The only place we currently take the WT_DHANDLE::close_lock is in __wt_conn_btree_sync_and_close. It used to protect cases where we did apply single when creating a checkpoint, but that code is no longer present.

I don't see anywhere that we protect against multiple threads calling __wt_conn_btree_sync_and_close in parallel. Callers often hold the schema lock, but not always.

I think the __wt_conn_btree_sync_and_close should re-check whether the handle is open after getting the close_lock and return if someone beat us to closing it.