On 08/07/2015 08:42 AM, Daniel Wagner wrote:

On 08/05/2015 03:30 PM, Daniel Wagner wrote:

quoted

My test system didn't crash or showed any obvious defects, so I
decided to apply some benchmarks utilizing mmtests. I have picked some

As it turns out, this is not really true. I forgot to enable lockdep:

[...]

If I decoded this correctly the call to rcu_future_gp_cleanup() is
supposed to run with IRQs disabled. swake_up_all() though will reenable the
IRQs:

rcu_gp_cleanup() 
	rcu_for_each_node_breadth_first(rsp, rnp) {
		raw_spin_lock_irq(&rnp->lock);

		nocb += rcu_future_gp_cleanup(rsp, rnp);
		raw_spin_unlock_irq(&rnp->lock);
	}

rcu_future_gp_cleanup()
	rcu_nocb_gp_cleanup()
		swake_up_all()


With IRQs enabled again and we end up in rcu_process_callbacks
under SOFTIRQ. rcu_process_callbacks aquires the RCU lock again.

Not sure what to do here.

Not really knowing if this is okay but I think the call to
rcu_nocb_gp_cleanup() inside rcu_future_gp_cleanup() doesn't need to be
protected by rnp->lock. At least lockdep and rcutorture is still happy.

diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index d424378..9411fc3 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c

@@ -1569,7 +1569,6 @@ static int rcu_future_gp_cleanup(struct rcu_state

*rsp, struct rcu_node *rnp)
 	int needmore;
 	struct rcu_data *rdp = this_cpu_ptr(rsp->rda);

-	rcu_nocb_gp_cleanup(rsp, rnp);
 	rnp->need_future_gp[c & 0x1] = 0;
 	needmore = rnp->need_future_gp[(c + 1) & 0x1];
 	trace_rcu_future_gp(rnp, rdp, c,

@@ -1992,6 +1991,7 @@ static void rcu_gp_cleanup(struct rcu_state *rsp)
 		/* smp_mb() provided by prior unlock-lock pair. */
 		nocb += rcu_future_gp_cleanup(rsp, rnp);
 		raw_spin_unlock_irq(&rnp->lock);
+		rcu_nocb_gp_cleanup(rsp, rnp);
 		cond_resched_rcu_qs();
 		WRITE_ONCE(rsp->gp_activity, jiffies);
 		rcu_gp_slow(rsp, gp_cleanup_delay);