x86/PV: don't commit debug register values early in arch_set_info_guest()

They're being taken care of later (via set_debugreg()), and temporarily
copying them into struct vcpu means that bad values may end up getting
loaded during context switch if the vCPU is already running and the
function errors out between the premature and real commit step, leading
to the same issue that XSA-12 dealt with.

Signed-off-by: Jan Beulich <jbeulich@xxxxxxxx>
Acked-by: Ian Campbell <ian.campbell@xxxxxxxxxx>

--- a/xen/arch/x86/domain.c
+++ b/xen/arch/x86/domain.c
@@ -740,11 +740,12 @@ int arch_set_info_guest(
             XLAT_trap_info(v->arch.pv_vcpu.trap_ctxt + i,
                            c.cmp->trap_ctxt + i);
     }
-    for ( i = 0; i < ARRAY_SIZE(v->arch.debugreg); ++i )
-        v->arch.debugreg[i] = c(debugreg[i]);
 
     if ( has_hvm_container_vcpu(v) )
     {
+        for ( i = 0; i < ARRAY_SIZE(v->arch.debugreg); ++i )
+            v->arch.debugreg[i] = c(debugreg[i]);
+
         /*
          * NB: TF_kernel_mode is set unconditionally for HVM guests,
          * so we always use the gs_base_kernel here. If we change this