[RFC PATCH 01/38] arm64: mpam: Context switch the MPAM registers

[Ben Horgan]

Hi James,

On 12/5/25 21:58, James Morse wrote:
> MPAM allows traffic in the SoC to be labeled by the OS, these labels
> are used to apply policy in caches and bandwidth regulators, and to
> monitor traffic in the SoC. The label is made up of a PARTID and PMG
> value. The x86 equivalent calls these CLOSID and RMID, but they don't
> map precisely.
> 
> MPAM has two CPU system registers that is used to hold the PARTID and PMG
> values that traffic generated at each exception level will use. These can be
> set per-task by the resctrl file system. (resctrl is the defacto interface
> for controlling this stuff).
> 
> Add a helper to switch this.
> 
> struct task_struct's separate CLOSID and RMID fields are insufficient
> to implement resctrl using MPAM, as resctrl can change the PARTID (CLOSID)
> and PMG (sort of like the RMID) separately. On x86, the rmid is an
> independent number, so a race that writes a mismatched  closid and rmid
> into hardware is benign. On arm64, the pmg bits extend the partid.
> (i.e. partid-5 has a pmg-0 that is not the same as partid-6's pmg-0).
> In this case, mismatching the values will 'dirty' a pmg value that
> resctrl believes is clean, and is not tracking with its 'limbo' code.
> 
> To avoid this, the partid and pmg are always read and written as a pair.
> Instead of making struct task_struct's closid and rmid fields an
> endian-unsafe union, add the value to struct thread_info and always use
> READ_ONCE()/WRITE_ONCE() when accessing this field.
> 
> Resctrl allows a per-cpu 'default' value to be set, this overrides the
> values when scheduling a task in the default control-group, which has
> PARTID 0. The way 'code data prioritisation' gets emulated means the
> register value for the default group needs to be a variable.
> 
> The current system register value is kept in a per-cpu variable to
> avoid writing to the system register if the value isn't going to change.
> Writes to this register may reset the hardware state for regulating
> bandwidth.
> 
> Finally, there is no reason to context switch these registers unless
> there is a driver changing the values in struct task_struct. Hide
> the whole thing behind a static key. This also allows the driver to
> disable MPAM in response to errors reported by hardware. Move the
> existing static key to belong to the arch code, as in the future
> the MPAM driver may become a loadable module.
> 
> All this should depend on whether there is an MPAM driver, hide
> it behind CONFIG_MPAM.
> 
> CC: Amit Singh Tomar <amitsinght at marvell.com>
> Signed-off-by: James Morse <james.morse at arm.com>
[...]
> +
> +static inline void mpam_thread_switch(struct task_struct *tsk)
> +{
> +	u64 oldregval;
> +	int cpu = smp_processor_id();
> +	u64 regval = mpam_get_regval(tsk);
> +
> +	if (!IS_ENABLED(CONFIG_ARM64_MPAM) ||
> +	    !static_branch_likely(&mpam_enabled))
> +		return;
> +
> +	if (regval == READ_ONCE(arm64_mpam_global_default))
> +		regval = READ_ONCE(per_cpu(arm64_mpam_default, cpu));
> +
> +	oldregval = READ_ONCE(per_cpu(arm64_mpam_current, cpu));
> +	if (oldregval == regval)
> +		return;
> +
> +	write_sysreg_s(regval, SYS_MPAM1_EL1);
> +	isb();
> +
> +	/* Synchronising the EL0 write is left until the ERET to EL0 */
> +	write_sysreg_s(regval, SYS_MPAM0_EL1);

SYS_MPAMSM_EL1 needs to be written here too to account for accesses
generated by SME loads. Also, when in streaming mode, SVE and FP loads,
stores and SVE prefetches.

SYS_MPAMSM_EL1 should also be considered in initialisation code too.

> +
> +	WRITE_ONCE(per_cpu(arm64_mpam_current, cpu), regval);
> +}
Thanks,

Ben