On 11.02.2025 9:02 AM, neil.armstrong@xxxxxxxxxx wrote:
> On 10/02/2025 19:29, Konrad Dybcio wrote:
>> On 10.02.2025 4:30 PM, neil.armstrong@xxxxxxxxxx wrote:
>>> On 10/02/2025 16:23, Konrad Dybcio wrote:
>>>> On 9.02.2025 3:44 PM, Neil Armstrong wrote:
>>>>> On 07/02/2025 21:30, Konrad Dybcio wrote:
>>>>>> On 7.02.2025 11:31 AM, Neil Armstrong wrote:
>>>>>>> The PMUs shares the same per-cpu (PPI) interrupt, so declare the proper
>>>>>>> interrupt partition maps and use the 4th interrupt cell to pass the
>>>>>>> partition phandle for each ARM PMU node.
>>>>>>>
>>>>>>> Signed-off-by: Neil Armstrong <neil.armstrong@xxxxxxxxxx>
>>>>>>> ---
>>>>>>
>>>>>>> @@ -5309,6 +5309,20 @@ intc: interrupt-controller@17100000 {
>>>>>>> #size-cells = <2>;
>>>>>>> ranges;
>>>>>>> + ppi-partitions {
>>>>>>> + ppi_cluster0: interrupt-partition-0 {
>>>>>>> + affinity = <&cpu0 &cpu1>;
>>>>>>> + };
>>>>>>> +
>>>>>>> + ppi_cluster1: interrupt-partition-1 {
>>>>>>> + affinity = <&cpu2 &cpu3 &cpu4 &cpu5 &cpu6>;
>>>>>>> + };
>>>>>>> +
>>>>>>> + ppi_cluster2: interrupt-partition-2 {
>>>>>>> + affinity = <&cpu7>;
>>>>>>> + };
>>>>>>
>>>>>> I'm not sure this is accurate.
>>>>>>
>>>>>> I *think* it's cores 0-1 and 2-7, but I can't find a concrete answer
>>>>>
>>>>> Core 7 is a Cortex-X4, and has a dedicated PMU node, look at the cpu compatibles.
>>>>
>>>> Look at what these compatibles do in code. Nothing special for the X.
>>>
>>> So you suggest to revert Rob's change ?
>>>
>>> https://lore.kernel.org/all/20240417204247.3216703-1-robh@xxxxxxxxxx/
>>>
>>> So what I understood is that yes the code is the same, but the perf
>>> attributes are not necessarily the same between heterogeneous cores,
>>> so each instance here would load the attributes for each core type
>>> correctly instead of only using the ones from the first core
>>>
>>> Again, other SoCs uses this same scheme so I wonder what's the issue here ?
>>
>> So I'm a little confused here. Is this partitioning scheme only describing
>> a set of same-kind cores to Linux so that the PMU interrupts only arrive at
>> one PMU device? Or does it reflect some actual physical topology of clusters
>> and how they're connected to the GIC?
>>
>> If the former, I have no issues with this version of the patch.
>
> The former, we simply partition the PPI interrupt for each device node, it
> has obviously something to do with the topology, but the goal is not to describe
> the topology.
Ok, good
Reviewed-by: Konrad Dybcio <konrad.dybcio@xxxxxxxxxxxxxxxx>
Konrad
