On 9 December 2016 at 14:52, Robert Richter [off-list ref] wrote:

On 09.12.16 21:15:12, Yisheng Xie wrote:

quoted

For invalid pages, their zone and node information is not initialized, and it
do have risk to trigger the BUG_ON, so I have a silly question,
why not just change the BUG_ON:

We need to get the page handling correct. Modifying the BUG_ON() just
hides that something is wrong.

Actually, I think this is a reasonable question. We are trying very
hard to work around the BUG_ON(), which arguably does something wrong
by calling page_to_nid() on a struct page without checking if it is a
valid page.

Looking at commit 344c790e3821 ("mm: make setup_zone_migrate_reserve()
aware of overlapping nodes"), the BUG_ON() has a specific purpose
related to adjacent zones.

What will go wrong if we ignore this check?

quoted

-----------

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 6de9440..af199b8 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c

@@ -1860,12 +1860,13 @@ int move_freepages(struct zone *zone,
         * Remove at a later date when no bug reports exist related to
         * grouping pages by mobility
         */
-       VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
+       VM_BUG_ON(early_pfn_valid(start_page) && early_pfn_valid(end_page) &&
+                       page_zone(start_page) != page_zone(end_page));
 #endif

        for (page = start_page; page <= end_page;) {
                /* Make sure we are not inadvertently changing nodes */
-               VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
+               VM_BUG_ON_PAGE(early_pfn_valid(page) && (page_to_nid(page) != zone_to_nid(zone)), page);

                if (!pfn_valid_within(page_to_pfn(page))) {
                        page++;