Linux Virtualization - Apr 2023 - [PATCH net] virtio-net: reject small vring sizes

On Mon, Apr 17, 2023 at 11:51:22AM +0000, Alvaro Karsz
wrote:
> > > I see your point.
> > > Regardless, we'll need to fail probe in some cases.
> > > ring size of 1 for example (if I'm not mistaken)
> > 
> > Hmm. We can make it work if we increase hard header size, then
> > there will always be room for vnet header.
> > 
> > > control vq even needs a bigger ring.
> > 
> > Why does it?
> 
> At the moment, most of the commands chain 3 descriptors:
> 1 - class + command
> 2 - command specific
> 3 - ack
> 
> We could merge 1 and 2 into a single one, both are read only for the
device, so I take it back, it won't need a bigger ring.
> But it will need 2 descriptors at least(1 read only for the device and 1
write only for the device), so we still need to fail probe sometimes.
> 
Yes that makes sense, it's architetural. We can disable ctrl vq though.

-- 
MST

> Yes that makes sense, it's architetural. We can disable ctrl vq though.
The problem here is that we know the vring size after calling virtnet_find_vqs,
so the number of VQs already includes the control VQ.

Actually, many variables/settings that are initialized before we call
virtnet_find_vqs may need modifications if we use small vrings.
For example has_rss_hash_report, has_rss, hdr_len etc..

We could have a fixup function to fix everything after we discover that we are
using small vrings, but, honestly, I think that this will be hard to maintain in
the future, and I don't like this approach much.

The ideal thing will be to discover if we use small vrings in probe's
beginning.

I'm looking for a way at the moment.