Hi,
One of the goals of Project Stratos is to enable hypervisor agnostic backends so we can enable as much re-use of code as possible and avoid repeating ourselves. This is the flip side of the front end where multiple front-end implementations are required - one per OS, assuming you don't just want Linux guests. The resultant guests are trivially movable between hypervisors modulo any abstracted paravirt type interfaces.
In my original thumb nail sketch of a solution I envisioned vhost-user daemons running in a broadly POSIX like environment. The interface to the daemon is fairly simple requiring only some mapped memory and some sort of signalling for events (on Linux this is eventfd). The idea was a stub binary would be responsible for any hypervisor specific setup and then launch a common binary to deal with the actual virtqueue requests themselves.
Since that original sketch we've seen an expansion in the sort of ways backends could be created. There is interest in encapsulating backends in RTOSes or unikernels for solutions like SCMI. There interest in Rust has prompted ideas of using the trait interface to abstract differences away as well as the idea of bare-metal Rust backends.
We have a card (STR-12) called "Hypercall Standardisation" which calls for a description of the APIs needed from the hypervisor side to support VirtIO guests and their backends. However we are some way off from that at the moment as I think we need to at least demonstrate one portable backend before we start codifying requirements. To that end I want to think about what we need for a backend to function.
Configuration =============
In the type-2 setup this is typically fairly simple because the host system can orchestrate the various modules that make up the complete system. In the type-1 case (or even type-2 with delegated service VMs) we need some sort of mechanism to inform the backend VM about key details about the system:
- where virt queue memory is in it's address space - how it's going to receive (interrupt) and trigger (kick) events - what (if any) resources the backend needs to connect to
Obviously you can elide over configuration issues by having static configurations and baking the assumptions into your guest images however this isn't scalable in the long term. The obvious solution seems to be extending a subset of Device Tree data to user space but perhaps there are other approaches?
Before any virtio transactions can take place the appropriate memory mappings need to be made between the FE guest and the BE guest. Currently the whole of the FE guests address space needs to be visible to whatever is serving the virtio requests. I can envision 3 approaches:
* BE guest boots with memory already mapped
This would entail the guest OS knowing where in it's Guest Physical Address space is already taken up and avoiding clashing. I would assume in this case you would want a standard interface to userspace to then make that address space visible to the backend daemon.
* BE guests boots with a hypervisor handle to memory
The BE guest is then free to map the FE's memory to where it wants in the BE's guest physical address space. To activate the mapping will require some sort of hypercall to the hypervisor. I can see two options at this point:
- expose the handle to userspace for daemon/helper to trigger the mapping via existing hypercall interfaces. If using a helper you would have a hypervisor specific one to avoid the daemon having to care too much about the details or push that complexity into a compile time option for the daemon which would result in different binaries although a common source base.
- expose a new kernel ABI to abstract the hypercall differences away in the guest kernel. In this case the userspace would essentially ask for an abstract "map guest N memory to userspace ptr" and let the kernel deal with the different hypercall interfaces. This of course assumes the majority of BE guests would be Linux kernels and leaves the bare-metal/unikernel approaches to their own devices.
Operation =========
The core of the operation of VirtIO is fairly simple. Once the vhost-user feature negotiation is done it's a case of receiving update events and parsing the resultant virt queue for data. The vhost-user specification handles a bunch of setup before that point, mostly to detail where the virt queues are set up FD's for memory and event communication. This is where the envisioned stub process would be responsible for getting the daemon up and ready to run. This is currently done inside a big VMM like QEMU but I suspect a modern approach would be to use the rust-vmm vhost crate. It would then either communicate with the kernel's abstracted ABI or be re-targeted as a build option for the various hypervisors.
One question is how to best handle notification and kicks. The existing vhost-user framework uses eventfd to signal the daemon (although QEMU is quite capable of simulating them when you use TCG). Xen has it's own IOREQ mechanism. However latency is an important factor and having events go through the stub would add quite a lot.
Could we consider the kernel internally converting IOREQ messages from the Xen hypervisor to eventfd events? Would this scale with other kernel hypercall interfaces?
So any thoughts on what directions are worth experimenting with?