On Fri, 22 May 2020 23:04:51 +0200 Halil Pasic <pa...@linux.ibm.com> wrote:
> On Wed, 20 May 2020 12:23:24 -0400 > "Michael S. Tsirkin" <m...@redhat.com> wrote: > > > On Fri, May 15, 2020 at 12:11:55AM +0200, Halil Pasic wrote: > > > The virtio specification tells that the device is to present > > > VIRTIO_F_ACCESS_PLATFORM (a.k.a. VIRTIO_F_IOMMU_PLATFORM) when the > > > device "can only access certain memory addresses with said access > > > specified and/or granted by the platform". This is the case for a > > > protected VMs, as the device can access only memory addresses that are > > > in pages that are currently shared (only the guest can share/unsare its > > > pages). > > > > > > No VM, however, starts out as a protected VM, but some VMs may be > > > converted to protected VMs if the guest decides so. > > > > > > Making the end user explicitly manage the VIRTIO_F_ACCESS_PLATFORM via > > > the property iommu_on is a minor disaster. Since the correctness of the > > > paravirtualized virtio devices depends (and thus in a sense the > > > correctness of the hypervisor) it, then the hypervisor should have the > > > last word about whether VIRTIO_F_ACCESS_PLATFORM is to be presented or > > > not. > > > > So, how about this: switch iommu to on/off/auto. > > Many thanks for the reveiw, and sorry about the delay on my side. We > have holidays here in Germany and I was not motivated enough up until > now to check on my mails. > > > I've actually played with the thought of switching iommu_platform to > 'on/off/auto', but I didn't find an easy way to do it. I will look > again. This would be the first property of this kind in QEMU, or? virtio-pci uses it for 'disable-legacy'. > > The 'on/off/auto' would be certainly much cleaner form user-interface > perspective. The downsides are that it is more invasive, and more > complicated. I'm afraid that it would also leave more possibilities for > user error. To me, on/off/auto sounds like a reasonable thing to do. What possibilities of 'user error' do you see? Shouldn't we fence off misconfigurations, if the consequences would be disastrous? > > > Add a property with a > > reasonable name "allow protected"? If set allow switch to protected > > memory and also set iommu auto to on by default. If not set then don't. > > > > I think we have "allow protected" already expressed via cpu models. I'm > also not sure how libvirt would react to the idea of a new machine > property for this. You did mean "allow protected" as machine property, > or? "Unpack facility in cpu model" means "guest may transition into pv mode", right? What does it look like when the guest actually has transitioned? > > AFAIU "allow protected" would be required for the !PV to PV switch, and > we would have to reject paravirtualized devices with iommu_platform='off' > on VM construction or hotplug (iommu_platform='auto/on' would be fine). > > Could you please confirm that I understood this correctly? > > > > This will come handy for other things like migrating to hosts without > > protected memory support. > > > > This is already covered by cpu model AFAIK. I don't think we'd want to migrate between pv and non-pv anyway? > > > > > Also, virtio now calls this PLATFORM_ACCESS, maybe we should rename > > the property (keeping old one around for compat)? > > You mean the like rename 'iommu_platform' to 'platform_access'? I like > the idea, but I'm not sure libvirt will like it as well. Boris any > opinions? > > > I feel this will address lots of complaints ... > > > > > Currently presenting a PV guest with a (paravirtualized) virtio-ccw > > > device has catastrophic consequences for the VM (after the hypervisors > > > access to protected memory). This is especially grave in case of device > > > hotplug (because in this case the guest is more likely to be in the > > > middle of something important). > > > > > > Let us manage the VIRTIO_F_ACCESS_PLATFORM virtio feature automatically > > > for virtio-ccw devices, i.e. force it before we start the protected VM. > > > If the VM should cease to be protected, the original value is restored. > > > > > > Signed-off-by: Halil Pasic <pa...@linux.ibm.com> > > > > > > I don't really understand things fully but it looks like you are > > changing features of a device. If so this bothers me, resets > > happen at random times while driver is active, and we never > > expect features to change. > > > > Changing the device features is IMHO all right because the features can > change only immediately after a system reset and before the first vCPU > is run. That is ensured by two facts. > > > First, the feature can only change when ms->pv changes. That is on the > first reset after the VM entered or left the "protected virtualization" > mode of operation. And that switch requires a system reset. Because the > PV switch is initiated by the guest, and the guest is rebooted as a > consequence, the guest will never observe the change in features. This really needs more comments, as it is not obvious to the casual reader. (I also stumbled over the resets.) But I wonder whether we are actually missing those subsystems resets today?