On Wed, 2024-06-26 at 10:25 +0800, Yan Zhao wrote: > > Maybe better than a comment...? Need to put the whole thing together and > > test it > > still. > Hmm, I think passing KVM_DIRECT_ROOTS only in kvm_tdp_mmu_zap_all() is ok, > because kvm_mmu_uninit_tdp_mmu() will zap all invalidated roots eventually, > though KVM_DIRECT_ROOTS | KVM_INVALID_ROOTS matches more to the function name > zap_all. > > I'm not convinced that the change in tdp_mmu_root_match() above is needed. > Once a root is marked invalid, it won't be restored later. Distinguishing > between invalid direct roots and invalid mirror roots will only result in more > unused roots lingering unnecessarily. The logic for direct roots is for normal VMs as well. In any case, we should not mix generic KVM MMU changes in with mirrored memory additions. So let's keep the existing direct root behavior the same.
