Linux Ethernet Bridging - Feb 2023 - [Bridge] [PATCH net-next 0/5] ATU and FDB synchronization on locked ports

On Mon, Jan 30, 2023 at 06:34:24PM +0100, Hans J. Schultz
wrote:
> This patch set makes it possible to have synchronized dynamic ATU and FDB
> entries on locked ports. As locked ports are not able to automatically
> learn, they depend on userspace added entries, where userspace can add
> static or dynamic entries. The lifetime of static entries are completely
> dependent on userspace intervention, and thus not of interest here. We
> are only concerned with dynamic entries, which can be added with a
> command like:
> 
> bridge fdb replace ADDR dev <DEV> master dynamic
> 
> We choose only to support this feature on locked ports, as it involves
> utilizing the CPU to handle ATU related switchcore events (typically
> interrupts) and thus can result in significant performance loss if
> exposed to heavy traffic.
Not sure I understand this reasoning. I was under the impression that
hostapd is installing dynamic entries instead of static ones since the
latter are not flushed when carrier is lost. Therefore, with static
entries it is possible to unplug a host (potentially plugging a
different one) and not lose authentication.
> 
> On locked ports it is important for userspace to know when an authorized
> station has become silent, hence not breaking the communication of a
> station that has been authorized based on the MAC-Authentication Bypass
> (MAB) scheme. Thus if the station keeps being active after authorization,
> it will continue to have an open port as long as it is active. Only after
> a silent period will it have to be reauthorized. As the ageing process in
> the ATU is dependent on incoming traffic to the switchcore port, it is
> necessary for the ATU to signal that an entry has aged out, so that the
> FDB can be updated at the correct time.
Why mention MAB at all? Don't you want user space to always use dynamic
entries to authenticate hosts regardless of 802.1X/MAB?
> 
> This patch set includes a solution for the Marvell mv88e6xxx driver, where
> for this driver we use the Hold-At-One feature so that an age-out
> violation interrupt occurs when a station has been silent for the
> system-set age time. The age out violation interrupt allows the switchcore
> driver to remove both the ATU and the FDB entry at the same time.
> 
> It is up to the maintainers of other switchcore drivers to implement the
> feature for their specific driver.
> 
> Hans J. Schultz (5):
>   net: bridge: add dynamic flag to switchdev notifier
>   net: dsa: propagate flags down towards drivers
>   drivers: net: dsa: add fdb entry flags incoming to switchcore drivers
>   net: bridge: ensure FDB offloaded flag is handled as needed
>   net: dsa: mv88e6xxx: implementation of dynamic ATU entries
Will try to review tomorrow, but it looks like this set is missing
selftests. What about extending bridge_locked_port.sh?

On 2023-01-31 20:25, Ido Schimmel wrote:
> 
> Will try to review tomorrow, but it looks like this set is missing
> selftests. What about extending bridge_locked_port.sh?
I knew you would take this up. :-)
But I am not sure that it's so easy to have selftests here as it is 
timing based and it would take the 5+ minutes just waiting to test in 
the stadard case, and there is opnly support for mv88e6xxx driver with 
this patch set.

On 2023-01-31 20:25, Ido Schimmel wrote:
>> command like:
>> 
>> bridge fdb replace ADDR dev <DEV> master dynamic
>> 
>> We choose only to support this feature on locked ports, as it involves
>> utilizing the CPU to handle ATU related switchcore events (typically
>> interrupts) and thus can result in significant performance loss if
>> exposed to heavy traffic.
> 
> Not sure I understand this reasoning. I was under the impression that
> hostapd is installing dynamic entries instead of static ones since the
> latter are not flushed when carrier is lost. Therefore, with static
> entries it is possible to unplug a host (potentially plugging a
> different one) and not lose authentication.
> 
Both auth schemes 802.1X and MAB install dynamic entries as you point 
out, and both use locked ports.
In the case of non locked ports, they just learn normally and age and 
refresh their entries, so the use case of a userspace added dynamic FDB 
entry is hard for me to see. And having userspace being notified of an 
ordinary event that a FDB entry has been aged out could maybe be used, 
but for the reasons mentioned it is not supported here.
>> 
>> On locked ports it is important for userspace to know when an 
>> authorized
>> station has become silent, hence not breaking the communication of a
>> station that has been authorized based on the MAC-Authentication 
>> Bypass
>> (MAB) scheme. Thus if the station keeps being active after 
>> authorization,
>> it will continue to have an open port as long as it is active. Only 
>> after
>> a silent period will it have to be reauthorized. As the ageing process 
>> in
>> the ATU is dependent on incoming traffic to the switchcore port, it is
>> necessary for the ATU to signal that an entry has aged out, so that 
>> the
>> FDB can be updated at the correct time.
> 
> Why mention MAB at all? Don't you want user space to always use dynamic
> entries to authenticate hosts regardless of 802.1X/MAB?
Yes, you are right about that. I guess it came about as this was 
developed much in the same time and with the code of MAB.