Re: [PATCH v3] powerpc/pseries: read the lpar name from the firmware

[Laurent Dufour <ldufour@linux.ibm.com>]

On 08/12/2021, 16:21:29, Nathan Lynch wrote:
> Laurent Dufour <ldufour@linux.ibm.com> writes:
>> On 07/12/2021, 18:07:50, Nathan Lynch wrote:
>>> Laurent Dufour <ldufour@linux.ibm.com> writes:
>>>> On 07/12/2021, 15:32:39, Nathan Lynch wrote:
>>>>> Is there a reasonable fallback for VMs where this parameter doesn't
>>>>> exist? PowerVM partitions should always have it, but what do we want the
>>>>> behavior to be on other hypervisors?
>>>>
>>>> In that case, there is no value displayed in the /proc/powerpc/lparcfg and
>>>> the lparstat -i command will fall back to the device tree value. I can't
>>>> see any valid reason to report the value defined in the device tree
>>>> here.
>>>
>>> Here's a valid reason :-)
>>>
>>> lparstat isn't the only possible consumer of the interface, and the
>>> 'ibm,partition-name' property and the dynamic system parameter clearly
>>> serve a common purpose. 'ibm,partition-name' is provided by qemu.
>>
>> If the hypervisor is not providing this value, this is not the goal of this
>> interface to fetch it from the device tree.
>>
>> Any consumer should be able to fall back on the device tree value, and
>> there is no added value to do such a trick in the kernel when it can be
>> done in the user space.
> 
> There is value in imposing a level of abstraction so that the semantics
> are:
> 
> * Report the name assigned to the guest by the hosting environment, if
>   available
> 
> as opposed to
> 
> * Return the string returned by a RTAS call to ibm,get-system-parameter
>   with token 55, if implemented
> 
> The benefit is that consumers of lparcfg do not have to be coded with
> the knowledge that "if a partition_name= line is absent, the
> ibm,get-system-parameter RTAS call must have failed, so now I should
> read /sys/firmware/devicetree/base/ibm,partition_name." That's the sort
> of esoterica that is appropriate for the kernel to encapsulate.
> 
> And I'd say the effort involved (falling back to a root node property
> lookup) is proportional to the benefit.
> 

I don't agree.
From the kernel point of view, I can't see any benefit, this is adding more
complexity to do in the kernel what can be done easily in user space.

This is typically what should be implemented in a user space shared library.