > > Is all of this applicable to CPU 'host' or only to built-in
> > > > cpu_models?
> > > 
> > > This applies only to the built-in CPU models and the stable (versioned)
> > > machine-types.
> > > 
> > > -cpu host is special and is expected to break the rules above. We could
> > > also create a "pc-next" machine-type where those rules could be broken.
> > Isn't default machine like pc-next?
> 
> No, the default machine is versioned (today it is "pc-i440fx-1.4").
> 
> > [...]
> 
Ok, lets drop this patch.
I'll prepare another one that moves default setting to class_init().