Re: [PATCH RT 4/6] rt/locking: Reenable migration accross schedule
From: Thomas Gleixner
Date: Thu Mar 24 2016 - 06:46:00 EST
On Thu, 24 Mar 2016, Mike Galbraith wrote:
> On Sun, 2016-03-20 at 09:43 +0100, Mike Galbraith wrote:
> > On Sat, 2016-02-13 at 00:02 +0100, Sebastian Andrzej Siewior wrote:
> > > From: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> > >
> > > We currently disable migration across lock acquisition. That includes the part
> > > where we block on the lock and schedule out. We cannot disable migration after
> > > taking the lock as that would cause a possible lock inversion.
> > >
> > > But we can be smart and enable migration when we block and schedule out. That
> > > allows the scheduler to place the task freely at least if this is the first
> > > migrate disable level. For nested locking this does not help at all.
> >
> > I met a problem while testing shiny new hotplug machinery.
> >
> > rt/locking: Fix rt_spin_lock_slowlock() vs hotplug migrate_disable() bug
> >
> > migrate_disable() -> pin_current_cpu() -> hotplug_lock() leads to..
> > > BUG_ON(rt_mutex_real_waiter(task->pi_blocked_on));
> > ..so let's call migrate_disable() after we acquire the lock instead.
>
> Well crap, that wasn't very clever A little voice kept nagging me, and
> yesterday I realized what it was grumbling about, namely that doing
> migrate_disable() after lock acquisition will resurrect a hotplug
> deadlock that we fixed up a while back.
Glad you found out yourself. Telling you that was on my todo list ....
> On the bright side, with the busted migrate enable business reverted,
> plus one dinky change from me [1], master-rt.today has completed 100
> iterations of Steven's hotplug stress script along side endless
> futexstress, and is happily doing another 900 as I write this, so the
> next -rt should finally be hotplug deadlock free.
>
> Thomas's state machinery seems to work wonders. 'course this being
> hotplug, the other shoe will likely apply itself to my backside soon.
That's a given :)
I really wonder what makes the change. The only thing which comes to my mind
is the enforcement of running the online and down_prepare callbacks on the
plugged cpu instead of doing it wherever the scheduler decides to run it.
> 1. nest module_mutex inside hotplug_lock to prevent bloody systemd
> -udevd from blocking in migrate_disable() while holding kernfs_mutex
> during module load, putting a quick end to hotplug stress testing.
Did I miss a patch here or is that still in your pile?
Thanks,
tglx