Hi Florian, On Sun, Oct 01, 2017 at 12:46:34PM -0700, Florian Fainelli wrote: > Hi all, > > This patch series is sent as RFC since I have only been able to test LAG > with dsa-loop and not with real HW yet (that should be tomorrow). I also > looked at how the Marvell DSDT API is defined for adding ports to "trunk" > groups and the API proposed here should work there too. Can't speak about > QCA, Mediatek or KSZ switches though.
Thanks for working on this. I've yet to look at the patches, but I thought I'll mention a few issues we bumped into with LAG devices: 1) It is possible for users to stack devices on top of the LAG and only then enslave your port. This means that the underlying driver might not be aware of all the necessary configuration. It's quite a complicated problem to solve properly, so we currently forbid enslavements to devices that already have uppers. There's also an issue with IP addresses and routes configured on top of the LAG, but I hope to fix that soon. I don't think you support L3 in DSA yet, so it shouldn't be a problem for you. 2) Similarly, you're no longer guaranteed to have the bridge do proper clean up in case you pull a port out of a bridged LAG, so you'll need to handle that. Any context you store for the bridge port needs to be destroyed upon the removal of the last port from the LAG. > Few open questions that may need solving now or later: > > - on Broadcom switches, we should allow enslaving a port as a LAG group > member if its speed does not match that of the other members of the group > > - not sure what to do with a switch fabric, naively, if adding two ports > of two distinct switches as a LAG group, we may have to propagate that > to "dsa" cross-chip interfaces as well At least in mlxsw case, enslaving switch and non-switch ports to the same LAG doesn't make sense. Any traffic routed by the switch will only be load-balanced between the switch ports. One way to solve that is to forbid such enslavements during NETDEV_PRECHANGEUPPER in case the lower devices in the adjacency list of the LAG don't belong to the same switch. Note that such configurations are bound to fail anyway, as the non-switch ports will not have `switchdev_ops` configured and thus fail during __switchdev_port_obj_add() / __switchdev_port_attr_set().