[Gluster-devel] Readdir d_off encoding

J. Bruce Fields bfields at fieldses.org
Mon Dec 22 21:36:33 UTC 2014


On Mon, Dec 22, 2014 at 04:20:58PM -0500, Jeff Darcy wrote:
> > Maybe, but that'd have a pretty high bar to clear: we definitely don't
> > want to break existing working xfs setups.
> 
> Darn right it's a high bar.  Is a single failed readdir under truly
> exceptional (and so far only theoretical) conditions "breakage"?

You're talking about the server-reboot case here?  I think failures
would be fairly easy to reproduce there, but without knowing what
exactly is proposed it's hard to be sure.

> If so,
> every distributed file system ever - including NFS - is broken in a
> hundred ways.  The severe and persistent errors that people have
> actually seen on ext4 matter.  If we say that we won't make even the
> smallest compromise on XFS behavior to fix them, we might as well come
> right out and say that ext4 users can go rot.  I'm not willing to say
> that.  I'd rather fix users' problems, in-tree if I can but out-of-tree
> if I have to.
> 
> I've said my piece.  Others can decide.  Merry Christmas.

Apologies for giving the impression that I've chosen a side here.  I
absolutely don't claim to know what the best approach is.

Mainly I'd just like to see a sketch of the design with a good
description of the failure cases so we're sure we understand the
tradeoffs.

--b.


More information about the Gluster-devel mailing list