On Fri, Feb 12, 2021 at 10:56:19AM +0100, David Hildenbrand wrote:

On 12.02.21 10:55, David Hildenbrand wrote:

quoted

On 08.02.21 12:08, Mike Rapoport wrote:

quoted

+#ifdef CONFIG_SPARSEMEM
+	/*
+	 * Sections in the memory map may not match actual populated
+	 * memory, extend the node span to cover the entire section.
+	 */
+	*start_pfn = round_down(*start_pfn, PAGES_PER_SECTION);
+	*end_pfn = round_up(*end_pfn, PAGES_PER_SECTION);

Does that mean that we might create overlapping zones when one node

s/overlapping zones/overlapping nodes/

quoted

starts in the middle of a section and the other one ends in the middle
of a section?

quoted

Could it be a problem? (e.g., would we have to look at neighboring nodes
when making the decision to extend, and how far to extend?)

Having a node end/start in a middle of a section would be a problem, but in
this case I don't see a way to detect how a node should be extended :(

We can return to a v4 [1] without x86 modifications.
With that we'll have struct pages corresponding to a hole in a middle of a
zone with correct zone link and a good guess for the node.

As for the pfn 0 on x86, it'll remain outside any node and zone, but since
it was the case since, like forever, I think we can live with it.

[1] https://lore.kernel.org/lkml/20210130221035.4169-1-rppt@kernel.org (local)

-- 
Sincerely yours,
Mike.