On Wed, Jan 11, 2023 at 03:07:36PM +0100, Christian König wrote: > Well no, we can perfectly fine enable this as we have done in the past. > What we can't do is rejecting it without driver specific knowledge, > because the hardware might still work correctly. It is an interesting point that any device using PASID for SVA must necessary also be doing ATS/PRI and thus must always be setting the translated bit in their Mem TLPs So, at least at this instant in the kernel we have no need for the ACS check as everything is SVA. This is forward looking where we are going to have non SVA uses of PASIDs and we cannot guarantee translated TLPs. Keep in mind this is pretty much an integrity problem, eg if I allow iommufd to assign page tables to a PASID without PRI we can get bad behaviors if the HW does not route properly. So we are justified to be conservative here to prevent data corruption in bad cases. Jason
