net: dsa: mv88e6xxx: keep the pvid at 0 when VLAN-unaware

The VLAN support in mv88e6xxx has a loaded history. Commit 2ea7a679ca2a
("net: dsa: Don't add vlans when vlan filtering is disabled") noticed
some issues with VLAN and decided the best way to deal with them was to
make the DSA core ignore VLANs added by the bridge while VLAN awareness
is turned off. Those issues were never explained, just presented as
"at least one corner case".

That approach had problems of its own, presented by
commit 54a0ed0df496 ("net: dsa: provide an option for drivers to always
receive bridge VLANs") for the DSA core, followed by
commit 1fb74191988f ("net: dsa: mv88e6xxx: fix vlan setup") which
applied ds->configure_vlan_while_not_filtering = true for mv88e6xxx in
particular.

We still don't know what corner case Andrew saw when he wrote
commit 2ea7a679ca2a ("net: dsa: Don't add vlans when vlan filtering is
disabled"), but Tobias now reports that when we use TX forwarding
offload, pinging an external station from the bridge device is broken if
the front-facing DSA user port has flooding turned off. The full
description is in the link below, but for short, when a mv88e6xxx port
is under a VLAN-unaware bridge, it inherits that bridge's pvid.
So packets ingressing a user port will be classified to e.g. VID 1
(assuming that value for the bridge_default_pvid), whereas when
tag_dsa.c xmits towards a user port, it always sends packets using a VID
of 0 if that port is standalone or under a VLAN-unaware bridge - or at
least it did so prior to commit d82f8ab0d874 ("net: dsa: tag_dsa:
offload the bridge forwarding process").

In any case, when there is a conversation between the CPU and a station
connected to a user port, the station's MAC address is learned in VID 1
but the CPU tries to transmit through VID 0. The packets reach the
intended station, but via flooding and not by virtue of matching the
existing ATU entry.

DSA has established (and enforced in other drivers: sja1105, felix,
mt7530) that a VLAN-unaware port should use a private pvid, and not
inherit the one from the bridge. The bridge's pvid should only be
inherited when that bridge is VLAN-aware, so all state transitions need
to be handled. On the other hand, all bridge VLANs should sit in the VTU
starting with the moment when the bridge offloads them via switchdev,
they are just not used.

This solves the problem that Tobias sees because packets ingressing on
VLAN-unaware user ports now get classified to VID 0, which is also the
VID used by tag_dsa.c on xmit.

Fixes: d82f8ab0d874 ("net: dsa: tag_dsa: offload the bridge forwarding process")
Link: https://patchwork.kernel.org/project/netdevbpf/patch/20211003222312.284175-2-vladimir.oltean@nxp.com/#24491503
Reported-by: Tobias Waldekranz <tobias@waldekranz.com>
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

1 files changed, 21 insertions, 0 deletions

diff --git a/drivers/net/dsa/mv88e6xxx/port.c b/drivers/net/dsa/mv88e6xxx/port.c
index 451028c57af8..d9817b20ea64 100644
--- a/drivers/net/dsa/mv88e6xxx/port.c
+++ b/drivers/net/dsa/mv88e6xxx/port.c
@@ -1257,6 +1257,27 @@ int mv88e6xxx_port_set_8021q_mode(struct mv88e6xxx_chip *chip, int port,
 	return 0;
 }
 
+int mv88e6xxx_port_drop_untagged(struct mv88e6xxx_chip *chip, int port,
+				 bool drop_untagged)
+{
+	u16 old, new;
+	int err;
+
+	err = mv88e6xxx_port_read(chip, port, MV88E6XXX_PORT_CTL2, &old);
+	if (err)
+		return err;
+
+	if (drop_untagged)
+		new = old | MV88E6XXX_PORT_CTL2_DISCARD_UNTAGGED;
+	else
+		new = old & ~MV88E6XXX_PORT_CTL2_DISCARD_UNTAGGED;
+
+	if (new == old)
+		return 0;
+
+	return mv88e6xxx_port_write(chip, port, MV88E6XXX_PORT_CTL2, new);
+}
+
 int mv88e6xxx_port_set_map_da(struct mv88e6xxx_chip *chip, int port)
 {
 	u16 reg;