On Tue, May 30, 2017 at 05:16:27PM -0700, Florian Fainelli wrote:
> On 05/30/2017 05:06 PM, Andrew Lunn wrote:
> >> - past the initial setup, if we start creating bridge devices and so on,
> >> we have no way to tell: group Ports 0-3 together and send traffic to CPU
> >> port 0, then let Port 5 alone and send traffic to CPU port 1, that's a
> >> DSA-only problem though, because we still have the CPU port(s) as
> >> independent network interfaces.
> >
> > What is the problem here? Frames come out the master interface, get
> > untagged and passed to the slave interface and go upto the bridge. It
> > should all just work. Same in the reverse direction.
>
> The problem is really that is you have multiple CPU ports, how do you
> define which one gets all the traffic by default? Ascending order of
> port number? Descending order?
I would probably default to round robin when allocating user ports to
CPU ports. That probably gives you the best default.
> I actually tend to think that most use cases our there are in the order
> of dedicating one CPU port to one corresponding switch port (user
> facing, or internal) in order to provided guaranteed bandwidth for that
> port.
Which is generally a waste of bandwidth. Best case, i get 40Mbps
Internet access. Meaning 960Mbps of a dedicated cpu port would be
wasted.
> But as an user, I want to choose how the grouping is going to
> work, and right now, I cannot, unless this is hardcoded in Device Tree,
> which sounds both wrong and inadequate.
So how about round-robin default, and then devlink to move a user port
to a specific cpu port?
We also need to watch out for asymmetry. I think newer marvell chips
don't support egress to multiple CPU ports. Ingress to the switch i
think is unlimited. The older chips are more flexible in this
respect. So we need some degree of flexibility here.
Andrew
