Re: [PATCH v2 net 4/4] net: dsa: mv88e6xxx: Limit rsvd2cpu policy to user ports on 6393X

[Vladimir Oltean <olteanv@gmail.com>]

On Thu, Dec 19, 2024 at 03:34:43PM +0100, Tobias Waldekranz wrote:
> On tor, dec 19, 2024 at 16:05, Vladimir Oltean <olteanv@gmail.com> wrote:
> > It looks like an oversight in the switchdev tx_fwd_offload scheme. Can't
> > we teach nbp_switchdev_frame_mark_tx_fwd_offload() to make an exception
> > for is_link_local_ether_addr() packets, and not set skb->offload_fwd_mark?
> 
> That sounds like a better option if it is acceptible to the broader
> community. I thought that this might be a quirk of mv88e6xxx's rsvd2cpu
> bits. But if more devices behave in the same way, then it would be
> better to just exempt this whole class from offloading.
> 
> Do you know how any other ASICs behave from this perspective?

The other driver with tx_fwd_offload, sja1105, is going to drop any
packet coming from the host_port which isn't sent through a management
route (set up by sja1105_defer_xmit()). So it's more than likely bugged.

We can't fix this from sja1105_xmit() by reordering sja1105_imprecise_xmit()
and sja1105_defer_xmit(). It's not just the order of operations in the
tagger. It's the fact that the bridge thinks it doesn't need to clone
the skb, and it does.

So yes, it's probably best to exclude link-local from skb->offload_fwd_mark.