Hi,

On Mon, Nov 14, 2011 at 01:11:02PM +0100, Lars Ellenberg wrote:
> On Mon, Nov 14, 2011 at 10:54:47AM +0100, Ulrich Windl wrote:
> > Hello!
> > 
> > We see significant performance issues with the LVM RA when writing to
> > an ext3 filesystem with default journaling and berriers enabled. So
> > anything that speeds up the RA monitor operation (it needs up to 90
> > seconds here accasionally) would be appreciated.
> > 
> > However I feel the best thing would eb to fix the perfomance issues in
> > the LVM commands, rather than not using them.
> > 
> > For the vgck: If the VG metadata are checked during startup, a
> > periodic re-check seems unnecessary to me: During normal (use)
> > operation of the VG's LVs (that's the typical use case for VGs in a
> > cluster configuration) the (checkable) metadata don't change a lot, so
> > it may not make much sense. I'd rely on the kernel to shutdown (i.e.
> > put into read-only) a faulty VG anyway.
> > 
> > The paranoid should monitor individual LVs anyway instead of just
> > monitoring the VG.
> > 
> > Eventually the documentations says:
> > "The 'monitor' operation reports whether the volume seems present"
> > 
> > So probably a [ -r /dev/mapper/VG-name ] should also do ;-)
> 
> As you usually have something *using* it, Filesystem, VirtualDomain or
> whatever, monitoring that should be enough.
> Just drop monitoring of the VG primitive,
> there is not much point doing so anyways.

Good idea :)

> > I see that vgck is used even after LVM_status was successful;
> > LVM-status uses vgdisplay to check the status. There's another funny
> > thing in my version (SLES11 SP1): The check is:
> > vgdisplay -v $1 2>&1 | grep -i 'Status[ \t]*available' 2>&1 >/dev/null
> > 
> > but there is not "available status" for the VG, only for the LV! 
> > # vgdisplay -v VG |grep Status
> >     Using volume group(s) on command line
> >     Finding volume group "VG"
> >     Do not wait for udev on information only command
> >     Do not wait for udev on information only command
> >     Do not wait for udev on information only command
> >     Do not wait for udev on information only command
> >     Do not wait for udev on information only command
> >     Do not wait for udev on information only command
> >     Do not wait for udev on information only command
> >     Do not wait for udev on information only command
> >     Do not wait for udev on information only command
> >     Do not wait for udev on information only command
> >     Do not wait for udev on information only command
> >     Do not wait for udev on information only command
> >     Do not wait for udev on information only command
> >     Do not wait for udev on information only command
> >     Do not wait for udev on information only command
> >     Do not wait for udev on information only command
> >     Do not wait for udev on information only command
> >     Do not wait for udev on information only command
> >     Do not wait for udev on information only command
> >   VG Status             resizable
> >   LV Status              available
> >   LV Status              available
> >   PV Status             allocatable
> > 
> > So the VG is not "available", but "resizable". Well, there are many
> > ways to write a resource agent, but only few of them are correct...

Hmm, did you actually have a problem here? Or do you just think
that there could be one?

> Thats actually ok: if the "VG" is "active" its "LV"s are "available".
> Well, unless you have a VG without any LVs, which, again, is a bit pointless.
> Just compare the output of an active an an inactive VG.

Xin Wei suggested to drop vgck from the monitor operation and
submitted the patch.

Thanks,

Dejan
_______________________________________________
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