KVM: add explicit barrier to kvm_vcpu_kick

kvm_vcpu_kick() must issue a general memory barrier prior to reading
vcpu->mode in order to ensure correctness of the mutual-exclusion
memory barrier pattern used with vcpu->requests.  While the cmpxchg
called from kvm_vcpu_kick():

 kvm_vcpu_kick
   kvm_arch_vcpu_should_kick
     kvm_vcpu_exiting_guest_mode
       cmpxchg

implies general memory barriers before and after the operation, that
implication is only valid when cmpxchg succeeds.  We need an explicit
barrier for when it fails, otherwise a VCPU thread on its entry path
that reads zero for vcpu->requests does not exclude the possibility
the requesting thread sees !IN_GUEST_MODE when it reads vcpu->mode.

kvm_make_all_cpus_request already had a barrier, so we remove it, as
now it would be redundant.

Signed-off-by: Andrew Jones <drjones@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

include/linux/kvm_host.h

@@ -270,6 +270,12 @@ struct kvm_vcpu {
 
 static inline int kvm_vcpu_exiting_guest_mode(struct kvm_vcpu *vcpu)
 {
+	/*
+	 * The memory barrier ensures a previous write to vcpu->requests cannot
+	 * be reordered with the read of vcpu->mode.  It pairs with the general
+	 * memory barrier following the write of vcpu->mode in VCPU RUN.
+	 */
+	smp_mb__before_atomic();
 	return cmpxchg(&vcpu->mode, IN_GUEST_MODE, EXITING_GUEST_MODE);
 }