KVM: nVMX: Further fixes for lazy FPU loading
authorNadav Har'El <nyh@il.ibm.com>
Wed, 25 May 2011 20:15:08 +0000 (23:15 +0300)
committerAvi Kivity <avi@redhat.com>
Tue, 12 Jul 2011 08:45:18 +0000 (11:45 +0300)
KVM's "Lazy FPU loading" means that sometimes L0 needs to set CR0.TS, even
if a guest didn't set it. Moreover, L0 must also trap CR0.TS changes and
NM exceptions, even if we have a guest hypervisor (L1) who didn't want these
traps. And of course, conversely: If L1 wanted to trap these events, we
must let it, even if L0 is not interested in them.

This patch fixes some existing KVM code (in update_exception_bitmap(),
vmx_fpu_activate(), vmx_fpu_deactivate()) to do the correct merging of L0's
and L1's needs. Note that handle_cr() was already fixed in the above patch,
and that new code in introduced in previous patches already handles CR0
correctly (see prepare_vmcs02(), prepare_vmcs12(), and nested_vmx_vmexit()).

Signed-off-by: Nadav Har'El <nyh@il.ibm.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
arch/x86/kvm/vmx.c

index 6e9bebcdd5a68e6e93309434be69a7711c0e3bfd..10b22619e7ecb4e12cb00479d03049ad0f1c8d99 100644 (file)
@@ -1179,6 +1179,15 @@ static void update_exception_bitmap(struct kvm_vcpu *vcpu)
                eb &= ~(1u << PF_VECTOR); /* bypass_guest_pf = 0 */
        if (vcpu->fpu_active)
                eb &= ~(1u << NM_VECTOR);
+
+       /* When we are running a nested L2 guest and L1 specified for it a
+        * certain exception bitmap, we must trap the same exceptions and pass
+        * them to L1. When running L2, we will only handle the exceptions
+        * specified above if L1 did not want them.
+        */
+       if (is_guest_mode(vcpu))
+               eb |= get_vmcs12(vcpu)->exception_bitmap;
+
        vmcs_write32(EXCEPTION_BITMAP, eb);
 }
 
@@ -1473,6 +1482,9 @@ static void vmx_fpu_activate(struct kvm_vcpu *vcpu)
        vmcs_writel(GUEST_CR0, cr0);
        update_exception_bitmap(vcpu);
        vcpu->arch.cr0_guest_owned_bits = X86_CR0_TS;
+       if (is_guest_mode(vcpu))
+               vcpu->arch.cr0_guest_owned_bits &=
+                       ~get_vmcs12(vcpu)->cr0_guest_host_mask;
        vmcs_writel(CR0_GUEST_HOST_MASK, ~vcpu->arch.cr0_guest_owned_bits);
 }
 
@@ -1496,12 +1508,29 @@ static inline unsigned long nested_read_cr4(struct vmcs12 *fields)
 
 static void vmx_fpu_deactivate(struct kvm_vcpu *vcpu)
 {
+       /* Note that there is no vcpu->fpu_active = 0 here. The caller must
+        * set this *before* calling this function.
+        */
        vmx_decache_cr0_guest_bits(vcpu);
        vmcs_set_bits(GUEST_CR0, X86_CR0_TS | X86_CR0_MP);
        update_exception_bitmap(vcpu);
        vcpu->arch.cr0_guest_owned_bits = 0;
        vmcs_writel(CR0_GUEST_HOST_MASK, ~vcpu->arch.cr0_guest_owned_bits);
-       vmcs_writel(CR0_READ_SHADOW, vcpu->arch.cr0);
+       if (is_guest_mode(vcpu)) {
+               /*
+                * L1's specified read shadow might not contain the TS bit,
+                * so now that we turned on shadowing of this bit, we need to
+                * set this bit of the shadow. Like in nested_vmx_run we need
+                * nested_read_cr0(vmcs12), but vmcs12->guest_cr0 is not yet
+                * up-to-date here because we just decached cr0.TS (and we'll
+                * only update vmcs12->guest_cr0 on nested exit).
+                */
+               struct vmcs12 *vmcs12 = get_vmcs12(vcpu);
+               vmcs12->guest_cr0 = (vmcs12->guest_cr0 & ~X86_CR0_TS) |
+                       (vcpu->arch.cr0 & X86_CR0_TS);
+               vmcs_writel(CR0_READ_SHADOW, nested_read_cr0(vmcs12));
+       } else
+               vmcs_writel(CR0_READ_SHADOW, vcpu->arch.cr0);
 }
 
 static unsigned long vmx_get_rflags(struct kvm_vcpu *vcpu)