[dm-devel] Re: [RFC] [PATCH] lvm2: mirroredlog support

[malahal at us.ibm.com]

Stefan Raspl [raspl at linux.vnet.ibm.com] wrote:
> If I understand Takahiro correctly, the main issue is that the whole
> log device becomes unusable in case one leg is broken...? Of course,
> we still have the other leg with the most recent data available and
> can read from that one. However, this would require that we can
> start a mirror in rw-mode (so we can keep the still-recent half up
> to date) even if one half has failed - a functionality which was
> recently added to LVM, as we discussed in our last call...?

Stefan, Takahiro is using the latest LVM that brings LVs in rw-mode.
That is not the problem. Actually, everything works if a slave leg of a
"log mirror" is broken. But if the master leg of the "log mirror" is
broken, entire log device becomes unusable. The reason is when we
activate the "log mirror", it always synchronizes the data from the
master leg to slave leg since it has no log device (as the "log mirror"
uses "corelog"). In other words, the "log mirror" always starts in "not
in sync" state and synchronizes.

If the master leg is broken for the "log mirror", it starts in "not in
sync" state and tries to copy data from the "broken leg" to the active
leg! This appears as the master leg failure while doing synchronization
which makes the whole device unusable.

> I'm not sure if I understand his statements about the
> synchronization correctly. Why would any synchronization within the
> data mirror occur at all? It's only the log which is broken, not the
> actual data in vg00-lv00_mimage_0/1. If any synchronization in the
> _data_ mirror occurs just because one leg in the _log_ mirror has a
> problem, I'd consider that a bug.

Since the entire log device is unusable with its "master device
broken", the "data mirror" goes for synchronization.