>>>> This patch solves the problems where it is possible for the kernel to >>>> have VMAs pointing at cachable memory without causing >>>> pfn_is_map_memory() to be true, eg DAX memremap cases and CXL/pre-CXL >>>> devices. This memory is now properly marked as cachable in KVM. >>> >>> Does this only imply in worse performance, or does this also affect >>> correctness? I suspect performance is the problem, correct? >> >> Correctness. Things like atomics don't work on non-cachable mappings. > > Hah! This needs to be highlighted in the patch description. And maybe > this even implies Fixes: etc? Understood. I'll put that in the patch description. >>> Likely you assume to never end up with COW VM_PFNMAP -- I think it's >>> possible when doing a MAP_PRIVATE /dev/mem mapping on systems that allow >>> for mapping /dev/mem. Maybe one could just reject such cases (if KVM PFN >>> lookup code not already rejects them, which might just be that case IIRC). >> >> At least VFIO enforces SHARED or it won't create the VMA. >> >> drivers/vfio/pci/vfio_pci_core.c: if ((vma->vm_flags & VM_SHARED) == 0) > > That makes a lot of sense for VFIO. So, I suppose we don't need to check this? Specially if we only extend the changes to the following case: - type is VM_PFNMAP && - user mapping is cacheable (MT_NORMAL or MT_NORMAL_TAGGED) && - The suggested VM_FORCE_CACHED is set. - Ankit Agrawal
