Re: [PATCH v5 1/4] target/i386: add missing vmx features for several CPU models

[Eduardo Habkost]

On Mon, Jul 13, 2020 at 03:45:55PM +0800, Xiaoyao Li wrote:
> On 7/13/2020 3:23 PM, Chenyi Qiang wrote:
> > 
> > 
> > On 7/11/2020 12:48 AM, Eduardo Habkost wrote:
> > > On Fri, Jul 10, 2020 at 09:45:49AM +0800, Chenyi Qiang wrote:
> > > > 
> > > > 
> > > > On 7/10/2020 6:12 AM, Eduardo Habkost wrote:
> > > > > 
> > > > > I'm very sorry for taking so long to review this.  Question
> > > > > below:
> > > > > 
> > > > > On Fri, Jun 19, 2020 at 03:31:11PM +0800, Chenyi Qiang wrote:
> > > > > > Add some missing VMX features in Skylake-Server,
> > > > > > Cascadelake-Server and
> > > > > > Icelake-Server CPU models based on the output of Paolo's script.
> > > > > > 
> > > > > > Signed-off-by: Chenyi Qiang <chenyi.qiang@intel.com>
> > > > > 
> > > > > Why are you changing the v1 definition instead adding those new
> > > > > features in a new version of the CPU model, just like you did in
> > > > > patch 3/4?
> > > > > 
> > > > 
> > > > I suppose these missing vmx features are not quite necessary for
> > > > customers.
> > > > Just post it here to see if they are worth being added.
> > > > Adding a new version is reasonable. Is it appropriate to put all
> > > > the missing
> > > > features in patch 1/4, 3/4, 4/4 in a same version?
> > > 
> > > Yes, it would be OK to add only one new version with all the new
> > > features.
> > > 
> > 
> > During the coding, I prefer to split the missing vmx features into a new
> > version of CPU model, because the vmx features depends on CPUID_EXT_VMX.
> > I think It would be better to distinguish it instead of enabling the vmx
> > transparently. i.e.
> > {
> >  Â Â Â Â .version = 4,
> >  Â Â Â Â .props = (PropValue[]) {
> >  Â Â Â Â Â Â Â  { "sha-ni", "on" },
> >  Â Â Â Â Â Â Â  ... ...
> >  Â Â Â Â Â Â Â  { "model", "106" },
> >  Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  { /* end of list */ }
> >  Â Â Â Â },
> > },
> > {
> >  Â Â Â Â .version = 5,
> >  Â Â Â Â .props = (PropValue[]) {
> >  Â Â Â Â Â Â Â  { "vmx", "on" }
> 
> Chenyi,
> 
> This is not we have discussed. I prefer to changing the logic of versioned
> CPU model to not add the features in versioned CPU model to
> env->user_features[]. They're not supposed to be added to
> env->user_features[] since they're not set by user through -feature/+feature
> 
> Eduardo,
> 
> What do you think?

If features added by the CPU model versions appear in
user_features, that's a bug.  What's the user-visible symptom you
are seeing because of it?

-- 
Eduardo