Open Stack

On 8/23/2018 7:10 AM, Matt Riedemann wrote:
> On 8/23/2018 6:34 AM, Moshe Levi wrote:
>> Recent change in tripleo [1] configure nova rx_queue_size 
>> tx_queue_size config by default.
>>
>> It seem that this config option breaks booting vm with SR-IOV. See [2]
>>
>> The issues is because of this code [3] which configure virtio queue 
>> size if the in the interface xml the driver is vhost or None.
>>
>> In case of SR-IOV the driver is also None and that why we get the error.
>>
>> A quick fix will be adding driver=vfio to [4]
>>
>> I just wonder if there are other interface in the libvirt xml which 
>> this can have the same issue.
>>
>> [1] - 
>> https://github.com/openstack/tripleo-heat-templates/commit/444fc042dca3f9a85e8f7076ce68114ac45478c7#diff-99a22d37b829681d157f41d35c38e4c5 
>>
>>
>> [2] - http://paste.openstack.org/show/728666/
>>
>> [3] - https://review.openstack.org/#/c/595592/
>>
>> [4] - 
>> https://github.com/openstack/nova/blob/34956bea4beb8e5ba474b42ba777eb88a5eadd76/nova/virt/libvirt/designer.py#L123 
>>
>>
> 
> Quick note, your [3] and [4] references are reversed.
> 
> Nice find on this, it's a regression in Rocky. As such, please report a 
> bug so we can track it as an RC3 potential issue. Note that RC3 is *today*.

Moshe had to leave for the day. The IRC conversation about this bug was 
confusing at best, and it sounds like we don't know what the correct 
solution to make the rx/tx queues work with vnic_type direct interfaces. 
Given that, I would like to know:

* What do we know actually does work with rx/tx queue sizes? Is it just 
macvtap ports? Is that was the feature was tested with?

* If we have a known good tested vnic_type with the rx/tx queue config 
options in Rocky, let's put out a known limitations release note and 
update the help text for those config options to mention that only known 
types of interfaces work with them.

Then people can work on fixing the configs to work with other types of 
vnics in Stein when there is actually time to test the changes other 
than unit tests.

-- 

Thanks,

Matt