On Thu, Jun 4, 2020 at 7:51 AM Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote: > > On 02/06/20 02:11, Jim Mattson wrote: > > I can only assume that no one cares that KVM_GET_SREGS/KVM_SET_SREGS > > is broken for PAE mode guests (i.e. KVM_GET_SREGS doesn't capture the > > PDPTRs and KVM_SET_SREGS re-reads them from memory). > > > > Presumably, since AMD's nested paging is broken for PAE mode guests, > > the kvm community has made the decision not to get things right for > > Intel either. Can anyone confirm? This was all before my time. > > Yes, pretty much. The PDPTRs are not part of the saved state, we just > treat them as a small third level in the radix tree. Of course, for > nested VMX they are properly synced to the VMCS12 and serialized by > KVM_SET_NESTED_STATE. > > Out of curiosity are there OSes that rely on the PDPTRs remaining cached > until the next CR3 load? None that I know of. It's interesting that Intel has taken great pains to virtualize the architected behavior, but AMD just shrugged it off.
