It seems that nfsd can return reply attributes with a link count of zero but without an NFS3ERR_STALE status. We've seen this actually happen for a write request to a file with a single link that is concurrently being removed without NLM lock protection. What is the proper behavior here?
On Thu, 2 Aug 2012 19:29:14 -0700 (PDT)
Andrei Warkentin <[email protected]> wrote:
> ----- Original Message -----
> > From: "Ed Goggin" <[email protected]>
> > To: [email protected]
> > Sent: Thursday, August 2, 2012 9:57:58 PM
> > Subject: stale or not stale
> >
> >
> > It seems that nfsd can return reply attributes with a link count of
> > zero but without an NFS3ERR_STALE status. We've seen this actually
> > happen for a write request to a file with a single link that is
> > concurrently being removed without NLM lock protection. What is the
> > proper behavior here?
>
> I think it would be worthwhile to add here that the the remove was not concurrent with the write, and at the time of the NFS write a new file with the same name existed, yet fh decoding picked up the old inode instead of reporting -ESTALE. FS was ext4.
>
> In fact seen two things happen - n_link = 0 and n_link = 1, and in both cases we knew the file was unlink()ed and re-creat()ed.
>
That's entirely valid behavior.
ESTALE from the server is its way of saying: "I don't recognize that
filehandle". It can't figure out how to match that filehandle to an
inode.
An i_nlink count of 0 means that there are no more hardlinks attached
to the inode. It's deleted from the namespace, but the inode still
exists until there are no more i_count references held on it.
--
Jeff Layton <[email protected]>
----- Original Message -----
> From: "Ed Goggin" <[email protected]>
> To: [email protected]
> Sent: Thursday, August 2, 2012 9:57:58 PM
> Subject: stale or not stale
>
>
> It seems that nfsd can return reply attributes with a link count of
> zero but without an NFS3ERR_STALE status. We've seen this actually
> happen for a write request to a file with a single link that is
> concurrently being removed without NLM lock protection. What is the
> proper behavior here?
I think it would be worthwhile to add here that the the remove was not concurrent with the write, and at the time of the NFS write a new file with the same name existed, yet fh decoding picked up the old inode instead of reporting -ESTALE. FS was ext4.
In fact seen two things happen - n_link = 0 and n_link = 1, and in both cases we knew the file was unlink()ed and re-creat()ed.
A