Skip to contentSkip to conversation

Fix hypervisor crash with unpopulated NUMA nodes

PATCH In review Andre Przywara 2 participants 3 messages Oct 7, 2009

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Fix hypervisor crash with unpopulated NUMA nodes

Hi Keir,

on NUMA systems with memory-less nodes Xen crashes quite early in the hypervisor (while initializing the heaps). This is not an issue if this happens to be the last node, but "inner" nodes trigger this reliably. On multi-node processors it is much more likely to leave a node unequipped.
The attached patch fixes this by enumerating the node via the node_online_map instead of counting from 0 to num_nodes.

The resulting NUMA setup is still somewhat strange, but at least it does not crash. In lowlevel/xc/xc.c there is again this enumeration bug, but I suppose we cannot access the HV's node_online_map from this context, so the xm info output is not correct (but xm debug-keys H is). I plan to rework the handling of memory-less nodes later.

Please also consider committing this to the -testing trees.

Regards,
Andre.

Signature · 9 lines
Andre Przywara
AMD-Operating System Research Center (OSRC), Dresden, Germany
Tel: +49 351 448 3567 12
----to satisfy European Law for business letters:
Advanced Micro Devices GmbH
Karl-Hammerschmidt-Str. 34, 85609 Dornach b. Muenchen
Geschaeftsfuehrer: Andrew Bowd; Thomas M. McCoy; Giuliano Meroni
Sitz: Dornach, Gemeinde Aschheim, Landkreis Muenchen
Registergericht Muenchen, HRB Nr. 43632
xen/common/page_alloc.c +5 -6
diff -r 440d22c01652 xen/common/page_alloc.c
@@ -294,7 +294,6 @@
         node = cpu_to_node(smp_processor_id());
 
     ASSERT(node >= 0);
-    ASSERT(node < num_nodes);
     ASSERT(zone_lo <= zone_hi);
     ASSERT(zone_hi < NR_ZONES);
 
@@ -323,8 +322,9 @@
         } while ( zone-- > zone_lo ); /* careful: unsigned zone may wrap */
 
         /* Pick next node, wrapping around if needed. */
-        if ( ++node == num_nodes )
-            node = 0;
+        node = next_node(node, node_online_map);
+        if (node == MAX_NUMNODES)
+            node = first_node(node_online_map);
     }
 
     /* Try to free memory from tmem */
@@ -466,7 +466,6 @@
 
     ASSERT(order <= MAX_ORDER);
     ASSERT(node >= 0);
-    ASSERT(node < num_online_nodes());
 
     for ( i = 0; i < (1 << order); i++ )
     {
@@ -817,13 +816,13 @@
 static unsigned long avail_heap_pages(
     unsigned int zone_lo, unsigned int zone_hi, unsigned int node)
 {
-    unsigned int i, zone, num_nodes = num_online_nodes();
+    unsigned int i, zone;
     unsigned long free_pages = 0;
 
     if ( zone_hi >= NR_ZONES )
         zone_hi = NR_ZONES - 1;
 
-    for ( i = 0; i < num_nodes; i++ )
+    for_each_online_node(i)
     {
         if ( !avail[i] )
             continue;

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

#1 lore raw
Quoted 7 lines
Quoted 1 line
Quoted 1 line

Andre Przywara <[email protected]> 07.10.09 11:45 >>>

on NUMA systems with memory-less nodes Xen crashes quite early in the hypervisor (while initializing the heaps). This is not an issue if this happens to be the last node, but "inner" nodes trigger this reliably. On multi-node processors it is much more likely to leave a node unequipped. The attached patch fixes this by enumerating the node via the node_online_map instead of counting from 0 to num_nodes.

While I do not see anything wrong with the patch, I still wonder why it would be needed: It seems to indicate that node_online_map represents only nodes with memory, but imo should be representing nodes with memory or processors (leaving aside pure I/O nodes for the moment). So perhaps there's rather a problem with the setup of node_online_map somewhere?

Jan

#2 lore raw
Quoted 10 lines · Jan Beulich
Quoted 4 lines
Quoted 1 line
Quoted 1 line

Andre Przywara <[email protected]> 07.10.09 11:45 >>>

on NUMA systems with memory-less nodes Xen crashes quite early in the hypervisor (while initializing the heaps). This is not an issue if this happens to be the last node, but "inner" nodes trigger this reliably. On multi-node processors it is much more likely to leave a node unequipped.
The attached patch fixes this by enumerating the node via the node_online_map instead of counting from 0 to num_nodes.

While I do not see anything wrong with the patch, I still wonder why it would be needed: It seems to indicate that node_online_map represents only nodes with memory, but imo should be representing nodes with memory or processors (leaving aside pure I/O nodes for the moment). So perhaps there's rather a problem with the setup of node_online_map somewhere?

Yes, because the map creation is callback driven by ACPI code.
The BIOS of my machine is omitting the memory entries for memory-less nodes, so there is no callback triggered for these nodes. Nevertheless Xen uses the SRAT provided node numbers, this creates the hole.
(My setup: 2 + 0 + 2 + 0 GB per node, Xen sees two nodes named 0 and 2).
I agree that should be changed (that is what I meant with "will rework later"), not only because the "lonely" cores will simply be added to another node. But since I will be not in the office for the next two weeks I would like to get this patch applied for the time being.

Regards,
Andre.

Signature · 9 lines
Andre Przywara
AMD-Operating System Research Center (OSRC), Dresden, Germany
Tel: +49 351 448 3567 12
----to satisfy European Law for business letters:
Advanced Micro Devices GmbH
Karl-Hammerschmidt-Str. 34, 85609 Dornach b. Muenchen
Geschaeftsfuehrer: Andrew Bowd; Thomas M. McCoy; Giuliano Meroni
Sitz: Dornach, Gemeinde Aschheim, Landkreis Muenchen
Registergericht Muenchen, HRB Nr. 43632