On Tue, 22 May 2012 13:03:54 +1000 Stephen Rothwell [off-list ref] wrote:

On Mon, 21 May 2012 18:53:37 -0700 (PDT) David Rientjes [off-list ref] wrote:

quoted

Yeah, it's sched/numa since that's what introduced numa_init().  It does 
for_each_node() for each node and does a kmalloc_node() even though that 
node may not be online.  Slub ends up passing this node to the page 
allocator through alloc_pages_exact_node().  CONFIG_DEBUG_VM would have 
caught this and your config confirms its not enabled.

sched/numa either needs a memory hotplug notifier or it needs to pass 
NUMA_NO_NODE for nodes that aren't online.  Until we get the former, the 
following should fix it.


sched, numa: Allocate node_queue on any node for offline nodes

struct node_queue must be allocated with NUMA_NO_NODE for nodes that are 
not (yet) online, otherwise the page allocator has a bad zonelist.

Signed-off-by: David Rientjes <rientjes@google.com>

Thanks, that fixes it.

Tested-by: Stephen Rothwell <redacted>

And I will put that patch in linux-next until it (or something better)
appears.

-- 
Cheers,
Stephen Rothwell                    sfr@canb.auug.org.au