On Tue, Oct 04, 2022 at 06:57:59PM -0400, Joel Fernandes wrote:
> >> needed after an entrain. Otherwise, the RCU barrier callback can wait in
> >> the queue for several seconds before the lazy callbacks in front of it
> >> are serviced.
> >
> > It's not about lazy callbacks here (but you can mention the fact that
> > waking nocb_gp if necessary after flushing bypass is a beneficial side
> > effect for further lazy implementation).
> >
> > So here is the possible bad scenario:
> >
> > 1) CPU 0 is nocb, it queues a callback
> > 2) CPU 0 goes idle (or userspace with nohz_full) forever
> > 3) The grace period related to that callback elapses
> > 4) The callback is moved to the done list (but is not invoked yet), there are no more pending for CPU 0
> > 5) CPU 1 calls rcu_barrier() and entrains to CPU 0 cblist
>
> CPU 1 can only entrain into CPU 0 if the CPU is offline:
>
> if (!rcu_rdp_cpu_online(rdp)) {
> rcu_barrier_entrain(rdp);
> WARN_ON_ONCE(READ_ONCE(rdp->barrier_seq_snap) != gseq);
> raw_spin_unlock_irqrestore(&rcu_state.barrier_lock,
> ...
> continue;
> }
Ah good point. So CPU 1 sends an IPI to CPU 0 which entrains itself.
And then looks like the result is the same.
>
> Otherwise an IPI does the entraining. So I do not see how CPU 0 being idle
> causes the cross-CPU entraining.
It doesn't but it shows that the CPU isn't going to enqueue any further
callback before a while. Though even if it did, it may not even solve the
situation, not until an RCU_NOCB_WAKE_FORCE is issued...
>
> > 6) CPU 1 waits forever
>
> But, I agree it can still wait forever, once the IPI handler does the
> entraining, since nothing will do the GP thread wakeup.
>
> >>
> >> Reported-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> >
> > Fixes: 5d6742b37727 ("rcu/nocb: Use rcu_segcblist for no-CBs CPUs")
>
> So, do you mind writing a proper patch with a proper commit message and Fixes
> tag then? It can independent of this series and add my Reported-by tag,
> thanks!
Ok will do.
Thanks!
>
> Thanks!
>
> - Joel
>
