"Bad page state" while freeing gigantic pages

The following "Bad page state" occurs on s390 when freeing gigantic pages:

[  276.681603] BUG: Bad page state in process bash  pfn:380001
[  276.681614] page:00000000c35f0856 refcount:0 mapcount:0 mapping:00000000126b68aa index:0x0 pfn:0x380001
[  276.681620] aops:0x0
[  276.681622] flags: 0x3ffff00000000000()
[  276.681626] raw: 3ffff00000000000 0000000000000100 0000000000000122 0000000100000000
[  276.681628] raw: 0000000000000000 0000000000000000 ffffffff00000000 0000000000000000
[  276.681630] page dumped because: non-NULL mapping
[  276.681632] Modules linked in:
[  276.681637] CPU: 6 PID: 616 Comm: bash Not tainted 5.10.0-rc7-next-20201208 #1
[  276.681639] Hardware name: IBM 3906 M03 703 (LPAR)
[  276.681641] Call Trace:
[  276.681648]  [<0000000458c252b6>] show_stack+0x6e/0xe8
[  276.681652]  [<000000045971cf60>] dump_stack+0x90/0xc8
[  276.681656]  [<0000000458e8b186>] bad_page+0xd6/0x130
[  276.681658]  [<0000000458e8cdea>] free_pcppages_bulk+0x26a/0x800
[  276.681661]  [<0000000458e8e67e>] free_unref_page+0x6e/0x90
[  276.681663]  [<0000000458e8ea6c>] free_contig_range+0x94/0xe8
[  276.681666]  [<0000000458ea5e54>] update_and_free_page+0x1c4/0x2c8
[  276.681669]  [<0000000458ea784e>] free_pool_huge_page+0x11e/0x138
[  276.681671]  [<0000000458ea8530>] set_max_huge_pages+0x228/0x300
[  276.681673]  [<0000000458ea86c0>] nr_hugepages_store_common+0xb8/0x130
[  276.681678]  [<0000000458fd5b6a>] kernfs_fop_write+0xd2/0x218
[  276.681681]  [<0000000458ef9da0>] vfs_write+0xb0/0x2b8
[  276.681684]  [<0000000458efa15c>] ksys_write+0xac/0xe0
[  276.681687]  [<000000045972c5ca>] system_call+0xe6/0x288
[  276.681730] Disabling lock debugging due to kernel taint

I bisected it to commit 1378a5ee451a ("mm: store compound_nr as well as compound_order"), and it seems that the new compound_nr overlaying page->mapping is not properly cleared, which then triggers the non-NULL mapping warning.

This is because only the compound_order is cleared in destroy_compound_gigantic_page(), and compound_nr is set to 1U << order == 1 for order 0 in set_compound_order(page, 0).

For some reason, I can not reproduce this on x86, but I do not see where this could be an arch-sepcific issue. Still, I might be missing something, and my proposed patch also looks a bit ugly (at least to me), hence this RFC. Any comments?

BTW, for "normal sized" hugepages, this is not an issue, as page->mapping seems to be cleared explicitly in this case, in free_tail_pages_check(), but the freeing path for normal hugepages is quite different from that for gigantic pages using free_contig_range(). So a "page[1].mapping = NULL" might also be an option, instead of the "page[1].compound_nr = 0" in my patch, but that looks even more ugly, since it would clear more than needed.

Gerald Schaefer (1):
mm/hugetlb: clear compound_nr before freeing gigantic pages

 mm/hugetlb.c | 1 +
 1 file changed, 1 insertion(+)

2.17.1

Commit 1378a5ee451a ("mm: store compound_nr as well as compound_order") added compound_nr counter to first tail struct page, overlaying with page->mapping. The overlay itself is fine, but while freeing gigantic hugepages via free_contig_range(), a "bad page" check will trigger for non-NULL page->mapping on the first tail page:

[  276.681603] BUG: Bad page state in process bash  pfn:380001
[  276.681614] page:00000000c35f0856 refcount:0 mapcount:0 mapping:00000000126b68aa index:0x0 pfn:0x380001
[  276.681620] aops:0x0
[  276.681622] flags: 0x3ffff00000000000()
[  276.681626] raw: 3ffff00000000000 0000000000000100 0000000000000122 0000000100000000
[  276.681628] raw: 0000000000000000 0000000000000000 ffffffff00000000 0000000000000000
[  276.681630] page dumped because: non-NULL mapping
[  276.681632] Modules linked in:
[  276.681637] CPU: 6 PID: 616 Comm: bash Not tainted 5.10.0-rc7-next-20201208 #1
[  276.681639] Hardware name: IBM 3906 M03 703 (LPAR)
[  276.681641] Call Trace:
[  276.681648]  [<0000000458c252b6>] show_stack+0x6e/0xe8
[  276.681652]  [<000000045971cf60>] dump_stack+0x90/0xc8
[  276.681656]  [<0000000458e8b186>] bad_page+0xd6/0x130
[  276.681658]  [<0000000458e8cdea>] free_pcppages_bulk+0x26a/0x800
[  276.681661]  [<0000000458e8e67e>] free_unref_page+0x6e/0x90
[  276.681663]  [<0000000458e8ea6c>] free_contig_range+0x94/0xe8
[  276.681666]  [<0000000458ea5e54>] update_and_free_page+0x1c4/0x2c8
[  276.681669]  [<0000000458ea784e>] free_pool_huge_page+0x11e/0x138
[  276.681671]  [<0000000458ea8530>] set_max_huge_pages+0x228/0x300
[  276.681673]  [<0000000458ea86c0>] nr_hugepages_store_common+0xb8/0x130
[  276.681678]  [<0000000458fd5b6a>] kernfs_fop_write+0xd2/0x218
[  276.681681]  [<0000000458ef9da0>] vfs_write+0xb0/0x2b8
[  276.681684]  [<0000000458efa15c>] ksys_write+0xac/0xe0
[  276.681687]  [<000000045972c5ca>] system_call+0xe6/0x288
[  276.681730] Disabling lock debugging due to kernel taint

This is because only the compound_order is cleared in destroy_compound_gigantic_page(), and compound_nr is set to 1U << order == 1 for order 0 in set_compound_order(page, 0).

Fix this by explicitly clearing compound_nr for first tail page after calling set_compound_order(page, 0).

---
 mm/hugetlb.c | 1 +
 1 file changed, 1 insertion(+)

index 1f3bf1710b66..3bcc0bc7e02a 100644
@@ -1216,6 +1216,7 @@ static void destroy_compound_gigantic_page(struct page *page,
 	}
 
 	set_compound_order(page, 0);
+	page[1].compound_nr = 0;
 	__ClearPageHead(page);
 }
 

2.17.1

I didn't realise there was a separate destroy path for gigantic pages. Thanks for catching and fixing this.

I cant see the patch?