freeBSD nullfs together with nfs and "silly rename"

[Kostik Belousov]

On Sat, Jun 12, 2010 at 11:56:10AM +0300, Mikolaj Golub wrote:
> 
> On Sun, 6 Jun 2010 16:44:43 +0200 Leon Me??ner wrote:
> 
>  LM> Hi,
>  LM> I hope this is not the wrong list to ask. Didn't get any answers on
>  LM> -questions.
> 
>  LM> When you try to do the following inside a nullfs mounted directory,
>  LM> where the nullfs origin is itself mounted via nfs you get an error:
> 
>  LM> # foo 
>  LM> # tail -f foo& 
>  LM> # rm -f foo 
>  LM> tail: foo: Stale NFS file handle
>  LM> # fg
> 
>  LM> This is really a problem when running services inside jails and using
>  LM> NFS as storage. As of [2] it looks like this problem is known for a
>  LM> while. On a normal NFS mount this does not happen as "silly renaming"
>  LM> [1] works there (producing nasty little .nfsXXXX files).
> 
> nfs_sillyrename() is called when vnode's usecount is more then 1. It is
> expected that unlink() syscall increases vnode's usecount in namei() and if
> the file has been already opened usecount will be more then 1.
> 
> But with nullfs layer present the reference counts are held by the upper node,
> not the lower (nfs) one, so when unlink() is called it increases usecount of
> the upper vnode, not nfs vnode and nfs_sillyrename() is never called.
> 
> The strightforward solution looks like to implement null_remove() that will
> increase lower vnode's refcount before calling null_bypass() and then
> decrement it after the call. See the attached patch (it works for me on both
> 8-STABLE and CURRENT).

The upper vnode holds a reference to the lower vnode, as you noted.
Now, with your patch, I believe that _all_ calls to the nfs_remove()
are happen with refcount > 1.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 196 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-stable/attachments/20100612/be8ba75c/attachment.pgp