Re: [PATCH xfstests] generic/035: Override output for NFS testing

[Trond Myklebust <trondmy@primarydata.com>]

On Tue, 2018-04-03 at 02:45 -0700, Christoph Hellwig wrote:
> On Thu, Mar 29, 2018 at 11:34:39AM -0400, Benjamin Coddington wrote:
> > We'd like to run generic tests for NFS, but often have slightly
> > different
> > output for our results.  One instance is that for the NFS client
> > the
> > removal of an open file or directory is handled differently than
> > for a
> > local filesystem.  We can expect nlink to be 1 for files, and to
> > receive
> > -ESTALE for operations on deleted directories, isn't that silly?
> 
> NFS is simply buggy in this case, and we should at least xfail the
> test
> case, not make it look fine.
> 
> I'd rather have a file that lists expected fails per file system with
> an
> explanation than a hack like this that papers over the issue.

IIRC that ESTALE test is hitting a protocol issue: NFS doesn't have
stateful readdir() (or any stateful directory operations), and so there
is nothing to tell the server to pin the removed directory while we
have it open in the VFS layer on the client.

I'm fine either way, but if we make the decision to call out protocol
issues as 'expected failure' then we need to make that a consistent
policy for all xfstests.

-- 
Trond Myklebust
Linux NFS client maintainer, PrimaryData
trond.myklebust@primarydata.com