Re: [06/44] numa: fix slab_node(MPOL_BIND)

[Lee Schermerhorn <Lee.Schermerhorn@hp.com>]

On Wed, 2010-12-08 at 05:33 +0100, Eric Dumazet wrote:
> Le mardi 07 décembre 2010 à 22:03 -0500, Lee Schermerhorn a écrit :
> > On Tue, 2010-12-07 at 16:04 -0800, Greg KH wrote:
> > > 2.6.27-stable review patch.  If anyone has any objections, please let us know.
> > > 
> > > ------------------
> > > 
> > > From: Eric Dumazet <eric.dumazet@gmail.com>
> > > 
> > > commit 800416f799e0723635ac2d720ad4449917a1481c upstream.
> > > 
> 
> > > 
> > > --- a/mm/mempolicy.c
> > > +++ b/mm/mempolicy.c
> > > @@ -1404,7 +1404,7 @@ unsigned slab_node(struct mempolicy *pol
> > >  		(void)first_zones_zonelist(zonelist, highest_zoneidx,
> > >  							&policy->v.nodes,
> > >  							&zone);
> > > -		return zone->node;
> > > +		return zone ? zone->node : numa_node_id();
> > 
> > I think this should be numa_mem_id().  Given the documented purpose of
> > slab_node(), we want a node from which page allocation is likely to
> > succeed.  numa_node_id() can return a memoryless node for, e.g.,  some
> > configurations of some HP ia64 platforms.  numa_mem_id() was introduced
> > to return that same node from which "local" mempolicy would allocate
> > pages.
> 
> Hmm... numa_mem_id() was introduced in 2.6.35 as an optimization.
> 
> When I did this patch (to fix a bug), mm/mempolicy.c only contained
> calls to numa_node_id() (and still is today)

Sometimes you want numa_node_id()--e.g., for use with a mempolicy-based
allocation that allows fallback.  When the node id will be used for a
'_THIS_NODE allocation, numa_mem_id() is preferred as it will always
return a node that contains or contained--maybe now oom--memory.  It's
the same as numa_node_id() on platforms that don't expose memoryless
nodes.  

> 
> By the way, anybody knows how I can emulate a memoryless node on a dual
> node x86_64 machine (with memory present on both nodes) ?
> 

You can use the mem= boot parameter and specify the amount of memory on
the 1st/boot node.  Or you can use the memmap parameter to reserve the
memory on the 2nd/non-boot node.  With the memmap parameter, you can
reserve the memory of nodes other than the highest numbered
one[s]--e.g., on a >2 node platform.  However, you'll probably a patch
to see the cpus on any node that you hide using memmap.  I have such a
patch if you're interested in going that route.

You can also reduce the amount of memory on any/each node by reserving
ranges of physical memory with memmap.  Use the 'SRAT.*PXM' boot
messages to find the nodes' physical memory ranges and reserve how ever
much you want off the top of the nodes.

Lee