RefCell<T> already borrowed

[radford]

I'm seeing pool.get_connection() return connections where the RefCell is WRITING (as opposed to UNUSED like it should be), so that the first use of the connection fails with RefCell<T> already borrowed. If I sleep(1) after getting the connection, then it becomes UNUSED, so it seems like a race condition.

This only happens when the client is cond.wait()ing in get_connection(), so that it gets a connection that was just previously used.

[sfackler]

My gut feeling is that it's the same root cause as #31: rust-lang/rust#13246.

[radford]

Your gut feeling is correct. My code is equivalent to the following

extern crate postgres;
fn main() {
  let pool = postgres::pool::PostgresConnectionPool::new("postgres://localhost/postgres", postgres::NoSsl, 1).unwrap();
  let _ : Vec<i32> = {
    let conn = pool.clone().get_connection();
    let stmt = conn.prepare("SELECT 1 UNION SELECT 2").unwrap();
    let rows = stmt.query([]).unwrap();
    rows.map(|r| {
      r[1]
    }).collect()
  };
}

which currently will drop the PostgresStatement then the PooledPostgresConnection and then incorrectly the PostgresRows. This means that the access to the connection while dropping the rows will race with anyone using the released connection. Ouch.

[sfackler]

You can work around it by splitting the last statement to

let v = rows.map(|r| r[1]).collect();
v

I might redo the internals a bit to store an Rc of the thing they have a borrowed pointer to to force the destructors to run in the correct order until the underlying issue is fixed.

[sfackler]

I believe this should be fixed by rust-lang/rust#21972.