linux-imx

net: mscc: ocelot: Workaround to allow traffic to CPU in standalone mode

The Ocelot switches have what is, in my opinion, a design flaw: their
DSA header is in front of the Ethernet header, which means that they
subvert the DSA master's RX filter, which for all practical purposes,
either needs to be in promiscuous mode, or the OCELOT_TAG_PREFIX_LONG
needs to be used for extraction, which makes the switch add a fake DMAC
of ff:ff:ff:ff:ff:ff so that the DSA master accepts the frame.

The issue with this design, of course, is that the CPU will be spammed
with frames that it doesn't want to respond to, and there isn't any
hardware offload in place by default to drop them.

What is being done in the VSC7514 Ocelot driver is a process of
selective whitelisting. The "MAC address" of each Ocelot switch net
device, with all VLANs installed on that port, is being added as a FDB
entry towards PGID_CPU.

Some background first: Port Group IDs (PGIDs) are masks of destination
ports. The switch performs 3 lookups in the PGID table for each frame,
and forwards the frame to the ports that are present in the logical AND
of all 3 PGIDs (for the most part, see below).

The first PGID lookup is for the destination masks and the PGID table is
indexed by the DEST_IDX field from the MAC table (FDB).
The PGID can be an unicast set: PGIDs 0-11 are the per-port PGIDs, and
by convention PGID i has only BIT(i) set, aka only this port is set in
the destination mask.
Or the PGID can be a multicast set: PGIDs 12-63 can (again, still by
convention) hold a richer destination mask comprised of multiple ports.

[ Ignoring the second PGID lookup, for aggregation, since it doesn't
  interfere. ]

The third PGID lookup is for source masks: PGID entries 80-91 answer the
question: is port i allowed to forward traffic to port j? If yes, then
BIT(j) of PGID 80+i will be found set.

What is interesting about the CPU port in this whole story is that, in
the way the driver sets up the PGIDs, its bit isn't set in any source
mask PGID of any other port (therefore, the third lookup would always
decide to exclude the CPU port from this list). So frames are never
_forwarded_ to the CPU.

There is a loophole in this PGID mechanism which is described in the
VSC7514 manual:

	If an entry is found in the MAC table entry of ENTRY_TYPE 0 or 1
	and the CPU port is set in the PGID pointed to by the MAC table
	entry, CPU extraction queue PGID.DST_PGID is added to the CPUQ.

In other words, the CPU port is special, and frames are "copied" to the
CPU, disregarding the source masks (third PGID lookup), if BIT(cpu) is
found to be set in the destination masks (first PGID lookup).

Now back to the story: what is PGID_CPU? It is a multicast set
containing only BIT(cpu). I don't know why it was chosen to be a
multicast PGID (59) and not simply the unicast one of this port, but it
doesn't matter.

The point is that frames that match the FDB will go to PGID_CPU by
virtue of the DEST_IDX from the respective MAC table entry, and frames
that don't will go to PGID_UC or PGID_MC, by virtue of the FLD_UNICAST,
FLD_BROADCAST etc settings for flooding. And that is where the
distinction is made: flooded frames will be subject to the third PGID
lookup, while frames that are whitelisted to the PGID_CPU by the MAC
table aren't.

So we can use this mechanism to simulate an RX filter, given that we are
subverting the DSA master's implicit one, as mentioned in the first
paragraph. But this has some limitations:

- In Ocelot each net device has its own MAC address. When simulating
  this with MAC table entries, it will practically result in having N
  MAC addresses for each of the N front-panel ports (because FDB entries
  are not per source port). A bit strange, I think.

- In DSA we don't have the infrastructure in place to support this
  whitelisting mechanism. Calling .port_fdb_add on the CPU port for each
  slave net device dev_addr isn't, in itself, hard. The problem is with
  the VLANs that this port is part of. We would need to keep a duplicate
  list of the VLANs from the bridge, plus the ones added from 8021q, for
  each port. And we would need reference counting on each MAC address,
  such that when a front-panel port changes its MAC address and we need
  to delete the old FDB entry, we don't actually delete it if the other
  front-panel ports are still using it. Not to mention that this FDB
  entry would have to be added on the whole net of upstream DSA switches.

So... it's complicated. What this patch does is to simply allow frames
to be flooded to the CPU, which is anyway what the Ocelot driver is
doing after removing the bridge from the net devices, see this snippet
from ocelot_bridge_stp_state_set:

	/* Apply FWD mask. The loop is needed to add/remove the current port as
	 * a source for the other ports.
	 */
	for (p = 0; p < ocelot->num_phys_ports; p++) {
		if (p == ocelot->cpu || (ocelot->bridge_fwd_mask & BIT(p))) {
			(...)
		} else {
			/* Only the CPU port, this is compatible with link
			 * aggregation.
			 */
			ocelot_write_rix(ocelot,
					 BIT(ocelot->cpu),
					 ANA_PGID_PGID, PGID_SRC + p);
		}

Otherwise said, the ocelot driver itself is already not self-coherent,
since immediately after probe time, and immediately after removal from a
bridge, it behaves in different ways, although the front panel ports are
standalone in both cases.

While standalone traffic _does_ work for the Felix DSA wrapper after
enslaving and removing the ports from a bridge, this patch makes
standalone traffic work at probe time too, with the caveat that even
irrelevant frames will get processed by software, making it more
susceptible to denial of service.

Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>