Description of problem:

Occasionally, ypserv dies with the message

    ypserv[686]: svc_run: - poll failed: No child processes

logged to syslog.  However, it doesn't happen very often.  Nevertheless, I have
set the severity to high, since when it *does* happen, the entire system
(including all YP clients) becomes more or less unusable.



Version-Release number of selected component (if applicable):

ypserv-2.8-0.9E



How reproducible:

Not very...  I installed a server with RedHat 9 and updates at the end of July,
and during August ypserv died perhaps ten times.  Then it ran fine for almost a
full month, and today (September 26th) it died again.



Steps to Reproduce:

1. Run ypserv on a busy ypserver
2. Wait for hours, days or months...


    
Actual results:

ypserv dies with the above message logged to syslog



Expected results:

ypserv should not die. :-)



Additional info:

That particular message only exists in one place in the ypserv source, in the
function ypserv_svc_run() in ypserv.c, the -1 branch of the switch statement.

The probable cause is that the process gets a SIGCHLD signal after poll()
returned -1, but before it gets the chance to actually look at the value of
errno.  The signal handler for SIGCHLD, sig_child(), calls waitpid(), which
clobbers the value of errno.

One way to work around this problem, is for sig_child() to save the value of
errno at entry and restore it when it returns:

==============================================================================
--- ypserv.c.orig       2003-09-26 14:31:18.000000000 +0200
+++ ypserv.c    2003-09-26 14:30:58.000000000 +0200
@@ -354,10 +354,11 @@
 static void
 sig_child (int sig)
 {
   int st;
   pid_t pid;
+  int save = errno;
 
   if (debug_flag)
     log_msg ("sig_child: got signal %i", sig);
 
 
@@ -371,10 +372,12 @@
 
   if (children < 0)
     log_msg ("children is lower 0 (%i)!", children);
   else if (debug_flag)
     log_msg ("children = %i", children);
+
+  errno = save;
 }
 
 
 static void
 Usage (int exitcode)
==============================================================================

However, that only helps for this particular signal.  Other signal handlers can
wreak havoc too, and not just with errno.  sig_hup() does a lot of things that I
believe are not safe to do in a signal handler; calling fopen() and malloc(),
for instance.  A better way would be to make the signal handlers only set a flag
(of the type 'volatile sig_atomic_t'), and then let the main loop detect that
the flag has been set.