Hi again,

Replying to myself because I didn't spot the problem the first
time.

On Fri, Apr 11, 2008 at 01:39:17PM +0200, Dejan Muhamedagic wrote:
> Hi Yamauchi-san,
> 
[snip]
> 
> There are three big time gaps where crmd was waiting:
> 
> tengine[558]: 2008/04/11_16:27:02 info: send_rsc_command: Initiating action 
> 7: prmIpPostgreSQLDB_start_0 on dl380g5d
> crmd[553]: 2008/04/11_16:27:21 info: handle_shutdown_request: Creating 
> shutdown request for dl380g5c
> tengine[558]: 2008/04/11_16:28:02 WARN: action_timer_callback: Timer popped 
> (abort_level=1000000, complete=false)
> tengine[558]: 2008/04/11_16:28:02 WARN: print_elem: Action missed its 
> timeout[Action 7]: In-flight (id: prmIpPostgreSQLDB_start_0, loc: dl380g5d, 
> priority: 0)

The action crmd is waiting for was supposed to happen on
dl380g5d, but the network is down so dl380g5c is waiting in
vain. This is what happened:

1. dl380g5c (the DC) looses connection
2. admin issues a heartbeat stop on dl380g5c
3. crmd, correctly, wants to move resources to dl380g5d before
   the shutdown
4. after "deadtime" ccm delivers the new connectivity
   information, i.e. a split brain occurs
5. crmd (actually tengine) is still waiting for the result of the
   operation it initiated even though it can't hear from that
   node anymore; first it times out after one minute and
   then after three minutes when it finally gives up

You should file a bugzilla for pacemaker for CRM/tengine
component. You can insert the above analysis.

BTW, if you were to wait with the shutdown until the "deadtime"
passed, then I guess that you would get very different results,
i.e. a much faster shutdown.

Thanks,

Dejan

[snip]
_______________________________________________
Linux-HA mailing list
[email protected]
http://lists.linux-ha.org/mailman/listinfo/linux-ha
See also: http://linux-ha.org/ReportingProblems

Reply via email to