Date: Tue, 15 Nov 2011 17:38:11 -0500
From: Matthew Treinish <treinish@linux.vnet.ibm.com>
To: Trond Myklebust <Trond.Myklebust@netapp.com>
Cc: NeilBrown <neilb@suse.de>, Christoph Hellwig <hch@infradead.org>,
        linux-nfs@vger.kernel.org
Subject: Re: [PATCH/RFC 0/7] Volatile Filehandle Client-side Support
Message-ID: <20111115223811.GA3687@Gelgoog.pok.ibm.com>
References: <1321052673-22171-1-git-send-email-treinish@linux.vnet.ibm.com>
 <1321056809.8733.2.camel@lade.trondhjem.org>
 <20111112144953.GA3740@infradead.org>
 <20111113145400.6c7a9be3@notabene.brown>
 <20111113180633.GA4084@Gelgoog.pok.ibm.com>
 <1321339791.8267.17.camel@lade.trondhjem.org>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
In-Reply-To: <1321339791.8267.17.camel@lade.trondhjem.org>
Sender: linux-nfs-owner@vger.kernel.org

On Tue, Nov 15, 2011 at 08:49:51AM +0200, Trond Myklebust wrote:
> On Sun, 2011-11-13 at 13:06 -0500, Matthew Treinish wrote: 
> > On Sun, Nov 13, 2011 at 02:54:00PM +1100, NeilBrown wrote:
> > > On Sat, 12 Nov 2011 09:49:53 -0500 Christoph Hellwig <hch@infradead.org>
> > > wrote:
> > > 
> > > > On Fri, Nov 11, 2011 at 07:13:29PM -0500, Trond Myklebust wrote:
> > > > > On Fri, 2011-11-11 at 18:04 -0500, Matthew Treinish wrote: 
> > > > > > This patch series implements client side support for volatile file handle
> > > > > > recovery (RFC 3530 section 4.2 and 4.3) with walk back using the dcache. To
> > > > > > test the client you either need a server that supports volatile file handles or 
> > > > > > you can hard code the server to output NFS4ERR_FHEXPIRED instead of
> > > > > > NFSERR_STALE. (See the last patch in the series)
> > > > > 
> > > > > WHY do we want to support this kind of "feature"? As you said, the RFC
> > > > > doesn't actually help in figuring out how this crap is supposed to work
> > > > > in practice, so why do we even consider starting to give a damn?
> > > > 
> > > > *nod*. Pretending we handle it seems fairly dangerous.  I'd much prefer
> > > > outright rejecting it.
> > > 
> > > Hence the suggested mount option.
> > > 
> > > A server might not be able to provide stable file handles, but can ensure
> > > that files don't get renamed - for these filesystems, the name is a
> > > reliable stable handle for the file (it just doesn't fit in the NFSv4 file
> > > handle structure).
> > > 
> > > So if you know the filesystem will only return FHEXPIRED for filehandles
> > > belonging to files that cannot be renamed, then it is perfectly reasonable to
> > > repeat the name lookup to re-access the file after the server forgets about
> > > an old filehandle.  The mount option is how you communicate this knowledge,
> > > because the RFC doesn't provide a way to communicate it.
> > > 
> > This was one of 2 reasons for implementing this, and we actually run into this with 
> > certain z/OS systems, because the z/OS NFS server currently uses FHEXPIRED in this way.
> 
> So you're both basically saying that 'we know that this is a bad idea,
> so let's punt it to the users and assume they will know those few
> exceptions when it is safe to use'?
> In that case, are you planning on documenting what constitutes safe
> usage? So far, I've seen nothing either in the discussion here or in the
> changelogs that explains precisely when you can safely enable this mount
> option.
> 
> Note that just disabling renames is, as I stated yesterday, not a
> sufficient condition. You pretty much need a read-only filesystem
> situation, in which case you can easily devise persistent filehandle
> solutions that work just as well. 
> 
Yes, I agree documenting the risks associated with the mount option is a 
necessity, but something that I clearly overlooked. How about something like:

   This option enables volatile filehandle recovery by re-lookup 
   on FHEXPIRED errors. Only use this mount option if the 
   filenames/paths on the server are not going to change from the
   initial expiration until all the recovery operations complete.
   Otherwise the validity of the files from the server can not be
   guaranteed. It can only truly be considered safe to use on a
   linux server, if the filesystem is read-only.

> > The other thought was that this could be used for migration/replication 
> > between file synced servers. So, if we wanted to switch/move to another server where 
> > the file names were the same but all the inode numbers were different you could use 
> > this to refresh the invalid file handles on the new server.
> 
> This runs into the rename problem. How do you guarantee that the files
> haven't been renamed before the migration event occurred? How does the
> client identify that the file is the same one when it looks it up on the
> new server?
> 

I don't think there is a way to guarantee that the files haven't been renamed
before the migration event. It would probably only be fully safe under the same
conditions as above.