__split_huge_pmd_locked() must check if the cleared huge pmd was dirty,
and propagate that to PageDirty: otherwise, data may be lost when a huge
tmpfs page is modified then split then reclaimed.
How has this taken so long to be noticed? Because there was no problem
when the huge page is written by a write system call (shmem_write_end()
calls set_page_dirty()), nor when the page is allocated for a write fault
(fault_dirty_shared_page() calls set_page_dirty()); but when allocated
for a read fault (which MAP_POPULATE simulates), no set_page_dirty().
Fixes: d21b9e57c74c ("thp: handle file pages in split_huge_pmd()")
Reported-by: Ashwin Chaugule <ashwinch@xxxxxxxxxx>
Signed-off-by: Hugh Dickins <hughd@xxxxxxxxxx>
Cc: "Kirill A. Shutemov" <kirill.shutemov@xxxxxxxxxxxxxxx>
Cc: "Huang, Ying" <ying.huang@xxxxxxxxx>
Cc: Yang Shi <yang.shi@xxxxxxxxxxxxxxxxx>
Cc: <stable@xxxxxxxxxxxxxxx> # v4.8+
---
mm/huge_memory.c | 2 ++
1 file changed, 2 insertions(+)
--- 4.18-rc4/mm/huge_memory.c 2018-06-16 18:48:22.029173363 -0700
+++ linux/mm/huge_memory.c 2018-07-10 20:11:29.991011603 -0700
@@ -2084,6 +2084,8 @@ static void __split_huge_pmd_locked(stru
if (vma_is_dax(vma))
return;
page = pmd_page(_pmd);
+ if (!PageDirty(page) && pmd_dirty(_pmd))
+ set_page_dirty(page);
if (!PageReferenced(page) && pmd_young(_pmd))
SetPageReferenced(page);
page_remove_rmap(page, true);
