Re: [PATCH net 2/2] net: switchdev: don't set port_obj_info->handled true when -EOPNOTSUPP

[Jakub Kicinski]

On Wed, 23 Dec 2020 15:45:33 +0100 Rasmus Villemoes wrote:
> It's not true that switchdev_port_obj_notify() only inspects the
> ->handled field of "struct switchdev_notifier_port_obj_info" if  
> call_switchdev_blocking_notifiers() returns 0 - there's a WARN_ON()
> triggering for a non-zero return combined with ->handled not being
> true. But the real problem here is that -EOPNOTSUPP is not being
> properly handled.
> 
> The wrapper functions switchdev_handle_port_obj_add() et al change a
> return value of -EOPNOTSUPP to 0, and the treatment of ->handled in
> switchdev_port_obj_notify() seems to be designed to change that back
> to -EOPNOTSUPP in case nobody actually acted on the notifier (i.e.,
> everybody returned -EOPNOTSUPP).
> 
> Currently, as soon as some device down the stack passes the check_cb()
> check, ->handled gets set to true, which means that
> switchdev_port_obj_notify() cannot actually ever return -EOPNOTSUPP.
> 
> This, for example, means that the detection of hardware offload
> support in the MRP code is broken - br_mrp_set_ring_role() always ends
> up setting mrp->ring_role_offloaded to 1, despite not a single
> mainline driver implementing any of the SWITCHDEV_OBJ_ID*_MRP. So
> since the MRP code thinks the generation of MRP test frames has been
> offloaded, no such frames are actually put on the wire.
> 
> So, continue to set ->handled true if any callback returns success or
> any error distinct from -EOPNOTSUPP. But if all the callbacks return
> -EOPNOTSUPP, make sure that ->handled stays false, so the logic in
> switchdev_port_obj_notify() can propagate that information.
> 
> Signed-off-by: Rasmus Villemoes <rasmus.villemoes@xxxxxxxxx>

Please make sure you CC the folks who may have something to say about
this - Jiri, Ivan, Ido, Florian, etc.