Re: [PATCH net-next 0/6] Disable host flooding for DSA ports under a bridge

[Tobias Waldekranz <tobias@waldekranz.com>]

On Fri, Apr 08, 2022 at 23:03, Vladimir Oltean <vladimir.oltean@nxp.com> wrote:
> For this patch series to make more sense, it should be reviewed from the
> last patch to the first. Changes were made in the order that they were
> just to preserve patch-with-patch functionality.
>
> A little while ago, some DSA switch drivers gained support for
> IFF_UNICAST_FLT, a mechanism through which they are notified of the
> MAC addresses required for local standalone termination.
> A bit longer ago, DSA also gained support for offloading BR_FDB_LOCAL
> bridge FDB entries, which are the MAC addresses required for local
> termination when under a bridge.
>
> So we have come one step closer to removing the CPU from the list of
> destinations for packets with unknown MAC DA.What remains is to check
> whether any software L2 forwarding is enabled, and that is accomplished
> by monitoring the neighbor bridge ports that DSA switches have.
>
> With these changes, DSA drivers that fulfill the requirements for
> dsa_switch_supports_uc_filtering() and dsa_switch_supports_mc_filtering()
> will keep flooding towards the CPU disabled for as long as no port is
> promiscuous. The bridge won't attempt to make its ports promiscuous
> anymore either if said ports are offloaded by switchdev (this series
> changes that behavior). Instead, DSA will fall back by its own will to
> promiscuous mode on bridge ports when the bridge itself becomes
> promiscuous, or a foreign interface is detected under the same bridge.

Hi Vladimir,

Great stuff! I've added Joachim to Cc. He has been working on a series
to add support for configuring the equivalent of BR_FLOOD,
BR_MCAST_FLOOD, and BR_BCAST_FLOOD on the bridge itself. I.e. allowing
the user to specify how local_rcv is managed in br_handle_frame_finish.

For switchdev drivers, being able to query whether a bridge will ingress
unknown unicast to the host or not seems like the missing piece that
makes this bullet proof. I.e. if you have...

- No foreign interfaces
- No promisc
_and_
- No BR_FLOOD on the bridge itself

..._then_ you can safely disable unicast flooding towards the CPU
port. The same would hold for multicast and BR_MCAST_FLOOD of course.

Not sure how close Joachim is to publishing his work. But I just thought
you two should know about the other one's work :)