[BUG] "sched: Remove rq->lock from the first half of ttwu()" locks up on ARM

[Marc.Zyngier@arm.com (Marc Zyngier)]

Peter,

I've experienced all kind of lock-ups on ARM SMP platforms recently, and
finally tracked it down to the following patch:

e4a52bcb9a18142d79e231b6733cabdbf2e67c1f [sched: Remove rq->lock from the first half of ttwu()].

Even on moderate load, the machine locks up, often silently, and
sometimes with a few messages like:
INFO: rcu_preempt_state detected stalls on CPUs/tasks: { 0} (detected by 1, t=12002 jiffies)

Another side effect of this patch is that the load average is always 0,
whatever load I throw at the system.

Reverting the sched changes up to that patch (included) gives me a
working system again, which happily survives parallel kernel
compilations without complaining.

My knowledge of the scheduler being rather limited, I haven't been able
to pinpoint the exact problem (though it probably have something to do
with __ARCH_WANT_INTERRUPTS_ON_CTXSW being defined on ARM). The enclosed
patch somehow papers over the load average problem, but the system ends
up locking up anyway:

diff --git a/kernel/sched.c b/kernel/sched.c
index d3ade54..5ab43c4 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -2526,8 +2526,13 @@ try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
 		 * to spin on ->on_cpu if p is current, since that would
 		 * deadlock.
 		 */
-		if (p == current)
+		if (p == current) {
+			p->sched_contributes_to_load = !!task_contributes_to_load(p);
+			p->state = TASK_WAKING;
+			if (p->sched_class->task_waking)
+				p->sched_class->task_waking(p);
 			goto out_activate;
+		}
 #endif
 		cpu_relax();
 	}

I'd be happy to test any patch you may have.

Cheers,

	M.
-- 
Reality is an implementation detail.