KVM: VMX: Reduce unnecessary saving of host msrs
authorAvi Kivity <avi@qumranet.com>
Tue, 17 Apr 2007 12:30:24 +0000 (15:30 +0300)
committerAvi Kivity <avi@qumranet.com>
Thu, 3 May 2007 07:52:29 +0000 (10:52 +0300)
THe automatically switched msrs are never changed on the host (with
the exception of MSR_KERNEL_GS_BASE) and thus there is no need to save
them on every vm entry.

This reduces vmexit latency by ~400 cycles on i386 and by ~900 cycles (10%)
on x86_64.

Signed-off-by: Avi Kivity <avi@qumranet.com>
drivers/kvm/vmx.c

index 8c0115b..3745e6c 100644 (file)
@@ -78,6 +78,10 @@ static const u32 vmx_msr_index[] = {
 };
 #define NR_VMX_MSR ARRAY_SIZE(vmx_msr_index)
 
+#ifdef CONFIG_X86_64
+static unsigned msr_offset_kernel_gs_base;
+#endif
+
 static inline int is_page_fault(u32 intr_info)
 {
        return (intr_info & (INTR_INFO_INTR_TYPE_MASK | INTR_INFO_VECTOR_MASK |
@@ -1129,6 +1133,10 @@ static int vmx_vcpu_setup(struct kvm_vcpu *vcpu)
                vcpu->host_msrs[j].reserved = 0;
                vcpu->host_msrs[j].data = data;
                vcpu->guest_msrs[j] = vcpu->host_msrs[j];
+#ifdef CONFIG_X86_64
+               if (index == MSR_KERNEL_GS_BASE)
+                       msr_offset_kernel_gs_base = j;
+#endif
                ++vcpu->nmsrs;
        }
 
@@ -1760,7 +1768,9 @@ again:
        fx_save(vcpu->host_fx_image);
        fx_restore(vcpu->guest_fx_image);
 
-       save_msrs(vcpu->host_msrs, vcpu->nmsrs);
+#ifdef CONFIG_X86_64
+       save_msrs(vcpu->host_msrs + msr_offset_kernel_gs_base, 1);
+#endif
        load_msrs(vcpu->guest_msrs, NR_BAD_MSRS);
 
        asm (