mm: thp: make the THP mapcount atomic against __split_huge_pmd_locked()

Write protect anon page faults require an accurate mapcount to decide
if to break the COW or not. This is implemented in the THP path with
reuse_swap_page() ->

If the COW triggers while the other processes sharing the page are
under a huge pmd split, to do an accurate reading, we must ensure the
mapcount isn't computed while it's being transferred from the head
page to the tail pages.

reuse_swap_cache() already runs serialized by the page lock, so it's
enough to add the page lock around __split_huge_pmd_locked too, in
order to add the missing serialization.

Note: the commit in "Fixes" is just to facilitate the backporting,
because the code before such commit didn't try to do an accurate THP
mapcount calculation and it instead used the page_count() to decide if
to COW or not. Both the page_count and the pin_count are THP-wide
refcounts, so they're inaccurate if used in
reuse_swap_page(). Reverting such commit (besides the unrelated fix to
the local anon_vma assignment) would have also opened the window for
memory corruption side effects to certain workloads as documented in
such commit header.

Signed-off-by: Andrea Arcangeli <>
Suggested-by: Jann Horn <>
Reported-by: Jann Horn <>
Acked-by: Kirill A. Shutemov <>
Fixes: 6d0a07edd17c ("mm: thp: calculate the mapcount correctly for THP pages during WP faults")
Signed-off-by: Linus Torvalds <>
diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 11fe0b4..dddc863 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -2385,6 +2385,8 @@
 	spinlock_t *ptl;
 	struct mmu_notifier_range range;
+	bool was_locked = false;
+	pmd_t _pmd;
 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
 				address & HPAGE_PMD_MASK,
@@ -2397,11 +2399,32 @@
 	 * pmd against. Otherwise we can end up replacing wrong page.
 	VM_BUG_ON(freeze && !page);
-	if (page && page != pmd_page(*pmd))
-	        goto out;
+	if (page) {
+		VM_WARN_ON_ONCE(!PageLocked(page));
+		was_locked = true;
+		if (page != pmd_page(*pmd))
+			goto out;
+	}
 	if (pmd_trans_huge(*pmd)) {
-		page = pmd_page(*pmd);
+		if (!page) {
+			page = pmd_page(*pmd);
+			if (unlikely(!trylock_page(page))) {
+				get_page(page);
+				_pmd = *pmd;
+				spin_unlock(ptl);
+				lock_page(page);
+				spin_lock(ptl);
+				if (unlikely(!pmd_same(*pmd, _pmd))) {
+					unlock_page(page);
+					put_page(page);
+					page = NULL;
+					goto repeat;
+				}
+				put_page(page);
+			}
+		}
 		if (PageMlocked(page))
 	} else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
@@ -2409,6 +2432,8 @@
 	__split_huge_pmd_locked(vma, pmd, range.start, freeze);
+	if (!was_locked && page)
+		unlock_page(page);
 	 * No need to double call mmu_notifier->invalidate_range() callback.
 	 * They are 3 cases to consider inside __split_huge_pmd_locked():