Re: [PATCH] rcu/nocb: Fix misordered rcu_barrier() while (de-)offloading

[Frederic Weisbecker]

On Mon, Oct 18, 2021 at 09:18:14AM -0700, Paul E. McKenney wrote:
> On Mon, Oct 18, 2021 at 01:32:59PM +0200, Frederic Weisbecker wrote:
> > When an rdp is in the process of (de-)offloading, rcu_core() and the
> > nocb kthreads can process callbacks at the same time. This leaves many
> > possible scenarios leading to an rcu barrier to execute before
> > the preceding callbacks. Here is one such example:
> > 
> >             CPU 0                                  CPU 1
> >        --------------                         ---------------
> >      call_rcu(callbacks1)
> >      call_rcu(callbacks2)
> >      // move callbacks1 and callbacks2 on the done list
> >      rcu_advance_callbacks()
> >      call_rcu(callbacks3)
> >      rcu_barrier_func()
> >          rcu_segcblist_entrain(...)
> >                                             nocb_cb_wait()
> >                                                 rcu_do_batch()
> >                                                     callbacks1()
> >                                                     cond_resched_tasks_rcu_qs()
> >      // move callbacks3 and rcu_barrier_callback()
> >      // on the done list
> >      rcu_advance_callbacks()
> >      rcu_core()
> >          rcu_do_batch()
> >              callbacks3()
> >              rcu_barrier_callback()
> >                                                     //MISORDERING
> >                                                     callbacks2()
> > 
> > Fix this with preventing two concurrent rcu_do_batch() on a  same rdp
> > as long as an rcu barrier callback is pending somewhere.
> > 
> > Reported-by: Paul E. McKenney <paulmck@xxxxxxxxxx>
> > Signed-off-by: Frederic Weisbecker <frederic@xxxxxxxxxx>
> > Cc: Josh Triplett <josh@xxxxxxxxxxxxxxxx>
> > Cc: Joel Fernandes <joel@xxxxxxxxxxxxxxxxx>
> > Cc: Boqun Feng <boqun.feng@xxxxxxxxx>
> > Cc: Neeraj Upadhyay <neeraju@xxxxxxxxxxxxxx>
> > Cc: Uladzislau Rezki <urezki@xxxxxxxxx>
> 
> Yow!
> 
> But how does the (de-)offloading procedure's acquisition of
> rcu_state.barrier_mutex play into this?  In theory, that mutex was
> supposed to prevent these sorts of scenarios.  In practice, it sounds
> like the shortcomings in this theory should be fully explained so that
> we don't get similar bugs in the future.  ;-)

I think you're right. The real issue is something I wanted to
fix next: RCU_SEGCBLIST_RCU_CORE isn't cleared when nocb is enabled on
boot so rcu_core() always run concurrently with nocb kthreads in TREE04,
without holding rcu_barrier mutex of course (I mean with the latest patchset).

Ok forget this patch, I'm testing again with simply clearing
RCU_SEGCBLIST_RCU_CORE on boot.

Thanks.