Re: [PATCH v7 11/14] mpam,x86/resctrl: Resolve INTEL_PMT_TELEMETRY symbols at runtime

[Luck, Tony]

On Thu, Jun 11, 2026 at 02:22:23PM -0700, Reinette Chatre wrote:
> Hi Tony,
> 
> On 6/11/26 11:01 AM, Luck, Tony wrote:
> > With renamed variable, and comment. Is this clearer?
> > 
> > 
> > /*
> >  * intel_aet_register_enumeration() provides the value of "THIS_MODULE"
> >  * from the the pmt_telemetry module. Note that this is NULL when
> >  * CONFIG_INTEL_PMT_TELEMETRY=y so it serves as a flag to decide whether
> >  * try_module_get() and module_put() need to be called.
> >  */
> > static struct module *pmt_is_a_module;
> > 
> 
> Using a single variable to contain both data and state has caused problem in
> the past. Could it be helpful to create two utilities, for example get_pmt()
> and put_pmt() although naming is not at issue at the moment, that contains
> and makes obvious these "is INTEL_PMT_TELEMETRY a module or not" checks? These would
> then be no-op when INTEL_PMT_TELEMETRY=y. This would simplify the code in the
> callers by containing this logic and not rely on callers doing this check so
> many times.

I can code this up to see how it looks.

> Even so, it is not clear to me how the AET lifetime will look like since 
> INTEL_PMT_TELEMETRY's .remove() may be called at other times so resctrl
> may need other hooks that could replace this module handling.

Case 1: No races

Resctrl mount: intel_aet_pre_mount() gets a success return from
try_module_get() - so the hold on the pmt_module will block the
kernel from removing the module (-EBUSY).

Case 2: Race between resctrl mount and rmmod pmt_telemetry

Mount code gets into intel_aet_pre_mount() and rmmod gets to intel_aet_unregister_enumeration()
where both try to acquire aet_register_lock.

	If intel_aet_pre_mount() gets the mutex it calls try_module_get(), which
	fails because if the pmt_telemetry code is in the .remove() functions,
	then the module state is already in MODULE_STATE_GOING.

	If rmmod process wins the mutex it sets get_feature and put_feature to
	NULL and releases the mutex. intel_aet_pre_mount() runs and returns
	false because get_feature and put_feature are NULL.

Is there some other race condition to consider?

> Reinette

-Tony

P.S. David Box is OK with setting suppress_bind_attrs for the telemetry
driver. So that will make the unbind issue moot.
>