[Pacemaker] [Question and Problem] In vSphere5.1 environment, IO blocking of pengine occurs at the time of shared disk trouble for a long time.

[renayama19661014 at ybb.ne.jp]

Hi Andrew,
Hi Vladislav,

I try whether this correction is effective for this problem.
 * https://github.com/beekhof/pacemaker/commit/eb6264bf2db395779e65dadf1c626e050a388c59

Best Regards,
Hideo Yamauchi.

--- On Thu, 2013/5/16, Andrew Beekhof <andrew at beekhof.net> wrote:

> 
> On 16/05/2013, at 3:49 PM, Vladislav Bogdanov <bubble at hoster-ok.com> wrote:
> 
> > 16.05.2013 02:46, Andrew Beekhof wrote:
> >> 
> >> On 15/05/2013, at 6:44 PM, Vladislav Bogdanov <bubble at hoster-ok.com> wrote:
> >> 
> >>> 15.05.2013 11:18, Andrew Beekhof wrote:
> >>>> 
> >>>> On 15/05/2013, at 5:31 PM, Vladislav Bogdanov <bubble at hoster-ok.com> wrote:
> >>>> 
> >>>>> 15.05.2013 10:25, Andrew Beekhof wrote:
> >>>>>> 
> >>>>>> On 15/05/2013, at 3:50 PM, Vladislav Bogdanov <bubble at hoster-ok.com> wrote:
> >>>>>> 
> >>>>>>> 15.05.2013 08:23, Andrew Beekhof wrote:
> >>>>>>>> 
> >>>>>>>> On 15/05/2013, at 3:11 PM, renayama19661014 at ybb.ne.jp wrote:
> >>>>>>>> 
> >>>>>>>>> Hi Andrew,
> >>>>>>>>> 
> >>>>>>>>> Thank you for comments.
> >>>>>>>>> 
> >>>>>>>>>>> The guest located it to the shared disk.
> >>>>>>>>>> 
> >>>>>>>>>> What is on the shared disk?  The whole OS or app-specific data (i.e. nothing pacemaker needs directly)?
> >>>>>>>>> 
> >>>>>>>>> Shared disk has all the OS and the all data.
> >>>>>>>> 
> >>>>>>>> Oh. I can imagine that being problematic.
> >>>>>>>> Pacemaker really isn't designed to function without disk access.
> >>>>>>>> 
> >>>>>>>> You might be able to get away with it if you turn off saving PE files to disk though.
> >>>>>>> 
> >>>>>>> I store CIB and PE files to tmpfs, and sync them to remote storage
> >>>>>>> (CIFS) with lsyncd level 1 config (I may share it on request). It copies
> >>>>>>> critical data like cib.xml, and moves everything else, symlinking it to
> >>>>>>> original place. The same technique may apply here, but with local fs
> >>>>>>> instead of cifs.
> >>>>>>> 
> >>>>>>> Btw, the following patch is needed for that, otherwise pacemaker
> >>>>>>> overwrites remote files instead of creating new ones on tmpfs:
> >>>>>>> 
> >>>>>>> --- a/lib/common/xml.c  2011-02-11 11:42:37.000000000 +0100
> >>>>>>> +++ b/lib/common/xml.c  2011-02-24 15:07:48.541870829 +0100
> >>>>>>> @@ -529,6 +529,8 @@ write_file(const char *string, const char *filename)
> >>>>>>>      return -1;
> >>>>>>>  }
> >>>>>>> 
> >>>>>>> +    unlink(filename);
> >>>>>> 
> >>>>>> Seems like it should be safe to include for normal operation.
> >>>>> 
> >>>>> Exactly.
> >>>> 
> >>>> Small flaw in that logic... write_file() is not used anywhere.
> >>> 
> >>> Heh, thanks for spotting this.
> >>> 
> >>> I recall write_file() was used for pengine, but some other function for
> >>> CIB. You probably optimized that but forgot to remove unused function,
> >>> that's why I was sure patch is still valid. And I did tests (CIFS
> >>> storage outage simulation) only after initial patch, but not last years,
> >>> that's why I didn't notice the regression - storage uses pacemaker too ;) .
> >>> 
> >>> This should go to write_xml_file() (And probably to other places just
> >>> before fopen(..., "w"), f.e. series).
> >> 
> >> I've consolidated the code, however adding the unlink() would break things for anyone intentionally symlinking cib.xml from somewhere else (like a git repo).
> >> So I'm not so sure I should make the unlink() change :(
> > 
> > Agree.
> > I originally made it specific to pengine files.
> > What do you prefer, simple wrapper in xml.c (f.e.
> > unlink_and_write_xml_file()) or just add unlink() call to pengine before
> > it calls write_xml_file()?
> 
> The last one :)
> _______________________________________________
> Pacemaker mailing list: Pacemaker at oss.clusterlabs.org
> http://oss.clusterlabs.org/mailman/listinfo/pacemaker
> 
> Project Home: http://www.clusterlabs.org
> Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf
> Bugs: http://bugs.clusterlabs.org
>