Re: [PATCH] namespaces: fix exit race by splitting exit

[Herbert Poetzl <herbert@13thfloor.at>]

On Thu, Jan 25, 2007 at 10:30:56PM -0800, Andrew Morton wrote:
> On Thu, 25 Jan 2007 23:26:59 -0600
> "Serge E. Hallyn" <serue@us.ibm.com> wrote:
> 
> > Fix exit race by splitting the nsproxy putting into two pieces.
> > First piece reduces the nsproxy refcount.  If we dropped the last
> > reference, then it puts the mnt_ns, and returns the nsproxy as a
> > hint to the caller.  Else it returns NULL.  The second piece of
> > exiting task namespaces sets tsk->nsproxy to NULL, and drops the
> > references to other namespaces and frees the nsproxy only if an
> > nsproxy was passed in.
> > 
> > A little awkward and should probably be reworked, but hopefully
> > it fixes the NFS oops.
> 
> I'm a bit worried about jamming something like this into 2.6.20.
> Could the usual culprits please review this carefully with 
> some urgency?

okay, after integrating this into two Linux-VServer
branches and some testing, I can confirm that it 
_seems_ to fix the nfs and related issues, but still, 
I do not like it :)

here my issues with this approach:

 - the code is quite hard to read and can easily
   lead to unexpected issues when spaces are
   manipulated

 - it breaks the basic get/put refcounting for
   nsproxy references outside the task struct
   i.e. we had to add a vs_put_nsproxy() which
   does what the put_nsproxy() did before, to
   keep and handle a reference to the nsproxy 
   from the context structure

 - the following scenario might become a problem
   for future spaces (especially the pid space?)

            A                       B

 exit_task_namespaces_early()
			exit_task_namespaces_early()
			exit_notify()
			exit_task_namespaces()
 ---------------------------------------------------
 exit_notify()
 exit_task_namespaces()

note: I still consider it the best available fix
for this issues, especially as 2.6.20 is in a 
late rc stage ... but IMHO the nfs threads should
be modified to handle the nsproxy disposal properly

> And Daniel, if you can find time to runtime test it please?

he did, looks like it works fine with vanilla too
(even when stressing the described cornercase)

best,
Herbert