[Linux-cluster] bug in GFS2?

[Vladislav Bogdanov]

26.09.2013 10:59, Pavel Herrmann wrote:
> Hi
> 
> On Thursday 26 of September 2013 08:28:50 Vladislav Bogdanov wrote:
>> 25.09.2013 17:25, Pavel Herrmann wrote:
>>> Hi
>>>
>>> I am trying to build a two-node cluster for samba, but I'm having some
>>> GFS2
>>> issues.
>>>
>>> The nodes themselves run as virtual machines in KVM (on different hosts),
>>> use gentoo kernel 3.10.7 (not sure what exact version of vanilla it is
>>> based on), and I use the cluster-next stack in somewhat minimal
>>> configuration (corosync-2 with DLM-4, no pacemaker).
>>
>> Just a note.
>> dlm-4 (and thus gfs) requires stonith-ng subsystem of pacemaker to be
>> running, otherwise it is unable to query/perform fencing and many funny
>> things may happen. I believe something was done in the pacemaker code to
>> allow stonithd (daemon implementing stonith-ng) to be run independently
>> of the rest of pacemaker.
> 
> From looking at a bit of fencing code, there was no (obvious) dependency on 
> pacemaker. I do have fencing set up, using a custom script that connects to 
> the other nodes qemu console and forcibly reboots it. In testing (that is 
> stopping one of the nodes from said console) it worked perfectly, but in the 
> case of this lockup it was not invoked (logging the date is of course part of 
> the script).

Ok, got it.
I actually meant default dlm setup, it runs /usr/sbin/dlm_stonith which
uses stonith_api_*_helper() functions defined inline in pacemaker's
stonith-ng API, which in turn dlopen libstonithd.so.2 and use symbols
from there.