<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0 Transitional//DE"><HTML><HEAD><META HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=us-ascii"><TITLE></TITLE></HEAD><BODY><div style="font-family:courier new,courier;font-size:8pt;">   Hello systemd experts and developers, I recently stumbled over the bug with the watchdog mechanism that has already been reported to free desktop bugzilla (56109). I analyzed the bug and came to a simple solution for solving it. First, what I think is going on: - watchdog timeout is detected in service_handle_watchdog(), service_enter_dead(…) is called - service_enter_dead() sets the service state to auto_restart - triggered by a timer, service_enter_restart is called - service_enter_restart schedules a restart job - systemd splits up the jobs into a stop and a start job and schedules both - the stop job lasts to a call of service_stop() - here it begins to get interesting: - based on the AUTO_RESTART state, this function decides to go directly into dead state, nothing of the normal stopping procedure is done. This is probably because in most cases that cause a restart to be scheduled the stop proceeding is done automatically (for instance in case of a killed or normally exiting service.). But this is not true for a watchdog timeout. Nothing of the stop proceeding is executed in case of such a timeout. So the process that missed to send the watchdog event is going on to life (in which state ever). No one is cleaning up. A second instance of the service is started. My suggestion to solve this: Changes are needed in service.c in service_stop(…). change: /* A restart will be scheduled or is in progress. */ if (s->state == SERVICE_AUTO_RESTART) { service_set_state(s, SERVICE_DEAD); return 0; } to: /* A restart will be scheduled or is in progress. In all cases but the watchdog timeout, stop is already progressed by systemd automatically*/ if (s->state == SERVICE_AUTO_RESTART && s->result != SERVICE_FAILURE_WATCHDOG) { service_set_state(s, SERVICE_DEAD); return 0; } and change: assert(s->state == SERVICE_RUNNING || s->state == SERVICE_EXITED); to: assert(s->state == SERVICE_RUNNING || s->state == SERVICE_AUTO_RESTART || s->state == SERVICE_EXITED); I tested the following: - the watchdog mechanism is now actually stopping / killing the service in case it is not sending the watchdog event right in time - a restart triggered by a killed service works like before Hopefully, I didn’t miss some side effects caused by my changes. Any opinions on my proposed changes? Kind regards, Marko Hoyer </div> --- <div style="font-family:arial;font-size:10pt;">Alle Postfächer an einem Ort. Jetzt wechseln und E-Mail-Adresse mitnehmen! <a style="font-family:arial;font-size:10pt;" href="http://email.freenet.de/produkte/basic/index.html" target="_blank">Rundum glücklich mit freenetMail</a></div> </BODY></HTML>