On further thought, your fork() and exec() tests were run in the same script under the same conditions, right? Where I'm going is that the target user may have had reached its process limits where in testing you (personally) still have processes left.

here's the bare bone of my code which forks the child process. I also added some debug messages.

===

sub fork_job
{
    my $job  = shift;
    my $system_fail = 0;
    system( "ls . " ) == 0 ||
                print "system failed before forking\n"

    my $pid = fork();

    if( !defined( $pid ) ) { # bad, bad, error
          return 0;
    } elsif ($pid != 0) {        # parent
          return $pid;
    } elsif ($pid == 0) {   # $child
          system( "ls . " ) == 0 ||
               print "system failed after forking(1)\n";
          $> = $job->{'uid'};
          $< = $job->{'uid'};
          system( "ls . " ) == 0 ||
               print "system failed after forking(2)\n";

          setpgrp(0, $$);
          system( "ls . " ) == 0 ||
              print "system failed after forking(3)\n";

          ## doesn't exec, but continues on
          my $runner = Runner->new( $job );
          $runner->run();
          exit 1;
     }
}
[download]

===

It seems that any system() call fails after setpgrp( 0, $$ ). Is setpgrp() an issue on solaris?

My bad, the above was completely inacurate. I started doing some searching around to the cause of the annoying error, and I reduced it down to this:

#!/usr/local/bin/perl
$SIG{ 'CHLD' } = sub {
        my $pid = wait();
        print "Got $pid\n";
};
system("/bin/ls > /dev/null") == 0 ||
        print "system exited with $?, error is \"$!\"\n";
[download]

If you take out the signal handler, it works. Now I'm completely lost...

By the way, this is the uname -a output:

SunOS xxxxxx 5.7 Generic sun4u sparc SUNW,Ultra-5_10

ANY ideas??

system() does a wait() for the /bin/ls process to complete. When the process ends, your handler jumps in and grabs it, and system() gets nothing, which isn't what it wanted. If you run a command that has a side effect, you should see that it actually does get executed.

It's interesting that it worked on Linux but not Solaris. Must be some difference in the way they deliver signals.

Best solution is probably not to mix $SIG{CHLD} with system(). There are a couple of threads around here that might be useful... try a search or two.