On 10/14/21 16:45, Kent Overstreet wrote:

On Wed, Oct 13, 2021 at 06:33:06PM +0200, David Hildenbrand wrote:

quoted

quoted

@@ -9317,6 +9319,7 @@ void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
 	struct page *page;
 	struct zone *zone;
 	unsigned int order;
+	unsigned int migratetype;
 	unsigned long flags;
 
 	offline_mem_sections(pfn, end_pfn);

@@ -9346,7 +9349,8 @@ void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
 		BUG_ON(page_count(page));
 		BUG_ON(!PageBuddy(page));
 		order = buddy_order(page);
-		del_page_from_free_list(page, zone, order);
+		migratetype = get_pfnblock_migratetype(page, pfn);

As the free pages are isolated, theoretically this should be
MIGRATE_ISOLATE.

Thanks for noticing that - I somehow missed the fact that pageblock migratetypes
change at runtime,

Not only that. Buddy merging will also merge pages from different
migratetypes. I think that's the main reason why nr_free is not per
migratetype. Your patch has been attempted few times already, e.g. here [1].

[1]
https://lore.kernel.org/all/alpine.DEB.2.10.1611161731350.17379@chino.kir.corp.google.com/ (local)

so my patch is wrong. I'm going to have to rework my patch to
store the migratetype of free pages in the page itself.

Yeah that would be the solution. Will likely bring some overhead to
alloc/free fastpaths, which would have to be worth it, so nobody attempted it.