Re: [PATCH] KVM: x86/tdp_mmu: Trigger the callback only when an interesting change

[Yan Zhao]

On Fri, Sep 27, 2024 at 06:51:58AM -0700, Sean Christopherson wrote:
> On Thu, Sep 26, 2024, Yan Zhao wrote:
> > On Thu, Sep 12, 2024 at 05:07:57PM -0700, Sean Christopherson wrote:
> > > > diff --git a/arch/x86/kvm/mmu/spte.h b/arch/x86/kvm/mmu/spte.h
> > > > index a72f0e3bde17..1726f8ec5a50 100644
> > > > --- a/arch/x86/kvm/mmu/spte.h
> > > > +++ b/arch/x86/kvm/mmu/spte.h
> > > > @@ -214,6 +214,12 @@ extern u64 __read_mostly shadow_nonpresent_or_rsvd_mask;
> > > >   */
> > > >  #define FROZEN_SPTE	(SHADOW_NONPRESENT_VALUE | 0x5a0ULL)
> > > >  
> > > > +#define EXTERNAL_SPTE_IGNORE_CHANGE_MASK		\
> > > > +	(shadow_acc_track_mask |			\
> > > > +	 (SHADOW_ACC_TRACK_SAVED_BITS_MASK <<		\
> > > > +	  SHADOW_ACC_TRACK_SAVED_BITS_SHIFT) |		\
> > > 
> > > Just make TDX require A/D bits, there's no reason to care about access tracking.
> > If KVM_PRE_FAULT_MEMORY is allowed for TDX and if
> > cpu_has_vmx_ept_ad_bits() is false in TDX's hardware (not sure if it's possible),
> 
> Make it a requirement in KVM that TDX hardware supports A/D bits and that KVM's
> module param is enabled.  EPT A/D bits have been supported in all CPUs since
> Haswell, I don't expect them to ever go away.
Got it!

> 
> > access tracking bit is possbile to be changed, as in below scenario:
> > 
> > 1. KVM_PRE_FAULT_MEMORY ioctl calls kvm_arch_vcpu_pre_fault_memory() to map
> >    a GFN, and make_spte() will call mark_spte_for_access_track() to
> >    remove shadow_acc_track_mask (i.e. RWX bits) and move R+X left to
> >    SHADOW_ACC_TRACK_SAVED_BITS_SHIFT.
> > 2. If a concurrent page fault occurs on the same GFN on another vCPU, then
> >    make_spte() in that vCPU will not see prefetch and the new_spte is
> >    with RWX bits and with no bits set in SHADOW_ACC_TRACK_SAVED_MASK.
> 
> This should be fixed by the mega-series.  I'll make sure to Cc you on that series.
Thanks!

> Thanks much for the input and feedback!
:)