Hi Andrew,
Andrew Lunn <and...@lunn.ch> writes:
> This is way too simplistic.
>
> The switchdev call is per port. This manipulates the whole switch. We
> need to somehow handle the difference.
>
> I've not look at the bridge code, but i assume it initially sets each
> port to a long age time, probably 5 minutes. When there is a topology
> change, it enables fast ageing by setting a shorter age time in each
> port. After a while it will return to the default age time. Although
> the switchdev call is per port, i think the age time is a property of
> the bridge, not a port.
>
> For the Marvell devices, we only have a global setting. It will apply
> to all bridges we create on the switch. So if one bridge requests fast
> ageing, we need to apply it to all bridges. We should only go back to
> slow ageing when all bridges are out of fast ageing. That is, we need
> some sort of reference counting.
>
> I'm not sure we have enough information to know why the bridge changed
> the age timing. Did the use change the forwarding delay, or have we
> entered fast ageing? So i think for Marvell devices, we need an
> additional property passed down. Is this a fast or a slow age time?
> We can then determine what is the fastest fast ageing, and the fastest
> slow ageing is, perform reference counting as appropriate, and set the
> global setting as needed.
Nope, the bridge ageing time is not per-port, even though switchdev ops
are per-port by design. This is a switch-wide attribute.
See f55ac58ae64c ("switchdev: add bridge ageing_time attribute") [1]
Rocker and mlxsw implement AGEING_TIME switch-wide too.
[1]
https://git.kernel.org/cgit/linux/kernel/git/davem/net.git/commit/?id=f55ac58ae64c
Thanks,
Vivien
