Re: virtio-msg inter-VM transport vs virtio-vhost-user

Demi Marie Obenour <[email protected]> Wed, 4 Mar 2026 04:18:16 -0500
Newsgroups	dev.linux.lists.virtio-comment,org.nongnu.qemu-devel
Message-ID	<[email protected]>
On 3/4/26 03:26, Bertrand Marquis wrote:
> Hi Demi,
> 
>> On 3 Mar 2026, at 18:56, Demi Marie Obenour <[email protected]> wrote:
>>
>> Spectrum (https://spectrum-os.org) is going to be implementing
>> virtio devices outside of the host.  One proposed method of doing
>> this is virtio-vhost-user, which is a virtio device that allows a
>> VM to expose a vhost-user device to another VM.  For instance, one
>> could assign a NIC to one VM and have it provide a vhost-user-net
>> device for use by a different VM.
>>
>> I brought this up on the KVM/QEMU community call today.  Alex Bennée
>> recommended using virtio-msg instead.  However, I have a few concerns
>> with this:
>>
>> 1. Virtio-msg buses are specific to a given hypervisor or (in the
>>   case of FF-A) to a given CPU architecture.  None of the current
>>   buses support KVM on platforms other than Arm64.  Therefore,
>>   a brand-new bus would be needed.
> 
> Even FF-A is not useable at the moment with KVM as there is no FF-A
> support for VM to VM in KVM (only host can communicate with the secure
> world).

Ah, I thought that pKVM had pVM <=> pVM communication implemented.

> MMIO or PCI based virtio in KVM case is working and was not really
> the target of our work.
> 
> pKVM is a target and is being worked on using FF-A but pVM to Host
> virtio is still using PCI at the moment

Makes sense.  Does that involve any confidential computing-specific
code in the PCI subsystem, or has the subsystem as a whole been
hardened against malicious devices?  That matters for some of my
use-cases, where a device may be present but not authorized for use.
It's supposed to be passed through to a VM.

> Now creating a KVM specific bus reusing the concept of a FIFO
> to transfer the messages between a VM and Host is definitely possible
> to do and should not be that complex.

Can you give a rough estimate of how much code you are referring to?

> Right now i am working on backend implementation where:
> - the bus implementation would be in linux kernel allowing several
> implementations like FF-A, xen or others to be done as linux drivers.

What is the advantage of having the bus implementation in the kernel
as opposed to userspace?  Is it because the bus implementation is
responsible for protecting the kernel from malicious userspace?

In case you can't tell already, I'm a fan of microkernels :).

> - have a bus interface provided to user land so that Qemu could contain
> the transport implementation but would not need to be modified for new
> bus implementations.
> 
> Reusing this, it should be fairly simple to define a KVM bus and reuse
> the other parts of the implementations.

How long do you think it will be before this could be included in
upstream, mainline Linux?  I imagine this would need to wait for your
existing virtio-msg work to be upstreamed, and then I would need to
upstream a separate driver and spec.

>> 2. Virtio-msg requires not-yet-upstream drivers in both the frontend
>>   (the VM using the device) and the backend (the VM providing the
>>   device).  Vhost-user uses any of the existing transports, such as
>>   PCI or MMIO.  This means that upstream drivers can be used in the
>>   frontend, and also enables supports for Windows and other guests
>>   that lack support for virtio-msg.
> 
> This is definitely true and will always stay true. To use virtio-msg you
> will need a new transport implementation for it when you want to use
> it and the bus implementation(s) you require. To be used in windows
> those part will also be needed.
> 
> In your example here you rely on existing MMIO or PCI transport and
> existing vhost implementations. This is definitely quicker to do and use.
> The goal is not replace what works but to provide solutions for use cases
> where MMIO or PCI currently do not work or need to be optimized.

My use-case doesn't have the same restrictions, as long as config
space access is rare and denial of service is not a concern.

>> 3. Vhost-user is already widely deployed, so frontend implementations
>>   are quite well tested.  A KVM-specific virtio-msg transport would
>>   serve only one purpose: driver VMs (with assigned devices) on
>>   non-Arm64 platforms.  This is a quite niche use-case.  Therefore,
>>   I'm concerned that the needed frontend code will be poorly tested
>>   and bitrot.
> 
> We are in the process of defining the specification for virtio-msg and we
> are working on implementations in parallel so our implementations are
> for now not widely tested that is clear.
> Now a specific KVM virtio message bus implementation would reuse
> the transport and driver implementations which would be used on any
> platforms in the future. I am not following your niche use-case here
> and the poorly tested argument. Maybe i am missing something.
> 
>>
>> Manos Pitsidianakis stated that vhost-user does not make sense in
>> this case.  Why is that?  Would it make sense to use virtio-msg
>> between VMM and its VM, and expose a vhost-user device to the
>> outside world?  What about having the virtio-vhost-user guest driver
>> emulate a virtio-msg transport, so that it can be used with any device
>> implementation supporting virtio-msg?
> 
> I am not following your point here. You want to do virtio on top of virtio ?

Yes, actually!  More specifically, I want to use a virtio device in
one VM to implement a virtio device for a different VM.

To avoid confusion, and to match Xen terminology, I'll call the
userspace VMM that acts as a vhost-user server the *backend* VMM.
The *frontend* VMM is the one that acts as a vhost-user client, and is
completely unaware that anything special is happening in the backend.

A virtio vhost-user device is a virtio device that acts as a vhost-user
server.  For each vhost-user device it wants its guest to implement,
a VMM listens on a AF_UNIX socket, just like any other userspace
process would.  It then creates a virtio device in this guest, with
type *vhost-user device backend*.  This device has a single virtqueue,
and vhost-user messages are forwarded between that virtqueue and the
AF_UNIX socket.

The backend VMM needs to handle ancillary data specially.
For instance, memory that it mmap's is placed in a PCI BAR that is
accessible to the guest, much like a virtio-GPU blob object.  Vring
kick file descriptors are registered with the kernel as irqfds, so
when the frontend signals them, the backend VM receives an interrupt.
Vring call, error, and log file descriptors are registered with the
kernel as ioeventfds, so that the backend VM can trigger them by
MMIO writes.  It also registers and unregisters irqfds in response
to interrupts being unmasked and masked.

After everything has been set up, the backend VMM is *not* involved in
performance-critical operations.  The vhost-user shared memory regions
are mapped into its address space, so it can access them just like any
other memory.  Writing to an ioeventfd in the frontend VM causes KVM to
trigger an interrupt in the backend VM and visa versa.  I expect that
performance should be almost as good as a vhost-user device implemented
in userspace, but with the full isolation guarantees of a VM.

https://stefanha.github.io/virtio/vhost-user-slave.html has the spec
I am currently using.  It does a much better job explaining things.
-- 
Sincerely,
Demi Marie Obenour (she/her/hers)
OpenPGP_0xB288B55FFF9C22C1.asc (application/pgp-keys, 7 KB)
OpenPGP_signature.asc (application/pgp-signature, 833 B)