[PATCH DNM v3 8/8] arm64: dts: qcom: kaanapali: Add MPAM MSC nodes for the L2 caches

[Yin Li]

On 10/26/2026 8:23 PM, Rob Herring wrote:
> On Fri, Oct 23, 2026 at 02:03:57PM +0200, Konrad Dybcio wrote:
>> On 10/14/26 9:09 AM, Yin Li wrote:
>>>
>>>
>>> On 10/13/2026 6:37 PM, Ben Horgan wrote:
>>>> Hi Yin,
>>
>> [...]
>>
>>> Hi Ben, Konrad,
>>>
>>> The original 0x10000 reflects the per-MSC register block size as
>>> allocated in the KNP hardware — as Konrad noted, Qualcomm DTs tend to
>>> use "how much register space is allocated for this block" rather than
>>> "what the last used register ends" to avoid holes.
>>>
>>> We'd like to update it to 0x321c to give other platform authors a more useful reference, without exposing KNP-internal details such as address holes between blocks. In our case, we don't implement MPAMCFG_MBW_PBM, and — as you noted — the registers from 0x3000 onward are domain-specific; since we have no domains, those aren't used either. So the actual footprint we need is smaller than 0x321c, but showing the full 0x321c gives other authors a complete picture of the allocated register space.
>>>
>>> That said, this patch is marked DNM and is intended as a local
>>> verification example only, not for merging.
>>
>> If you don't intend it to get merged, you can even put 0x12345
>> in there. But for future submissions that will have the intent to
>> be merged, please stick to the 0x10000 size
> 
> Why? That just wastes virtual memory space. I know it is "infinite" on
> 64-bit and *only* 96 bytes of PTEs, but still.
> 

Hi Rob, Konrad,

In v4 we settled on 0x2000 for the Kaanapali DTS. This reflects the
actual register footprint: the platform does not implement
MPAMCFG_MBW_PBM, so per the MPAM specification the feature page is
permitted to end at 0x1FFF. The remaining space up to 0x10000 is
reserved for this MSC on the Kaanapali hardware. Compared to 0x10000
(the full block size as allocated in the KNP hardware), 0x2000 avoids
mapping unused space as Rob noted. Compared to 0x321c (the full MPAM
feature page per the spec), 0x2000 more accurately reflects the
platform's actual implementation, while still being grounded in the
specification rather than an arbitrary value.

We have now posted v4, which also reduces the binding and driver to the
minimal set actually used on the platform: aside from Kaanapali, there
are no other concrete platform users, so memory controller MSC support,
RIS child node support, and related properties (arm,mpam-device, cpus,
interrupt-names) have all been dropped.

https://lore.kernel.org/all/20261028-mpam-resctrl-dt-knp-support-v4-6-bdda1e445cba@oss.qualcomm.com/

Thanks,
Yin

> Rob
> 

-- 
Thx and BRs,
Yin