KVM: SVM: obey guest PAT

commit 15038e14724799b8c205beb5f20f9e54896013c3 upstream.

For many years some users of assigned devices have reported worse
performance on AMD processors with NPT than on AMD without NPT,
Intel or bare metal.

The reason turned out to be that SVM is discarding the guest PAT
setting and uses the default (PA0=PA4=WB, PA1=PA5=WT, PA2=PA6=UC-,
PA3=UC).  The guest might be using a different setting, and
especially might want write combining but isn't getting it
(instead getting slow UC or UC- accesses).

Thanks a lot to geoff@hostfission.com for noticing the relation
to the g_pat setting.  The patch has been tested also by a bunch
of people on VFIO users forums.

Fixes: 709ddebf81cb40e3c36c6109a7892e8b93a09464
Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=196409
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Tested-by: Nick Sarnie <commendsarnex@gmail.com>
Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
Signed-off-by: Ben Hutchings <ben@decadent.org.uk>

diff --git a/arch/x86/kvm/svm.c b/arch/x86/kvm/svm.c
index fba31b7..4bdd0ac 100644 (file)
--- a/arch/x86/kvm/svm.c
+++ b/arch/x86/kvm/svm.c
@@ -3040,6 +3040,13 @@ static int svm_set_msr(struct kvm_vcpu *vcpu, unsigned ecx, u64 data)
        struct vcpu_svm *svm = to_svm(vcpu);
 
        switch (ecx) {
+       case MSR_IA32_CR_PAT:
+               if (!kvm_mtrr_valid(vcpu, MSR_IA32_CR_PAT, data))
+                       return 1;
+               vcpu->arch.pat = data;
+               svm->vmcb->save.g_pat = data;
+               mark_dirty(svm->vmcb, VMCB_NPT);
+               break;
        case MSR_IA32_TSC:
                kvm_write_tsc(vcpu, data);
                break;