On Thu, Jun 20, 2024, Rick P Edgecombe wrote: > On Tue, 2024-06-18 at 07:34 -0700, Sean Christopherson wrote: > > There's also option: > > > > c) Init disabled_quirks based on VM type. > > > > I.e. let userspace enable the quirk. If the VMM wants to shoot its TDX VM > > guests, then so be it. That said, I don't like this option because it > > would create a very bizarre ABI. > > I think we actually need to force it on for TDX because kvm_mmu_zap_all_fast() > only zaps the direct (shared) root. If userspace decides to not enable the > quirk, mirror/private memory will not be zapped on memslot deletion. Then later > if there is a hole punch it will skip zapping that range because there is no > memslot. Then won't it let the pages get freed while they are still mapped in > the TD? > > If I got that right (not 100% sure on the gmem hole punch page freeing), I think > KVM needs to force the behavior for TDs. What I was suggesting is that we condition the skipping of the mirror/private EPT pages tables on the quirk, i.e. zap *everything* for TDX VMs if the quirk is enabled. Hence the very bizarre ABI.
