[PATCH v2 0/2] document ->sighand protection, rcu_read_unlock() deadlocks

[Oleg Nesterov <oleg@redhat.com>]

Paul, could you take these 2 doc patches? Assuming that you agree
with the comments, of course.

On 09/24, Paul E. McKenney wrote:
>
> On Tue, Sep 23, 2014 at 09:03:48PM +0200, Oleg Nesterov wrote:
> >
> > Paul, will you agree if we turn it into
> > ...
> >      /*
> >       * On the succesfull return we hold ->siglock. According to comment
> >       * above rcu_read_unlock() this is against the rules, but scheduler
> >       * locks are fine under this lock, signal_wake_up() takes them too.
> >       */
> >      rcu_read_unlock();
>
> If you guys continue the guarantee of no deadlock, I am OK with this change.

Heh. Contrary to what I said (and you all were agree ;), this deadlock
is actually possible, so we can not remove the deadlock-avoidance from
__lock_task_sighand(). And I do not see how we can cleanup this code
because preempt_disable() + spin_lock() is not -rt friendly.

I think this deserves a bit of documentation, see 2/2. Perhaps this is
just me, but imo the current comment is a bit misleading.

"if the caller of rcu_read_unlock() already holds one of these locks ..."
is not a problem in fact. I mean, pi_lock or rq->lock are special enough,
nobody should ever call the outermost rcu_read_unlock() with these locks
held. rt_mutex->wait_lock should be fine too, also because ->boost_mtx
is private to rcu_boost() and rcu_read_unlock_special().

But. They can race with each other, and that is why rcu_read_unlock()
under (say) ->siglock can actually lead to deadlock. And only because
rt_mutex->wait_lock doesn't disable irqs. Or I am totally confused.

Perhaps we can change rtmutex.c to use raw_spin_lock_irqsave(), or do
something else...

Oleg.

 include/linux/rcupdate.h |    4 +++-
 kernel/fork.c            |    5 ++++-
 kernel/signal.c          |   12 +++++++++++-
 3 files changed, 18 insertions(+), 3 deletions(-)