On Mon, Aug 10, 2020 at 12:08:05PM +0200, Greg Kurz wrote:
> Similarly to what was done with XICS-on-XIVE and XIVE native KVM devices
> with commit 5422e95103cf ("KVM: PPC: Book3S HV: XIVE: Replace the 'destroy'
> method by a 'release' method"), convert the historical XICS KVM device to
> implement the 'release' method. This is needed to run nested guests with
> an in-kernel IRQ chip. A typical POWER9 guest can select XICS or XIVE
> during boot, which requires to be able to destroy and to re-create the
> KVM device. Only the historical XICS KVM device is available under pseries
> at the current time and it still uses the legacy 'destroy' method.
>
> Switching to 'release' means that vCPUs might still be running when the
> device is destroyed. In order to avoid potential use-after-free, the
> kvmppc_xics structure is allocated on first usage and kept around until
> the VM exits. The same pointer is used each time a KVM XICS device is
> being created, but this is okay since we only have one per VM.
>
> Clear the ICP of each vCPU with vcpu->mutex held. This ensures that the
> next time the vCPU resumes execution, it won't be going into the XICS
> code anymore.
>
> Signed-off-by: Greg Kurz <groug@xxxxxxxx>
Thanks, applied to my kvm-ppc-next branch.
Paul.
