Re: [RFC PATCH 00/10] arm_mpam: Add extra MPAM controls in resctrl

[Reinette Chatre]

Hi Jinjie,

On 10/22/26 7:48 PM, Jinjie Ruan wrote:
> Hi all,
> 
> Resctrl today only exposes CBM and the MBA bandwidth percentage.
> But MPAM MSCs also have per-PARTID fields for internal priority, capacity
> limits, a bandwidth minimum and the MBW_MAX hard-limit bit. The driver
> already writes these fields, but resctrl has no schemata line for them,
> so userspace can't choose the values.
> 
> So this series adds those schemata lines:
> 
> - Patch 1 fixes maybe_update_config(): a feature set with value zero
>   counts as no change, so the register is not written again.
>   The controls below can be zero, so this comes first.
> 
> - Patches 2-4 add mbw_min/max_limit, cmax/cmin and intpri to
>   mpam_config and write them from mpam_reprogram_ris_partid().
> 
> - Patch 5 adds the resource levels, selects the range path by schema_fmt
>   instead of the MBA/SMBA rid, skips the new resources in info/, and sets
>   their default value when a group is created.
> 
> - Patches 6-7 adjust the fixed-point helpers and clamp cmax_wd.
> 
> - Patches 8-10 add the schemata lines:
> 
>   L2MAX/L3MAX       CMAX, percent
>   L2MIN/L3MIN       CMIN, percent
>   MBMIN             MBW_MIN, percent
>   L2PRI/L3PRI/MBPRI INTPRI, raw value
>   MBHDL             HARDLIM, 0 or 1
>   MBOPT             MBW_MAX, raw fraction
> 
> The new resources reuse the class, domains and components of L2/L3/MB; only
> the mpam_config field written differs. MBOPT and MBA both write MBW_MAX.
> 
> So before this series userspace can only set CBM and the MBA percentage.
> After it, you can do:
> 
>   # echo "L3MAX:33751040=50" > schemata
>   # echo "MBHDL:33751040=1" > schemata
> 
> A new group starts from each resource's default value, so the new controls
> stay out of the way until schemata is written.
> 
> Tested this with a QEMU that has one L2 MSC, one L3 MSC and one
> memory-controller MSC (CMAX_WD 4, INTPRI_WD 2, BWA_WD 16). The guest writes
> every new line and reads it back, for example:
> 
>   WRITE-OK   [L3MAX:33751040=50]
>   SCHEMATA: L3MAX:33751040=50
> 
> Then reads the MSC registers:
> 
>   L3 MSC  partid0: PRI(IN)=2  CMIN=0x3000  CMAX=0x7000
>   mem MSC partid0: PRI(IN)=1  MBW_MIN=0x3fff  MBW_MAX=0x1fff  HARDLIM=Y
> 
> CMIN/CMAX are 4-bit fractions (3/16 for 25%, 7/16 for 50%); MBW_MIN is a
> 16-bit fraction (0x3fff for 25%); MBOPT writes the raw fraction (8191).

How does user space know what the valid control values are? What if a new
platform comes along with the same control but different properties of the
control value?


Please consider the ongoing work around adding support for multiple controls
to resctrl. Most recent RFC can be found at 

	https://lore.kernel.org/lkml/d258a32f-12d5-464d-abe9-4720fb3e44b3@xxxxxxxxx/

An update on this can be found in LPC slides:

	https://lpc.events/event/20/contributions/2368/attachments/2066/4625/resctrl_2026_combined.pdf

The work aims to make the distinction between a resource and its control clear.
It will expose the controls in the schemata file but also support it
with a hierarchy in the info/ directory that makes clear to user space how the
control is associated with the resource while also providing an interface from which
user space can learn what the valid control values are as well as what a domain value
represents.

Reinette