Re: [PATCH net-next 3/3] selftests: net: Add port split test

[Ido Schimmel <idosch@idosch.org>]

On Tue, May 19, 2020 at 04:15:41PM +0200, Andrew Lunn wrote:
> > +# Test port split configuration using devlink-port width attribute.
> > +# The test is skipped in case the attribute is not available.
> > +#
> > +# First, check that all the ports with a width of 1 fail to split.
> > +# Second, check that all the ports with a width larger than 1 can be split
> > +# to all valid configurations (e.g., split to 2, split to 4 etc.)
> 
> Hi Ido

Hi Andrew,

> 
> I know very little about splitting ports. So these might be dumb
> questions.
> 
> Is there a well defined meaning of width? Is it something which can be
> found in an 802.3 standard?

It's basically the number of lanes: If a port is a 100Gbps port and has
a width of 4, then every lane is running at 25Gbps. Splitting this port
to 4 (via 'devlink port split') allows you to get 4 ports each capable
of supporting speeds up to 25Gbps (the original netdev disappears).

Example splitter cable:
https://www.mellanox.com/related-docs/prod_cables/PB_MCP7F00-A0xxRyyz_100GbE_QSFP28_to_4x25GbE_4xSFP28_DAC_Splitter.pdf

Some documentation from mlxsw Wiki:
https://github.com/Mellanox/mlxsw/wiki/Switch-Port-Configuration#port-splitting

> Is it well defined that all splits of the for 2, 4, 8 have to be
> supported?

That I don't actually know. It is true for Mellanox and I can only
assume it holds for other vendors. So far beside mlxsw only nfp
implemented port_split() callback. I see it has this check:

```
        if (eth_port.is_split || eth_port.port_lanes % count) {
                ret = -EINVAL;
                goto out;
        }
```

So it seems to be consistent with mlxsw. Jakub will hopefully chime in
and keep me honest.

> Must all 40Gbps ports with a width of 4, be splitable to 2x
> 20Mps? It seems like some hardware might only allow 4x 10G?

Possible. There are many vendor-specific quirks in this area, as I'm
sure you know :)

> 
> If 20Gbps is supported, can you then go recursive and split one of the
> 20G ports into 2x 10G, leaving the other as a 20G port?

Quite certain this is not supported by any vendor.

I assume you're asking because you are trying to see if the test is not
making some vendor-specific assumptions? Personally, I think it's not.
We decided to put it under net/ rather than drivers/net/mlxsw because we
always prefer to write tests that can be shared with others. This is
what actually motivated this work. We had a very Mellanox-specific test
in our regression and we wanted to remove it, but it was not possible to
write such a test without this attribute.