DIR1 on Fedora 17 and recent Ubuntu

[Jeremy Allison]

On Fri, Nov 02, 2012 at 10:32:23AM +1100, Andrew Bartlett wrote:
> On Thu, 2012-11-01 at 12:20 -0700, Jeremy Allison wrote:
> > On Thu, Nov 01, 2012 at 12:53:26PM +0100, Jelmer Vernooij wrote:
> > > On Thu, 2012-11-01 at 14:13 +1100, Andrew Bartlett wrote:
> > > > I forget to mention this a while back, but the DIR1 test (as part of
> > > > make test, eg "make test TESTS=DIR1" seems to get into an infinite loop
> > > > on Fedora 17.  
> > > > 
> > > > What made me remember is that Jelmer mentioned it happens for him on
> > > > Ubuntu.  Presumably something has changed in the kernel, and I'm
> > > > wondering if someone who groks the low level filesystem stuff could have
> > > > a look into this?
> > > Thanks for sending an email about this.
> > > 
> > > I'm actually using Debian, running Linux 3.5.5-1~experimental.1 at the moment.
> > 
> > Is this with a 64-bit long ? I'm guessing this is the
> > problem when the reply from telldir() exceeds 32-bits,
> > which only hits clients doing the "old" SMBSearch calls
> > (DOS clients essentially). The DIR1 test does this also.
> > 
> > We have an outstanding bug on this but it's fiddly to
> > fix (essentially we need to keep a mapping of 64-bit
> > offsets to 32-bit indexes).
> 
> My filesystem is ext4 and I'm on 64 bit.

Then yeah, that's the bug.

> If it's unusual, hard, and difficult to fix, is there something we could
> do to avoid the CPU spin in make test?

We'll have to remove DIR1 from make test until we fix this I think.

Jeremy.