Re: [PATCH v12 15/25] x86,fs/resctrl: Handle systems where AET is the only resource

[Luck, Tony]

On Fri, Sep 25, 2026 at 07:37:20AM -0700, Reinette Chatre wrote:
> Hi Tony,
> 
> On 9/24/26 5:06 PM, Luck, Tony wrote:
> > On Thu, Sep 24, 2026 at 01:41:48PM -0700, Reinette Chatre wrote:
> >> On 9/16/26 4:13 PM, Tony Luck wrote:
> 
> >>> @@ -1001,27 +1001,29 @@ static __init bool get_rdt_mon_resources(void)
> >>>  
> >>>  	/* Any of the L3 monitoring features? */
> >>>  	if (!cpu_feature_enabled(X86_FEATURE_CQM_LLC))
> >>> -		return false;
> >>> +		goto skip_l3_feature_checks;
> >>
> >> This goto looks unnecessary. Why not just "return true"? Even so, this 
> >> also changes behavior from the previous version in a way that is not clear
> >> to me. I was expecting this to consider the number of RMIDs supported by the system,
> >> now that this function added:
> >>
> >> 	pqr_assoc_num_rmid = cpuid_ebx(0xf) + 1;
> >>
> >> Should this be:
> >>
> >> 	if (!cpu_feature_enabled(X86_FEATURE_CQM_LLC))
> >> 		return pqr_assoc_num_rmid > 1;
> >>
> >> Although ... looking at this closer it does look strange for pqr_assoc_num_rmid
> >> to be 1, thus reflecting that the system supports one RMID, when zero may be more
> >> accurate?
> > 
> > If X86_FEATURE_CQM_LLC is set, then monitoring is supported and the system
> > must support at least one RMID. I don't see a need to insist on more than one
> > RMID as a precondition for using resctrl. It wouldn't be super interesting
> 
> My suggestion was not to use "more than one RMID as a precondition". By copying the
> relevant code not shown in the hunk
> 
> 	pqr_assoc_num_rmid = cpuid_ebx(0xf) + 1;
> 
> I aimed to highlight that if the system returns zero then pqr_assoc_num_rmid
> will contain 1. The "pqr_assoc_num_rmid > 1" is thus not a check for "more
> than one RMID" but instead a check that the system did not return zero.

Architecturally the cpuid_ebx(0xf) can legally return 0 meaning that the
system supports one RMID. Here's the description from SDM Volume 1 Table
21-38 "Leaf 0FH.00H Intel® Resource Director Technology (Intel® RDT) Monitoring":

	Register	Field Name	Description
	EBX[31:0]	MAX_RMID	Maximum range (zero-based) of RMID within
					this physical processor of all types.

This code path is only executed if the test for X86_FEATURE_CQM
succeeded. That should guarantee that leaf 0xF exists[1]. There's
no way for leaf 0xF to indicate zero RMIDs because the MAX_RMID
field gives a zero-based result.

So a test for "pqr_assoc_num_rmid > 1" really is checking that two or
more RMIDs are supported.

If you want resctrl to refuse to support a crazy system with only one
RMID, then the check should as soon as pqr_assoc_num_rmid is set:

	pqr_assoc_num_rmid = cpuid_ebx(0xf) + 1;

	/* Systems with only one RMID can't usefully support resctrl */
	if (pqr_assoc_num_rmid == 1)
		return false;

> 
> > as all activity would be captured by that single RMID.  But it would be
> > keeping counts of events.
> > 
> > So you first instinct above to "just return true" seems reasonable.
> > 
> Reinette

-Tony

[1] Technically should check "cpuid_eax(0) >= 0xF" but it feels like only
a buggy h/w implementation would set X86_FEATURE_CQM on a platform that
didn't have leaf 0xF.