[Gluster-devel] Readdir d_off encoding

[J. Bruce Fields]

On Mon, Dec 22, 2014 at 03:20:48PM -0500, Jeff Darcy wrote:
> > I don't love the bit-stealing hack either, but in practice keep in mind
> > this all seems to be about ext4.  If you want reliable nfs readdir with
> > 16M-entry directories and all the rest you can get that already with
> > xfs.
> 
> Do we know it's just ext4?

That (and ext3) is the only place I've seen the problem.  But we're
definitely depending on undocumented behavior, so I certainly can't rule
out the possibility some other (current or future) filesystem doing
something weird.

> I'm not even sure that's all of the newer
> or less well known Linux file systems are immune, let alone those on
> BSD or other platforms where GlusterFS must run.  The only other FS that
> I'm sure doesn't have this problem is ZFS.  My worry is that, to someone
> who has already been affected by this, telling them it doesn't happen on
> XFS is like pouring salt in the wound.  They're more likely to abandon
> GlusterFS than embrace XFS.
> 
> *In practice* I believe that ext4 is a reality we must deal with.  Any
> discussion of this bug's frequency, or of solutions' efficacy, has to
> take that into account.  If solution X works perfectly on XFS but still
> fails regularly on ext4, then for evaluation purposes it's the "fails
> regularly" part that counts.  A solution Y which fails less regularly
> on both platforms might well be preferable, even if it's worse than X
> on XFS.

Maybe, but that'd have a pretty high bar to clear: we definitely don't
want to break existing working xfs setups.

--b.