Re: [PATCH 16/20] ceph: nfs re-export support

["J. Bruce Fields" <bfields@fieldses.org>]

On Fri, Jul 17, 2009 at 09:49:06AM -0700, Sage Weil wrote:
> On Fri, 17 Jul 2009, J. Bruce Fields wrote:
> > On Thu, Jul 16, 2009 at 03:07:35PM -0700, Sage Weil wrote:
> > > On Thu, 16 Jul 2009, Trond Myklebust wrote:
> > > > On Thu, 2009-07-16 at 12:50 -0700, Sage Weil wrote:
> > > > > On Thu, 16 Jul 2009, J. Bruce Fields wrote:
> > > > > > On Wed, Jul 15, 2009 at 02:24:46PM -0700, Sage Weil wrote:
> > > > > > > Basic NFS re-export support is included.  This mostly works.  However,
> > > > > > > Ceph's MDS design precludes the ability to generate a (small)
> > > > > > > filehandle that will be valid forever, so this is of limited utility.
> > > > > > 
> > > > > > Is there any hope of fixing that?
> > > > > 
> > > > > Yes, but it requires some additional ondisk metadata the MDS isn't 
> > > > > maintaining yet (a parent directory backpointer on file objects).  
> > > > > 
> > > > > The MDS changes will mean more random IO for rename intensive workloads, 
> > > > > but the backpointers would also be useful for rebuilding the directory 
> > > > > tree in the event of some catastrophic metadata loss or corruption.  
> > > > > (Currently they're only there for directories, not all files.)
> > > > 
> > > > Note that a filehandle that contains parent directory information is
> > > > still not one that is valid forever. It will change in the case of a
> > > > cross-directory rename, and so isn't a filehandle in the NFSv2/v3 sense.
> > > > Even in the NFSv4 case, it would have to be labelled as 'volatile'.
> > > 
> > > Right.  The parent directory information in the fh it used as a hint, but 
> > > can't be relied on because of the rename problem.  That's exactly why the 
> > > Ceph MDS will need to be changed to maintain backpointers on all files, 
> > > not just directories.  When that happens, reexporting via NFS will work 
> > > reliably.  Until then, old and idle filehandles for renamed files will 
> > > eventually go stale.
> > 
> > Maybe I should look again at the patch instead of continuing to ask,
> > but....  I'm confused: how will having a backpointers from inodes to
> > directories help do the filehandle-to-inode lookup?  (If you can't look
> > up the inode in the first place, what use is any pointer stored in that
> > inode?)
> 
> The backpointers will be on the first file data or directory metadata 
> object in the object store, which is random access.  The inode itself is 
> embedded in the containing directory's metadata object, sono backpointer 
> is needed there.  (The MDS' embedded inodes trade random access to inodes 
> for directory prefetching/locality.)

Oh, OK, I think what I didn't understand was the distinction between
"file data object" and "inode".  So the filehandle will be some kind of
pointer to the file data object, from which you'll then be able to look
up the directory and the inode that's stored in it.  And rename will
require modifying that file data object, not the filehandle, so
filehandles will remain stable over rename.

Thanks for the explanation!

> But this is all outside the scope of the client, so you wouldn't find it 
> in the patch.  The result is just that there is no reliable lookup-by-ino 
> operation, but it's only needed for NFS reexport.  (Hard links are 
> resolved by the server using a slightly different mechanism.)

Got it.

--b.