Re: Device is ineligible for IOMMU domain attach due to platform RMRR requirement

[Alex Williamson <alex.williamson@redhat.com>]

On Sat, 2015-03-07 at 05:13 -0500, Steven DuChene wrote:
> Alex:
> What would be the result of running an earlier kernel that did not have 
> your RMRR patch on a system that was known to have these problems with 
> RMRR issues? Would there possibly be some instability when trying to do 
> PCI passthrough of these same NVidia devices?
> 
> We have a debian install on one of these same systems and it is running 
> a 3.14.23-2 kernel and we are seeing some issues with PCI passthrough.

The potential problems depend on whether the RMRR memory region is
actually used, whether the device or the platform depend on that ongoing
access, and the address space consumed by the RMRR.  In the case of a
VM, the fear is that the RMRR is used as a data reporting location by
the device, that use continues after the device is assigned to the VM,
and that the RMRR memory region overlaps guest RAM.  If all of those
conditions hold true, then we have a memory integrity issue in the VM.
On the platform side, whatever data the platform was depending on the
device to report is now disconnected from specified data reporting range
and lost.

In the case of an Nvidia device, I'd speculate that the more likely
cause of issues for that kernel would be that the VT-d hardware may not
support snoop-control and the kvm-vfio interface that manages whether
KVM emulates cache coherence instructions wasn't added until kernel
v3.15.  Thanks,

Alex