On Mon, Aug 29, 2022 at 06:13:14PM +0200, netdev@xxxxxxxxxxxxxxxxxxxx wrote: > On 2022-08-29 18:03, Ido Schimmel wrote: > > On Mon, Aug 29, 2022 at 05:08:23PM +0200, netdev@xxxxxxxxxxxxxxxxxxxx > > wrote: > > > On 2022-08-29 16:37, Ido Schimmel wrote: > > > > On Mon, Aug 29, 2022 at 02:04:42PM +0200, netdev@xxxxxxxxxxxxxxxxxxxx > > > > wrote: > > > > > On 2022-08-29 13:32, Ido Schimmel wrote: > > > > > Port association is needed for MAB to work at all on mv88e6xxx, but > > > > > for > > > > > 802.1X port association is only needed for dynamic ATU entries. > > > > > > > > Ageing of dynamic entries in the bridge requires learning to be on as > > > > well, but in these test cases you are only using static entries and > > > > there is no reason to enable learning in the bridge for that. I prefer > > > > not to leak this mv88e6xxx implementation detail to user space and > > > > instead have the driver enable port association based on whether > > > > "learning" or "mab" is on. > > > > > > > > > > Then it makes most sense to have the mv88e6xxx driver enable port > > > association when then port is locked, as it does now. > > > > As you wish, but like you wrote "802.1X port association is only needed > > for dynamic ATU entries" and in this case user space needs to enable > > learning (for refresh only) so you can really key off learning on > > "learning || mab". User space can decide to lock the port and work with > > static entries and then learning is not required. > > I will of course remove all "learning on" in the selftests, which is what I > think you are referring to. In the previous I am referring to the code in > the driver itself which I understand shall turn on port association with > locked ports, e.g. no need for "learning on" when using the feature in > general outside selftests... "learning on" is needed when dynamic FDB entries are used to authorize hosts. Without learning being enabled, the bridge driver (or the underlying hardware) will not refresh the entries during forwarding and they will age out, resulting in packet loss until the hosts are re-authorized. Given the current test cases only use static entries, there is no need to enable learning on locked ports. This will change when test cases are added with dynamic entries. Regarding mv88e6xxx, my understanding is that you also need learning enabled for MAB (I assume for the violation interrupts). Therefore, for mv88e6xxx, learning can be enabled if learning is on or MAB is on. Enabling it based on whether the port is locked or not seems inaccurate.
