Re: [PATCH v4 0/6] arm-mpam: Add basic device tree support for resctrl

From: Yin Li

Date: Sun Nov 08 2026 - 22:00:00 EST




On 11/6/2026 11:08 PM, Rob Herring wrote:
On Thu, Nov 05, 2026 at 01:53:47PM +0800, Ping Li wrote:
Hi Yin,

Sure. Our memory controller MSC is described as a top-level node.
The MPAM register window is mapped independently in the SoC (the
component sits on the DDR egress path rather than being a sub-block
of a memory-controller DT node), so the node is a direct child of
the root:

ddr_msc: msc@2238000 {
compatible = "<vendor>,<soc>-ddr-qos",
"arm,mpam-memory-controller-msc",
"arm,mpam-msc";
reg = <0x0 0x02238000 0x0 0x1000>;
arm,not-ready-us = <10>;
};

The first compatible in the list is for our platform glue; I have
replaced it with a placeholder here since it is not relevant to the
binding discussion.

Notes on the properties, in case they matter for the binding:

- reg: the MPAM register window. Only a subset of the architectural
register space is implemented, so the size is the truncated window.
- single RIS: the component reports HAS_RIS=0, so there are no
ris@N child nodes.
- no interrupts: the component has no MPAM ESR block, so there is
no interrupts property.
- arm,not-ready-us: 10.

The verified probe path: the node matches via "arm,mpam-msc", and
patch 6/8 in v3 classifies it as MPAM_CLASS_MEMORY and treats it as
accessible from all CPUs, which is exactly the behavior we need.

One question in return: with the node being a direct child of the
root rather than a child of a memory-controller node, is that a
form the binding should cover? If the binding expects the MSC to
always be a sub-node of a memory controller, we would need the
top-level form to be allowed as well.

This would be why we're not upstreaming the binding without users. We
need actual h/w to validate the binding works.


Hi Ping, Rob,

Ping, thank you for the detailed description and for confirming this
on hardware.

To answer your question: the top-level form should be supported.
Unlike cache MSCs, where both class_id and component_id are derived
from the parent cache node, memory controller MSCs derive both from
the MSC node itself via of_node_to_nid(). There is no requirement for
the MSC to be a sub-node of a memory-controller node.

Rob, with Ping's hardware-validated use case covering the top-level
form, we would like to restore memory controller MSC support in v5.
We will use Ping's DT as a reference for the binding example.

Ping, a Tested-by from you would be very welcome once v5 is posted.

Thanks,
Yin


Rob

--
Thx and BRs,
Yin