Hi,
This is one of several emails to follow up on Linaro's internal KWG sprint last week in Cambridge where a number of Project Stratos hackers discussed what next steps we have and started to think about future work. I am splitting the update into several emails so I can freely CC the relevant lists for each without too much cross-posting spam.
Intro =====
We've made good progress over the last year and have up-streamed a number of device models as vhost-user daemons. We have also gotten our first proof of concept build of the xen-vhost-master which has allowed us to reuse these backends on the Xen hypervisor.
Outstanding work ================
vm-virtio definitions ---------------------
Given our vhost-user daemons were not re-implementing existing virtio device models a number of the queue handling definitions are in the vhost-device repository itself. As discussed before now we have these working we should migrate common definitions to the vm-virtio crate so in-VMM virtio emulation can re-use this code.
Get outstanding vsock PR merged -------------------------------
We actually have two outstanding PR's against the vhost-device repository which implement virtio-vsock and virtio-scsi. They were done as GSoC projects but didn't get merged at the time due to lack of review. They currently have outstanding requests for code changes but due to the nature of GSoC it looks like the original authors don't have time to make the changes which is understandable given changes the repository has gone through over the last two years.
I'm agnostic about virtio-scsi but given the usefulness of virtio-vsock it seems a shame to leave an implementation to wither on a branch. There has been some work on vm-virtio to improve the queue handling and with Andreea's help I have a branch that uses that. Should we just pick up the branch and finish the pull request process?
Sort out an official vhost-master repository in rust-vmm --------------------------------------------------------
The rust-vmm project has the vhost-user-backend which implements the core backend behaviour for handling vhost-user messages. There is also an abstraction for vhost (user and kernel handling) from the VMM side in the vhost repository. However it doesn't provide everything needed to implement a full vhost-master. Currently Viresh is using:
https://github.com/vireshk/vhost-user-master
is the xen-vhost-master project which is constructed from the in-VMM vhost-master bits from Cloud Hypervisor. We should get this properly up-streamed into the rust-vmm project.
Should this be merged into the existing rust-vmm/vhost repository or does it require it's own repository?
Properly document and support cross-compilation -----------------------------------------------
Currently most of our testing is on Arm systems and currently we are either:
- hacking up the local repo for cross-compilation or - doing a "native" build in an QEMU emulated Aarch64 system
the second option is potentially quite slow, at least for the first build. Given building backends for non-x86 systems is core to Linaro's goals we should properly support cross compilation for the vhost-device repository and document it. This should be also be enabled in the CI to ensure the configuration doesn't bitrot.
See Also ========
Other subjects discussed will be the subject of other emails today with different distribution lists. These are:
- Xen specific enabling work - Additional virtio devices - Integrating rust-vmm with QEMU
Happy reading ;-)