[Qemu-devel] [2.5 issue] virtio-1 in virtio-net and old vhost

[Cornelia Huck]

Some of our test folks tried to run a recent-ish qemu (nearly 2.5)
combined with an old host kernel (and a virtio-1 capable guest).

In that setup, we had the transport (in that case, virtio-ccw)
advertise VERSION_1 as it is a revision 1 device. However, the old
vhost driver did not support virtio-1 and therefore cleared the
VERSION_1 bit. In the end, qemu did not offer VERSION_1 to the guest
for a revision 1 device, which the guest treats as a fatal error.

It looks to me as if virtio-pci has the same problem: The kernel will
detect a modern device as by the I/O layout and then barf at the
missing VERSION_1 feature bit.

We _could_ make this missing VERSION_1 bit non-fatal in the guest, but
that does not fix guests that are already out there.

The problem is that the transport cannot know whether the VERSION_1 bit
will be pulled from under it later during device setup: This is only
done in the ->get_features() callback when virtio-net will handle the
features supported by vhost.

I'm currently lacking a good idea on how to fix this, but I think that
is an issue that should be dealt with for 2.5...