vfio-pci module not loaded if vswitch_type=none

[Steven Webster]

Brief Description
-----------------
If the vswitch_type of the system is set to "none", the vfio-pci module is not loaded on worker nodes. This makes sense for an openstack enabled system. However, with the introduction of the multus/sriov CNI plugins, this can cause an issue on a non-openstack enabled node if the user wants to use a DPDK enabled NetworkAttachmentDefinition.

The sriov-cni will try to bind a device to the vfio-pci driver, but the pod will be unable to launch as the sriov-cni will report an error binding to the unloaded module.

The workaround is to simply load the module manually, but we want to have this occur without user intervention.

Severity
--------
Major: Unable to use a DPDK enabled NetworkAttachmentDefinition in a pod without loading the vfio-pci module.

Steps to Reproduce
------------------
- Configure system with vswitch_type="none"
- Enable the SRIOV device plugin: system host-label-assign sriovdp=enabled
- Create a DPDK enabled NetworkAttachmentDefinition
- Launch a pod referencing the NetworkAttachmentDefinition

Expected Behavior
------------------
We should automatically load the vfio-pci module if the label sriovdp=enabled is set. This can probably be achieved with an sriov-cni init container.

Actual Behavior
----------------
User must manually load the module

Reproducibility
---------------
Reproducible

System Configuration
--------------------
Any system with non-openstack enabled workers and sriovdp=enabled label.

Branch/Pull Time/Commit
-----------------------
Master

[Ghada Khalil (gkhalil)]

Marking as release gating; related to new multus/sriov container support

[OpenStack Infra (hudson-openstack)]

Reviewed: https://review.opendev.org/661757
Committed: https://git.openstack.org/cgit/starlingx/config/commit/?id=13b142ff8ba25b6035b461218ef86d9fb14db2ad
Submitter: Zuul
Branch: master

commit 13b142ff8ba25b6035b461218ef86d9fb14db2ad
Author: Steven Webster <email address hidden>
Date: Mon May 27 14:25:28 2019 -0500

    Integration with latest SR-IOV CNI images

    As part of the ongoing development of the sriov-cni and
    sriov-device-plugin, the DPDK NetworkAttachmentDefinition
    configuration options have been deprecated.

    Previously, we used this functionality to have the sriov-cni
    plugin perform the device bind from netdevice (kernel) to
    vfio (userspace), and simply set sriov-device-plugin
    deviceType configuration parameter to 'netdevice'.

    Going forward, we must add a mechanism for a user to define
    the deviceType at the interface configuration level. This
    means an SR-IOV enabled device can no longer have a mix of
    netdevice, vfio chosen by the NeworkAttachmentDefinition.
    That is, it must be determined by the user beforehand which
    type of virtual function driver (kernel or DPDK) a device's
    VFs should have.

    This commit includes the cgtsclient, API, DB and puppet
    related changes required for a user to set the VF driver type.

    In terms of the cgts-client, the following parameter has been
    added: --vf-driver. Example usage for a device intended to
    be used with a DPDK application is as follows:

    system host-if-modify -m 1500 -n sriov0 -d ${DATANET} \
      -c pci-sriov -N ${NUM_VFS} --vf-driver=vfio ${WORKER_NAME} \
      ${INTERFACE_UUID}

    If the user does not specify a vf-driver, the default device
    type will remain as it is today as 'netdevice'. The user can
    also choose to explicitly set the --vf-driver to 'netdevice'
    for the same effect. In this case, a check is made to ensure
    the VF driver has been detected and reported by the sysinv
    agent.

    Story: 2005208
    Task: 33485
    Closes-Bug: 1829565
    Change-Id: I8f6f27b79c7fafa03873e71473f7694991142e50
    Signed-off-by: Steven Webster <email address hidden>