Re: [virtio] [PATCH] virtio-net subcommittee proposal

[Cornelia Huck <cohuck@redhat.com>]

On Mon, Apr 03 2023, Parav Pandit <parav@nvidia.com> wrote:

> On 4/3/2023 12:59 PM, Cornelia Huck wrote:
>> On Fri, Mar 31 2023, Parav Pandit <parav@nvidia.com> wrote:
>> 
>>> Modern applications demand low latency and high-performance
>>> networking devices at scale. Modernizing virtio networking devices
>>> involves developing many complex features. While the current methodology
>>> for developing the specification has worked well, there is an
>>> opportunity for better collaboration to support the complex scope of modernization.
>>>
>>> Many hardware vendors and cloud providers have expressed interest in such virtio enhancements.
>>> Hence, we propose to have a virtio net subcommittee.
>>>
>>> Subcommittee mission: Produce backward compatible, low latency, scalable virtio net device spec which works well on hardware NIC and enables efficient use of CPU and memory.
>>>
>>> Top objectives of the subcommittee (statement of purpose):
>>> 1. Chartered to develop the improved virtio-net device
>>> (features, scale, performance) which adapts to ubiquitous PCIe + PCIe virtualization
>> 
>> So this is about virtio-net-pci rather than virtio-net?
>> 
> Mostly yes, but may be it can be done mmio too.
> It is envisioned to cover _real_ pci devices that do DMAs.
> so some pieces may be usable to mmio too in future.
>
>> Is anyone interested in mmio/ccw involved? Would some hypothetical
>> ccw-originating virtio-net feature align with the charter?
>
> to my knowledge, at this point charter likely do not have ccw expertise. 
> It is hard to guess what would that ccw-feature be.
>
> If you see a use cases to modernize it and take advantage like pcie, may 
> be ccw expert should participate too.
>
> Will that ccw-origin feature be used by the ccw?
> Will someone implement it on ccw guest driver side?
>
> If so, we would need a ccw contributor that owns spec writing to actual 
> implementation.

One of my main issues with explicitly talking about PCI(e) is the risk
of introducing something that is not only something that can optionally
improve things for pci, but that might be detrimental to other
transports without the proponents realizing it: if you focus on one
transport, you can easily miss issues that may come up for other
transports. virtio-net is a very fundamental type of device that
everybody needs; we must not make things worse for non-pci
users. (Devices like virtio-scmi that are very platform specific from
the start are a different story.)

Regarding current ccw development: I'm not that much directly involved
with it anymore, maybe Halil can provide some insights.