Re: [RFC PATCH] virtio_net: Extend virtio to use VF datapath when available

On Thu, 21 Dec 2017 02:15:31 +0200, Michael S. Tsirkin wrote:

On Wed, Dec 20, 2017 at 02:33:34PM -0800, Jakub Kicinski wrote:

quoted

On Mon, 18 Dec 2017 16:40:36 -0800, Sridhar Samudrala wrote:  

quoted

+static int virtio_netdev_event(struct notifier_block *this,
+			       unsigned long event, void *ptr)
+{
+	struct net_device *event_dev = netdev_notifier_info_to_dev(ptr);
+
+	/* Skip our own events */
+	if (event_dev->netdev_ops == &virtnet_netdev)
+		return NOTIFY_DONE;  

I wonder how does this work WRT loop prevention.  What if I have two
virtio devices with the same MAC, what is preventing them from claiming
each other?  Is it only the check above (not sure what is "own" in
comment referring to)?  

I expect we will add a feature bit (VIRTIO_NET_F_MASTER) and it will
be host's responsibility not to add more than 1 such device.

Do you mean a virtio_net feature bit?  That won't stop the loop with
a hyperv device.  Unless we want every paravirt device to have such
bit and enforce there can be only one magic bond enslavement active in
the system.

Making the bonding information separate from the slaves seems so much
cleaner...  No limitation on device types, or how many bonds user
chooses to create.  No MAC matching...

quoted

I'm worried the check above will not stop virtio from enslaving hyperv
interfaces and vice versa, potentially leading to a loop, no?  There is
also the fact that it would be preferable to share the code between
paravirt drivers, to avoid duplicated bugs.

My suggestion during the previous discussion was to create a paravirt
bond device, which will explicitly tell the OS which interfaces to
bond, regardless of which driver they're using.  Could be some form of
ACPI/FW driver too, I don't know enough about x86 FW to suggest
something fully fleshed :(