arm_mpam: selecting hard or soft MBW_MAX limits

arm_mpam: selecting hard or soft MBW_MAX limits

[Ping Li]

Hi Ben,

mbw_max support in the driver always programs MPAMCFG_MBW_MAX without
setting MPAMCFG_MBW_MAX.HARDLIM, so the limit is always soft: a
PARTID above its MAX contends with low priority for bandwidth beyond
the limit, but is not prevented from exceeding it. On our platform we
would like to be able to select the hard limit, where a PARTID above
its MAX is throttled even when the system is otherwise idle - soft
limits let a noisy task exceed its limit and disturb higher-priority
work when bandwidth contention appears.

MPAMF_MBW_IDR.MAXLIM reports whether the MSC implements a hard limit,
a soft limit, or both, but the driver does not read it today.

Is there a preferred way to expose this choice? The options we see:

  - a mount option or info flag, if the choice must be global
  - a per-resource or per-group resctrl knob
  - selecting hard limits by default whenever MAXLIM reports the MSC
    supports them

Thanks,
Ping

Re: arm_mpam: selecting hard or soft MBW_MAX limits

[Ben Horgan]

Hi Ping,

On 06/11/2026 04:36, Ping Li wrote:
> Hi Ben,
> 
> mbw_max support in the driver always programs MPAMCFG_MBW_MAX without
> setting MPAMCFG_MBW_MAX.HARDLIM, so the limit is always soft: a

Just to clarify. The limit is always soft whenever a soft limit is supported but if it isn't
MPAMCFG_MBW_MAX.HARDLIM is RAO/WI and so even though the driver writes 0 to MPAMCFG_MBW_MAX.HARDLIM
you would still get a hard limit.

> PARTID above its MAX contends with low priority for bandwidth beyond
> the limit, but is not prevented from exceeding it. On our platform we
> would like to be able to select the hard limit, where a PARTID above
> its MAX is throttled even when the system is otherwise idle - soft
> limits let a noisy task exceed its limit and disturb higher-priority
> work when bandwidth contention appears.
> 
> MPAMF_MBW_IDR.MAXLIM reports whether the MSC implements a hard limit,
> a soft limit, or both, but the driver does not read it today.
> 
> Is there a preferred way to expose this choice? The options we see:
> 
>   - a mount option or info flag, if the choice must be global
>   - a per-resource or per-group resctrl knob
>   - selecting hard limits by default whenever MAXLIM reports the MSC
>     supports them

The current plan is that this will be an extra line in the schemata file. Fenghua was looking into
this and there is some discussion here [1]. This relies on the generic schemata changes to have
multiple controls for resources. The same series as linked although there is a rfc v2 now.

[1] https://lore.kernel.org/lkml/e13cf2e7-3956-4069-ac43-c65e874c6109@intel.com/

Thanks,

Ben

> 
> Thanks,
> Ping