[PATCH 1/3] powerpc/pseries: Simplify cpu readd to use drc_index

[Tyrel Datwyler]

On 05/20/2019 08:01 AM, Nathan Lynch wrote:
> Tyrel Datwyler <tyreld at linux.vnet.ibm.com> writes:
> 
>> On 05/16/2019 12:17 PM, Nathan Lynch wrote:
>>> Tyrel Datwyler <tyreld at linux.ibm.com> writes:
>>>> The current dlpar_cpu_readd() takes in a cpu_id and uses that to look up
>>>> the cpus device_node so that we can get at the ibm,my-drc-index
>>>> property. The only user of cpu readd is an OF notifier call back. This
>>>> call back already has a reference to the device_node and therefore can
>>>> retrieve the drc_index from the device_node.
>>>
>>> dlpar_cpu_readd is a hack to try to change the CPU-node relationship at
>>> runtime without destabilizing the system. It doesn't accomplish that and
>>> it should just be removed (and I'm working on that).
>>>
>>
>> I will politely disagree. We've done exactly this from userspace for
>> years. My experience still suggests that memory affinity is the
>> problem area, and that the work to push this all into the kernel
>> originally was poorly tested.
> 
> Kernel implementation details aside, how do you change the cpu-node
> relationship at runtime without breaking NUMA-aware applications? Is
> this not a fundamental issue to address before adding code like this?
> 

If that is the concern then hotplug in general already breaks them. Take for
example the removal of a faulty processor and then adding a new processor back.
It is quite possible that the new processor is in a different NUMA node. Keep in
mind that in this scenario the new processor and threads gets the same logical
cpu ids as the faulty processor we just removed.

Now we have to ask the question who is right and who is wrong. In this case the
kernel data structures reflect the correct NUMA topology. However, did the NUMA
aware application or libnuma make an assumption that specific sets of logical
cpu ids are always in the same NUMA node?

-Tyrel