Re: [PATCH 6/6] Use one zonelist that is filtered by nodemask

[mel@skynet.ie (Mel Gorman)]

On (09/11/07 07:45), Christoph Lameter didst pronounce:
> On Fri, 9 Nov 2007, Mel Gorman wrote:
> 
> >  struct page * fastcall
> >  __alloc_pages(gfp_t gfp_mask, unsigned int order,
> >  		struct zonelist *zonelist)
> >  {
> > +	/*
> > +	 * Use a temporary nodemask for __GFP_THISNODE allocations. If the
> > +	 * cost of allocating on the stack or the stack usage becomes
> > +	 * noticable, allocate the nodemasks per node at boot or compile time
> > +	 */
> > +	if (unlikely(gfp_mask & __GFP_THISNODE)) {
> > +		nodemask_t nodemask;
> 
> Hmmm.. This places a potentially big structure on the stack. nodemask can 
> contain up to 1024 bits which means 128 bytes. Maybe keep an array of 
> gfp_thisnode nodemasks (node_nodemask?) and use node_nodemask[nid]?
> 

That is what I was hinting at in the comment as a possible solution.

> > +
> > +		return __alloc_pages_internal(gfp_mask, order,
> > +			zonelist, nodemask_thisnode(numa_node_id(), &nodemask));
> 
> Argh.... GFP_THISNODE must use the nid passed to alloc_pages_node and 
> *not* the local numa node id. Only if the node specified to alloc_pages 
> nodes is -1 will this work.
> 

alloc_pages_node() calls __alloc_pages_nodemask() though where in this
function if I'm reading it right is called without a node id. Given no
other details on the nid, the current one seemed a logical choice.

What I did notice when rechecking is I left the warning about THISNODE
in by accident :(

-- 
-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab