On Thu 25-03-21 16:35:58, Michal Hocko wrote:
[...]

So there is indeed a difference. One way around that would be to mark
vmemmap pages (e.g. PageReserved && magic value stored somewhere in the
struct page - resembling bootmem vmemmaps) or mark section fully backing
vmemmaps as online (ugly).

I am not yet ready to give up on this. Here is a quick stab at the
pfn_to_online_page approach. It is not great but it is not really
terrible either. I think we can do better and skip
find_memory_block_by_id in most cases because nr_vmemmap_pages should be
constant with our current implementation.

diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
index 754026a9164d..0d2bc22c29d3 100644
--- a/mm/memory_hotplug.c
+++ b/mm/memory_hotplug.c

@@ -304,8 +304,20 @@ struct page *pfn_to_online_page(unsigned long pfn)
 		return NULL;
 
 	ms = __nr_to_section(nr);
-	if (!online_section(ms))
+	if (!online_section(ms)) {
+		if (memmap_on_memory) {
+			struct memory_block *mem;
+			unsigned long section_pfn = section_nr_to_pfn(nr);
+
+			mem = find_memory_block_by_id(memory_block_id(nr));
+			if (!mem || !mem->nr_vmemmap_pages || pfn - section_pfn > mem->nr_vmemmap_pages)
+				return NULL;
+
+			return pfn_to_page(pfn);
+
+		}
 		return NULL;
+	}
 
 	/*
 	 * Save some code text when online_section() +

-- 
Michal Hocko
SUSE Labs