On 04/07/18 10:38, Christoffer Dall wrote: > Implement the required MMIO accessors for GICv2 and GICv3 for the > IGROUPR distributor and redistributor registers. > > This can allow guests to change behavior compared to running on previous > versions of KVM, but only to align with the architecture and hardware > implementations. > > This also allows userspace to configure the groups for interrupts. Note > that this potentially results in GICv2 guests not receiving interrupts > after migration if migrating from an older kernel that exposes GICv2 > interrupts as group 1. > > Cc: Andrew Jones <drjones@xxxxxxxxxx> > Signed-off-by: Christoffer Dall <christoffer.dall@xxxxxxx> > --- > I implemented (but stashed) a version of this which predicated the > behavior based on the value of GICD_IIDR revision field, falling back to > ignoring writes and resetting GICv2 groups to 0 if the guest wrote a > revision less than 2. However, current QEMU implementations simply > don't write the GICD_IIDR, so this doesn't help at all without changing > QEMU anyhow. > > The only actual fix I can see here to work around the problem in the > kernel is to require an opt-in to allow restoring groups from userspace, > but that's a lot of logic to support cross-kernel version migration. > > Question: Do we expect that cross-kernel version migration is a critical > feature that people really expect to work, and do we actually have > examples of catering to this in the kernel elsewhere? (Also, how would > then that relate to the whole 'adding a new sysreg breaks migration' > situation?) I don't really get why QEMU doesn't try to restore GICD_IIDR, while it is definitely trying to restore RO sysregs (and that's how we detect incompatibilities). I think we should at least give userspace a chance to do the right thing. If it doesn't, well, too bad. How bad is that "writable GICD_IIDR" patch? Because at this stage, and in the absence of any comment, I'm close to just pick that series and merge it for 4.19. Thanks, M. -- Jazz is not dead. It just smells funny...
