From: Kedar Sovani <kedars@hotpop.com>
Subject: Re: [PATCH / RFC] nfs-utils: High Availability NFS
Date: 03 Sep 2004 12:58:45 +0530
Sender: nfs-admin@lists.sourceforge.net
Message-ID: <1094196525.2283.3.camel@localhost.localdomain>
References: <4124DB86.9060505@steeleye.com>
	 <16677.22269.988036.787320@cse.unsw.edu.au> <412E1C10.5020703@steeleye.com>
	 <20040827073049.GA23324@suse.de> <412F3362.5040707@steeleye.com>
	 <20040830082510.GD28454@suse.de>
Mime-Version: 1.0
Content-Type: text/plain
Cc: Paul Clements <paul.clements@steeleye.com>,
	Neil Brown <neilb@cse.unsw.edu.au>, nfs@lists.sourceforge.net
To: Olaf Kirch <okir@suse.de>
In-Reply-To: <20040830082510.GD28454@suse.de>
Errors-To: nfs-admin@lists.sourceforge.net

NFS Server keeps state about the recent responses given to the clients,
in a reply cache.

Shouldn't that state be replicated to the failed-over NFS server as well
?

Kedar.

On Mon, 2004-08-30 at 13:55, Olaf Kirch wrote:
> On Fri, Aug 27, 2004 at 09:13:06AM -0400, Paul Clements wrote:
> > With so many things moving out of the kernel to userland, it's 
> > surprising to see statd moving into the kernel. And it certainly does 
> > make modifications, such as this, much more difficult.
> 
> Well, the issue with statd is that it exists _only_ to make lockd happy.
> But what it really does is make lockd awfully complicated because we
> have to do upcalls all the time.
> 
> So my rationale for moving statd into the kernel is to actually eliminate
> a lot of code and have a minimal set up functionality in there that does
> exactly what lockd needs, no more. In particular, no more SM_MON and
> SM_UNMON support that allows untrusted hosts to do sneaky stuff - instead,
> lockd writes the /var/lib/nfs/sm files directly. The only thing my kernel
> statd supports is SM_NOTIFY, and that even is delivered directly to the
> locking code.
> 
> > But the callout happens before statd replies to the client, so there's 
> > no chance that we lose any clients if there's a failure. If we use 
> > dnotify then we only find out there's been a change after the change has 
> > occurred, which means the client may have already gotten a reply saying 
> > he's been added to the notify list. If we get a failure at this point, 
> > we've lost a client, and upon failover the client's locks are no longer 
> > valid.
> 
> Doesn't a HA NFS deployment require a shared disk anyway? Why not just
> move /var/lib/nfs to this file system so it gets shared the same way you
> share the rest of your data?
> 
> Olaf


-------------------------------------------------------
This SF.Net email is sponsored by BEA Weblogic Workshop
FREE Java Enterprise J2EE developer tools!
Get your free copy of BEA WebLogic Workshop 8.1 today.
http://ads.osdn.com/?ad_id=5047&alloc_id=10808&op=click
_______________________________________________
NFS maillist  -  NFS@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/nfs