CMSC 421 Operating Systems Lecture Notes
(c) 1994 Howard E. Motteler


Message Passing
================

The goal of critical sections, monitors, etc. is to allow
processes to _communicate_

Two general approaches to communication
  - shared memory (with sem's or monitors)
  - messages

Distributed systems must use message passing

Typical message passing operations are
  - send(destination, message)
  - receive(source)

Messages can be
  - sent to specific destinations
  - sent on `channels'
  - sent one at a time  (as in CSP's rendezvous protocol)
  - queued at source, dest, or in-between  (e.g., email)
  - point to point  (one sender and one receiver)
  - broadcast  (sent to multiple destinations)
  - acknowledged or unacknowledged

Destinations can be
  - internet addresses
  - process id's
  - any process doing a receive()

Channels can be
  - static (like programming language variables)
  - dynamic (ie., they need an `open' operation)
  - 1-way or 2-way



Example: Bounded Buffer with message passing
---------------------------------------------

Assumptions:
  - Message are buffered
  - the message buffer acts as the queue
  - buffer size is N

Producer:

while (alive) {
   generate an item V;
   receive(consumer);      // get `empty' from the consumer
   send(consumer, V);      // send data to the consumer
   }

Consumer:

for (i=0; i<N; i++) 
   send(producer, dummy);  // send N `empties' to producer
while (alive) {
   V = receive(producer);  // get value from producer
   send(producer, dummy);  // send an `empty' back
   use the item V;
   }


Q: Why not have
  - Producer: send(consumer, V);
  - Consumer: V = receive(producer);

A: This should give *unbounded* buffer 
   (unbounded except by system limitations)



Example: CSP 
-------------

CSP (`communicating sequential processes') is a simple 
message passing language

CSP processes communicate through _channels_

CSP channels are 
  - statically allocated (declared like variables)
  - point to point
  - one way

CSP processes do not share memory space, and can run on either
one or several processors

Suppose C is a channel variable shared by processes P1 and P2

  - C ! x means send x on channel C
  - C ? y means receive a value on channel C and save in y

   P1                      P2
      :                       :
      C ! x                   C ? y
      :                       :

If P1 does C ! x first
  - P1 waits for C ? y
  - both P1 and P2 continue
  - y contains the value from x
  
If P2 does C ? y first
  - P2 waits for C ! x, 
  - both P1 and P2 continue
  - y contains the value from x

CSP communication is a _rendezvous_, and no queues are needed

(In practice, this works best over relatively short distances)

CSP has a special construct ``ALT'' that makes it possible to
wait for input from more than one channel at a time

    ALT
        C1 ? x         # wait for something on C1
          C3 ! x       # do this if we get something on C1
        C2 ? x         # wait for something on C2
          C3 ! x       # do this if we get something on C2

This example acts as a `merge', taking input from either C1 or
C2 at any time, and sending the value received out on C3

CSP channels can be implemented with sem's and shared memory

Each channel needs two semaphores S1 and S2, both initially 0,
and a shared memory location, L

    C ! x                 C ? y

    L = x;                wait(S1)
    signal(S1)            y = L;
    wait(S2)              signal(S2)


Example: bounded buffer in CSP
-------------------------------


Producer    ---->    Buffer     ---->    Consumer
process    channel   process   channel   process
           `pchan'             `cchan'


Producer:  
   WHILE true DO BEGIN
      generate x
      pchan ! x       // send x to the buffer
      END

Consumer:  
   WHILE true DO BEGIN
      cchan ? y      // receive y from the buffer
      use y
      END

Buffer:    
   INTEGER buf[N], in=0, out=0;
   WHILE true DO
     ALT
      // test buffer not full and sender ready:
      (in < out + N) & pchan ? buf[in mod N]  
         in := in + 1

      // test buffer not empty and receiver ready:
      (out < in) & cchan ! buf[out mod N]
         out := out + 1

Unfortunately, the buffer here is busy-waiting; if it is not
running on a seperate processor, a sleep could be added to the
while-loop

The actual implementations of CSP (such as the `Occam' language)
may not provide the `guarded send' that we used here



Equivalence of IPC Primitives
==============================

Some IPC primitives can be used to simulate others

  - CSP channels with with sem's and shared memory
  - Monitors with semaphores (example in text)
  - Message passing with sem's and shared memory
  - Semaphores with messages 


Messages with Semaphores and Shared Memory
-------------------------------------------

Most forms of message passing can be simulated with semaphores and
shared memory

Message Passing implemented with semaphores is similar to the
bounded buffer example:


                  P1  -->  |||||||  -->  P1
               producer     buffer     consumer
               or sender              or receiver


The buffer in a message passing system may be
  - part of the sender
  - part of a separate process
  - part of the receiver

The details of the semaphore/shared memory simulation depend on
what sort of message passing system we want


Semaphores with messages 
-------------------------

Semaphores can be implemented in a message passing system, by using
a synchronization process.

Note that since semaphores are usually used with shared memory, this
example would not bee not too useful for a true distributed system.
Further, in any shared memory system, there are probably easier ways
to implement semaphores.

Assumptions:
  - messages are queued
  - message are sent to process id's

(example for a single semaphore)

signal() {
  msg[] = [SIG, pid];
  send(synch, msg);
  }

wait() {
  msg[] = [WAIT, pid];
  send(synch, msg);
  receive();
  }

synch process:
  while (alive) {
    msg = receive();
    fnx = msg[1];
    pid = msg[2];
    if (fnx == SIG) 
      s = s + 1;
      if (s <= 0) {
         get p from list L;
         send(p, dummy);
         }
      }
    else if (fnx == WAIT) {
       s = s - 1;
       if (s < 0)
          add pid to L;
       else
          send(pid);
       }
   }
          

Q: do the synch process or signal and wait procedures
   need mutual exclusion?

A: No, since the synch process only receives and processes one
   message at a time.