virt_type=qemu does not support cpu affinity.

[sean mooney]

Description
===========
QEMU support a number of different backends for code generation.
Qemu with the TCG(tiny code generator) backend does not support
cpu affinity. the TCG backened is used when the nova
virt_type is set to qemu.

Currently by default when we
request a numa topology nova generates vcpupin elements
that confine the guest vcpus to float only across
the core that are associated with a numa node.

Currently by default when we request hugepage memory
nova implictly generates a numa topology of 1 if no
numa topology is specified.

as a result of the above behavior it is not possible
to spawn a guest a guest with a numa topology or
hugepage memory backing on non kvm/xen acclerated qemu backends
as qemu with the TCG backend does not support cpu affinity.

Steps to reproduce
==================

- deploy nova as normal and set virt_type=qemu in nova.conf.
- set hw:numa_nodes=1 or hw:mem_page_size=large in the nova flavor.
- boot an instance with this modified flavor.

Expected result
===============

vm starts with a virtual numa topology of 1 numa node(hw:numa_nodes=1)
or with hugepage backed memory (hw:mem_page_size=large)

Actual result
=============

nova reports libvirt error that cpu affinity is not support by TCG backend.

logs
====
http://paste.openstack.org/show/609710/

[sean mooney (sean-k-mooney)]

i will submit a partial fix for this shortly for the numa/hugepage case.

open question should the scheduler be modified to exclude host with hypervior=qemu if dedicated cpus are requested.

[sean mooney (sean-k-mooney)]

addtional logs

http://paste.openstack.org/show/609710/

[Tetsuro Nakamura (tetsuro0907)]

Link to gerrit manually...

Fix proposed to branch: master
Review: https://review.openstack.org/#/c/465160

[OpenStack Infra (hudson-openstack)]

Change abandoned by sean mooney (<email address hidden>) on branch: master
Review: https://review.openstack.org/465160
Reason: im going to abandon this as it now work if you use a cpu_model that is supported by the mttcg backend which is enable by default on new enough qemu versions.
see
https://review.openstack.org/#/c/652197/
as an example of a ci job that uses that fact