WAITing for specific process

[Bill Shannon]

> > The wait() system call returns the process id of zombie child, but a
> > process may have more than one of these outstanding at a time. Wait()
> > is free (it appears) to return the pid of any of these zombie children.
> > What if you want to wait for a particular child?

	...

> The facility IS provided, to some degree:
> 
> 	void
> 	waitfor( pid )			/* wait for process to exit */
> 		int	pid;		/* ID of process to wait on */
> 		{
> 		extern unsigned	sleep();
> 	#define	DELAY	2		/* test interval, in seconds */
> 
> 		while ( kill( pid, 0 ) == 0 )
> 			sleep( (unsigned)DELAY );
> 		}

Polling for a child's death is really gross.  If the child will run
for a long time, this keeps the parent "hot" (in memory) and wastes
cpu time.  If the child will finish quickly this wastes real time.

> The only children your process should have other than
> those it created itself are the ones it is given by the shell when
> it is last in a pipeline.  It doesn't matter if you eat those zombies.

Ah, but what about the processes created by library routines that you
called?  As you start to build larger programs that make heavy use of
multiple processes (both directly in the program itself and indirectly
in library routines called by the program) you discover that the existing
facilities provided by UNIX are not nearly powerful enough to allow all
the users of sub-processes to cooperate without interfering with each
other.  However, they are powerful enough to build such a facility on
top of.  What's needed is the ability to start a process and be notified
of its termination, without interfering with any other such uses by other
parts of the same parent process.  A list of terminated processes' statuses
needs to be kept in the parent process, so that it can be examined by
something like the "waitfor" proposed above.  Switching to such a mechanism
almost certainly precludes the standard use of wait(), SIGCHLD, etc.  A
layer on top of them is needed and everyone needs to be convinced to use
only that layer (e.g. fread vs. read).  Maybe we should entertain proposals
for such a layer?

					Bill Shannon