kvm_arch_vcpu_ioctl_get_fpu does not set the mxcsr in the kvm_fpu
struct; conversely kvm_arch_vcpu_ioctl_set_fpu does not set the mxcsr
value in the fxregs_state struct of the vcpu.
This leads to the KVM_GET_FPU ioctl returning 0 as the mxcsr value,
regardless of the actual value on the vcpu; while KVM_SET_FPU leaves the
MXCSR on the vcpu untouched.

Fix kvm_arch_vcpu_ioctl_{get,set}_fpu to properly handle MXCSR.

Signed-off-by: Justinien Bouron <justinien.bouron@gmail.com>
---
 arch/x86/kvm/x86.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 43a6a7efc6ec..c33a2599a497 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -11462,6 +11462,7 @@ int kvm_arch_vcpu_ioctl_get_fpu(struct kvm_vcpu *vcpu, struct kvm_fpu *fpu)
 	fpu->last_ip = fxsave->rip;
 	fpu->last_dp = fxsave->rdp;
 	memcpy(fpu->xmm, fxsave->xmm_space, sizeof(fxsave->xmm_space));
+	fpu->mxcsr = fxsave->mxcsr;
 
 	vcpu_put(vcpu);
 	return 0;
@@ -11486,6 +11487,7 @@ int kvm_arch_vcpu_ioctl_set_fpu(struct kvm_vcpu *vcpu, struct kvm_fpu *fpu)
 	fxsave->rip = fpu->last_ip;
 	fxsave->rdp = fpu->last_dp;
 	memcpy(fxsave->xmm_space, fpu->xmm, sizeof(fxsave->xmm_space));
+	fxsave->mxcsr = fpu->mxcsr;
 
 	vcpu_put(vcpu);
 	return 0;
-- 
2.25.1

On Sat, Sep 17, 2022, Justinien Bouron wrote:
> kvm_arch_vcpu_ioctl_get_fpu does not set the mxcsr in the kvm_fpu
> struct; conversely kvm_arch_vcpu_ioctl_set_fpu does not set the mxcsr
> value in the fxregs_state struct of the vcpu.
> This leads to the KVM_GET_FPU ioctl returning 0 as the mxcsr value,
> regardless of the actual value on the vcpu; while KVM_SET_FPU leaves the
> MXCSR on the vcpu untouched.
> 
> Fix kvm_arch_vcpu_ioctl_{get,set}_fpu to properly handle MXCSR.
> 
> Signed-off-by: Justinien Bouron <justinien.bouron@gmail.com>
> ---
>  arch/x86/kvm/x86.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 43a6a7efc6ec..c33a2599a497 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -11462,6 +11462,7 @@ int kvm_arch_vcpu_ioctl_get_fpu(struct kvm_vcpu *vcpu, struct kvm_fpu *fpu)
>  	fpu->last_ip = fxsave->rip;
>  	fpu->last_dp = fxsave->rdp;
>  	memcpy(fpu->xmm, fxsave->xmm_space, sizeof(fxsave->xmm_space));
> +	fpu->mxcsr = fxsave->mxcsr;
>  
>  	vcpu_put(vcpu);
>  	return 0;
> @@ -11486,6 +11487,7 @@ int kvm_arch_vcpu_ioctl_set_fpu(struct kvm_vcpu *vcpu, struct kvm_fpu *fpu)
>  	fxsave->rip = fpu->last_ip;
>  	fxsave->rdp = fpu->last_dp;
>  	memcpy(fxsave->xmm_space, fpu->xmm, sizeof(fxsave->xmm_space));
> +	fxsave->mxcsr = fpu->mxcsr;

The incoming MXCSR needs to be vetted, e.g. see fpu_copy_uabi_to_guest_fpstate().

That said, this code is incredibly ancient and has been obsolete for years.  I
wonder if we can get away with deprecating KVM_{G,S}ET_FPU on x86 and leaving the
broken behavior as-is.