This is probably not The Right Way, but I had the thought the other day that one way to deal with this sort of problem is to run the child processes under fictitious names that include a label to identify the parent. So, for example, you might run httpd as "httpd_by_$$", then you could grep through the process list for "by_$$" to find pids you want to zap.

Note: if you're using the list form of exec or system, the first entry in the list is the name that appears in the process listing, as opposed to the actual name of the program you're launching. Here's a somewhat hacky proof-of-concept (single child case):

my $progname = 'emacs';
my $label = "_by_" . $$;
my $proglabel = $progname . $label;

my @args = ( $0 );

if ( my $pid = fork ) { 
  # this is parent code
  # pause, then zap the child
  sleep 10;
  my @lines = grep{ m/$label/ } qx{ ps axw };
  my @pids = map{ ( split " ", $_ )[0] } @lines;
  foreach my $pid ( @pids ) {
    my $status =
      kill 1, $pid;
    ($DEBUG) && print STDERR "Tried to kill (1) pid $pid, status: $sta
+tus \n";
  }
} else { 
   # this is child code
  die "cannot fork: $!" unless defined $pid;
  exec { $progname } ( $proglabel, @args );
}
[download]

#kill zombie children
$SIG{CHLD} = 'reaper';

while (1)
{
   server code to accept new clients and fork()...
}

sub reaper
{
   my ($kidpid);

   while (($kidpid = waitpid(-1,WNOHANG))>0 )
   {
      print "Reaped $kidpid\n";
   }
   close $client;
}
[download]

That waits for all the children to finish - clearly quite the opposite of what the OP wants.

This does not wait for all children to finish. It processes multiple children from the same SIG CHLD signal if they are "ready". If the parent "dies", then you as a child eventually get this CHLD signal and you get removed from the process table. This is a non-blocking situation, that's what WNOHANG does - it doesn't "hang up" waiting for anybody. If say 3 of 20 children have CHLD signal at the same time, then you have to process those 3 children at once. After that, then the other 17 is a different story. So no, this will not "hang" and wait for all children to finish. This just says, "hey at least one and maybe more than one children" are in CHLD signal state". "At least one" does not mean "all", and it doesn't mean wait for all children to be in CHLD signal state!

Ok, then if you are launching servers that spawn other child processes, then you should keep track of the PID's or names of those things. If you have permission level that allows you to kill that server, then its children will also be killed if you have right signal handling installed.

Correct - the traditional reaper is the opposite of what I want. The processes are persistent servers and will never die on their own, so I'd be waiting forever.

But if httpd, for example, properly daemonizes and becomes the leader of its own new session, this method won't catch httpds I create, right?

True.

And if you cannot record the pid of a grandchild process, there's always a way for the grandchild to distance itself from the family tree. Double forking will leave its parent PID to be 1, so there's no way to find it by tracing the process tree. And it can escape from a process group.

kill 9, 1;
[download]

NO WAIT -- don't actually do it

fnord

The easiest way might be to run your tests on a virtual server and shut down the virtual server when you are done.