Re: [PATCH v2 7/7] x86/resctrl: Determine if Sub-NUMA Cluster is enabled and initialize.

[Tony Luck <tony.luck@intel.com>]

On Fri, Jun 23, 2023 at 05:19:37PM +0200, Peter Newman wrote:
> Hi Tony,
> 
> On Thu, Jun 22, 2023 at 6:02 PM Luck, Tony <tony.luck@intel.com> wrote:
> >
> > > Unfortunately I'm not getting as good of results with the new series.
> > > The main difference seems to be updating the 0xca0 MSR instead of
> > > applying the offset to PQR_ASSOC.
> >
> > I think I may have reversed the actions to update the MSR in one of
> > my refactor/rebase. The comment here is correct, but that's not
> > what the code is doing :-(
> >
> > Can you swap the bodies of these two functions and retest?
> 
> It's a small improvement, but still not great. Still only node 0
> giving believable results, but at least no more empty results from the
> second package.
> 
> I poked around in /proc/kcore and noticed that my snc_ways is still 1, though.

Below is the patch I applied to reverse the add/remove package actions
together with some debug to make double sure SNC is being detected as
I expect and the right actions taken.

When booting the debug messages say:

[    9.458624] resctrl: SNC_ways = 2
[    9.458801] resctrl: CPU0: set MSR_RMID_SNC_CONFIG to 0x0
[    9.461986] resctrl: CPU56: set MSR_RMID_SNC_CONFIG to 0x0

which is all good and correct.

For my tests I have a memory hog process that loops on a memcpy()
of 100 MBytes to generate enough traffic to be totally obvious when
looking at the mbm counters.

Test 1: I used taskset(1) to start a copy on the first CPU of node0
and checked the MBM counters. Both local and remote showed around
25 GB/s on node 0. Killed this process.

Tests 2, 3, 4: Same as 1, but started the process on first CPU of node
1, 2, 3.  Same result. Around 25 GB/s appeared in the MBM counts for
the right node in each cycle.

Test 5: Started on node0, then periodically used taskset to bind the
running process onto a CPU on another node.  It looks like Linux
migrates the memory for the job shortly after the affinity of the
process is changed. A few seconds after each process migration, the
MBM counters reflect traffic on the node that is the new home of the
process.

-Tony

From 414341db02cd51daaf4a9ea8bd68b22a23cf4b59 Mon Sep 17 00:00:00 2001
From: Tony Luck <tony.luck@intel.com>
Date: Fri, 23 Jun 2023 08:57:57 -0700
Subject: [PATCH] Fix inverted SNC enable/disable MSR writes. Add some debug
 too.

---
 arch/x86/kernel/cpu/resctrl/core.c | 21 ++++++++++++---------
 1 file changed, 12 insertions(+), 9 deletions(-)

diff --git a/arch/x86/kernel/cpu/resctrl/core.c b/arch/x86/kernel/cpu/resctrl/core.c
index 869cfb46e8e4..e66b2b84fe6f 100644
--- a/arch/x86/kernel/cpu/resctrl/core.c
+++ b/arch/x86/kernel/cpu/resctrl/core.c
@@ -546,18 +546,20 @@ static void snc_add_pkg(void)
 {
 	u64	msrval;
 
-	rdmsrl(MSR_RMID_SNC_CONFIG, msrval);
-	msrval |= BIT_ULL(0);
-	wrmsrl(MSR_RMID_SNC_CONFIG, msrval);
-}
-
-static void snc_remove_pkg(void)
-{
-	u64	msrval;
-
 	rdmsrl(MSR_RMID_SNC_CONFIG, msrval);
 	msrval &= ~BIT_ULL(0);
 	wrmsrl(MSR_RMID_SNC_CONFIG, msrval);
+pr_info("CPU%d: set MSR_RMID_SNC_CONFIG to 0x%llx\n", raw_smp_processor_id(), msrval);
+}
+
+static void snc_remove_pkg(void)
+{
+	u64	msrval;
+
+	rdmsrl(MSR_RMID_SNC_CONFIG, msrval);
+	msrval |= BIT_ULL(0);
+	wrmsrl(MSR_RMID_SNC_CONFIG, msrval);
+pr_info("CPU%d: set MSR_RMID_SNC_CONFIG to 0x%llx\n", raw_smp_processor_id(), msrval);
 }
 
 /*
@@ -987,6 +989,7 @@ static __init int find_snc_ways(void)
 	if (ret > 1)
 		rdt_resources_all[RDT_RESOURCE_PKG].r_resctrl.pkg_actions = true;
 
+pr_info("SNC_ways = %d\n", ret);
 	return ret;
 }
 
-- 
2.40.1