[dm-devel] [RFC] multipath userspace support

[Kevin Corry]

Hi Christophe,

On Wednesday 10 March 2004 5:11 pm, christophe varoqui wrote:
> Hello,
>
> I just read in the EVMS 2.3.0 changelog that some sort of support for
> the DM multipath was introduced.

EVMS has actually had multipath support for quite a while now, which utilized 
the MD driver's multipath personality. However, we've been trying to migrate 
more stuff off of MD and onto DM. So for the latest EVMS release, we 
converted the multipath portion of the MD plugin to use the new dm-mpath 
module. (We also switched to using DM for raid-linear and raid-0 devices. Our 
hope is that the remaining raid levels could be switched over in the future 
as well...or at least some sort of merging of the two drivers.)

> It was previously stated on this mailing list that LVM2 tools may be
> extended to support it too.
>
> Which leads me to ask : given the current status of the multipath-tools,
> don't you agree that the multipath userspace handling does not belong in
> volume managers like LVM2 or EVMS ?
>
> The multipath-tools package tends to prove that this handling can be
> done 100% automagically by leveraging the hotplug subsys and the core DM
> events. Most users won't need more than that. Why force them to use a
> sophisticated volume manager they don't necessarily need ?

At this point, I see the multipath support in EVMS and the multipath-tools 
package as somewhat complimentary. From my rather brief review of 
multipath-tools, it appears to do MP detection based on SCSI information from 
sysfs. EVMS, on the other hand, does MP detection based on on-disk metadata 
(such as MD superblocks). *Ideally* the scsi-based MP detection will work for 
most people, but it's not clear to me yet that it will work for everyone. And 
since there are already people using the on-disk metadata approach (who have 
specifically asked us to continue supporting it), for the time being we will 
still need to provide this alternative discovery method.

As multipath-tools matures, we'll be able to get these people to test with it 
and determine if it will handle their setups. Over time, hopefully the 
on-disk metadata approach will become unnecessary.

-- 
Kevin Corry
kevcorry at us.ibm.com
http://evms.sourceforge.net/