Re: [PATCH RFC v1 0/6] arm_mpam,resctrl: expose MPAM MBW_MAX hard limit
From: Ben Horgan
Date: Tue Nov 17 2026 - 08:44:04 EST
Hi Fenghua,
On 09/11/2026 21:29, Fenghua Yu wrote:
> This series exposes the ARM MPAM MBW_MAX HARDLIM bit through resctrl.
> The MBA schemata line sets the maximum bandwidth. HARDLIM selects
> whether that maximum is a hard limit. It is a per-domain 0/1 value, not
> another percentage, so it is a separate control on the MBA resource
> rather than a second bandwidth percentage.
>
> MPAMF_MBW_IDR.MAX_LIM reports how the implementation encodes that limit.
> When the probed value is 0, this kernel treats HARDLIM as read/write.
> When MAX_LIM is nonzero, or the platform has no hard-limit control, the
> schemata line is omitted.
>
> The line follows the MSC, the same way the bandwidth line does. A cache
> MSC uses cache ids:
>
> MB:1=100;2=100
> MB_MAXHLM:1=0;2=0
>
> A memory-level MSC uses NUMA node ids:
>
> MB_NODE:0=100;1=100
> MB_MAXHLM_NODE:0=0;1=0
>
> Please advise if "MB_MAXHLM" on cache ids and "MB_MAXHLM_NODE" on
> NUMA ids meet the new naming conventions.
It doesn't look to match the naming convention to me.
I was expecting: <resource_type>_<scope>_<control>_<control_param>.
Reinette asked you your view on this here:
https://lore.kernel.org/all/e13cf2e7-3956-4069-ac43-c65e874c6109@xxxxxxxxx/
Thanks,
Ben
>
> 0 clears HARDLIM (soft limit). 1 sets HARDLIM. New groups start with the
> hard limit off. info/MB/schemata/MB_MAXHLM/max and
> info/MB/schemata/MB_MAXHLM_NODE/max are 1. The bandwidth control's max
> stays 100. The control directory under info/MB/schemata/ is MB_MAXHLM or
> MB_MAXHLM_NODE.
>
> The 6 patches are grouped as follows:
>
> 1. resctrl hard-limit control (patch 1)
>
> Parse MB_MAXHLM as a 0/1 value on the MBA resource, and initialize
> new groups with HARDLIM clear.
>
> 2. MPAM probe and config (patches 2-3)
>
> Read MAX_LIM from MPAMF_MBW_IDR, derive HARDLIM read/write when that
> encoding is 0, and keep the HARDLIM bit in mpam_config so later
> MBW_MAX programming does not drop it.
>
> 3. Wire the control (patch 4)
>
> Create the MAXHLM control on the MBA resource when HARDLIM is
> read/write, and read and write it through resctrl_arch_get_config()
> and resctrl_arch_update_one().
>
> 4. Node scope and documentation (patches 5-6)
>
> On a memory-level MSC, name the control MB_MAXHLM_NODE, set the
> resource scope to the NUMA node, and give the hard-limit control the
> range 0..1. Document both schemata lines.
>
> The branch is on top CPU-less RFC V3 and resource controls RFC V2:
> https://github.com/fyu1/linux/commits/hlim.rfc.v1
>
> Testing
>
> Brought up on an ARM MPAM platform whose MBA resource is a memory-level
> MSC. resctrl mounts, and schemata lists MB_NODE and MB_MAXHLM_NODE with
> NUMA node ids. Not tested on x86 beyond compilation of the resctrl
> changes.
>
> Fenghua Yu (6):
> resctrl: wire MB_MAXHLM schemata parsing and group setup
> arm_mpam: resctrl: probe MPAMF_MBW_IDR MAX_LIM and hardlim_rw
> arm_mpam: resctrl: track MBW max hard-limit in config
> arm_mpam: resctrl: add MAXHLM control on the MBA resource
> arm_mpam: Add MB_MAXHLM_NODE and Node scope
> Documentation: resctrl: document MB_MAXHLM for MPAM MBA
>
> Documentation/filesystems/resctrl.rst | 28 ++++
> drivers/resctrl/mpam_devices.c | 67 ++++++++-
> drivers/resctrl/mpam_internal.h | 8 ++
> drivers/resctrl/mpam_resctrl.c | 191 +++++++++++++++++++++-----
> fs/resctrl/ctrlmondata.c | 36 +++++
> fs/resctrl/rdtgroup.c | 17 +++
> include/linux/resctrl.h | 16 ++-
> 7 files changed, 325 insertions(+), 38 deletions(-)
>