Re: [PATCH net-next v2 2/4] net: dsa: mt7530: use independent VLAN learning on VLAN-unaware bridges

On Wed, Aug 04, 2021 at 12:04:02AM +0800, DENG Qingfang wrote:

Consider the following bridge configuration, where bond0 is not
offloaded:

         +-- br0 --+
        / /   |     \
       / /    |      \
      /  |    |     bond0
     /   |    |     /   \
   swp0 swp1 swp2 swp3 swp4
     .        .       .
     .        .       .
     A        B       C

Ideally, when the switch receives a packet from swp3 or swp4, it should
forward the packet to the CPU, according to the port matrix and unknown
unicast flood settings.

But packet loss will happen if the destination address is at one of the
offloaded ports (swp0~2). For example, when client C sends a packet to
A, the FDB lookup will indicate that it should be forwarded to swp0, but
the port matrix of swp3 and swp4 is configured to only allow the CPU to
be its destination, so it is dropped.

However, this issue does not happen if the bridge is VLAN-aware. That is
because VLAN-aware bridges use independent VLAN learning, i.e. use VID
for FDB lookup, on offloaded ports. As swp3 and swp4 are not offloaded,
shared VLAN learning with default filter ID of 0 is used instead. So the
lookup for A with filter ID 0 never hits and the packet can be forwarded
to the CPU.

In the current code, only two combinations were used to toggle user
ports' VLAN awareness: one is PCR.PORT_VLAN set to port matrix mode with
PVC.VLAN_ATTR set to transparent port, the other is PCR.PORT_VLAN set to
security mode with PVC.VLAN_ATTR set to user port.

It turns out that only PVC.VLAN_ATTR contributes to VLAN awareness, and
port matrix mode just skips the VLAN table lookup. The reference manual
is somehow misleading when describing PORT_VLAN modes. It states that
PORT_MEM (VLAN port member) is used for destination if the VLAN table
lookup hits, but actually **PORT_MEM & PORT_MATRIX** (bitwise AND of
VLAN port member and port matrix) is used instead, which means we can
have two or more separate VLAN-aware bridges with the same PVID and
traffic won't leak between them.

Therefore, to solve this, enable independent VLAN learning with PVID 0
on VLAN-unaware bridges, by setting their PCR.PORT_VLAN to fallback
mode, while leaving standalone ports in port matrix mode. The CPU port
is always set to fallback mode to serve those bridges.

During testing, it is found that FDB lookup with filter ID of 0 will
also hit entries with VID 0 even with independent VLAN learning. To
avoid that, install all VLANs with filter ID of 1.

Signed-off-by: DENG Qingfang <dqfext@gmail.com>
---
v1 -> v2: use FID enum instead of hardcoding.

Reviewed-by: Vladimir Oltean <olteanv@gmail.com>

After this patch set gets merged, can you also please take a look at the
following:

Documentation/networking/switchdev.rst says:

When the bridge has VLAN filtering enabled and a PVID is not configured on the
ingress port, untagged and 802.1p tagged packets must be dropped. When the bridge
has VLAN filtering enabled and a PVID exists on the ingress port, untagged and
priority-tagged packets must be accepted and forwarded according to the
bridge's port membership of the PVID VLAN. When the bridge has VLAN filtering
disabled, the presence/lack of a PVID should not influence the packet
forwarding decision.

I'm not sure if this happens or not with mt7530, since the driver
attempts to change the pvid back to 0. You are not changing this
behavior in this series, so no reason to deal with it as part of it.

 
-	mt7530_rmw(priv, MT7530_PPBV1_P(port), G0_PORT_VID_MASK, pvid);
-	priv->ports[port].pvid = pvid;
 
 	mutex_unlock(&priv->reg_mutex);