[PATCH] rcutorture: add random preemption

[PATCH] rcutorture: add random preemption

[Lai Jiangshan]

Add random preemption to help we to torture the preemptable rcu.

srcu_read_delay() also calls rcu_read_delay() for shorter delays.

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
---
diff --git a/kernel/rcutorture.c b/kernel/rcutorture.c
index 2e2726d..7c81d07 100644
--- a/kernel/rcutorture.c
+++ b/kernel/rcutorture.c
@@ -303,6 +303,10 @@ static void rcu_read_delay(struct rcu_random_state *rrsp)
 		mdelay(longdelay_ms);
 	if (!(rcu_random(rrsp) % (nrealreaders * 2 * shortdelay_us)))
 		udelay(shortdelay_us);
+#ifdef CONFIG_PREEMPT
+	if (!preempt_count() && !(rcu_random(rrsp) % (nrealreaders * 20000)))
+		preempt_schedule();
+#endif
 }
 
 static void rcu_torture_read_unlock(int idx) __releases(RCU)
@@ -536,6 +540,8 @@ static void srcu_read_delay(struct rcu_random_state *rrsp)
 	delay = rcu_random(rrsp) % (nrealreaders * 2 * longdelay * uspertick);
 	if (!delay)
 		schedule_timeout_interruptible(longdelay);
+	else
+		rcu_read_delay(rrsp);
 }
 
 static void srcu_torture_read_unlock(int idx) __releases(&srcu_ctl)

Re: [PATCH] rcutorture: add random preemption

[Paul E. McKenney]

On Mon, Jun 21, 2010 at 04:57:42PM +0800, Lai Jiangshan wrote:
> Add random preemption to help we to torture the preemptable rcu.
> 
> srcu_read_delay() also calls rcu_read_delay() for shorter delays.

I do like the change to srcu_read_delay(), good to fall back to the
normal rcu_read_delay() behavior when a long delay is not selected.
The change to rcu_read_delay() looks promising as well, but please see
below for some comments on the other change.

And the big question: did you find any failures when testing with this
change?  ;-)

							Thanx, Paul

> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
> ---
> diff --git a/kernel/rcutorture.c b/kernel/rcutorture.c
> index 2e2726d..7c81d07 100644
> --- a/kernel/rcutorture.c
> +++ b/kernel/rcutorture.c
> @@ -303,6 +303,10 @@ static void rcu_read_delay(struct rcu_random_state *rrsp)
>  		mdelay(longdelay_ms);
>  	if (!(rcu_random(rrsp) % (nrealreaders * 2 * shortdelay_us)))
>  		udelay(shortdelay_us);
> +#ifdef CONFIG_PREEMPT
> +	if (!preempt_count() && !(rcu_random(rrsp) % (nrealreaders * 20000)))
> +		preempt_schedule();
> +#endif

This one scared me for a bit -- then I realized that preempt_schedule()
won't actually schedule if preemption is in any way disabled.  So the
above really is OK, because Classic RCU and RCU-bh disable preemption.

So, should we have a comment to this effect, or is my hypersensitivity to
RCU semantics unique to me?

>  }
> 
>  static void rcu_torture_read_unlock(int idx) __releases(RCU)
> @@ -536,6 +540,8 @@ static void srcu_read_delay(struct rcu_random_state *rrsp)
>  	delay = rcu_random(rrsp) % (nrealreaders * 2 * longdelay * uspertick);
>  	if (!delay)
>  		schedule_timeout_interruptible(longdelay);
> +	else
> +		rcu_read_delay(rrsp);
>  }
> 
>  static void srcu_torture_read_unlock(int idx) __releases(&srcu_ctl)
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/

Re: [PATCH] rcutorture: add random preemption

[Peter Zijlstra]

On Mon, 2010-06-21 at 09:43 -0700, Paul E. McKenney wrote:
> > +#ifdef CONFIG_PREEMPT
> > +     if (!preempt_count() && !(rcu_random(rrsp) % (nrealreaders * 20000)))
> > +             preempt_schedule();
> > +#endif
> 
> This one scared me for a bit -- then I realized that preempt_schedule()
> won't actually schedule if preemption is in any way disabled.  So the
> above really is OK, because Classic RCU and RCU-bh disable preemption.
> 
> So, should we have a comment to this effect, or is my hypersensitivity to
> RCU semantics unique to me? 

Well it seems to do a !preempt_count() test too, so I wouldn't worry too
much about it, still using preempt_schedule() doesn't seem right, why
not use cond_resched()?

Re: [PATCH] rcutorture: add random preemption

[Paul E. McKenney]

On Mon, Jun 21, 2010 at 06:50:11PM +0200, Peter Zijlstra wrote:
> On Mon, 2010-06-21 at 09:43 -0700, Paul E. McKenney wrote:
> > > +#ifdef CONFIG_PREEMPT
> > > +     if (!preempt_count() && !(rcu_random(rrsp) % (nrealreaders * 20000)))
> > > +             preempt_schedule();
> > > +#endif
> > 
> > This one scared me for a bit -- then I realized that preempt_schedule()
> > won't actually schedule if preemption is in any way disabled.  So the
> > above really is OK, because Classic RCU and RCU-bh disable preemption.
> > 
> > So, should we have a comment to this effect, or is my hypersensitivity to
> > RCU semantics unique to me? 
> 
> Well it seems to do a !preempt_count() test too, so I wouldn't worry too
> much about it, still using preempt_schedule() doesn't seem right, why
> not use cond_resched()?

My guess is that Lai wants to force a call to rcu_sched_qs() even if
!need_resched().  One reason for doing this would be to put more stress
on preemptible RCU's handling of context switches in RCU read-side
critical sections.

Another approach would be to have some high-priority threads that wake up
periodically, as I plan to make rcutorture do for RCU priority boosting.
Of course, this is not necessarily mutually exclusive with Lai's approach.

							Thanx, Paul

Re: [PATCH] rcutorture: add random preemption

[Peter Zijlstra]

On Mon, 2010-06-21 at 10:05 -0700, Paul E. McKenney wrote:
> On Mon, Jun 21, 2010 at 06:50:11PM +0200, Peter Zijlstra wrote:
> > On Mon, 2010-06-21 at 09:43 -0700, Paul E. McKenney wrote:
> > > > +#ifdef CONFIG_PREEMPT
> > > > +     if (!preempt_count() && !(rcu_random(rrsp) % (nrealreaders * 20000)))
> > > > +             preempt_schedule();
> > > > +#endif
> > > 
> > > This one scared me for a bit -- then I realized that preempt_schedule()
> > > won't actually schedule if preemption is in any way disabled.  So the
> > > above really is OK, because Classic RCU and RCU-bh disable preemption.
> > > 
> > > So, should we have a comment to this effect, or is my hypersensitivity to
> > > RCU semantics unique to me? 
> > 
> > Well it seems to do a !preempt_count() test too, so I wouldn't worry too
> > much about it, still using preempt_schedule() doesn't seem right, why
> > not use cond_resched()?
> 
> My guess is that Lai wants to force a call to rcu_sched_qs() even if
> !need_resched(). 

Ah, d'0h yeah, I'm not having the best of days.. monday or something :-)