Re: 2.6.35 - INFO: kernel/signal.c:660 invoked rcu_dereference_check() without protection!

Re: 2.6.35 - INFO: kernel/signal.c:660 invoked rcu_dereference_check() without protection!

[Paul E. McKenney]

On Tue, Aug 03, 2010 at 11:18:07PM -0400, Miles Lane wrote:
> [ INFO: suspicious rcu_dereference_check() usage. ]
> ---------------------------------------------------
> kernel/signal.c:660 invoked rcu_dereference_check() without protection!
> 
> other info that might help us debug this:
> 
> rcu_scheduler_active = 1, debug_locks = 1
> 1 lock held by udevd/670:
>  #0:  (tasklist_lock){.+.+..}, at: [<ffffffff8104628c>] do_exit+0x252/0x671
>  stack backtrace:
>  Pid: 670, comm: udevd Not tainted 2.6.35 #15
>  Call Trace:
>  [<ffffffff8106759c>] lockdep_rcu_dereference+0x9d/0xa6
>  [<ffffffff8104f297>] check_kill_permission+0x9d/0x132
>  [<ffffffff8105010e>] group_send_sig_info+0x1a/0x3d
>  [<ffffffff81046388>] do_exit+0x34e/0x671
>  [<ffffffff81046723>] do_group_exit+0x78/0xa3
>  [<ffffffff81046760>] sys_exit_group+0x12/0x16
>  [<ffffffff81009c82>] system_call_fastpath+0x16/0x1b

Thank you again for your testing efforts!

This one is a bit odd.  exit_signals() holds either the tasklist_lock
or the ->sighand->siglock over its signal manipulations, so this is
not the code path triggering the above.  I might be missing something,
but I don't see either hrtimer_cancel() or exit_itimers() attempting
to send signals.

I am copying Oleg, as he might see something that I am missing.

							Thanx, Paul

Re: 2.6.35 - INFO: kernel/signal.c:660 invoked rcu_dereference_check() without protection!

[Oleg Nesterov]

On 08/09, Paul E. McKenney wrote:
>
> On Tue, Aug 03, 2010 at 11:18:07PM -0400, Miles Lane wrote:
> > [ INFO: suspicious rcu_dereference_check() usage. ]
> > ---------------------------------------------------
> > kernel/signal.c:660 invoked rcu_dereference_check() without protection!
> >
> > other info that might help us debug this:
> >
> > rcu_scheduler_active = 1, debug_locks = 1
> > 1 lock held by udevd/670:
> >  #0:  (tasklist_lock){.+.+..}, at: [<ffffffff8104628c>] do_exit+0x252/0x671
> >  stack backtrace:
> >  Pid: 670, comm: udevd Not tainted 2.6.35 #15
> >  Call Trace:
> >  [<ffffffff8106759c>] lockdep_rcu_dereference+0x9d/0xa6
> >  [<ffffffff8104f297>] check_kill_permission+0x9d/0x132
> >  [<ffffffff8105010e>] group_send_sig_info+0x1a/0x3d
> >  [<ffffffff81046388>] do_exit+0x34e/0x671
> >  [<ffffffff81046723>] do_group_exit+0x78/0xa3
> >  [<ffffffff81046760>] sys_exit_group+0x12/0x16
> >  [<ffffffff81009c82>] system_call_fastpath+0x16/0x1b
>
> Thank you again for your testing efforts!
>
> This one is a bit odd.  exit_signals() holds either the tasklist_lock
> or the ->sighand->siglock over its signal manipulations, so this is
> not the code path triggering the above.  I might be missing something,
> but I don't see either hrtimer_cancel() or exit_itimers() attempting
> to send signals.

Agreed, this is strange and shouldn't happen.

Perhaps, this kernel lack's David's 694f690d27dadccc8cb9d90532e76593b61fe098
"CRED: Fix RCU warning due to previous patch fixing __task_cred()'s checks"
patch?

Oleg.