Re: [PATCH v15 21/23] KVM: MMU: Disable fast path for private memslots

[Sean Christopherson]

On Thu, May 09, 2024, Michael Roth wrote:
> ---
>  arch/x86/kvm/mmu/mmu.c | 30 ++++++++++++++++++++++++++++--
>  1 file changed, 28 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> index 62ad38b2a8c9..cecd8360378f 100644
> --- a/arch/x86/kvm/mmu/mmu.c
> +++ b/arch/x86/kvm/mmu/mmu.c
> @@ -3296,7 +3296,7 @@ static int kvm_handle_noslot_fault(struct kvm_vcpu *vcpu,
>  	return RET_PF_CONTINUE;
>  }
>  
> -static bool page_fault_can_be_fast(struct kvm_page_fault *fault)
> +static bool page_fault_can_be_fast(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault)
>  {
>  	/*
>  	 * Page faults with reserved bits set, i.e. faults on MMIO SPTEs, only
> @@ -3307,6 +3307,32 @@ static bool page_fault_can_be_fast(struct kvm_page_fault *fault)
>  	if (fault->rsvd)
>  		return false;
>  
> +	/*
> +	 * For hardware-protected VMs, certain conditions like attempting to
> +	 * perform a write to a page which is not in the state that the guest
> +	 * expects it to be in can result in a nested/extended #PF. In this
> +	 * case, the below code might misconstrue this situation as being the
> +	 * result of a write-protected access, and treat it as a spurious case
> +	 * rather than taking any action to satisfy the real source of the #PF
> +	 * such as generating a KVM_EXIT_MEMORY_FAULT. This can lead to the
> +	 * guest spinning on a #PF indefinitely.
> +	 *
> +	 * For now, just skip the fast path for hardware-protected VMs since
> +	 * they don't currently utilize any of this machinery anyway. In the
> +	 * future, these considerations will need to be taken into account if
> +	 * there's any need/desire to re-enable the fast path for
> +	 * hardware-protected VMs.
> +	 *
> +	 * Since software-protected VMs don't have a notion of a shared vs.
> +	 * private that's separate from what KVM is tracking, the above
> +	 * KVM_EXIT_MEMORY_FAULT condition wouldn't occur, so avoid the
> +	 * special handling for that case for now.

Very technically, it can occur if userspace _just_ modified the attributes.  And
as I've said multiple times, at least for now, I want to avoid special casing
SW-protected VMs unless it is *absolutely* necessary, because their sole purpose
is to allow testing flows that are impossible to excercise without SNP/TDX hardware.

> +	 */
> +	if (kvm_slot_can_be_private(fault->slot) &&
> +	    !(IS_ENABLED(CONFIG_KVM_SW_PROTECTED_VM) &&
> +	      vcpu->kvm->arch.vm_type == KVM_X86_SW_PROTECTED_VM))

Heh, !(x && y) kills me, I misread this like 4 times.

Anyways, I don't like the heuristic.  It doesn't tie the restriction back to the
cause in any reasonable way.  Can't this simply be?

	if (fault->is_private != kvm_mem_is_private(vcpu->kvm, fault->gfn)
		return false;

Which is much, much more self-explanatory.