inetd fd leak?

[Denys Vlasenko]

On Friday 04 September 2009 00:32, Denys Vlasenko wrote:
> > The  leak appears
> > to be related to syslog.  For example, descriptor 6 was leaked, and 
> > strace shows
> > only these operations on that descriptor:
> > 
> > socket(PF_FILE, SOCK_DGRAM, 0)          = 6
> > fcntl64(6, F_SETFD, FD_CLOEXEC)         = 0
> > connect(6, {sa_family=AF_FILE, path="/dev/log"}, 110) = 0
> > send(6, "<27>Sep  3 17:49:45 inetd[1938]:"..., 75, MSG_NOSIGNAL) = 75
> > 
> > After another run of bury.sh, the same thing happens on descriptor 7 
> > (without 6
> > being closed first).  The message being logged is this one:
> > 
> > Sep  3 17:49:45 uml0 inetd[1938]: telnet/tcp: too many connections, pausing
> > 
> > This is so totally weird that I figured it had to have something to do with
> > vfork (just because vfork is so totally weird itself).  Replacing the 
> > vfork in
> > inetd.c with a fork instead does fix my leak.
> 
> 
> Hmmm. Try inetd with option -e. If it stops leaking, we found it.

The tentative fix is to add LOG_NDELAY here in inetd.c:

        if (!(opt & 4)) {
                /* LOG_NDELAY: connect to syslog daemon NOW.
                 * Otherwise, we may open syslog socket
                 * in vforked child, making opened fds and syslog()
                 * internal state inconsistent.
                 * This was observed to leak file descriptors. */
                openlog(applet_name, LOG_PID | LOG_NDELAY, LOG_DAEMON);
                logmode = LOGMODE_SYSLOG;
        }

--
vda