Re: [PATCH] hwmon (occ): Avoid polling OCC during probe

[Eddie James <eajames@linux.ibm.com>]

On 4/27/22 08:45, Guenter Roeck wrote:
> On 4/26/22 08:26, Eddie James wrote:
>>
>> On 4/26/22 10:01, Guenter Roeck wrote:
>>> On 4/26/22 07:45, Eddie James wrote:
>>>> Instead of polling the OCC during the probe, use the "occ_active"
>>>> sysfs file to control when the driver polls the OCC for sensor data.
>>>> The reason for this change is that the SBE, which is the device by
>>>> which the driver communicates with the OCC, cannot handle 
>>>> communications
>>>> during certain system state transitions.
>>>>
>>>
>>> This is hackish, to say the least. Why not just instantiate the driver
>>> manually instead if userspace interaction is indeed needed, and there
>>> is no means to auto-instantiate it only after the hardware is ready ?
>>
>>
>> Well, the occ-hwmon platform devices are currently created by the 
>> parent FSI occ device (that's a bit of a hack too). I poked through 
>> the platform/device/bus interfaces but couldn't work out how to 
>> create the device but not probe it, so that userspace could later 
>> bind the device to the driver. The current userspace does just that, 
>> but it relies on the the initial probe failing since SBE can't handle 
>> the communications yet. The error scenario is that the SBE can get 
>> into a bad state, so communications shouldn't even be attempted.
>>
>> So to answer your question, there's no reason not to just instantiate 
>> the driver manually, but I don't actually know how to do so with 
>> these drivers.
>>
>>
>>>
>>> On the other side, what means does userspace have to determine that
>>> the SBE/OCC is ready ? Can't that same mechanism not be used in the
>>> kernel to auto-instantiate the driver ?
>>
>>
>> No it's impossible, it's not just some bit we can check in HW 
>> unfortunately, the signal is PLDM (platform level data model) over 
>> the network.
>>
>
> What a mess.


It is, yes. Just sent a v2. Thanks for taking a second look.


Eddie


>
> The subject and description are misleading. The code doesn't
> "avoid polling during probe", it delays hwmon registration until
> directed to do it by writing into a new sysfs attribute.
> Please update accordingly.


It's actually an existing sysfs file, I've just made it writeable.


>
> Thanks,
> Guenter