mm/huge_memory: Avoid calling pmd_page() on a non-leaf PMD

Calling try_to_unmap() with TTU_SPLIT_HUGE_PMD and a folio that's not
mapped by a PMD causes oopses on arm64 because we now call page_folio()
on an invalid page.  pmd_page() returns a valid page for non-leaf PMDs on
some architectures, so this bug escaped testing before now.  Fix this bug
by delaying the call to pmd_page() until after we know the PMD is a leaf.

Link: https://bugzilla.kernel.org/show_bug.cgi?id=215804
Fixes: af28a988b313 ("mm/huge_memory: Convert __split_huge_pmd() to take a folio")
Reported-by: Zorro Lang <zlang@redhat.com>
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Tested-by: Zorro Lang <zlang@redhat.com>

diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 2fe3821..c468fee 100644 (file)
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -2145,15 +2145,14 @@ void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
         * pmd against. Otherwise we can end up replacing wrong folio.
         */
        VM_BUG_ON(freeze && !folio);
-       if (folio) {
-               VM_WARN_ON_ONCE(!folio_test_locked(folio));
-               if (folio != page_folio(pmd_page(*pmd)))
-                       goto out;
-       }
+       VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
 
        if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
-           is_pmd_migration_entry(*pmd))
+           is_pmd_migration_entry(*pmd)) {
+               if (folio && folio != page_folio(pmd_page(*pmd)))
+                       goto out;
                __split_huge_pmd_locked(vma, pmd, range.start, freeze);
+       }
 
 out:
        spin_unlock(ptl);