Re: [PATCH v8 12/21] x86/resctrl: Move the filesystem bits to headers visible to fs/resctrl

[James Morse]

Hi Reinette,

On 17/04/2025 23:46, Reinette Chatre wrote:
>> diff --git a/arch/x86/include/asm/resctrl.h b/arch/x86/include/asm/resctrl.h
>> index 7a39728b0743..6eb7d5c94c7a 100644
>> --- a/arch/x86/include/asm/resctrl.h
>> +++ b/arch/x86/include/asm/resctrl.h
>> @@ -210,6 +210,9 @@ int resctrl_arch_measure_l2_residency(void *_plr);
>>  int resctrl_arch_measure_l3_residency(void *_plr);
>>  void resctrl_cpu_detect(struct cpuinfo_x86 *c);
>>  
>> +bool resctrl_arch_get_cdp_enabled(enum resctrl_res_level l);
>> +int resctrl_arch_set_cdp_enabled(enum resctrl_res_level l, bool enable);

> I noticed this while reading through the telemetry work ... the custom
> appears to be for all arch helpers to be declared in include/linux/resctrl.h
> making these two stand out. What motivated their inclusion into
> arch/x86/include/asm/resctrl.h?

Doing some archaeology ... it looks like this dates from the era when MPAM had stub
versions of these because it didn't support/emulate CDP. Having these in the arch code's
header meant the static-inline stubs could be inlined into resctrl. x86 has an inline
version before this patch, but I moved it out of line to avoid exposing the array of all
resources. (which in turn would suggest changing the name, which isn't worth the churn)

I was assuming only x86 would ever support CDP, but since then, emulating CDP using the
separate I/D PARTID values MPAM has proved feasible.

I'll move them.


> If they move to include/linux/resctrl.h it looks like enum resctrl_res_level is
> no longer required to be in include/linux/resctrl_types.h.

Yes. I'll add a patch move it, but it feels like churn.


> Looking further, from commit f16adbaf9272 ("x86/resctrl: Move resctrl types to a separate header")
> the motivation for including enum resctrl_event_id is to support 
> resctrl_arch_mon_ctx_alloc() and resctrl_arch_mon_ctx_free(), but looking
> at their definitions in arch/x86/include/asm/resctrl.h they are using "int evtid"
> instead of enum resctrl_event_id. Looks like their definitions need to
> change?

Bother. The MPAM definition of those has the enum type, I'd missed that the x86 stub
swallowed the type. I'll fix that.


> I assume the pseudo-locking arch helpers are placed in arch/x86/include/asm/resctrl.h
> with the implicit knowledge that only x86 will set CONFIG_RESCTRL_FS_PSEUDO_LOCK so
> avoiding that extra handling in include/linux/resctrl.h?

Given how specific that is to x86, I think it will only ever be enabled there.
For arm, things like the prefetcher can't be disabled by linux. I would assume risc-v is
structurally the same.

Those definitions went into the arch header so that some architectures could inline the
stubs to let the compiler remove code that will never run.

I guess these could be defined both ways round in linux/resctrl.h depending on the Kconfig
symbol. At this point its churn, but this saves a third architecture having to define the
same list of stubs. (I reckon that is likely for risc-v).


Thanks,

James