Re: [PATCH v2 13/15] KVM: s390: Configure the guest's CRYCB

[Tony Krowiak]

On 03/02/2018 05:08 AM, David Hildenbrand wrote:
> On 01.03.2018 21:42, Tony Krowiak wrote:
> > On 03/01/2018 04:37 AM, David Hildenbrand wrote:
> > > On 28.02.2018 21:45, Tony Krowiak wrote:
> > > > On 02/28/2018 04:49 AM, David Hildenbrand wrote:
> > > > > > +static int vfio_ap_mdev_open(struct mdev_device *mdev)
> > > > > > +{
> > > > > > +	struct ap_matrix_mdev *matrix_mdev = mdev_get_drvdata(mdev);
> > > > > > +	unsigned long events;
> > > > > > +	int ret;
> > > > > > +
> > > > > > +	matrix_mdev->group_notifier.notifier_call = vfio_ap_mdev_group_notifier;
> > > > > > +	events = VFIO_GROUP_NOTIFY_SET_KVM;
> > > > > > +	ret = vfio_register_notifier(mdev_dev(mdev), VFIO_GROUP_NOTIFY,
> > > > > > +				     &events, &matrix_mdev->group_notifier);
> > > > > > +
> > > > > > +	ret = kvm_ap_configure_matrix(matrix_mdev->kvm,
> > > > > > +				      matrix_mdev->matrix);
> > > > > > +	if (ret)
> > > > > > +		return ret;
> > > > > > +
> > > > > > +	ret = kvm_ap_enable_ie_mode(matrix_mdev->kvm);
> > > > > Can't this happen while the guest is already running? Or what hinders us
> > > > > from doing that?
> > > > I'm not sure exactly what you're asking here. Are you asking if the
> > > > vfio_ap_mdev_open()
> > > > function can be called multiple times while the guest is running? AFAIK
> > > > this will be
> > > > called only once when the mediated device's file descriptor is opened.
> > > > This happens in
> > > > QEMU when the -device vfio-ap device is realized.
> > > Okay, but from a pure interface point of view, this could happen any
> > > time, even while the guest is already running. Patching in the SCB of a
> > > running VCPU is evil.
> > How can this happen while the guest is running? QEMU opens the fd when the
> > device is realized and AFAIK vfio mdev will not allow any other process to
> > open it until the guest is terminated. What am I missing?
> It can't happen right now (the way QEMU uses it), but the kernel
> interface allows it, no?
>
> Anyhow, as discussed this should be handled directly while creating a
> VCPU. Then also CPU hotplug is properly covered.
Here is what I think we should do:

* Set ECA.28 in the VCPU setup function based on whether the CPU model 
feature
  has been turned on by user space as you suggest.

* Replace the kvm_ap_enable_ie_mode() call in the open() above with a 
query of
  the CPU model feature and return an error if it is not turned on.

Does this sound reasonable?

Would it be more appropriate in this case to rename the feature to
KVM_S390_VM_CPU_FEAT_APIE?
>