Re: [PATCH v2 1/6] KVM: x86/mmu: add a new mmu zap helper to indicate memtype changes

[Sean Christopherson <seanjc@google.com>]

On Wed, May 24, 2023, Yan Zhao wrote:
> On Tue, May 23, 2023 at 03:51:49PM -0700, Sean Christopherson wrote:
> > diff --git a/arch/x86/kvm/mtrr.c b/arch/x86/kvm/mtrr.c
> > index 3eb6e7f47e96..a67c28a56417 100644
> > --- a/arch/x86/kvm/mtrr.c
> > +++ b/arch/x86/kvm/mtrr.c
> > @@ -320,7 +320,7 @@ static void update_mtrr(struct kvm_vcpu *vcpu, u32 msr)
> >  	struct kvm_mtrr *mtrr_state = &vcpu->arch.mtrr_state;
> >  	gfn_t start, end;
> >  
> > -	if (!tdp_enabled || !kvm_arch_has_noncoherent_dma(vcpu->kvm))
> > +	if (!kvm_mmu_honors_guest_mtrrs(vcpu->kvm))
> Could we also add another helper kvm_mmu_cap_honors_guest_mtrrs(), which
> does not check kvm_arch_has_noncoherent_dma()?
> 
> +static inline bool kvm_mmu_cap_honors_guest_mtrrs(struct kvm *kvm)
> +{
> +       return !!shadow_memtype_mask;
> +}
> 
> This is because in patch 4 I plan to do the EPT zap when
> noncoherent_dma_count goes from 1 to 0.

Hrm, the 1->0 transition is annoying.  Rather than trying to capture the "everything
except non-coherent DMA" aspect, what about this?

mmu.c:

bool __kvm_mmu_honors_guest_mtrrs(struct kvm *kvm, bool vm_has_noncoherent_dma)
{
	/*
	 * If the TDP is enabled, the host MTRRs are ignored by TDP
	 * (shadow_memtype_mask is non-zero), and the VM has non-coherent DMA
	 * (DMA doesn't snoop CPU caches), KVM's ABI is to honor the memtype
	 * from the guest's MTRRs so that guest accesses to memory that is
	 * DMA'd aren't cached against the guest's wishes.
	 *
	 * Note, KVM may still ultimately ignore guest MTRRs for certain PFNs,
	 * e.g. KVM will force UC memtype for host MMIO.
	 */
	return vm_has_noncoherent_dma && tdp_enabled && shadow_memtype_mask;
}

mmu.h:

bool __kvm_mmu_honors_guest_mtrrs(struct kvm *kvm, bool vm_has_noncoherent_dma);

static inline bool kvm_mmu_honors_guest_mtrrs(struct kvm *kvm)
{
	
	return __kvm_mmu_honors_guest_mtrrs(kvm, kvm_arch_has_noncoherent_dma(kvm));
}

> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 41d7bb51a297..ad0c43d7f532 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -13146,13 +13146,19 @@ EXPORT_SYMBOL_GPL(kvm_arch_has_assigned_device);
> 
>  void kvm_arch_register_noncoherent_dma(struct kvm *kvm)
>  {
> -       atomic_inc(&kvm->arch.noncoherent_dma_count);
> +       if (atomic_inc_return(&kvm->arch.noncoherent_dma_count) == 1) {
> +               if (kvm_mmu_cap_honors_guest_mtrrs(kvm))
> +                       kvm_zap_gfn_range(kvm, 0, ~0ULL);

No need for multiple if statements.  Though rather than have identical code in
both the start/end paths, how about this?  That provides a single location for a
comment.  Or maybe first/last instead of start/end?

static void kvm_noncoherent_dma_start_or_end(struct kvm *kvm)
{
	/* comment goes here. */
	if (__kvm_mmu_honors_guest_mtrrs(kvm, true))
		kvm_zap_gfn_range(kvm, 0, ~0ULL);
}

void kvm_arch_register_noncoherent_dma(struct kvm *kvm)
{
	if (atomic_inc_return(&kvm->arch.noncoherent_dma_count) == 1)
		kvm_noncoherent_dma_start_or_end(kvm);
}
EXPORT_SYMBOL_GPL(kvm_arch_register_noncoherent_dma);

void kvm_arch_unregister_noncoherent_dma(struct kvm *kvm)
{
	if (!atomic_dec_return(&kvm->arch.noncoherent_dma_count))
		kvm_noncoherent_dma_start_or_end(kvm);
}
EXPORT_SYMBOL_GPL(kvm_arch_unregister_noncoherent_dma);