watchfrr: don't wait forever at startup

If we wait forever for all daemons to come up, we can hang the entire
boot process, especially on init.d based systems.

Signed-off-by: David Lamparter <equinox@diac24.net>

diff --git a/watchfrr/watchfrr.c b/watchfrr/watchfrr.c
index 5230ec383f2bd9fad47473ed262666d60058843e..7e75bed2d6edd4592b974f05f551aea6ddce983c 100644 (file)
--- a/watchfrr/watchfrr.c
+++ b/watchfrr/watchfrr.c
@@ -83,6 +83,7 @@ static const char *phase_str[] = {
 };
 
 #define PHASE_TIMEOUT (3*gs.restart_timeout)
+#define STARTUP_TIMEOUT        55 * 1000
 
 struct restart_info {
        const char *name;
@@ -97,6 +98,7 @@ struct restart_info {
 static struct global_state {
        restart_phase_t phase;
        struct thread *t_phase_hanging;
+       struct thread *t_startup_timeout;
        const char *vtydir;
        long period;
        long timeout;
@@ -630,23 +632,38 @@ static int handle_read(struct thread *t_read)
  * Wait till we notice that all daemons are ready before
  * we send we are ready to systemd
  */
-static void daemon_send_ready(void)
+static void daemon_send_ready(int exitcode)
 {
+       FILE *fp;
        static int sent = 0;
-       if (!sent && gs.numdown == 0) {
-               FILE *fp;
 
+       if (sent)
+               return;
+
+       if (exitcode == 0)
                zlog_notice("all daemons up, doing startup-complete notify");
-               frr_detach();
+       else if (gs.numdown < gs.numdaemons)
+               flog_err(WATCHFRR_ERR_CONNECTION,
+                        "startup did not complete within timeout"
+                        " (%d/%d daemons running)",
+                        gs.numdaemons - gs.numdown, gs.numdaemons);
+       else {
+               flog_err(WATCHFRR_ERR_CONNECTION,
+                        "all configured daemons failed to start"
+                        " -- exiting watchfrr");
+               exit(exitcode);
+
+       }
 
-               fp = fopen(DAEMON_VTY_DIR "/watchfrr.started", "w");
-               if (fp)
-                       fclose(fp);
+       frr_detach();
+
+       fp = fopen(DAEMON_VTY_DIR "/watchfrr.started", "w");
+       if (fp)
+               fclose(fp);
 #if defined HAVE_SYSTEMD
-               systemd_send_started(master, 0);
+       systemd_send_started(master, 0);
 #endif
-               sent = 1;
-       }
+       sent = 1;
 }
 
 static void daemon_up(struct daemon *dmn, const char *why)
@@ -655,7 +672,8 @@ static void daemon_up(struct daemon *dmn, const char *why)
        gs.numdown--;
        dmn->connect_tries = 0;
        zlog_notice("%s state -> up : %s", dmn->name, why);
-       daemon_send_ready();
+       if (gs.numdown == 0)
+               daemon_send_ready(0);
        SET_WAKEUP_ECHO(dmn);
        phase_check();
 }
@@ -1030,6 +1048,12 @@ static char *translate_blanks(const char *cmd, const char *blankstr)
        return res;
 }
 
+static int startup_timeout(struct thread *t_wakeup)
+{
+       daemon_send_ready(1);
+       return 0;
+}
+
 static void watchfrr_init(int argc, char **argv)
 {
        const char *special = "zebra";
@@ -1037,6 +1061,9 @@ static void watchfrr_init(int argc, char **argv)
        struct daemon *dmn, **add = &gs.daemons;
        char alldaemons[512] = "", *p = alldaemons;
 
+       thread_add_timer_msec(master, startup_timeout, NULL, STARTUP_TIMEOUT,
+                             &gs.t_startup_timeout);
+
        for (i = optind; i < argc; i++) {
                dmn = XCALLOC(MTYPE_WATCHFRR_DAEMON, sizeof(*dmn));