Re: [PATCH v2] gfp.h: GFP_THISNODE can go to other nodes if some are unpopulated

[Christoph Lameter <clameter@sgi.com>]

On Mon, 11 Jun 2007, Lee Schermerhorn wrote:

> When the hugepages patch was evolving, I suggested that we might want to
> export the "populated map" to applications so that they could ask to
> bind to or interleave across only populated nodes.  We never pursued
> that.  Maybe just eliminate nodes that are unpopulated in the "policy
> zone" from the node masks for MPOL_BIND and MPOL_INTERLEAVE in the
> system calls?  Saves checking the populated node set in the allocation
> paths.  Would need appropriate error return if this resulted in empty
> nodemask.

That would work for the MPOL_BIND case since it has a zonelist. However, 
MPOL_INTERLEAVE does not have a zonelist. I think we need the populated 
map for interleave. The hacky way in how I checked for an unpopulated 
node in the patch just posted is not that effective.

> Of course, memory hotplug could result in nodes becoming empty after the
> nodemasks are adjusted, so we probably can't avoid checks in the
> allocation paths if we want to avoid the bind and interleave issues you
> mention above.

Right.

--
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>