Re: [PATCH v2 01/25] KVM: x86/mmu: avoid indirect call for get_cr3

[Sean Christopherson <seanjc@google.com>]

On Tue, Mar 08, 2022, Paolo Bonzini wrote:
> On 3/8/22 17:16, Sean Christopherson wrote:
> > 
> > > +static inline unsigned long kvm_mmu_get_guest_pgd(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu)
> > Wrap the params, no reason to make this line so long.
> > 
> > > +{
> > > +#ifdef CONFIG_RETPOLINE
> > > +	if (mmu->get_guest_pgd == kvm_get_guest_cr3)
> > > +		return kvm_read_cr3(vcpu);
> > This is unnecessarily fragile and confusing at first glance.  Compilers are smart
> > enough to generate a non-inline version of functions if they're used for function
> > pointers, while still inlining where appropriate.  In other words, just drop
> > kvm_get_guest_cr3() entirely, a al get_pdptr => kvm_pdptr_read().
> 
> Unfortunately this isn't entirely true.  The function pointer will not match
> between compilation units, in this case between the one that calls
> kvm_mmu_get_guest_pgd and the one that assigned kvm_read_cr3 to the function
> pointer.

Ooh, that's a nasty gotcha.  And that's why your v1 used a NULL entry as a sentinel
for rerouting to kvm_read_cr3().  Hrm, I'm torn between disliking the NULL behavior
and disliking the subtle redirect :-)

Aha!  An idea that would provide line of sight to avoiding retpoline in all cases
once we use static_call() for nested_ops, which I really want to do...  Drop the
mmu hook entirely and replace it with:

static inline kvm_mmu_get_guest_pgd(struct kvm_vcpu *vcpu)
{
	if (!mmu_is_nested(vcpu))
		return kvm_read_cr3(vcpu);
	else
		return kvm_x86_ops.nested_ops->get_guest_pgd(vcpu);
}