On Mon, 2014-04-28 at 07:09 +0200, Mike Galbraith wrote:
> Hi Nicholas,
>
> On Sat, 2014-04-26 at 15:58 +0200, Mike Galbraith wrote:
> > On Sat, 2014-04-26 at 10:38 +0200, Mike Galbraith wrote:
> > > On Fri, 2014-04-25 at 09:40 +0200, Mike Galbraith wrote:
> > >
> > > > Hotplug can still deadlock in rt trees too, and will if you beat it
> > > > hard.
> > >
> > > Box actually deadlocks like so.
> >
> > ...
> >
> > 3.12-rt looks a bit busted migrate_disable/enable() wise.
> >
> > /me eyeballs 3.10-rt (looks better), confirms 3.10-rt hotplug works,
> > swipes working code, confirms 3.12-rt now works. Yup, that was it.
>
> My boxen, including 64 core DL980 that ran hotplug stress for 3 hours
> yesterday with pre-pushdown rwlocks, say the migrate_disable/enable
> pushdown patches are very definitely busted.
migrate_disable-pushd-down-in-atomic_dec_and_spin_lo.patch
bug: migrate_disable() after blocking is too late.
@@ -1028,12 +1028,12 @@ int atomic_dec_and_spin_lock(atomic_t *a
/* Subtract 1 from counter unless that drops it to 0 (ie. it was 1) */
if (atomic_add_unless(atomic, -1, 1))
return 0;
- migrate_disable();
rt_spin_lock(lock);
- if (atomic_dec_and_test(atomic))
+ if (atomic_dec_and_test(atomic)){
+ migrate_disable();
return 1;
+ }
rt_spin_unlock(lock);
- migrate_enable();
return 0;
}
EXPORT_SYMBOL(atomic_dec_and_spin_lock);
read_lock-migrate_disable-pushdown-to-rt_read_lock.patch
bug: ditto.
@@ -244,8 +246,10 @@ void __lockfunc rt_read_lock(rwlock_t *r
/*
* recursive read locks succeed when current owns the lock
*/
- if (rt_mutex_owner(lock) != current)
+ if (rt_mutex_owner(lock) != current) {
__rt_spin_lock(lock);
+ migrate_disable();
+ }
rwlock->read_depth++;
}
Moving that migrate_disable() up will likely fix my hotplug troubles.
I'll find out when I get back from physical torture (therapy) session.
-Mike
--
To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
