Hi,
On 3/26/21 10:55 AM, Henning Schild wrote:
> Am Thu, 18 Mar 2021 12:45:01 +0100
> schrieb Hans de Goede <hdegoede@xxxxxxxxxx>:
>
>> Hi,
>>
>> On 3/18/21 12:30 PM, Enrico Weigelt, metux IT consult wrote:
>>> On 17.03.21 21:03, Hans de Goede wrote:
>>>
>>> Hi,
>>>
>>>>> It just identifies the box and tells subsequent drivers which one
>>>>> it is, which watchdog and LED path to take. Moving the knowledge
>>>>> of which box has which LED/watchdog into the respective drivers
>>>>> seems to be the better way to go.
>>>>>
>>>>> So we would end up with a LED and a watchdog driver both
>>>>> MODULE_ALIAS("dmi:*:svnSIEMENSAG:*");
>>>
>>> Uh, isn't that a bit too broad ? This basically implies that Siemens
>>> will never produce boards with different configurations.
>>
>> There is a further check done in probe() based on some Siemens
>> specific DMI table entries.
>>
>>>>> and doing the identification with the inline dmi from that header,
>>>>> doing p2sb with the support to come ... possibly a
>>>>> "//TODO\ninline" in the meantime.
>>>>>
>>>>> So no "main platform" driver anymore, but still central platform
>>>>> headers.
>>>>>
>>>>> Not sure how this sounds, but i think making that change should be
>>>>> possible. And that is what i will try and go for in v3.
>>>>
>>>> Dropping the main drivers/platform/x86 driver sounds good to me,
>>>> I was already wondering a bit about its function since it just
>>>> instantiates devs to which the other ones bind to then instantiate
>>>> more devs (in the LED case).
>>>
>>> hmm, IMHO that depends on whether the individual sub-devices can be
>>> more generic than just that specific machine. (@Hanning: could you
>>> tell us more about that ?).
>>>
>>> Another question is how they're actually probed .. only dmi or maybe
>>> also pci dev ? (i've seen some refs to pci stuff in the led driver,
>>> but missed the other code thats called here).
>>>
>>> IMHO, if the whole thing lives on some PCI device (which can be
>>> probed via pci ID), and that device has the knowledge, where the
>>> LED registers actually are (eg. based on device ID, pci mmio
>>> mapping, ...) then there should be some parent driver that
>>> instantiates the led devices (and possibly other board specific
>>> stuff). That would be a clear separation, modularization. In that
>>> case, maybe this LED driver could even be replaced by some really
>>> generic "register-based-LED" driver, which just needs to be fed
>>> with some parameters like register ranges, bitmasks, etc.
>>>
>>> OTOH, if everything can be derived entirely from DMI match, w/o
>>> things like pci mappings involved (IOW: behaves like directly wired
>>> to the cpu's mem/io bus, no other "intelligent" bus involved), and
>>> it's all really board specific logic (no generic led or gpio
>>> controllers involved), then it might be better to have entirely
>>> separate drivers.
>
> In fact it does dmi and not "common" but unfortunately vendor-specific.
> On top it does pci, so it might be fair to call it "intelligent" and
> keep it.
>
>> FWIW I'm fine with either solution, and if we go the "parent driver"
>> route I'm happy to have that driver sit in drivers/platform/x86
>> (once all the discussions surrounding this are resolved).
>>
>> My reply was because I noticed that the Led driver seemed to sort of
>> also act as a parent driver (last time I looked) and instantiated
>> a bunch of stuff, so then we have 2 parent(ish) drivers. If things
>> stay that way then having 2 levels of parent drivers seems a bit too
>> much to me, esp. if it can all be done cleanly in e.g. the LED driver.
>
> One "leds" driver doing multiple leds seems to be a common pattern. So
> that "1 parent N children" maybe does not count as parentish.
>
>> But as said I'm fine either way as long as the code is reasonably
>> clean and dealing with this sort of platform specific warts happens
>> a lot in drivers/platform/x86 .
>
> I thought about it again and also prefer the "parent driver" idea as it
> is. That parent identifies the machine and depending on it, causes
> device drivers to be loaded. At the moment LED and watchdog, but with
> nvram, hwmon to come.
>
> I will stick with "platform" instead of "mfd" because it is really a
> machine having multiple devices. Not a device having multiple functions.
Ok, sticking with the current separate "platform" parent driver design
is fine by me.
Regards,
Hans
