public inbox for linux-kernel@vger.kernel.org
 help / color / mirror / Atom feed
* rcu: Avoid irq disable in rcu_cpu_kthread
@ 2013-12-05 21:06 Christoph Lameter
  2013-12-10  0:05 ` Paul E. McKenney
  0 siblings, 1 reply; 3+ messages in thread
From: Christoph Lameter @ 2013-12-05 21:06 UTC (permalink / raw)
  To: Paul E. McKenney; +Cc: linux-kernel

Once we have the per cpu patchset merged we could do the following [it
even works without that patchset but the __this_cpu ops will not do
preemption checks]. Would this work?


Subject: rcu: Avoid irq disable in rcu_cpu_kthread

The use of this_cpu ops avoids numerous address calculations
and allows to avoid the irq enable/disable sequence through a
low latency non locking this_cpu_xchg.

Signed-off-by: Christoph Lameter <cl@linux.com>

Index: linux/kernel/rcu/tree_plugin.h
===================================================================
--- linux.orig/kernel/rcu/tree_plugin.h	2013-12-03 11:32:23.322999660 -0600
+++ linux/kernel/rcu/tree_plugin.h	2013-12-03 11:32:23.312999941 -0600
@@ -1417,33 +1417,29 @@ static int rcu_cpu_kthread_should_run(un
  */
 static void rcu_cpu_kthread(unsigned int cpu)
 {
-	unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
-	char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
+	char work;
 	int spincnt;

 	for (spincnt = 0; spincnt < 10; spincnt++) {
 		trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
 		local_bh_disable();
-		*statusp = RCU_KTHREAD_RUNNING;
-		this_cpu_inc(rcu_cpu_kthread_loops);
-		local_irq_disable();
-		work = *workp;
-		*workp = 0;
-		local_irq_enable();
+		__this_cpu_write(rcu_cpu_kthread_status, RCU_KTHREAD_RUNNING);
+		__this_cpu_inc(rcu_cpu_kthread_loops);
+		work = this_cpu_xchg(rcu_cpu_has_work, 0);
 		if (work)
 			rcu_kthread_do_work();
 		local_bh_enable();
-		if (*workp == 0) {
+		if (__this_cpu_read(rcu_cpu_has_work) == 0) {
 			trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
-			*statusp = RCU_KTHREAD_WAITING;
+			__this_cpu_write(rcu_cpu_kthread_status, RCU_KTHREAD_WAITING);
 			return;
 		}
 	}
-	*statusp = RCU_KTHREAD_YIELDING;
+	__this_cpu_write(rcu_cpu_kthread_status, RCU_KTHREAD_YIELDING);
 	trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
 	schedule_timeout_interruptible(2);
 	trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
-	*statusp = RCU_KTHREAD_WAITING;
+	__this_cpu_write(rcu_cpu_kthread_status, RCU_KTHREAD_WAITING);
 }

 /*

^ permalink raw reply	[flat|nested] 3+ messages in thread

* Re: rcu: Avoid irq disable in rcu_cpu_kthread
  2013-12-05 21:06 rcu: Avoid irq disable in rcu_cpu_kthread Christoph Lameter
@ 2013-12-10  0:05 ` Paul E. McKenney
  2013-12-10 15:24   ` Christoph Lameter
  0 siblings, 1 reply; 3+ messages in thread
From: Paul E. McKenney @ 2013-12-10  0:05 UTC (permalink / raw)
  To: Christoph Lameter; +Cc: linux-kernel

On Thu, Dec 05, 2013 at 09:06:55PM +0000, Christoph Lameter wrote:
> Once we have the per cpu patchset merged we could do the following [it
> even works without that patchset but the __this_cpu ops will not do
> preemption checks]. Would this work?

Looks plausible at first glance.  But are you really seeing performance
issues with this code?  It is only compiled into the kernel when you build
with CONFIG_RCU_BOOST=y -- are you actually using that for your workloads?

							Thanx, Paul

> Subject: rcu: Avoid irq disable in rcu_cpu_kthread
> 
> The use of this_cpu ops avoids numerous address calculations
> and allows to avoid the irq enable/disable sequence through a
> low latency non locking this_cpu_xchg.
> 
> Signed-off-by: Christoph Lameter <cl@linux.com>
> 
> Index: linux/kernel/rcu/tree_plugin.h
> ===================================================================
> --- linux.orig/kernel/rcu/tree_plugin.h	2013-12-03 11:32:23.322999660 -0600
> +++ linux/kernel/rcu/tree_plugin.h	2013-12-03 11:32:23.312999941 -0600
> @@ -1417,33 +1417,29 @@ static int rcu_cpu_kthread_should_run(un
>   */
>  static void rcu_cpu_kthread(unsigned int cpu)
>  {
> -	unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
> -	char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
> +	char work;
>  	int spincnt;
> 
>  	for (spincnt = 0; spincnt < 10; spincnt++) {
>  		trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
>  		local_bh_disable();
> -		*statusp = RCU_KTHREAD_RUNNING;
> -		this_cpu_inc(rcu_cpu_kthread_loops);
> -		local_irq_disable();
> -		work = *workp;
> -		*workp = 0;
> -		local_irq_enable();
> +		__this_cpu_write(rcu_cpu_kthread_status, RCU_KTHREAD_RUNNING);
> +		__this_cpu_inc(rcu_cpu_kthread_loops);
> +		work = this_cpu_xchg(rcu_cpu_has_work, 0);
>  		if (work)
>  			rcu_kthread_do_work();
>  		local_bh_enable();
> -		if (*workp == 0) {
> +		if (__this_cpu_read(rcu_cpu_has_work) == 0) {
>  			trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
> -			*statusp = RCU_KTHREAD_WAITING;
> +			__this_cpu_write(rcu_cpu_kthread_status, RCU_KTHREAD_WAITING);
>  			return;
>  		}
>  	}
> -	*statusp = RCU_KTHREAD_YIELDING;
> +	__this_cpu_write(rcu_cpu_kthread_status, RCU_KTHREAD_YIELDING);
>  	trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
>  	schedule_timeout_interruptible(2);
>  	trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
> -	*statusp = RCU_KTHREAD_WAITING;
> +	__this_cpu_write(rcu_cpu_kthread_status, RCU_KTHREAD_WAITING);
>  }
> 
>  /*
> 


^ permalink raw reply	[flat|nested] 3+ messages in thread

* Re: rcu: Avoid irq disable in rcu_cpu_kthread
  2013-12-10  0:05 ` Paul E. McKenney
@ 2013-12-10 15:24   ` Christoph Lameter
  0 siblings, 0 replies; 3+ messages in thread
From: Christoph Lameter @ 2013-12-10 15:24 UTC (permalink / raw)
  To: Paul E. McKenney; +Cc: linux-kernel

On Mon, 9 Dec 2013, Paul E. McKenney wrote:

> On Thu, Dec 05, 2013 at 09:06:55PM +0000, Christoph Lameter wrote:
> > Once we have the per cpu patchset merged we could do the following [it
> > even works without that patchset but the __this_cpu ops will not do
> > preemption checks]. Would this work?
>
> Looks plausible at first glance.  But are you really seeing performance
> issues with this code?  It is only compiled into the kernel when you build
> with CONFIG_RCU_BOOST=y -- are you actually using that for your workloads?

I have not done any benchmarking. Just looking for more use cases for the
this_cpu ops. There is a lot of use of per cpu operations in the rcu code
which seems to be areas in which these operations can help.


^ permalink raw reply	[flat|nested] 3+ messages in thread

end of thread, other threads:[~2013-12-10 15:24 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2013-12-05 21:06 rcu: Avoid irq disable in rcu_cpu_kthread Christoph Lameter
2013-12-10  0:05 ` Paul E. McKenney
2013-12-10 15:24   ` Christoph Lameter

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox