Re: [PATCH 0/2] KVM: x86/mmu: .change_pte() optimization in TDP MMU

[Yan Zhao]

On Wed, Sep 06, 2023 at 05:51:36PM -0700, Sean Christopherson wrote:
> On Thu, Sep 07, 2023, Paolo Bonzini wrote:
> > On Tue, Sep 5, 2023 at 10:18 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
> > > Ooh, actually, maybe we could do
> > >
> > >         static bool <name_tbd> = !IS_ENABLED(CONFIG_KSM);
> > >
> > > and then cross our fingers that that doesn't regress some other funky setups.
> > 
> > It probably breaks gvisor-like setups that use MAP_PRIVATE mmap for
> > memslots? It would instantly break CoW even if memory is never
> > written.
> 
> Doh, I completely forgot about gvisor and the like.
> 
> Yan, I don't think this is worth pursuing.  My understanding is that only legacy,
> relatively slow devices need DMA32.  And as Robin pointed out, swiotlb=force isn't
> something that's likely deployed and certainly isn't intended for performance
> sensitive environments.

Yes. Then will you reconsider my patch 2?
I think it can remove the second EPT violation for each page write and
doesn't break KSM, and COW.