Re: [PATCH tip/core/rcu 1/2] rcu: Parallelize and economize NOCB kthread wakeups

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Sat, Aug 23, 2014 at 03:43:38AM -0400, Pranith Kumar wrote:
> On Fri, Aug 22, 2014 at 5:53 PM, Paul E. McKenney
> <paulmck@linux.vnet.ibm.com> wrote:
> >
> > Hmmm...  Please try replacing the synchronize_rcu() in
> > __sysrq_swap_key_ops() with (say) schedule_timeout_interruptible(HZ / 10).
> > I bet that gets rid of the hang.  (And also introduces a low-probability
> > bug, but should be OK for testing.)
> >
> > The other thing to try is to revert your patch that turned my event
> > traces into printk()s, then put an ftrace_dump(DUMP_ALL); just after
> > the synchronize_rcu() -- that might make it so that the ftrace data
> > actually gets dumped out.
> >
> 
> I was able to reproduce this error on my Ubuntu 14.04 machine. I think
> I found the root cause of the problem after several kvm runs.
> 
> The problem is that earlier we were waiting on nocb_head and now we
> are waiting on nocb_leader_wake.
> 
> So there are a lot of nocb callbacks which are enqueued before the
> nocb thread is spawned. This sets up nocb_head to be non-null, because
> of which the nocb kthread used to wake up immediately after sleeping.
> 
> Now that we have switched to nocb_leader_wake, this is not being set
> when there are pending callbacks, unless the callbacks overflow the
> qhimark. The pending callbacks were around 7000 when the boot hangs.
> 
> So setting the qhimark using the boot parameter rcutree.qhimark=5000
> is one way to allow us to boot past the point by forcefully waking up
> the nocb kthread. I am not sure this is fool-proof.

Unfortunately, not in all cases.  A small kernel for embedded use might
register only a few callbacks during boot, which could still result
in a hang.

> Another option to start the nocb kthreads with nocb_leader_wake set,
> so that it can handle any pending callbacks. The following patch also
> allows us to boot properly.

This seems like a much better approach.

> Phew! Let me know if this makes any sense :)

It might well!  Another possibility is that the early_initcall function
doing the synchronize_rcu() is happening before the early_initcall
creating the RCU grace-period kthreads.

Seems like we need to close both holes.  Let's see how your patch works
for Amit, and I am testing a patch for the possible early_initcall
ordering issue.

> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> index 00dc411..4c397aa 100644
> --- a/kernel/rcu/tree_plugin.h
> +++ b/kernel/rcu/tree_plugin.h
> @@ -2386,6 +2386,9 @@ static int rcu_nocb_kthread(void *arg)
>         struct rcu_head **tail;
>         struct rcu_data *rdp = arg;
> 
> +       if (rdp->nocb_leader == rdp)
> +               rdp->nocb_leader_wake = true;
> +

Not that it matters all that much, but given that the followers don't
ever reference ->nocb_leader_wake, we should be able to set this flag
unconditionally.

						Thanx, Paul

>         /* Each pass through this loop invokes one batch of callbacks */
>         for (;;) {
>                 /* Wait for callbacks. */
> 
> 
> -- 
> Pranith
>