On 28.09.2007 [15:25:27 +0100], Mel Gorman wrote:
quoted text
> 
> Two zonelists exist so that GFP_THISNODE allocations will be guaranteed
> to use memory only from a node local to the CPU. As we can now filter the
> zonelist based on a nodemask, we filter the standard node zonelist for zones
> on the local node when GFP_THISNODE is specified.
> 
> When GFP_THISNODE is used, a temporary nodemask is created with only the
> node local to the CPU set. This allows us to eliminate the second zonelist.
> 
> Signed-off-by: Mel Gorman <mel@csn.ul.ie>
> Acked-by: Christoph Lameter <clameter@sgi.com>

<snip>

quoted text
> diff -rup -X /usr/src/patchset-0.6/bin//dontdiff linux-2.6.23-rc8-mm2-030_filter_nodemask/include/linux/gfp.h linux-2.6.23-rc8-mm2-040_use_one_zonelist/include/linux/gfp.h
> --- linux-2.6.23-rc8-mm2-030_filter_nodemask/include/linux/gfp.h	2007-09-28 15:49:57.000000000 +0100
> +++ linux-2.6.23-rc8-mm2-040_use_one_zonelist/include/linux/gfp.h	2007-09-28 15:55:03.000000000 +0100

[Reordering the chunks to make my comments a little more logical]

<snip>

quoted text
> -static inline struct zonelist *node_zonelist(int nid, gfp_t flags)
> +static inline struct zonelist *node_zonelist(int nid)
>  {
> -	return NODE_DATA(nid)->node_zonelists + gfp_zonelist(flags);
> +	return &NODE_DATA(nid)->node_zonelist;
>  }
> 
>  #ifndef HAVE_ARCH_FREE_PAGE
> @@ -198,7 +186,7 @@ static inline struct page *alloc_pages_n
>  	if (nid < 0)
>  		nid = numa_node_id();
> 
> -	return __alloc_pages(gfp_mask, order, node_zonelist(nid, gfp_mask));
> +	return __alloc_pages(gfp_mask, order, node_zonelist(nid));
>  }

This is alloc_pages_node(), and converting the nid to a zonelist means
that lower levels (specifically __alloc_pages() here) are not aware of
nids, as far as I can tell. This isn't a change, I just want to make
sure I understand...

<snip>

quoted text
>  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;
> +
> +		return __alloc_pages_internal(gfp_mask, order,
> +				zonelist, nodemask_thisnode(&nodemask));
> +	}
> +
>  	return __alloc_pages_internal(gfp_mask, order, zonelist, NULL);
>  }

<snip>

So alloc_pages_node() calls here and for THISNODE allocations, we go ask
nodemask_thisnode() for a nodemask...

quoted text
> +static nodemask_t *nodemask_thisnode(nodemask_t *nodemask)
> +{
> +	/* Build a nodemask for just this node */
> +	int nid = numa_node_id();
> +
> +	nodes_clear(*nodemask);
> +	node_set(nid, *nodemask);
> +
> +	return nodemask;
> +}

<snip>

And nodemask_thisnode() always gives us a nodemask with only the node
the current process is running on set, I think?

That seems really wrong -- and would explain what Lee was seeing while
using my patches for the hugetlb pool allocator to use THISNODE
allocations. All the allocations would end up coming from whatever node
the process happened to be running on. This obviously messes up hugetlb
accounting, as I rely on THISNODE requests returning NULL if they go
off-node.

I'm not sure how this would be fixed, as __alloc_pages() no longer has
the nid to set in the mask.

Am I wrong in my analysis?

Thanks,
Nish

-- 
Nishanth Aravamudan <nacc@us.ibm.com>
IBM Linux Technology Center
-
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/