[RFE] Add support for IPoIB interface driver

[Adrian Chiris]

To support DHCP and Routing services in an infiniband fabric, it is required to leverage the kernel IPoIB driver to create IP interfaces that provide network connectivity to router ports and DHCP clients.

This RFE proposes to add a new IPoIB interface driver to be used by L3 and DHCP agents to provide IP connectivity over Infiniband fabric.

Some background links:

https://tools.ietf.org/html/rfc4390 - DHCP over infiniband
https://tools.ietf.org/html/rfc4391 - Transmission of IPoIB
https://tools.ietf.org/html/rfc4392 - IPoIB
https://www.kernel.org/doc/Documentation/infiniband/ipoib.txt

[Bence Romsics (bence-romsics)]

Hi Adrian,

Are you familiar with the neutron RFE process? Usually they are discussed first on the neutron-drivers meeting (weekly on Fridays):

http://eavesdrop.openstack.org/#Neutron_drivers_Meeting

It may be a good idea to also ping Miguel (mlavalle in irc) - he usually chairs that meeting - so he can plan for it.

[Bence Romsics (bence-romsics)]

And while I'm not closely familiar with Infiniband I'm wondering why do you propose to directly modify l3 and dhcp-agents instead of creating an ML2 mechanism driver (and maybe a type driver too)? Isn't Infiniband a layer 2 technology?

[Adrian Chiris (adrian.chiris)]

Hi Bence,
I know the RFE needs to be approved by neutron drivers, but one would first need to create it :).

I don't want to modify L3 or DHCP agent but to add support for a new interface driver that is able to create/delete (plug/unplug) IPoIB interfaces so, for example, a DHCP server (created by dhcp agent) would be able to serve DHCP requests when the underlying fabric is infiniband.

will ping on IRC.

[Nate Johnston (nate-johnston)]

Just to make sure I understand the use case and what you mean by "underlying fabric", do you have servers with hardware Infiniband network interfaces in them, and you want to make sure L3/DHCP can serve virtual interfaces created on those networks?

[Adrian Chiris (adrian.chiris)]

Hi Nate,
Yes, for both virtual and physical interfaces (i.e for baremetal instances)

While the L2 wiring today in an infiniband network is vendor specific, there is a standard way to provide L3 connectivity using ipoib interfaces.

[OpenStack Infra (hudson-openstack)]

Fix proposed to branch: master
Review: https://review.opendev.org/668510

[Slawek Kaplonski (slaweq)]

I understand Your use case here (I don't know inifiband protocol but I can imagine what You want to achieve) but I'm not sure if this new driver should be in neutron repo and I have this doubts because:
1. I don't think we have a lot of expertise about ininiband in our core team, so we don't have many people who will be able to review such patches probably,
2. How You want to test this new driver upstream?

So maybe better place for such driver would be some external project, something like "networking-infiniband".

[Miguel Lavalle (minsel)]

After discussing this RFE, it was decided that the proposed code for the interface driver should reside in the networking-mellanox repo

[Adrian Chiris (adrian.chiris)]

relevant IRC logs:

http://eavesdrop.openstack.org/irclogs/%23openstack-meeting/%23openstack-meeting.2019-07-12.log.html#t2019-07-12T14:00:07

[OpenStack Infra (hudson-openstack)]

Change abandoned by Adrian Chiris (<email address hidden>) on branch: master
Review: https://review.opendev.org/668510
Reason: a different set of PS will be proposed

[Rodolfo Alonso (rodolfo-alonso-hernandez)]

Bug closed due to lack of activity, please feel free to reopen if needed.