Re: [PATCH] ns: move free_nsproxy() out of do_exit() path

["Kirill A. Shutemov" <kirill@shutemov.name>]

On Fri, Jul 13, 2012 at 02:08:06PM -0700, Andrew Morton wrote:
> On Fri, 13 Jul 2012 14:48:08 +0300
> "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> wrote:
> 
> > From: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
> > 
> > free_nsproxy() is too heavy to be on exit path. Let's free namespaces
> > asynchronously to not block exit_group() syscall.
> 
> Please be specific, and complete.
> 
> Why is it "too heavy"?  Where is the time being spent?  Is it spent in
> D state or is it spent burning CPU cycles?  Does the patch simply
> offload the work into kernel threads, providing no net gain?

Unpatched switch_task_namespaces() takes 0.010 - 0.011 seconds on my
machine. About 0.008 of the time is synchronize_rcu().

So it's mostly waiting with wait_for_completion() in wait_rcu_gp().
It means D state.

> > The patch also fixes bug with free namespace without synchronize_rcu() through
> > put_nsproxy().
> 
> I just don't understand this description.

IIUC current locking model requires synchronize_rcu() before
free_nsproxy(). put_nsproxy() calls free_nsproxy() without
synchronize_rcu(). So it's racy.

I guess it was missed during switch to RCU (see cf7b708).
Pavel, am I right?

> Please send a new one which
> includes all details about the bug, including a description of
> the user-visible effects of the bug.

Okay, I will.

-- 
 Kirill A. Shutemov