Re: [PATCH net-next 5/5 v2] net: dsa: rtl8366rb: Support fast aging

On Tue, Sep 07, 2021 at 07:48:43PM +0200, Linus Walleij wrote:

On Tue, Aug 31, 2021 at 12:46 AM Vladimir Oltean [off-list ref] wrote:

quoted

quoted

+     /* This will age out any L2 entries */

Clarify "any L2 entries". The fdb flushing process should remove the
dynamically learned FDB entries, it should keep the static ones. Did you
say "any" because rtl8366rb does not implement static FDB entries via
.port_fdb_add, and therefore all entries are dynamic, or does it really
delete static FDB entries?

It's what Realtek calls "L2 entries" sadly I do not fully understand
their lingo.

The ASIC can do static L2 entries as well, but I haven't looked into
that. The (confused) docs for the function that set these bits is
the following:

"ASIC will age out L2 entry with fields Static, Auth and IP_MULT are 0.
 Aging out function is for new address learning LUT update because
 size of LUT is limited, old address information should not be kept and
 get more new learning SA information. Age field of L2 LUT is updated
 by following sequence {0b10,0b11,0b01,0b00} which means the LUT

This is {3, 2, 1, 0} in 2-bit Gray code, really curious why they went
for that coding scheme. It is common to designate the states of an FSM
in Gray code because of the single bit change required on state
transitions, but in this case, every ageing state should have a
transition path back to 3 when a packet with that {MAC DA, VLAN ID} is
received on the port, and the L2 entry is refreshed. The Gray code
is a micro-optimization that doesn't seem to help for the primary state
transition there. Anyway, doesn't make a difference.

 entries with age value 0b00 is free for ASIC. ASIC will use this aging
 sequence to decide which entry to be replace by new SA learning
 information. This function can be replace by setting STP state each
 port."

Next it sets the bit for the port in register
RTL8366RB_SECURITY_CTRL.

Realtek talks about "LUT" which I think is the same as "FDB"
(which I assume is forwarding database, I'm not good with this stuff).

LUT is "look-up table", any piece of memory which translates between a
key and a value is a look-up table. In this case, the forwarding database
would qualify as a look-up table where the key is the {MAC DA, VLAN ID}
tuple, and the value is the destination port mask.

My interpretation of this convoluted text is that static, auth and ip_mult
will *not* be affected ("are 0"), but only learned entries in the LUT/FDB
will be affected.

Same interpretation here. This behavior is to be expected.

The sequence listed in the comment I interpret as a reference to what
the ASIC is doing with the age field for the entry internally to
achieve this. Then I guess they say that one can also do fast aging by
stopping the port (duh).

I'll update the doc to say "any learned L2 entries", but eager to hear
what you say about it too :)

Your interpretation seems correct (I can't think of anything else being meant),
but I don't know why you say "duh" about the update of STP state
resulting in the port losing its dynamic L2 entries. Sure, it makes
sense, but many other vendors do not do that automatically, and DSA
calls .port_fast_age whenever the STP state transitions from a value
capable of learning (LEARNING, FORWARDING) to one incapable of learning
(DISABLED, BLOCKING, LISTENING).

To prove/disprove, it would be interesting to implement port STP states,
without implementing .port_fast_age, force a port down and then back up,
and then run "bridge fdb" and see whether it is true that STP state
changes also lead to FDB flushing but at a hardware level (whether there
is any 'self' entry being reported).

By this logic, the hardware should also auto-age its dynamically learned
L2 entries when address learning is turned off, and there would be no
purpose at all in implementing the .port_fast_age method. Also curious
if that is the case.