Re: mmotm 2010-04-05 - another RCU whinge (not network this time)

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Fri, Apr 09, 2010 at 11:22:32PM -0400, Valdis.Kletnieks@vt.edu wrote:
> On Fri, 09 Apr 2010 16:16:14 PDT, "Paul E. McKenney" said:
> > On Thu, Apr 08, 2010 at 07:57:28PM -0400, Valdis.Kletnieks@vt.edu wrote:
> > > On Mon, 05 Apr 2010 16:09:45 PDT, akpm@linux-foundation.org said:
> > > > The mm-of-the-moment snapshot 2010-04-05-16-09 has been uploaded to
> > > > 
> > > >    http://userweb.kernel.org/~akpm/mmotm/
> > > 
> > > Hit another one.  I seem to be on a roll...
> > > 
> > > Seen in dmesg, happened near end of the initrd..
> > > 
> > > [   26.756864] 
> > > [   26.756866] ===================================================
> > > [   26.756869] [ INFO: suspicious rcu_dereference_check() usage. ]
> > > [   26.756871] ---------------------------------------------------
> > > [   26.756874] fs/proc/array.c:241 invoked rcu_dereference_check() without protection!
> > 
> > Color me confused.  I cloned James Toy's git repository at
> > git://zen-kernel.org/kernel/mmotm.git, and gitk claims that I am on tag
> > 2010-04-05-16-09, which matches the string above.  But when I look at
> > fs/proc/array.c near line 241, I see:
> 
> Andrew's -mm tree has 3 patches from Oleg Nesterov that hit that file, so the
> code is different from what you show. Color *me* confused why your clone of
> mmotm.git doesn't seem to contain them - I'm not sure how James Toy builds
> that git tree.  Perhaps the tag is applied before those patches are - the
> 'mm.patch' that updates the Makefile with the version is usually in the
> *middle* of the 'series' file.  What does HEAD of that tree look like?

Good point...  The last commit is branch "master" and tagged
2010-04-05-16-09, but the commit line is "Linux 2.6.34-rc3", which seems
unlikely to me.

> My tree has:
> 
> /* needs ->siglock or rcu_read_lock() */
> static void collect_sigign_sigcatch(struct task_struct *p, sigset_t *ign,
>                                     sigset_t *catch)
> {
>         struct sighand_struct *sighand = rcu_dereference(p->sighand);
> 
> And that rcu_dereference() does it.

Thank you!!!

> Oleg, looks like proc-make-collect_sigign_sigcatch-rcu-safe.patch is the
> offender here, it added the line that causes the whinge.

If collect_sigign_sigcatch() is OK to call by updaters as well as
readers, we need something like:

	struct sighand_struct *sighand;
	
	sighand = rcu_dereference_check(p->sighand,
					rcu_read_lock_held() ||
					lockdep_is_held(&???));

Where the "???" is replaced with whichever of the two locks is protecting
updates.  My guess would be the sighand lock, but I would not rely on
my guesses in this case.  ;-)

							Thanx, Paul