[Rust-VMM] [Rust-vmm] Goals for this list

[Anthony Liguori]

Paolo Bonzini <pbonzini at redhat.com> writes:

> On 18/12/18 00:30, Liguori, Anthony wrote:
>> As a side note, I think having OS X hypervisor framework bindings
>> and whatever the new Windows thing is would be pretty cool.
>
> Yes, indeed.  Hypervisor.framework however is much more complex than KVM
> or WHP because you deal manually with VMCSes and have to do instruction
> emulation in userspace.  QEMU takes a stab at it, but it's not as stable
> as KVM.
>
> *However* Google does have patches for KVM to do instruction emulation
> in userspace, and I'd like to apply them upstream too now that KVM has
> an API test framework (and thus we can know they won't bitrot).
> (Steve, you are in Cc because hint, hint :)).  Once that is in place, I
> guess a minimal x86 emulator written in Rust, porting the emulator code
> that QEMU has for Hypervisor.framework, would be a fun GSoC project for
> a very good student.

That does sound like an awesome project.

>> 2) The crosvm data_model crate.  This one is super critical but easy to
>> misunderstand as it allows for safe access to volatile memory.  Somewhat
>> related is the mmap() bits from sys_util.  Not sure how the crosvm folks
>> feel but I think there is some refactoring here that could be useful to
>> build a memory crate.
>> 
>> 3) Some traits for device model implementations.  It's easy to really
>> bike shed here so I reckon it's best to start with a concrete device
>> model like a UART, work through what is required for interfaces, and
>> then iterate from there.
>> 
>> 4) Common device models with only a single implementation (i.e.
>> 16650A).  Not sure about virtio, maybe.
>
> virtio would be interesting.  One initial target could be a demo
> vhost-user client, it has to set up a memory map, parse vrings, handle
> endianness, etc.  It would be an interesting benchmark for a DMA API.

Maybe.  I'll poke around.

One challenge with DMA is that you have to decide whether you are only
supporting synchronous DMA or you are also trying to support
asynchronous DMA.

My general preference is to avoid complex infrastructure and keep to
simple traits.  A DMA trait can be pretty darn simple with all of the
gooey logic being up to the implementor.  Not sure you would get
meaningful benchmarks there.

Regards,

Anthony Liguori

> The control plane (your item 3) by comparison is a bit less interesting.
>
> Paolo