Re: [PATCH 1/3] fixed_phy: handle link-down case

[Florian Fainelli <f.fainelli@gmail.com>]

On 17/07/15 16:24, Stas Sergeev wrote:
> 18.07.2015 01:01, Florian Fainelli пишет:
>> On 17/07/15 13:03, Stas Sergeev wrote:
>>> 17.07.2015 21:50, Florian Fainelli пишет:
>>>> On 17/07/15 04:26, Stas Sergeev wrote:
>>>>> 17.07.2015 02:25, Florian Fainelli пишет:
>>>>>> On 16/07/15 07:50, Stas Sergeev wrote:
>>>>>>> Currently fixed_phy driver recognizes only the link-up state.
>>>>>>> This simple patch adds an implementation of link-down state.
>>>>>>> It fixes the status registers when link is down, and also allows
>>>>>>> to register the fixed-phy with link down without specifying the
>>>>>>> speed.
>>>>>> This patch still breaks my setups here, e.g:
>>>>>> drivers/net/dsa/bcm_sf2.c,
>>>>>> but I will look into it.
>>>>>>
>>>>>> Do we really need this for now for your two other patches to work
>>>>>> properly, or is it just nicer to have?
>>>>> Yes, absolutely.
>>>>> Otherwise registering fixed phy will return -EINVAL
>>>>> because of the missing link speed (even though the link
>>>>> is down).
>>>> Ok, I see the problem that you have now. Arguably you could say that
>>>> according to the fixed-link binding, speed needs to be specified and
>>>> the
>>>> code correctly errors out with such an error if you do not specify
>>>> it. I
>>> Aren't you missing the fact that .link=0?
>>> I think what you say is true only for the link-up case, no?
>>> .speed==0 is valid for link-down IMHO: no link - zero speed.
>> Pardon me being very dense and stupid here, but your problem is that the
>> "speed" parameter is not specified in your DT,
> Not even a fixed-link at all, since the latest patches.
> I removed fixed-link defs from my DT.

Hummm, okay, so you just have the inband-status property and that's it,
not even a fixed-link node anymore, right? How does
mvneta_fixed_link_update() work then since it needs a fixed PHY to be
registered?

> 
>>   and we end-up returning
>> -EINVAL from of_phy_register_fixed_link(), is that what is happening?
> Yes.
> 
>> And even if we silenced that error,
> I don't agree with calling it an error silencing.
> To me it is a fix. It will also return a more correct status when
> link is down.
> 
>>   we would end-up calling
>> fixed_phy_add() which would also return -EINVAL because then, we would
>> have status.link = 1, but no speed.
> Why link=1 and no speed? This is not valid, should never
> be used. The error checking is still there to prevent it.
> 
>>   So I better understand what is it
>> that you are after here, and that is also a broken Device Tree, is not
>> it?
> I don't understand. If you didn't specify the in-band status, you
> _must_ set the speed. There is no broken DT in either case.



> 
>>   So this was the reason why in earlier versions of the patchset you
>> ended-up with a given speed which would make us pass this condition,
>> right?
> As explained earlier, yes.
> 
> 
>>>> So is different is that I use a link_update callback, and so we rely on
>>>> at least one call of this function to initialize the hardware in
>>>> drivers/net/dsa/bcm_sf2.c
>>> Do you mean this?:
>>> core_writel(priv, reg, CORE_STS_OVERRIDE_GMIIP_PORT(port));
>>> Maybe just moving the HW initialization bits to some init func
>>> will be a quick fix?
>> Well, the problem with that is that to know how we should be configuring
>> the hardware in the adjust_link function, we need to run the link_update
>> function first. By default, there is no auto-negotiation on these fixed
>> links at all, so we cannot rely on any value being programmed other than
>> those specified in DT.
> Ah, so is my understanding correct that in fixed_link_update()
> you set .link=0 and as a result get wrong speed in adjust_link(),
> which gets then written to init HW?

Yes, that's what happens.

> AFAIK when link is down, you are not allowed to rely on the PHY
> status registers to read speed from, or am I wrong? So if my
> understanding is correct, this was working by a pure luck.

Well, it's more like it is undefined, and before this patch, the fixed
PHY would update everything except the link status indication.

> As for the quick fix - why not to do this pre-init in fixed_link_update()
> instead of adjust_link()? In fixed_link_update() you'll get the speed
> right from DT, so it will be correct.

fixed_link_update() only gets called once you start your PHY state
machine, unfortunately, not upon fixed PHY device registration, and it
runs before your adjust_link callback does, that's why starting with
correct parameters is kind of important here. Of course, this could be
fixed.

> 
>> The changes are not trivial, it took a while to get that logic done
> For a longer term fix,
> how about adding a *status arg to of_phy_register_fixed_link() to
> always get the status back to the driver, unless NULL is provided?
> Using an update callback for that doesn't look like the best thing
> to do. And besides, if we move to my fixed_phy_update_state(),
> this will be needed anyway.

I agree that the link_update callback is not the best thing, it polls
the hardware and comes with that problem that it may or may not have yet
run to configure your fixed_phy_status appropriately.
-- 
Florian