[Cluster-devel] Re: [NFS] [PATCH 0/3] NLM lock failover

[James Yarbrough <jmy@sgi.com>]

> > take-over server:
> > B-1. mount the subject filesystem
> > B-2. "echo 1234 > /proc/fs/nfsd/nlm_set_ip_grace"
> > B-3. "rpc.statd -n 10.10.1.1 -N -P /shared_storage/sm_10.10.1.1"
> > B-4. bring up 10.10.1.1
> > B-5. re-export the filesystem
> 
> Umm, don't you want to do B-3 after B-4 and B-5 ?  Otherwise
> clients might racily fail on the first try.

I don't think they will necessrily fail.  It depends on whether the
server sends ICMP unreachable messages and how the client responds to
those.  In any case, I think the ordering should be B-5, B-4, and B-3
last.  One can argue about the ordering of B-3 and B-4, but if exporting
(B-5) does not happen before bringing up the IP address (B-4), clients
can get ESTALE replies.  For better transparency, it's probably best
to avoid ESTALE.

It's probably OK to do step B-3 after bringing up the IP address since
that will mimic what happens during boot.

> 
> Also, just curious here, when do you purge the clients' ARP caches?

I don't think you can actually do a purge from the server explicitly.
You should get the desired result when the IP address (10.10.1.1 in
the above example) is brought up.  There's a gratuitous ARP that goes
with that step.

-- 
jmy at sgi.com
650 933 3124

Blow up the mailbox!