Re: [PATCH] Add nid sanity on alloc_pages_node

[Hugh Dickins <hugh@veritas.com>]

On Thu, 12 Jul 2007, Andrew Morton wrote:
> 
> It'd be much better to fix the race within alloc_fresh_huge_page().  That
> function is pretty pathetic.
> 
> Something like this?
> 
> --- a/mm/hugetlb.c~a
> +++ a/mm/hugetlb.c
> @@ -105,13 +105,20 @@ static void free_huge_page(struct page *
>  
>  static int alloc_fresh_huge_page(void)
>  {
> -	static int nid = 0;
> +	static int prev_nid;
> +	static DEFINE_SPINLOCK(nid_lock);
>  	struct page *page;
> -	page = alloc_pages_node(nid, htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
> -					HUGETLB_PAGE_ORDER);
> -	nid = next_node(nid, node_online_map);
> +	int nid;
> +
> +	spin_lock(&nid_lock);
> +	nid = next_node(prev_nid, node_online_map);
>  	if (nid == MAX_NUMNODES)
>  		nid = first_node(node_online_map);
> +	prev_nid = nid;
> +	spin_unlock(&nid_lock);
> +
> +	page = alloc_pages_node(nid, htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
> +					HUGETLB_PAGE_ORDER);
>  	if (page) {
>  		set_compound_page_dtor(page, free_huge_page);
>  		spin_lock(&hugetlb_lock);

Now that it's gone into the tree, I look at it and wonder, does your
nid_lock really serve any purpose?  We're just doing a simple assignment
to prev_nid, and it doesn't matter if occasionally two racers choose the
same node, and there's no protection here against a node being offlined
before the alloc_pages_node anyway (unsupported? I'm ignorant).

Hugh