This note describes the logic of mas-slave.c
Some notable global variables are:
  pending[]
  ...
The typical calling sequence seen by a task is:

main() [User]
  [enter mas-slave.c]
  init_master_slave() [Is this called here?]
  master_slave()
    SET par_loop_get_task  [MAYBE get_task_global A BETTER NAME?]
    raw_master_slave()
      SET get_task_result_global
      SET update_environment_global
      init_master_slave() [if not yet called]
      par_call()  [Defined in threads-mpi.c or threads.c; Par. stuff begins]
        body_master_slave()
          runtime()
          if (is_master()) goto MASTER else goto SLAVE
==========================

MASTER:
walltime()
parallel_loop()
  LOOP:
  wait_for_slave() [while num_free_slaves == 0]
  task = get_task() [same as par_loop_get_task()]
  [NOTE:  set_task() callable directly by user, if raw_master_slave() invoked
          by user, and user submits his own version of parallel_loop()]
  set_task(task) [if mas_slave_is_task  (if there is a current task)]
    wait_for_slave() [if num_free_slaves < num_slaves]
      recv_msg()
      action = get_task_result_global()
      switch(action)
        case REDO: send_task()
        case CONTINUE: send_msg()
        case NO_ACTION:  add_free_slave()
        case UPDATE:
          add_free_slave()
          broadcast_command(result, task) [if not is_shared_memory]
            [CURRENTLY NOT IMPLEMENTED -- ALSO NOT CALLED CURRENTLY]
          update_environment()  [same as update_environment_global()]
    send_task(task)
wait_for_all_slaves()
  wait_for_slave() [until num_free_slaves = num_slaves]
  send_msg(doneMessage)
master_stats()

SLAVE:
  LOOP:
  result = recv_msg()
  tag = get_last_tag()
  result == doneMessage:  break
  result == confirmMessage:  /* ping better than confirm ? */
    send_msg(answerMessage)
  tag == BROADCAST_TAG:
    update_environment()
  else:
    do_task(result)
    send_msg(task)
  LOOP_END:
  slave_stats()
  finalize()

==================================================================

MSG() sets these:
mas_slave_is_msg set to TRUE
mas_slave_msg_size
mas_slave_msg_ptr

send_msg() uses them and resets msg_slave_is_msg to FALSE

==================================================================
enum tag {DUMMY_TAG = NUM_PEND_TAGS, PING_TAG, ACK_TAG, DONE_TAG,
	  BCAST_RESULT_TAG, BCAST_TASK_TAG};
tasks and results are always accompanied by a tag less than NUM_PEND_TAGS
Currently, the MPI_datatype is always assumed to be MPI_CHAR
This will not work, of course, for general heterogeneous architectures,
  but then the user would have to declare the types of his tasks/results

==================================================================
mas-slave.c calls attach_new_slave().  This is defined in threads-mpi.c
  and threads.c by default as a function returning 0.
  If threads-mpi.c is compiled with LOADER defined, attach_new_slave()
  is defined to call poll_new_slave() and MPI_Spawn2().
  Also, init_loaderctl() is called from threads-mpi.c in that case.
  MPI_Spawn2() is defined in mpinu/spawn2.c.  If it is not linked
  into the final executable, then it will not require symbols from loader/*
  loader/* defines poll_new_slave() and other utilities.

When invoked, init_loaderctl() sets up a well-known port to be used
  by the binary loader/loader.  That binary then downloads an unexec'ed
  checkpoint file (called topcjob locally), and starts it running.
  ./topcjob starts at main(), invokes MPI_Init(), where it talks
  to MPI_Spawn2() on the master process, which integrates it as a new slave.