diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index a513861a9037..26769465af63 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -1746,7 +1746,39 @@ int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 	if (prot_numa && pmd_protnone(*pmd))
 		goto unlock;
 
-	entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
+	/*
+	 * In case prot_numa, we are under down_read(mmap_sem). It's critical
+	 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
+	 * which is also under down_read(mmap_sem):
+	 *
+	 *	CPU0:				CPU1:
+	 *				change_huge_pmd(prot_numa=1)
+	 *				 pmdp_huge_get_and_clear_notify()
+	 * madvise_dontneed()
+	 *  zap_pmd_range()
+	 *   pmd_trans_huge(*pmd) == 0 (without ptl)
+	 *   // skip the pmd
+	 *				 set_pmd_at();
+	 *				 // pmd is re-established
+	 *
+	 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
+	 * which may break userspace.
+	 *
+	 * pmdp_invalidate() is required to make sure we don't miss
+	 * dirty/young flags set by hardware.
+	 */
+	entry = *pmd;
+	pmdp_invalidate(vma, addr, pmd);
+
+	/*
+	 * Recover dirty/young flags.  It relies on pmdp_invalidate to not
+	 * corrupt them.
+	 */
+	if (pmd_dirty(*pmd))
+		entry = pmd_mkdirty(entry);
+	if (pmd_young(*pmd))
+		entry = pmd_mkyoung(entry);
+
 	entry = pmd_modify(entry, newprot);
 	if (preserve_write)
 		entry = pmd_mk_savedwrite(entry);