Hi Jason, On 3/19/25 1:31 AM, Jason Gunthorpe wrote: > On Tue, Mar 18, 2025 at 07:31:36PM +0100, Eric Auger wrote: >> Nevertheless I don't think anything prevents the acceleration granted >> device from also working with virtio/vhost devices for instance unless >> you unplug the existing infra. > If the accel mode is using something like vcmdq then it is not > possible to work since the invalidations won't even be trapped.
I acknowledged I was more focused on the case without vcmdq which was addressed in the past and now I better see the problem. > > Even in the case where we trap the invalidations it sure is > complicated.. invalidation is done by ASID which is not obviously > related to any specific device. An ASID could be hidden inside a CD > table that is being HW accessed and also inside a CD table that is SW > accessed. The VMM has no way to know what is going on so you'd end up > forced to replicate all the ASID invalidations. :\ Nevertheless I think we shall also support the case without vcmdq (currently supported in this series). And this one looks more compatible with emulated devices althout less optimized. > > It just doesn't seem worthwhile to try to make it all work. > > I'd suggest arranging to share some of the SMMUv3 emulation code, > maybe with a library/headerfile or something, but I think it does make > sense they would be different implementations given how completely > different they should be. I agree with can do our utmost to separate implementations. I more concerned about having libvirt guessing what kind of devices it shall use. on x86 libvirt needs to use -device intel-iommu,caching-mode=on if one wants to protect a VFIO device. So this looks like similar to adding accel=on on ARM. Eric > > Jason >
