2021-10-28 Project Stratos Sync Meeting Notes

Participants

• @Mathieu Poirier

• @Alex Bennée

• @Vincent Guittot

• @Don Harbin

• @Arnd Bergmann

• Christopher Clark (Xen/Argo)

• @Randy Linnell

• Azzadine (QC)

• Srinvas (QC)

Agenda

• Argo Research Card (STR-66: Research hypervisor mediated copy approaches including ArgoClosed )

• General Updates

• Action Follow Ups

Discussion

@Arnd Bergmann had an update on the Fat VirtQueue work.

A number of issues had been found that showed some leaky abstractions between the FE drivers and the VirtIO transports. His experimentation with virtio-console had it assuming a small virtq size (32 queue entries) but without knowledge of the total size of the queue attempts to allocate more pages than the queue could hold. As a result this caused everything to fall over.

There was discussion about back pressure and if there are mechanisms like the network drivers use that could prevent pushing more packets than could be consumed.

There are also assumptions being made that appear to be outside the virtio specification but happen to currently work. This doesn’t bode well for new virtio transports.

Christopher Clark gave a quick overview of the Argo interface and thought the problems Arnd described may well impact a virtio-over-argo transport.

There are Argo implementations for Xen, micro Xen, and something like it in hyperv so it has seen some portability. We discussed a VM-to-VM communication using the Argo transport in Linux kernel and userspace LD_PRELOAD library. This has successfully exposed d-bus across guests.

We had a brief diversion to discuss exposing wifi via virtio and discussed a number of options

• exposing the host netlink to guest doesn’t really abstract stuff

• exposing the host NM to the guest was deemed overly complication. CC mentioned an OpenXT implementation of the model which had left engineers not recommending such a model again

• the consensus was extending the existing virtio-net interface made most sense as the common kernel interface would be exposed to user space.

  • it was acknowledged there is a risk of baking linux-isms into virtio here

  • the discussion on wifi was driven by AGL/automotive discussions AZ suggested we should talk to the Android folks about their thoughts on this. @Arnd Bergmann mentioned there is an upcoming meeting that this could be brought up in