Re: problems with reiserfs + nfs using 2.4.2-pre4

[Neil Brown]

On Tuesday February 20, [e-mail] wrote:
> Hi,
> 
> On Tue, 20 Feb 2001, Neil Brown wrote:
> 
> > 2/ lookup("..").
> 
> A small question:
> Why exactly is this needed?
> 
> bye, Roman

Having read the subsequent posts, I now see what you are thinking and
know how to answer this.

The problem that I see is indeed related to rename, but not in the way
that Trond mentions.  (Tronds issue is a real, though very different,
issue). 

Suppose I have a directory path

   /a/b/c/d/e/f/g

Where filehandle X refers to /a/b and filehande Y refers to
/a/b/c/d/e/f/g

Then an NFS request comes in saying

   RENAME X/c to Y/h

If this was performed, then you would end up with a directory path
 
 /a/b

and a disconnected loop:

    d/e/f/g/h
    ^       v
    +-------+

which is not good.
This possibility is protected against by the VFS layer.  It serialises
all directory renames using s_vfs_rename_sem and then checks that the
source of the rename isn't an ancestor of the target.
For this check to be reliable, each dentry for a directory must be
properly connected to the root.

Now if you have a filesystem that:

 - cannot do ".." lookups efficiently, or doesn't want to and
 - can protect against this sort of loop (and any other issues that
    the VFS usually protects against) itself

then it can (with my patch) simply define decode_fh and encode_fh and
do whatever it likes, such as create a dentry that isn't properly
connected.

get_parent is only called by nfsd_find_fh_dentry which is a helper
routine which is normally called by the decode_fh function (and is
called by the default decode_fh function).
If a filesystem provides it's own decode_fh which doesn't call
nfsd_find_fh_dentry, then get_parent won't be called and isn't needed.

NeilBrown