On Wed, Jul 31, 2024 at 11:04:36AM -0600, Alex Williamson wrote: > I'd also be careful about command line parameters. I think we require > one for the vIOMMU to enable PASID support, but I'd prefer to avoid one > on the vfio-pci device, instead simply enabling support when both the > vIOMMU support is enabled and the device is detected to support it. > Each command line option requires support in the upper level tools to > enable it. FWIW AMD seems to have an issue here where if they don't intend to use PASID then they get a bunch more performance if they leave it turned off. So cases like no vIOMMU, or a vIOMMU with no PASID capability, should automatically tell the kernel to disable PASID support for the device. These decisions needs to be made when the HWPT's are first allocated. Jason
