On Fri, Apr 29, 2016 at 9:09 AM, Jan Kiszka <address@hidden> wrote:
On 2016-04-25 00:12, David Kiarie wrote:
Add IVRS table for AMD IOMMU. Generate IVRS or DMAR
depending on emulated IOMMU
It seems you lack scope descriptions for the PCI devices in the system.
At least, this is what our jailhouse config generator complains about
right now (didn't look into details yet). If so, the guest OS will
likely not configure the IOMMU appropriately as it thinks that the
devices are passed through anyway.
On Intel, there is an easy way to state "catch them all" in the ACPI
table. Maybe AMD has this as well so that you don't need to add
individual devices.
This is what am currently doing. From what I can see, linux correctly
detects the range but I will have a look at it again before sending
the next version.
There was another issue raised by Michael about denying VFIO in case a
user requests for IOMMU support. I did look at this but since VFIO is
started with '-device' I didn't find a way to deny it in case of IOMMU
support except by parsing command line arguments which turned out to
be a bit ugly.
We could exclude VFIO from translation by IOMMU by excluding it from
the scope descriptions but am not quite sure this will remedy all the
problems relating to VFIO and IOMMU. This will also require the device
scope to explicitly list all the devices.