Hi, tj

On 09/11/2015 03:29 AM, Tejun Heo wrote:

Hello,

On Thu, Sep 10, 2015 at 12:27:45PM +0800, Tang Chen wrote:

quoted

diff --git a/include/linux/gfp.h b/include/linux/gfp.h
index ad35f30..1a1324f 100644
--- a/include/linux/gfp.h
+++ b/include/linux/gfp.h

@@ -307,13 +307,19 @@ static inline struct page *alloc_pages_node(int nid, gfp_t gfp_mask,
  	if (nid < 0)
  		nid = numa_node_id();
  
+	if (!node_online(nid))
+		nid = get_near_online_node(nid);
+
  	return __alloc_pages(gfp_mask, order, node_zonelist(nid, gfp_mask));
  }

Why not just update node_data[]->node_zonelist in the first place?

zonelist will be rebuilt in __offline_pages() when the zone is not 
populated any more.

Here, getting the best near online node is for those cpus on memory-less 
nodes.

In the original code, if nid is NUMA_NO_NODE, the node the current cpu 
resides in
will be chosen. And if the node is memory-less node, the cpu will be 
mapped to its
best near online node.

But this patch-set will map the cpu to its original node, so 
numa_node_id() may return
a memory-less node to allocator. And then memory allocation may fail.

Also, what's the synchronization rule here?  How are allocators
synchronized against node hot [un]plugs?

The rule is: node_to_near_node_map[] array will be updated each time 
node [un]hotplug happens.

Now it is not protected by a lock. But I think acquiring a lock may 
cause performance regression
to memory allocator.

When rebuilding zonelist, stop_machine is used. So I think maybe 
updating the
node_to_near_node_map[] array at the same time when zonelist is rebuilt 
could be a good idea.

Thanks.



--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>