Re: [RFC] net: towards a generic PON framework

[Gaoyang Wei <yhyxwgy@gmail.com>]

Hi Matheus,

Thanks, that clarifies the test setup.

The numbers are useful as an implementation comparison, although I
think they are not yet enough to attribute the difference specifically
to the kernel/userspace boundary.

In particular, the old SDK-style userspace path and the current
in-kernel implementation differ in more than just where the G.988
agent runs.  A useful follow-up experiment may be to compare the
in-kernel agent against the AN7581/AN7583 userspace implementation
using an ARPHRD_NONE packet interface and AF_PACKET, with the same MIB
size and workload.

That would help separate:

   - cost of userspace/kernel crossings;
   - cost of the transport mechanism;
   - cost of the MIB representation and implementation itself.

Regarding the old SDK netdev, your description also makes me suspect
that the bridge behavior may have come from the way that interface was
constructed as a normal Ethernet-like netdev, rather than from the
net_device abstraction itself.

Andrew has now raised the broader question of whether we need a
separate management net_device at all, so I think that is probably the
more useful interface question to pursue from here.

Thanks,
Gaoyang