[DRBD-user] RH Cluster Integration

Florian Haas florian.haas at linbit.com
Thu Mar 13 20:25:40 CET 2008

Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.


On Monday 10 March 2008 13:00:27 drbd at bobich.net wrote:
> Hi,
>
> Is there a way to integrate DRBD with RHC? I've found documentation on
> heartbeat integration, but nothing specific for RHCS.
>
> In particular, if I have 2 nodes connected, and one of them disconnects
> briefly, DLM will lock up and wait for the nodes to reconnect or for one
> node to get fenced. DRBD will also disconnect. If, before fencing occurs,
> the network connection is re-established, GFS/DLM will re-connect, but
> DRBD will not. GFS will remain convinced that the file system is shared
> when the underlying DRBD volume is actually split-brained.

There is a section on GFS in the user's guide. Does that not help? 

> Is there a way to make DRBD force fencing of the other node on disconnect
> using the RHC hooks? Failing that, a direct fencing script hook would do.
>
> I am currently using these SB resolutions:
>
> after-sb-0pri   discard-younger-primary;
> after-sb-1pri   discard-secondary;
> after-sb-2pri   call-pri-lost-after-sb;
>
> I can see in the docs that if it is decided that the current secondary is
> current, the primary gets panicked. But what happens of both nodes
> disconnect and decide to go into Primary/Unknown mode? Is there a hook
> that gets called when this is detected (if it is detected) that can be
> used to initiate fencing? This would be avoided if the nodes tried to
> re-sync after the connection is established, but they don't appear to do
> that for whatever reason. Am I missing an important setting somewhere?

Yes. "fencing resource-and-stonith" combined with an outdate-peer handler that 
invokes an RHCS compliant fencing script. That ought to do it. It's brutal, 
but will prevent split brain. The two nodes will attempt to kill each other 
and one node wins. Since the RHCS fencing scripts don't return unless the 
peer is confirmed dead, it shouldn't ever happen that both nodes die.

See if that works for you and if it does, I'll add it to the GFS chapter in 
the guide. :-)

Florian

-- 
: Florian G. Haas 
: LINBIT Information Technologies GmbH
: Vivenotgasse 48, A-1120 Vienna, Austria

Please note: when replying, there is no need to CC my 
personal address. Replying to the list is fine. Thank you.



More information about the drbd-user mailing list