Re: [PATCH v8 01/13] rcu: Fix missing nocb gp wake on rcu_barrier()

[Frederic Weisbecker <frederic@kernel.org>]

On Fri, Oct 14, 2022 at 07:21:27AM -0700, Paul E. McKenney wrote:
> On Tue, Oct 11, 2022 at 06:01:30PM +0000, Joel Fernandes (Google) wrote:
> > From: Frederic Weisbecker <frederic@kernel.org>
> > 
> > Upon entraining a callback to a NOCB CPU, no further wake up is
> > issued on the corresponding nocb_gp kthread. As a result, the callback
> > and all the subsequent ones on that CPU may be ignored, at least until
> > an RCU_NOCB_WAKE_FORCE timer is ever armed or another NOCB CPU belonging
> > to the same group enqueues a callback on an empty queue.
> > 
> > Here is a possible bad scenario:
> > 
> > 1) CPU 0 is NOCB unlike all other CPUs.
> > 2) CPU 0 queues a callback
> 
> Call it CB1.
> 
> > 2) The grace period related to that callback elapses
> > 3) The callback is moved to the done list (but is not invoked yet),
> >    there are no more pending callbacks for CPU 0
> 
> So CB1 is on ->cblist waiting to be invoked, correct?
> 
> > 4) CPU 1 calls rcu_barrier() and sends an IPI to CPU 0
> > 5) CPU 0 entrains the callback but doesn't wake up nocb_gp
> 
> And CB1 must still be there because otherwise the IPI handler would not
> have entrained the callback, correct?  If so, we have both CB1 and the
> rcu_barrier() callback (call it CB2) in ->cblist, but on the done list.
> 
> > 6) CPU 1 blocks forever, unless CPU 0 ever queues enough further
> >    callbacks to arm an RCU_NOCB_WAKE_FORCE timer.
> 
> Except that -something- must have already been prepared to wake up in
> order to invoke CB1.  And that something would invoke CB2 along with CB1,
> given that they are both on the done list.  If there is no such wakeup
> already, then the hang could occur with just CB1, without the help of CB2.

Heh good point. I was confused with CB1 on RCU_DONE_TAIL and the possibility
for CB2 to be entrained on RCU_WAIT_TAIL. But that's indeed not supposed to
happen. Ok so this patch indeed doesn't make sense outside lazy.

> > This is also required to make sure lazy callbacks in future patches
> > don't end up making rcu_barrier() wait for multiple seconds.
> 
> But I do see that the wakeup is needed in the lazy case, and if I remember
> correctly, the ten-second rcu_barrier() delay really did happen.  If I
> understand correctly, for this to happen, all of the callbacks must be
> in the bypass list, that is, ->cblist must be empty.
> 
> So has the scenario steps 1-6 called out above actually happened in the
> absence of lazy callbacks?

Nope, so I guess we can have the pending check around rcu_nocb_flush_bypass()
only...

Thanks!