mm: fix crashes from deferred split racing folio migration

Even on 6.10-rc6, I've been seeing elusive "Bad page state"s (often on
flags when freeing, yet the flags shown are not bad: PG_locked had been
set and cleared??), and VM_BUG_ON_PAGE(page_ref_count(page) == 0)s from
deferred_split_scan()'s folio_put(), and a variety of other BUG and WARN
symptoms implying double free by deferred split and large folio migration.

6.7 commit 9bcef59 ("mm: memcg: fix split queue list crash when large
folio migration") was right to fix the memcg-dependent locking broken in
85ce2c5 ("memcontrol: only transfer the memcg data for migration"),
but missed a subtlety of deferred_split_scan(): it moves folios to its own
local list to work on them without split_queue_lock, during which time
folio->_deferred_list is not empty, but even the "right" lock does nothing
to secure the folio and the list it is on.

Fortunately, deferred_split_scan() is careful to use folio_try_get(): so
folio_migrate_mapping() can avoid the race by folio_undo_large_rmappable()
while the old folio's reference count is temporarily frozen to 0 - adding
such a freeze in the !mapping case too (originally, folio lock and
unmapping and no swap cache left an anon folio unreachable, so no freezing
was needed there: but the deferred split queue offers a way to reach it).

Link: https://lkml.kernel.org/r/[email protected]
Fixes: 9bcef59 ("mm: memcg: fix split queue list crash when large folio migration")
Signed-off-by: Hugh Dickins <[email protected]>
Reviewed-by: Baolin Wang <[email protected]>
Cc: Barry Song <[email protected]>
Cc: David Hildenbrand <[email protected]>
Cc: Hugh Dickins <[email protected]>
Cc: Kefeng Wang <[email protected]>
Cc: Matthew Wilcox (Oracle) <[email protected]>
Cc: Nhat Pham <[email protected]>
Cc: Yang Shi <[email protected]>
Cc: Zi Yan <[email protected]>
Cc: <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>

mm/memcontrol.c

@@ -7823,17 +7823,6 @@ void mem_cgroup_migrate(struct folio *old, struct folio *new)
 
 	/* Transfer the charge and the css ref */
 	commit_charge(new, memcg);
-	/*
-	 * If the old folio is a large folio and is in the split queue, it needs
-	 * to be removed from the split queue now, in case getting an incorrect
-	 * split queue in destroy_large_folio() after the memcg of the old folio
-	 * is cleared.
-	 *
-	 * In addition, the old folio is about to be freed after migration, so
-	 * removing from the split queue a bit earlier seems reasonable.
-	 */
-	if (folio_test_large(old) && folio_test_large_rmappable(old))
-		folio_undo_large_rmappable(old);
 	old->memcg_data = 0;
 }
 

mm/migrate.c

@@ -415,6 +415,15 @@ int folio_migrate_mapping(struct address_space *mapping,
 		if (folio_ref_count(folio) != expected_count)
 			return -EAGAIN;
 
+		/* Take off deferred split queue while frozen and memcg set */
+		if (folio_test_large(folio) &&
+		    folio_test_large_rmappable(folio)) {
+			if (!folio_ref_freeze(folio, expected_count))
+				return -EAGAIN;
+			folio_undo_large_rmappable(folio);
+			folio_ref_unfreeze(folio, expected_count);
+		}
+
 		/* No turning back from here */
 		newfolio->index = folio->index;
 		newfolio->mapping = folio->mapping;
@@ -433,6 +442,10 @@ int folio_migrate_mapping(struct address_space *mapping,
 		return -EAGAIN;
 	}
 
+	/* Take off deferred split queue while frozen and memcg set */
+	if (folio_test_large(folio) && folio_test_large_rmappable(folio))
+		folio_undo_large_rmappable(folio);
+
 	/*
 	 * Now we know that no one else is looking at the folio:
 	 * no turning back from here.