Skip to contentSkip to conversation

core: fix unbound watchdog-notify for timeouts <2s

PATCH In review David Herrmann 2 participants 5 messages Sep 13, 2014 – Sep 14, 2014 Bluetooth drivers

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 core: fix unbound watchdog-notify for timeouts <2s

If the watchdog timeout is below 2s, we end up with a timeout of 0s as glib event source. This causes unbound watchdog notifications. Avoid this by never using event-timeouts below 1s.

Reported by Michael Biebl.

---
 src/main.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
src/main.c +2 -1
index e6bac6e..da0bd12 100644
@@ -596,9 +596,10 @@ int main(int argc, char *argv[])
 
 		seconds = atoi(watchdog_usec) / (1000 * 1000);
 		info("Watchdog timeout is %d seconds", seconds);
+		seconds = (seconds < 4) ? 1 : seconds / 2;
 
 		watchdog = g_timeout_add_seconds_full(G_PRIORITY_HIGH,
-							seconds / 2,
+							seconds,
 							watchdog_callback,
 							NULL, NULL);
 	} else
Signature · 1 line
2.1.0
maintainer
#1 lore raw

Hi David,

Quoted 15 lines

If the watchdog timeout is below 2s, we end up with a timeout of 0s as glib event source. This causes unbound watchdog notifications. Avoid this by never using event-timeouts below 1s.

Reported by Michael Biebl.

---
src/main.c | 3 ++-

1 file changed, 2 insertions(+), 1 deletion(-)
src/main.c +0 -0
index e6bac6e..da0bd12 100644

src/main.c int main(int argc, char *argv[]) · line 604

Quoted 10 lines
src/main.c +2 -1
@@ -596,9 +596,10 @@ int main(int argc, char *argv[])
 
 		seconds = atoi(watchdog_usec) / (1000 * 1000);
 		info("Watchdog timeout is %d seconds", seconds);
+		seconds = (seconds < 4) ? 1 : seconds / 2;
 
 		watchdog = g_timeout_add_seconds_full(G_PRIORITY_HIGH,
-							seconds / 2,
+							seconds,
 							watchdog_callback,
 							NULL, NULL);

while setting WatchdogSec=1 is pretty insane interval anyway, this now also means that setting it that low will cause a race condition between systemd's watchdog killing us and we are able to be quickly enough to respond with a keep alive message.

I get the feeling that if > 2s, then we should use g_timeout_add_seconds to get the advantage of being woken up with all other timeouts in our daemon. And when it is <= 2s, then we better use a high precision g_timeout_add.

One alternative is to actually use timerfd natively here and use kernel based range timers.

Regards

Marcel

#2 lore raw

Hi

Quoted 1 line · Marcel Holtmann

while setting WatchdogSec=1 is pretty insane interval anyway, this now also means that setting it that low will cause a race condition between systemd's watchdog killing us and we are able to be quickly enough to respond with a keep alive message.

Whoops, indeed. The patch is stupid. And yes, WatchdogSec=1 is insane, but I guess we should at least have a sane behavior. I mean, it's user-config so someone will (and currently is!) trigger this.

Quoted 1 line

I get the feeling that if > 2s, then we should use g_timeout_add_seconds to get the advantage of being woken up with all other timeouts in our daemon. And when it is <= 2s, then we better use a high precision g_timeout_add.

This sounds actually good. I will do that in v2.

Thanks
David

Quoted 6 lines

One alternative is to actually use timerfd natively here and use kernel based range timers.

Regards

Marcel

#3 lore raw

Hi

Quoted 3 lines · David Herrmann
Quoted 1 line

I get the feeling that if > 2s, then we should use g_timeout_add_seconds to get the advantage of being woken up with all other timeouts in our daemon. And when it is <= 2s, then we better use a high precision g_timeout_add.

This sounds actually good. I will do that in v2.

Ok, glib only supports millisecond precision, so whatever we do, there will be a WATCHDOG_USEC range that we cannot react to properly. Ok, to be honest, if WATCHDOG_USEC is below 1ms, we're screwed anyway, so maybe we should just keep our current behavior.. I mean there is nothing sane to do for use if users specify such low watchdog ranges. At some point, we have to resort to either run watchdogs without a timeout or bail out and tells users to stop using such low values. Whether this is at 1s or at 1ms doesn't matter much, does it?

So I think we can just keep the current behavior.

Thanks
David

maintainer
#4 lore raw

Hi David,

Quoted 4 lines
Quoted 3 lines
Quoted 1 line

I get the feeling that if > 2s, then we should use g_timeout_add_seconds to get the advantage of being woken up with all other timeouts in our daemon. And when it is <= 2s, then we better use a high precision g_timeout_add.

This sounds actually good. I will do that in v2.

Quoted 8 lines

Ok, glib only supports millisecond precision, so whatever we do, there will be a WATCHDOG_USEC range that we cannot react to properly. Ok, to be honest, if WATCHDOG_USEC is below 1ms, we're screwed anyway, so maybe we should just keep our current behavior.. I mean there is nothing sane to do for use if users specify such low watchdog ranges. At some point, we have to resort to either run watchdogs without a timeout or bail out and tells users to stop using such low values. Whether this is at 1s or at 1ms doesn't matter much, does it?

I am fine with just bailing out and terminating the daemon with an error exit code in case the watchdog timeout is < 2 seconds.

Regards

Marcel