Nova has opinions on which VIF types are acceptable, and these types vary by circumstance. vhost-user can be used if QEMU is in use and hugepages are used on the VM (specifically, if the VM's memory is backed by an open file, in fact); exclusively virtio mechanisms like vhost-user cannot be used when QEMU emulates phsyical hardware such as E1000s; different hosts can have different hypervisors; and so on. It cannot, today, express those preferences to Neutron, so Neutron generally returns a fixed and preconfigured plugging type that may or may not actually be useful to Nova - and when it isn't, VMs fail to boot.
https://review.openstack.org/#/c/390512/ is a Nova spec that describes how Nova will tell Neutron which types are acceptable to it for a specific port when plugging is initiated. This is the companion RFE for Neutron that asks that Neutron's core code and ML2 driver be changed to offer up those preferences over the plugin and ML2 interfaces, respectively (likely no change if we use the port structure to pass the information as we bind) and change the in-tree OVS and LB mechanism drivers to respect the passed types as appropriate.
(I'm not sure how much change this actually requires in Neutron, if any, becase the LB and OVS drivers are not terribly flexible as regards plugging types; but if any example code that can be written for the LB and OVS drivers I want to make sure that I'm covered as I introduce the changes.)
This is fine. File a Neutron spec