I was right, there were race conditions [1] in monnet. Mark [2] helped me in locating them and my only comment on the whole thing is: Unix signal semantics suck. Although Mark assures me that any form synchronization is nasty, although I still don't see why it has to be so difficult.
To be portable, the only thing you can do in a signal handler is do a simple assignment to a variable declared as volatile sig_atomic_t. Anything else could lead to problems. So, in monnet I now have:
volatile sig_atomic_t g_sigint = 0;
volatile sig_atomic_t g_sighup = 0;
volatile sig_atomic_t g_sigchld = 0;
static void handler_int()
{
g_sigint = 1;
}
static void handler_hup()
{
g_sighup = 1;
}
static void handler_chld()
{
int status;
g_sigchld = 1;
wait(&status);
}
Granted, that isn't proper ANSI C function headers, but there is no real consensus as to what signal handlers take (on some systems, a single integer parameter, others, no parameters, others several parameters) so that's about the best you can do. I am taking a risk with handler_chld() in doing the wait() but POSIX lists wait() as being callable via a signal handler so hey, why not live on the edge here. Now, elsewhere, the main code:
while(1)
{
while(1)
{
s = read( /* ... */ )
if (s <= 0) break;
/* ... */
}
if (g_sighup)
{
g_sighup = 0;
generate_report();
}
if (g_sigchld)
{
g_sigchld = 0;
g_child = 0;
}
if (g_sigint)
break;
}
/* ... */
static void generate_report(void)
{
if (g_child != 0)
return;
g_child = fork();
if (g_child > 0) /* parent resumes */
return;
else if (g_child < 0) /* error? just resume */
{
g_child = 0;
return;
}
/* ... */
}
SIGINT just breaks out of the main loop and terminates the program (with some cleanup). SIGHUP generates a call to generate_report() which creates a new process (if one hasn't already been created) to generate the actual report.
If I didn't handle SIGCHLD, I would end up with zombie processes (lovely in that even if I don't care about the child, I still have to wait() for it). Now, it is conceivable that a SIGHUP sent at the right time would fail to create a report file, but would only happen if a previous SIGHUP had been given to generate a report file and was just finishing up. But I can live with that.