Date: Mon, 31 Mar 2003 11:32:29 -0500
From: Joseph Formoso <jformoso@stevens-tech.edu>
Subject: Possible patch for stunnel 4.04 running on IRIX

Folks,

        First, much praise for providing stunnel.  It's way cool.

        The "patch" I've got is extremely experimental (though it also
has the benefit of being very small), and likely applies only to people
running stunnel on an IRIX system.
        We compiled stunnel 4.04 against a patched OpenSSL 0.9.7a, and
it exhibited odd behavior -- it keeled over dead after the first
connection (we're using it to wrap IMAP and POP3 connections).  After
poking around a little, we traced the process and found that, on the
exit of the first child process, there were many hundreds of "SIGCHLD
received" messages in the trace log, and then a SIGKILL.
        The cause for this, it seems, is that IRIX (for reasons unknown
but likely dubious) chose, as its means of compatibility with other
systems, to implement SYSV SIGCLD signal semantics and map SIGCHLD to
it.  In short, at the time that a signal handler for SIGCHLD (or SIGCLD;
under IRIX, they're synonyms) is established, the kernel checks to see if a
SIGCHLD is pending for the process.  The sigchld_handler in sselect.c of
stunnel is thus:

--------------------------------------------------------------
static void sigchld_handler(int sig) { /* SIGCHLD detected */
    int save_errno=errno;

    write(signal_pipe[1], signal_buffer, 1);
    signal(SIGCHLD, sigchld_handler);
    errno=save_errno;
}
--------------------------------------------------------------

        Since there's no wait() before the signal() call reestablishes
the handler, reestablishing the handler causes the kernel to check
immediately to see if a SIGCHLD is pending (which it is), and fire
*another* SIGCHLD signal.  Which is similarly caught and causes another
one, until the stack blows and the kernel kills the process.  The simplest
way to fix this is to make sure there's some style of wait() before the
handler is reestablished, like this:

--------------------------------------------------------------
static void sigchld_handler(int sig) { /* SIGCHLD detected */
    int save_errno=errno;
    pid_t pid;
    int status;

    write(signal_pipe[1], signal_buffer, 1);
    pid = waitpid(-1, &status, WNOHANG);
    while (pid > 0) {
      pid = waitpid(-1, &status, WNOHANG);
    }
    signal(SIGCHLD, sigchld_handler);
    errno=save_errno;
}
--------------------------------------------------------------

        This, it seems, allows stunnel to run normally on our system (at
least, so far as we've tested it).  I'm not sure how you get away
without having *any* wait()-type call, so I'm obviously not entirely
clear on how everything is working.  Still, the (trivial) patch to make
the change is offered (into the public domain) in case it happens to be
useful to you or anyone else: