[RFC][PATCH 5/6] x86/topo: Fix SNC topology mess

[Peter Zijlstra]

So per 4d6dd05d07d0 ("sched/topology: Fix sched domain build error for GNR, CWF in SNC-3 mode")

The original crazy SNC-3 SLIT table was:

node distances:
node     0    1    2    3    4    5
    0:   10   15   17   21   28   26
    1:   15   10   15   23   26   23
    2:   17   15   10   26   23   21
    3:   21   28   26   10   15   17
    4:   23   26   23   15   10   15
    5:   26   23   21   17   15   10

And per:

  https://lore.kernel.org/lkml/20250825075642.GQ3245006@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/

My suggestion was to average the off-trace clusters to restore sanity.

However, 4d6dd05d07d0 implements this under various assumptions:

 - there will never be more than 2 packages;
 - the off-trace cluster will have distance >20

And then HPE shows up with a machine that matches the
Vendor-Family-Model checks but looks like this:

Here's an 8 socket (2 chassis) HPE system with SNC enabled:

node   0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15
  0:  10  12  16  16  16  16  18  18  40  40  40  40  40  40  40  40
  1:  12  10  16  16  16  16  18  18  40  40  40  40  40  40  40  40
  2:  16  16  10  12  18  18  16  16  40  40  40  40  40  40  40  40
  3:  16  16  12  10  18  18  16  16  40  40  40  40  40  40  40  40
  4:  16  16  18  18  10  12  16  16  40  40  40  40  40  40  40  40
  5:  16  16  18  18  12  10  16  16  40  40  40  40  40  40  40  40
  6:  18  18  16  16  16  16  10  12  40  40  40  40  40  40  40  40
  7:  18  18  16  16  16  16  12  10  40  40  40  40  40  40  40  40
  8:  40  40  40  40  40  40  40  40  10  12  16  16  16  16  18  18
  9:  40  40  40  40  40  40  40  40  12  10  16  16  16  16  18  18
 10:  40  40  40  40  40  40  40  40  16  16  10  12  18  18  16  16
 11:  40  40  40  40  40  40  40  40  16  16  12  10  18  18  16  16
 12:  40  40  40  40  40  40  40  40  16  16  18  18  10  12  16  16
 13:  40  40  40  40  40  40  40  40  16  16  18  18  12  10  16  16
 14:  40  40  40  40  40  40  40  40  18  18  16  16  16  16  10  12
 15:  40  40  40  40  40  40  40  40  18  18  16  16  16  16  12  10

 10 = Same chassis and socket
 12 = Same chassis and socket (SNC)
 16 = Same chassis and adjacent socket
 18 = Same chassis and non-adjacent socket
 40 = Different chassis

*However* this is SNC-2.

This completely invalidates all the earlier assumptions and trips
WARNs.

Now that the topology code has a sensible measure of
nodes-per-package, we can use that to divinate the SNC mode at hand,
and only fix up SNC-3 topologies.

With the only assumption that there are no CPU-less nodes -- is this
a valid assumption ?

Fixes: 4d6dd05d07d0 ("sched/topology: Fix sched domain build error for GNR, CWF in SNC-3 mode")
Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
---
 arch/x86/kernel/smpboot.c |   64 +++++++++++++++++-----------------------------
 1 file changed, 25 insertions(+), 39 deletions(-)

--- a/arch/x86/kernel/smpboot.c
+++ b/arch/x86/kernel/smpboot.c
@@ -506,33 +506,32 @@ static void __init build_sched_topology(
 }
 
 #ifdef CONFIG_NUMA
-static int sched_avg_remote_distance;
-static int avg_remote_numa_distance(void)
+static int slit_cluster_distance(int i, int j)
 {
-	int i, j;
-	int distance, nr_remote, total_distance;
-
-	if (sched_avg_remote_distance > 0)
-		return sched_avg_remote_distance;
-
-	nr_remote = 0;
-	total_distance = 0;
-	for_each_node_state(i, N_CPU) {
-		for_each_node_state(j, N_CPU) {
-			distance = node_distance(i, j);
-
-			if (distance >= REMOTE_DISTANCE) {
-				nr_remote++;
-				total_distance += distance;
-			}
+	int u = __num_nodes_per_package;
+	long d = 0;
+	int x, y;
+
+	/*
+	 * Is this a unit cluster on the trace?
+	 */
+	if ((i / u) == (j / u))
+		return node_distance(i, j);
+
+	/*
+	 * Off-trace cluster, return average of the cluster to force symmetry.
+	 */
+	x = i - (i % u);
+	y = j - (j % u);
+
+	for (i = x; i < x + u; i++) {
+		for (j = y; j < y + u; j++) {
+			d += node_distance(i, j);
+			d += node_distance(j, i);
 		}
 	}
-	if (nr_remote)
-		sched_avg_remote_distance = total_distance / nr_remote;
-	else
-		sched_avg_remote_distance = REMOTE_DISTANCE;
 
-	return sched_avg_remote_distance;
+	return d / (2*u*u);
 }
 
 int arch_sched_node_distance(int from, int to)
@@ -542,13 +541,11 @@ int arch_sched_node_distance(int from, i
 	switch (boot_cpu_data.x86_vfm) {
 	case INTEL_GRANITERAPIDS_X:
 	case INTEL_ATOM_DARKMONT_X:
-
-		if (topology_max_packages() == 1 || __num_nodes_per_package == 1 ||
-		    d < REMOTE_DISTANCE)
+		if (topology_max_packages() == 1 || __num_nodes_per_package < 3)
 			return d;
 
 		/*
-		 * With SNC enabled, there could be too many levels of remote
+		 * With SNC-3 enabled, there could be too many levels of remote
 		 * NUMA node distances, creating NUMA domain levels
 		 * including local nodes and partial remote nodes.
 		 *
@@ -557,19 +554,8 @@ int arch_sched_node_distance(int from, i
 		 * in the remote package in the same sched group.
 		 * Simplify NUMA domains and avoid extra NUMA levels including
 		 * different remote NUMA nodes and local nodes.
-		 *
-		 * GNR and CWF don't expect systems with more than 2 packages
-		 * and more than 2 hops between packages. Single average remote
-		 * distance won't be appropriate if there are more than 2
-		 * packages as average distance to different remote packages
-		 * could be different.
 		 */
-		WARN_ONCE(topology_max_packages() > 2,
-			  "sched: Expect only up to 2 packages for GNR or CWF, "
-			  "but saw %d packages when building sched domains.",
-			  topology_max_packages());
-
-		d = avg_remote_numa_distance();
+		return slit_cluster_distance(from, to);
 	}
 	return d;
 }