bcachefs: Clear should_be_locked before unlock in key_cache_drop()

We're adding new should_be_locked assertions, also add a comment
explaining why clearing should_be_locked is safe here.

Signed-off-by: Kent Overstreet <[email protected]>

Author: Kent Overstreet

fs/bcachefs/btree_key_cache.c

@@ -646,9 +646,16 @@ void bch2_btree_key_cache_drop(struct btree_trans *trans,
 	unsigned i;
 	trans_for_each_path(trans, path2, i)
 		if (path2->l[0].b == (void *) ck) {
+			/*
+			 * It's safe to clear should_be_locked here because
+			 * we're evicting from the key cache, and we still have
+			 * the underlying btree locked: filling into the key
+			 * cache would require taking a write lock on the btree
+			 * node
+			 */
+			path2->should_be_locked = false;
 			__bch2_btree_path_unlock(trans, path2);
 			path2->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_drop);
-			path2->should_be_locked = false;
 			btree_path_set_dirty(path2, BTREE_ITER_NEED_TRAVERSE);
 		}