Re: [PATCH v19 059/130] KVM: x86/tdp_mmu: Don't zap private pages for unsupported cases

[Isaku Yamahata]

On Wed, Mar 20, 2024 at 12:56:38AM +0000,
"Edgecombe, Rick P" <rick.p.edgecombe@xxxxxxxxx> wrote:

> On Tue, 2024-03-19 at 16:56 -0700, Isaku Yamahata wrote:
> > When we zap a page from the guest, and add it again on TDX even with
> > the same
> > GPA, the page is zeroed.  We'd like to keep memory contents for those
> > cases.
> > 
> > Ok, let me add those whys and drop migration part. Here is the
> > updated one.
> > 
> > TDX supports only write-back(WB) memory type for private memory
> > architecturally so that (virtualized) memory type change doesn't make
> > sense for private memory.  When we remove the private page from the
> > guest
> > and re-add it with the same GPA, the page is zeroed.
> > 
> > Regarding memory type change (mtrr virtualization and lapic page
> > mapping change), the current implementation zaps pages, and populate
>                                                                      s^
> > the page with new memory type on the next KVM page fault.  
>                                ^s
> 
> > It doesn't work for TDX to have zeroed pages.
> What does this mean? Above you mention how all the pages are zeroed. Do
> you mean it doesn't work for TDX to zero a running guest's pages. Which
> would happen for the operations that would expect the pages could get
> faulted in again just fine.

(non-TDX part of) KVM assumes that page contents are preserved after zapping and
re-populate.  This isn't true for TDX.  The guest would suddenly see zero pages
instead of the old memory contents and would be upset.


> > Because TDX supports only WB, we
> > ignore the request for MTRR and lapic page change to not zap private
> > pages on unmapping for those two cases
> 
> Hmm. I need to go back and look at this again. It's not clear from the
> description why it is safe for the host to not zap pages if requested
> to. I see why the guest wouldn't want them to be zapped.

KVM siltently ignores the request to change memory types.


> > TDX Secure-EPT requires removing the guest pages first and leaf
> > Secure-EPT pages in order. It doesn't allow zap a Secure-EPT entry
> > that has child pages.  It doesn't work with the current TDP MMU
> > zapping logic that zaps the root page table without touching child
> > pages.  Instead, zap only leaf SPTEs for KVM mmu that has a shared
> > bit
> > mask.
> 
> Could this be better as two patches that each address a separate thing?
> 1. Leaf only zapping
> 2. Don't zap for MTRR, etc.

Makes sense. Let's split it.


> > > There seems to be an attempt to abstract away the existence of
> > > Secure-
> > > EPT in mmu.c, that is not fully successful. In this case the code
> > > checks kvm_gfn_shared_mask() to see if it needs to handle the
> > > zapping
> > > in a way specific needed by S-EPT. It ends up being a little
> > > confusing
> > > because the actual check is about whether there is a shared bit. It
> > > only works because only S-EPT is the only thing that has a
> > > kvm_gfn_shared_mask().
> > > 
> > > Doing something like (kvm->arch.vm_type == KVM_X86_TDX_VM) looks
> > > wrong,
> > > but is more honest about what we are getting up to here. I'm not
> > > sure
> > > though, what do you think?
> > 
> > Right, I attempted and failed in zapping case.  This is due to the
> > restriction
> > that the Secure-EPT pages must be removed from the leaves.  the VMX
> > case (also
> > NPT, even SNP) heavily depends on zapping root entry as optimization.
> > 
> > I can think of
> > - add TDX check. Looks wrong
> > - Use kvm_gfn_shared_mask(kvm). confusing
> > - Give other name for this check like zap_from_leafs (or better
> > name?)
> >   The implementation is same to kvm_gfn_shared_mask() with comment.
> >   - Or we can add a boolean variable to struct kvm
> 
> Hmm, maybe wrap it in a function like:
> static inline bool kvm_can_only_zap_leafs(const struct kvm *kvm)
> {
> 	/* A comment explaining what is going on */
> 	return kvm->arch.vm_type == KVM_X86_TDX_VM;
> }
> 
> But KVM seems to be a bit more on the open coded side when it comes to
> things like this, so not sure what maintainers would prefer. My opinion
> is the kvm_gfn_shared_mask() check is too strange and it's worth a new
> helper. If that is bad, then just open coded kvm->arch.vm_type ==
> KVM_X86_TDX_VM is the second best I think.
> 
> I feel both strongly that it should be changed, and unsure what
> maintainers would prefer. Hopefully one will chime in.

Now compile time config is dropped, open code is option.
-- 
Isaku Yamahata <isaku.yamahata@xxxxxxxxx>