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

[Peter Zijlstra]

On Thu, Feb 26, 2026 at 11:00:38AM -0800, Tim Chen wrote:
> May be a good idea to sanity check that the nodes in the first unit cluster
> has the same package id and give a WARNING if that's not the case.

But then we'd also have check the second cluster is another package. And
if we're checking that, we might as well check they're symmetric.

Is this sufficiently paranoid for you? :-)


---
--- a/arch/x86/kernel/smpboot.c
+++ b/arch/x86/kernel/smpboot.c
@@ -61,6 +61,7 @@
 #include <linux/cpuhotplug.h>
 #include <linux/mc146818rtc.h>
 #include <linux/acpi.h>
+#include <linux/once_lite.h>
 
 #include <asm/acpi.h>
 #include <asm/cacheinfo.h>
@@ -506,12 +507,58 @@ static void __init build_sched_topology(
 }
 
 #ifdef CONFIG_NUMA
+static bool slit_cluster_symmetric(int N)
+{
+	for (int k = 0; k < __num_nodes_per_package; k++) {
+		for (int l = k; l < __num_nodes_per_package; l++) {
+			if (node_distance(N + k, N + l) != 
+			    node_distance(N + l, N + k))
+				return false;
+		}
+	}
+
+	return true;
+}
+
+static u32 slit_cluster_package(int N)
+{
+	u32 pkg_id = ~0;
+
+	for (int n = 0; n < __num_nodes_per_package; n++) {
+		const struct cpumask *cpus = cpumask_of_node(N + n);
+		int cpu;
+
+		for_each_cpu(cpu, cpus) {
+			u32 id = topology_logical_package_id(cpu);
+			if (pkg_id == ~0)
+				pkg_id = id;
+			if (pkg_id != id)
+				return ~0;
+		}
+	}
+
+	return pkg_id;
+}
+
+/* If you NUMA_EMU on top of SNC, you get to keep the pieces */
+static void slit_validate(void)
+{
+	u32 pkg1 = slit_cluster_package(0);
+	u32 pkg2 = slit_cluster_package(__num_nodes_per_package);
+	WARN_ON(pkg1 == ~0 || pkg2 == ~0 || pkg1 == pkg2);
+
+	WARN_ON(!slit_cluster_symmetric(0));
+	WARN_ON(!slit_cluster_symmetric(__num_nodes_per_package));
+}
+
 static int slit_cluster_distance(int i, int j)
 {
 	int u = __num_nodes_per_package;
 	long d = 0;
 	int x, y;
 
+	DO_ONCE_LITE(slit_validate);
+
 	/*
 	 * Is this a unit cluster on the trace?
 	 */