Re: [PATCH] NFS: Retry the CLOSE if the embedded GETATTR is rejected with ERR_STALE

[Trond Myklebust]

On Wed, 2020-11-18 at 21:29 +0000, Anchal Agarwal wrote:
> On Wed, Nov 18, 2020 at 03:17:20AM +0000, Trond Myklebust wrote:
> > CAUTION: This email originated from outside of the organization. Do
> > not click links or open attachments unless you can confirm the
> > sender and know the content is safe.
> > 
> > 
> > 
> > On Wed, 2020-11-18 at 00:24 +0000, Anchal Agarwal wrote:
> > > If our CLOSE RPC call is rejected with an ERR_STALE error, then
> > > we
> > > should remove the GETATTR call from the compound RPC and retry.
> > > This could happen in a scenario where two clients tries to access
> > > the same file. One client opens the file and the other client
> > > removes
> > > the file while it's opened by first client. When the first client
> > > attempts to close the file the server returns ESTALE and the file
> > > ends
> > > up being leaked on the server. This depends on how nfs server is
> > > configured and is not reproducible if running against nfsd.
> > 
> > That would be a seriously broken server. If you return
> > NFS4ERR_STALE to
> > the client, you cannot expect any further interaction with that
> > file
> > from the client. It won't try to send CLOSE or DELEGRETURN or any
> > other
> > stateful operation.
> > 
> In this scenario, the setup we have at EFS is more of a distributed
> fashion. Multiple
> clients are connected to multiple servers with a common filesystem.
> So the above
> scenario leads to leaked open file handles on the client that tries
> to close deleted
> file. So I was of the view, in that case client could retry close
> without getattr
> in the close sequence without anything to do on server side.


If you send the client an NFS4ERR_STALE, you are telling it that its
access to the file has been revoked. That is not a temporary error, it
is a fatal one. The client is not responsible for cleaning up any
state.

-- 
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@xxxxxxxxxxxxxxx