Re: [PATCH net-next v2] net: dsa: make "label" property optional for dsa2

From: Jiri Pirko
Date: Wed Jan 11 2017 - 02:26:20 EST


Tue, Jan 10, 2017 at 06:58:18PM CET, f.fainelli@xxxxxxxxx wrote:
>On 01/10/2017 01:55 AM, Jiri Pirko wrote:
>> Mon, Jan 09, 2017 at 07:06:39PM CET, f.fainelli@xxxxxxxxx wrote:
>>> On 01/09/2017 09:58 AM, Jiri Pirko wrote:
>>>> Mon, Jan 09, 2017 at 06:42:07PM CET, f.fainelli@xxxxxxxxx wrote:
>>>>> On 01/09/2017 08:06 AM, Jiri Pirko wrote:
>>>>>> Mon, Jan 09, 2017 at 04:45:33PM CET, vivien.didelot@xxxxxxxxxxxxxxxxxxxx wrote:
>>>>>>> Hi Jiri,
>>>>>>>
>>>>>>> Jiri Pirko <jiri@xxxxxxxxxxx> writes:
>>>>>>>
>>>>>>>>> Extra question: shouldn't phys_port_{id,name} be switchdev attributes in
>>>>>>>>
>>>>>>>> Again, phys_port_id has nothing to do with switches. Should be removed
>>>>>>>> from dsa because its use there is incorrect.
>>>>>>>
>>>>>>> Florian, since 3a543ef just got in, can it be reverted?
>>>>>>
>>>>>> Yes, please revert it. It is only in net-next.
>>>>>
>>>>> Maybe the use case can be understood before reverting the change. How do
>>>>> we actually the physical port number of an Ethernet switch per-port
>>>>> network device? The name is not enough, because there are plenty of
>>>>> cases where we need to manipulate a physical port number (be it just for
>>>>> informational purposes).
>>>>
>>>> Like what?
>>>
>>> Specifying the physical port number (and derive a queue number
>>> eventually) for some ethtool (e.g: rxnfc)/tc (queue mapping) operations
>>> where there is an action/queue/port destination argument that gets
>>> programmed into the hardware.
>>
>> Could you point me to a real example? User command?
>
>ethtool --config-nfc moca flow-type udp4 src-ip 192.168.1.20 dst-ip \
> 192.168.1.10 src-port 49884 dst-port 5001 action 2
>
>Where 2 here designates a port number, users need to be able to look up
>the physical port number corresponding to an interface to know which
>value to put in this command.

2 is not a port number but RX queue number. I believe you need to
ditinguish that and port_name. Not sure how are they related.



>
>Yes I know we can do the same thing with cls_flower, possibly by
>referencing network devices directly.

Yes, that is what you should do. I believe that using config-nfc is not
correct for this use-case.


>
>>
>>
>>>
>>> You already have the originating port number from the interface you call
>>> the method against, but you also need the destination port number since
>>> that is what the HW understands.
>>
>> This is internal to kernel? I fail to understand what you mean exactly.
>
>See the command above, from using the "moca" netdev here, we can access
>the DSA private network device (dsa_slave_priv) structure and get the
>port number from there, and pass this down to the switch driver. The
>switch driver also takes another port number (and eventually a queue
>number) to program classification filters.
>
>>
>>
>>>
>>> Aside from that, it is useful for allowing interface naming in user
>>> space if you don't want to use labels.
>>>
>>>>
>>>> Why the name is not enough? This is something propagated to userspace
>>>> and never used internally in kernel.
>>>
>>> Because the name is not reflective of the port number in some switches.
>>> In my case for instance, we have 5 ports that are named after the
>>> entities they connect to (an integrated Gigabit PHY, two RGMII pads, one
>>> MoCA interface, and the CPU)
>>>
>>
>> Again, I'm missing why you need a portnumber as a Integer to userspace.
>> From driver, you can expose phys_port_name:
>
>If we are exposing the port name here, we may as well expose the DSA
>"label" instead of the physical port number number?

Yeah, that makes sense.


>
>I don't deny my change may be misusing what phys_port_id was originally
>designed for, but providing "p0" instead of "0" to user-space, what
>value is there in adding the "p" in front really?

It's up to a driver. He knows how the front panel names look like.


>--
>Florian