Set cpu_mode to "none" when virt_type is qemu

The Nova code states that the default value for cpu_mode when using
qemu as virt_type should be host-model [1]. However, this is not
actually tested in the gate due to [2], and with QEMU 2.6.0 it is
known to cause issues [3], because there is a chance a new CPU
flag is used, but not emulated by QEMU TCG code.

Setting cpu_mode to "none" in this case as default would prevent us
from hitting issues when this code is most exercised, i.e. CI
environments.

[1] f419825c21/nova/conf/libvirt.py (L442-L448)
[2] https://github.com/openstack-dev/devstack/blob/master/lib/nova_plugins/hypervisor-libvirt#L42
[3] https://bugzilla.redhat.com/show_bug.cgi?id=1404627

Depends-On: I687bc765b77638096b47f823e3a784ce7c8c120f
Change-Id: Ic95755a927eef0580433cd5c8eca1c4c7cad2f69

manifests/compute/libvirt.pp

@@ -24,8 +24,8 @@
 # [*libvirt_cpu_mode*]
 #   (optional) The libvirt CPU mode to configure.  Possible values
 #   include custom, host-model, none, host-passthrough.
-#   Defaults to 'host-model' if libvirt_virt_type is set to either
+#   Defaults to 'host-model' if libvirt_virt_type is set to kvm,
-#   kvm or qemu, otherwise defaults to 'none'.
+#   otherwise defaults to 'none'.
 #
 # [*libvirt_cpu_model*]
 #   (optional) The named libvirt CPU model (see names listed in
@@ -132,7 +132,7 @@ class nova::compute::libvirt (
   # libvirt_cpu_mode has different defaults depending on hypervisor.
   if !$libvirt_cpu_mode {
     case $libvirt_virt_type {
-      'kvm','qemu': {
+      'kvm': {
         $libvirt_cpu_mode_real = 'host-model'
       }
       default: {

releasenotes/notes/cpu_mode-for-virt_type=qemu-is-now-none-da2f1ada88d6d2c0.yaml

@@ -0,0 +1,10 @@
+---
+upgrade:
+  - The default value for nova::compute::libvirt::libvirt_cpu_mode
+    when nova::compute::libvirt::libvirt_virt_type is set to 'qemu'
+    has been changed from 'host-model' to 'none'. The Nova gate
+    does not really test the 'host-model' option in their CI, but
+    'none' due to Devstack configuration, and it is known to create
+    issues with QEMU 2.6 (see 
+    https://bugzilla.redhat.com/show_bug.cgi?id=1404627).
+

spec/acceptance/nova_wsgi_apache_spec.rb

@@ -71,6 +71,12 @@ describe 'basic nova' do
         virtlock_service_name => false,
         virtlog_service_name  => false,
       }
+      # FIXME(jpena): this is only here to avoid an attempted downgrade
+      # of qemu-kvm-ev. Remove after https://review.openstack.org/411179
+      # is merged
+      if $::osfamily == 'RedHat' {
+          Package['qemu-kvm-ev'] -> Class['nova::compute::libvirt::services']
+      }
       class { '::nova::scheduler': }
       class { '::nova::vncproxy': }
 

spec/classes/nova_compute_libvirt_spec.rb

@@ -123,6 +123,14 @@ describe 'nova::compute::libvirt' do
       it { is_expected.to contain_nova_config('libvirt/cpu_model').with_value('kvm64')}
     end
 
+    describe 'with qemu as virt_type' do
+      let :params do
+        { :libvirt_virt_type => 'qemu' }
+      end
+
+      it { is_expected.to contain_nova_config('libvirt/cpu_mode').with_value('none')}
+    end
+
     describe 'with migration_support enabled' do
 
       context 'with vncserver_listen set to 0.0.0.0' do