On Wed, Jun 07, 2017 at 04:10:02PM -0500, Rob Herring wrote: > On Tue, May 30, 2017 at 02:32:29PM -0700, Florian Fainelli wrote: > > On 05/30/2017 03:44 AM, John Crispin wrote: > > > Extend the DSA binding documentation, adding the new property required > > > when there is more than one CPU port attached to the switch. > > > > > > Cc: Rob Herring <robh+dt@xxxxxxxxxx> > > > Cc: devicetree@xxxxxxxxxxxxxxx > > > Signed-off-by: John Crispin <john@xxxxxxxxxxx> > > > --- > > > Documentation/devicetree/bindings/net/dsa/dsa.txt | 61 ++++++++++++++++++++++- > > > 1 file changed, 60 insertions(+), 1 deletion(-) > > > > > > diff --git a/Documentation/devicetree/bindings/net/dsa/dsa.txt b/Documentation/devicetree/bindings/net/dsa/dsa.txt > > > index cfe8f64eca4f..c164eb38ccc5 100644 > > > --- a/Documentation/devicetree/bindings/net/dsa/dsa.txt > > > +++ b/Documentation/devicetree/bindings/net/dsa/dsa.txt > > > @@ -55,6 +55,11 @@ A user port has the following optional property: > > > - label : Describes the label associated with this port, which > > > will become the netdev name. > > > > > > +- cpu : Option for non "cpu"/"dsa" ports. A phandle to a > > > + "cpu" port, which will be used for passing packets > > > + from this port to the host. If not present, the first > > > + "cpu" port will be used. > > > > So this option essentially allow us to "partition" the switch between > > vectors of ports and their upstream/CPU port. > > Could this be more generic? This is basically saying route all packets > on this port to another port. Hi Rob No, it is not saying that. The CPU port of the switch is special. It is used by the switch for frames it does not know what to do with. e.g, it has not learned the destination MAC address, it is an IGMP management packet, etc. Or the MAC address is that of the CPU, or the CPU needs to bridge it out another interface, e.g. a L2 VPN. The switch will add an additional header indicating what the ingress port was, and pass it to the CPU via this port. There is a presentation Florian, Vivien and I made at netdev 2.1 earlier this year which talks about this. If you want to mirror all packets from one port to another, you can use tc and the mirror action. > > While using Device Tree is an obvious choice for making the initial > > partitioning, it seems like we are missing a configuration mechanism > > whereby we can properly assign ports to a specific upstream CPU port. > > What determines how things are routed/partitioned? If it is purely user > choice then I don't think this should be in DT. > > Let's move the actual discussion into patch 2 in order not to pollute > > the DT maintainers' inbox. We are aiming to load balance traffic to/from the CPU and the switch. The ports of a switch can very in characteristics. Sometimes two are able to do 10Gbps, while others are just 1Gbps. So it would make sense to give those higher speed ports a bigger fraction of the available CPU bandwidth, etc. This binding gives us the option to start with a sensible default for a typical application of the hardware. For something like a WiFi box, this is probably sufficient. However, there is also a lot of usage of DSA in industrial application, and more flexibility is needed. For that we probably need a user API of some sort which allows the defaults in the device tree to be modified to a specific user case. Andrew -- To unsubscribe from this list: send the line "unsubscribe devicetree" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
