On 12/03/18 12:13, Yang Yingliang wrote:
>
>
> On 2018/3/12 17:55, Marc Zyngier wrote:
>> On 12/03/18 06:49, Yang Yingliang wrote:
>>> This patch set adjust struct rdists to support CommonLPIAff field.
>>> This field is described in spec:
>>>
>>> CommonLPIAff, bits [25:24]
>>> The affinity level at which Redistributors share a LPI Configuration table.
>>> 00 All Redistributors must share a LPI Configuration table.
>>> 01 All Redistributors with the same Aff3 value must share an LPI
>>> Configurationt table.
>>> 10 All Redistributors with the same Aff3.Aff2 value must share an LPI
>>> Configuration table.
>>> 11 All Redistributors with the same Aff3.Aff2.Aff1 value must share an LPI
>>> Configuration
>>> table.
>>
>> What are you trying to achieve here? We already share the same
>> configuration table across all the redistributors, irrelevant the of
>> CommonLPIAff. Why would we need to do anything else?
>
> In some cases, such as "with the same Aff3.Aff2 value must share an LPI
> Configuration table", redistributors will access local memory node
> only, it may improve the performance. And I don't have any beneficial
> data now. I may provide it later.
These are two different things.
- "must share" is what the kernel does *today*. No changes are required.
- NUMA behaviour has nothing to do with CommonLPIAff, and cannot be
based on that field.
Thanks,
M.
--
Jazz is not dead. It just smells funny...
