Re: [PATCH net-next 5/5] net: dsa: mv88e6xxx: Offload bridge broadcast flooding flag

On Wed, Mar 17, 2021 at 12:14:18PM +0100, Tobias Waldekranz wrote:

Good point. I see two ways forward:

- My first idea was to cache a vector per switch that would act as the
  template when creating a new entry. This avoids having the driver
  layer knowing about stacked netdevs etc. But I think that Andrew is
  generally opposed to caching?

- Add a new helper at the dsa layer that takes a dp and returns the
  netdev that is attached to the bridge, if any:

  struct net_device *dsa_port_to_bridge_port(struct dsa_port *dp)

Any preference or other ideas?

I vote for dsa_port_to_bridge_port. We'll need it anyway for my software
bridging series with sandwiched net devices.

quoted

Speaking of, shouldn't mv88e6xxx_port_vlan_join also be called from
mv88e6xxx_port_bridge_join somehow, or are we waiting for the bridge
facility to replay VLANs added to the LAG when we emit the offload
notification for it?

I do not think so. VLANs are always added via the .port_vlan_add
callback, no?

I got things mixed up in my head while thinking about it.
Yes, of course, the bridge pvid is added via .port_vlan_add as soon as
the port joins the bridge.
What I meant to say is that this sequence of events:

ip link add br0 type bridge
ip link add bond0 type bonding
ip link set bond0 master br0
ip link set lan0 master bond0

will cause lan0 to miss all sorts of information about the bridge port
switchdev objects, including the pvid. My patch series for
SWITCHDEV_BRPORT_OFFLOADED catches this case and asks for a replay.

Potentially the bridge is of the non-filtering variety, so it could be
that no VLANs are ever added.

That will not happen unless you set ds->configure_vlan_while_not_filtering = false,
which mv88e6xxx no longer does, so we're in the clear.

Or do you mean (the most confusingly named feature Marvell LinkStreet
devices) port-based VLANs? Those are setup on a bridge join via
mv88e6xxx_port_vlan_map and mv88e6xxx_pvt_map.

Nope, I wasn't thinking about PVTs.

quoted

quoted

+			/* Skip bridged user ports where broadcast
+			 * flooding is disabled.
+			 */
+			continue;
+
 		err = mv88e6xxx_port_add_broadcast(chip, port, vid);
 		if (err)
 			return err;

@@ -1958,6 +1970,51 @@ static int mv88e6xxx_broadcast_setup(struct mv88e6xxx_chip *chip, u16 vid)
 	return 0;
 }
 
+struct mv88e6xxx_port_broadcast_sync_ctx {
+	int port;
+	bool flood;
+};
+
+static int
+mv88e6xxx_port_broadcast_sync_vlan(struct mv88e6xxx_chip *chip,
+				   const struct mv88e6xxx_vtu_entry *vlan,
+				   void *_ctx)
+{
+	const char broadcast[6] = { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff };

MAC addresses are usually defined as unsigned char[ETH_ALEN]. You can
also use eth_broadcast_addr(broadcast) for initialization.

I was going for uniformity with mv88e6xxx_port_add_broadcast. But I will
add a clean-up commit that fixes the existing code first, and then adds
this definition in the proper way.

Thanks.