we are not properly having port.done() in a finally close.
There are at least 2 spots where port is not properly put back in pool.

One spot is pretty obvious:

• in DBTCPConnector.call(), if there is any ioexception during call, then _error() is called
• in _error(), does a checkMaster(), which will throw an exception if there is no new master
• due to this exception, the following mp.error() is not called, so port is not put back in pool.
  Consequently if you have a replset, and master dies but it takes some time to elect new master, any find operation will take away a port from pool.
  You just need 10 find calls to make pool unusable for that server, until the application is restarted.

Other spot:

• in DBTCPConnector.say(), in case that there is a writeconcern, getlasterror() is called
• if there is ioexception in getlasterror, it is converted into mongoexception
• the catch block of mongoexception does not put port back in pool.
  Consequently if you have any ioexception in during the getlasterror call, the port will get lost.