Re: Q: NFSD readdir in linux-2.6.28

[Matthew Wilcox <matthew@wil.cx>]

On Thu, Mar 19, 2009 at 03:17:17PM +0000, David Woodhouse wrote:
> > In 2.6.27, when nfs3svc_encode_entry_plus() calls lookup_one_len(), the
> > i_mutex lock was acquired by vfs_readdir() and it was not a problem.
> > 
> > After the commit (above), nfsd_readdir/nfsd_buffered_readdir/vfs_readdir
> > calls nfsd_buffered_filldir(), and nfs3svc_encode_entry_plus() is called
> > later.
> > In this sequence, lookup_one_len() is called without i_mutex held.
> > 
> > Isn't it a problem?
> 
> Yes, well spotted. It didn't matter when the buffered readdir() was
> purely internal to XFS, because it didn't matter there that we called
> ->lookup() without i_mutex set. But now we're exposing arbitrary file
> systems to it, we need to make sure we follow the locking rules. 
> 
> I _think_ it's sufficient to make the affected callers of
> lookup_one_len() lock the parent's i_mutex for themselves before calling
> it. I'll take a closer look...

Should we also add this?

---

Ensure inode is locked in lookup_one_len()

Signed-off-by: Matthew Wilcox <willy@linux.intel.com>

diff --git a/fs/namei.c b/fs/namei.c
index bbc15c2..476b1d0 100644
--- a/fs/namei.c
+++ b/fs/namei.c
@@ -1244,6 +1244,7 @@ struct dentry *lookup_one_len(const char *name, struct dentry *base, int len)
 	int err;
 	struct qstr this;
 
+	BUG_ON(!mutex_is_locked(&base->d_inode->i_mutex));
 	err = __lookup_one_len(name, &this, base, len);
 	if (err)
 		return ERR_PTR(err);

-- 
Matthew Wilcox				Intel Open Source Technology Centre
"Bill, look, we understand that you're interested in selling us this
operating system, but compare it to ours.  We can't possibly take such
a retrograde step."