[PATCH] ia64: default the NUMA node distance when there is no ACPI SLIT

[PATCH] ia64: default the NUMA node distance when there is no ACPI SLIT

[Andy Whitcroft]

If during boot no ACPI 2.0 SLIT table is found we do not initialise
the NUMA node distance table, leaving all node distances as 0.
This causes IA-64 to fail the regression tests for libnuma.  Also,
when we have a bad formatted table we will set all node distances
to 10.  Both of these forms lose all NUMA information by indicating
all nodes are as close as the "local node", and are inconsistant
in their values.

When an architecture does not offer NUMA topology support the default
node_distance is a distance of LOCAL_DISTANCE (10) to the local node
and REMOTE_DISTANCE (20) to all others.  It seems more correct to
follow this model when the SLIT information is missing or corrupt.

This patch updates the SLIT failure paths to fill in the NUMA node
distance table with this default.

Signed-off-by: Andy Whitcroft <apw@shadowen.org>
---
 arch/ia64/kernel/acpi.c |   62 ++++++++++++++++++++++++++++++++++------------
 1 files changed, 46 insertions(+), 16 deletions(-)
diff --git a/arch/ia64/kernel/acpi.c b/arch/ia64/kernel/acpi.c
index 3d45d24..3cd3e5b 100644
--- a/arch/ia64/kernel/acpi.c
+++ b/arch/ia64/kernel/acpi.c
@@ -445,7 +445,6 @@ void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
 		printk(KERN_ERR
 		       "ACPI 2.0 SLIT: size mismatch: %d expected, %d actual\n",
 		       len, slit->header.length);
-		memset(numa_slit, 10, sizeof(numa_slit));
 		return;
 	}
 	slit_table = slit;
@@ -507,9 +506,50 @@ acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma)
 	num_node_memblks++;
 }
 
+void __init acpi_numa_distance_from_slit(void)
+{
+	int i, j;
+	int node_from;
+	int node_to;
+	int nodes = slit_table->locality_count;
+
+	printk(KERN_INFO "Building NUMA distance from ACPI 2.0 SLIT\n");
+
+	for (i = 0; i < nodes; i++) {
+		if (!pxm_bit_test(i))
+			continue;
+		node_from = pxm_to_node(i);
+		for (j = 0; j < nodes; j++) {
+			if (!pxm_bit_test(j))
+				continue;
+			node_to = pxm_to_node(j);
+			node_distance(node_from, node_to) +					  slit_table->entry[i * nodes + j];
+		}
+	}
+}
+
+void __init acpi_numa_distance_default(void)
+{
+	int i, j;
+	int node_from;
+	int node_to;
+
+	printk(KERN_INFO "No SLIT table, defaulting NUMA distance\n");
+
+	for_each_online_node(i) {
+		node_from = pxm_to_node(i);
+		for_each_online_node(j) {
+			node_to = pxm_to_node(j);
+			node_distance(node_from, node_to) = (i = j)?
+				LOCAL_DISTANCE : REMOTE_DISTANCE;
+		}
+	}
+}
+
 void __init acpi_numa_arch_fixup(void)
 {
-	int i, j, node_from, node_to;
+	int i, j;
 
 	/* If there's no SRAT, fix the phys_id and mark node 0 online */
 	if (srat_num_cpus = 0) {
@@ -553,21 +593,11 @@ void __init acpi_numa_arch_fixup(void)
 	printk(KERN_INFO "Number of memory chunks in system = %d\n",
 	       num_node_memblks);
 
-	if (!slit_table)
-		return;
 	memset(numa_slit, -1, sizeof(numa_slit));
-	for (i = 0; i < slit_table->locality_count; i++) {
-		if (!pxm_bit_test(i))
-			continue;
-		node_from = pxm_to_node(i);
-		for (j = 0; j < slit_table->locality_count; j++) {
-			if (!pxm_bit_test(j))
-				continue;
-			node_to = pxm_to_node(j);
-			node_distance(node_from, node_to) -			    slit_table->entry[i * slit_table->locality_count + j];
-		}
-	}
+	if (slit_table)
+		acpi_numa_distance_from_slit();
+	else
+		acpi_numa_distance_default();
 
 #ifdef SLIT_DEBUG
 	printk("ACPI 2.0 SLIT locality table:\n");

RE: [PATCH] ia64: default the NUMA node distance when there is no ACPI SLIT

[Luck, Tony]

+	printk(KERN_INFO "Building NUMA distance from ACPI 2.0 SLIT\n");

This printk just looks like noise during boot.  Surely this
is normal behavior on a NUMA system?

+	printk(KERN_INFO "No SLIT table, defaulting NUMA distance\n");

But this one deserves more prominence than just KERN_INFO.  I'd
say that it deserves a KERN_ERR rating (without accurate information
we'll make some highly questionable scheduling and memory placement
decisions).  On the other hand the only way for the end-user to fix
this is to complain to their system vendor to fix firmware.  So perhaps
it is just a KERN_NOTICE.

-Tony