On 03/21/2017 02:04 AM, Niklas Cassel wrote:
> On 03/20/2017 11:07 PM, Florian Fainelli wrote:
>>
>> (snip)
>>>
>>> However, it is kind of sad that drivers are so inconsistent of what goes
>>> in probe and what goes in ndo_open...which is tied together with the
>>> whole mess of when certain ethtool commands work or do not work.
>> Well, inconsistent here is kind of big statement, what I meant to say is
>> that your proposed change actually makes thing inconsistent.
>
> What I mean is that it is about 50/50 if something e.g. phy_connect
> is called from probe or from ndo_open.
> This is what I think is inconsistent.
>
> git grep -p -E "phy_connect\(|phy_connect_direct\(" drivers/net/ethernet/ |
> grep open | wc -l
> 12
>
> git grep -p -E "phy_connect\(|phy_connect_direct\(" drivers/net/ethernet/ |
> grep probe | wc -l
> 27
> even if we exclude *_mii_probe/*_mdio_probe which are done
> git grep -p mii_probe drivers/net/ethernet/ | grep open | wc -l
> 6
> git grep -p mdio_probe drivers/net/ethernet/ | grep open | wc -l
> 2
>
> phy_connect done from probe:
> 27-6-2=19
>
> phy_connect done from ndo_open:
> 12+6+2=20
>
>
>>
>>> Do you know of a good way to avoid the -EBUSY in
>>> stmmac_ethtool_get_link_ksettings,
>>> but still keep phy_connect in ndo_open?
>> Let me rephrase this: doing an ethtool operation on an interface that is
>> not open is not a defined behavior which adheres to a contract with the
>> kernel. ethtool operations on interfaces that are DOWN, may succeed if
>> they do not involve a piece of hardware that needs to be active (e.g:
>> the PHY) but there is no guarantee that a) the change is immediately
>> applied, or b) that the results are consistent.
>>
>> The best thing to do IMHO is just silence the warning, return an error
>> coed, and come back configuring the interface at a later time when it is
>> guaranteed to be operational.
>
> Thank you for your feedback. I agree with you,
> silencing the warning would be the best way forward.
> David, please ignore this patch.
>
>
>
> However, since ethtool is a user interface, I think that it would
> have been nice if it wasn't so driver specific regarding to if a
> command works before the interface is open or not.
On premise I agree, in practice, every HW and driver is different, and
enforcing standard behaviors is even more difficult as there are
multiple pieces of HW interconnected to each other: MAC, MDIO bus, PHY
etc. all with different power management capabilities and properties,
and different requirements, and of course, varying degrees of brokenness.
>
> If the user does some ethtool operation that affects e.g. the PHY
> before the interface is open, perhaps there is a way to save this
> setting so it could be applied when the resource is available.
Well behaving driver will connect to the PHY as one of their final
driver's ndo_open() step, any call in between will result in phydev
being NULL which will return a proper error code and should be acted
upon by the application doing the ethtool calls (or the one doing
ioctl() calls similar to ethtool).
>
> For the PHY case, net_device has a pointer to a phy_device,
> register_netdev could create a dummy phy_device if
> ndev->phy_device is NULL, that way changes could be saved
> and applied when phy_connect replaces the dummy device
> (or perhaps when phy_start() is called).
> (If it isn't a dummy phy_device, changes could be applied immediately.)
> This way it would not matter if phy_connect is done from probe
> or from ndo_open.
>
I don't think there is a need for a dummy PHY device, you can do this in
your driver if you want to, but this adds a lot of complexity for little
value, and, keep in mind that before ndo_open() is complete, most
operations against your network driver can be undefined.
--
Florian
