2011-06-10 07:50:56

by Shaohua Li

Subject: [PATCH]rcu: avoid unnecessary thread wakeup

invoke_rcu_cpu_kthread could be called in the thread itself. In this case,
we don't need call wakeup, which is just wasting CPU.

Signed-off-by: Shaohua Li <[email protected]>

diff --git a/kernel/rcutree.c b/kernel/rcutree.c
index 89419ff..f9bd051 100644
--- a/kernel/rcutree.c
+++ b/kernel/rcutree.c
@@ -1475,7 +1475,8 @@ static void invoke_rcu_cpu_kthread(void)
local_irq_restore(flags);
return;
}
- wake_up_process(__this_cpu_read(rcu_cpu_kthread_task));
+ if (current != __this_cpu_read(rcu_cpu_kthread_task))
+ wake_up_process(__this_cpu_read(rcu_cpu_kthread_task));
local_irq_restore(flags);
}

2011-06-10 16:38:49

by Paul E. McKenney

Subject: Re: [PATCH]rcu: avoid unnecessary thread wakeup

On Fri, Jun 10, 2011 at 03:50:51PM +0800, Shaohua Li wrote:
> invoke_rcu_cpu_kthread could be called in the thread itself. In this case,
> we don't need call wakeup, which is just wasting CPU.
>
> Signed-off-by: Shaohua Li <[email protected]>
>
> diff --git a/kernel/rcutree.c b/kernel/rcutree.c
> index 89419ff..f9bd051 100644
> --- a/kernel/rcutree.c
> +++ b/kernel/rcutree.c
> @@ -1475,7 +1475,8 @@ static void invoke_rcu_cpu_kthread(void)
> local_irq_restore(flags);
> return;
> }
> - wake_up_process(__this_cpu_read(rcu_cpu_kthread_task));
> + if (current != __this_cpu_read(rcu_cpu_kthread_task))
> + wake_up_process(__this_cpu_read(rcu_cpu_kthread_task));
> local_irq_restore(flags);
> }

Excellent point, thank you!

But how about combining the tests, perhaps something like the
following?

Unless you have objections or spot problems with it (or it breaks during
testing), I will queue the patch below with your SOB, since I derived
it from your patch.

Thanx, Paul

------------------------------------------------------------------------

diff --git a/kernel/rcutree.c b/kernel/rcutree.c
index b4b254d..eda3986 100644
--- a/kernel/rcutree.c
+++ b/kernel/rcutree.c
@@ -1523,11 +1523,9 @@ static void invoke_rcu_cpu_kthread(void)

local_irq_save(flags);
__this_cpu_write(rcu_cpu_has_work, 1);
- if (__this_cpu_read(rcu_cpu_kthread_task) == NULL) {
- local_irq_restore(flags);
- return;
- }
- wake_up_process(__this_cpu_read(rcu_cpu_kthread_task));
+ if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
+ current != __this_cpu_read(rcu_cpu_kthread_task))
+ wake_up_process(__this_cpu_read(rcu_cpu_kthread_task));
local_irq_restore(flags);
}

2011-06-13 00:36:54

by Shaohua Li

Subject: Re: [PATCH]rcu: avoid unnecessary thread wakeup

On Sat, 2011-06-11 at 00:38 +0800, Paul E. McKenney wrote:
> On Fri, Jun 10, 2011 at 03:50:51PM +0800, Shaohua Li wrote:
> > invoke_rcu_cpu_kthread could be called in the thread itself. In this case,
> > we don't need call wakeup, which is just wasting CPU.
> >
> > Signed-off-by: Shaohua Li <[email protected]>
> >
> > diff --git a/kernel/rcutree.c b/kernel/rcutree.c
> > index 89419ff..f9bd051 100644
> > --- a/kernel/rcutree.c
> > +++ b/kernel/rcutree.c
> > @@ -1475,7 +1475,8 @@ static void invoke_rcu_cpu_kthread(void)
> > local_irq_restore(flags);
> > return;
> > }
> > - wake_up_process(__this_cpu_read(rcu_cpu_kthread_task));
> > + if (current != __this_cpu_read(rcu_cpu_kthread_task))
> > + wake_up_process(__this_cpu_read(rcu_cpu_kthread_task));
> > local_irq_restore(flags);
> > }
>
> Excellent point, thank you!
>
> But how about combining the tests, perhaps something like the
> following?
>
> Unless you have objections or spot problems with it (or it breaks during
> testing), I will queue the patch below with your SOB, since I derived
> it from your patch.
that's better, thanks.

Thanks,
Shaohua