2.6.xx: NFS: directory motion/cam2 contains a readdir loop

[Christoph Hellwig]

On Wed, Jul 27, 2011 at 04:02:40PM -0400, Christoph Hellwig wrote:
> But looking closer at it it only prints the directory name and not that
> of any of the matching cookies, making it pretty useless to debug any
> problem.  (and it makes my previous question to Justin look stupid..).
> 
> 
> But so far I still stick to my previous theory that this sounds like
> a directory offset getting reused.  How is cache invalidation for
> the array supposed to work?  And maybe more importantly, given that he
> can only reproduce it with a .38 client did any bugs get fixed in that
> code recently that might lead to issues with the cache invalidation?

Actually we won't even need cache invalidation bugs, see
nfsd_buffered_readdir() - we might do multiple vfs_readdir calls to
fill a single nfs reply, and between these two directory contents might
have been completely replaced, in the worst (pathological case) you
might get a second readdir having exactly the same offsets, but pointing
to completely different inodes.