Re: [RFC] readdir mess

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

On Sun, Aug 24, 2008 at 12:03:26PM +0100, Al Viro wrote:
> On Sun, Aug 24, 2008 at 11:10:14AM +0100, Al Viro wrote:
> 
> > I agree that such transition plan makes sense, but that'll take more
> > preliminary work than in your patch; there are other vfs_readdir() and
> > ->readdir() callers, not just the obvious syscall ones.
> 
> BTW, nfsd4_list_rec_dir() is FUBAR.  Its users, actually - they try to use it
> for lovely things like kernel-side rm -rf /var/lib/nfs/v4recovery/* and
> screw up in rather amusing ways..  I'm not even talking about the effects
> of OOM (dentry leak); if you rename something away from that directory,
> you'll get vfs_rmdir(dir, dentry) with dentry not being a child of dir,
> which means deadlock if you are lucky and underlying fs corruption if you
> are not...
> 
> I really wonder WTF is that doing in the kernel, anyway.  Looks like an
> obvious candidate for userland helper...

Yes.  Christoph complained about this before (err, 3 years ago!), though
without (as far as I can tell) catching those particular bugs:

	http://marc.info/?l=linux-fsdevel&m=112703894118581&w=2

We eventually agreed that it was a problem and worked on a userspace
replacement, but it never got to the point where I was happy enough with
it to commit to the new user interface, and the effort died.  I'll take
a look at what we last had.

--b.