On May 7, 2016 3:56:34 PM PDT, Philippe Reynes <trem...@gmail.com> wrote:
>On 07/05/16 13:59, Ben Hutchings wrote:
>> On Sat, 2016-05-07 at 01:18 +0200, Philippe Reynes wrote:
>>> The callback {get|set}_link_ksettings are often defined
>>> in a very close way. There are mainly two differences in
>>> those callback:
>>> - the name of the netdev private structure
>>> - the name of the struct phydev in the private structure
>>>
>>> We add two defines ethtool_phy_{get|set}_link_ksettings
>>> to avoid writing severals times almost the same function.
>> [...]
>>
>> I don't think there's no need to access a private structure, as
>there's
>> a phydev pointer in struct net_device. If some drivers don't
>maintain
>> that pointer, they should be changed to do so. Then they can
>> use generic implementations of {get,set}_link_ksettings provided by
>> phylib.
>
>If we could use the phydev in the struct net_device, we could write a
>generic function for {get|set}_link_ksettings. It's a good idea.
>
>But I've quickly looked and a lot of ethernet driver use the private
>structure to store the phydev. If the ethernet driver may use the
>struct net_device for phydev, do you know why so many drivers use
>the private structure ?
The introduction of a phy_device pointer in net_device came much later than the
introduction and use of PHYLIB by most drivers so it is probably just an
oversight.
>
>If everybody agree, I can send a new version with a generic
>{get|set}_link_ksettings
>and a update of fec to use the phydev store in the structure
>net_device.
Yes, that sounds very reasonable. It might be possible to cook a coccinelle
patch which replaces the use of a private phy_device pointer and utilize the
one from net_device.
--
Florian
