Re: [PATCH 04/18] KVM: x86: hyper-v: Introduce VTL awareness to Hyper-V's PV-IPIs

[Sean Christopherson <seanjc@google.com>]

On Sun, Jun 09, 2024, Nicolas Saenz Julienne wrote:
> HvCallSendSyntheticClusterIpi and HvCallSendSyntheticClusterIpiEx allow
> sending VTL-aware IPIs. Honour the hcall by exiting to user-space upon
> receiving a request with a valid VTL target. This behaviour is only
> available if the VSM CPUID flag is available and exposed to the guest.
> It doesn't introduce a behaviour change otherwise.
> 
> User-space is accountable for the correct processing of the PV-IPI
> before resuming execution.
> 
> Signed-off-by: Nicolas Saenz Julienne <nsaenz@amazon.com>
> ---
>  arch/x86/kvm/hyperv.c | 19 ++++++++++++++++++-
>  1 file changed, 18 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/x86/kvm/hyperv.c b/arch/x86/kvm/hyperv.c
> index 42f44546fe79c..d00baf3ffb165 100644
> --- a/arch/x86/kvm/hyperv.c
> +++ b/arch/x86/kvm/hyperv.c
> @@ -2217,16 +2217,20 @@ static void kvm_hv_send_ipi_to_many(struct kvm *kvm, u32 vector,
>  
>  static u64 kvm_hv_send_ipi(struct kvm_vcpu *vcpu, struct kvm_hv_hcall *hc)
>  {
> +	bool vsm_enabled = kvm_hv_cpuid_vsm_enabled(vcpu);
>  	struct kvm_vcpu_hv *hv_vcpu = to_hv_vcpu(vcpu);
>  	u64 *sparse_banks = hv_vcpu->sparse_banks;
>  	struct kvm *kvm = vcpu->kvm;
>  	struct hv_send_ipi_ex send_ipi_ex;
>  	struct hv_send_ipi send_ipi;
> +	union hv_input_vtl *in_vtl;
>  	u64 valid_bank_mask;
> +	int rsvd_shift;
>  	u32 vector;
>  	bool all_cpus;
>  
>  	if (hc->code == HVCALL_SEND_IPI) {
> +		in_vtl = &send_ipi.in_vtl;

I don't see any value in having a local pointer to a union.  Just use send_ipi.in_vtl.

>  		if (!hc->fast) {
>  			if (unlikely(kvm_read_guest(kvm, hc->ingpa, &send_ipi,
>  						    sizeof(send_ipi))))
> @@ -2235,16 +2239,22 @@ static u64 kvm_hv_send_ipi(struct kvm_vcpu *vcpu, struct kvm_hv_hcall *hc)
>  			vector = send_ipi.vector;
>  		} else {
>  			/* 'reserved' part of hv_send_ipi should be 0 */
> -			if (unlikely(hc->ingpa >> 32 != 0))
> +			rsvd_shift = vsm_enabled ? 40 : 32;
> +			if (unlikely(hc->ingpa >> rsvd_shift != 0))
>  				return HV_STATUS_INVALID_HYPERCALL_INPUT;

The existing error handling doesn't make any sense to me.  Why is this the _only_
path that enforces reserved bits?

Regarding the shift, I think it makes more sense to do:

			/* Bits 63:40 are always reserved. */
			if (unlikely(hc->ingpa >> 40 != 0))
				return HV_STATUS_INVALID_HYPERCALL_INPUT;

			send_ipi.in_vtl.as_uint8 = (u8)(hc->ingpa >> 32);
			if (unlikely(!vsm_enabled && send_ipi.in_vtl.as_uint8))
				return HV_STATUS_INVALID_HYPERCALL_INPUT;

so that it's more obvious exactly what is/isn't reserved when VSM isn't/is enabled.

> +			in_vtl->as_uint8 = (u8)(hc->ingpa >> 32);
>  			sparse_banks[0] = hc->outgpa;
>  			vector = (u32)hc->ingpa;
>  		}
>  		all_cpus = false;
>  		valid_bank_mask = BIT_ULL(0);
>  
> +		if (in_vtl->use_target_vtl)

Due to the lack of error checking for the !hc->fast case, this will do the wrong
thing if vsm_enabled=false.

> +			return -ENODEV;
> +
>  		trace_kvm_hv_send_ipi(vector, sparse_banks[0]);
>  	} else {
> +		in_vtl = &send_ipi_ex.in_vtl;
>  		if (!hc->fast) {
>  			if (unlikely(kvm_read_guest(kvm, hc->ingpa, &send_ipi_ex,
>  						    sizeof(send_ipi_ex))))
> @@ -2253,8 +2263,12 @@ static u64 kvm_hv_send_ipi(struct kvm_vcpu *vcpu, struct kvm_hv_hcall *hc)
>  			send_ipi_ex.vector = (u32)hc->ingpa;
>  			send_ipi_ex.vp_set.format = hc->outgpa;
>  			send_ipi_ex.vp_set.valid_bank_mask = sse128_lo(hc->xmm[0]);
> +			in_vtl->as_uint8 = (u8)(hc->ingpa >> 32);
>  		}
>  
> +		if (vsm_enabled && in_vtl->use_target_vtl)
> +			return -ENODEV;
> +
>  		trace_kvm_hv_send_ipi_ex(send_ipi_ex.vector,
>  					 send_ipi_ex.vp_set.format,
>  					 send_ipi_ex.vp_set.valid_bank_mask);
> @@ -2682,6 +2696,9 @@ int kvm_hv_hypercall(struct kvm_vcpu *vcpu)
>  			break;
>  		}
>  		ret = kvm_hv_send_ipi(vcpu, &hc);
> +		/* VTL-enabled ipi, let user-space handle it */
> +		if (ret == -ENODEV)

I generally don't love "magic" error codes, but I don't see an obvious better
solution either.  The other weird thing is that "ret" is a u64, versus the more
common int or even long.  I doubt it's problematic in practice, just a bit odd.

> +			goto hypercall_userspace_exit;
>  		break;
>  	case HVCALL_POST_DEBUG_DATA:
>  	case HVCALL_RETRIEVE_DEBUG_DATA:
> -- 
> 2.40.1
>