klogd: fix the problem of delayed handling of ^C and SIGTERM

Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>

sysklogd/klogd.c

@@ -150,12 +150,41 @@ int klogd_main(int argc UNUSED_PARAM, char **argv)
 	 */
 	klogd_open();
 	openlog("kernel", 0, LOG_KERN);
+	/*
+	 * glibc problem: for some reason, glibc changes LOG_KERN to LOG_USER
+	 * above. The logic behind this is that standard
+	 * http://pubs.opengroup.org/onlinepubs/9699919799/functions/syslog.html
+	 * says the following about openlog and syslog:
+	 * "LOG_USER
+	 *  Messages generated by arbitrary processes.
+	 *  This is the default facility identifier if none is specified."
+	 *
+	 * I believe glibc misinterpreted this text as "if openlog's
+	 * third parameter is 0 (=LOG_KERN), treat it as LOG_USER".
+	 * Whereas it was meant to say "if *syslog* is called with facility
+	 * 0 in its 1st parameter without prior call to openlog, then perform
+	 * implicit openlog(LOG_USER)".
+	 *
+	 * As a result of this, eh, feature, standard klogd was forced
+	 * to open-code its own openlog and syslog implementation (!).
+	 *
+	 * Note that prohibiting openlog(LOG_KERN) on libc level does not
+	 * add any security: any process can open a socket to "/dev/log"
+	 * and write a string "<0>Voila, a LOG_KERN + LOG_EMERG message"
+	 *
+	 * Google code search tells me there is no widespread use of
+	 * openlog("foo", 0, 0), thus fixing glibc won't break userspace.
+	 *
+	 * The bug against glibc was filed:
+	 * bugzilla.redhat.com/show_bug.cgi?id=547000
+	 */
 
 	if (i)
 		klogd_setloglevel(i);
 
-	bb_signals(BB_FATAL_SIGS, record_signo);
 	signal(SIGHUP, SIG_IGN);
+	/* We want klogd_read to not be restarted, thus _norestart: */
+	bb_signals_recursive_norestart(BB_FATAL_SIGS, record_signo);
 
 	syslog(LOG_NOTICE, "klogd started: %s", bb_banner);