Re: [PATCH v3 2/4] target/i386: Remove monitor from some CPU models

[Eduardo Habkost <ehabkost@redhat.com>]

On Mon, Mar 02, 2020 at 07:47:28PM +0800, Tao Xu wrote:
> On 2/29/2020 5:39 AM, Eduardo Habkost wrote:
> > On Wed, Feb 12, 2020 at 04:13:26PM +0800, Tao Xu wrote:
> > > Add new version of Snowridge, Denverton, Opteron_G3, EPYC, and Dhyana
> > > CPU model to uremove MONITOR/MWAIT featre.
> > > 
> > > After QEMU/KVM use "-overcommit cpu-pm=on" to expose MONITOR/MWAIT
> > > (commit id 6f131f13e68d648a8e4f083c667ab1acd88ce4cd), the MONITOR/MWAIT
> > > feature in these CPU model is unused.
> > > 
> > > Signed-off-by: Tao Xu <tao3.xu@intel.com>
> > 
> > What exactly is the problem you are trying to fix?
> > 
> > No CPU model will ever have monitor=on set by default with KVM,
> > because kvm_default_props has a monitor=off element.
> > 
> 
> Maybe it is not a fix. For example, when we boot a guest with Denverton
> cpu model, guest cannot detect MONITOR/MWAIT and boot with no warning,
> because of "monitor=off" by default. The MONITOR/MWAIT feature in these CPU
> model is unused,but no harm. I am wondering if we should remove it from
> existing CPU models.

As monitor=off is on kvm_default_props, changing the CPU model
table will only affect other accelerators (e.g. TCG, where
MONITOR/MWAIT support is advertised as supported).

We shouldn't be dictating policy for other accelerators just
because KVM doesn't support it.  Removing the feature on
kvm_default_props is sufficient.

-- 
Eduardo