Re: rcu_sched stalls in idle task introduced in pre-4.8?

[Rich Felker]

On Wed, Aug 03, 2016 at 09:16:31AM -0700, Paul E. McKenney wrote:
> On Tue, Aug 02, 2016 at 01:45:04PM -0700, Paul E. McKenney wrote:
> > On Tue, Aug 02, 2016 at 04:32:17PM -0400, Rich Felker wrote:
> > > On Tue, Aug 02, 2016 at 12:48:02PM -0700, Paul E. McKenney wrote:
> 
> [ . . . ]
> 
> > > > Does the problem reproduces easily?
> > > 
> > > Yes, it happens right after boot and repeats every 30-90 seconds or
> > > so.
> > 
> > Well, that at least makes it easier to test any patches!
> > 
> > > > A bisection might be very helpful.
> > > 
> > > Bisection would require some manual work to setup because the whole
> > > reason I was rebasing on Linus's tree was to adapt the drivers to
> > > upstream infrastructure changes (the new cpuhp stuff replacing
> > > notifier for cpu starting). The unfortunate way it was done, each
> > > driver adds an enum to linux/cpuhotplug.h so all the patches have
> > > gratuitous conflicts. In addition, for older revisions in Linus's
> > > tree, there's at least one show-stopping (hang during boot) bug that
> > > needs a cherry-pick to fix. There may be other small issues too. I
> > > don't think they're at all insurmountible but it requires an annoying
> > > amount of scripting.
> > 
> > I had to ask!  Might eventually be necessary, but let's see what we
> > can learn from what you currently have.
> 
> And at first glance, my overnight run looks uglier than I would expect.
> I am now running tests at v4.7, and will run other tests to see if
> there really is a statistically significant degradation.  If there is,
> then I might be able to bisect, though with nine-hour runs this could
> take quite some time.

Any more thoughts on this? I'm testing v4.8-rc5 (plus jcore drivers
not yet upstream) and it's still happening.

Rich